Imported Upstream version 1.20.0upstream/1.20.0submit/tizen/20220415.103159

149 files changed, 7681 insertions, 289 deletions

diff --git a/compiler/luci-interpreter/README.md b/compiler/luci-interpreter/README.md
index 4a9a34e6d..77ec5c81c 100644
--- a/compiler/luci-interpreter/README.md
+++ b/compiler/luci-interpreter/README.md
@@ -111,7 +111,7 @@ Note that one memory manager could be shared between multiple interpreter instan
 
 List of predefined memory managers:
 - `SimpleMemoryManager` This is a simple wrapper around new/delete, default one.
-- `TestMemoryManager` Memorizes all allocated memory and releases it in Manager desctuctor, used in kernel unit tests.
+- `TestMemoryManager` Memorizes all allocated memory and releases it in Manager destructor, used in kernel unit tests.
 - `BuddyMemoryManager` Implements Buddy algorithm, uses external buffer for tensor data allocations, does not need new/delete.
 - `StaticMemoryManger` Uses precomputed memory allocation plan. Requires preparation with MemoryPlanner, but could reduce memory consumption in restricted environments (like MCUs).
 
diff --git a/compiler/luci-interpreter/include/luci_interpreter/GraphBuilderRegistry.h b/compiler/luci-interpreter/include/luci_interpreter/GraphBuilderRegistry.h
new file mode 100644
index 000000000..375b1ae20
--- /dev/null
+++ b/compiler/luci-interpreter/include/luci_interpreter/GraphBuilderRegistry.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_INTERPRETER_GRAPH_BUILDER_REGISTRY__
+#define __LUCI_INTERPRETER_GRAPH_BUILDER_REGISTRY__
+
+#include <luci/Import/GraphBuilderRegistry.h>
+
+namespace luci_interpreter
+{
+
+/**
+ * @brief Creates and returns GraphBuilderSource, which allows to not copy constant buffers from
+ * model's file.
+ *
+ * @warning Use this source only in case when model's buffer alive longer than Interpreter.
+ */
+std::unique_ptr<luci::GraphBuilderSource> source_without_constant_copying();
+
+} // namespace luci_interpreter
+
+#endif // __LUCI_INTERPRETER_GRAPH_BUILDER_REGISTRY__
diff --git a/compiler/luci-interpreter/include/luci_interpreter/Interpreter.h b/compiler/luci-interpreter/include/luci_interpreter/Interpreter.h
index 7dee8a7f2..8e2f457a5 100644
--- a/compiler/luci-interpreter/include/luci_interpreter/Interpreter.h
+++ b/compiler/luci-interpreter/include/luci_interpreter/Interpreter.h
@@ -50,7 +50,9 @@ public:
 class Interpreter
 {
 public:
-  explicit Interpreter(const luci::Module *module, IMemoryManager *memory_manager = nullptr);
+  explicit Interpreter(const luci::Module *module);
+
+  explicit Interpreter(const luci::Module *module, IMemoryManager *memory_manager);
 
   ~Interpreter();
 
@@ -69,7 +71,6 @@ private:
   // the order of deletion in the destructor
   std::unique_ptr<IMemoryManager> _default_memory_manager = nullptr;
   std::unique_ptr<class RuntimeModule> _runtime_module;
-  IMemoryManager *_memory_manager = nullptr;
 
   // Observer functionality support.
   std::unique_ptr<struct RuntimeToIR> _runtime_to_ir;
diff --git a/compiler/luci-interpreter/pal/cmsisnn/KernelsToBuild.lst b/compiler/luci-interpreter/pal/cmsisnn/KernelsToBuild.lst
index 771974afe..d134a6b95 100644
--- a/compiler/luci-interpreter/pal/cmsisnn/KernelsToBuild.lst
+++ b/compiler/luci-interpreter/pal/cmsisnn/KernelsToBuild.lst
@@ -7,9 +7,11 @@ REGISTER_KERNEL(Concatenation)
 REGISTER_KERNEL(Conv2D)
 REGISTER_KERNEL(DepthToSpace)
 REGISTER_KERNEL(DepthwiseConv2D)
+REGISTER_KERNEL(Dequantize)
 REGISTER_KERNEL(Div)
 REGISTER_KERNEL(Elu)
 REGISTER_KERNEL(Exp)
+REGISTER_KERNEL(ExpandDims)
 REGISTER_KERNEL(Floor)
 REGISTER_KERNEL(FloorDiv)
 REGISTER_KERNEL(Equal)
@@ -37,6 +39,7 @@ REGISTER_KERNEL(NotEqual)
 REGISTER_KERNEL(Pad)
 REGISTER_KERNEL(PadV2)
 REGISTER_KERNEL(PRelu)
+REGISTER_KERNEL(Quantize)
 REGISTER_KERNEL(Reshape)
 REGISTER_KERNEL(ResizeBilinear)
 REGISTER_KERNEL(ResizeNearestNeighbor)
@@ -50,6 +53,7 @@ REGISTER_KERNEL(Square)
 REGISTER_KERNEL(SquaredDifference)
 REGISTER_KERNEL(Squeeze)
 REGISTER_KERNEL(Sub)
+REGISTER_KERNEL(SVDF)
 REGISTER_KERNEL(Tanh)
 REGISTER_KERNEL(Transpose)
 REGISTER_KERNEL(TransposeConv)
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALAveragePool2d.h b/compiler/luci-interpreter/pal/cmsisnn/PALAveragePool2d.h
new file mode 100644
index 000000000..a274afb7e
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALAveragePool2d.h
@@ -0,0 +1,124 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+#define LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h>
+#include <tensorflow/lite/kernels/internal/reference/pooling.h>
+#include <arm_nn_types.h>
+#include <arm_nnfunctions.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void AveragePool(const tflite::PoolParams &params,
+                               const tflite::RuntimeShape &input_shape, const T *input_data,
+                               const tflite::RuntimeShape &output_shape, T *output_data,
+                               const tflite::RuntimeShape &scratchpad_shape, T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation is not supported
+    assert(false && "AveragePool NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void AveragePool<int8_t>(const tflite::PoolParams &params,
+                                const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                                const tflite::RuntimeShape &output_shape, int8_t *output_data,
+                                const tflite::RuntimeShape &scratchpad_shape,
+                                int8_t *scratchpad_data)
+{
+  assert(input_shape.DimensionsCount() == 4);
+  assert(output_shape.DimensionsCount() == 4);
+  assert(scratchpad_data != nullptr);
+
+  const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+  assert(batches == 1);
+
+  const int depth = tflite::MatchingDim(input_shape, 3, output_shape, 3);
+
+  cmsis_nn_dims input_dims;
+  input_dims.n = 1;
+  input_dims.h = input_shape.Dims(1);
+  input_dims.w = input_shape.Dims(2);
+  input_dims.c = depth;
+
+  cmsis_nn_dims output_dims;
+  output_dims.n = 1;
+  output_dims.h = output_shape.Dims(1);
+  output_dims.w = output_shape.Dims(2);
+  output_dims.c = depth;
+
+  cmsis_nn_pool_params pool_params;
+  pool_params.stride.h = params.stride_height;
+  pool_params.stride.w = params.stride_width;
+  pool_params.padding.h = params.padding_values.height;
+  pool_params.padding.w = params.padding_values.width;
+  pool_params.activation.min = params.quantized_activation_min;
+  pool_params.activation.max = params.quantized_activation_max;
+
+  cmsis_nn_dims filter_dims;
+  filter_dims.n = 1;
+  filter_dims.h = params.filter_height;
+  filter_dims.w = params.filter_width;
+  filter_dims.c = 1;
+
+  cmsis_nn_context ctx;
+  ctx.buf = scratchpad_data;
+  ctx.size = scratchpad_shape.Dims(0);
+  auto res = arm_avgpool_s8(&ctx, &pool_params, &input_dims, input_data, &filter_dims, &output_dims,
+                            output_data);
+  assert(res == ARM_MATH_SUCCESS);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &output_shape)
+
+{
+  if (input_data_type == luci_interpreter::DataType::S8)
+  {
+    assert(input_shape.DimensionsCount() == 4);
+    assert(output_shape.DimensionsCount() == 4);
+
+    const int32_t output_width = output_shape.Dims(2);
+    const int32_t depth = tflite::MatchingDim(input_shape, 3, output_shape, 3);
+
+    const int32_t buf_size = arm_avgpool_s8_get_buffer_size(output_width, depth);
+    auto data_type_size = static_cast<int32_t>(luci_interpreter::getDataTypeSize(input_data_type));
+
+    luci_interpreter::Shape scratchpad_shape{buf_size * data_type_size};
+    scratchpad->resize(scratchpad_shape);
+  }
+  else
+  {
+    scratchpad->set_allocatable(false);
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALConv2d.h b/compiler/luci-interpreter/pal/cmsisnn/PALConv2d.h
index 0a8ae4e48..cfb84ea60 100644
--- a/compiler/luci-interpreter/pal/cmsisnn/PALConv2d.h
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALConv2d.h
@@ -19,6 +19,8 @@
 
 #include <tensorflow/lite/kernels/internal/reference/conv.h>
 #include <tensorflow/lite/kernels/internal/reference/integer_ops/conv.h>
+#include <arm_nn_types.h>
+#include <arm_nnfunctions.h>
 
 namespace luci_interpreter_pal
 {
@@ -26,11 +28,11 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const float *input_data, const tflite::RuntimeShape &filter_shape,
                         const float *filter_data, const tflite::RuntimeShape &bias_shape,
                         const float *bias_data, const tflite::RuntimeShape &output_shape,
-                        float *output_data, const tflite::RuntimeShape &im2col_shape,
-                        float *im2col_data)
+                        float *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        float *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
                               bias_shape, bias_data, output_shape, output_data,
                               tflite::RuntimeShape(), nullptr);
@@ -40,14 +42,14 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const uint8 *input_data, const tflite::RuntimeShape &filter_shape,
                         const uint8 *filter_data, const tflite::RuntimeShape &bias_shape,
                         const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                        uint8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                        uint8 *im2col_data)
+                        uint8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        uint8 *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
-                              bias_shape, bias_data, output_shape, output_data, im2col_shape,
-                              im2col_data, nullptr);
+                              bias_shape, bias_data, output_shape, output_data, scratchpad_shape,
+                              scratchpad_data, nullptr);
 }
 
 static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_t *mult,
@@ -55,14 +57,141 @@ static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_
                                   const int8 *input_data, const tflite::RuntimeShape &filter_shape,
                                   const int8 *filter_data, const tflite::RuntimeShape &bias_shape,
                                   const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                                  int8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                                  int8 *im2col_data)
+                                  int8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                                  int8 *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
-  tflite::reference_integer_ops::ConvPerChannel(params, mult, shifts, input_shape, input_data,
-                                                filter_shape, filter_data, bias_shape, bias_data,
-                                                output_shape, output_data);
+  if (scratchpad_data)
+  {
+    cmsis_nn_conv_params conv_params;
+    conv_params.dilation.h = params.dilation_height_factor;
+    conv_params.dilation.w = params.dilation_width_factor;
+
+    assert(conv_params.dilation.h == 1);
+    assert(conv_params.dilation.w == 1);
+
+    conv_params.input_offset = params.input_offset;
+    conv_params.output_offset = params.output_offset;
+    conv_params.stride.h = params.stride_height;
+    conv_params.stride.w = params.stride_width;
+    conv_params.padding.h = params.padding_values.height;
+    conv_params.padding.w = params.padding_values.width;
+    conv_params.activation.min = params.quantized_activation_min;
+    conv_params.activation.max = params.quantized_activation_max;
+
+    cmsis_nn_per_channel_quant_params quant_params;
+    quant_params.multiplier = const_cast<int32_t *>(mult);
+    quant_params.shift = const_cast<int32_t *>(shifts);
+
+    assert(conv_params.activation.min <= conv_params.activation.max);
+    assert(input_shape.DimensionsCount() == 4);
+    assert(filter_shape.DimensionsCount() == 4);
+    assert(output_shape.DimensionsCount() == 4);
+    const int batch_size = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int input_depth = tflite::MatchingDim(input_shape, 3, filter_shape, 3);
+    const int output_depth = tflite::MatchingDim(filter_shape, 0, output_shape, 3);
+    if (bias_data)
+    {
+      assert(bias_shape.FlatSize() == output_depth);
+    }
+
+    cmsis_nn_dims input_dims;
+    input_dims.n = batch_size;
+    input_dims.h = input_shape.Dims(1);
+    input_dims.w = input_shape.Dims(2);
+    input_dims.c = input_depth;
+
+    cmsis_nn_dims filter_dims;
+    filter_dims.n = output_depth;
+    filter_dims.h = filter_shape.Dims(1);
+    filter_dims.w = filter_shape.Dims(2);
+    filter_dims.c = input_depth;
+
+    cmsis_nn_dims bias_dims;
+    bias_dims.n = 1;
+    bias_dims.h = 1;
+    bias_dims.w = 1;
+    bias_dims.c = output_depth;
+
+    cmsis_nn_dims output_dims;
+    output_dims.n = batch_size;
+    output_dims.h = output_shape.Dims(1);
+    output_dims.w = output_shape.Dims(2);
+    output_dims.c = output_depth;
+
+    cmsis_nn_context ctx;
+    ctx.buf = scratchpad_data;
+    ctx.size = scratchpad_shape.Dims(0);
+
+    auto res = arm_convolve_wrapper_s8(&ctx, &conv_params, &quant_params, &input_dims, input_data,
+                                       &filter_dims, filter_data, &bias_dims, bias_data,
+                                       &output_dims, output_data);
+    assert(res == ARM_MATH_SUCCESS);
+  }
+  else
+  {
+    tflite::reference_integer_ops::ConvPerChannel(params, mult, shifts, input_shape, input_data,
+                                                  filter_shape, filter_data, bias_shape, bias_data,
+                                                  output_shape, output_data);
+  }
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::ConvParams &params,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+{
+  cmsis_nn_conv_params conv_params;
+  conv_params.dilation.h = params.dilation_height_factor;
+  conv_params.dilation.w = params.dilation_width_factor;
+
+  if (input_data_type == loco::DataType::S8 && conv_params.dilation.h == 1 &&
+      conv_params.dilation.w == 1)
+  {
+    const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int32_t input_depth = tflite::MatchingDim(input_shape, 3, filter_shape, 3);
+    const int32_t output_depth = tflite::MatchingDim(filter_shape, 0, output_shape, 3);
+    const int32_t filter_height = filter_shape.Dims(1);
+    const int32_t filter_width = filter_shape.Dims(2);
+    const int32_t output_height = output_shape.Dims(1);
+    const int32_t output_width = output_shape.Dims(2);
+
+    conv_params.input_offset = params.input_offset;
+    conv_params.output_offset = params.output_offset;
+    conv_params.stride.h = params.stride_height;
+    conv_params.stride.w = params.stride_width;
+    conv_params.padding.h = params.padding_values.height;
+    conv_params.padding.w = params.padding_values.width;
+
+    cmsis_nn_dims input_dims;
+    input_dims.n = batches;
+    input_dims.h = input_shape.Dims(1);
+    input_dims.w = input_shape.Dims(2);
+    input_dims.c = input_depth;
+
+    cmsis_nn_dims filter_dims;
+    filter_dims.n = output_depth;
+    filter_dims.h = filter_height;
+    filter_dims.w = filter_width;
+    filter_dims.c = input_depth;
+
+    cmsis_nn_dims output_dims;
+    output_dims.n = batches;
+    output_dims.h = output_height;
+    output_dims.w = output_width;
+    output_dims.c = output_depth;
+
+    const int32_t buf_size = arm_convolve_wrapper_s8_get_buffer_size(&conv_params, &input_dims,
+                                                                     &filter_dims, &output_dims);
+
+    luci_interpreter::Shape scratchpad_shape{buf_size};
+    scratchpad->resize(scratchpad_shape);
+  }
+  else
+  {
+    scratchpad->set_allocatable(false);
+  }
 }
 
 } // namespace luci_interpreter_pal
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALDepthwiseConv2d.h b/compiler/luci-interpreter/pal/cmsisnn/PALDepthwiseConv2d.h
new file mode 100644
index 000000000..120dcd803
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALDepthwiseConv2d.h
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+#define LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h>
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h>
+#include <arm_nnfunctions.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void
+DepthwiseConvPerChannel(const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+                        const int32_t *output_shift, const tflite::RuntimeShape &input_shape,
+                        const T *input_data, const tflite::RuntimeShape &filter_shape,
+                        const T *filter_data, const tflite::RuntimeShape &bias_shape,
+                        const int32_t *bias_data, const tflite::RuntimeShape &output_shape,
+                        T *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation is not supported
+    assert(false && "DepthwiseConvPerChannel NYI");
+    (void)params;
+    (void)output_multiplier;
+    (void)output_shift;
+    (void)input_shape;
+    (void)output_data;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void DepthwiseConvPerChannel<int8_t>(
+  const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+  const int32_t *output_shift, const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+  const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+  const tflite::RuntimeShape &output_shape, int8_t *output_data,
+  const tflite::RuntimeShape &scratchpad_shape, int8_t *scratchpad_data)
+{
+  if (scratchpad_data)
+  {
+    cmsis_nn_dw_conv_params dw_conv_params;
+    dw_conv_params.dilation.h = params.dilation_height_factor;
+    dw_conv_params.dilation.w = params.dilation_width_factor;
+    assert(dw_conv_params.dilation.h == 1);
+    assert(dw_conv_params.dilation.w == 1);
+
+    dw_conv_params.input_offset = params.input_offset;
+    dw_conv_params.output_offset = params.output_offset;
+    dw_conv_params.stride.h = params.stride_height;
+    dw_conv_params.stride.w = params.stride_width;
+    dw_conv_params.padding.h = params.padding_values.height;
+    dw_conv_params.padding.w = params.padding_values.width;
+
+    dw_conv_params.activation.min = params.quantized_activation_min;
+    dw_conv_params.activation.max = params.quantized_activation_max;
+    dw_conv_params.ch_mult = params.depth_multiplier;
+
+    cmsis_nn_per_channel_quant_params quant_params;
+    int32_t output_multiplier = params.output_multiplier;
+    int32_t output_shift = params.output_shift;
+
+    quant_params.multiplier = &output_multiplier;
+    quant_params.shift = &output_shift;
+
+    assert(dw_conv_params.activation.min <= dw_conv_params.activation.max);
+    const int batch_size = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = tflite::MatchingDim(filter_shape, 3, output_shape, 3);
+    if (bias_data)
+    {
+      assert(bias_shape.FlatSize() == output_depth);
+    }
+
+    cmsis_nn_dims input_dims;
+    input_dims.n = batch_size;
+    input_dims.h = input_shape.Dims(1);
+    input_dims.w = input_shape.Dims(2);
+    input_dims.c = input_shape.Dims(3);
+
+    cmsis_nn_dims filter_dims;
+    filter_dims.n = filter_shape.Dims(0);
+    filter_dims.h = filter_shape.Dims(1);
+    filter_dims.w = filter_shape.Dims(2);
+    filter_dims.c = output_depth;
+
+    cmsis_nn_dims bias_dims;
+    bias_dims.n = 1;
+    bias_dims.h = 1;
+    bias_dims.w = 1;
+    bias_dims.c = output_depth;
+
+    cmsis_nn_dims output_dims;
+    output_dims.n = batch_size;
+    output_dims.h = output_shape.Dims(1);
+    output_dims.w = output_shape.Dims(2);
+    output_dims.c = output_depth;
+
+    cmsis_nn_context ctx;
+    ctx.buf = scratchpad_data;
+    ctx.size = scratchpad_shape.Dims(0);
+
+    auto res = arm_depthwise_conv_wrapper_s8(&ctx, &dw_conv_params, &quant_params, &input_dims,
+                                             input_data, &filter_dims, filter_data, &bias_dims,
+                                             bias_data, &output_dims, output_data);
+    assert(res == ARM_MATH_SUCCESS);
+  }
+  else
+  {
+    tflite::reference_integer_ops::DepthwiseConvPerChannel(
+      params, output_multiplier, output_shift, input_shape, input_data, filter_shape, filter_data,
+      bias_shape, bias_data, output_shape, output_data);
+  }
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const tflite::DepthwiseParams &params,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+{
+  cmsis_nn_dw_conv_params dw_conv_params;
+  dw_conv_params.dilation.h = params.dilation_height_factor;
+  dw_conv_params.dilation.w = params.dilation_width_factor;
+
+  if (input_data_type == loco::DataType::S8 && dw_conv_params.dilation.h == 1 &&
+      dw_conv_params.dilation.w == 1)
+  {
+    const int batch_size = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = tflite::MatchingDim(filter_shape, 3, output_shape, 3);
+
+    cmsis_nn_dims input_dims;
+    input_dims.n = batch_size;
+    input_dims.h = input_shape.Dims(1);
+    input_dims.w = input_shape.Dims(2);
+    input_dims.c = input_shape.Dims(3);
+
+    cmsis_nn_dims filter_dims;
+    filter_dims.n = filter_shape.Dims(0);
+    filter_dims.h = filter_shape.Dims(1);
+    filter_dims.w = filter_shape.Dims(2);
+    filter_dims.c = output_depth;
+
+    cmsis_nn_dims output_dims;
+    output_dims.n = batch_size;
+    output_dims.h = output_shape.Dims(1);
+    output_dims.w = output_shape.Dims(2);
+    output_dims.c = output_depth;
+
+    const int32_t buf_size = arm_depthwise_conv_wrapper_s8_get_buffer_size(
+      &dw_conv_params, &input_dims, &filter_dims, &output_dims);
+
+    auto data_type_size = static_cast<int32_t>(luci_interpreter::getDataTypeSize(input_data_type));
+
+    luci_interpreter::Shape scratchpad_shape{buf_size * data_type_size};
+    scratchpad->resize(scratchpad_shape);
+  }
+  else
+  {
+    scratchpad->set_allocatable(false);
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALDequantize.h b/compiler/luci-interpreter/pal/cmsisnn/PALDequantize.h
new file mode 100644
index 000000000..15ff0327b
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALDequantize.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+#define LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/dequantize.h"
+#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
+
+namespace luci_interpreter_pal
+{
+
+template <typename T>
+static inline void Dequantize(tflite::DequantizationParams &params,
+                              const tflite::RuntimeShape &input_shape, const T *input_data,
+                              const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_integer_ops::Dequantize<T>(params, input_shape, input_data, output_shape,
+                                               output_data);
+}
+
+static inline void Dequantize(tflite::DequantizationParams &params,
+                              const tflite::RuntimeShape &input_shape, const uint8_t *input_data,
+                              const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_ops::Dequantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEQUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALFullyConnected.h b/compiler/luci-interpreter/pal/cmsisnn/PALFullyConnected.h
new file mode 100644
index 000000000..32e905761
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALFullyConnected.h
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+#define LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+
+#include <tensorflow/lite/kernels/internal/reference/fully_connected.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h>
+#include <arm_nnfunctions.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void FullyConnected(const tflite::FullyConnectedParams &params,
+                                  const tflite::RuntimeShape &input_shape, const T *input_data,
+                                  const tflite::RuntimeShape &filter_shape, const T *filter_data,
+                                  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                                  const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  {
+    // MARK: At this moment this operation doesn't support
+    assert(false && "FullyConnected NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+  }
+}
+
+template <>
+inline void
+FullyConnected<int8_t>(const tflite::FullyConnectedParams &params,
+                       const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                       const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+                       const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                       const tflite::RuntimeShape &output_shape, int8_t *output_data)
+{
+  assert(output_shape.DimensionsCount() == 2);
+
+  const int batches = output_shape.Dims(0);
+  const int output_depth = output_shape.Dims(1);
+
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+
+  cmsis_nn_fc_params fc_params;
+  fc_params.input_offset = params.input_offset;
+  fc_params.output_offset = params.output_offset;
+  fc_params.filter_offset = params.weights_offset;
+  fc_params.activation.min = params.quantized_activation_min;
+  fc_params.activation.max = params.quantized_activation_max;
+
+  cmsis_nn_per_tensor_quant_params quant_params;
+  quant_params.multiplier = params.output_multiplier;
+  quant_params.shift = params.output_shift;
+
+  cmsis_nn_dims input_dims;
+  input_dims.n = batches;
+  input_dims.h = 1;
+  input_dims.w = 1;
+  input_dims.c = accum_depth;
+
+  cmsis_nn_dims filter_dims;
+  filter_dims.n = accum_depth;
+  filter_dims.h = 1;
+  filter_dims.w = 1;
+  filter_dims.c = output_depth;
+
+  cmsis_nn_dims bias_dims;
+  bias_dims.n = 1;
+  bias_dims.h = 1;
+  bias_dims.w = 1;
+  bias_dims.c = output_depth;
+
+  cmsis_nn_dims output_dims;
+  output_dims.n = batches;
+  output_dims.h = 1;
+  output_dims.w = 1;
+  output_dims.c = output_depth;
+
+  int32_t buf_size = arm_fully_connected_s8_get_buffer_size(&filter_dims);
+  auto buffer = std::make_unique<int8_t[]>(buf_size);
+  assert(buffer != nullptr);
+
+  cmsis_nn_context ctx;
+  ctx.buf = buffer.get();
+  ctx.size = buf_size;
+
+  auto res =
+    arm_fully_connected_s8(&ctx, &fc_params, &quant_params, &input_dims, input_data, &filter_dims,
+                           filter_data, &bias_dims, bias_data, &output_dims, output_data);
+  assert(res == ARM_MATH_SUCCESS);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALMul.h b/compiler/luci-interpreter/pal/cmsisnn/PALMul.h
index 2b46b100c..347a97a83 100644
--- a/compiler/luci-interpreter/pal/cmsisnn/PALMul.h
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALMul.h
@@ -21,21 +21,21 @@
 
 namespace luci_interpreter_pal
 {
+template <typename T>
 static inline void Mul(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
-                       const float *input1_data, const tflite::RuntimeShape &input2_shape,
-                       const float *input2_data, const tflite::RuntimeShape &output_shape,
-                       float *output_data)
+                       const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                       const T *input2_data, const tflite::RuntimeShape &output_shape,
+                       T *output_data)
 {
   tflite::reference_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
                                             input2_data, output_shape, output_data);
 }
 
-static inline void BroadcastMul4DSlow(tflite::ArithmeticParams &params,
-                                      const tflite::RuntimeShape &input1_shape,
-                                      const float *input1_data,
-                                      const tflite::RuntimeShape &input2_shape,
-                                      const float *input2_data,
-                                      const tflite::RuntimeShape &output_shape, float *output_data)
+template <typename T>
+static inline void
+BroadcastMul4DSlow(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
+                   const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                   const T *input2_data, const tflite::RuntimeShape &output_shape, T *output_data)
 {
   tflite::reference_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
                                             input2_data, output_shape, output_data);
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALQuantize.h b/compiler/luci-interpreter/pal/cmsisnn/PALQuantize.h
new file mode 100644
index 000000000..6046789ae
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALQuantize.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_QUANTIZE_H
+#define LUCI_INTERPRETER_PAL_QUANTIZE_H
+
+#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void Quantize(tflite::QuantizationParams &params,
+                            const tflite::RuntimeShape &input_shape, const float *input_data,
+                            const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  tflite::reference_ops::AffineQuantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+template <typename Input, typename Output>
+static inline void Requantize(const Input *input_data, int32_t size,
+                              int32_t effective_scale_multiplier, int32_t effective_scale_shift,
+                              int32_t input_zero_point, int32_t output_zero_point,
+                              Output *output_data)
+{
+  tflite::reference_ops::Requantize(input_data, size, effective_scale_multiplier,
+                                    effective_scale_shift, input_zero_point, output_zero_point,
+                                    output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_QUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h b/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h
new file mode 100644
index 000000000..a4a5b2a78
--- /dev/null
+++ b/compiler/luci-interpreter/pal/cmsisnn/PALSVDF.h
@@ -0,0 +1,190 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_SVDF_H
+#define LUCI_INTERPRETER_PAL_SVDF_H
+
+#include <arm_nn_types.h>
+#include <arm_nnfunctions.h>
+
+namespace luci_interpreter_pal
+{
+static inline void
+IntegerSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+            const int8_t *input_data, const tflite::RuntimeShape &weight_feature_shape,
+            const int8_t *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+            const int16_t *weight_time_data, const tflite::RuntimeShape &bias_shape,
+            const int32_t *bias_data, int16_t *activation_state_data,
+            const tflite::RuntimeShape &output_shape, int8_t *output_data, int32_t *scratchpad_data,
+            int32_t *output_temp_data, int32_t scale_1_a, int scale_1_b, int32_t scale_2_a,
+            int scale_2_b, int32_t input_zp, int32_t output_zp)
+{
+  const int32_t rank = params.rank;
+  const int32_t batch_size = input_shape.Dims(0);
+  const int32_t num_filters = weight_feature_shape.Dims(0);
+  const int32_t memory_size = weight_time_shape.Dims(1);
+
+  cmsis_nn_dims input_dims;
+  input_dims.n = input_shape.Dims(0);
+  input_dims.h = input_shape.Dims(1);
+
+  cmsis_nn_dims weights_feature_dims;
+  weights_feature_dims.n = weight_feature_shape.Dims(0);
+  weights_feature_dims.h = weight_feature_shape.Dims(1);
+
+  cmsis_nn_dims weights_time_dims;
+  weights_time_dims.n = weight_time_shape.Dims(0);
+  weights_time_dims.h = weight_time_shape.Dims(1);
+
+  cmsis_nn_dims bias_dims;
+  bias_dims.n = bias_shape.Dims(0);
+
+  cmsis_nn_dims state_dims;
+  state_dims.n = batch_size;
+  state_dims.h = memory_size * num_filters;
+
+  cmsis_nn_dims output_dims;
+  output_dims.n = output_shape.Dims(0);
+  output_dims.h = output_shape.Dims(1);
+
+  cmsis_nn_svdf_params svdf_params;
+  svdf_params.rank = params.rank;
+  svdf_params.input_offset = input_zp;
+  svdf_params.output_offset = output_zp;
+
+  svdf_params.input_activation.min = INT16_MIN;
+  svdf_params.input_activation.max = INT16_MAX;
+
+  svdf_params.output_activation.min = INT8_MIN;
+  svdf_params.output_activation.max = INT8_MAX;
+
+  cmsis_nn_per_tensor_quant_params in_quant_params;
+  in_quant_params.multiplier = scale_1_a;
+  in_quant_params.shift = scale_1_b;
+
+  cmsis_nn_per_tensor_quant_params out_quant_params;
+  out_quant_params.multiplier = scale_2_a;
+  out_quant_params.shift = scale_2_b;
+
+  cmsis_nn_context scratch_ctx;
+  scratch_ctx.buf = scratchpad_data;
+
+  cmsis_nn_context scratch_output_ctx;
+  scratch_output_ctx.buf = output_temp_data;
+
+  arm_svdf_s8(&scratch_ctx, &scratch_output_ctx, &svdf_params, &in_quant_params, &out_quant_params,
+              &input_dims, input_data, &state_dims, activation_state_data, &weights_feature_dims,
+              weight_feature_data, &weights_time_dims, weight_time_data, &bias_dims, bias_data,
+              &output_dims, output_data);
+}
+static inline void
+FloatSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+          const float *input_data, const tflite::RuntimeShape &weight_feature_shape,
+          const float *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+          const float *weight_time_data, const tflite::RuntimeShape &bias_shape,
+          const float *bias_data, float *scratchpad_data, float *activation_state_data,
+          const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  const int32_t rank = params.rank;
+  const int32_t batch_size = input_shape.Dims(0);
+  const int32_t input_size = input_shape.Dims(1);
+  const int32_t num_filters = weight_feature_shape.Dims(0);
+  const int32_t num_units = num_filters / rank;
+  const int32_t memory_size = weight_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  {
+    float *new_state_start = activation_state_data;
+    const float *old_state_start = activation_state_data + 1;
+    const float *old_state_end = activation_state_data + batch_size * num_filters * memory_size;
+    while (old_state_start != old_state_end)
+    {
+      *new_state_start++ = *old_state_start++;
+    }
+  }
+
+  // Note: no need to clear the latest activation, matmul is not accumulative.
+
+  // Compute conv1d(inputs, weights_feature).
+  // The activation_state's rightmost column is used to save current cycle
+  // activation. This is achieved by starting at state_ptr[memory_size - 1] and
+  // having the stride equal to memory_size.
+
+  // Perform batched matrix vector multiply operation:
+  {
+    const float *matrix = weight_feature_data;
+    const float *vector = input_data;
+    float *result = &activation_state_data[memory_size - 1];
+    float *result_in_batch = result;
+    for (int i = 0; i < batch_size; ++i)
+    {
+      const float *matrix_ptr = matrix;
+      for (int j = 0; j < num_filters; ++j)
+      {
+        float dot_prod = 0.0f;
+        const float *vector_in_batch = vector + i * input_size;
+        for (int k = 0; k < input_size; ++k)
+        {
+          dot_prod += *matrix_ptr++ * *vector_in_batch++;
+        }
+        *result_in_batch = dot_prod;
+        result_in_batch += memory_size;
+      }
+    }
+  }
+
+  tflite::reference_ops::ApplyTimeWeightsBiasAndActivation(
+    batch_size, memory_size, num_filters, num_units, rank, weight_time_data, bias_data,
+    params.activation, activation_state_data, scratchpad_data, output_data);
+}
+
+static inline void SetupScratchpadTensor(
+  const luci_interpreter::DataType &input_data_type,
+  const luci_interpreter::DataType &weight_feature_data_type,
+  luci_interpreter::Tensor *scratchpad_1, luci_interpreter::Tensor *scratchpad_2,
+  luci_interpreter::Tensor *scratchpad_3, luci_interpreter::Tensor *scratchpad_4,
+  luci_interpreter::Tensor *scratchpad_5, luci_interpreter::Tensor *scratchpad_6,
+  const luci_interpreter::Shape input_shape, const luci_interpreter::Shape weight_time_shape,
+  const int32_t batch_size, const int32_t num_filters, const int32_t num_units)
+{
+  if (input_data_type == loco::DataType::FLOAT32 &&
+      (weight_feature_data_type == loco::DataType::S8 ||
+       weight_feature_data_type == loco::DataType::U8))
+  {
+    (void)input_shape;
+    (void)weight_time_shape;
+    (void)scratchpad_3;
+    (void)scratchpad_4;
+    (void)scratchpad_5;
+    (void)scratchpad_6;
+
+    throw std::runtime_error("Hybrid type is not supported for cmsisnn");
+  }
+
+  // Resize scratchpad_1 tensor
+  scratchpad_1->resize({batch_size, num_filters});
+
+  if (input_data_type == loco::DataType::S8)
+  {
+    // Resize scratchpad_2 for full_integer op
+    scratchpad_2->resize({batch_size, num_units});
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_SVDF_H
diff --git a/compiler/luci-interpreter/pal/cmsisnn/pal.cmake b/compiler/luci-interpreter/pal/cmsisnn/pal.cmake
index 9a25a3c5d..a68b363d9 100644
--- a/compiler/luci-interpreter/pal/cmsisnn/pal.cmake
+++ b/compiler/luci-interpreter/pal/cmsisnn/pal.cmake
@@ -42,9 +42,12 @@ macro(add_pal_to_target TGT)
             "${TensorFlowSource_DIR}")
     target_include_directories(${TGT} PRIVATE ${LUCI_INTERPRETER_PAL_DIR})
 
-    set(PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc)
+    file(GLOB_RECURSE PAL_SOURCES "${CMSISSource_DIR}/CMSIS/NN/Source/*.c")
+    list(APPEND PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/tensor_utils.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc)
     add_library(luci_interpreter_cmsisnn_pal STATIC ${PAL_SOURCES})
-    set_target_properties(luci_interpreter_cmsisnn_pal PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    set_property(TARGET luci_interpreter_cmsisnn_pal PROPERTY POSITION_INDEPENDENT_CODE ON)
     target_include_directories(luci_interpreter_cmsisnn_pal PRIVATE
             "${TensorFlowRuySource_DIR}"
             "${TensorFlowGEMMLowpSource_DIR}"
@@ -53,7 +56,7 @@ macro(add_pal_to_target TGT)
     )
 
     add_subdirectory(${CMSISSource_DIR}/CMSIS/NN ${CMAKE_CURRENT_BINARY_DIR}/CMSISNN)
-    target_include_directories(luci_interpreter_cmsisnn_pal PRIVATE
+    target_include_directories(luci_interpreter_cmsisnn_pal PUBLIC
             "${CMSISSource_DIR}/CMSIS/NN/Include"
             "${CMSISSource_DIR}/CMSIS/DSP/Include"
             "${CMSISSource_DIR}/CMSIS/Core/Include")
diff --git a/compiler/luci-interpreter/pal/linux/KernelsToBuild.lst b/compiler/luci-interpreter/pal/linux/KernelsToBuild.lst
index 9d541276c..428b15ee0 100644
--- a/compiler/luci-interpreter/pal/linux/KernelsToBuild.lst
+++ b/compiler/luci-interpreter/pal/linux/KernelsToBuild.lst
@@ -1,19 +1,23 @@
 REGISTER_KERNEL(Add)
 REGISTER_KERNEL(ArgMax)
 REGISTER_KERNEL(AveragePool2D)
+REGISTER_KERNEL(BatchMatMul)
 REGISTER_KERNEL(BatchToSpaceND)
 REGISTER_KERNEL(Cast)
 REGISTER_KERNEL(Concatenation)
 REGISTER_KERNEL(Conv2D)
 REGISTER_KERNEL(DepthToSpace)
 REGISTER_KERNEL(DepthwiseConv2D)
+REGISTER_KERNEL(Dequantize)
 REGISTER_KERNEL(Div)
 REGISTER_KERNEL(Elu)
 REGISTER_KERNEL(Exp)
+REGISTER_KERNEL(ExpandDims)
 REGISTER_KERNEL(Floor)
 REGISTER_KERNEL(FloorDiv)
 REGISTER_KERNEL(Equal)
 REGISTER_KERNEL(FullyConnected)
+REGISTER_KERNEL(Gather)
 REGISTER_KERNEL(Greater)
 REGISTER_KERNEL(GreaterEqual)
 REGISTER_KERNEL(If)
@@ -37,11 +41,13 @@ REGISTER_KERNEL(MirrorPad)
 REGISTER_KERNEL(Mul)
 REGISTER_KERNEL(Neg)
 REGISTER_KERNEL(NotEqual)
+REGISTER_KERNEL(OneHot)
 REGISTER_KERNEL(Pack)
 REGISTER_KERNEL(Pad)
 REGISTER_KERNEL(PadV2)
 REGISTER_KERNEL(Pow)
 REGISTER_KERNEL(PRelu)
+REGISTER_KERNEL(Quantize)
 REGISTER_KERNEL(Relu)
 REGISTER_KERNEL(Relu6)
 REGISTER_KERNEL(Reshape)
@@ -61,6 +67,7 @@ REGISTER_KERNEL(Square)
 REGISTER_KERNEL(SquaredDifference)
 REGISTER_KERNEL(Squeeze)
 REGISTER_KERNEL(Sub)
+REGISTER_KERNEL(SVDF)
 REGISTER_KERNEL(Tanh)
 REGISTER_KERNEL(Transpose)
 REGISTER_KERNEL(TransposeConv)
diff --git a/compiler/luci-interpreter/pal/linux/PALAveragePool2d.h b/compiler/luci-interpreter/pal/linux/PALAveragePool2d.h
new file mode 100644
index 000000000..cce30601f
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALAveragePool2d.h
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+#define LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h>
+#include <tensorflow/lite/kernels/internal/reference/pooling.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void AveragePool(const tflite::PoolParams &params,
+                               const tflite::RuntimeShape &input_shape, const T *input_data,
+                               const tflite::RuntimeShape &output_shape, T *output_data,
+                               const tflite::RuntimeShape &scratchpad_shape, T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation doesn't support
+    assert(false && "AveragePool NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void AveragePool<int8_t>(const tflite::PoolParams &params,
+                                const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                                const tflite::RuntimeShape &output_shape, int8_t *output_data,
+                                const tflite::RuntimeShape &scratchpad_shape,
+                                int8_t *scratchpad_data)
+{
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
+
+  tflite::reference_integer_ops::AveragePool(params, input_shape, input_data, output_shape,
+                                             output_data);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &output_shape)
+
+{
+  (void)input_data_type;
+  (void)input_shape;
+  (void)output_shape;
+
+  scratchpad->set_allocatable(false);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
diff --git a/compiler/luci-interpreter/pal/linux/PALBatchMatMul.h b/compiler/luci-interpreter/pal/linux/PALBatchMatMul.h
new file mode 100644
index 000000000..3894f2d92
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALBatchMatMul.h
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_BATCHMATMUL_H
+#define LUCI_INTERPRETER_PAL_BATCHMATMUL_H
+
+#include <tensorflow/lite/kernels/internal/reference/batch_matmul.h>
+
+namespace luci_interpreter_pal
+{
+inline void BatchMatMul(const tflite::RuntimeShape &lhs_shape, const float *lhs_data,
+                        const tflite::RuntimeShape &rhs_shape, const float *rhs_data,
+                        const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_ops::BatchMatMul(lhs_shape, lhs_data, rhs_shape, rhs_data, output_shape,
+                                     output_data);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *lhs_scratchpad,
+                                         luci_interpreter::Tensor *rhs_scratchpad,
+                                         const tflite::RuntimeShape &lhs_shape,
+                                         const tflite::RuntimeShape &rhs_shape)
+{
+  // Scratchpad for transposed LHS
+  {
+    auto lhs_rank = lhs_shape.DimensionsCount();
+    luci_interpreter::Shape scratchpad_size(lhs_rank);
+    for (int i = 0; i < lhs_rank - 2; ++i)
+    {
+      scratchpad_size.dim(i) = lhs_shape.Dims(i);
+    }
+    scratchpad_size.dim(lhs_rank - 2) = lhs_shape.Dims(lhs_rank - 1);
+    scratchpad_size.dim(lhs_rank - 1) = lhs_shape.Dims(lhs_rank - 2);
+
+    lhs_scratchpad->resize(scratchpad_size);
+  }
+  // Scratchpad for transposed RHS
+  {
+    auto rhs_rank = rhs_shape.DimensionsCount();
+    luci_interpreter::Shape scratchpad_size(rhs_rank);
+    for (int i = 0; i < rhs_rank - 2; ++i)
+    {
+      scratchpad_size.dim(i) = rhs_shape.Dims(i);
+    }
+    scratchpad_size.dim(rhs_rank - 2) = rhs_shape.Dims(rhs_rank - 1);
+    scratchpad_size.dim(rhs_rank - 1) = rhs_shape.Dims(rhs_rank - 2);
+
+    rhs_scratchpad->resize(scratchpad_size);
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_BATCHMATMUL_H
diff --git a/compiler/luci-interpreter/pal/linux/PALConv2d.h b/compiler/luci-interpreter/pal/linux/PALConv2d.h
index 2550dd5d7..985a15f39 100644
--- a/compiler/luci-interpreter/pal/linux/PALConv2d.h
+++ b/compiler/luci-interpreter/pal/linux/PALConv2d.h
@@ -26,14 +26,24 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const float *input_data, const tflite::RuntimeShape &filter_shape,
                         const float *filter_data, const tflite::RuntimeShape &bias_shape,
                         const float *bias_data, const tflite::RuntimeShape &output_shape,
-                        float *output_data, const tflite::RuntimeShape &im2col_shape,
-                        float *im2col_data)
+                        float *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        float *scratchpad_data)
 {
-  if (im2col_data)
+  (void)scratchpad_shape;
+  if (scratchpad_data)
   {
+    const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int32_t input_depth = tflite::MatchingDim(input_shape, 3, filter_shape, 3);
+    const int32_t output_height = output_shape.Dims(1);
+    const int32_t output_width = output_shape.Dims(2);
+    const int32_t filter_height = filter_shape.Dims(1);
+    const int32_t filter_width = filter_shape.Dims(2);
+    tflite::RuntimeShape im2col_shape{batches, output_height, output_width,
+                                      input_depth * filter_height * filter_width};
+
     tflite::optimized_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
                                 bias_shape, bias_data, output_shape, output_data, im2col_shape,
-                                im2col_data);
+                                scratchpad_data);
   }
   else
     tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
@@ -45,8 +55,8 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const uint8 *input_data, const tflite::RuntimeShape &filter_shape,
                         const uint8 *filter_data, const tflite::RuntimeShape &bias_shape,
                         const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                        uint8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                        uint8 *im2col_data)
+                        uint8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        uint8 *scratchpad_data)
 {
   // TODO This should only be done once (although it takes only a few microseconds).
   //  Also, the user should be able to adjust the number of threads.
@@ -54,8 +64,8 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
   gemmlowp_context->set_max_num_threads(static_cast<int>(std::thread::hardware_concurrency()));
 
   tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
-                              bias_shape, bias_data, output_shape, output_data, im2col_shape,
-                              im2col_data, gemmlowp_context.get());
+                              bias_shape, bias_data, output_shape, output_data, scratchpad_shape,
+                              scratchpad_data, gemmlowp_context.get());
 }
 
 static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_t *mult,
@@ -63,17 +73,55 @@ static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_
                                   const int8 *input_data, const tflite::RuntimeShape &filter_shape,
                                   const int8 *filter_data, const tflite::RuntimeShape &bias_shape,
                                   const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                                  int8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                                  int8 *im2col_data)
+                                  int8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                                  int8 *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   // TODO enable optimized version
   tflite::reference_integer_ops::ConvPerChannel(params, mult, shifts, input_shape, input_data,
                                                 filter_shape, filter_data, bias_shape, bias_data,
                                                 output_shape, output_data);
 }
 
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::ConvParams &params,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+{
+  const int32_t filter_height = filter_shape.Dims(1);
+  const int32_t filter_width = filter_shape.Dims(2);
+
+  // Allocate tensor for scratchpad, if needed.
+  // The checks here should be aligned with the actual implementation.
+  const bool need_dilated_scratchpad =
+    params.dilation_height_factor != 1 || params.dilation_width_factor != 1;
+  const bool need_non_dilated_scratchpad = params.stride_height != 1 || params.stride_width != 1 ||
+                                           filter_height != 1 || filter_width != 1;
+  auto _need_scratchpad = input_data_type != luci_interpreter::DataType::S16 &&
+                          (need_dilated_scratchpad || need_non_dilated_scratchpad);
+
+  if (_need_scratchpad)
+  {
+    const int32_t batches = tflite::MatchingDim(input_shape, 0, output_shape, 0);
+    const int32_t input_depth = tflite::MatchingDim(input_shape, 3, filter_shape, 3);
+    const int32_t output_height = output_shape.Dims(1);
+    const int32_t output_width = output_shape.Dims(2);
+
+    auto data_type_size = static_cast<int32_t>(luci_interpreter::getDataTypeSize(input_data_type));
+    int32_t scratchpad_size = batches * output_width * output_height * input_depth * filter_height *
+                              filter_width * data_type_size;
+    luci_interpreter::Shape scratchpad_shape{scratchpad_size};
+    scratchpad->resize(scratchpad_shape);
+  }
+  else
+  {
+    scratchpad->set_allocatable(false);
+  }
+}
+
 } // namespace luci_interpreter_pal
 
 #endif // LUCI_INTERPRETER_PAL_CONV2D_H
diff --git a/compiler/luci-interpreter/pal/linux/PALDepthwiseConv2d.h b/compiler/luci-interpreter/pal/linux/PALDepthwiseConv2d.h
new file mode 100644
index 000000000..c9d1a2948
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALDepthwiseConv2d.h
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+#define LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h>
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void
+DepthwiseConvPerChannel(const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+                        const int32_t *output_shift, const tflite::RuntimeShape &input_shape,
+                        const T *input_data, const tflite::RuntimeShape &filter_shape,
+                        const T *filter_data, const tflite::RuntimeShape &bias_shape,
+                        const int32_t *bias_data, const tflite::RuntimeShape &output_shape,
+                        T *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation is not supported
+    assert(false && "DepthwiseConvPerChannel NYI");
+    (void)params;
+    (void)output_multiplier;
+    (void)output_shift;
+    (void)input_shape;
+    (void)output_data;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void DepthwiseConvPerChannel<int8_t>(
+  const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+  const int32_t *output_shift, const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+  const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+  const tflite::RuntimeShape &output_shape, int8_t *output_data,
+  const tflite::RuntimeShape &scratchpad_shape, int8_t *scratchpad_data)
+{
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
+  tflite::reference_integer_ops::DepthwiseConvPerChannel(
+    params, output_multiplier, output_shift, input_shape, input_data, filter_shape, filter_data,
+    bias_shape, bias_data, output_shape, output_data);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const tflite::DepthwiseParams &params,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+
+{
+  (void)params;
+  (void)input_data_type;
+  (void)input_shape;
+  (void)filter_shape;
+  (void)output_shape;
+
+  scratchpad->set_allocatable(false);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
diff --git a/compiler/luci-interpreter/pal/linux/PALDequantize.h b/compiler/luci-interpreter/pal/linux/PALDequantize.h
new file mode 100644
index 000000000..3af6d0777
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALDequantize.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+#define LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+
+#include <tensorflow/lite/kernels/internal/optimized/optimized_ops.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void Dequantize(tflite::DequantizationParams &params,
+                              const tflite::RuntimeShape &input_shape, const T *input_data,
+                              const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::optimized_ops::Dequantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEQUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/linux/PALFullyConnected.h b/compiler/luci-interpreter/pal/linux/PALFullyConnected.h
new file mode 100644
index 000000000..62970dbf7
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALFullyConnected.h
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+#define LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+
+#include <tensorflow/lite/kernels/internal/reference/fully_connected.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void FullyConnected(const tflite::FullyConnectedParams &params,
+                                  const tflite::RuntimeShape &input_shape, const T *input_data,
+                                  const tflite::RuntimeShape &filter_shape, const T *filter_data,
+                                  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                                  const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  {
+    // MARK: At this moment this operation doesn't support
+    assert(false && "FullyConnected NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+  }
+}
+
+template <>
+inline void
+FullyConnected<int8_t>(const tflite::FullyConnectedParams &params,
+                       const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                       const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+                       const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                       const tflite::RuntimeShape &output_shape, int8_t *output_data)
+{
+  tflite::reference_integer_ops::FullyConnected(params, input_shape, input_data, filter_shape,
+                                                filter_data, bias_shape, bias_data, output_shape,
+                                                output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
diff --git a/compiler/luci-interpreter/pal/linux/PALGather.h b/compiler/luci-interpreter/pal/linux/PALGather.h
new file mode 100644
index 000000000..49ac35f93
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALGather.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_GATHER_H
+#define LUCI_INTERPRETER_PAL_GATHER_H
+
+#include <tensorflow/lite/kernels/internal/optimized/optimized_ops.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T, typename CoordsT = int32>
+static inline void Gather(const tflite::GatherParams &op_params,
+                          const tflite::RuntimeShape &input_shape, const T *input_data,
+                          const tflite::RuntimeShape &coords_shape, const CoordsT *coords_data,
+                          const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  tflite::optimized_ops::Gather(op_params, input_shape, input_data, coords_shape, coords_data,
+                                output_shape, output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_GATHER_H
diff --git a/compiler/luci-interpreter/pal/linux/PALMul.h b/compiler/luci-interpreter/pal/linux/PALMul.h
index cfaec1b58..a8a9d4abc 100644
--- a/compiler/luci-interpreter/pal/linux/PALMul.h
+++ b/compiler/luci-interpreter/pal/linux/PALMul.h
@@ -21,21 +21,31 @@
 
 namespace luci_interpreter_pal
 {
+template <typename T>
 static inline void Mul(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
-                       const float *input1_data, const tflite::RuntimeShape &input2_shape,
-                       const float *input2_data, const tflite::RuntimeShape &output_shape,
-                       float *output_data)
+                       const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                       const T *input2_data, const tflite::RuntimeShape &output_shape,
+                       T *output_data)
 {
   tflite::optimized_ops::Mul(params, input1_shape, input1_data, input2_shape, input2_data,
                              output_shape, output_data);
 }
 
-static inline void BroadcastMul4DSlow(tflite::ArithmeticParams &params,
-                                      const tflite::RuntimeShape &input1_shape,
-                                      const float *input1_data,
-                                      const tflite::RuntimeShape &input2_shape,
-                                      const float *input2_data,
-                                      const tflite::RuntimeShape &output_shape, float *output_data)
+template <>
+inline void Mul(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
+                const int64_t *input1_data, const tflite::RuntimeShape &input2_shape,
+                const int64_t *input2_data, const tflite::RuntimeShape &output_shape,
+                int64_t *output_data)
+{
+  tflite::optimized_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
+                                            input2_data, output_shape, output_data);
+}
+
+template <typename T>
+static inline void
+BroadcastMul4DSlow(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
+                   const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                   const T *input2_data, const tflite::RuntimeShape &output_shape, T *output_data)
 {
   tflite::optimized_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
                                             input2_data, output_shape, output_data);
diff --git a/compiler/luci-interpreter/pal/linux/PALQuantize.h b/compiler/luci-interpreter/pal/linux/PALQuantize.h
new file mode 100644
index 000000000..bf1d7954e
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALQuantize.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_QUANTIZE_H
+#define LUCI_INTERPRETER_PAL_QUANTIZE_H
+
+#include <tensorflow/lite/kernels/internal/optimized/optimized_ops.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void Quantize(tflite::QuantizationParams &params,
+                            const tflite::RuntimeShape &input_shape, const float *input_data,
+                            const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  tflite::optimized_ops::AffineQuantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+template <typename Input, typename Output>
+static inline void Requantize(const Input *input_data, int32_t size,
+                              int32_t effective_scale_multiplier, int32_t effective_scale_shift,
+                              int32_t input_zero_point, int32_t output_zero_point,
+                              Output *output_data)
+{
+  tflite::optimized_ops::Requantize(input_data, size, effective_scale_multiplier,
+                                    effective_scale_shift, input_zero_point, output_zero_point,
+                                    output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_QUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/linux/PALSVDF.h b/compiler/luci-interpreter/pal/linux/PALSVDF.h
new file mode 100644
index 000000000..0ffba14f0
--- /dev/null
+++ b/compiler/luci-interpreter/pal/linux/PALSVDF.h
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_SVDF_H
+#define LUCI_INTERPRETER_PAL_SVDF_H
+
+#include <tensorflow/lite/kernels/internal/reference/svdf.h>
+
+namespace luci_interpreter_pal
+{
+static inline void
+IntegerSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+            const int8_t *input_data, const tflite::RuntimeShape &weight_feature_shape,
+            const int8_t *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+            const int16_t *weight_time_data, const tflite::RuntimeShape &bias_shape,
+            const int32_t *bias_data, int16_t *activation_state_data,
+            const tflite::RuntimeShape &output_shape, int8_t *output_data, int32_t *scratchpad_data,
+            int32_t *output_temp_data, int32_t scale_1_a, int scale_1_b, int32_t scale_2_a,
+            int scale_2_b, int32_t input_zp, int32_t output_zp)
+{
+  tflite::reference_ops::EvalIntegerSVDF(&params, input_shape, input_data, weight_feature_shape,
+                                         weight_feature_data, weight_time_shape, weight_time_data,
+                                         bias_shape, bias_data, activation_state_data, output_shape,
+                                         output_data, scratchpad_data, output_temp_data, scale_1_a,
+                                         scale_1_b, scale_2_a, scale_2_b, input_zp, output_zp);
+}
+static inline void
+FloatSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+          const float *input_data, const tflite::RuntimeShape &weight_feature_shape,
+          const float *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+          const float *weight_time_data, const tflite::RuntimeShape &bias_shape,
+          const float *bias_data, float *scratchpad_data, float *activation_state_data,
+          const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_ops::EvalFloatSVDF(&params, input_shape, input_data, weight_feature_shape,
+                                       weight_feature_data, weight_time_shape, weight_time_data,
+                                       bias_shape, bias_data, scratchpad_data,
+                                       activation_state_data, output_shape, output_data);
+}
+
+static inline void SetupScratchpadTensor(
+  const luci_interpreter::DataType &input_data_type,
+  const luci_interpreter::DataType &weight_feature_data_type,
+  luci_interpreter::Tensor *scratchpad_1, luci_interpreter::Tensor *scratchpad_2,
+  luci_interpreter::Tensor *scratchpad_3, luci_interpreter::Tensor *scratchpad_4,
+  luci_interpreter::Tensor *scratchpad_5, luci_interpreter::Tensor *scratchpad_6,
+  const luci_interpreter::Shape input_shape, const luci_interpreter::Shape weight_time_shape,
+  const int32_t batch_size, const int32_t num_filters, const int32_t num_units)
+{
+
+  if (input_data_type == loco::DataType::FLOAT32 &&
+      (weight_feature_data_type == loco::DataType::S8 ||
+       weight_feature_data_type == loco::DataType::U8))
+  {
+    (void)input_shape;
+    (void)weight_time_shape;
+    (void)scratchpad_3;
+    (void)scratchpad_4;
+    (void)scratchpad_5;
+    (void)scratchpad_6;
+
+    throw std::runtime_error("Hybrid type is not currently supported for linux platform");
+  }
+
+  // Resize scratchpad_1 tensor
+  scratchpad_1->resize({batch_size, num_filters});
+
+  if (input_data_type == loco::DataType::S8)
+  {
+    // Resize scratchpad_2 for full_integer op
+    scratchpad_2->resize({batch_size, num_units});
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_SVDF_H
diff --git a/compiler/luci-interpreter/pal/linux/pal.cmake b/compiler/luci-interpreter/pal/linux/pal.cmake
index 84349e0bf..185700cf9 100644
--- a/compiler/luci-interpreter/pal/linux/pal.cmake
+++ b/compiler/luci-interpreter/pal/linux/pal.cmake
@@ -40,7 +40,35 @@ macro(add_pal_to_target TGT)
 
     # TODO put it back, I changed my mind.
     # instead add sources with visitors in this library
-    set(PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc)
+    set(PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/tensor_utils.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc)
+
+    if(BUILD_ARM32_NEON)
+        # NOTE may need to revise this list for version upgrade
+        set(PAL_SOURCES ${PAL_SOURCES}
+                ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc
+                ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/optimized/cpu_check.cc
+                ${TensorFlowRuySource_DIR}/ruy/allocator.cc
+                ${TensorFlowRuySource_DIR}/ruy/block_map.cc
+                ${TensorFlowRuySource_DIR}/ruy/blocking_counter.cc
+                ${TensorFlowRuySource_DIR}/ruy/context_get_ctx.cc
+                ${TensorFlowRuySource_DIR}/ruy/cpuinfo.cc
+                ${TensorFlowRuySource_DIR}/ruy/ctx.cc
+                ${TensorFlowRuySource_DIR}/ruy/denormal.cc
+                ${TensorFlowRuySource_DIR}/ruy/frontend.cc
+                ${TensorFlowRuySource_DIR}/ruy/pack_arm.cc
+                ${TensorFlowRuySource_DIR}/ruy/prepacked_cache.cc
+                ${TensorFlowRuySource_DIR}/ruy/prepare_packed_matrices.cc
+                ${TensorFlowRuySource_DIR}/ruy/system_aligned_alloc.cc
+                ${TensorFlowRuySource_DIR}/ruy/thread_pool.cc
+                ${TensorFlowRuySource_DIR}/ruy/trmul.cc
+                ${TensorFlowRuySource_DIR}/ruy/tune.cc
+                ${TensorFlowRuySource_DIR}/ruy/wait.cc
+                ${TensorFlowRuySource_DIR}/ruy/kernel_arm32.cc
+                )
+    endif(BUILD_ARM32_NEON)
+
     add_library(luci_interpreter_linux_pal STATIC ${PAL_SOURCES})
     set_target_properties(luci_interpreter_linux_pal PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_include_directories(luci_interpreter_linux_pal SYSTEM PRIVATE
diff --git a/compiler/luci-interpreter/pal/mcu/KernelsToBuild.lst b/compiler/luci-interpreter/pal/mcu/KernelsToBuild.lst
index 771974afe..d134a6b95 100644
--- a/compiler/luci-interpreter/pal/mcu/KernelsToBuild.lst
+++ b/compiler/luci-interpreter/pal/mcu/KernelsToBuild.lst
@@ -7,9 +7,11 @@ REGISTER_KERNEL(Concatenation)
 REGISTER_KERNEL(Conv2D)
 REGISTER_KERNEL(DepthToSpace)
 REGISTER_KERNEL(DepthwiseConv2D)
+REGISTER_KERNEL(Dequantize)
 REGISTER_KERNEL(Div)
 REGISTER_KERNEL(Elu)
 REGISTER_KERNEL(Exp)
+REGISTER_KERNEL(ExpandDims)
 REGISTER_KERNEL(Floor)
 REGISTER_KERNEL(FloorDiv)
 REGISTER_KERNEL(Equal)
@@ -37,6 +39,7 @@ REGISTER_KERNEL(NotEqual)
 REGISTER_KERNEL(Pad)
 REGISTER_KERNEL(PadV2)
 REGISTER_KERNEL(PRelu)
+REGISTER_KERNEL(Quantize)
 REGISTER_KERNEL(Reshape)
 REGISTER_KERNEL(ResizeBilinear)
 REGISTER_KERNEL(ResizeNearestNeighbor)
@@ -50,6 +53,7 @@ REGISTER_KERNEL(Square)
 REGISTER_KERNEL(SquaredDifference)
 REGISTER_KERNEL(Squeeze)
 REGISTER_KERNEL(Sub)
+REGISTER_KERNEL(SVDF)
 REGISTER_KERNEL(Tanh)
 REGISTER_KERNEL(Transpose)
 REGISTER_KERNEL(TransposeConv)
diff --git a/compiler/luci-interpreter/pal/mcu/PALAveragePool2d.h b/compiler/luci-interpreter/pal/mcu/PALAveragePool2d.h
new file mode 100644
index 000000000..cce30601f
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALAveragePool2d.h
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+#define LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h>
+#include <tensorflow/lite/kernels/internal/reference/pooling.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void AveragePool(const tflite::PoolParams &params,
+                               const tflite::RuntimeShape &input_shape, const T *input_data,
+                               const tflite::RuntimeShape &output_shape, T *output_data,
+                               const tflite::RuntimeShape &scratchpad_shape, T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation doesn't support
+    assert(false && "AveragePool NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void AveragePool<int8_t>(const tflite::PoolParams &params,
+                                const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                                const tflite::RuntimeShape &output_shape, int8_t *output_data,
+                                const tflite::RuntimeShape &scratchpad_shape,
+                                int8_t *scratchpad_data)
+{
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
+
+  tflite::reference_integer_ops::AveragePool(params, input_shape, input_data, output_shape,
+                                             output_data);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &output_shape)
+
+{
+  (void)input_data_type;
+  (void)input_shape;
+  (void)output_shape;
+
+  scratchpad->set_allocatable(false);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_AVERAGEPOOL2D_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALConv2d.h b/compiler/luci-interpreter/pal/mcu/PALConv2d.h
index 0a8ae4e48..13976877a 100644
--- a/compiler/luci-interpreter/pal/mcu/PALConv2d.h
+++ b/compiler/luci-interpreter/pal/mcu/PALConv2d.h
@@ -26,11 +26,11 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const float *input_data, const tflite::RuntimeShape &filter_shape,
                         const float *filter_data, const tflite::RuntimeShape &bias_shape,
                         const float *bias_data, const tflite::RuntimeShape &output_shape,
-                        float *output_data, const tflite::RuntimeShape &im2col_shape,
-                        float *im2col_data)
+                        float *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        float *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
                               bias_shape, bias_data, output_shape, output_data,
                               tflite::RuntimeShape(), nullptr);
@@ -40,14 +40,14 @@ static inline void Conv(const tflite::ConvParams &params, const tflite::RuntimeS
                         const uint8 *input_data, const tflite::RuntimeShape &filter_shape,
                         const uint8 *filter_data, const tflite::RuntimeShape &bias_shape,
                         const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                        uint8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                        uint8 *im2col_data)
+                        uint8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        uint8 *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   tflite::reference_ops::Conv(params, input_shape, input_data, filter_shape, filter_data,
-                              bias_shape, bias_data, output_shape, output_data, im2col_shape,
-                              im2col_data, nullptr);
+                              bias_shape, bias_data, output_shape, output_data, scratchpad_shape,
+                              scratchpad_data, nullptr);
 }
 
 static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_t *mult,
@@ -55,16 +55,31 @@ static inline void ConvPerChannel(const tflite::ConvParams &params, const int32_
                                   const int8 *input_data, const tflite::RuntimeShape &filter_shape,
                                   const int8 *filter_data, const tflite::RuntimeShape &bias_shape,
                                   const int32 *bias_data, const tflite::RuntimeShape &output_shape,
-                                  int8 *output_data, const tflite::RuntimeShape &im2col_shape,
-                                  int8 *im2col_data)
+                                  int8 *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                                  int8 *scratchpad_data)
 {
-  (void)im2col_shape;
-  (void)im2col_data;
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
   tflite::reference_integer_ops::ConvPerChannel(params, mult, shifts, input_shape, input_data,
                                                 filter_shape, filter_data, bias_shape, bias_data,
                                                 output_shape, output_data);
 }
 
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::ConvParams &params,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+{
+  (void)input_data_type;
+  (void)params;
+  (void)input_shape;
+  (void)filter_shape;
+  (void)output_shape;
+  scratchpad->set_allocatable(false);
+}
+
 } // namespace luci_interpreter_pal
 
 #endif // LUCI_INTERPRETER_PAL_CONV2D_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALDepthwiseConv2d.h b/compiler/luci-interpreter/pal/mcu/PALDepthwiseConv2d.h
new file mode 100644
index 000000000..c9d1a2948
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALDepthwiseConv2d.h
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+#define LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
+
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h>
+#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void
+DepthwiseConvPerChannel(const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+                        const int32_t *output_shift, const tflite::RuntimeShape &input_shape,
+                        const T *input_data, const tflite::RuntimeShape &filter_shape,
+                        const T *filter_data, const tflite::RuntimeShape &bias_shape,
+                        const int32_t *bias_data, const tflite::RuntimeShape &output_shape,
+                        T *output_data, const tflite::RuntimeShape &scratchpad_shape,
+                        T *scratchpad_data)
+{
+  {
+    // MARK: At this moment this operation is not supported
+    assert(false && "DepthwiseConvPerChannel NYI");
+    (void)params;
+    (void)output_multiplier;
+    (void)output_shift;
+    (void)input_shape;
+    (void)output_data;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+    (void)scratchpad_shape;
+    (void)scratchpad_data;
+  }
+}
+
+template <>
+inline void DepthwiseConvPerChannel<int8_t>(
+  const tflite::DepthwiseParams &params, const int32_t *output_multiplier,
+  const int32_t *output_shift, const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+  const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+  const tflite::RuntimeShape &output_shape, int8_t *output_data,
+  const tflite::RuntimeShape &scratchpad_shape, int8_t *scratchpad_data)
+{
+  (void)scratchpad_shape;
+  (void)scratchpad_data;
+  tflite::reference_integer_ops::DepthwiseConvPerChannel(
+    params, output_multiplier, output_shift, input_shape, input_data, filter_shape, filter_data,
+    bias_shape, bias_data, output_shape, output_data);
+}
+
+static inline void SetupScratchpadTensor(luci_interpreter::Tensor *scratchpad,
+                                         const tflite::DepthwiseParams &params,
+                                         const luci_interpreter::DataType &input_data_type,
+                                         const tflite::RuntimeShape &input_shape,
+                                         const tflite::RuntimeShape &filter_shape,
+                                         const tflite::RuntimeShape &output_shape)
+
+{
+  (void)params;
+  (void)input_data_type;
+  (void)input_shape;
+  (void)filter_shape;
+  (void)output_shape;
+
+  scratchpad->set_allocatable(false);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEPTHWISECONV2D_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALDequantize.h b/compiler/luci-interpreter/pal/mcu/PALDequantize.h
new file mode 100644
index 000000000..15ff0327b
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALDequantize.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+#define LUCI_INTERPRETER_PAL_DEQUANTIZE_H
+
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/dequantize.h"
+#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
+
+namespace luci_interpreter_pal
+{
+
+template <typename T>
+static inline void Dequantize(tflite::DequantizationParams &params,
+                              const tflite::RuntimeShape &input_shape, const T *input_data,
+                              const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_integer_ops::Dequantize<T>(params, input_shape, input_data, output_shape,
+                                               output_data);
+}
+
+static inline void Dequantize(tflite::DequantizationParams &params,
+                              const tflite::RuntimeShape &input_shape, const uint8_t *input_data,
+                              const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  tflite::reference_ops::Dequantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_DEQUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALFullyConnected.h b/compiler/luci-interpreter/pal/mcu/PALFullyConnected.h
new file mode 100644
index 000000000..048624d74
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALFullyConnected.h
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+#define LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
+
+#include <tensorflow/lite/kernels/internal/reference/fully_connected.h>
+#include <tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h>
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void FullyConnected(const tflite::FullyConnectedParams &params,
+                                  const tflite::RuntimeShape &input_shape, const T *input_data,
+                                  const tflite::RuntimeShape &filter_shape, const T *filter_data,
+                                  const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                                  const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  {
+    // MARK: At this moment this operation is not supported
+    assert(false && "FullyConnected NYI");
+    (void)params;
+    (void)input_shape;
+    (void)input_data;
+    (void)filter_shape;
+    (void)filter_data;
+    (void)bias_shape;
+    (void)bias_data;
+    (void)output_shape;
+    (void)output_data;
+  }
+}
+
+template <>
+inline void
+FullyConnected<int8_t>(const tflite::FullyConnectedParams &params,
+                       const tflite::RuntimeShape &input_shape, const int8_t *input_data,
+                       const tflite::RuntimeShape &filter_shape, const int8_t *filter_data,
+                       const tflite::RuntimeShape &bias_shape, const int32_t *bias_data,
+                       const tflite::RuntimeShape &output_shape, int8_t *output_data)
+{
+  tflite::reference_integer_ops::FullyConnected(params, input_shape, input_data, filter_shape,
+                                                filter_data, bias_shape, bias_data, output_shape,
+                                                output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_FULLYCONNECTED_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALMul.h b/compiler/luci-interpreter/pal/mcu/PALMul.h
index 2b46b100c..347a97a83 100644
--- a/compiler/luci-interpreter/pal/mcu/PALMul.h
+++ b/compiler/luci-interpreter/pal/mcu/PALMul.h
@@ -21,21 +21,21 @@
 
 namespace luci_interpreter_pal
 {
+template <typename T>
 static inline void Mul(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
-                       const float *input1_data, const tflite::RuntimeShape &input2_shape,
-                       const float *input2_data, const tflite::RuntimeShape &output_shape,
-                       float *output_data)
+                       const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                       const T *input2_data, const tflite::RuntimeShape &output_shape,
+                       T *output_data)
 {
   tflite::reference_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
                                             input2_data, output_shape, output_data);
 }
 
-static inline void BroadcastMul4DSlow(tflite::ArithmeticParams &params,
-                                      const tflite::RuntimeShape &input1_shape,
-                                      const float *input1_data,
-                                      const tflite::RuntimeShape &input2_shape,
-                                      const float *input2_data,
-                                      const tflite::RuntimeShape &output_shape, float *output_data)
+template <typename T>
+static inline void
+BroadcastMul4DSlow(tflite::ArithmeticParams &params, const tflite::RuntimeShape &input1_shape,
+                   const T *input1_data, const tflite::RuntimeShape &input2_shape,
+                   const T *input2_data, const tflite::RuntimeShape &output_shape, T *output_data)
 {
   tflite::reference_ops::BroadcastMul4DSlow(params, input1_shape, input1_data, input2_shape,
                                             input2_data, output_shape, output_data);
diff --git a/compiler/luci-interpreter/pal/mcu/PALQuantize.h b/compiler/luci-interpreter/pal/mcu/PALQuantize.h
new file mode 100644
index 000000000..6046789ae
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALQuantize.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_QUANTIZE_H
+#define LUCI_INTERPRETER_PAL_QUANTIZE_H
+
+#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
+
+namespace luci_interpreter_pal
+{
+template <typename T>
+static inline void Quantize(tflite::QuantizationParams &params,
+                            const tflite::RuntimeShape &input_shape, const float *input_data,
+                            const tflite::RuntimeShape &output_shape, T *output_data)
+{
+  tflite::reference_ops::AffineQuantize(params, input_shape, input_data, output_shape, output_data);
+}
+
+template <typename Input, typename Output>
+static inline void Requantize(const Input *input_data, int32_t size,
+                              int32_t effective_scale_multiplier, int32_t effective_scale_shift,
+                              int32_t input_zero_point, int32_t output_zero_point,
+                              Output *output_data)
+{
+  tflite::reference_ops::Requantize(input_data, size, effective_scale_multiplier,
+                                    effective_scale_shift, input_zero_point, output_zero_point,
+                                    output_data);
+}
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_QUANTIZE_H
diff --git a/compiler/luci-interpreter/pal/mcu/PALSVDF.h b/compiler/luci-interpreter/pal/mcu/PALSVDF.h
new file mode 100644
index 000000000..3bba668fb
--- /dev/null
+++ b/compiler/luci-interpreter/pal/mcu/PALSVDF.h
@@ -0,0 +1,258 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_PAL_SVDF_H
+#define LUCI_INTERPRETER_PAL_SVDF_H
+
+#include <tensorflow/lite/kernels/internal/reference/svdf.h>
+
+namespace luci_interpreter_pal
+{
+static inline void
+IntegerSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+            const int8_t *input_data, const tflite::RuntimeShape &weight_feature_shape,
+            const int8_t *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+            const int16_t *weight_time_data, const tflite::RuntimeShape &bias_shape,
+            const int32_t *bias_data, int16_t *activation_state_data,
+            const tflite::RuntimeShape &output_shape, int8_t *output_data, int32_t *scratchpad_data,
+            int32_t *output_temp_data, int32_t scale_1_a, int scale_1_b, int32_t scale_2_a,
+            int scale_2_b, int32_t input_zp, int32_t output_zp)
+{
+  const int n_rank = params.rank;
+  const int n_batch = input_shape.Dims(0);
+  const int n_input = input_shape.Dims(1);
+  const int n_filter = weight_feature_shape.Dims(0);
+  const int n_unit = n_filter / n_rank;
+  const int n_memory = weight_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  {
+    int16_t *new_state_start = activation_state_data;
+    const int16_t *old_state_start = activation_state_data + 1;
+    const int16_t *old_state_end = activation_state_data + n_batch * n_filter * n_memory;
+    while (old_state_start != old_state_end)
+    {
+      *new_state_start++ = *old_state_start++;
+    }
+  }
+
+  // Note: no need to clear the latest activation, matmul is not accumulative.
+
+  // Feature matmul.
+  {
+    const int32_t output_max = std::numeric_limits<int16_t>::max();
+    const int32_t output_min = std::numeric_limits<int16_t>::min();
+    int16_t *result_in_batch = activation_state_data + (n_memory - 1);
+    for (int b = 0; b < n_batch; b++)
+    {
+      const int8_t *matrix_ptr = weight_feature_data;
+      for (int r = 0; r < n_filter; r++)
+      {
+        int32_t dot_prod = 0;
+        const int8_t *vector_in_batch = input_data + b * n_input;
+        for (int c = 0; c < n_input; c++)
+        {
+          dot_prod += *matrix_ptr++ * (*vector_in_batch++ - input_zp);
+        }
+        dot_prod = tflite::MultiplyByQuantizedMultiplier(dot_prod, scale_1_a, scale_1_b);
+        dot_prod = std::min(std::max(output_min, dot_prod), output_max);
+        // This assumes state is symmetrically quantized. Otherwise last bit of
+        // state should be initialized to its zero point and accumulate the
+        // dot_prod.
+        // Equivalent as the following:
+        //     result_in_batch = zero point, which happens to be zero.
+        //     result_in_batch += dot_prod_56.
+        *result_in_batch = dot_prod;
+        result_in_batch += n_memory;
+      }
+    }
+  }
+
+  // Time.
+  {
+    for (int b = 0; b < n_batch; ++b)
+    {
+      int32_t *scratch_ptr_batch = scratchpad_data + b * n_filter;
+
+      // Perform batched vector dot product:
+      const int16_t *vector1_ptr = weight_time_data;
+      const int16_t *vector2_ptr = activation_state_data + b * n_memory * n_filter;
+
+      for (int i = 0; i < n_filter; i++)
+      {
+        *scratch_ptr_batch = 0;
+        for (int j = 0; j < n_memory; j++)
+        {
+          *scratch_ptr_batch += *vector1_ptr++ * *vector2_ptr++;
+        }
+        scratch_ptr_batch++;
+      }
+    }
+  }
+
+  // Reduce, add bias, rescale, activation.
+  {
+    // Add bias.
+    if (bias_data)
+    {
+      // Vector batch assign:
+      for (int i = 0; i < n_batch; ++i)
+      {
+        int32_t *output_ptr = output_temp_data + i * n_unit;
+        const int32_t *bias_ptr = bias_data;
+        for (int j = 0; j < n_unit; ++j)
+        {
+          *output_ptr++ = *bias_ptr++;
+        }
+      }
+    }
+    else
+    {
+      int32_t *output_ptr = output_temp_data;
+      for (int i = 0; i < n_batch * n_unit; ++i)
+      {
+        *output_ptr++ = 0;
+      }
+    }
+
+    // Reduce.
+    for (int b = 0; b < n_batch; ++b)
+    {
+      int32_t *output_temp_ptr = output_temp_data + b * n_unit;
+      int32_t *scratch_ptr_batch = scratchpad_data + b * n_filter;
+
+      // Reduction sum vector
+      for (int i = 0; i < n_unit; ++i)
+      {
+        for (int j = 0; j < n_rank; ++j)
+        {
+          output_temp_ptr[i] += *scratch_ptr_batch++;
+        }
+      }
+    }
+
+    // Rescale.
+    const int32_t output_max = std::numeric_limits<int8_t>::max();
+    const int32_t output_min = std::numeric_limits<int8_t>::min();
+    for (int i = 0; i < n_batch * n_unit; ++i)
+    {
+      int32_t x1 = output_temp_data[i];
+      int32_t x2 = tflite::MultiplyByQuantizedMultiplier(x1, scale_2_a, scale_2_b);
+      int32_t x3 = x2 + output_zp;
+      int32_t x4 = std::min(std::max(output_min, x3), output_max);
+      output_data[i] = static_cast<int8_t>(x4);
+    }
+  }
+}
+static inline void
+FloatSVDF(const TfLiteSVDFParams &params, const tflite::RuntimeShape &input_shape,
+          const float *input_data, const tflite::RuntimeShape &weight_feature_shape,
+          const float *weight_feature_data, const tflite::RuntimeShape &weight_time_shape,
+          const float *weight_time_data, const tflite::RuntimeShape &bias_shape,
+          const float *bias_data, float *scratchpad_data, float *activation_state_data,
+          const tflite::RuntimeShape &output_shape, float *output_data)
+{
+  const int32_t rank = params.rank;
+  const int32_t batch_size = input_shape.Dims(0);
+  const int32_t input_size = input_shape.Dims(1);
+  const int32_t num_filters = weight_feature_shape.Dims(0);
+  const int32_t num_units = num_filters / rank;
+  const int32_t memory_size = weight_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  {
+    float *new_state_start = activation_state_data;
+    const float *old_state_start = activation_state_data + 1;
+    const float *old_state_end = activation_state_data + batch_size * num_filters * memory_size;
+    while (old_state_start != old_state_end)
+    {
+      *new_state_start++ = *old_state_start++;
+    }
+  }
+
+  // Note: no need to clear the latest activation, matmul is not accumulative.
+
+  // Compute conv1d(inputs, weights_feature).
+  // The activation_state's rightmost column is used to save current cycle
+  // activation. This is achieved by starting at state_ptr[memory_size - 1] and
+  // having the stride equal to memory_size.
+
+  // Perform batched matrix vector multiply operation:
+  {
+    const float *matrix = weight_feature_data;
+    const float *vector = input_data;
+    float *result = &activation_state_data[memory_size - 1];
+    float *result_in_batch = result;
+    for (int i = 0; i < batch_size; ++i)
+    {
+      const float *matrix_ptr = matrix;
+      for (int j = 0; j < num_filters; ++j)
+      {
+        float dot_prod = 0.0f;
+        const float *vector_in_batch = vector + i * input_size;
+        for (int k = 0; k < input_size; ++k)
+        {
+          dot_prod += *matrix_ptr++ * *vector_in_batch++;
+        }
+        *result_in_batch = dot_prod;
+        result_in_batch += memory_size;
+      }
+    }
+  }
+
+  tflite::reference_ops::ApplyTimeWeightsBiasAndActivation(
+    batch_size, memory_size, num_filters, num_units, rank, weight_time_data, bias_data,
+    params.activation, activation_state_data, scratchpad_data, output_data);
+}
+
+static inline void SetupScratchpadTensor(
+  const luci_interpreter::DataType &input_data_type,
+  const luci_interpreter::DataType &weight_feature_data_type,
+  luci_interpreter::Tensor *scratchpad_1, luci_interpreter::Tensor *scratchpad_2,
+  luci_interpreter::Tensor *scratchpad_3, luci_interpreter::Tensor *scratchpad_4,
+  luci_interpreter::Tensor *scratchpad_5, luci_interpreter::Tensor *scratchpad_6,
+  const luci_interpreter::Shape input_shape, const luci_interpreter::Shape weight_time_shape,
+  const int32_t batch_size, const int32_t num_filters, const int32_t num_units)
+{
+
+  if (input_data_type == loco::DataType::FLOAT32 &&
+      (weight_feature_data_type == loco::DataType::S8 ||
+       weight_feature_data_type == loco::DataType::U8))
+  {
+    (void)input_shape;
+    (void)weight_time_shape;
+    (void)scratchpad_3;
+    (void)scratchpad_4;
+    (void)scratchpad_5;
+    (void)scratchpad_6;
+
+    throw std::runtime_error("Hybrid type is not currently supported for mcu platform");
+  }
+
+  // Resize scratchpad_1 tensor
+  scratchpad_1->resize({batch_size, num_filters});
+
+  if (input_data_type == loco::DataType::S8)
+  {
+    // Resize scratchpad_2 for full_integer op
+    scratchpad_2->resize({batch_size, num_units});
+  }
+}
+
+} // namespace luci_interpreter_pal
+
+#endif // LUCI_INTERPRETER_PAL_SVDF_H
diff --git a/compiler/luci-interpreter/pal/mcu/pal.cmake b/compiler/luci-interpreter/pal/mcu/pal.cmake
index a479d407b..907d51de6 100644
--- a/compiler/luci-interpreter/pal/mcu/pal.cmake
+++ b/compiler/luci-interpreter/pal/mcu/pal.cmake
@@ -39,7 +39,9 @@ macro(add_pal_to_target TGT)
 
     # TODO put it back, I changed my mind.
     # instead add sources with visitors in this library
-    set(PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc)
+    set(PAL_SOURCES ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/quantization_util.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/tensor_utils.cc
+            ${TensorFlowSource_DIR}/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc)
     add_library(luci_interpreter_mcu_pal STATIC ${PAL_SOURCES})
     set_target_properties(luci_interpreter_mcu_pal PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_include_directories(luci_interpreter_mcu_pal PRIVATE
diff --git a/compiler/luci-interpreter/src/CMakeLists.txt b/compiler/luci-interpreter/src/CMakeLists.txt
index e37150336..997b75a84 100644
--- a/compiler/luci-interpreter/src/CMakeLists.txt
+++ b/compiler/luci-interpreter/src/CMakeLists.txt
@@ -13,6 +13,7 @@ set(LUCI_INTERPRETER_BINARY "luci_interpreter${LUCI_INTERPRETER_SUFFIX}")
 set(LUCI_INTERPRETER_CORE "luci_interpreter_core${LUCI_INTERPRETER_SUFFIX}")
 set(LUCI_INTERPRETER_KERNELS "luci_interpreter_kernels${LUCI_INTERPRETER_SUFFIX}")
 set(LUCI_INTERPRETER_LOADER "luci_interpreter_loader${LUCI_INTERPRETER_SUFFIX}")
+set(LUCI_INTERPRETER_IMPORT "luci_interpreter_import${LUCI_INTERPRETER_SUFFIX}")
 
 add_subdirectory(core)
 message(STATUS "LUCI INTERPRETER CORE")
@@ -20,6 +21,8 @@ add_subdirectory(kernels)
 message(STATUS "LUCI INTERPRETER KERNELS")
 add_subdirectory(loader)
 message(STATUS "LUCI INTERPRETER LOADER")
+add_subdirectory(import)
+message(STATUS "LUCI INTERPRETER IMPORT")
 
 message(STATUS "LUCI INTERPTER INITALIZED")
 
diff --git a/compiler/luci-interpreter/src/Interpreter.cpp b/compiler/luci-interpreter/src/Interpreter.cpp
index 1b8792a6c..8cf272efd 100644
--- a/compiler/luci-interpreter/src/Interpreter.cpp
+++ b/compiler/luci-interpreter/src/Interpreter.cpp
@@ -70,25 +70,30 @@ private:
 
 } // namespace
 
+Interpreter::Interpreter(const luci::Module *module)
+{
+  _runtime_to_ir = std::make_unique<RuntimeToIR>();
+  _event_notifier = std::make_unique<EventNotifierImpl>(*_runtime_to_ir, _observers);
+  _runtime_module = std::make_unique<RuntimeModule>(_event_notifier.get());
+
+  _default_memory_manager = std::make_unique<SimpleMemoryManager>();
+
+  ModuleLoader loader(module, _runtime_module.get(), *_runtime_to_ir, _node_to_tensor,
+                      _default_memory_manager.get());
+  loader.load();
+}
+
 Interpreter::Interpreter(const luci::Module *module,
                          luci_interpreter::IMemoryManager *memory_manager)
 {
+  assert(memory_manager && "Use Interpreter::Interpreter(module) constructor instead");
+
   _runtime_to_ir = std::make_unique<RuntimeToIR>();
   _event_notifier = std::make_unique<EventNotifierImpl>(*_runtime_to_ir, _observers);
   _runtime_module = std::make_unique<RuntimeModule>(_event_notifier.get());
 
-  if (memory_manager == nullptr)
-  {
-    _default_memory_manager = std::make_unique<SimpleMemoryManager>();
-    _memory_manager = _default_memory_manager.get();
-  }
-  else
-  {
-    _memory_manager = memory_manager;
-  }
-
   ModuleLoader loader(module, _runtime_module.get(), *_runtime_to_ir, _node_to_tensor,
-                      _memory_manager);
+                      memory_manager);
   loader.load();
 }
 
diff --git a/compiler/luci-interpreter/src/core/CMakeLists.txt b/compiler/luci-interpreter/src/core/CMakeLists.txt
index 4430cba11..c2471e01c 100644
--- a/compiler/luci-interpreter/src/core/CMakeLists.txt
+++ b/compiler/luci-interpreter/src/core/CMakeLists.txt
@@ -10,7 +10,9 @@ set(SOURCES
     Tensor.cpp)
 
 add_library(${LUCI_INTERPRETER_CORE} STATIC ${SOURCES})
-set_target_properties(${LUCI_INTERPRETER_CORE} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+if (NOT NNCC_LIBRARY_NO_PIC)
+    set_target_properties(${LUCI_INTERPRETER_CORE} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif(NOT NNCC_LIBRARY_NO_PIC)
 target_include_directories(${LUCI_INTERPRETER_CORE} PUBLIC "${LUCI_INTERPRETER_INCLUDE_DIR}")
 target_include_directories(${LUCI_INTERPRETER_CORE} PUBLIC "${LUCI_INTERPRETER_SOURCE_DIR}")
 target_link_libraries(${LUCI_INTERPRETER_CORE} PUBLIC luci_lang)
diff --git a/compiler/luci-interpreter/src/core/KernelParams.h b/compiler/luci-interpreter/src/core/KernelParams.h
index ee0390fcc..958fd4b74 100644
--- a/compiler/luci-interpreter/src/core/KernelParams.h
+++ b/compiler/luci-interpreter/src/core/KernelParams.h
@@ -43,6 +43,12 @@ struct ArgMaxParams
   DataType output_type;
 };
 
+struct BatchMatMulParams
+{
+  bool adj_x;
+  bool adj_y;
+};
+
 struct ConcatenationParams
 {
   int axis;
@@ -83,6 +89,13 @@ struct DivParams
 struct FullyConnectedParams
 {
   Activation activation;
+  bool keep_num_dims = false;
+};
+
+struct GatherParams
+{
+  int32_t axis;
+  int32_t batch_dims;
 };
 
 struct InstanceNormParams
@@ -119,6 +132,11 @@ struct MulParams
   Activation activation;
 };
 
+struct OneHotParams
+{
+  int32_t axis;
+};
+
 struct PackParams
 {
   int32_t values_count;
@@ -157,6 +175,13 @@ struct SubParams
   Activation activation;
 };
 
+struct SVDFParams
+{
+  bool asymmetric_quantize_inputs;
+  int32_t svdf_rank;
+  Activation activation;
+};
+
 struct SpaceToDepthParams
 {
   int block_size;
diff --git a/compiler/luci-interpreter/src/import/CMakeLists.txt b/compiler/luci-interpreter/src/import/CMakeLists.txt
new file mode 100644
index 000000000..dd9733f92
--- /dev/null
+++ b/compiler/luci-interpreter/src/import/CMakeLists.txt
@@ -0,0 +1,15 @@
+set(SOURCES
+    "${LUCI_INTERPRETER_INCLUDE_DIR}/luci_interpreter/GraphBuilderRegistry.h"
+    GraphBuilderRegistry.cpp)
+
+# include specific builders
+file(GLOB_RECURSE NODES "Nodes/*")
+list(APPEND SOURCES ${NODES})
+
+add_library(${LUCI_INTERPRETER_IMPORT} STATIC ${SOURCES})
+if (NOT NNCC_LIBRARY_NO_PIC)
+  set_target_properties(${LUCI_INTERPRETER_IMPORT} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif(NOT NNCC_LIBRARY_NO_PIC)
+
+target_include_directories(${LUCI_INTERPRETER_IMPORT} PUBLIC "${LUCI_INTERPRETER_INCLUDE_DIR}")
+target_link_libraries(${LUCI_INTERPRETER_IMPORT} PUBLIC luci_import)
diff --git a/compiler/luci-interpreter/src/import/GraphBuilderRegistry.cpp b/compiler/luci-interpreter/src/import/GraphBuilderRegistry.cpp
new file mode 100644
index 000000000..a33bca6a4
--- /dev/null
+++ b/compiler/luci-interpreter/src/import/GraphBuilderRegistry.cpp
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "luci_interpreter/GraphBuilderRegistry.h"
+#include "Nodes/CircleReferencingConst.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<luci::GraphBuilderSource> source_without_constant_copying()
+{
+  auto builder = std::make_unique<luci::GraphBuilderRegistry>();
+  {
+    // redefine NodeBuilder of BUFFER type
+    builder->add(std::make_unique<CircleReferencingConstNodeBuilder>());
+  }
+
+  return builder;
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.cpp b/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.cpp
new file mode 100644
index 000000000..14e90f240
--- /dev/null
+++ b/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.cpp
@@ -0,0 +1,113 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "CircleReferencingConst.h"
+
+#include <vector>
+
+namespace
+{
+
+// helper struct which describes data loaded to custom_options of CircleReferencingConst node
+struct ConstDataReference
+{
+  const uint8_t *data = nullptr;
+  uint32_t size = 0;
+};
+
+} // namespace
+
+namespace luci_interpreter
+{
+using namespace luci;
+
+CircleNode *CircleReferencingConstNodeBuilder::build(TensorIndex tensor_index,
+                                                     GraphBuilderContext *context) const
+{
+  assert(tensor_index >= 0);
+
+  const auto graph = context->graph();
+  const auto reader = context->reader();
+  const auto tensors = reader->tensors();
+  auto const const_tensor = tensors[tensor_index];
+  assert(const_tensor != nullptr);
+  if (const_tensor->is_variable())
+  {
+    // Create CircleVariable for variable
+    return nullptr;
+  }
+
+  auto const buffer = wrap(reader->buffers()[const_tensor->buffer()]->data());
+  auto const const_dims = wrap(const_tensor->shape()); // in NHWC
+  if (const_dims.empty() && buffer.empty())
+  {
+    // unknown shape tensor and scalar tensor
+    return nullptr;
+  }
+
+  // if tensor_index is used as output to some other operator, this is not a constant
+  auto tensoroutputs = context->tensoroutputs();
+  if (tensoroutputs->find(tensor_index))
+  {
+    // other operator output tensor
+    return nullptr;
+  }
+
+  uint32_t num_elements = 1;
+  for (uint32_t r = 0; r < const_dims.size(); ++r)
+  {
+    num_elements = num_elements * const_dims[r];
+  }
+
+  if (buffer.empty() && num_elements > 0)
+  {
+    // normal empty tensor
+    return nullptr;
+  }
+
+  // create CircleReferencingConst
+  auto custom_node = graph->nodes()->create<CircleCustom>(0, 1);
+  {
+    custom_node->custom_code("CircleReferencingConst");
+
+    copy_tensor_attributes(const_tensor, custom_node);
+    custom_node->shape_status(luci::ShapeStatus::VALID);
+
+    // custom options stores size of buffer and pointer's value to buffer's data
+    {
+      std::vector<uint8_t> custom_options(sizeof(ConstDataReference));
+      {
+        auto &const_data_ref = *reinterpret_cast<ConstDataReference *>(custom_options.data());
+        const_data_ref = {buffer.data(), buffer.size()};
+      }
+      custom_node->custom_options(custom_options);
+    }
+  }
+
+  // Output of CircleCustom node presented with CircleConstNode
+  auto out_node = graph->nodes()->create<CircleCustomOut>();
+  {
+    out_node->index(0);
+    out_node->input(custom_node);
+
+    copy_tensor_attributes(const_tensor, out_node);
+    out_node->shape_status(luci::ShapeStatus::VALID);
+  }
+
+  return out_node;
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.h b/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.h
new file mode 100644
index 000000000..ed8f95124
--- /dev/null
+++ b/compiler/luci-interpreter/src/import/Nodes/CircleReferencingConst.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_INTERPRETER_IMPORT_OP_CIRCLE_REFERENCING_CONST_H__
+#define __LUCI_INTERPRETER_IMPORT_OP_CIRCLE_REFERENCING_CONST_H__
+
+#include <luci/Import/NodeBuilder.h>
+
+#include <luci/IR/Nodes/CircleConst.h>
+
+namespace luci_interpreter
+{
+using namespace luci;
+
+/**
+ * @brief Builder creates CircleCustom node with pointer to constants data from Tensor with buffer.
+ */
+class CircleReferencingConstNodeBuilder : public TypedNodeBuilder<NodeBuilderType::BUFFER>
+{
+public:
+  CircleNode *build(TensorIndex tensor_index, GraphBuilderContext *ctx) const final;
+};
+
+} // namespace luci_interpreter
+
+#endif // __LUCI_INTERPRETER_IMPORT_OP_CIRCLE_REFERENCING_CONST_H__
diff --git a/compiler/luci-interpreter/src/kernels/Add.cpp b/compiler/luci-interpreter/src/kernels/Add.cpp
index 7381c3849..d7bf3084f 100644
--- a/compiler/luci-interpreter/src/kernels/Add.cpp
+++ b/compiler/luci-interpreter/src/kernels/Add.cpp
@@ -38,8 +38,11 @@ Add::Add(const Tensor *input1, const Tensor *input2, Tensor *output, const AddPa
 void Add::configure()
 {
   LUCI_INTERPRETER_CHECK(input1()->element_type() == input2()->element_type());
+  LUCI_INTERPRETER_CHECK(input1()->element_type() == output()->element_type());
   if (input1()->element_type() == DataType::S16)
   {
+    LUCI_INTERPRETER_CHECK(input1()->zero_points().size() == 1 &&
+                           input2()->zero_points().size() == 1);
     LUCI_INTERPRETER_CHECK(input1()->zero_point() == 0 && input2()->zero_point() == 0 &&
                            output()->zero_point() == 0);
   }
@@ -54,6 +57,12 @@ void Add::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -67,13 +76,8 @@ void Add::execute() const
 
 void Add::evalFloat() const
 {
-  float activation_min{};
-  float activation_max{};
-  calculateActivationRange(_params.activation, &activation_min, &activation_max);
-
   tflite::ArithmeticParams params{};
-  params.float_activation_min = activation_min;
-  params.float_activation_max = activation_max;
+  fillArithmeticActivationRange<float>(params, _params.activation);
 
   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
     getTensorShape(input1()), getTensorShape(input2()), &params);
@@ -92,6 +96,28 @@ void Add::evalFloat() const
   }
 }
 
+template <typename T> void Add::evalInteger() const
+{
+  tflite::ArithmeticParams params{};
+  fillArithmeticActivationRange<T>(params, _params.activation);
+
+  const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
+    getTensorShape(input1()), getTensorShape(input2()), &params);
+
+  if (need_broadcast)
+  {
+    tflite::reference_ops::BroadcastAdd4DSlow(
+      params, getTensorShape(input1()), getTensorData<T>(input1()), getTensorShape(input2()),
+      getTensorData<T>(input2()), getTensorShape(output()), getTensorData<T>(output()));
+  }
+  else
+  {
+    tflite::reference_ops::Add(params, getTensorShape(input1()), getTensorData<T>(input1()),
+                               getTensorShape(input2()), getTensorData<T>(input2()),
+                               getTensorShape(output()), getTensorData<T>(output()));
+  }
+}
+
 void Add::evalQuantized() const
 {
   const auto input1_scale = static_cast<double>(input1()->scale());
diff --git a/compiler/luci-interpreter/src/kernels/Add.h b/compiler/luci-interpreter/src/kernels/Add.h
index 79518845d..91d95b6af 100644
--- a/compiler/luci-interpreter/src/kernels/Add.h
+++ b/compiler/luci-interpreter/src/kernels/Add.h
@@ -39,6 +39,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
   void evalQuantizedS16() const;
 };
diff --git a/compiler/luci-interpreter/src/kernels/Add.test.cpp b/compiler/luci-interpreter/src/kernels/Add.test.cpp
index 847b65667..b8b1c3089 100644
--- a/compiler/luci-interpreter/src/kernels/Add.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Add.test.cpp
@@ -166,6 +166,69 @@ TEST_F(AddTest, Float)
   }
 }
 
+template <loco::DataType DType> void CheckInteger(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  Shape base_shape = {2, 3, 1, 2};
+  std::vector<Shape> test_shapes{{1, 1, 3, 2}, {1, 3, 1, 2}, {2, 1, 3, 1}, {2, 3, 1, 1}};
+  std::vector<std::vector<dtype>> test_outputs = {
+    {3, 3, 0, 1, 0, 8, 5,  1, 0, 0, 2, 6, 8,  0, 1, 0, 5, 1,
+     5, 4, 0, 2, 2, 9, 11, 0, 4, 0, 8, 5, 11, 2, 4, 0, 8, 7},
+    {3, 3, 0, 0, 5, 1, 5, 4, 4, 0, 8, 7},
+    {3, 6, 0, 3, 0, 0, 5, 4, 2, 1, 0,  0, 8, 0, 5, 0, 1,  0,
+     0, 2, 2, 4, 7, 9, 6, 0, 8, 0, 13, 5, 6, 0, 8, 2, 13, 7},
+    {3, 6, 2, 1, 1, 0, 0, 2, 8, 0, 13, 7}};
+  std::vector<dtype> input1_data{-1, 2, 1, 0, 4, -5, 1, 3, 7, -1, 7, 1};
+  std::vector<dtype> input2_data{4, 1, -3, -1, 1, 6};
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    AddParams params{};
+    params.activation = Activation::RELU;
+
+    Add kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+  // Re-run with exchanged inputs.
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    AddParams params{};
+    params.activation = Activation::RELU;
+
+    Add kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+};
+
+TEST_F(AddTest, SInt32)
+{
+  CheckInteger<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(AddTest, SInt64)
+{
+  CheckInteger<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 TEST_F(AddTest, SInt16)
 {
   Shape base_shape = {2, 3, 1, 2};
@@ -248,11 +311,24 @@ TEST_F(AddTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
-TEST_F(AddTest, Invalid_Input_Type_NEG)
+TEST_F(AddTest, Invalid_Output_Type_NEG)
 {
   Tensor input1_tensor = makeInputTensor<DataType::S64>({1}, {1}, _memory_manager.get());
   Tensor input2_tensor = makeInputTensor<DataType::S64>({1}, {2}, _memory_manager.get());
-  Tensor output_tensor = makeOutputTensor(DataType::S64);
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  AddParams params{};
+  params.activation = Activation::RELU;
+
+  Add kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(AddTest, Invalid_Input_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::U64>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::U64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U64);
 
   AddParams params{};
   params.activation = Activation::RELU;
@@ -263,6 +339,19 @@ TEST_F(AddTest, Invalid_Input_Type_NEG)
   EXPECT_ANY_THROW(kernel.execute());
 }
 
+TEST_F(AddTest, Invalid_Quantization_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::S16>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S16>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16);
+
+  AddParams params{};
+  params.activation = Activation::NONE;
+
+  Add kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/ArgMax.test.cpp b/compiler/luci-interpreter/src/kernels/ArgMax.test.cpp
index 119c69ccf..474f4b321 100644
--- a/compiler/luci-interpreter/src/kernels/ArgMax.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/ArgMax.test.cpp
@@ -57,7 +57,7 @@ template <typename T> class ArgMaxTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(ArgMaxTest, DataTypes);
+TYPED_TEST_SUITE(ArgMaxTest, DataTypes);
 
 TYPED_TEST(ArgMaxTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/AveragePool2D.cpp b/compiler/luci-interpreter/src/kernels/AveragePool2D.cpp
index 5545fb4d4..d3bade9e4 100644
--- a/compiler/luci-interpreter/src/kernels/AveragePool2D.cpp
+++ b/compiler/luci-interpreter/src/kernels/AveragePool2D.cpp
@@ -18,8 +18,7 @@
 
 #include "kernels/Utils.h"
 
-#include <tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h>
-#include <tensorflow/lite/kernels/internal/reference/pooling.h>
+#include "PALAveragePool2d.h"
 
 #include <stdexcept>
 
@@ -29,8 +28,9 @@ namespace luci_interpreter
 namespace kernels
 {
 
-AveragePool2D::AveragePool2D(const Tensor *input, Tensor *output, const Pool2DParams &params)
-  : KernelWithParams<Pool2DParams>({input}, {output}, params)
+AveragePool2D::AveragePool2D(const Tensor *input, Tensor *output, Tensor *scratchpad,
+                             const Pool2DParams &params)
+  : KernelWithParams<Pool2DParams>({input}, {output, scratchpad}, params)
 {
 }
 
@@ -76,6 +76,10 @@ void AveragePool2D::configure()
     LUCI_INTERPRETER_CHECK(output()->zero_point() == input()->zero_point());
   }
   output()->resize({batches, output_height, output_width, depth});
+
+  auto scratchpad = getOutputTensors()[1];
+  luci_interpreter_pal::SetupScratchpadTensor(scratchpad, input()->element_type(),
+                                              getTensorShape(input()), getTensorShape(output()));
 }
 
 void AveragePool2D::execute() const
@@ -155,9 +159,14 @@ void AveragePool2D::evalSInt8() const
   params.quantized_activation_min = activation_min;
   params.quantized_activation_max = activation_max;
 
-  tflite::reference_integer_ops::AveragePool(
+  auto scratchpad = getOutputTensors()[1];
+  int8_t *scratchpad_data = nullptr;
+  if (scratchpad->is_allocatable())
+    scratchpad_data = scratchpad->data<int8_t>();
+
+  luci_interpreter_pal::AveragePool<int8_t>(
     params, getTensorShape(input()), getTensorData<int8_t>(input()), getTensorShape(output()),
-    getTensorData<int8_t>(output()));
+    getTensorData<int8_t>(output()), getTensorShape(scratchpad), scratchpad_data);
 }
 
 void AveragePool2D::evalSInt16() const
diff --git a/compiler/luci-interpreter/src/kernels/AveragePool2D.h b/compiler/luci-interpreter/src/kernels/AveragePool2D.h
index b98367f31..2c8fe16e7 100644
--- a/compiler/luci-interpreter/src/kernels/AveragePool2D.h
+++ b/compiler/luci-interpreter/src/kernels/AveragePool2D.h
@@ -28,7 +28,8 @@ namespace kernels
 class AveragePool2D : public KernelWithParams<Pool2DParams>
 {
 public:
-  AveragePool2D(const Tensor *input, Tensor *output, const Pool2DParams &params);
+  AveragePool2D(const Tensor *input, Tensor *output, Tensor *scratchpad,
+                const Pool2DParams &params);
 
   const Tensor *input() const { return _inputs[0]; }
   Tensor *output() const { return _outputs[0]; }
diff --git a/compiler/luci-interpreter/src/kernels/AveragePool2D.test.cpp b/compiler/luci-interpreter/src/kernels/AveragePool2D.test.cpp
index 7ed421129..478bfa68e 100644
--- a/compiler/luci-interpreter/src/kernels/AveragePool2D.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/AveragePool2D.test.cpp
@@ -46,6 +46,7 @@ TEST_F(AveragePool2DTest, Float)
   Tensor input_tensor =
     makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -55,8 +56,9 @@ TEST_F(AveragePool2DTest, Float)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   kernel.configure();
+  _memory_manager->allocate_memory(scratchpad);
   _memory_manager->allocate_memory(output_tensor);
   kernel.execute();
 
@@ -78,6 +80,7 @@ TEST_F(AveragePool2DTest, Uint8_0)
   Tensor input_tensor = makeInputTensor<DataType::U8>(
     {1, 2, 4, 1}, quant_param.first, quant_param.second, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::U8, quant_param.first, quant_param.second);
+  Tensor scratchpad(DataType::U8, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -87,8 +90,9 @@ TEST_F(AveragePool2DTest, Uint8_0)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   kernel.configure();
+  _memory_manager->allocate_memory(scratchpad);
   _memory_manager->allocate_memory(output_tensor);
   kernel.execute();
 
@@ -107,6 +111,7 @@ TEST_F(AveragePool2DTest, Uint8_1)
   Tensor input_tensor = makeInputTensor<DataType::U8>(
     {1, 2, 4, 1}, quant_param.first, quant_param.second, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::U8, quant_param.first, quant_param.second);
+  Tensor scratchpad(DataType::U8, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -116,9 +121,10 @@ TEST_F(AveragePool2DTest, Uint8_1)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   EXPECT_THAT(dequantizeTensorData(output_tensor), FloatArrayNear({2.75, 6.0}));
@@ -141,6 +147,7 @@ TEST_F(AveragePool2DTest, SInt16)
   Tensor input_tensor =
     makeInputTensor<DataType::S16>(input_shape, 0.5, 0, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::S16, 0.5, 0);
+  Tensor scratchpad(DataType::S16, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -150,8 +157,9 @@ TEST_F(AveragePool2DTest, SInt16)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   kernel.configure();
+  _memory_manager->allocate_memory(scratchpad);
   _memory_manager->allocate_memory(output_tensor);
   kernel.execute();
 
@@ -174,6 +182,7 @@ TEST_F(AveragePool2DTest, SInt8)
   Tensor input_tensor = makeInputTensor<DataType::S8>(
     input_shape, quant_param.first, quant_param.second, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::S8, quant_param.first, quant_param.second);
+  Tensor scratchpad(DataType::S8, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -183,8 +192,9 @@ TEST_F(AveragePool2DTest, SInt8)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   kernel.configure();
+  _memory_manager->allocate_memory(scratchpad);
   _memory_manager->allocate_memory(output_tensor);
   kernel.execute();
 
@@ -203,6 +213,7 @@ TEST_F(AveragePool2DTest, Invalid_Input_Shape_NEG)
   Tensor input_tensor =
     makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -212,7 +223,7 @@ TEST_F(AveragePool2DTest, Invalid_Input_Shape_NEG)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -227,6 +238,7 @@ TEST_F(AveragePool2DTest, In_Out_Type_NEG)
   Tensor input_tensor =
     makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::U8);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -236,7 +248,7 @@ TEST_F(AveragePool2DTest, In_Out_Type_NEG)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -252,6 +264,7 @@ TEST_F(AveragePool2DTest, Quant_Param_NEG)
   Tensor input_tensor = makeInputTensor<DataType::U8>(
     {1, 2, 4, 1}, quant_param1.first, quant_param1.second, input_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::U8, quant_param2.first, quant_param2.second);
+  Tensor scratchpad(DataType::U8, Shape({}), {}, "");
 
   Pool2DParams params{};
   params.padding = Padding::VALID;
@@ -261,7 +274,7 @@ TEST_F(AveragePool2DTest, Quant_Param_NEG)
   params.stride_width = 2;
   params.activation = Activation::RELU6;
 
-  AveragePool2D kernel(&input_tensor, &output_tensor, params);
+  AveragePool2D kernel(&input_tensor, &output_tensor, &scratchpad, params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
diff --git a/compiler/luci-interpreter/src/kernels/BatchMatMul.cpp b/compiler/luci-interpreter/src/kernels/BatchMatMul.cpp
new file mode 100644
index 000000000..24ca22996
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/BatchMatMul.cpp
@@ -0,0 +1,188 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/BatchMatMul.h"
+#include "kernels/Utils.h"
+
+#include "PALBatchMatMul.h"
+
+#include <tensorflow/lite/kernels/internal/reference/transpose.h>
+
+#include <stdexcept>
+
+namespace
+{
+
+tflite::RuntimeShape SwapRowColumnDims(const tflite::RuntimeShape &shape)
+{
+  tflite::RuntimeShape swapped_shape(shape);
+  const int32_t dims = shape.DimensionsCount();
+  swapped_shape.SetDim(dims - 2, shape.Dims(dims - 1));
+  swapped_shape.SetDim(dims - 1, shape.Dims(dims - 2));
+  return swapped_shape;
+}
+
+} // namespace
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+BatchMatMul::BatchMatMul(const Tensor *x, const Tensor *y, Tensor *output, Tensor *x_tmp,
+                         Tensor *y_tmp, const BatchMatMulParams &params)
+  : KernelWithParams({x, y}, {output, x_tmp, y_tmp}, params)
+{
+}
+
+void BatchMatMul::configure()
+{
+  auto lhs = x();
+  auto rhs = y();
+  auto adj_x = params().adj_x;
+  auto adj_y = params().adj_y;
+
+  // TODO Support non-float types
+  if (lhs->element_type() != DataType::FLOAT32 || rhs->element_type() != DataType::FLOAT32)
+    throw std::runtime_error("Unsupported type.");
+
+  LUCI_INTERPRETER_CHECK(lhs->element_type() == rhs->element_type());
+
+  auto lhs_rank = lhs->shape().num_dims();
+  auto rhs_rank = rhs->shape().num_dims();
+  LUCI_INTERPRETER_CHECK(lhs_rank >= 2 && lhs_rank <= 4);
+  LUCI_INTERPRETER_CHECK(rhs_rank >= 2 && rhs_rank <= 4);
+
+  auto lhs_scratchpad = temp_lhs();
+  auto rhs_scratchpad = temp_rhs();
+  luci_interpreter_pal::SetupScratchpadTensor(lhs_scratchpad, rhs_scratchpad, getTensorShape(lhs),
+                                              getTensorShape(rhs));
+
+  auto output_rank = std::max(lhs_rank, rhs_rank);
+
+  auto extended_lhs_shape = tflite::RuntimeShape::ExtendedShape(output_rank, getTensorShape(lhs));
+  auto extended_rhs_shape = tflite::RuntimeShape::ExtendedShape(output_rank, getTensorShape(rhs));
+
+  // Ensure any batch dimensions obey broacasting rules.
+  for (int i = 0; i < output_rank - 2; ++i)
+  {
+    const int lhs_dim = extended_lhs_shape.Dims(i);
+    const int rhs_dim = extended_rhs_shape.Dims(i);
+    if (lhs_dim != rhs_dim)
+    {
+      if (lhs_dim != 1)
+      {
+        LUCI_INTERPRETER_CHECK(rhs_dim == 1);
+      }
+    }
+  }
+
+  // Ensure other dimensions work for matrix multiplication.
+  int accum_dim_lhs =
+    adj_x ? extended_lhs_shape.Dims(output_rank - 2) : extended_lhs_shape.Dims(output_rank - 1);
+  int accum_dim_rhs =
+    adj_y ? extended_rhs_shape.Dims(output_rank - 1) : extended_rhs_shape.Dims(output_rank - 2);
+  LUCI_INTERPRETER_CHECK(accum_dim_lhs == accum_dim_rhs);
+
+  Shape output_shape(output_rank);
+  // Fill in any broadcast dimensions.
+  for (int i = 0; i < output_rank - 2; ++i)
+  {
+    const int lhs_dim = extended_lhs_shape.Dims(i);
+    const int rhs_dim = extended_rhs_shape.Dims(i);
+    int broadcast_dim = lhs_dim;
+    if ((lhs_dim != rhs_dim) && (lhs_dim == 1))
+    {
+      broadcast_dim = rhs_dim;
+    }
+    output_shape.dim(i) = broadcast_dim;
+  }
+  // Fill in the matmul dimensions.
+  int lhs_rows_index = adj_x ? output_rank - 1 : output_rank - 2;
+  int rhs_cols_index = adj_y ? output_rank - 2 : output_rank - 1;
+
+  output_shape.dim(output_rank - 2) = extended_lhs_shape.Dims(lhs_rows_index);
+  output_shape.dim(output_rank - 1) = extended_rhs_shape.Dims(rhs_cols_index);
+
+  output()->resize(output_shape);
+}
+
+void TransposeRowsColumns(const Tensor *tensor_in, Tensor *tensor_out)
+{
+  tflite::RuntimeShape transposed_shape(getTensorShape(tensor_in));
+  tflite::RuntimeShape shape(getTensorShape(tensor_in));
+  tflite::TransposeParams params;
+  int rank = shape.DimensionsCount();
+  params.perm_count = rank;
+  for (int i = 0; i < rank - 2; ++i)
+  {
+    params.perm[i] = i;
+  }
+  // Transpose the last two dimensions.
+  params.perm[rank - 2] = rank - 1;
+  params.perm[rank - 1] = rank - 2;
+  transposed_shape.SetDim(rank - 1, shape.Dims(rank - 2));
+  transposed_shape.SetDim(rank - 2, shape.Dims(rank - 1));
+  switch (tensor_in->element_type())
+  {
+    case DataType::FLOAT32:
+      tflite::reference_ops::Transpose(params, shape, getTensorData<float>(tensor_in),
+                                       transposed_shape, getTensorData<float>(tensor_out));
+      break;
+    default:
+      throw std::runtime_error("Only suppport fp32 BatchMatMul for now.");
+  }
+}
+
+void BatchMatMul::execute() const
+{
+  auto lhs = x();
+  auto rhs = y();
+
+  bool adj_x = params().adj_x;
+  bool adj_y = params().adj_y;
+
+  auto orig_lhs_shape = getTensorShape(lhs);
+  auto orig_rhs_shape = getTensorShape(rhs);
+
+  auto rhs_tensor = adj_y ? rhs : temp_rhs();
+  auto lhs_tensor = adj_x ? temp_lhs() : lhs;
+  if (not adj_y)
+  {
+    TransposeRowsColumns(rhs, temp_rhs());
+  }
+  if (adj_x)
+  {
+    TransposeRowsColumns(lhs, temp_lhs());
+  }
+  tflite::RuntimeShape rhs_shape = adj_y ? orig_rhs_shape : SwapRowColumnDims(orig_rhs_shape);
+  tflite::RuntimeShape lhs_shape = adj_x ? orig_lhs_shape : SwapRowColumnDims(orig_lhs_shape);
+
+  switch (x()->element_type())
+  {
+    case DataType::FLOAT32:
+      luci_interpreter_pal::BatchMatMul(rhs_shape, getTensorData<float>(rhs_tensor), lhs_shape,
+                                        getTensorData<float>(lhs_tensor), getTensorShape(output()),
+                                        getTensorData<float>(output()));
+      break;
+    default:
+      throw std::runtime_error("Unsupported type.");
+  }
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/BatchMatMul.h b/compiler/luci-interpreter/src/kernels/BatchMatMul.h
new file mode 100644
index 000000000..744f49795
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/BatchMatMul.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_BATCHMATMUL_H
+#define LUCI_INTERPRETER_KERNELS_BATCHMATMUL_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class BatchMatMul : public KernelWithParams<BatchMatMulParams>
+{
+public:
+  BatchMatMul(const Tensor *x, const Tensor *y, Tensor *output, Tensor *x_tmp, Tensor *y_tmp,
+              const BatchMatMulParams &params);
+
+  const Tensor *x() const { return _inputs[0]; }
+  const Tensor *y() const { return _inputs[1]; }
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+
+private:
+  Tensor *temp_lhs() const { return _outputs[1]; }
+  Tensor *temp_rhs() const { return _outputs[2]; }
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_BATCHMATMUL_H
diff --git a/compiler/luci-interpreter/src/kernels/BatchMatMul.test.cpp b/compiler/luci-interpreter/src/kernels/BatchMatMul.test.cpp
new file mode 100644
index 000000000..edfa3a685
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/BatchMatMul.test.cpp
@@ -0,0 +1,272 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/BatchMatMul.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class BatchMatMulTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(BatchMatMulTest, Float)
+{
+  std::vector<float> lhs_data = {1, 2, 3, 4, 5, 6};
+  std::vector<float> rhs_data = {7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18};
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 2, 3}, lhs_data, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 3, 4}, rhs_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(lhs_scratch);
+  _memory_manager->allocate_memory(rhs_scratch);
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              FloatArrayNear({74., 80., 86., 92., 173., 188., 203., 218.}));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 4}));
+}
+
+TEST_F(BatchMatMulTest, Float_SimpleRHSAdjoint)
+{
+  std::vector<float> lhs_data = {1, 2, 3, 4, 5, 6};
+  std::vector<float> rhs_data = {7, 11, 15, 8, 12, 16, 9, 13, 17, 10, 14, 18};
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 2, 3}, lhs_data, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 4, 3}, rhs_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = true;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(lhs_scratch);
+  _memory_manager->allocate_memory(rhs_scratch);
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              FloatArrayNear({74., 80., 86., 92., 173., 188., 203., 218.}));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 4}));
+}
+
+TEST_F(BatchMatMulTest, Float_SimpleLHSAdjoint)
+{
+  std::vector<float> lhs_data = {1, 4, 2, 5, 3, 6};
+  std::vector<float> rhs_data = {7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18};
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 3, 2}, lhs_data, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 3, 4}, rhs_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = true;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(lhs_scratch);
+  _memory_manager->allocate_memory(rhs_scratch);
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              FloatArrayNear({74., 80., 86., 92., 173., 188., 203., 218.}));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 4}));
+}
+
+TEST_F(BatchMatMulTest, Float_BatchSizeTwo)
+{
+  std::vector<float> lhs_data = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
+  std::vector<float> rhs_data = {7,  8,  9,  10, 11, 12, 13, 14, 15, 16, 17, 18,
+                                 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30};
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 2, 3}, lhs_data, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 3, 4}, rhs_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(lhs_scratch);
+  _memory_manager->allocate_memory(rhs_scratch);
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              FloatArrayNear({74., 80., 86., 92., 173., 188., 203., 218., 560., 584., 608., 632.,
+                              767., 800., 833., 866.}));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 2, 4}));
+}
+
+TEST_F(BatchMatMulTest, Float_DiffBatch)
+{
+  std::vector<float> lhs_data = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
+  std::vector<float> rhs_data = {7,  8,  9,  10, 11, 12, 13, 14, 15, 16, 17, 18,
+                                 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30};
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 1, 6}, lhs_data, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 6, 4}, rhs_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(lhs_scratch);
+  _memory_manager->allocate_memory(rhs_scratch);
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              FloatArrayNear({427., 448., 469., 490., 1039., 1096., 1153., 1210.}));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 1, 4}));
+}
+
+TEST_F(BatchMatMulTest, Invalid_Shape_NEG)
+{
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 2, 2}, {1, 2, 3, 4}, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 3, 2}, {5, 6, 7, 8, 9, 10}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(BatchMatMulTest, Invalid_Batch_NEG)
+{
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 1, 3}, {1, 2, 3, 4, 5, 6}, _memory_manager.get());
+  Tensor rhs_tensor = makeInputTensor<DataType::FLOAT32>({3, 3, 1}, {5, 6, 7, 8, 9, 10, 11, 12, 13},
+                                                         _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(BatchMatMulTest, Invalid_Rank_NEG)
+{
+  Tensor lhs_tensor = makeInputTensor<DataType::FLOAT32>({4}, {1, 2, 3, 4}, _memory_manager.get());
+  Tensor rhs_tensor = makeInputTensor<DataType::FLOAT32>({1, 4, 2}, {5, 6, 7, 8, 9, 10, 11, 12},
+                                                         _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(BatchMatMulTest, Invalid_Rank2_NEG)
+{
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 1, 1, 1, 4}, {1, 2, 3, 4}, _memory_manager.get());
+  Tensor rhs_tensor = makeInputTensor<DataType::FLOAT32>({1, 4, 2}, {5, 6, 7, 8, 9, 10, 11, 12},
+                                                         _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(BatchMatMulTest, TypeMisMatch_NEG)
+{
+  Tensor lhs_tensor =
+    makeInputTensor<DataType::U8>({1, 2, 3}, {1, 2, 3, 4, 5, 6}, _memory_manager.get());
+  Tensor rhs_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 3, 2}, {5, 6, 7, 8, 9, 10}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor lhs_scratch(DataType::U8, Shape({}), {}, "");
+  Tensor rhs_scratch(DataType::FLOAT32, Shape({}), {}, "");
+
+  BatchMatMulParams params;
+  params.adj_x = false;
+  params.adj_y = false;
+
+  BatchMatMul kernel(&lhs_tensor, &rhs_tensor, &output_tensor, &lhs_scratch, &rhs_scratch, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/BatchToSpaceND.test.cpp b/compiler/luci-interpreter/src/kernels/BatchToSpaceND.test.cpp
index f3a344974..52647a763 100644
--- a/compiler/luci-interpreter/src/kernels/BatchToSpaceND.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/BatchToSpaceND.test.cpp
@@ -58,7 +58,7 @@ template <typename T> class BatchToSpaceNDTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(BatchToSpaceNDTest, DataTypes);
+TYPED_TEST_SUITE(BatchToSpaceNDTest, DataTypes);
 
 TYPED_TEST(BatchToSpaceNDTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/CMakeLists.txt b/compiler/luci-interpreter/src/kernels/CMakeLists.txt
index 1b7d0f66a..9f4ba0e0b 100644
--- a/compiler/luci-interpreter/src/kernels/CMakeLists.txt
+++ b/compiler/luci-interpreter/src/kernels/CMakeLists.txt
@@ -15,7 +15,9 @@ endmacro(REGISTER_KERNEL)
 include(${KERNEL_REGISTER_FILE})
 
 add_library(${LUCI_INTERPRETER_KERNELS} STATIC ${SOURCES})
-set_target_properties(${LUCI_INTERPRETER_KERNELS} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+if (NOT NNCC_LIBRARY_NO_PIC)
+  set_target_properties(${LUCI_INTERPRETER_KERNELS} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif(NOT NNCC_LIBRARY_NO_PIC)
 target_include_directories(${LUCI_INTERPRETER_KERNELS} PUBLIC ${LUCI_INTERPRETER_SOURCE_DIR})
 
 target_link_libraries(${LUCI_INTERPRETER_KERNELS} PUBLIC ${LUCI_INTERPRETER_CORE})
diff --git a/compiler/luci-interpreter/src/kernels/Cast.test.cpp b/compiler/luci-interpreter/src/kernels/Cast.test.cpp
index 731260522..4713ad34c 100644
--- a/compiler/luci-interpreter/src/kernels/Cast.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Cast.test.cpp
@@ -79,7 +79,7 @@ template <typename T> class CastTest : public ::testing::Test
 
 using IntDataTypes =
   ::testing::Types<uint8_t, uint16_t, uint32_t, uint64_t, int8_t, int16_t, int32_t, int64_t>;
-TYPED_TEST_CASE(CastTest, IntDataTypes);
+TYPED_TEST_SUITE(CastTest, IntDataTypes);
 
 TYPED_TEST(CastTest, FloatToInt)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Concatenation.cpp b/compiler/luci-interpreter/src/kernels/Concatenation.cpp
index 7cfdf34b9..46ee5941e 100644
--- a/compiler/luci-interpreter/src/kernels/Concatenation.cpp
+++ b/compiler/luci-interpreter/src/kernels/Concatenation.cpp
@@ -69,11 +69,21 @@ void Concatenation::configure()
   Shape output_shape = t0->shape();
   output_shape.dim(axis) = sum_axis;
 
-  // TODO S8 type needs more checking: quantization parameters of all input tensors and the output
-  //  tensor should be the same. Note that there is no such requirement for U8 type.
-  if (t0->element_type() == DataType::S8)
-    throw std::runtime_error("Unsupported type.");
+  // If input tensors are INT8 type then quantization parameters of all input tensors and the output
+  // should be the same
+  for (auto current_tensor : _inputs)
+  {
+    if (current_tensor->element_type() == DataType::S8)
+    {
+      LUCI_INTERPRETER_CHECK(current_tensor->quantized_dimension() ==
+                             output()->quantized_dimension());
 
+      LUCI_INTERPRETER_CHECK(current_tensor->zero_points().size() ==
+                             current_tensor->scales().size());
+      LUCI_INTERPRETER_CHECK(current_tensor->zero_points() == output()->zero_points());
+      LUCI_INTERPRETER_CHECK(current_tensor->scales() == output()->scales());
+    }
+  }
   output()->resize(output_shape);
 }
 
diff --git a/compiler/luci-interpreter/src/kernels/Concatenation.test.cpp b/compiler/luci-interpreter/src/kernels/Concatenation.test.cpp
index e4b50611a..f893b38fd 100644
--- a/compiler/luci-interpreter/src/kernels/Concatenation.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Concatenation.test.cpp
@@ -183,12 +183,12 @@ TEST_F(ConcatenationTest, Mismatching_Input_Dimension_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
-TEST_F(ConcatenationTest, Unsupported_Configure_Type_NEG)
+TEST_F(ConcatenationTest, Int8_Mismatching_Input_Type_NEG)
 {
-  std::vector<int8_t> input1_data{1, 2, 3, 4, 5, 6};
-  std::vector<int8_t> input2_data{7, 8, 9, 10, 11, 12};
-  Tensor input1_tensor = makeInputTensor<DataType::S8>({2, 3}, input1_data, _memory_manager.get());
-  Tensor input2_tensor = makeInputTensor<DataType::S8>({2, 3}, input2_data, _memory_manager.get());
+  std::vector<uint8_t> input1_data{1, 2, 3, 4};
+  std::vector<int8_t> input2_data{5, 6, 7, 8};
+  Tensor input1_tensor = makeInputTensor<DataType::U8>({2, 2}, input1_data, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S8>({2, 2}, input2_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::S8);
   ConcatenationParams params{};
 
@@ -199,6 +199,51 @@ TEST_F(ConcatenationTest, Unsupported_Configure_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(ConcatenationTest, Int8_Mismatching_Input_Output_Quant_Params_NEG)
+{
+  std::vector<float> input1_data{1, 2, 3, 4, 5, 6};
+  std::vector<float> input2_data{7, 8, 9, 10, 11, 12};
+  int quantized_dimension = 3;
+  std::vector<float> scales{0.1, 0.2, 0.3};
+  std::vector<int32_t> zero_points{1, -1, 1};
+
+  Tensor input1_tensor = makeInputTensor<DataType::S8>(
+    {1, 1, 2, 3}, scales, zero_points, quantized_dimension, input1_data, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S8>(
+    {1, 1, 2, 3}, scales, zero_points, quantized_dimension, input2_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, scales.at(0), zero_points.at(0));
+  ConcatenationParams params{};
+
+  params.axis = -1;
+  params.activation = luci::FusedActFunc::NONE;
+
+  Concatenation kernel({&input1_tensor, &input2_tensor}, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(ConcatenationTest, Int8_Mismatching_Zero_Point_NEG)
+{
+  std::vector<float> input1_data{1, 2, 3, 4};
+  std::vector<float> input2_data{5, 6, 7, 8};
+  float scale = 0.1;
+  int32_t zero_point_1 = 1;
+  int32_t zero_point_2 = -1;
+
+  Tensor input1_tensor =
+    makeInputTensor<DataType::S8>({2, 2}, scale, zero_point_1, input1_data, _memory_manager.get());
+  Tensor input2_tensor =
+    makeInputTensor<DataType::S8>({2, 2}, scale, zero_point_2, input2_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::S8, scale, zero_point_1);
+  ConcatenationParams params{};
+
+  params.axis = -1;
+  params.activation = luci::FusedActFunc::NONE;
+
+  Concatenation kernel({&input1_tensor, &input2_tensor}, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 // TODO: Remove this test when concat w/ fused_activation is supported
 TEST_F(ConcatenationTest, With_Fused_Activation_NEG)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Conv2D.cpp b/compiler/luci-interpreter/src/kernels/Conv2D.cpp
index 5647f4c44..234f95425 100644
--- a/compiler/luci-interpreter/src/kernels/Conv2D.cpp
+++ b/compiler/luci-interpreter/src/kernels/Conv2D.cpp
@@ -30,8 +30,8 @@ namespace kernels
 {
 
 Conv2D::Conv2D(const Tensor *input, const Tensor *filter, const Tensor *bias, Tensor *output,
-               Tensor *im2col, const Conv2DParams &params)
-  : KernelWithParams<Conv2DParams>({input, filter, bias}, {output, im2col}, params)
+               Tensor *scratchpad, const Conv2DParams &params)
+  : KernelWithParams<Conv2DParams>({input, filter, bias}, {output, scratchpad}, params)
 {
 }
 
@@ -108,27 +108,18 @@ void Conv2D::configure()
 
   output()->resize({batches, output_height, output_width, output_depth});
 
-  // Allocate tensor for Im2Col, if needed.
-  // The checks here should be aligned with the actual implementation.
-  const bool need_dilated_im2col =
-    _params.dilation_height_factor != 1 || _params.dilation_width_factor != 1;
-  const bool need_non_dilated_im2col = _params.stride_height != 1 || _params.stride_width != 1 ||
-                                       filter_height != 1 || filter_width != 1;
-  _need_im2col =
-    input()->element_type() != DataType::S16 && (need_dilated_im2col || need_non_dilated_im2col);
-  if (_need_im2col)
-  {
-    const int input_depth = input_shape.dim(3);
-    Shape im2col_shape{batches, output_height, output_width,
-                       input_depth * filter_height * filter_width};
-    auto im2col = getOutputTensors()[1];
-    im2col->resize(im2col_shape);
-  }
-  else
-  {
-    auto im2col = getOutputTensors()[1];
-    im2col->set_allocatable(false);
-  }
+  // Allocate tensor for scratchpad, if needed.
+  tflite::ConvParams params{};
+  params.padding_values.height = _padding_height;
+  params.padding_values.width = _padding_width;
+  params.stride_height = _params.stride_height;
+  params.stride_width = _params.stride_width;
+  params.dilation_height_factor = _params.dilation_height_factor;
+  params.dilation_width_factor = _params.dilation_width_factor;
+  auto scratchpad = getOutputTensors()[1];
+  luci_interpreter_pal::SetupScratchpadTensor(scratchpad, input()->element_type(), params,
+                                              getTensorShape(input()), getTensorShape(filter()),
+                                              getTensorShape(output()));
 
   switch (_params.activation)
   {
@@ -193,16 +184,16 @@ void Conv2D::evalFloat() const
   params.float_activation_min = activation_min;
   params.float_activation_max = activation_max;
 
-  float *im2col_data = nullptr;
-  auto im2col = getOutputTensors()[1];
-  if (_need_im2col)
-  {
-    im2col_data = im2col->data<float>();
-  }
-  luci_interpreter_pal::Conv(
-    params, getTensorShape(input()), getTensorData<float>(input()), getTensorShape(filter()),
-    getTensorData<float>(filter()), getTensorShape(bias()), getTensorData<float>(bias()),
-    getTensorShape(output()), getTensorData<float>(output()), getTensorShape(im2col), im2col_data);
+  auto scratchpad = getOutputTensors()[1];
+  float *scratchpad_data = nullptr;
+  if (scratchpad->is_allocatable())
+    scratchpad_data = scratchpad->data<float>();
+
+  luci_interpreter_pal::Conv(params, getTensorShape(input()), getTensorData<float>(input()),
+                             getTensorShape(filter()), getTensorData<float>(filter()),
+                             getTensorShape(bias()), getTensorData<float>(bias()),
+                             getTensorShape(output()), getTensorData<float>(output()),
+                             getTensorShape(scratchpad), scratchpad_data);
 }
 
 void Conv2D::evalQuantized() const
@@ -236,12 +227,12 @@ void Conv2D::evalQuantized() const
   params.quantized_activation_min = activation_min;
   params.quantized_activation_max = activation_max;
 
-  auto im2col = getOutputTensors()[1];
+  auto scratchpad = getOutputTensors()[1];
   luci_interpreter_pal::Conv(params, getTensorShape(input()), getTensorData<uint8_t>(input()),
                              getTensorShape(filter()), getTensorData<uint8_t>(filter()),
                              getTensorShape(bias()), getTensorData<int32_t>(bias()),
                              getTensorShape(output()), getTensorData<uint8_t>(output()),
-                             getTensorShape(im2col), getTensorData<uint8_t>(im2col));
+                             getTensorShape(scratchpad), getTensorData<uint8_t>(scratchpad));
 }
 
 void Conv2D::evalQuantizedPerChannel() const
@@ -364,18 +355,16 @@ void Conv2D::evalQuantizedS8PerChannel() const
                  std::back_inserter(multipliers),
                  [](ChannelQuantMultipliers cm) { return cm.multiplier; });
 
-  int8_t *im2col_data = nullptr;
-  auto im2col = getOutputTensors()[1];
-  if (_need_im2col)
-  {
-    im2col_data = im2col->data<int8_t>();
-  }
+  auto scratchpad = getOutputTensors()[1];
+  int8_t *scratchpad_data = nullptr;
+  if (scratchpad->is_allocatable())
+    scratchpad_data = scratchpad->data<int8_t>();
 
   luci_interpreter_pal::ConvPerChannel(
     params, multipliers.data(), shifts.data(), getTensorShape(input()),
     getTensorData<int8_t>(input()), getTensorShape(filter()), getTensorData<int8_t>(filter()),
     getTensorShape(bias()), getTensorData<int32_t>(bias()), getTensorShape(output()),
-    getTensorData<int8_t>(output()), getTensorShape(im2col), im2col_data);
+    getTensorData<int8_t>(output()), getTensorShape(scratchpad), scratchpad_data);
 }
 
 void Conv2D::evalQuantizedS16() const
diff --git a/compiler/luci-interpreter/src/kernels/Conv2D.h b/compiler/luci-interpreter/src/kernels/Conv2D.h
index 5f1317638..330bf3a2a 100644
--- a/compiler/luci-interpreter/src/kernels/Conv2D.h
+++ b/compiler/luci-interpreter/src/kernels/Conv2D.h
@@ -31,7 +31,7 @@ class Conv2D : public KernelWithParams<Conv2DParams>
 {
 public:
   Conv2D(const Tensor *input, const Tensor *filter, const Tensor *bias, Tensor *output,
-         Tensor *im2col, const Conv2DParams &params);
+         Tensor *scratchpad, const Conv2DParams &params);
 
   const Tensor *input() const { return _inputs[0]; }
   const Tensor *filter() const { return _inputs[1]; }
@@ -49,7 +49,6 @@ private:
   void evalQuantizedS16() const;
 
 private:
-  bool _need_im2col = false;
   int32_t _padding_height{};
   int32_t _padding_width{};
 };
diff --git a/compiler/luci-interpreter/src/kernels/DepthToSpace.test.cpp b/compiler/luci-interpreter/src/kernels/DepthToSpace.test.cpp
index 9b1c09ba9..88e6e07f1 100644
--- a/compiler/luci-interpreter/src/kernels/DepthToSpace.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/DepthToSpace.test.cpp
@@ -32,7 +32,7 @@ template <typename T> class DepthToSpaceTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(DepthToSpaceTest, DataTypes);
+TYPED_TEST_SUITE(DepthToSpaceTest, DataTypes);
 
 TYPED_TEST(DepthToSpaceTest, SimpleCase)
 {
diff --git a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.cpp b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.cpp
index f2dbf6c68..c554c309d 100644
--- a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.cpp
+++ b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.cpp
@@ -18,9 +18,7 @@
 
 #include "kernels/Utils.h"
 
-#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h>
-#include <tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h>
-#include <tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h>
+#include "PALDepthwiseConv2d.h"
 
 #include <stdexcept>
 
@@ -30,8 +28,9 @@ namespace kernels
 {
 
 DepthwiseConv2D::DepthwiseConv2D(const Tensor *input, const Tensor *filter, const Tensor *bias,
-                                 Tensor *output, const DepthwiseConv2DParams &params)
-  : KernelWithParams<DepthwiseConv2DParams>({input, filter, bias}, {output}, params)
+                                 Tensor *output, Tensor *scratchpad,
+                                 const DepthwiseConv2DParams &params)
+  : KernelWithParams<DepthwiseConv2DParams>({input, filter, bias}, {output, scratchpad}, params)
 {
 }
 
@@ -109,6 +108,16 @@ void DepthwiseConv2D::configure()
                                   filter_width, output_width);
 
   output()->resize({batches, output_height, output_width, channels_out});
+
+  tflite::DepthwiseParams params{};
+
+  params.dilation_height_factor = _params.dilation_height_factor;
+  params.dilation_width_factor = _params.dilation_width_factor;
+
+  auto scratchpad = getOutputTensors()[1];
+  luci_interpreter_pal::SetupScratchpadTensor(scratchpad, params, input()->element_type(),
+                                              getTensorShape(input()), getTensorShape(filter()),
+                                              getTensorShape(output()));
 }
 
 void DepthwiseConv2D::execute() const
@@ -337,11 +346,16 @@ void DepthwiseConv2D::evalQuantizedS8PerChannel() const
                  std::back_inserter(multipliers),
                  [](ChannelQuantMultipliers cm) { return cm.multiplier; });
 
-  tflite::reference_integer_ops::DepthwiseConvPerChannel(
+  auto scratchpad = getOutputTensors()[1];
+  int8_t *scratchpad_data = nullptr;
+  if (scratchpad->is_allocatable())
+    scratchpad_data = scratchpad->data<int8_t>();
+
+  luci_interpreter_pal::DepthwiseConvPerChannel<int8_t>(
     params, multipliers.data(), shifts.data(), getTensorShape(input()),
     getTensorData<int8_t>(input()), getTensorShape(filter()), getTensorData<int8_t>(filter()),
     getTensorShape(bias()), getTensorData<int32_t>(bias()), getTensorShape(output()),
-    getTensorData<int8_t>(output()));
+    getTensorData<int8_t>(output()), getTensorShape(scratchpad), scratchpad_data);
 }
 
 void DepthwiseConv2D::evalQuantizedS16() const
diff --git a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.h b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.h
index 6cffd6583..3d1faf6c1 100644
--- a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.h
+++ b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.h
@@ -29,7 +29,7 @@ class DepthwiseConv2D : public KernelWithParams<DepthwiseConv2DParams>
 {
 public:
   DepthwiseConv2D(const Tensor *input, const Tensor *filter, const Tensor *bias, Tensor *output,
-                  const DepthwiseConv2DParams &params);
+                  Tensor *scratchpad, const DepthwiseConv2DParams &params);
 
   const Tensor *input() const { return _inputs[0]; }
   const Tensor *filter() const { return _inputs[1]; }
diff --git a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.test.cpp b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.test.cpp
index 74975899a..6b4673f3e 100644
--- a/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/DepthwiseConv2D.test.cpp
@@ -59,6 +59,7 @@ TEST_F(DepthwiseConv2DTest, Float)
     makeInputTensor<DataType::FLOAT32>(filter_shape, filter_data, _memory_manager.get());
   Tensor bias_tensor =
     makeInputTensor<DataType::FLOAT32>(bias_shape, bias_data, _memory_manager.get());
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
 
   DepthwiseConv2DParams params{};
@@ -70,8 +71,10 @@ TEST_F(DepthwiseConv2DTest, Float)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
+  _memory_manager->allocate_memory(scratchpad);
   _memory_manager->allocate_memory(output_tensor);
   kernel.execute();
 
@@ -111,6 +114,7 @@ TEST_F(DepthwiseConv2DTest, Uint8)
     {4}, input_quant_param.first * input_quant_param.first, 0, bias_data, _memory_manager.get());
   Tensor output_tensor =
     makeOutputTensor(DataType::U8, output_quant_param.first, output_quant_param.second);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -121,9 +125,11 @@ TEST_F(DepthwiseConv2DTest, Uint8)
   params.dilation_width_factor = 1;
   params.activation = Activation::NONE;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   std::vector<float> ref_output_data{
@@ -166,6 +172,7 @@ TEST_F(DepthwiseConv2DTest, SInt16)
   Tensor bias_tensor =
     makeInputTensor<DataType::S64>(bias_shape, 0.25 * 0.2, 0, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::S16, 0.5, 0);
+  Tensor scratchpad(DataType::S64, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -176,9 +183,11 @@ TEST_F(DepthwiseConv2DTest, SInt16)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
@@ -224,6 +233,7 @@ TEST_F(DepthwiseConv2DTest, SInt16_CWQ_weights)
   Tensor bias_tensor = makeInputTensor<DataType::S64>(bias_shape, bias_scales, zerop, 0, bias_data,
                                                       _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::S16, 0.5, 0);
+  Tensor scratchpad(DataType::S16, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -234,9 +244,11 @@ TEST_F(DepthwiseConv2DTest, SInt16_CWQ_weights)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
@@ -299,6 +311,7 @@ TEST_F(DepthwiseConv2DTest, Uint8_CWQ_weights)
                                                       _memory_manager.get());
   Tensor output_tensor =
     makeOutputTensor(DataType::U8, output_quant_param.first, output_quant_param.second);
+  Tensor scratchpad(DataType::U8, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -309,9 +322,11 @@ TEST_F(DepthwiseConv2DTest, Uint8_CWQ_weights)
   params.dilation_width_factor = 1;
   params.activation = Activation::NONE;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
@@ -375,6 +390,7 @@ TEST_F(DepthwiseConv2DTest, SInt8_CWQ_weights)
                                                       _memory_manager.get());
   Tensor output_tensor =
     makeOutputTensor(DataType::S8, output_quant_param.first, output_quant_param.second);
+  Tensor scratchpad(DataType::S8, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -385,9 +401,11 @@ TEST_F(DepthwiseConv2DTest, SInt8_CWQ_weights)
   params.dilation_width_factor = 1;
   params.activation = Activation::NONE;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   kernel.configure();
   _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad);
   kernel.execute();
 
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
@@ -419,6 +437,7 @@ TEST_F(DepthwiseConv2DTest, InvalidBiasType_NEG)
     makeInputTensor<DataType::FLOAT32>(filter_shape, filter_data, _memory_manager.get());
   Tensor bias_tensor = makeInputTensor<DataType::S32>(bias_shape, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -429,7 +448,8 @@ TEST_F(DepthwiseConv2DTest, InvalidBiasType_NEG)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -458,6 +478,7 @@ TEST_F(DepthwiseConv2DTest, InOutTypeMismatch_NEG)
   Tensor bias_tensor =
     makeInputTensor<DataType::FLOAT32>(bias_shape, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::U8);
+  Tensor scratchpad(DataType::U8, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -468,7 +489,8 @@ TEST_F(DepthwiseConv2DTest, InOutTypeMismatch_NEG)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -497,6 +519,7 @@ TEST_F(DepthwiseConv2DTest, InvalidInputShape_NEG)
   Tensor bias_tensor =
     makeInputTensor<DataType::FLOAT32>(bias_shape, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -507,7 +530,8 @@ TEST_F(DepthwiseConv2DTest, InvalidInputShape_NEG)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -536,6 +560,7 @@ TEST_F(DepthwiseConv2DTest, InvalidFilterShape_NEG)
   Tensor bias_tensor =
     makeInputTensor<DataType::FLOAT32>(bias_shape, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -546,7 +571,8 @@ TEST_F(DepthwiseConv2DTest, InvalidFilterShape_NEG)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
@@ -575,6 +601,7 @@ TEST_F(DepthwiseConv2DTest, InvalidBiasDim_NEG)
   Tensor bias_tensor =
     makeInputTensor<DataType::FLOAT32>(bias_shape, bias_data, _memory_manager.get());
   Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  Tensor scratchpad(DataType::FLOAT32, Shape({}), {}, "");
 
   DepthwiseConv2DParams params{};
   params.padding = Padding::VALID;
@@ -585,7 +612,8 @@ TEST_F(DepthwiseConv2DTest, InvalidBiasDim_NEG)
   params.dilation_width_factor = 1;
   params.activation = Activation::RELU;
 
-  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, params);
+  DepthwiseConv2D kernel(&input_tensor, &filter_tensor, &bias_tensor, &output_tensor, &scratchpad,
+                         params);
   EXPECT_ANY_THROW(kernel.configure());
 }
 
diff --git a/compiler/luci-interpreter/src/kernels/Dequantize.cpp b/compiler/luci-interpreter/src/kernels/Dequantize.cpp
new file mode 100644
index 000000000..96399e5c7
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Dequantize.cpp
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Dequantize.h"
+#include "kernels/Utils.h"
+#include "PALDequantize.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+Dequantize::Dequantize(const Tensor *input, Tensor *output) : Kernel({input}, {output}) {}
+
+void Dequantize::configure()
+{
+  LUCI_INTERPRETER_CHECK(input()->element_type() == loco::DataType::S8 ||
+                         input()->element_type() == loco::DataType::U8 ||
+                         input()->element_type() == loco::DataType::S16);
+
+  LUCI_INTERPRETER_CHECK(input()->scales().size() == 1);
+
+  if (input()->element_type() == loco::DataType::S16)
+    LUCI_INTERPRETER_CHECK(input()->zero_point() == 0);
+
+  LUCI_INTERPRETER_CHECK(output()->element_type() == loco::DataType::FLOAT32);
+
+  output()->resize(input()->shape());
+}
+
+void Dequantize::execute() const
+{
+  tflite::DequantizationParams op_params;
+  op_params.zero_point = input()->zero_point();
+  op_params.scale = input()->scale();
+
+  switch (input()->element_type())
+  {
+    case loco::DataType::U8:
+    {
+      luci_interpreter_pal::Dequantize(op_params, getTensorShape(input()),
+                                       getTensorData<uint8_t>(input()), getTensorShape(output()),
+                                       getTensorData<float>(output()));
+      break;
+    }
+    case loco::DataType::S8:
+    {
+      luci_interpreter_pal::Dequantize(op_params, getTensorShape(input()),
+                                       getTensorData<int8_t>(input()), getTensorShape(output()),
+                                       getTensorData<float>(output()));
+      break;
+    }
+    case loco::DataType::S16:
+    {
+      luci_interpreter_pal::Dequantize(op_params, getTensorShape(input()),
+                                       getTensorData<int16_t>(input()), getTensorShape(output()),
+                                       getTensorData<float>(output()));
+      break;
+    }
+    default:
+      throw std::runtime_error("Unsupported type.");
+  }
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Dequantize.h b/compiler/luci-interpreter/src/kernels/Dequantize.h
new file mode 100644
index 000000000..5565df0e4
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Dequantize.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_DEQUANTIZE_H
+#define LUCI_INTERPRETER_KERNELS_DEQUANTIZE_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class Dequantize : public Kernel
+{
+public:
+  Dequantize(const Tensor *input, Tensor *output);
+
+  const Tensor *input() const { return _inputs[0]; }
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_DEQUANTIZE_H
diff --git a/compiler/luci-interpreter/src/kernels/Dequantize.test.cpp b/compiler/luci-interpreter/src/kernels/Dequantize.test.cpp
new file mode 100644
index 000000000..0cab633d6
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Dequantize.test.cpp
@@ -0,0 +1,149 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Dequantize.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class DequantizeTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(DequantizeTest, Uint8)
+{
+  std::vector<uint8_t> input_data{0, 1, 2, 3, 4, 251, 252, 253, 254, 255};
+
+  std::vector<float> ref_output_data{-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64};
+
+  Tensor input_tensor(loco::DataType::U8, {2, 5}, {{0.5}, {127}}, "");
+
+  _memory_manager->allocate_memory(input_tensor);
+  input_tensor.writeData(input_data.data(), input_data.size() * sizeof(uint8_t));
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(DequantizeTest, Sint8)
+{
+  std::vector<int8_t> input_data{-128, -127, -126, -125, -124, 123, 124, 125, 126, 127};
+
+  std::vector<float> ref_output_data{-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64};
+
+  Tensor input_tensor(loco::DataType::S8, {2, 5}, {{0.5}, {-1}}, "");
+
+  _memory_manager->allocate_memory(input_tensor);
+  input_tensor.writeData(input_data.data(), input_data.size() * sizeof(int8_t));
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(DequantizeTest, Sint16)
+{
+  std::vector<int16_t> input_data{-129, -126, -125, -124, -123, 124, 125, 126, 127, 131};
+
+  std::vector<float> ref_output_data{-64.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 65.5};
+
+  Tensor input_tensor(loco::DataType::S16, {2, 5}, {{0.5}, {0}}, "");
+
+  _memory_manager->allocate_memory(input_tensor);
+  input_tensor.writeData(input_data.data(), input_data.size() * sizeof(int16_t));
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(DequantizeTest, InvalidInputType_NEG)
+{
+  std::vector<float> input_data{-129, -126, -125, -124, -123, 124, 125, 126, 127, 131};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 5}, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(DequantizeTest, InvalidOutputType_NEG)
+{
+  std::vector<int16_t> input_data{-129, -126, -125, -124, -123, 124, 125, 126, 127, 131};
+
+  Tensor input_tensor(loco::DataType::S16, {2, 5}, {{0.5}, {0}}, "");
+
+  _memory_manager->allocate_memory(input_tensor);
+  input_tensor.writeData(input_data.data(), input_data.size() * sizeof(int16_t));
+
+  Tensor output_tensor = makeOutputTensor(DataType::S8, /*scale*/ 0.5, /*zero_point*/ -1);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(DequantizeTest, InvalidInputZeroPoint_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S16>({2, 5}, 0.5, -1, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Dequantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Div.cpp b/compiler/luci-interpreter/src/kernels/Div.cpp
index 0e52ba1f0..dd1532278 100644
--- a/compiler/luci-interpreter/src/kernels/Div.cpp
+++ b/compiler/luci-interpreter/src/kernels/Div.cpp
@@ -46,6 +46,12 @@ void Div::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -56,13 +62,9 @@ void Div::execute() const
 
 void Div::evalFloat() const
 {
-  float activation_min{};
-  float activation_max{};
-  calculateActivationRange(_params.activation, &activation_min, &activation_max);
-
   tflite::ArithmeticParams params{};
-  params.float_activation_min = activation_min;
-  params.float_activation_max = activation_max;
+  fillArithmeticActivationRange<float>(params, _params.activation);
+
   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
     getTensorShape(input1()), getTensorShape(input2()), &params);
 
@@ -80,6 +82,28 @@ void Div::evalFloat() const
   }
 }
 
+template <typename T> void Div::evalInteger() const
+{
+  tflite::ArithmeticParams params{};
+  fillArithmeticActivationRange<T>(params, _params.activation);
+
+  const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
+    getTensorShape(input1()), getTensorShape(input2()), &params);
+
+  if (need_broadcast)
+  {
+    tflite::reference_ops::BroadcastDivSlow(
+      params, getTensorShape(input1()), getTensorData<T>(input1()), getTensorShape(input2()),
+      getTensorData<T>(input2()), getTensorShape(output()), getTensorData<T>(output()));
+  }
+  else
+  {
+    tflite::reference_ops::Div(params, getTensorShape(input1()), getTensorData<T>(input1()),
+                               getTensorShape(input2()), getTensorData<T>(input2()),
+                               getTensorShape(output()), getTensorData<T>(output()));
+  }
+}
+
 void Div::evalQuantized() const
 {
   const auto input1_scale = static_cast<double>(input1()->scale());
diff --git a/compiler/luci-interpreter/src/kernels/Div.h b/compiler/luci-interpreter/src/kernels/Div.h
index 6040cdd02..c1bf3e10b 100644
--- a/compiler/luci-interpreter/src/kernels/Div.h
+++ b/compiler/luci-interpreter/src/kernels/Div.h
@@ -39,6 +39,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 };
 
diff --git a/compiler/luci-interpreter/src/kernels/Div.test.cpp b/compiler/luci-interpreter/src/kernels/Div.test.cpp
index 021d68d06..85cd8b90a 100644
--- a/compiler/luci-interpreter/src/kernels/Div.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Div.test.cpp
@@ -134,6 +134,56 @@ TEST_F(DivTest, Uint8)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(output_shape));
 }
 
+template <loco::DataType DType> void checkInteger(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  Shape base_shape = {2, 3, 1, 2};
+  std::vector<Shape> test_shapes{{1, 1, 3, 2}, {1, 3, 1, 2}, {2, 1, 3, 1}, {2, 3, 1, 1}};
+
+  std::vector<std::vector<dtype>> test_outputs = {{5,  6,  2, 0,  10, 3, //
+                                                   10, 0,  4, 5,  20, 0, //
+                                                   0,  0,  0, 2,  0,  0, //
+                                                   2,  0,  1, 10, 5,  0, //
+                                                   2,  3,  1, 0,  5,  1, //
+                                                   18, 20, 7, 0,  37, 10},
+                                                  {5, 6, 4, 5, 0, 0, 2, 0, 1, 0, 37, 10},
+                                                  {5, 7, 4, 6, 2, 3, 10, 0,  8,  0,  4, 0,
+                                                   0, 0, 0, 0, 0, 0, 0,  10, 5,  0,  1, 0,
+                                                   0, 0, 5, 9, 1, 1, 0,  0,  37, 50, 7, 10},
+                                                  {5, 7, 8, 0, 0, 0, 0, 10, 5, 9, 7, 10}};
+  std::vector<dtype> input1_data{20, 30, 40, -17, -4, -7, 11, -31, 10, 19, 75, 100};
+  std::vector<dtype> input2_data{4, 5, 10, -3, 2, 10};
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    DivParams params{};
+    params.activation = Activation::RELU;
+
+    Div kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+}
+
+TEST_F(DivTest, SInt64)
+{
+  checkInteger<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(DivTest, SInt32)
+{
+  checkInteger<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
 TEST_F(DivTest, Input_Output_Type_NEG)
 {
   Tensor input1_tensor = makeInputTensor<DataType::FLOAT32>({1}, {1.f}, _memory_manager.get());
@@ -149,9 +199,9 @@ TEST_F(DivTest, Input_Output_Type_NEG)
 
 TEST_F(DivTest, Invalid_Input_Type_NEG)
 {
-  Tensor input1_tensor = makeInputTensor<DataType::S64>({1}, {1}, _memory_manager.get());
-  Tensor input2_tensor = makeInputTensor<DataType::S64>({1}, {2}, _memory_manager.get());
-  Tensor output_tensor = makeOutputTensor(DataType::S64);
+  Tensor input1_tensor = makeInputTensor<DataType::U64>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::U64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U64);
 
   DivParams params{};
   params.activation = Activation::RELU;
@@ -162,6 +212,19 @@ TEST_F(DivTest, Invalid_Input_Type_NEG)
   EXPECT_ANY_THROW(kernel.execute());
 }
 
+TEST_F(DivTest, Invalid_Output_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::S32>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S32>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S64);
+
+  DivParams params{};
+  params.activation = Activation::RELU;
+
+  Div kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Equal.cpp b/compiler/luci-interpreter/src/kernels/Equal.cpp
index f58de1250..a57e127b7 100644
--- a/compiler/luci-interpreter/src/kernels/Equal.cpp
+++ b/compiler/luci-interpreter/src/kernels/Equal.cpp
@@ -49,6 +49,12 @@ void Equal::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -79,6 +85,29 @@ void Equal::evalFloat() const
   }
 }
 
+template <typename T> void Equal::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                                         getTensorShape(y()), y_data,
+                                                         getTensorShape(output()), output_data);
+  }
+  else
+  {
+    tflite::reference_ops::EqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                          getTensorShape(y()), y_data, getTensorShape(output()),
+                                          output_data);
+  }
+}
+
 void Equal::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/Equal.h b/compiler/luci-interpreter/src/kernels/Equal.h
index 11f025eac..c9be32cc0 100644
--- a/compiler/luci-interpreter/src/kernels/Equal.h
+++ b/compiler/luci-interpreter/src/kernels/Equal.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/Equal.test.cpp b/compiler/luci-interpreter/src/kernels/Equal.test.cpp
index 46a0f97d8..5870e5460 100644
--- a/compiler/luci-interpreter/src/kernels/Equal.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Equal.test.cpp
@@ -99,6 +99,82 @@ TEST_F(EqualTest, FloatBroardcast)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
 }
 
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value, -2, max_value};
+
+  std::vector<bool> ref_output_data{true, false, true};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Equal kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,         // Row 1
+    4,         5,  max_value, // Row 2
+    -1,        -2, -3,        // Row 3
+    min_value, -2, max_value, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value, -2, max_value, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    true,  false, false, // Row 1
+    false, false, true,  // Row 2
+    false, true,  false, // Row 3
+    true,  true,  true,  // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Equal kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(EqualTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(EqualTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
 const float F_MAX = 127.0 / 128.0;
@@ -195,6 +271,36 @@ TEST_F(EqualTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(EqualTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Equal kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(EqualTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Equal kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(EqualTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Equal kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/ExpandDims.cpp b/compiler/luci-interpreter/src/kernels/ExpandDims.cpp
new file mode 100644
index 000000000..ba35c99fa
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/ExpandDims.cpp
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/ExpandDims.h"
+#include "kernels/Utils.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+ExpandDims::ExpandDims(const Tensor *input, const Tensor *axis, Tensor *output)
+  : Kernel({input, axis}, {output})
+{
+}
+
+void ExpandDims::configure()
+{
+  int32_t axis_value;
+
+  switch (axis()->element_type())
+  {
+    case loco::DataType::S32:
+      axis_value = *getTensorData<int32_t>(axis());
+      break;
+    case loco::DataType::S64:
+      axis_value = static_cast<int32_t>(*getTensorData<int64_t>(axis()));
+      break;
+    default:
+      throw std::runtime_error("Unsupported type.");
+  }
+
+  const auto input_shape = input()->shape();
+
+  if (axis_value < 0)
+  {
+    axis_value += input_shape.num_dims() + 1;
+  }
+
+  LUCI_INTERPRETER_CHECK(axis_value <= input_shape.num_dims() and axis_value >= 0);
+
+  Shape output_shape(input_shape.num_dims() + 1);
+  for (int32_t i = 0; i < output_shape.num_dims(); ++i)
+  {
+    if (i < axis_value)
+    {
+      output_shape.dim(i) = input_shape.dim(i);
+    }
+    else if (i == axis_value)
+    {
+      output_shape.dim(i) = 1;
+    }
+    else
+    {
+      LUCI_INTERPRETER_CHECK(i >= 1);
+      output_shape.dim(i) = input_shape.dim(i - 1);
+    }
+  }
+
+  output()->resize(output_shape);
+}
+
+void ExpandDims::execute() const
+{
+  // Just copy input to output
+  const auto *input_data = input()->data<void>();
+  auto *output_data = output()->data<void>();
+
+  const size_t element_size = getDataTypeSize(input()->element_type());
+  const int32_t num_elements = input()->shape().num_elements();
+  std::memcpy(output_data, input_data, num_elements * element_size);
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/ExpandDims.h b/compiler/luci-interpreter/src/kernels/ExpandDims.h
new file mode 100644
index 000000000..e510b1160
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/ExpandDims.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_EXPAND_DIMS_H
+#define LUCI_INTERPRETER_KERNELS_EXPAND_DIMS_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class ExpandDims : public Kernel
+{
+public:
+  ExpandDims(const Tensor *input, const Tensor *axis, Tensor *output);
+
+  const Tensor *input() const { return _inputs[0]; }
+  const Tensor *axis() const { return _inputs[1]; }
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_EXPAND_DIMS_H
diff --git a/compiler/luci-interpreter/src/kernels/ExpandDims.test.cpp b/compiler/luci-interpreter/src/kernels/ExpandDims.test.cpp
new file mode 100644
index 000000000..df9eaccc0
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/ExpandDims.test.cpp
@@ -0,0 +1,115 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/ExpandDims.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class ExpandDimsTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(ExpandDimsTest, PositiveAxis)
+{
+  std::vector<int32_t> input_data{-1, 1, -2, 2};
+  std::initializer_list<int32_t> input_shape = {2, 2};
+
+  std::initializer_list<int32_t> axis_value = {0};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>(input_shape, input_data, _memory_manager.get());
+  Tensor axis_tensor = makeInputTensor<DataType::S32>({1}, axis_value, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  ExpandDims kernel(&input_tensor, &axis_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int32_t>(output_tensor), ::testing::ElementsAreArray(input_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 2, 2}));
+}
+
+TEST_F(ExpandDimsTest, NegAxis)
+{
+  std::vector<int32_t> input_data{-1, 1, -2, 2};
+  std::initializer_list<int32_t> input_shape = {2, 2};
+
+  std::initializer_list<int32_t> axis_value = {-1};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>(input_shape, input_data, _memory_manager.get());
+  Tensor axis_tensor = makeInputTensor<DataType::S32>({1}, axis_value, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  ExpandDims kernel(&input_tensor, &axis_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int32_t>(output_tensor), ::testing::ElementsAreArray(input_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 2, 1}));
+}
+
+TEST_F(ExpandDimsTest, InvalidAxisType_NEG)
+{
+  std::vector<int32_t> input_data{-1, 1, -2, 2};
+  std::initializer_list<int32_t> input_shape = {2, 2};
+
+  std::initializer_list<float> axis_value = {1.0};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>(input_shape, input_data, _memory_manager.get());
+  Tensor axis_tensor = makeInputTensor<DataType::FLOAT32>({1}, axis_value, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  ExpandDims kernel(&input_tensor, &axis_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(ExpandDimsTest, InvalidAxisValue_NEG)
+{
+  std::vector<int32_t> input_data{-1, 1, -2, 2};
+  std::initializer_list<int32_t> input_shape = {2, 2};
+
+  std::initializer_list<int32_t> axis_value = {3};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>(input_shape, input_data, _memory_manager.get());
+  Tensor axis_tensor = makeInputTensor<DataType::S32>({1}, axis_value, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  ExpandDims kernel(&input_tensor, &axis_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/FullyConnected.cpp b/compiler/luci-interpreter/src/kernels/FullyConnected.cpp
index cfe8f8bf2..bd2bb2f35 100644
--- a/compiler/luci-interpreter/src/kernels/FullyConnected.cpp
+++ b/compiler/luci-interpreter/src/kernels/FullyConnected.cpp
@@ -18,8 +18,7 @@
 
 #include "kernels/Utils.h"
 
-#include <tensorflow/lite/kernels/internal/reference/fully_connected.h>
-#include <tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h>
+#include "PALFullyConnected.h"
 
 #include <stdexcept>
 
@@ -74,7 +73,18 @@ void FullyConnected::configure()
   if (bias())
     LUCI_INTERPRETER_CHECK(bias()->shape().num_elements() == weights()->shape().dim(0));
 
-  output()->resize({batch_size, num_units});
+  if (params().keep_num_dims == false)
+  {
+    output()->resize({batch_size, num_units});
+  }
+  else
+  {
+    luci_interpreter::Shape output_shape(input_shape.num_dims());
+    for (int i = 0; i < input_shape.num_dims(); ++i)
+      output_shape.dim(i) = input_shape.dim(i);
+    output_shape.dim(input_shape.num_dims() - 1) = num_units;
+    output()->resize(output_shape);
+  }
 }
 
 void FullyConnected::execute() const
@@ -172,7 +182,7 @@ void FullyConnected::evalQuantizedS8() const
   op_params.quantized_activation_max = output_activation_max;
   op_params.lhs_cacheable = false;
   op_params.rhs_cacheable = false;
-  tflite::reference_integer_ops::FullyConnected(
+  luci_interpreter_pal::FullyConnected<int8_t>(
     op_params, getTensorShape(input()), getTensorData<int8_t>(input()), getTensorShape(weights()),
     getTensorData<int8_t>(weights()), getTensorShape(bias()), getTensorData<int32_t>(bias()),
     getTensorShape(output()), getTensorData<int8_t>(output()));
diff --git a/compiler/luci-interpreter/src/kernels/FullyConnected.test.cpp b/compiler/luci-interpreter/src/kernels/FullyConnected.test.cpp
index b0eda0145..4474cc4fb 100644
--- a/compiler/luci-interpreter/src/kernels/FullyConnected.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/FullyConnected.test.cpp
@@ -133,7 +133,7 @@ template <typename T> class FullyConnectedTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t, int8_t>;
-TYPED_TEST_CASE(FullyConnectedTest, DataTypes);
+TYPED_TEST_SUITE(FullyConnectedTest, DataTypes);
 
 TYPED_TEST(FullyConnectedTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Gather.cpp b/compiler/luci-interpreter/src/kernels/Gather.cpp
new file mode 100644
index 000000000..f1256660f
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Gather.cpp
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Gather.h"
+#include "kernels/Utils.h"
+#include "PALGather.h"
+
+#include <stdexcept>
+#include <cassert>
+
+namespace luci_interpreter
+{
+
+namespace kernels
+{
+
+Gather::Gather(const Tensor *params, const Tensor *indices, Tensor *output,
+               const GatherParams &gparams)
+  : KernelWithParams<GatherParams>({params, indices}, {output}, gparams)
+{
+}
+
+void Gather::configure()
+{
+  if (params()->element_type() == DataType::FLOAT32)
+  {
+    LUCI_INTERPRETER_CHECK(output()->element_type() == DataType::FLOAT32);
+  }
+  else
+  {
+    throw std::runtime_error("Unsupported type.");
+  }
+
+  LUCI_INTERPRETER_CHECK(indices()->element_type() == DataType::S32 ||
+                         indices()->element_type() == DataType::S64);
+
+  // refer tensorflow/lite/kernels/gather.cc
+
+  const Shape &params_shape = params()->shape();
+  const Shape &indices_shape = indices()->shape();
+
+  int axis = _params.axis;
+  if (axis < 0)
+  {
+    axis += params_shape.num_dims();
+  }
+  LUCI_INTERPRETER_CHECK(0 <= axis && axis < params_shape.num_dims());
+
+  int batch_dims = _params.batch_dims;
+  // batch_dims should be in range: [-rank(indices), rank(indices)].
+  // Negative batch_dims is added with rank of positions.
+  if (batch_dims < 0)
+  {
+    batch_dims += indices_shape.num_dims();
+  }
+  LUCI_INTERPRETER_CHECK(batch_dims <= axis);
+  LUCI_INTERPRETER_CHECK(0 <= batch_dims && batch_dims < params_shape.num_dims());
+  LUCI_INTERPRETER_CHECK(batch_dims <= indices_shape.num_dims());
+  for (int i = 0; i < batch_dims; ++i)
+  {
+    LUCI_INTERPRETER_CHECK(params_shape.dim(i) == indices_shape.dim(i));
+  }
+
+  const int num_dimensions = params_shape.num_dims() + indices_shape.num_dims() - 1 - batch_dims;
+
+  Shape output_shape(num_dimensions);
+  int output_index = 0;
+  for (int i = 0; i < axis; ++i)
+  {
+    output_shape.dim(output_index++) = params_shape.dim(i);
+  }
+  for (int i = batch_dims; i < indices_shape.num_dims(); ++i)
+  {
+    output_shape.dim(output_index++) = indices_shape.dim(i);
+  }
+  for (int i = axis + 1; i < params_shape.num_dims(); ++i)
+  {
+    output_shape.dim(output_index++) = params_shape.dim(i);
+  }
+  output()->resize(output_shape);
+}
+
+void Gather::execute() const
+{
+  switch (params()->element_type())
+  {
+    case DataType::FLOAT32:
+      evalFloat();
+      break;
+    default:
+      throw std::runtime_error("Unsupported type.");
+  }
+}
+
+void Gather::evalFloat() const
+{
+  assert(indices()->element_type() == DataType::S32 || indices()->element_type() == DataType::S64);
+
+  const auto params_data = getTensorData<float>(params());
+  auto output_data = getTensorData<float>(output());
+
+  tflite::GatherParams tparams;
+  tparams.axis = _params.axis;
+  tparams.batch_dims = _params.batch_dims;
+
+  if (indices()->element_type() == DataType::S32)
+  {
+    const auto indices_data = getTensorData<int32_t>(indices());
+
+    luci_interpreter_pal::Gather<float, int32_t>(tparams, getTensorShape(params()), params_data,
+                                                 getTensorShape(indices()), indices_data,
+                                                 getTensorShape(output()), output_data);
+  }
+  else
+  {
+    const auto indices_data = getTensorData<int64_t>(indices());
+
+    luci_interpreter_pal::Gather<float, int64_t>(tparams, getTensorShape(params()), params_data,
+                                                 getTensorShape(indices()), indices_data,
+                                                 getTensorShape(output()), output_data);
+  }
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Gather.h b/compiler/luci-interpreter/src/kernels/Gather.h
new file mode 100644
index 000000000..cc02d64fb
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Gather.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_GATHER_H
+#define LUCI_INTERPRETER_KERNELS_GATHER_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class Gather : public KernelWithParams<GatherParams>
+{
+public:
+  Gather(const Tensor *params, const Tensor *indices, Tensor *output, const GatherParams &gparams);
+
+  const Tensor *params() const { return _inputs[0]; }
+  const Tensor *indices() const { return _inputs[1]; }
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+
+private:
+  void evalFloat() const;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_GATHER_H
diff --git a/compiler/luci-interpreter/src/kernels/Gather.test.cpp b/compiler/luci-interpreter/src/kernels/Gather.test.cpp
new file mode 100644
index 000000000..4b3dda708
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Gather.test.cpp
@@ -0,0 +1,137 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Gather.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class GatherTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(GatherTest, Simple)
+{
+  std::vector<float> params_data{1.f, 2.f, 3.f, 4.f, 5.f, 6.f};
+  std::vector<int32_t> indices_data{1, 0, 1, 5};
+  std::vector<float> ref_output_data{2.f, 1.f, 2.f, 6.f};
+
+  Tensor params_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 6}, params_data, _memory_manager.get());
+  Tensor indices_tensor = makeInputTensor<DataType::S32>({4}, indices_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  GatherParams gparams;
+
+  gparams.axis = 1;
+  gparams.batch_dims = 0;
+
+  Gather kernel(&params_tensor, &indices_tensor, &output_tensor, gparams);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 4}));
+}
+
+TEST_F(GatherTest, Simple_Batch)
+{
+  Shape params_shape = {3, 5};
+  Shape indices_shape = {3, 2};
+  std::vector<float> params_data{0., 0., 1., 0., 2., 3., 0., 0., 0., 4., 0., 5., 0., 6., 0.};
+  std::vector<int32_t> indices_data{2, 4, 0, 4, 1, 3};
+  std::vector<float> ref_output_data{1., 2., 3., 4., 5., 6.};
+
+  Tensor params_tensor =
+    makeInputTensor<DataType::FLOAT32>(params_shape, params_data, _memory_manager.get());
+  Tensor indices_tensor =
+    makeInputTensor<DataType::S32>(indices_shape, indices_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  GatherParams gparams;
+
+  gparams.axis = 1;
+  gparams.batch_dims = 1;
+
+  Gather kernel(&params_tensor, &indices_tensor, &output_tensor, gparams);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3, 2}));
+}
+
+TEST_F(GatherTest, Simple_NEG)
+{
+  Tensor params_tensor = makeInputTensor<DataType::S32>({1}, {1}, _memory_manager.get());
+  Tensor indices_tensor = makeInputTensor<DataType::S32>({1}, {0}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  GatherParams gparams;
+
+  Gather kernel(&params_tensor, &indices_tensor, &output_tensor, gparams);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GatherTest, Axis_NEG)
+{
+  Tensor params_tensor = makeInputTensor<DataType::FLOAT32>({1}, {1.f}, _memory_manager.get());
+  Tensor indices_tensor = makeInputTensor<DataType::S32>({1}, {0}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  GatherParams gparams;
+
+  gparams.axis = 100;
+  gparams.batch_dims = 0;
+
+  Gather kernel(&params_tensor, &indices_tensor, &output_tensor, gparams);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GatherTest, Batch_NEG)
+{
+  std::vector<float> params_data{1.f, 2.f, 3.f, 4.f, 5.f, 6.f};
+  std::vector<int32_t> indices_data{1, 0, 1, 5};
+  std::vector<float> ref_output_data{2.f, 1.f, 2.f, 6.f};
+
+  Tensor params_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 6}, params_data, _memory_manager.get());
+  Tensor indices_tensor = makeInputTensor<DataType::S32>({4}, indices_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+  GatherParams gparams;
+
+  gparams.axis = 0;
+  gparams.batch_dims = 1;
+
+  Gather kernel(&params_tensor, &indices_tensor, &output_tensor, gparams);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Greater.cpp b/compiler/luci-interpreter/src/kernels/Greater.cpp
index f0dd2db36..5ccae3c38 100644
--- a/compiler/luci-interpreter/src/kernels/Greater.cpp
+++ b/compiler/luci-interpreter/src/kernels/Greater.cpp
@@ -49,6 +49,12 @@ void Greater::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -79,6 +85,29 @@ void Greater::evalFloat() const
   }
 }
 
+template <typename T> void Greater::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowGreaterNoScaling(op_params, getTensorShape(x()), x_data,
+                                                           getTensorShape(y()), y_data,
+                                                           getTensorShape(output()), output_data);
+  }
+  else
+  {
+    tflite::reference_ops::GreaterNoScaling(op_params, getTensorShape(x()), x_data,
+                                            getTensorShape(y()), y_data, getTensorShape(output()),
+                                            output_data);
+  }
+}
+
 void Greater::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/Greater.h b/compiler/luci-interpreter/src/kernels/Greater.h
index 877c139c9..065f76d7b 100644
--- a/compiler/luci-interpreter/src/kernels/Greater.h
+++ b/compiler/luci-interpreter/src/kernels/Greater.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/Greater.test.cpp b/compiler/luci-interpreter/src/kernels/Greater.test.cpp
index ba3925f17..a48080124 100644
--- a/compiler/luci-interpreter/src/kernels/Greater.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Greater.test.cpp
@@ -97,6 +97,82 @@ TEST_F(GreaterTest, FloatBroardcast)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3, 3}));
 }
 
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value + 1, -2, max_value};
+
+  std::vector<bool> ref_output_data{false, true, false};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Greater kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,         // Row 1
+    4,         5,  max_value, // Row 2
+    -1,        -4, -3,        // Row 3
+    min_value, -2, max_value, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value + 1, -2, max_value - 1, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    false, true,  false, // Row 1
+    true,  true,  true,  // Row 2
+    true,  false, false, // Row 3
+    false, false, true,  // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Greater kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(GreaterTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(GreaterTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
 const float F_MAX = 127.0 / 128.0;
@@ -223,6 +299,36 @@ TEST_F(GreaterTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(GreaterTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Greater kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GreaterTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Greater kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GreaterTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Greater kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/GreaterEqual.cpp b/compiler/luci-interpreter/src/kernels/GreaterEqual.cpp
index e7c1b4afe..27e42c971 100644
--- a/compiler/luci-interpreter/src/kernels/GreaterEqual.cpp
+++ b/compiler/luci-interpreter/src/kernels/GreaterEqual.cpp
@@ -52,6 +52,12 @@ void GreaterEqual::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -82,6 +88,29 @@ void GreaterEqual::evalFloat() const
   }
 }
 
+template <typename T> void GreaterEqual::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowGreaterEqualNoScaling(
+      op_params, getTensorShape(x()), x_data, getTensorShape(y()), y_data, getTensorShape(output()),
+      output_data);
+  }
+  else
+  {
+    tflite::reference_ops::GreaterEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                                 getTensorShape(y()), y_data,
+                                                 getTensorShape(output()), output_data);
+  }
+}
+
 void GreaterEqual::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/GreaterEqual.h b/compiler/luci-interpreter/src/kernels/GreaterEqual.h
index 4a0f48748..e333c30a6 100644
--- a/compiler/luci-interpreter/src/kernels/GreaterEqual.h
+++ b/compiler/luci-interpreter/src/kernels/GreaterEqual.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/GreaterEqual.test.cpp b/compiler/luci-interpreter/src/kernels/GreaterEqual.test.cpp
index a9d172301..35bf88eab 100644
--- a/compiler/luci-interpreter/src/kernels/GreaterEqual.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/GreaterEqual.test.cpp
@@ -96,6 +96,81 @@ TEST_F(GreaterEqualTest, FloatBroardcast)
   EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3, 3}));
 }
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value + 1, -2, max_value};
+
+  std::vector<bool> ref_output_data{false, true, true};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  GreaterEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,             // Row 1
+    4,         5,  max_value,     // Row 2
+    -1,        -4, -3,            // Row 3
+    min_value, -2, max_value - 1, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value + 1, -2, max_value - 1, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    false, true,  false, // Row 1
+    true,  true,  true,  // Row 2
+    true,  false, false, // Row 3
+    false, true,  true,  // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  GreaterEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(GreaterEqualTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(GreaterEqualTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
 
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
@@ -223,6 +298,36 @@ TEST_F(GreaterEqualTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(GreaterEqualTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  GreaterEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GreaterEqualTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  GreaterEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(GreaterEqualTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  GreaterEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/L2Normalize.test.cpp b/compiler/luci-interpreter/src/kernels/L2Normalize.test.cpp
index 1e565e358..6f960e8b4 100644
--- a/compiler/luci-interpreter/src/kernels/L2Normalize.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/L2Normalize.test.cpp
@@ -81,7 +81,7 @@ template <typename T> class L2NormalizeTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(L2NormalizeTest, DataTypes);
+TYPED_TEST_SUITE(L2NormalizeTest, DataTypes);
 
 TYPED_TEST(L2NormalizeTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/L2Pool2D.test.cpp b/compiler/luci-interpreter/src/kernels/L2Pool2D.test.cpp
index 289742a50..7245456cb 100644
--- a/compiler/luci-interpreter/src/kernels/L2Pool2D.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/L2Pool2D.test.cpp
@@ -206,7 +206,8 @@ TEST_F(L2Pool2DTest, FloatPaddingSameStride)
   kernel.execute();
 
   std::vector<float> ref_output_data{3.5, 6.0, 6.5, 5.70088, 2.54951, 7.2111, 8.63134, 7.0};
-  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data));
+  // NOTE with NEON+ruy, error is #1=-1.14441e-05, #6=-1.81198e-05
+  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data, 1.0e-4f));
   // TODO make a Shape checking of output_tensor.
 }
 
diff --git a/compiler/luci-interpreter/src/kernels/LeakyRelu.test.cpp b/compiler/luci-interpreter/src/kernels/LeakyRelu.test.cpp
index 6ec8a348a..0f6263b57 100644
--- a/compiler/luci-interpreter/src/kernels/LeakyRelu.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/LeakyRelu.test.cpp
@@ -83,7 +83,7 @@ template <typename T> class LeakReluTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(LeakReluTest, DataTypes);
+TYPED_TEST_SUITE(LeakReluTest, DataTypes);
 
 TYPED_TEST(LeakReluTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Less.cpp b/compiler/luci-interpreter/src/kernels/Less.cpp
index 041444926..8d26ff297 100644
--- a/compiler/luci-interpreter/src/kernels/Less.cpp
+++ b/compiler/luci-interpreter/src/kernels/Less.cpp
@@ -49,6 +49,12 @@ void Less::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -79,6 +85,29 @@ void Less::evalFloat() const
   }
 }
 
+template <typename T> void Less::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowLessNoScaling(op_params, getTensorShape(x()), x_data,
+                                                        getTensorShape(y()), y_data,
+                                                        getTensorShape(output()), output_data);
+  }
+  else
+  {
+    tflite::reference_ops::LessNoScaling(op_params, getTensorShape(x()), x_data,
+                                         getTensorShape(y()), y_data, getTensorShape(output()),
+                                         output_data);
+  }
+}
+
 void Less::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/Less.h b/compiler/luci-interpreter/src/kernels/Less.h
index 293740e72..e27bb689c 100644
--- a/compiler/luci-interpreter/src/kernels/Less.h
+++ b/compiler/luci-interpreter/src/kernels/Less.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/Less.test.cpp b/compiler/luci-interpreter/src/kernels/Less.test.cpp
index e9d09b288..8c5963363 100644
--- a/compiler/luci-interpreter/src/kernels/Less.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Less.test.cpp
@@ -97,6 +97,82 @@ TEST_F(LessTest, FloatBroardcast)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3, 3}));
 }
 
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value + 1, -2, max_value};
+
+  std::vector<bool> ref_output_data{true, false, false};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Less kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,         // Row 1
+    4,         5,  max_value, // Row 2
+    -1,        -4, -3,        // Row 3
+    min_value, -2, max_value, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value + 1, -2, max_value - 1, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    true,  false, true,  // Row 1
+    false, false, false, // Row 2
+    false, true,  true,  // Row 3
+    true,  false, false, // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Less kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(LessTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(LessTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
 const float F_MAX = 127.0 / 128.0;
@@ -223,6 +299,36 @@ TEST_F(LessTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(LessTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Less kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(LessTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Less kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(LessTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  Less kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/LessEqual.cpp b/compiler/luci-interpreter/src/kernels/LessEqual.cpp
index 5f4c7f7aa..b474bc47a 100644
--- a/compiler/luci-interpreter/src/kernels/LessEqual.cpp
+++ b/compiler/luci-interpreter/src/kernels/LessEqual.cpp
@@ -49,6 +49,12 @@ void LessEqual::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -79,6 +85,29 @@ void LessEqual::evalFloat() const
   }
 }
 
+template <typename T> void LessEqual::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowLessEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                                             getTensorShape(y()), y_data,
+                                                             getTensorShape(output()), output_data);
+  }
+  else
+  {
+    tflite::reference_ops::LessEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                              getTensorShape(y()), y_data, getTensorShape(output()),
+                                              output_data);
+  }
+}
+
 void LessEqual::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/LessEqual.h b/compiler/luci-interpreter/src/kernels/LessEqual.h
index b6da1a2a8..f82ea90d4 100644
--- a/compiler/luci-interpreter/src/kernels/LessEqual.h
+++ b/compiler/luci-interpreter/src/kernels/LessEqual.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/LessEqual.test.cpp b/compiler/luci-interpreter/src/kernels/LessEqual.test.cpp
index 0558003dd..b2e2fa7a1 100644
--- a/compiler/luci-interpreter/src/kernels/LessEqual.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/LessEqual.test.cpp
@@ -97,6 +97,82 @@ TEST_F(LessEqualTest, FloatBroardcast)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3, 3}));
 }
 
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value + 1, -2, max_value};
+
+  std::vector<bool> ref_output_data{true, false, true};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  LessEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,         // Row 1
+    4,         5,  max_value, // Row 2
+    -1,        -4, -3,        // Row 3
+    min_value, -2, max_value, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value + 1, -2, max_value - 1, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    true,  false, true,  // Row 1
+    false, false, false, // Row 2
+    false, true,  true,  // Row 3
+    true,  true,  false, // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  LessEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(LessEqualTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(LessEqualTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
 const float F_MAX = 127.0 / 128.0;
@@ -223,6 +299,36 @@ TEST_F(LessEqualTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(LessEqualTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  LessEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(LessEqualTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  LessEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(LessEqualTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  LessEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Logistic.test.cpp b/compiler/luci-interpreter/src/kernels/Logistic.test.cpp
index 70227563f..5a1ea669c 100644
--- a/compiler/luci-interpreter/src/kernels/Logistic.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Logistic.test.cpp
@@ -76,7 +76,7 @@ template <typename T> class LogisticTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(LogisticTest, DataTypes);
+TYPED_TEST_SUITE(LogisticTest, DataTypes);
 
 TYPED_TEST(LogisticTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/MirrorPad.cpp b/compiler/luci-interpreter/src/kernels/MirrorPad.cpp
index 89049c96c..2fbeefce4 100644
--- a/compiler/luci-interpreter/src/kernels/MirrorPad.cpp
+++ b/compiler/luci-interpreter/src/kernels/MirrorPad.cpp
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,8 +19,6 @@
 
 #include "kernels/Utils.h"
 
-#include <tensorflow/lite/kernels/internal/reference/pad.h>
-
 namespace luci_interpreter
 {
 namespace kernels
@@ -59,44 +58,25 @@ void MirrorPad::configure()
   output()->resize(output_shape);
 }
 
+template <typename T>
+inline void MirrorPadImpl(const Tensor &input, const Tensor &paddings, MirrorPadMode mode,
+                          Tensor &output);
+
 void MirrorPad::execute() const
 {
-  const int num_dims = input()->shape().num_dims();
-
-  tflite::PadParams params{};
-  params.left_padding_count = num_dims;
-  params.right_padding_count = num_dims;
-
-  const auto *paddings_data = getTensorData<int32_t>(paddings());
-  for (int i = num_dims - 1; i >= 0; --i)
-  {
-    params.left_padding[i] = paddings_data[i * 2];
-    params.right_padding[i] = paddings_data[i * 2 + 1];
-  }
-
   switch (input()->element_type())
   {
     case DataType::FLOAT32:
     {
-      const float pad_value = 0;
-
-      // NOTE: this implementation only obtains min-max values for quantization
-      // TODO: calculate proper inference values
-      tflite::reference_ops::Pad(params, getTensorShape(input()), getTensorData<float>(input()),
-                                 &pad_value, getTensorShape(output()),
-                                 getTensorData<float>(output()));
+      MirrorPadImpl<float>(*input(), *paddings(), params().mode, *output());
       break;
     }
     case DataType::U8:
     {
-      // NOTE: this implementation only obtains min-max values for quantization
-      // TODO: calculate proper inference values
       assert(output()->zero_point() >= std::numeric_limits<uint8_t>::min());
       assert(output()->zero_point() <= std::numeric_limits<uint8_t>::max());
-      const auto pad_value = static_cast<uint8_t>(output()->zero_point());
-      tflite::reference_ops::Pad(params, getTensorShape(input()), getTensorData<uint8_t>(input()),
-                                 &pad_value, getTensorShape(output()),
-                                 getTensorData<uint8_t>(output()));
+
+      MirrorPadImpl<uint8_t>(*input(), *paddings(), params().mode, *output());
       break;
     }
     default:
@@ -104,5 +84,87 @@ void MirrorPad::execute() const
   }
 }
 
+template <typename T>
+inline void MirrorPadImpl(const Tensor &input, const Tensor &paddings, MirrorPadMode mode,
+                          Tensor &output)
+{
+  auto const input_dims = input.shape().num_dims();
+  auto const input_data = input.data<T>();
+  auto const paddings_data = paddings.data<int32_t>();
+  auto const output_data = output.data<T>();
+
+  auto const input_b = input_dims > 3 ? input.shape().dim(input_dims - 4) : 1;
+  auto const input_h = input_dims > 2 ? input.shape().dim(input_dims - 3) : 1;
+  auto const input_w = input_dims > 1 ? input.shape().dim(input_dims - 2) : 1;
+  auto const input_d = input.shape().dim(input_dims - 1);
+
+  auto const input_h_offset = input_d * input_w;
+  auto const input_b_offset = input_h_offset * input_h;
+
+  auto const output_b = input_dims > 3 ? output.shape().dim(input_dims - 4) : 1;
+  auto const output_h = input_dims > 2 ? output.shape().dim(input_dims - 3) : 1;
+  auto const output_w = input_dims > 1 ? output.shape().dim(input_dims - 2) : 1;
+  auto const output_d = output.shape().dim(input_dims - 1);
+
+  auto const left_b_pad = paddings_data[2 * (input_dims - 4)];
+  auto const left_h_pad = paddings_data[2 * (input_dims - 3)];
+  auto const left_w_pad = paddings_data[2 * (input_dims - 2)];
+  auto const left_d_pad = paddings_data[2 * (input_dims - 1)];
+
+  auto const right_b_pad = paddings_data[2 * (input_dims - 4) + 1];
+  auto const right_h_pad = paddings_data[2 * (input_dims - 3) + 1];
+  auto const right_w_pad = paddings_data[2 * (input_dims - 2) + 1];
+  auto const right_d_pad = paddings_data[2 * (input_dims - 1) + 1];
+
+  const auto positive_mod = [](auto a, auto b) { return (a % b + b) % b; };
+  const auto offset_index = [input_d, input_h_offset, input_b_offset](auto d, auto w, auto h,
+                                                                      auto b) {
+    return d + w * input_d + h * input_h_offset + b * input_b_offset;
+  };
+
+  const auto symmetric_dim = [&positive_mod](auto i, auto left_pad, auto input) {
+    bool reflected = (((i < left_pad ? i + 1 - input : i) - left_pad) / input & 1) == 1;
+    return positive_mod(reflected ? input + left_pad - i - 1 : i - left_pad, input);
+  };
+
+  const T *in_ptr = input_data;
+  T *out_ptr = output_data;
+
+  for (int32_t b = 0; b < output_b; ++b)
+  {
+    for (int32_t h = 0; h < output_h; ++h)
+    {
+      for (int32_t w = 0; w < output_w; ++w)
+      {
+        for (int32_t d = 0; d < output_d; ++d)
+        {
+          if (b < left_b_pad || b >= output_b - right_b_pad || //
+              h < left_h_pad || h >= output_h - right_h_pad || //
+              w < left_w_pad || w >= output_w - right_w_pad || //
+              d < left_d_pad || d >= output_d - right_d_pad)
+          {
+            if (mode == MirrorPadMode::REFLECT)
+            {
+              *out_ptr++ = input_data[offset_index(
+                positive_mod(d - left_d_pad, input_d), positive_mod(w - left_w_pad, input_w),
+                positive_mod(h - left_h_pad, input_h), positive_mod(b - left_b_pad, input_b))];
+            }
+            else
+            {
+              *out_ptr++ = input_data[offset_index(
+                symmetric_dim(d, left_d_pad, input_d), symmetric_dim(w, left_w_pad, input_w),
+                symmetric_dim(h, left_h_pad, input_h), symmetric_dim(b, left_b_pad, input_b))];
+            }
+          }
+          else
+          {
+            *out_ptr++ = *in_ptr++;
+          }
+        }
+      }
+    }
+  }
+}
+
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/MirrorPad.test.cpp b/compiler/luci-interpreter/src/kernels/MirrorPad.test.cpp
index de9da5051..740d8cb22 100644
--- a/compiler/luci-interpreter/src/kernels/MirrorPad.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/MirrorPad.test.cpp
@@ -14,4 +14,212 @@
  * limitations under the License.
  */
 
-// TODO: Add tests for MirrorPad
+#include "kernels/MirrorPad.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class MirrorPadTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  void Execute(const Tensor &input, const Tensor &padding, Tensor &output, MirrorPadMode mode)
+  {
+    MirrorPadParams params{};
+    params.mode = mode;
+
+    MirrorPad kernel(&input, &padding, &output, params);
+    kernel.configure();
+    _memory_manager->allocate_memory(output);
+    kernel.execute();
+  }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(MirrorPadTest, FloatReflect)
+{
+  Shape input_shape = {1, 2, 2, 1};
+  Shape padding_shape = {4, 2};
+
+  std::vector<float> input_data{1.0f, 2.0f,  //
+                                3.0f, 4.0f}; //
+  std::vector<int> padding_data{0, 0, 2, 1, 1, 2, 0, 0};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>(padding_shape, padding_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::REFLECT);
+
+  std::vector<float> ref_output_data{2.0f, 1.0f, 2.0f, 1.0f, 2.0f,  //
+                                     4.0f, 3.0f, 4.0f, 3.0f, 4.0f,  //
+                                     2.0f, 1.0f, 2.0f, 1.0f, 2.0f,  //
+                                     4.0f, 3.0f, 4.0f, 3.0f, 4.0f,  //
+                                     2.0f, 1.0f, 2.0f, 1.0f, 2.0f}; //
+  std::initializer_list<int32_t> ref_output_shape{1, 5, 5, 1};
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(MirrorPadTest, FloatSymmetric)
+{
+  Shape input_shape = {1, 2, 2, 1};
+  Shape padding_shape = {4, 2};
+
+  std::vector<float> input_data{1.0f, 2.0f,  //
+                                3.0f, 4.0f}; //
+  std::vector<int> padding_data{0, 0, 2, 1, 1, 2, 0, 0};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>(padding_shape, padding_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::SYMMETRIC);
+
+  std::vector<float> ref_output_data{3.0, 3.0, 4.0, 4.0, 3.0,  //
+                                     1.0, 1.0, 2.0, 2.0, 1.0,  //
+                                     1.0, 1.0, 2.0, 2.0, 1.0,  //
+                                     3.0, 3.0, 4.0, 4.0, 3.0,  //
+                                     3.0, 3.0, 4.0, 4.0, 3.0}; //
+  std::initializer_list<int32_t> ref_output_shape{1, 5, 5, 1};
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(MirrorPadTest, FloatSymmetric2Dim)
+{
+  Shape input_shape = {3, 1};
+  Shape padding_shape = {2, 2};
+
+  std::vector<float> input_data{1.0f, 2.0f, 3.0f};
+  std::vector<int> padding_data{1, 2, 0, 0};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>(padding_shape, padding_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::SYMMETRIC);
+
+  std::vector<float> ref_output_data{1.0, 1.0, 2.0, 3.0, 3.0, 2.0};
+  std::initializer_list<int32_t> ref_output_shape{6, 1};
+
+  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(MirrorPadTest, Uint8Reflect)
+{
+  Shape input_shape = {1, 2, 3, 1};
+  Shape padding_shape = {4, 2};
+
+  float quant_tolerance = getTolerance(0.0f, 6.0f, 255);
+  std::pair<float, int32_t> quant_param = quantizationParams<uint8_t>(0.0f, 6.0f);
+
+  std::vector<float> input_data{1.0f, 2.0f, 3.0f,  //
+                                4.0f, 5.0f, 6.0f}; //
+  std::vector<int> padding_data{0, 0, 2, 1, 1, 3, 0, 0};
+
+  Tensor input_tensor = makeInputTensor<DataType::U8>(
+    input_shape, quant_param.first, quant_param.second, input_data, _memory_manager.get());
+
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>(padding_shape, padding_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::U8, quant_param.first, quant_param.second);
+
+  Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::REFLECT);
+
+  std::vector<float> ref_output_data{
+    3.0f, 1.0f, 2.0f, 3.0f, 1.0f, 2.0f, 3.0f, //
+    6.0f, 4.0f, 5.0f, 6.0f, 4.0f, 5.0f, 6.0f, //
+    3.0f, 1.0f, 2.0f, 3.0f, 1.0f, 2.0f, 3.0f, //
+    6.0f, 4.0f, 5.0f, 6.0f, 4.0f, 5.0f, 6.0f, //
+    3.0f, 1.0f, 2.0f, 3.0f, 1.0f, 2.0f, 3.0f, //
+  };
+  std::initializer_list<int32_t> ref_output_shape{1, 5, 7, 1};
+
+  EXPECT_THAT(dequantizeTensorData(output_tensor),
+              FloatArrayNear(ref_output_data, quant_tolerance));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(MirrorPadTest, Uint8Symmetric)
+{
+  Shape input_shape = {1, 2, 3, 1};
+  Shape padding_shape = {4, 2};
+
+  float quant_tolerance = getTolerance(0.0f, 6.0f, 255);
+  std::pair<float, int32_t> quant_param = quantizationParams<uint8_t>(0.0f, 6.0f);
+
+  std::vector<float> input_data{1.0f, 2.0f, 3.0f,  //
+                                4.0f, 5.0f, 6.0f}; //
+  std::vector<int> padding_data{0, 0, 2, 1, 1, 3, 0, 0};
+
+  Tensor input_tensor = makeInputTensor<DataType::U8>(
+    input_shape, quant_param.first, quant_param.second, input_data, _memory_manager.get());
+
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>(padding_shape, padding_data, _memory_manager.get());
+
+  Tensor output_tensor = makeOutputTensor(DataType::U8, quant_param.first, quant_param.second);
+
+  Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::SYMMETRIC);
+
+  std::vector<float> ref_output_data{
+    4.0f, 4.0f, 5.0f, 6.0f, 6.0f, 5.0f, 4.0f, //
+    1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 2.0f, 1.0f, //
+    1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 2.0f, 1.0f, //
+    4.0f, 4.0f, 5.0f, 6.0f, 6.0f, 5.0f, 4.0f, //
+    4.0f, 4.0f, 5.0f, 6.0f, 6.0f, 5.0f, 4.0f, //
+  };
+  std::initializer_list<int32_t> ref_output_shape{1, 5, 7, 1};
+
+  EXPECT_THAT(dequantizeTensorData(output_tensor),
+              FloatArrayNear(ref_output_data, quant_tolerance));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(MirrorPadTest, UnsupportedDim_NEG)
+{
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 1, 1, 1, 1}, {1.0f}, _memory_manager.get());
+  Tensor padding_tensor =
+    makeInputTensor<DataType::S32>({5, 2}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  EXPECT_ANY_THROW(Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::REFLECT));
+}
+
+TEST_F(MirrorPadTest, InvalidInputType_NEG)
+{
+  Tensor input_tensor = makeInputTensor<DataType::S64>({1}, {1}, _memory_manager.get());
+  Tensor padding_tensor = makeInputTensor<DataType::S32>({1, 2}, {0, 0}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S64);
+
+  EXPECT_ANY_THROW(Execute(input_tensor, padding_tensor, output_tensor, MirrorPadMode::REFLECT));
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Mul.cpp b/compiler/luci-interpreter/src/kernels/Mul.cpp
index bc855de0f..531fb4fa1 100644
--- a/compiler/luci-interpreter/src/kernels/Mul.cpp
+++ b/compiler/luci-interpreter/src/kernels/Mul.cpp
@@ -42,6 +42,8 @@ void Mul::configure()
   LUCI_INTERPRETER_CHECK(output()->element_type() == input1()->element_type());
   if (input1()->element_type() == DataType::S16)
   {
+    LUCI_INTERPRETER_CHECK(input1()->zero_points().size() == 1 &&
+                           input2()->zero_points().size() == 1)
     LUCI_INTERPRETER_CHECK(input1()->zero_point() == 0 && input2()->zero_point() == 0 &&
                            output()->zero_point() == 0);
   }
@@ -56,6 +58,12 @@ void Mul::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::S16:
       evalQuantizedS16();
       break;
@@ -66,13 +74,8 @@ void Mul::execute() const
 
 void Mul::evalFloat() const
 {
-  float activation_min{};
-  float activation_max{};
-  calculateActivationRange(_params.activation, &activation_min, &activation_max);
-
   tflite::ArithmeticParams params{};
-  params.float_activation_min = activation_min;
-  params.float_activation_max = activation_max;
+  fillArithmeticActivationRange<float>(params, _params.activation);
 
   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
     getTensorShape(input1()), getTensorShape(input2()), &params);
@@ -91,6 +94,28 @@ void Mul::evalFloat() const
   }
 }
 
+template <typename T> void Mul::evalInteger() const
+{
+  tflite::ArithmeticParams params{};
+  fillArithmeticActivationRange<T>(params, _params.activation);
+
+  const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
+    getTensorShape(input1()), getTensorShape(input2()), &params);
+
+  if (need_broadcast)
+  {
+    luci_interpreter_pal::BroadcastMul4DSlow(
+      params, getTensorShape(input1()), getTensorData<T>(input1()), getTensorShape(input2()),
+      getTensorData<T>(input2()), getTensorShape(output()), getTensorData<T>(output()));
+  }
+  else
+  {
+    luci_interpreter_pal::Mul(params, getTensorShape(input1()), getTensorData<T>(input1()),
+                              getTensorShape(input2()), getTensorData<T>(input2()),
+                              getTensorShape(output()), getTensorData<T>(output()));
+  }
+}
+
 void Mul::evalQuantizedS16() const
 {
   const auto input1_scale = static_cast<double>(input1()->scale());
diff --git a/compiler/luci-interpreter/src/kernels/Mul.h b/compiler/luci-interpreter/src/kernels/Mul.h
index 2ccf60f3a..c0cf817df 100644
--- a/compiler/luci-interpreter/src/kernels/Mul.h
+++ b/compiler/luci-interpreter/src/kernels/Mul.h
@@ -42,6 +42,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantizedS16() const;
 };
 
diff --git a/compiler/luci-interpreter/src/kernels/Mul.test.cpp b/compiler/luci-interpreter/src/kernels/Mul.test.cpp
index 471f6ac86..fc0e60614 100644
--- a/compiler/luci-interpreter/src/kernels/Mul.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Mul.test.cpp
@@ -93,6 +93,78 @@ TEST_F(MulTest, Float)
   }
 }
 
+template <loco::DataType DType> void checkInteger(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  Shape base_shape = {2, 3, 1, 2};
+  std::vector<Shape> test_shapes{{1, 1, 3, 2}, {1, 3, 1, 2}, {2, 1, 3, 1}, {2, 3, 1, 1}};
+
+  dtype max_value = std::numeric_limits<dtype>::max();
+  dtype res_max = max_value - max_value % 10;
+
+  std::vector<std::vector<dtype>> test_outputs = {
+    {8,  0, 20,  0, 4,  30,  //
+     16, 0, 40,  3, 8,  0,   //
+     0,  0, 0,   6, 0,  0,   //
+     4,  0, 10,  9, 2,  0,   //
+     40, 0, 100, 0, 20, 150, //
+     28, 0, 70,  0, 14, res_max},
+    {8, 0, 40, 3, 0, 0, 4, 0, 100, 0, 14, res_max},
+    {8,  12,     0, 0, 20, 30, 16, 0, 0, 0,  40, 0,   0,   0, 0, 0,  0,
+     0,  0,      9, 2, 0,  10, 0,  0, 0, 20, 30, 100, 150, 0, 0, 14, max_value / 10 * 2,
+     70, res_max},
+    {8, 12, 0, 0, 0, 0, 0, 9, 20, 30, 70, res_max}};
+  std::vector<dtype> input1_data{2, 3, 4, -1, -3, -2, 1, -3, 10, 15, 7, max_value / 10};
+  std::vector<dtype> input2_data{4, 0, 10, -3, 2, 10};
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    MulParams params{};
+    params.activation = Activation::RELU;
+
+    Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+  // Re-run with exchanged inputs.
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    MulParams params{};
+    params.activation = Activation::RELU;
+
+    Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+}
+
+TEST_F(MulTest, SInt64)
+{
+  checkInteger<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(MulTest, SInt32)
+{
+  checkInteger<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
 TEST_F(MulTest, SInt16)
 {
   Shape base_shape = {2, 3, 1, 2};
@@ -161,6 +233,60 @@ TEST_F(MulTest, SInt16)
   }
 }
 
+TEST_F(MulTest, Input_Output_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::FLOAT32>({1}, {1.f}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S32>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  MulParams params{};
+  params.activation = Activation::RELU;
+
+  Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(MulTest, Invalid_Output_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::S64>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  MulParams params{};
+  params.activation = Activation::RELU;
+
+  Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(MulTest, Invalid_Input_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::U64>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::U64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U64);
+
+  MulParams params{};
+  params.activation = Activation::RELU;
+
+  Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  EXPECT_ANY_THROW(kernel.execute());
+}
+
+TEST_F(MulTest, Invalid_Quantization_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::S16>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S16>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16);
+
+  MulParams params{};
+  params.activation = Activation::NONE;
+
+  Mul kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/NotEqual.cpp b/compiler/luci-interpreter/src/kernels/NotEqual.cpp
index 99d5e0fa0..54e5eee34 100644
--- a/compiler/luci-interpreter/src/kernels/NotEqual.cpp
+++ b/compiler/luci-interpreter/src/kernels/NotEqual.cpp
@@ -49,6 +49,12 @@ void NotEqual::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -79,6 +85,29 @@ void NotEqual::evalFloat() const
   }
 }
 
+template <typename T> void NotEqual::evalInteger() const
+{
+  const auto x_data = getTensorData<T>(x());
+  const auto y_data = getTensorData<T>(y());
+  auto output_data = getTensorData<bool>(output());
+
+  tflite::ComparisonParams op_params;
+  op_params.is_broadcast = x()->shape() != y()->shape();
+
+  if (op_params.is_broadcast)
+  {
+    tflite::reference_ops::Broadcast4DSlowNotEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                                            getTensorShape(y()), y_data,
+                                                            getTensorShape(output()), output_data);
+  }
+  else
+  {
+    tflite::reference_ops::NotEqualNoScaling(op_params, getTensorShape(x()), x_data,
+                                             getTensorShape(y()), y_data, getTensorShape(output()),
+                                             output_data);
+  }
+}
+
 void NotEqual::evalQuantized() const
 {
   const auto x_data = getTensorData<uint8_t>(x());
diff --git a/compiler/luci-interpreter/src/kernels/NotEqual.h b/compiler/luci-interpreter/src/kernels/NotEqual.h
index 247874df7..d2aafe893 100644
--- a/compiler/luci-interpreter/src/kernels/NotEqual.h
+++ b/compiler/luci-interpreter/src/kernels/NotEqual.h
@@ -38,6 +38,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 
 private:
diff --git a/compiler/luci-interpreter/src/kernels/NotEqual.test.cpp b/compiler/luci-interpreter/src/kernels/NotEqual.test.cpp
index 763f86893..45bf4022a 100644
--- a/compiler/luci-interpreter/src/kernels/NotEqual.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/NotEqual.test.cpp
@@ -99,6 +99,82 @@ TEST_F(NotEqualTest, FloatBroardcast)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
 }
 
+template <loco::DataType DType>
+void checkIntegerSimple(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{min_value, 2, max_value};
+
+  std::vector<dtype> y_data{min_value, -2, max_value};
+
+  std::vector<bool> ref_output_data{false, true, false};
+
+  Tensor x_tensor = makeInputTensor<DType>({3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  NotEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({3}));
+}
+
+template <loco::DataType DType>
+void checkIntegerBroadcast(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  dtype min_value = std::numeric_limits<dtype>::min();
+  dtype max_value = std::numeric_limits<dtype>::max();
+  std::vector<dtype> x_data{
+    min_value, 2,  3,         // Row 1
+    4,         5,  max_value, // Row 2
+    -1,        -2, -3,        // Row 3
+    min_value, -2, max_value, // Row 4
+  };
+
+  std::vector<dtype> y_data{
+    min_value, -2, max_value, // Row 1
+  };
+
+  std::vector<bool> ref_output_data{
+    false, true,  true,  // Row 1
+    true,  true,  false, // Row 2
+    true,  false, true,  // Row 3
+    false, false, false, // Row 4
+  };
+
+  Tensor x_tensor = makeInputTensor<DType>({4, 3}, x_data, memory_manager);
+  Tensor y_tensor = makeInputTensor<DType>({3}, y_data, memory_manager);
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  NotEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<bool>(output_tensor), ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({4, 3}));
+}
+
+TEST_F(NotEqualTest, Int32)
+{
+  checkIntegerSimple<loco::DataType::S32>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(NotEqualTest, Int64)
+{
+  checkIntegerSimple<loco::DataType::S64>(_memory_manager.get());
+  checkIntegerBroadcast<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 // Choose min / max in such a way that there are exactly 256 units to avoid rounding errors.
 const float F_MIN = -128.0 / 128.0;
 const float F_MAX = 127.0 / 128.0;
@@ -195,6 +271,36 @@ TEST_F(NotEqualTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
+TEST_F(NotEqualTest, Float_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::FLOAT32>({2}, {1.f, 2.f}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::FLOAT32>({3}, {1.f, 2.f, 3.f}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  NotEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(NotEqualTest, Int32_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S32>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S32>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  NotEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(NotEqualTest, Int64_Broadcast_NEG)
+{
+  Tensor x_tensor = makeInputTensor<DataType::S64>({2}, {1, 2}, _memory_manager.get());
+  Tensor y_tensor = makeInputTensor<DataType::S64>({3}, {1, 2, 3}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::BOOL);
+
+  NotEqual kernel(&x_tensor, &y_tensor, &output_tensor);
+  ASSERT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/OneHot.cpp b/compiler/luci-interpreter/src/kernels/OneHot.cpp
new file mode 100644
index 000000000..4d3e5f2ef
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/OneHot.cpp
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/OneHot.h"
+#include "kernels/Utils.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+namespace
+{
+
+template <typename T>
+void OneHotComputeImpl(const Tensor *indices_tensor, const Tensor *on_value_tensor,
+                       const Tensor *off_value_tensor, int32_t depth, int32_t axis,
+                       Tensor *output_tensor)
+{
+  // define input shape and correct axis
+  auto const &input_shape = indices_tensor->shape();
+  axis = axis == -1 ? input_shape.num_dims() : axis;
+
+  // TODO support other integer input types
+  auto const *indices = getTensorData<int32_t>(indices_tensor);
+  auto const on_value = getTensorData<T>(on_value_tensor)[0];
+  auto const off_value = getTensorData<T>(off_value_tensor)[0];
+  auto *output = getTensorData<T>(output_tensor);
+
+  // prefix_dim_size == # of elements before the axis
+  // depth == # of elements per axis
+  // suffix_dim_size == # of elements after the axis
+  auto prefix_dim_size = 1;
+  for (int32_t i = 0; i < axis; ++i)
+  {
+    prefix_dim_size *= input_shape.dim(i);
+  }
+  assert(prefix_dim_size > 0);
+  auto const suffix_dim_size = input_shape.num_elements() / prefix_dim_size;
+
+  // View the indices as a matrix of size:
+  //     prefix_dim_size x suffix_dim_size
+  // View the output as a matrix of size:
+  //     prefix_dim_size x depth x suffix_dim_size
+  // Then the output is:
+  //     output(i, j, k) == (indices(i, k) == j) ? on : off
+  for (int32_t i = 0; i < prefix_dim_size; ++i)
+    for (int32_t j = 0; j < depth; ++j)
+      for (int32_t k = 0; k < suffix_dim_size; ++k, ++output)
+        *output = indices[i * suffix_dim_size + k] == j ? on_value : off_value;
+}
+
+} // namespace
+
+OneHot::OneHot(const Tensor *indices, const Tensor *depth, const Tensor *on_value,
+               const Tensor *off_value, Tensor *output, const OneHotParams &params)
+  : KernelWithParams<OneHotParams>({indices, depth, on_value, off_value}, {output}, params)
+{
+  // Do nothing
+}
+
+void OneHot::configure()
+{
+  // check types
+  LUCI_INTERPRETER_CHECK(indices()->element_type() == DataType::S32);
+  LUCI_INTERPRETER_CHECK(depth()->element_type() == DataType::S32);
+  LUCI_INTERPRETER_CHECK(on_value()->element_type() == off_value()->element_type());
+  LUCI_INTERPRETER_CHECK(output()->element_type() == on_value()->element_type());
+
+  // check shape dependent parameters
+  LUCI_INTERPRETER_CHECK(on_value()->shape().num_elements() == 1);
+  LUCI_INTERPRETER_CHECK(off_value()->shape().num_elements() == 1);
+  LUCI_INTERPRETER_CHECK(depth()->shape().num_elements() == 1);
+  LUCI_INTERPRETER_CHECK(params().axis >= -1 && params().axis <= indices()->shape().num_dims());
+
+  // define parameters that affect the output shape
+  auto const depth_value = getTensorData<int32_t>(depth())[0];
+  auto const &input_shape = indices()->shape();
+  auto const input_dims = input_shape.num_dims();
+  auto const axis = params().axis == -1 ? input_dims : params().axis;
+
+  // define output shape
+  Shape output_shape(input_shape.num_dims() + 1);
+  {
+    for (int32_t d = 0; d < axis; ++d)
+      output_shape.dim(d) = input_shape.dim(d);
+
+    output_shape.dim(axis) = depth_value;
+
+    for (int32_t d = axis + 1; d < output_shape.num_dims(); ++d)
+      output_shape.dim(d) = input_shape.dim(d - 1);
+  }
+
+  // reshape output
+  output()->resize(output_shape);
+}
+
+void OneHot::execute() const
+{
+  auto const depth_value = getTensorData<int32_t>(depth())[0];
+  auto const axis = params().axis;
+
+  switch (output()->element_type())
+  {
+    case loco::DataType::FLOAT32:
+      OneHotComputeImpl<float>(indices(), on_value(), off_value(), depth_value, axis, output());
+      break;
+    case loco::DataType::U8:
+      OneHotComputeImpl<uint8_t>(indices(), on_value(), off_value(), depth_value, axis, output());
+      break;
+    case loco::DataType::S16:
+      OneHotComputeImpl<int16_t>(indices(), on_value(), off_value(), depth_value, axis, output());
+      break;
+    default:
+      // TODO Support other data types
+      throw std::runtime_error("Not supported, yet!");
+      break;
+  }
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/OneHot.h b/compiler/luci-interpreter/src/kernels/OneHot.h
new file mode 100644
index 000000000..572f857ae
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/OneHot.h
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_ONEHOT_H
+#define LUCI_INTERPRETER_KERNELS_ONEHOT_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class OneHot : public KernelWithParams<OneHotParams>
+{
+public:
+  OneHot(const Tensor *indices, const Tensor *depth, const Tensor *on_value,
+         const Tensor *off_value, Tensor *output, const OneHotParams &params);
+
+  const Tensor *indices() const { return _inputs[0]; }
+  const Tensor *depth() const { return _inputs[1]; }
+  const Tensor *on_value() const { return _inputs[2]; }
+  const Tensor *off_value() const { return _inputs[3]; }
+
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_ONEHOT_H
diff --git a/compiler/luci-interpreter/src/kernels/OneHot.test.cpp b/compiler/luci-interpreter/src/kernels/OneHot.test.cpp
new file mode 100644
index 000000000..45b6968fa
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/OneHot.test.cpp
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/OneHot.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+template <typename T1, typename T2>
+void Check(std::initializer_list<int32_t> input_shape, std::initializer_list<int32_t> output_shape,
+           std::initializer_list<T1> input_data, std::initializer_list<int32_t> depth_data,
+           std::initializer_list<T2> on_value_data, std::initializer_list<T2> off_value_data,
+           int32_t axis, std::initializer_list<T2> output_data)
+{
+  std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
+
+  constexpr auto input_type = getElementType<T1>();
+  constexpr auto output_type = getElementType<T2>();
+
+  Tensor input_tensor = makeInputTensor<input_type>(input_shape, input_data, memory_manager.get());
+  Tensor depth_tensor = makeInputTensor<DataType::S32>({}, depth_data, memory_manager.get());
+  Tensor on_value_tensor = makeInputTensor<output_type>({}, on_value_data, memory_manager.get());
+  Tensor off_value_tensor = makeInputTensor<output_type>({}, off_value_data, memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(output_type);
+
+  OneHotParams params{};
+  params.axis = axis;
+
+  OneHot kernel(&input_tensor, &depth_tensor, &on_value_tensor, &off_value_tensor, &output_tensor,
+                params);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorShape(output_tensor), output_shape);
+  EXPECT_THAT(extractTensorData<T2>(output_tensor), ::testing::ElementsAreArray(output_data));
+}
+
+template <typename T> class OneHotTest : public ::testing::Test
+{
+};
+
+using DataTypes = ::testing::Types<float, uint8_t, int16_t>;
+TYPED_TEST_SUITE(OneHotTest, DataTypes);
+
+TYPED_TEST(OneHotTest, BasicPattern)
+{
+  // axis 0
+  Check<int32_t, TypeParam>(/*input_shape=*/{2, 3}, /*output_shape=*/{4, 2, 3},
+                            /*input_data=*/
+                            {
+                              0, 3, 5, //
+                              7, 3, 0, //
+                            },
+                            /*depth_data=*/{4}, /*on_value_data=*/{1}, /*off_value_data=*/{0},
+                            /*axis=*/0,
+                            /*output_data=*/
+                            {
+                              1, 0, 0, //
+                              0, 0, 1, //
+
+                              0, 0, 0, //
+                              0, 0, 0, //
+
+                              0, 0, 0, //
+                              0, 0, 0, //
+
+                              0, 1, 0, //
+                              0, 1, 0, //
+                            });
+  // axis 1
+  Check<int32_t, TypeParam>(/*input_shape=*/{2, 3}, /*output_shape=*/{2, 4, 3},
+                            /*input_data=*/
+                            {
+                              0, 3, 5, //
+                              7, 3, 0, //
+                            },
+                            /*depth_data=*/{4}, /*on_value_data=*/{1}, /*off_value_data=*/{0},
+                            /*axis=*/1,
+                            /*output_data=*/
+                            {
+                              1, 0, 0, //
+                              0, 0, 0, //
+                              0, 0, 0, //
+                              0, 1, 0, //
+
+                              0, 0, 1, //
+                              0, 0, 0, //
+                              0, 0, 0, //
+                              0, 1, 0, //
+                            });
+  // axis -1
+  Check<int32_t, TypeParam>(/*input_shape=*/{2, 3}, /*output_shape=*/{2, 3, 4},
+                            /*input_data=*/
+                            {
+                              0, 3, 5, //
+                              7, 3, 0, //
+                            },
+                            /*depth_data=*/{4}, /*on_value_data=*/{1}, /*off_value_data=*/{0},
+                            /*axis=*/-1,
+                            /*output_data=*/
+                            {
+                              1, 0, 0, 0, //
+                              0, 0, 0, 1, //
+                              0, 0, 0, 0, //
+
+                              0, 0, 0, 0, //
+                              0, 0, 0, 1, //
+                              1, 0, 0, 0, //
+                            });
+}
+
+TEST(OneHotTest, UnsupportedInputType_NEG)
+{
+  std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
+
+  // input type should be integer
+  Tensor input_tensor = makeInputTensor<DataType::FLOAT32>({1}, {0}, memory_manager.get());
+
+  Tensor depth_tensor = makeInputTensor<DataType::S32>({}, {1}, memory_manager.get());
+  Tensor on_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {1.0}, memory_manager.get());
+  Tensor off_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {0.0}, memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  OneHotParams params = {-1};
+
+  OneHot kernel(&input_tensor, &depth_tensor, &on_value_tensor, &off_value_tensor, &output_tensor,
+                params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST(OneHotTest, OutputTypeMismatch_NEG)
+{
+  std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
+
+  Tensor input_tensor = makeInputTensor<DataType::S32>({1}, {0}, memory_manager.get());
+  Tensor depth_tensor = makeInputTensor<DataType::S32>({}, {1}, memory_manager.get());
+
+  // type of on_value, off_value and output_tensor should be same
+  Tensor on_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {1.0}, memory_manager.get());
+  Tensor off_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {0.0}, memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16);
+
+  OneHotParams params = {-1};
+
+  OneHot kernel(&input_tensor, &depth_tensor, &on_value_tensor, &off_value_tensor, &output_tensor,
+                params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST(OneHotTest, InvalidAxis_NEG)
+{
+  std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
+
+  Tensor input_tensor = makeInputTensor<DataType::S32>({1}, {0}, memory_manager.get());
+  Tensor depth_tensor = makeInputTensor<DataType::S32>({}, {1}, memory_manager.get());
+  Tensor on_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {1.0}, memory_manager.get());
+  Tensor off_value_tensor = makeInputTensor<DataType::FLOAT32>({}, {0.0}, memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  // axis should be in [-1, input_shape.rank]
+  OneHotParams params = {-2};
+
+  OneHot kernel(&input_tensor, &depth_tensor, &on_value_tensor, &off_value_tensor, &output_tensor,
+                params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Pack.test.cpp b/compiler/luci-interpreter/src/kernels/Pack.test.cpp
index 90a0f894e..2404e4303 100644
--- a/compiler/luci-interpreter/src/kernels/Pack.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Pack.test.cpp
@@ -80,7 +80,7 @@ template <typename T> class PackTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<uint8_t, float>;
-TYPED_TEST_CASE(PackTest, DataTypes);
+TYPED_TEST_SUITE(PackTest, DataTypes);
 
 TYPED_TEST(PackTest, ThreeInputs)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Pad.cpp b/compiler/luci-interpreter/src/kernels/Pad.cpp
index 700448e7a..fe172884b 100644
--- a/compiler/luci-interpreter/src/kernels/Pad.cpp
+++ b/compiler/luci-interpreter/src/kernels/Pad.cpp
@@ -93,6 +93,16 @@ void Pad::execute() const
                                  getTensorData<uint8_t>(output()));
       break;
     }
+    case DataType::S8:
+    {
+      assert(output()->zero_point() >= std::numeric_limits<int8_t>::min());
+      assert(output()->zero_point() <= std::numeric_limits<int8_t>::max());
+      const auto pad_value = static_cast<int8_t>(output()->zero_point());
+      tflite::reference_ops::Pad(params, getTensorShape(input()), getTensorData<int8_t>(input()),
+                                 &pad_value, getTensorShape(output()),
+                                 getTensorData<int8_t>(output()));
+      break;
+    }
     default:
       throw std::runtime_error("Unsupported type.");
   }
diff --git a/compiler/luci-interpreter/src/kernels/Pad.test.cpp b/compiler/luci-interpreter/src/kernels/Pad.test.cpp
index 7994263e2..dd3ce947c 100644
--- a/compiler/luci-interpreter/src/kernels/Pad.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Pad.test.cpp
@@ -54,6 +54,32 @@ TEST(Pad, Uint8)
   EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 4, 7, 1}));
 }
 
+TEST(Pad, Int8)
+{
+  std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
+  float kQuantizedTolerance = GetTolerance(-1.0, 1.0);
+  std::pair<float, int32_t> quant_param = quantizationParams<int8_t>(-1.0f, 1.0f);
+  std::vector<float> input_data{-0.2, 0.4, 0.5, -0.7, -0.1, -0.9, 0.7, 0.1, 0.2};
+  std::vector<int32_t> paddings_data{0, 0, 1, 2, 2, 1, 0, 0};
+  Tensor input_tensor = makeInputTensor<DataType::S8>(
+    {1, 3, 3, 1}, quant_param.first, quant_param.second, input_data, memory_manager.get());
+  Tensor paddings_tensor =
+    makeInputTensor<DataType::S32>({4, 2}, paddings_data, memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, quant_param.first, quant_param.second);
+
+  Pad kernel(&input_tensor, &paddings_tensor, &output_tensor);
+  kernel.configure();
+  memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  std::vector<float> ref_output_data{0, 0, 0,    0,    0,    0, 0, 0, -0.2, 0.4, 0.5, 0,
+                                     0, 0, -0.7, -0.1, -0.9, 0, 0, 0, 0.7,  0.1, 0.2, 0,
+                                     0, 0, 0,    0,    0,    0, 0, 0, 0,    0,   0,   0};
+  EXPECT_THAT(dequantizeTensorData(output_tensor),
+              FloatArrayNear(ref_output_data, kQuantizedTolerance));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 6, 6, 1}));
+}
+
 TEST(Pad, Float)
 {
   std::unique_ptr<IMemoryManager> memory_manager = std::make_unique<TestMemoryManager>();
diff --git a/compiler/luci-interpreter/src/kernels/Quantize.cpp b/compiler/luci-interpreter/src/kernels/Quantize.cpp
new file mode 100644
index 000000000..0c8544a65
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Quantize.cpp
@@ -0,0 +1,160 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Quantize.h"
+#include "kernels/Utils.h"
+#include "PALQuantize.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+namespace
+{
+
+template <typename input_dtype> void call_requantize(const Tensor *input, Tensor *output)
+{
+  int32_t multiplier;
+  int shift;
+
+  const double effective_output_scale = input->scale() / output->scale();
+  quantizeMultiplier(effective_output_scale, &multiplier, &shift);
+
+  const auto input_shape = getTensorShape(input);
+  const auto output_shape = getTensorShape(output);
+  const auto size = tflite::MatchingFlatSize(input_shape, output_shape);
+
+  const auto input_data = getTensorData<input_dtype>(input);
+
+  switch (output->element_type())
+  {
+    case loco::DataType::S8:
+      luci_interpreter_pal::Requantize(input_data, size, multiplier, shift, input->zero_point(),
+                                       output->zero_point(), getTensorData<int8_t>(output));
+      break;
+    case loco::DataType::U8:
+      luci_interpreter_pal::Requantize(input_data, size, multiplier, shift, input->zero_point(),
+                                       output->zero_point(), getTensorData<uint8_t>(output));
+      break;
+    case loco::DataType::S16:
+      luci_interpreter_pal::Requantize(input_data, size, multiplier, shift, input->zero_point(),
+                                       output->zero_point(), getTensorData<int16_t>(output));
+      break;
+    default:
+      throw std::runtime_error("Unsupported quantized type, yet!");
+  }
+}
+
+} // namespace
+
+Quantize::Quantize(const Tensor *input, Tensor *output) : Kernel({input}, {output}) {}
+
+void Quantize::configure()
+{
+
+  if (input()->element_type() == loco::DataType::S16)
+    LUCI_INTERPRETER_CHECK(input()->zero_point() == 0);
+
+  switch (input()->element_type())
+  {
+    case loco::DataType::FLOAT32:
+    {
+      LUCI_INTERPRETER_CHECK(output()->element_type() == loco::DataType::U8 ||
+                             output()->element_type() == loco::DataType::S8 ||
+                             output()->element_type() == loco::DataType::S16);
+      break;
+    }
+    case loco::DataType::S16:
+    case loco::DataType::S8:
+    case loco::DataType::U8:
+    {
+      LUCI_INTERPRETER_CHECK(output()->element_type() == loco::DataType::S8 ||
+                             output()->element_type() == loco::DataType::U8 ||
+                             output()->element_type() == loco::DataType::S16);
+      if (output()->element_type() == loco::DataType::S16)
+      {
+        LUCI_INTERPRETER_CHECK(output()->zero_point() == 0);
+      }
+      break;
+    }
+    default:
+      throw std::runtime_error("Unsupported type");
+  }
+
+  output()->resize(input()->shape());
+}
+
+void Quantize::execute() const
+{
+  switch (input()->element_type())
+  {
+    case loco::DataType::FLOAT32:
+    {
+      tflite::QuantizationParams op_params;
+      op_params.zero_point = output()->zero_point();
+      op_params.scale = output()->scale();
+      const auto input_data = getTensorData<float>(input());
+
+      switch (output()->element_type())
+      {
+        case loco::DataType::S8:
+        {
+          luci_interpreter_pal::Quantize(op_params, getTensorShape(input()), input_data,
+                                         getTensorShape(output()), getTensorData<int8_t>(output()));
+          break;
+        }
+        case loco::DataType::U8:
+        {
+          luci_interpreter_pal::Quantize(op_params, getTensorShape(input()), input_data,
+                                         getTensorShape(output()),
+                                         getTensorData<uint8_t>(output()));
+          break;
+        }
+        case loco::DataType::S16:
+        {
+          luci_interpreter_pal::Quantize(op_params, getTensorShape(input()), input_data,
+                                         getTensorShape(output()),
+                                         getTensorData<int16_t>(output()));
+          break;
+        }
+        default:
+          throw std::runtime_error("Unsupported type.");
+      }
+      break;
+    }
+    case loco::DataType::S16:
+    {
+      call_requantize<int16_t>(input(), output());
+      break;
+    }
+    case loco::DataType::S8:
+    {
+      call_requantize<int8_t>(input(), output());
+      break;
+    }
+    case loco::DataType::U8:
+    {
+      call_requantize<uint8_t>(input(), output());
+      break;
+    }
+    default:
+      throw std::runtime_error("Unsupported type.");
+  }
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Quantize.h b/compiler/luci-interpreter/src/kernels/Quantize.h
new file mode 100644
index 000000000..006c5366f
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Quantize.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_QUANTIZE_H
+#define LUCI_INTERPRETER_KERNELS_QUANTIZE_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class Quantize : public Kernel
+{
+public:
+  Quantize(const Tensor *input, Tensor *output);
+
+  const Tensor *input() const { return _inputs[0]; }
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_QUANTIZE_H
diff --git a/compiler/luci-interpreter/src/kernels/Quantize.test.cpp b/compiler/luci-interpreter/src/kernels/Quantize.test.cpp
new file mode 100644
index 000000000..22e67fe3f
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/Quantize.test.cpp
@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/Quantize.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class QuantizeTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(QuantizeTest, FloatUint8)
+{
+  std::vector<float> input_data{-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64};
+
+  std::vector<uint8_t> ref_output_data{0, 1, 2, 3, 4, 251, 252, 253, 254, 255};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 5}, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U8, /*scale*/ 0.5, /*zero_point*/ 127);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<uint8_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(QuantizeTest, FloatInt8)
+{
+  std::vector<float> input_data{-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64};
+
+  std::vector<int8_t> ref_output_data{-128, -127, -126, -125, -124, 123, 124, 125, 126, 127};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 5}, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, /*scale*/ 0.5, /*zero_point*/ -1);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int8_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(QuantizeTest, FloatInt16)
+{
+  std::vector<float> input_data{-63.5, -63, -3, -2, -1, 1, 2, 3, 63.5, 64};
+
+  std::vector<int16_t> ref_output_data{-12700, -12600, -600, -400,  -200,
+                                       200,    400,    600,  12700, 12800};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({2, 5}, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16, /*scale*/ 0.005, /*zero_point*/ 0);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int16_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({2, 5}));
+}
+
+TEST_F(QuantizeTest, Int16Int16)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  std::vector<int16_t> ref_output_data{2, 4, 6, 8, 10, 12, 14, 16, 18, 20};
+
+  Tensor input_tensor = makeInputTensor<DataType::S16>(
+    {1, 1, 2, 5}, /*scale*/ 1.0, /*zero_point*/ 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16, /*scale*/ 0.5, /*zero_point*/ 0);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int16_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 1, 2, 5}));
+}
+
+TEST_F(QuantizeTest, Int8Int8)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  std::vector<int8_t> ref_output_data{1, 3, 5, 7, 9, 11, 13, 15, 17, 19};
+
+  Tensor input_tensor = makeInputTensor<DataType::S8>(
+    {1, 1, 2, 5}, /*scale*/ 0.5, /*zero_point*/ -1, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, /*scale*/ 0.5, /*zero_point*/ -1);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int8_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 1, 2, 5}));
+}
+
+TEST_F(QuantizeTest, Uint8Uint8)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  std::vector<uint8_t> ref_output_data{129, 131, 133, 135, 137, 139, 141, 143, 145, 147};
+
+  Tensor input_tensor = makeInputTensor<DataType::U8>(
+    {1, 1, 2, 5}, /*scale*/ 0.5, /*zero_point*/ 127, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U8, /*scale*/ 0.5, /*zero_point*/ 127);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<uint8_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 1, 2, 5}));
+}
+
+TEST_F(QuantizeTest, Int16Int8)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  std::vector<int8_t> ref_output_data{1, 3, 5, 7, 9, 11, 13, 15, 17, 19};
+
+  Tensor input_tensor = makeInputTensor<DataType::S16>(
+    {1, 1, 2, 5}, /*scale*/ 1.0, /*zero_point*/ 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, /*scale*/ 0.5, /*zero_point*/ -1);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  kernel.execute();
+
+  EXPECT_THAT(extractTensorData<int8_t>(output_tensor),
+              ::testing::ElementsAreArray(ref_output_data));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray({1, 1, 2, 5}));
+}
+
+TEST_F(QuantizeTest, InvalidInputType_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>({1, 1, 2, 5}, 0.5, 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S8, /*scale*/ 0.5, /*zero_point*/ -1);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(QuantizeTest, InvalidOutputTypeForFloatInput_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>({1, 1, 2, 5}, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(QuantizeTest, InvalidOutputTypeForInt16Input_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S16>({1, 1, 2, 5}, 0.5, 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(QuantizeTest, InvalidOutputTypeForInt8Input_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S8>({1, 1, 2, 5}, 0.5, 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(QuantizeTest, InvalidOutputTypeForUint8Input_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::U8>({1, 1, 2, 5}, 0.5, 0, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(QuantizeTest, InvalidInputZeroPoint_NEG)
+{
+  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S16>({1, 1, 2, 5}, 0.5, -1, input_data, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S16, 0.5, 0);
+
+  Quantize kernel(&input_tensor, &output_tensor);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/ResizeBilinear.test.cpp b/compiler/luci-interpreter/src/kernels/ResizeBilinear.test.cpp
index 7af20f8c4..933a1128c 100644
--- a/compiler/luci-interpreter/src/kernels/ResizeBilinear.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/ResizeBilinear.test.cpp
@@ -90,7 +90,7 @@ template <typename T> class ResizeBilinearTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(ResizeBilinearTest, DataTypes);
+TYPED_TEST_SUITE(ResizeBilinearTest, DataTypes);
 
 TYPED_TEST(ResizeBilinearTest, SimpleTest)
 {
diff --git a/compiler/luci-interpreter/src/kernels/ResizeNearestNeighbor.test.cpp b/compiler/luci-interpreter/src/kernels/ResizeNearestNeighbor.test.cpp
index 0e9017c78..7ade02a6f 100644
--- a/compiler/luci-interpreter/src/kernels/ResizeNearestNeighbor.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/ResizeNearestNeighbor.test.cpp
@@ -92,7 +92,7 @@ template <typename T> class ResizeNearestNeighborTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(ResizeNearestNeighborTest, DataTypes);
+TYPED_TEST_SUITE(ResizeNearestNeighborTest, DataTypes);
 
 TYPED_TEST(ResizeNearestNeighborTest, SimpleTest)
 {
diff --git a/compiler/luci-interpreter/src/kernels/ReverseV2.test.cpp b/compiler/luci-interpreter/src/kernels/ReverseV2.test.cpp
index 2bd94875b..c0025faca 100644
--- a/compiler/luci-interpreter/src/kernels/ReverseV2.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/ReverseV2.test.cpp
@@ -33,7 +33,7 @@ template <typename T> class ReverseV2Test : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(ReverseV2Test, DataTypes);
+TYPED_TEST_SUITE(ReverseV2Test, DataTypes);
 
 TYPED_TEST(ReverseV2Test, MultiDimensions)
 {
diff --git a/compiler/luci-interpreter/src/kernels/SVDF.cpp b/compiler/luci-interpreter/src/kernels/SVDF.cpp
new file mode 100644
index 000000000..40d79aaa3
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/SVDF.cpp
@@ -0,0 +1,241 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ * Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/SVDF.h"
+#include "kernels/Utils.h"
+#include "PALSVDF.h"
+
+#include <tensorflow/lite/kernels/internal/quantization_util.h>
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+namespace
+{
+TfLiteFusedActivation get_tflite_activation(Activation activation)
+{
+  switch (activation)
+  {
+    case luci::FusedActFunc::RELU:
+      return kTfLiteActRelu;
+    case luci::FusedActFunc::RELU6:
+      return kTfLiteActRelu6;
+    case luci::FusedActFunc::RELU_N1_TO_1:
+      return kTfLiteActReluN1To1;
+    case luci::FusedActFunc::TANH:
+      return kTfLiteActTanh;
+    case luci::FusedActFunc::SIGN_BIT:
+      return kTfLiteActSignBit;
+    case luci::FusedActFunc::NONE:
+      return kTfLiteActNone;
+    default:
+      throw std::runtime_error("Unsupported activation type");
+  }
+}
+} // namespace
+
+SVDF::SVDF(const Tensor *input, const Tensor *weight_feature, const Tensor *weight_time,
+           const Tensor *bias, const Tensor *input_activation_state, Tensor *output,
+           Tensor *scratchpad_activation_state, Tensor *scratchpad_1, Tensor *scratchpad_2,
+           Tensor *scratchpad_3, Tensor *scratchpad_4, Tensor *scratchpad_5, Tensor *scratchpad_6,
+           const SVDFParams &params)
+  : KernelWithParams<SVDFParams>({input, weight_feature, weight_time, bias, input_activation_state},
+                                 {output, scratchpad_activation_state, scratchpad_1, scratchpad_2,
+                                  scratchpad_3, scratchpad_4, scratchpad_5, scratchpad_6},
+                                 params)
+{
+  // Do nothing
+}
+
+void SVDF::configure()
+{
+  const Shape &input_shape = input()->shape();
+  const Shape &weight_features_shape = weight_feature()->shape();
+  const Shape &weight_time_shape = weight_time()->shape();
+
+  // Validate Input Tensor:
+  LUCI_INTERPRETER_CHECK(input()->element_type() == loco::DataType::FLOAT32 ||
+                         input()->element_type() == loco::DataType::S8);
+  LUCI_INTERPRETER_CHECK(input_shape.num_dims() == 2);
+
+  // Validate inputs and output types
+  if (input()->element_type() == loco::DataType::S8)
+  {
+    LUCI_INTERPRETER_CHECK(weight_feature()->element_type() == loco::DataType::S8);
+    LUCI_INTERPRETER_CHECK(weight_time()->element_type() == loco::DataType::S16 ||
+                           weight_time()->element_type() == loco::DataType::S8);
+    if (bias())
+      LUCI_INTERPRETER_CHECK(bias()->element_type() == loco::DataType::S32);
+
+    LUCI_INTERPRETER_CHECK(input_activation_state()->element_type() == loco::DataType::S16 ||
+                           input_activation_state()->element_type() == loco::DataType::S8);
+    LUCI_INTERPRETER_CHECK(output()->element_type() == loco::DataType::S8);
+
+    // Note: now tflite support only ReLU activation for integer SVDF
+    LUCI_INTERPRETER_CHECK(params().activation == luci::FusedActFunc::RELU);
+  }
+  else if (weight_feature()->element_type() == loco::DataType::FLOAT32)
+  {
+    LUCI_INTERPRETER_CHECK(weight_feature()->element_type() == loco::DataType::FLOAT32);
+    LUCI_INTERPRETER_CHECK(weight_time()->element_type() == loco::DataType::FLOAT32);
+    LUCI_INTERPRETER_CHECK(input_activation_state()->element_type() == loco::DataType::FLOAT32);
+    if (bias())
+      LUCI_INTERPRETER_CHECK(bias()->element_type() == loco::DataType::FLOAT32);
+    LUCI_INTERPRETER_CHECK(output()->element_type() == loco::DataType::FLOAT32);
+  }
+  else if ((weight_feature()->element_type() == loco::DataType::U8 ||
+            weight_feature()->element_type() == loco::DataType::S8) &&
+           input()->element_type() == loco::DataType::FLOAT32)
+  {
+    // TODO:: support hybrid SVDF op
+    throw std::runtime_error("Hybrid type is not currently supported");
+  }
+  else
+  {
+    throw std::runtime_error("Unsupported type.");
+  }
+
+  // Check all the parameters of tensor match within themselves and match the
+  // input configuration.
+  const int rank = params().svdf_rank;
+  const int batch_size = input_shape.dim(0);
+  const int num_filters = weight_features_shape.dim(0);
+  LUCI_INTERPRETER_CHECK(rank != 0);
+  LUCI_INTERPRETER_CHECK(num_filters % rank == 0);
+
+  const int num_units = num_filters / rank;
+  const int memory_size = weight_time_shape.dim(1);
+
+  // Validate Weight_Feature Input Tensor:
+  LUCI_INTERPRETER_CHECK(weight_features_shape.num_dims() == 2);
+  LUCI_INTERPRETER_CHECK(weight_features_shape.dim(1) == input_shape.dim(1));
+
+  // Validate Weight_Time Input Tensor:
+  LUCI_INTERPRETER_CHECK(weight_time_shape.num_dims() == 2);
+  LUCI_INTERPRETER_CHECK(weight_time_shape.dim(0) == num_filters);
+
+  // Validate Bias
+  if (bias())
+    LUCI_INTERPRETER_CHECK(bias()->shape().dim(0) == num_units);
+
+  // Validate Input Activation State
+  LUCI_INTERPRETER_CHECK(input_activation_state()->shape().num_dims() == 2);
+  LUCI_INTERPRETER_CHECK(input_activation_state()->shape().dim(0) == batch_size);
+  LUCI_INTERPRETER_CHECK(input_activation_state()->shape().dim(1) == memory_size * num_filters);
+
+  // Resize scratchpad_state to input_activation_state
+  auto scratchpad_activation_state = getOutputTensors()[1];
+  scratchpad_activation_state->resize({batch_size, memory_size * num_filters});
+
+  // Resize output tensor
+  output()->resize({batch_size, num_units});
+
+  luci_interpreter_pal::SetupScratchpadTensor(
+    input()->element_type(), weight_feature()->element_type(), getOutputTensors()[2],
+    getOutputTensors()[3], getOutputTensors()[4], getOutputTensors()[5], getOutputTensors()[6],
+    getOutputTensors()[7], input_shape, weight_time_shape, batch_size, num_filters, num_units);
+}
+
+void SVDF::execute() const
+{
+  switch (weight_feature()->element_type())
+  {
+    case loco::DataType::FLOAT32:
+      evalFloat();
+      break;
+    case loco::DataType::S8:
+    {
+      if (input()->element_type() == loco::DataType::S8)
+        evalInteger();
+      else
+        // TODO:: support hybrid SVDF op
+        throw std::runtime_error("Hybrid type is not currently supported");
+      break;
+    }
+    default:
+      throw std::runtime_error("Unsupported type");
+  }
+}
+
+void SVDF::evalInteger() const
+{
+  const auto effective_scale_1 = static_cast<double>(input()->scale() * weight_feature()->scale() /
+                                                     input_activation_state()->scale());
+  const auto effective_scale_2 = static_cast<double>(input_activation_state()->scale() *
+                                                     weight_time()->scale() / output()->scale());
+
+  int32_t effective_scale_1_a;
+  int effective_scale_1_b;
+  int32_t effective_scale_2_a;
+  int effective_scale_2_b;
+
+  tflite::QuantizeMultiplier(effective_scale_1, &effective_scale_1_a, &effective_scale_1_b);
+  tflite::QuantizeMultiplier(effective_scale_2, &effective_scale_2_a, &effective_scale_2_b);
+
+  TfLiteSVDFParams params_svdf{};
+  params_svdf.asymmetric_quantize_inputs = params().asymmetric_quantize_inputs;
+  params_svdf.rank = params().svdf_rank;
+  params_svdf.activation = get_tflite_activation(params().activation);
+
+  auto scratchpad_activation_state = getOutputTensors()[1];
+  // Note: it is expected that activation_state input variable tensor reset to zero,
+  // also expected that this variable tensor doesn't have buffer
+  auto scratchpad_data = getTensorData<int16_t>(scratchpad_activation_state);
+  std::fill_n(scratchpad_data, scratchpad_activation_state->shape().num_elements(), 0);
+
+  auto scratchpad = getOutputTensors()[2];
+  auto output_temp = getOutputTensors()[3];
+
+  int32_t input_zp = input()->zero_point();
+  int32_t output_zp = output()->zero_point();
+  luci_interpreter_pal::IntegerSVDF(
+    params_svdf, getTensorShape(input()), getTensorData<int8_t>(input()),
+    getTensorShape(weight_feature()), getTensorData<int8_t>(weight_feature()),
+    getTensorShape(weight_time()), getTensorData<int16_t>(weight_time()), getTensorShape(bias()),
+    getTensorData<int32_t>(bias()), scratchpad_data, getTensorShape(output()),
+    getTensorData<int8_t>(output()), getTensorData<int32_t>(scratchpad),
+    getTensorData<int32_t>(output_temp), effective_scale_1_a, effective_scale_1_b,
+    effective_scale_2_a, effective_scale_2_b, input_zp, output_zp);
+}
+
+void SVDF::evalFloat() const
+{
+  TfLiteSVDFParams params_svdf{};
+  params_svdf.asymmetric_quantize_inputs = params().asymmetric_quantize_inputs;
+  params_svdf.rank = params().svdf_rank;
+  params_svdf.activation = get_tflite_activation(params().activation);
+
+  auto scratchpad_activation_state = getOutputTensors()[1];
+  // Note: it is expected that activation_state input variable tensor reset to zero,
+  // also expected that this variable tensor doesn't have buffer
+  auto scratchpad_data = getTensorData<float>(scratchpad_activation_state);
+  std::fill_n(scratchpad_data, scratchpad_activation_state->shape().num_elements(), 0);
+
+  auto scratchpad_1 = getOutputTensors()[2];
+
+  luci_interpreter_pal::FloatSVDF(
+    params_svdf, getTensorShape(input()), getTensorData<float>(input()),
+    getTensorShape(weight_feature()), getTensorData<float>(weight_feature()),
+    getTensorShape(weight_time()), getTensorData<float>(weight_time()), getTensorShape(bias()),
+    getTensorData<float>(bias()), getTensorData<float>(scratchpad_1), scratchpad_data,
+    getTensorShape(output()), getTensorData<float>(output()));
+}
+
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/SVDF.h b/compiler/luci-interpreter/src/kernels/SVDF.h
new file mode 100644
index 000000000..335a6cd8f
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/SVDF.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LUCI_INTERPRETER_KERNELS_SVDF_H
+#define LUCI_INTERPRETER_KERNELS_SVDF_H
+
+#include "core/Kernel.h"
+#include "core/KernelParams.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+
+class SVDF : public KernelWithParams<SVDFParams>
+{
+public:
+  SVDF(const Tensor *input, const Tensor *weight_feature, const Tensor *weight_time,
+       const Tensor *bias, const Tensor *input_activation_state, Tensor *output,
+       Tensor *scratchpad_activation_state, Tensor *scratchpad_1, Tensor *scratchpad_2,
+       Tensor *scratchpad_3, Tensor *scratchpad_4, Tensor *scratchpad_5, Tensor *scratchpad_6,
+       const SVDFParams &params);
+
+  const Tensor *input() const { return _inputs[0]; }
+  const Tensor *weight_feature() const { return _inputs[1]; }
+  const Tensor *weight_time() const { return _inputs[2]; }
+  const Tensor *bias() const { return _inputs[3]; }
+  const Tensor *input_activation_state() const { return _inputs[4]; }
+
+  Tensor *output() const { return _outputs[0]; }
+
+  void configure() override;
+  void execute() const override;
+
+private:
+  void evalFloat() const;
+  void evalInteger() const;
+};
+
+} // namespace kernels
+} // namespace luci_interpreter
+
+#endif // LUCI_INTERPRETER_KERNELS_SVDF_H
diff --git a/compiler/luci-interpreter/src/kernels/SVDF.test.cpp b/compiler/luci-interpreter/src/kernels/SVDF.test.cpp
new file mode 100644
index 000000000..82bd9b009
--- /dev/null
+++ b/compiler/luci-interpreter/src/kernels/SVDF.test.cpp
@@ -0,0 +1,341 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernels/SVDF.h"
+#include "kernels/TestUtils.h"
+#include "luci_interpreter/TestMemoryManager.h"
+
+namespace luci_interpreter
+{
+namespace kernels
+{
+namespace
+{
+
+using namespace testing;
+
+class SVDFTest : public ::testing::Test
+{
+protected:
+  void SetUp() override { _memory_manager = std::make_unique<TestMemoryManager>(); }
+
+  std::unique_ptr<IMemoryManager> _memory_manager;
+};
+
+TEST_F(SVDFTest, FullIntegerTest)
+{
+  const int32_t batches = 2;
+  const int32_t input_size = 3;
+  const int32_t units = 4;
+  const int32_t memory_size = 10;
+  const int32_t rank = 1;
+  const int32_t num_filters = units * rank;
+
+  Shape input_shape{batches, input_size};
+  Shape weight_feature_shape{num_filters, input_size};
+  Shape weight_time_shape{num_filters, memory_size};
+  Shape bias_shape{units};
+  Shape activation_state_shape{batches, memory_size * num_filters};
+
+  std::vector<float> input_data{0.49837467, 0.19278903, 0.26584083,
+                                0.17660543, 0.52949083, -0.77931279};
+
+  std::vector<float> weight_feature_data{-0.31930989, -0.36118156, 0.0079667,   0.37613347,
+                                         0.22197971,  0.12416199,  0.27901134,  0.27557442,
+                                         0.3905206,   -0.36137494, -0.06634006, -0.10640851};
+
+  std::vector<float> weight_time_data{
+    -0.31930989, 0.37613347,  0.27901134,  -0.36137494, -0.36118156,
+    0.22197971,  0.27557442,  -0.06634006, 0.0079667,   0.12416199,
+
+    0.3905206,   -0.10640851, -0.0976817,  0.15294972,  0.39635518,
+    -0.02702999, 0.39296314,  0.15785322,  0.21931258,  0.31053296,
+
+    -0.36916667, 0.38031587,  -0.21580373, 0.27072677,  0.23622236,
+    0.34936687,  0.18174365,  0.35907319,  -0.17493086, 0.324846,
+
+    -0.10781813, 0.27201805,  0.14324132,  -0.23681851, -0.27115166,
+    -0.01580888, -0.14943552, 0.15465137,  0.09784451,  -0.0337657};
+
+  std::vector<float> bias_data{-0.0976817, 0.15294972, 0.39635518, -0.02702999};
+
+  std::pair<float, int32_t> input_quant_param = quantizationParams<int8_t>(-1, 1);
+  std::pair<float, int32_t> weight_feature_quant_param = quantizationParams<int8_t>(-0.5, 0.5);
+  std::pair<float, int32_t> weight_time_quant_param = quantizationParams<int16_t>(-1, 1);
+  std::pair<float, int32_t> bias_quant_param = quantizationParams<int32_t>(-512, 512);
+  std::pair<float, int32_t> activation_state_quant_param = quantizationParams<int16_t>(-16, 16);
+
+  std::pair<float, int32_t> output_quant_param = quantizationParams<int8_t>(-0.5, 0.5);
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S8>(input_shape, input_quant_param.first, input_quant_param.second,
+                                  input_data, _memory_manager.get());
+  Tensor weight_feature_tensor = makeInputTensor<DataType::S8>(
+    weight_feature_shape, weight_feature_quant_param.first, weight_feature_quant_param.second,
+    weight_feature_data, _memory_manager.get());
+  Tensor weight_time_tensor = makeInputTensor<DataType::S16>(
+    weight_time_shape, weight_time_quant_param.first, weight_time_quant_param.second,
+    weight_time_data, _memory_manager.get());
+  Tensor bias_tensor = makeInputTensor<DataType::S32>(
+    bias_shape, bias_quant_param.first, bias_quant_param.second, bias_data, _memory_manager.get());
+  Tensor activation_state_tensor = makeOutputTensor(
+    DataType::S16, activation_state_quant_param.first, activation_state_quant_param.second);
+  activation_state_tensor.resize(activation_state_shape);
+  Tensor output_tensor =
+    makeOutputTensor(DataType::S8, output_quant_param.first, output_quant_param.second);
+
+  Tensor scratchpad_activation_state(DataType::S16, Shape({}), {}, "");
+  Tensor scratchpad_1(DataType::S32, Shape({}), {}, "");
+  Tensor scratchpad_2(DataType::S32, Shape({}), {}, "");
+  Tensor scratchpad_3(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_4(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_5(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_6(DataType::FLOAT32, Shape({}), {}, "");
+
+  SVDFParams params{};
+  params.activation = Activation::RELU;
+  params.asymmetric_quantize_inputs = false;
+  params.svdf_rank = rank;
+
+  SVDF kernel(&input_tensor, &weight_feature_tensor, &weight_time_tensor, &bias_tensor,
+              &activation_state_tensor, &output_tensor, &scratchpad_activation_state, &scratchpad_1,
+              &scratchpad_2, &scratchpad_3, &scratchpad_4, &scratchpad_5, &scratchpad_6, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad_activation_state);
+  _memory_manager->allocate_memory(scratchpad_1);
+  _memory_manager->allocate_memory(scratchpad_2);
+  _memory_manager->allocate_memory(scratchpad_3);
+  _memory_manager->allocate_memory(scratchpad_4);
+  _memory_manager->allocate_memory(scratchpad_5);
+  _memory_manager->allocate_memory(scratchpad_6);
+  kernel.execute();
+
+  std::vector<int8_t> ref_output_data{-9, 24, 31, 1, -10, 10, -3, 0};
+
+  std::vector<int32_t> ref_output_shape{batches, units};
+  EXPECT_THAT(extractTensorData<int8_t>(output_tensor), ref_output_data);
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(SVDFTest, FloatTest)
+{
+  const int32_t batches = 2;
+  const int32_t input_size = 3;
+  const int32_t units = 4;
+  const int32_t memory_size = 10;
+  const int32_t rank = 1;
+  const int32_t num_filters = units * rank;
+
+  Shape input_shape{batches, input_size};
+  Shape weight_feature_shape{num_filters, input_size};
+  Shape weight_time_shape{num_filters, memory_size};
+  Shape activation_state_shape{batches, memory_size * num_filters};
+
+  std::vector<float> input_data{0.12609188, -0.46347019, -0.89598465,
+                                0.35867718, 0.36897406,  0.73463392};
+
+  std::vector<float> weight_feature_data{-0.31930989, -0.36118156, 0.0079667,   0.37613347,
+                                         0.22197971,  0.12416199,  0.27901134,  0.27557442,
+                                         0.3905206,   -0.36137494, -0.06634006, -0.10640851};
+
+  std::vector<float> weight_time_data{
+    -0.31930989, 0.37613347,  0.27901134,  -0.36137494, -0.36118156,
+    0.22197971,  0.27557442,  -0.06634006, 0.0079667,   0.12416199,
+
+    0.3905206,   -0.10640851, -0.0976817,  0.15294972,  0.39635518,
+    -0.02702999, 0.39296314,  0.15785322,  0.21931258,  0.31053296,
+
+    -0.36916667, 0.38031587,  -0.21580373, 0.27072677,  0.23622236,
+    0.34936687,  0.18174365,  0.35907319,  -0.17493086, 0.324846,
+
+    -0.10781813, 0.27201805,  0.14324132,  -0.23681851, -0.27115166,
+    -0.01580888, -0.14943552, 0.15465137,  0.09784451,  -0.0337657};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
+  Tensor weight_feature_tensor = makeInputTensor<DataType::FLOAT32>(
+    weight_feature_shape, weight_feature_data, _memory_manager.get());
+  Tensor weight_time_tensor =
+    makeInputTensor<DataType::FLOAT32>(weight_time_shape, weight_time_data, _memory_manager.get());
+  Tensor activation_state_tensor = makeOutputTensor(DataType::FLOAT32);
+  activation_state_tensor.resize(activation_state_shape);
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Tensor scratchpad_activation_state(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_1(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_2(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_3(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_4(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_5(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_6(DataType::FLOAT32, Shape({}), {}, "");
+
+  SVDFParams params{};
+  params.activation = Activation::NONE;
+  params.asymmetric_quantize_inputs = false;
+  params.svdf_rank = rank;
+
+  SVDF kernel(&input_tensor, &weight_feature_tensor, &weight_time_tensor, nullptr,
+              &activation_state_tensor, &output_tensor, &scratchpad_activation_state, &scratchpad_1,
+              &scratchpad_2, &scratchpad_3, &scratchpad_4, &scratchpad_5, &scratchpad_6, params);
+  kernel.configure();
+  _memory_manager->allocate_memory(output_tensor);
+  _memory_manager->allocate_memory(scratchpad_activation_state);
+  _memory_manager->allocate_memory(scratchpad_1);
+  _memory_manager->allocate_memory(scratchpad_2);
+  _memory_manager->allocate_memory(scratchpad_3);
+  _memory_manager->allocate_memory(scratchpad_4);
+  _memory_manager->allocate_memory(scratchpad_5);
+  _memory_manager->allocate_memory(scratchpad_6);
+  kernel.execute();
+
+  std::vector<float> ref_output_data{0.014899,    -0.0517661, -0.143725, -0.00271883,
+                                     -0.03004015, 0.09565311, 0.1587342, 0.00784263};
+
+  std::vector<float> ref_output_shape{batches, units};
+  const float tolerance = 1e-5;
+  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(ref_output_data, tolerance));
+  EXPECT_THAT(extractTensorShape(output_tensor), ::testing::ElementsAreArray(ref_output_shape));
+}
+
+TEST_F(SVDFTest, Unsupported_Type_Configure_NEG)
+{
+  const int32_t batches = 2;
+  const int32_t input_size = 3;
+  const int32_t units = 4;
+  const int32_t memory_size = 10;
+  const int32_t rank = 1;
+  const int32_t num_filters = units * rank;
+
+  Shape input_shape{batches, input_size};
+  Shape weight_feature_shape{num_filters, input_size};
+  Shape weight_time_shape{num_filters, memory_size};
+  Shape activation_state_shape{batches, memory_size * num_filters};
+
+  std::vector<int32_t> input_data{0, 1, 3, 4, 4, -2};
+
+  std::vector<float> weight_feature_data{-0.31930989, -0.36118156, 0.0079667,   0.37613347,
+                                         0.22197971,  0.12416199,  0.27901134,  0.27557442,
+                                         0.3905206,   -0.36137494, -0.06634006, -0.10640851};
+
+  std::vector<float> weight_time_data{
+    -0.31930989, 0.37613347,  0.27901134,  -0.36137494, -0.36118156,
+    0.22197971,  0.27557442,  -0.06634006, 0.0079667,   0.12416199,
+
+    0.3905206,   -0.10640851, -0.0976817,  0.15294972,  0.39635518,
+    -0.02702999, 0.39296314,  0.15785322,  0.21931258,  0.31053296,
+
+    -0.36916667, 0.38031587,  -0.21580373, 0.27072677,  0.23622236,
+    0.34936687,  0.18174365,  0.35907319,  -0.17493086, 0.324846,
+
+    -0.10781813, 0.27201805,  0.14324132,  -0.23681851, -0.27115166,
+    -0.01580888, -0.14943552, 0.15465137,  0.09784451,  -0.0337657};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::S32>(input_shape, input_data, _memory_manager.get());
+  Tensor weight_feature_tensor = makeInputTensor<DataType::FLOAT32>(
+    weight_feature_shape, weight_feature_data, _memory_manager.get());
+  Tensor weight_time_tensor =
+    makeInputTensor<DataType::FLOAT32>(weight_time_shape, weight_time_data, _memory_manager.get());
+  Tensor activation_state_tensor = makeOutputTensor(DataType::FLOAT32);
+  activation_state_tensor.resize(activation_state_shape);
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Tensor scratchpad_activation_state(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_1(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_2(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_3(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_4(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_5(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_6(DataType::FLOAT32, Shape({}), {}, "");
+
+  SVDFParams params{};
+  params.activation = Activation::NONE;
+  params.asymmetric_quantize_inputs = false;
+  params.svdf_rank = rank;
+
+  SVDF kernel(&input_tensor, &weight_feature_tensor, &weight_time_tensor, nullptr,
+              &activation_state_tensor, &output_tensor, &scratchpad_activation_state, &scratchpad_1,
+              &scratchpad_2, &scratchpad_3, &scratchpad_4, &scratchpad_5, &scratchpad_6, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(SVDFTest, Invalid_Input_Shape_NEG)
+{
+  const int32_t batches = 2;
+  const int32_t right_input_size = 3;
+  const int32_t wrong_input_size = 4;
+  const int32_t units = 4;
+  const int32_t memory_size = 10;
+  const int32_t rank = 1;
+  const int32_t num_filters = units * rank;
+
+  Shape input_shape{batches, wrong_input_size};
+  Shape weight_feature_shape{num_filters, right_input_size};
+  Shape weight_time_shape{num_filters, memory_size};
+  Shape activation_state_shape{batches, memory_size * num_filters};
+
+  std::vector<float> input_data{0, 1, 3, 2, 4, 4, -2, 1};
+
+  std::vector<float> weight_feature_data{-0.31930989, -0.36118156, 0.0079667,   0.37613347,
+                                         0.22197971,  0.12416199,  0.27901134,  0.27557442,
+                                         0.3905206,   -0.36137494, -0.06634006, -0.10640851};
+
+  std::vector<float> weight_time_data{
+    -0.31930989, 0.37613347,  0.27901134,  -0.36137494, -0.36118156,
+    0.22197971,  0.27557442,  -0.06634006, 0.0079667,   0.12416199,
+
+    0.3905206,   -0.10640851, -0.0976817,  0.15294972,  0.39635518,
+    -0.02702999, 0.39296314,  0.15785322,  0.21931258,  0.31053296,
+
+    -0.36916667, 0.38031587,  -0.21580373, 0.27072677,  0.23622236,
+    0.34936687,  0.18174365,  0.35907319,  -0.17493086, 0.324846,
+
+    -0.10781813, 0.27201805,  0.14324132,  -0.23681851, -0.27115166,
+    -0.01580888, -0.14943552, 0.15465137,  0.09784451,  -0.0337657};
+
+  Tensor input_tensor =
+    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
+  Tensor weight_feature_tensor = makeInputTensor<DataType::FLOAT32>(
+    weight_feature_shape, weight_feature_data, _memory_manager.get());
+  Tensor weight_time_tensor =
+    makeInputTensor<DataType::FLOAT32>(weight_time_shape, weight_time_data, _memory_manager.get());
+  Tensor activation_state_tensor = makeOutputTensor(DataType::FLOAT32);
+  activation_state_tensor.resize(activation_state_shape);
+  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);
+
+  Tensor scratchpad_activation_state(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_1(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_2(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_3(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_4(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_5(DataType::FLOAT32, Shape({}), {}, "");
+  Tensor scratchpad_6(DataType::FLOAT32, Shape({}), {}, "");
+
+  SVDFParams params{};
+  params.activation = Activation::NONE;
+  params.asymmetric_quantize_inputs = false;
+  params.svdf_rank = rank;
+
+  SVDF kernel(&input_tensor, &weight_feature_tensor, &weight_time_tensor, nullptr,
+              &activation_state_tensor, &output_tensor, &scratchpad_activation_state, &scratchpad_1,
+              &scratchpad_2, &scratchpad_3, &scratchpad_4, &scratchpad_5, &scratchpad_6, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+} // namespace
+} // namespace kernels
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Slice.cpp b/compiler/luci-interpreter/src/kernels/Slice.cpp
index 37a834a18..2fe2c5471 100644
--- a/compiler/luci-interpreter/src/kernels/Slice.cpp
+++ b/compiler/luci-interpreter/src/kernels/Slice.cpp
@@ -139,6 +139,11 @@ void Slice::execute() const
                                   getTensorData<uint8_t>(input()), getTensorShape(output()),
                                   getTensorData<uint8_t>(output()));
       break;
+    case DataType::S8:
+      luci_interpreter_pal::Slice(op_params, getTensorShape(input()),
+                                  getTensorData<int8_t>(input()), getTensorShape(output()),
+                                  getTensorData<int8_t>(output()));
+      break;
     default:
       throw std::runtime_error("Unsupported input type.");
   }
diff --git a/compiler/luci-interpreter/src/kernels/Slice.test.cpp b/compiler/luci-interpreter/src/kernels/Slice.test.cpp
index 3e0d0b0d7..517982990 100644
--- a/compiler/luci-interpreter/src/kernels/Slice.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Slice.test.cpp
@@ -31,8 +31,8 @@ template <typename T> class SliceTest : public ::testing::Test
 {
 };
 
-using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(SliceTest, DataTypes);
+using DataTypes = ::testing::Types<float, uint8_t, int8_t>;
+TYPED_TEST_SUITE(SliceTest, DataTypes);
 
 TYPED_TEST(SliceTest, SimpleTest)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Softmax.test.cpp b/compiler/luci-interpreter/src/kernels/Softmax.test.cpp
index 9de40b6ec..08e70672d 100644
--- a/compiler/luci-interpreter/src/kernels/Softmax.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Softmax.test.cpp
@@ -93,7 +93,7 @@ template <typename T> class SoftmaxTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t, int8_t>;
-TYPED_TEST_CASE(SoftmaxTest, DataTypes);
+TYPED_TEST_SUITE(SoftmaxTest, DataTypes);
 
 TYPED_TEST(SoftmaxTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/SpaceToBatchND.test.cpp b/compiler/luci-interpreter/src/kernels/SpaceToBatchND.test.cpp
index e06501c8c..3a8b0a812 100644
--- a/compiler/luci-interpreter/src/kernels/SpaceToBatchND.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/SpaceToBatchND.test.cpp
@@ -90,7 +90,7 @@ template <typename T> class SpaceToBatchNDTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(SpaceToBatchNDTest, DataTypes);
+TYPED_TEST_SUITE(SpaceToBatchNDTest, DataTypes);
 
 TYPED_TEST(SpaceToBatchNDTest, Simple)
 {
diff --git a/compiler/luci-interpreter/src/kernels/SpaceToDepth.test.cpp b/compiler/luci-interpreter/src/kernels/SpaceToDepth.test.cpp
index 735c010b9..4af488618 100644
--- a/compiler/luci-interpreter/src/kernels/SpaceToDepth.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/SpaceToDepth.test.cpp
@@ -32,7 +32,7 @@ template <typename T> class SpaceToDepthTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(SpaceToDepthTest, DataTypes);
+TYPED_TEST_SUITE(SpaceToDepthTest, DataTypes);
 
 TYPED_TEST(SpaceToDepthTest, SimpleCase)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Split.test.cpp b/compiler/luci-interpreter/src/kernels/Split.test.cpp
index 74d57aed3..283cd9aa9 100644
--- a/compiler/luci-interpreter/src/kernels/Split.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Split.test.cpp
@@ -73,7 +73,7 @@ template <typename T> class SplitTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(SplitTest, DataTypes);
+TYPED_TEST_SUITE(SplitTest, DataTypes);
 
 TYPED_TEST(SplitTest, FourDimensional)
 {
diff --git a/compiler/luci-interpreter/src/kernels/SplitV.test.cpp b/compiler/luci-interpreter/src/kernels/SplitV.test.cpp
index aac0567d7..035bc2122 100644
--- a/compiler/luci-interpreter/src/kernels/SplitV.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/SplitV.test.cpp
@@ -77,7 +77,7 @@ template <typename T> class SplitVTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t, int16_t>;
-TYPED_TEST_CASE(SplitVTest, DataTypes);
+TYPED_TEST_SUITE(SplitVTest, DataTypes);
 
 TYPED_TEST(SplitVTest, ThreeDimensional)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Squeeze.test.cpp b/compiler/luci-interpreter/src/kernels/Squeeze.test.cpp
index d3326fe98..1bc0b6459 100644
--- a/compiler/luci-interpreter/src/kernels/Squeeze.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Squeeze.test.cpp
@@ -56,7 +56,7 @@ template <typename T> class SqueezeTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(SqueezeTest, DataTypes);
+TYPED_TEST_SUITE(SqueezeTest, DataTypes);
 
 TYPED_TEST(SqueezeTest, TotalTest)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Sub.cpp b/compiler/luci-interpreter/src/kernels/Sub.cpp
index 603c62d0f..24b6a72e5 100644
--- a/compiler/luci-interpreter/src/kernels/Sub.cpp
+++ b/compiler/luci-interpreter/src/kernels/Sub.cpp
@@ -37,6 +37,7 @@ Sub::Sub(const Tensor *input1, const Tensor *input2, Tensor *output, const SubPa
 void Sub::configure()
 {
   LUCI_INTERPRETER_CHECK(!(input1()->element_type() != input2()->element_type()))
+  LUCI_INTERPRETER_CHECK(!(input1()->element_type() != output()->element_type()))
   output()->resize(calculateShapeForBroadcast(input1()->shape(), input2()->shape()));
 }
 
@@ -47,6 +48,12 @@ void Sub::execute() const
     case DataType::FLOAT32:
       evalFloat();
       break;
+    case DataType::S64:
+      evalInteger<int64_t>();
+      break;
+    case DataType::S32:
+      evalInteger<int32_t>();
+      break;
     case DataType::U8:
       evalQuantized();
       break;
@@ -57,13 +64,8 @@ void Sub::execute() const
 
 void Sub::evalFloat() const
 {
-  float activation_min{};
-  float activation_max{};
-  calculateActivationRange(_params.activation, &activation_min, &activation_max);
-
   tflite::ArithmeticParams params{};
-  params.float_activation_min = activation_min;
-  params.float_activation_max = activation_max;
+  fillArithmeticActivationRange<float>(params, _params.activation);
 
   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
     getTensorShape(input1()), getTensorShape(input2()), &params);
@@ -82,6 +84,28 @@ void Sub::evalFloat() const
   }
 }
 
+template <typename T> void Sub::evalInteger() const
+{
+  tflite::ArithmeticParams params{};
+  fillArithmeticActivationRange<T>(params, _params.activation);
+
+  const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
+    getTensorShape(input1()), getTensorShape(input2()), &params);
+
+  if (need_broadcast)
+  {
+    tflite::reference_ops::BroadcastSubSlow(
+      params, getTensorShape(input1()), getTensorData<T>(input1()), getTensorShape(input2()),
+      getTensorData<T>(input2()), getTensorShape(output()), getTensorData<T>(output()));
+  }
+  else
+  {
+    tflite::reference_ops::Sub(params, getTensorShape(input1()), getTensorData<T>(input1()),
+                               getTensorShape(input2()), getTensorData<T>(input2()),
+                               getTensorShape(output()), getTensorData<T>(output()));
+  }
+}
+
 void Sub::evalQuantized() const
 {
   const auto input1_scale = static_cast<double>(input1()->scale());
diff --git a/compiler/luci-interpreter/src/kernels/Sub.h b/compiler/luci-interpreter/src/kernels/Sub.h
index d7940b5c6..23952b3bd 100644
--- a/compiler/luci-interpreter/src/kernels/Sub.h
+++ b/compiler/luci-interpreter/src/kernels/Sub.h
@@ -39,6 +39,7 @@ public:
 
 private:
   void evalFloat() const;
+  template <typename T> void evalInteger() const;
   void evalQuantized() const;
 };
 
diff --git a/compiler/luci-interpreter/src/kernels/Sub.test.cpp b/compiler/luci-interpreter/src/kernels/Sub.test.cpp
index c189f4481..9abafd49a 100644
--- a/compiler/luci-interpreter/src/kernels/Sub.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Sub.test.cpp
@@ -162,6 +162,51 @@ TEST_F(SubTest, Float)
   }
 }
 
+template <loco::DataType DType> void CheckInteger(luci_interpreter::IMemoryManager *memory_manager)
+{
+  using dtype = typename loco::DataTypeImpl<DType>::Type;
+  Shape base_shape = {2, 3, 1, 2};
+  std::vector<Shape> test_shapes{{1, 1, 3, 2}, {1, 3, 1, 2}, {2, 1, 3, 1}, {2, 3, 1, 1}};
+  std::vector<std::vector<dtype>> test_outputs = {
+    {0, 1, 2, 3, 0, 0, 0, 0, 4,  1, 0, 0, 0, 0, 7,  0, 3, 0,
+     0, 2, 4, 4, 0, 0, 3, 0, 10, 0, 6, 0, 3, 0, 10, 2, 6, 0},
+    {0, 1, 4, 1, 3, 0, 0, 2, 10, 0, 6, 0},
+    {0, 0, 0, 1, 2, 5, 0, 0, 0, 0, 4, 3, 0, 0, 3, 0, 7, 0,
+     2, 4, 0, 2, 0, 0, 8, 0, 6, 0, 1, 0, 8, 2, 6, 0, 1, 0},
+    {0, 0, 0, 0, 7, 0, 2, 4, 6, 0, 1, 0}};
+  std::vector<dtype> input1_data{-1, 2, 1, 0, 4, -5, 1, 3, 7, -1, 7, 1};
+  std::vector<dtype> input2_data{4, 1, -3, -1, 1, 6};
+  for (size_t i = 0; i < test_shapes.size(); ++i)
+  {
+    Tensor input1_tensor = makeInputTensor<DType>(base_shape, input1_data, memory_manager);
+    Tensor input2_tensor = makeInputTensor<DType>(test_shapes[i], input2_data, memory_manager);
+    Tensor output_tensor = makeOutputTensor(DType);
+
+    SubParams params{};
+    params.activation = Activation::RELU;
+
+    Sub kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+    kernel.configure();
+    memory_manager->allocate_memory(output_tensor);
+    kernel.execute();
+
+    EXPECT_THAT(extractTensorData<dtype>(output_tensor), test_outputs[i])
+      << "With shape number " << i;
+  }
+};
+
+TEST_F(SubTest, SInt32)
+{
+  CheckInteger<loco::DataType::S32>(_memory_manager.get());
+  SUCCEED();
+}
+
+TEST_F(SubTest, SInt64)
+{
+  CheckInteger<loco::DataType::S64>(_memory_manager.get());
+  SUCCEED();
+}
+
 TEST_F(SubTest, Input_Output_Type_NEG)
 {
   Tensor input1_tensor = makeInputTensor<DataType::FLOAT32>({1}, {1.f}, _memory_manager.get());
@@ -175,11 +220,24 @@ TEST_F(SubTest, Input_Output_Type_NEG)
   EXPECT_ANY_THROW(kernel.configure());
 }
 
-TEST_F(SubTest, Invalid_Input_Type_NEG)
+TEST_F(SubTest, Invalid_Output_Type_NEG)
 {
   Tensor input1_tensor = makeInputTensor<DataType::S64>({1}, {1}, _memory_manager.get());
   Tensor input2_tensor = makeInputTensor<DataType::S64>({1}, {2}, _memory_manager.get());
-  Tensor output_tensor = makeOutputTensor(DataType::S64);
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  SubParams params{};
+  params.activation = Activation::RELU;
+
+  Sub kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
+TEST_F(SubTest, Invalid_Input_Type_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::U64>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::U64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::U64);
 
   SubParams params{};
   params.activation = Activation::RELU;
@@ -190,6 +248,19 @@ TEST_F(SubTest, Invalid_Input_Type_NEG)
   EXPECT_ANY_THROW(kernel.execute());
 }
 
+TEST_F(SubTest, Mismatching_Input_Int_Types_NEG)
+{
+  Tensor input1_tensor = makeInputTensor<DataType::S32>({1}, {1}, _memory_manager.get());
+  Tensor input2_tensor = makeInputTensor<DataType::S64>({1}, {2}, _memory_manager.get());
+  Tensor output_tensor = makeOutputTensor(DataType::S32);
+
+  SubParams params{};
+  params.activation = Activation::NONE;
+
+  Sub kernel(&input1_tensor, &input2_tensor, &output_tensor, params);
+  EXPECT_ANY_THROW(kernel.configure());
+}
+
 } // namespace
 } // namespace kernels
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/kernels/Transpose.test.cpp b/compiler/luci-interpreter/src/kernels/Transpose.test.cpp
index 107179910..43be8f8b9 100644
--- a/compiler/luci-interpreter/src/kernels/Transpose.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Transpose.test.cpp
@@ -52,7 +52,7 @@ template <typename T> class TransposeTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(TransposeTest, DataTypes);
+TYPED_TEST_SUITE(TransposeTest, DataTypes);
 
 TYPED_TEST(TransposeTest, Small3D)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Unpack.test.cpp b/compiler/luci-interpreter/src/kernels/Unpack.test.cpp
index 4f22c9f30..9384ddc83 100644
--- a/compiler/luci-interpreter/src/kernels/Unpack.test.cpp
+++ b/compiler/luci-interpreter/src/kernels/Unpack.test.cpp
@@ -75,7 +75,7 @@ template <typename T> class UnpackTest : public ::testing::Test
 };
 
 using DataTypes = ::testing::Types<float, uint8_t>;
-TYPED_TEST_CASE(UnpackTest, DataTypes);
+TYPED_TEST_SUITE(UnpackTest, DataTypes);
 
 TYPED_TEST(UnpackTest, ThreeOutputs)
 {
diff --git a/compiler/luci-interpreter/src/kernels/Utils.cpp b/compiler/luci-interpreter/src/kernels/Utils.cpp
index 586cfa1e1..5d8e5db83 100644
--- a/compiler/luci-interpreter/src/kernels/Utils.cpp
+++ b/compiler/luci-interpreter/src/kernels/Utils.cpp
@@ -27,17 +27,18 @@ namespace luci_interpreter
 namespace kernels
 {
 
-void calculateActivationRange(Activation activation, float *activation_min, float *activation_max)
+template <typename T>
+void calculateActivationRange(Activation activation, T *activation_min, T *activation_max)
 {
   switch (activation)
   {
     case Activation::NONE:
-      *activation_min = std::numeric_limits<float>::lowest();
-      *activation_max = std::numeric_limits<float>::max();
+      *activation_min = std::numeric_limits<T>::lowest();
+      *activation_max = std::numeric_limits<T>::max();
       break;
     case Activation::RELU:
       *activation_min = 0;
-      *activation_max = std::numeric_limits<float>::max();
+      *activation_max = std::numeric_limits<T>::max();
       break;
     case Activation::RELU_N1_TO_1:
       *activation_min = -1;
@@ -52,6 +53,13 @@ void calculateActivationRange(Activation activation, float *activation_min, floa
   }
 }
 
+template void calculateActivationRange(Activation activation, float *activation_min,
+                                       float *activation_max);
+template void calculateActivationRange(Activation activation, int32_t *activation_min,
+                                       int32_t *activation_max);
+template void calculateActivationRange(Activation activation, int64_t *activation_min,
+                                       int64_t *activation_max);
+
 static void calculateActivationRangeQuantizedImpl(Activation activation, int32_t qmin, int32_t qmax,
                                                   const Tensor *output, int32_t *activation_min,
                                                   int32_t *activation_max)
@@ -175,7 +183,11 @@ Shape calculateShapeForBroadcast(const Shape &input1_shape, const Shape &input2_
   {
     const int32_t input1_dim = i < num_input1_dims ? input1_shape.dim(num_input1_dims - i - 1) : 1;
     const int32_t input2_dim = i < num_input2_dims ? input2_shape.dim(num_input2_dims - i - 1) : 1;
-    assert(input1_dim == input2_dim || input1_dim == 1 || input2_dim == 1);
+
+    bool need_broadcast = input1_dim != input2_dim;
+    bool can_broadcast = input1_dim == 1 || input2_dim == 1;
+    LUCI_INTERPRETER_CHECK(!need_broadcast || can_broadcast);
+
     output_shape.dim(num_out_dims - i - 1) = std::max(input1_dim, input2_dim);
   }
 
diff --git a/compiler/luci-interpreter/src/kernels/Utils.h b/compiler/luci-interpreter/src/kernels/Utils.h
index 817a42f83..ebeb20e66 100644
--- a/compiler/luci-interpreter/src/kernels/Utils.h
+++ b/compiler/luci-interpreter/src/kernels/Utils.h
@@ -76,11 +76,42 @@ inline int32_t calcOffset(const Shape &shape, int32_t d0, int32_t d1, int32_t d2
   return ((d0 * shape.dim(1) + d1) * shape.dim(2) + d2) * shape.dim(3) + d3;
 }
 
-void calculateActivationRange(Activation activation, float *activation_min, float *activation_max);
+template <typename T>
+void calculateActivationRange(Activation activation, T *activation_min, T *activation_max);
 
 void calculateActivationRangeQuantized(Activation activation, const Tensor *output,
                                        int32_t *activation_min, int32_t *activation_max);
 
+template <typename T> constexpr bool one_of_types() { return false; }
+
+// Checks if T is equal to one of {U,Other} types
+template <typename T, typename U, typename... Other> constexpr bool one_of_types()
+{
+  return std::is_same<T, U>::value || one_of_types<T, Other...>();
+}
+
+/**
+ * Fills activation min and max parameters depending on given data type and activation
+ *
+ * T is a template parameter, so after optimization this code left with only required if case
+ *
+ * @tparam T data type of arithmetic operation output tensor
+ * @param params tflite params to fill
+ * @param activation luci_interpreter::Activation of arithmetic operation
+ */
+template <typename T>
+void fillArithmeticActivationRange(tflite::ArithmeticParams &p, Activation act)
+{
+  static_assert(one_of_types<T, float, int32_t, int64_t>(), "Unsupported dtype");
+
+  if (std::is_same<T, float>::value)
+    calculateActivationRange(act, &p.float_activation_min, &p.float_activation_max);
+  if (std::is_same<T, int32_t>::value)
+    calculateActivationRange(act, &p.quantized_activation_min, &p.quantized_activation_max);
+  else
+    calculateActivationRange(act, &p.int64_activation_min, &p.int64_activation_max);
+}
+
 // Decompose a double multiplier into a Q0.31 int32 representation of its
 // significand, and shift representation of its exponent.
 //
diff --git a/compiler/luci-interpreter/src/loader/CMakeLists.txt b/compiler/luci-interpreter/src/loader/CMakeLists.txt
index 2cde99f5d..292771592 100644
--- a/compiler/luci-interpreter/src/loader/CMakeLists.txt
+++ b/compiler/luci-interpreter/src/loader/CMakeLists.txt
@@ -17,7 +17,9 @@ endmacro(REGISTER_KERNEL)
 include(${KERNEL_REGISTER_FILE})
 
 add_library(${LUCI_INTERPRETER_LOADER} STATIC ${SOURCES})
-set_target_properties(${LUCI_INTERPRETER_LOADER} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+if (NOT NNCC_LIBRARY_NO_PIC)
+  set_target_properties(${LUCI_INTERPRETER_LOADER} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif(NOT NNCC_LIBRARY_NO_PIC)
 target_include_directories(${LUCI_INTERPRETER_LOADER} PUBLIC "${LUCI_INTERPRETER_PAL_DIR}")
 target_include_directories(${LUCI_INTERPRETER_LOADER} PUBLIC "${LUCI_INTERPRETER_SOURCE_DIR}")
 
diff --git a/compiler/luci-interpreter/src/loader/GraphLoader.cpp b/compiler/luci-interpreter/src/loader/GraphLoader.cpp
index a14442ed5..dba39050c 100644
--- a/compiler/luci-interpreter/src/loader/GraphLoader.cpp
+++ b/compiler/luci-interpreter/src/loader/GraphLoader.cpp
@@ -73,6 +73,26 @@ const void *getNodeData(const luci::CircleConst *node, size_t *data_size)
   }
 }
 
+const void *getNodeData(const luci::CircleCustom *node, size_t *data_size)
+{
+  if (node->custom_code() != "CircleReferencingConst")
+    return nullptr;
+
+  // helper struct which describes data loaded to custom_options of CircleReferencingConst node
+  // TODO move this struct to header
+  struct ConstDataReference
+  {
+    const uint8_t *data = nullptr;
+    uint32_t size = 0;
+  };
+
+  const auto &custom_options = node->custom_options();
+  const auto &const_data_ref = *reinterpret_cast<const ConstDataReference *>(custom_options.data());
+
+  *data_size = const_data_ref.size;
+  return const_data_ref.data;
+}
+
 bool isExecutableNode(const luci::CircleNode *node)
 {
   switch (node->opcode())
@@ -83,12 +103,30 @@ bool isExecutableNode(const luci::CircleNode *node)
     case luci::CircleOpcode::CIRCLEOUTPUT:
     case luci::CircleOpcode::CIRCLEOUTPUTEXCLUDE:
     // The following nodes denote outputs of multiple-output nodes.
+    case luci::CircleOpcode::CIRCLEBIDIRECTIONAL_SEQUENCE_LSTM_OUT:
+    case luci::CircleOpcode::CIRCLECUSTOMOUT:
     case luci::CircleOpcode::CIRCLEIFOUT:
+    case luci::CircleOpcode::CIRCLENONMAXSUPPRESSIONV4OUT:
+    case luci::CircleOpcode::CIRCLENONMAXSUPPRESSIONV5OUT:
     case luci::CircleOpcode::CIRCLESPLITOUT:
     case luci::CircleOpcode::CIRCLESPLITVOUT:
+    case luci::CircleOpcode::CIRCLETOPKV2OUT:
+    case luci::CircleOpcode::CIRCLEUNIQUEOUT:
     case luci::CircleOpcode::CIRCLEUNPACKOUT:
+    case luci::CircleOpcode::CIRCLEVARIABLE:
     case luci::CircleOpcode::CIRCLEWHILEOUT:
       return false;
+    // Custom nodes may be executable and non-executable
+    case luci::CircleOpcode::CUSTOM:
+    {
+      auto const custom_node = loco::must_cast<const luci::CircleCustom *>(node);
+
+      // TODO handle more non-executable Custom ops here
+      if (custom_node->custom_code() == "CircleReferencingConst")
+        return false;
+
+      return true;
+    }
     default:
       return true;
   }
@@ -102,15 +140,34 @@ bool isTensorProducingNode(const luci::CircleNode *node)
     case luci::CircleOpcode::CIRCLEOUTPUT:
     // The following nodes are multiple-output nodes. They do not produce tensors, the tensors
     // are produced by the corresponding *Out nodes instead.
+    case luci::CircleOpcode::BIDIRECTIONAL_SEQUENCE_LSTM:
+    case luci::CircleOpcode::CUSTOM:
     case luci::CircleOpcode::IF:
+    case luci::CircleOpcode::NON_MAX_SUPPRESSION_V4:
+    case luci::CircleOpcode::NON_MAX_SUPPRESSION_V5:
     case luci::CircleOpcode::SPLIT:
+    case luci::CircleOpcode::SPLIT_V:
+    case luci::CircleOpcode::TOPK_V2:
+    case luci::CircleOpcode::UNIQUE:
     case luci::CircleOpcode::UNPACK:
+    case luci::CircleOpcode::WHILE:
       return false;
     default:
       return true;
   }
 }
 
+bool isSupportedCustomNode(const luci::CircleNode *node)
+{
+  const auto custom_node = loco::must_cast<const luci::CircleCustom *>(node);
+
+  // TODO handle more Custom ops here
+  if (custom_node->custom_code() == "CircleReferencingConst")
+    return true;
+
+  return false;
+}
+
 } // namespace
 
 GraphLoader::GraphLoader(
@@ -129,18 +186,25 @@ void GraphLoader::loadTensors()
   {
     const auto *node = loco::must_cast<const luci::CircleNode *>(_graph->nodes()->at(i));
 
+    if (node->opcode() == luci::CircleOpcode::CUSTOM && !isSupportedCustomNode(node))
+      throw std::runtime_error("Unknown Custom Node, yet.");
+
     if (!isTensorProducingNode(node))
       continue;
 
-    // Only Input and Const nodes have shapes. Shapes of intermediate tensors will be inferred.
+    // Only Input, Const, Custom and Variable nodes have shapes. Shapes of intermediate tensors will
+    // be inferred.
     Shape shape{};
-    if (const auto *input_node = dynamic_cast<const luci::CircleInput *>(node))
+    switch (node->opcode())
     {
-      shape = getNodeShape(input_node);
-    }
-    else if (const auto *const_node = dynamic_cast<const luci::CircleConst *>(node))
-    {
-      shape = getNodeShape(const_node);
+      case luci::CircleOpcode::CIRCLECONST:
+      case luci::CircleOpcode::CIRCLECUSTOMOUT:
+      case luci::CircleOpcode::CIRCLEINPUT:
+      case luci::CircleOpcode::CIRCLEVARIABLE:
+        shape = getNodeShape(node);
+        break;
+      default:
+        break;
     }
 
     AffineQuantization quantization;
@@ -175,6 +239,22 @@ void GraphLoader::loadTensors()
         tensor->writeData(const_data, data_size);
       }
     }
+    else if (const auto *custom_out_node = dynamic_cast<const luci::CircleCustomOut *>(node))
+    {
+      const auto *custom_node =
+        loco::must_cast<const luci::CircleCustom *>(custom_out_node->input());
+
+      if (custom_node->custom_code() == "CircleReferencingConst")
+      {
+        size_t data_size{};
+        const void *const_data = getNodeData(custom_node, &data_size);
+        if (const_data != nullptr)
+        {
+          _memory_manager->allocate_memory(*tensor);
+          tensor->writeData(const_data, data_size);
+        }
+      }
+    }
 
     _node_to_tensor.emplace(node, tensor.get());
     _runtime_to_ir.tensor_to_node.emplace(tensor.get(), node);
diff --git a/compiler/luci-interpreter/src/loader/KernelBuilder.test.cpp b/compiler/luci-interpreter/src/loader/KernelBuilder.test.cpp
index 7a457a62f..b221b6921 100644
--- a/compiler/luci-interpreter/src/loader/KernelBuilder.test.cpp
+++ b/compiler/luci-interpreter/src/loader/KernelBuilder.test.cpp
@@ -21,6 +21,7 @@
 #include <kernels/Add.h>
 #include <kernels/ArgMax.h>
 #include <kernels/AveragePool2D.h>
+#include <kernels/BatchMatMul.h>
 #include <kernels/Cast.h>
 #include <kernels/Concatenation.h>
 #include <kernels/Conv2D.h>
@@ -54,6 +55,7 @@
 #include <kernels/Mul.h>
 #include <kernels/Neg.h>
 #include <kernels/NotEqual.h>
+#include <kernels/OneHot.h>
 #include <kernels/Pad.h>
 #include <kernels/PadV2.h>
 #include <kernels/Pow.h>
@@ -209,6 +211,27 @@ TEST_F(KernelBuilderTest, AveragePool2D)
   EXPECT_THAT(kernel->params().activation, Eq(op->fusedActivationFunction()));
 }
 
+TEST_F(KernelBuilderTest, BatchMatMul)
+{
+  auto *lhs = createInputNode();
+  auto *rhs = createInputNode();
+
+  auto *op = createNode<luci::CircleBatchMatMul>();
+  op->x(lhs);
+  op->y(rhs);
+  op->adj_x(false);
+  op->adj_y(false);
+
+  auto kernel = buildKernel<kernels::BatchMatMul>(op);
+  ASSERT_THAT(kernel, NotNull());
+
+  checkTensor(kernel->x(), lhs);
+  checkTensor(kernel->y(), rhs);
+  checkTensor(kernel->output(), op);
+  EXPECT_THAT(kernel->params().adj_x, Eq(op->adj_x()));
+  EXPECT_THAT(kernel->params().adj_y, Eq(op->adj_y()));
+}
+
 TEST_F(KernelBuilderTest, Cast)
 {
   auto *input = createInputNode();
@@ -832,6 +855,31 @@ TEST_F(KernelBuilderTest, NotEqual)
   checkTensor(kernel->output(), op);
 }
 
+TEST_F(KernelBuilderTest, OneHot)
+{
+  auto *indices = createInputNode();
+  auto *depth = createInputNode();
+  auto *on_value = createInputNode();
+  auto *off_value = createInputNode();
+  auto axis = 1;
+
+  auto *op = createNode<luci::CircleOneHot>();
+  op->indices(indices);
+  op->depth(depth);
+  op->on_value(on_value);
+  op->off_value(off_value);
+  op->axis(axis);
+
+  auto kernel = buildKernel<kernels::OneHot>(op);
+  ASSERT_THAT(kernel, NotNull());
+
+  checkTensor(kernel->indices(), indices);
+  checkTensor(kernel->depth(), depth);
+  checkTensor(kernel->on_value(), on_value);
+  checkTensor(kernel->off_value(), off_value);
+  EXPECT_THAT(kernel->params().axis, Eq(op->axis()));
+}
+
 TEST_F(KernelBuilderTest, Pad)
 {
   auto *input = createInputNode();
diff --git a/compiler/luci-interpreter/src/loader/nodes/AveragePool2D.cpp b/compiler/luci-interpreter/src/loader/nodes/AveragePool2D.cpp
index 5bc37bd4a..efb011257 100644
--- a/compiler/luci-interpreter/src/loader/nodes/AveragePool2D.cpp
+++ b/compiler/luci-interpreter/src/loader/nodes/AveragePool2D.cpp
@@ -17,6 +17,7 @@
 #include "Builders.h"
 
 #include "kernels/AveragePool2D.h"
+#include <luci/Plan/CircleNodeExecutionPlan.h>
 
 namespace luci_interpreter
 {
@@ -40,7 +41,26 @@ std::unique_ptr<Kernel> build_kernel_CircleAveragePool2D(const luci::CircleNode
   params.stride_width = node->stride()->w();
   params.activation = node->fusedActivationFunction();
 
-  return std::make_unique<kernels::AveragePool2D>(input, output, params);
+  // It is unknown what data will be stored in scratchpad tensor,
+  // using UINT8 as a most general option
+  auto scratchpad = std::make_unique<Tensor>(DataType::U8, Shape({}), AffineQuantization{}, "");
+  scratchpad->set_observable(false);
+  scratchpad->set_data_buffer(nullptr);
+  // If node has execution plan then read memory offsets for scratchpad temporary tensor
+  // from the beginning of shared memory buffer.
+  // Used in Static Memory Manager.
+  // TODO move tensors offset initialization to one place
+  if (luci::has_execution_plan(node))
+  {
+    const auto execution_plan = luci::get_execution_plan(node);
+    // Check whether the offset for the current CircleConv2D temporary was found.
+    if (execution_plan.offsets().size() > 1)
+      // If this is true, then we keep this offset in scratchpad.
+      scratchpad->set_offset(execution_plan.offsets().at(1));
+  }
+  Tensor *tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad));
+
+  return std::make_unique<kernels::AveragePool2D>(input, output, tmp, params);
 }
 
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/BatchMatMul.cpp b/compiler/luci-interpreter/src/loader/nodes/BatchMatMul.cpp
new file mode 100644
index 000000000..aae3dbab1
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/BatchMatMul.cpp
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/BatchMatMul.h"
+#include <luci/Plan/CircleNodeExecutionPlan.h>
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleBatchMatMul(const luci::CircleNode *circle_node,
+                                                       KernelBuilderHelper &helper)
+{
+  const auto *node = dynamic_cast<const luci::CircleBatchMatMul *>(circle_node);
+  if (node == nullptr)
+    throw std::runtime_error("wrong builder for operation");
+  assert(node->arity() == 2);
+
+  const Tensor *lhs = helper.getInputTensor(node->x());
+  const Tensor *rhs = helper.getInputTensor(node->y());
+  Tensor *output = helper.getOutputTensor(node);
+
+  auto lhs_scratchpad =
+    std::make_unique<Tensor>(lhs->element_type(), Shape({}), AffineQuantization{}, "");
+  lhs_scratchpad->set_observable(false);
+  lhs_scratchpad->set_data_buffer(nullptr);
+  auto rhs_scratchpad =
+    std::make_unique<Tensor>(rhs->element_type(), Shape({}), AffineQuantization{}, "");
+  rhs_scratchpad->set_observable(false);
+  rhs_scratchpad->set_data_buffer(nullptr);
+  // If node has execution plan then read memory offsets for scratchpad temporary tensor
+  // from the beginning of shared memory buffer.
+  // Used in Static Memory Manager.
+  // TODO move tensors offset initialization to one place
+  if (luci::has_execution_plan(node))
+  {
+    const auto execution_plan = luci::get_execution_plan(node);
+    // Check whether the offset for the current BatchMatMul temporary was found.
+    if (execution_plan.offsets().size() > 1)
+    {
+      assert(execution_plan.offsets().size() == 3);
+
+      // If this is true, then we keep this offset in scratchpad.
+      lhs_scratchpad->set_offset(execution_plan.offsets().at(1));
+      rhs_scratchpad->set_offset(execution_plan.offsets().at(2));
+    }
+  }
+  Tensor *lhs_tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(lhs_scratchpad));
+  Tensor *rhs_tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(rhs_scratchpad));
+
+  BatchMatMulParams params;
+  params.adj_x = node->adj_x();
+  params.adj_y = node->adj_y();
+
+  return std::make_unique<kernels::BatchMatMul>(lhs, rhs, output, lhs_tmp, rhs_tmp, params);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/Conv2D.cpp b/compiler/luci-interpreter/src/loader/nodes/Conv2D.cpp
index 22fd1aca4..b48d97d19 100644
--- a/compiler/luci-interpreter/src/loader/nodes/Conv2D.cpp
+++ b/compiler/luci-interpreter/src/loader/nodes/Conv2D.cpp
@@ -35,11 +35,12 @@ std::unique_ptr<Kernel> build_kernel_CircleConv2D(const luci::CircleNode *circle
   const Tensor *bias = helper.getOptionalInputTensor(node->bias());
   Tensor *output = helper.getOutputTensor(node);
 
-  auto im2col =
-    std::make_unique<Tensor>(input->element_type(), Shape({}), AffineQuantization{}, "");
-  im2col->set_observable(false);
-  im2col->set_data_buffer(nullptr);
-  // If node has execution plan then read memory offsets for im2col temporary tensor
+  // It is unknown what data will be stored in scratchpad tensor,
+  // using UINT8 as a most general option
+  auto scratchpad = std::make_unique<Tensor>(DataType::U8, Shape({}), AffineQuantization{}, "");
+  scratchpad->set_observable(false);
+  scratchpad->set_data_buffer(nullptr);
+  // If node has execution plan then read memory offsets for scratchpad temporary tensor
   // from the beginning of shared memory buffer.
   // Used in Static Memory Manager.
   // TODO move tensors offset initialization to one place
@@ -48,10 +49,10 @@ std::unique_ptr<Kernel> build_kernel_CircleConv2D(const luci::CircleNode *circle
     const auto execution_plan = luci::get_execution_plan(node);
     // Check whether the offset for the current CircleConv2D temporary was found.
     if (execution_plan.offsets().size() > 1)
-      // If this is true, then we keep this offset in im2col.
-      im2col->set_offset(execution_plan.offsets().at(1));
+      // If this is true, then we keep this offset in scratchpad.
+      scratchpad->set_offset(execution_plan.offsets().at(1));
   }
-  Tensor *tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(im2col));
+  Tensor *tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad));
 
   Conv2DParams params{};
   params.padding = node->padding();
diff --git a/compiler/luci-interpreter/src/loader/nodes/DepthwiseConv2D.cpp b/compiler/luci-interpreter/src/loader/nodes/DepthwiseConv2D.cpp
index c2f0346a2..db26ecf2e 100644
--- a/compiler/luci-interpreter/src/loader/nodes/DepthwiseConv2D.cpp
+++ b/compiler/luci-interpreter/src/loader/nodes/DepthwiseConv2D.cpp
@@ -17,6 +17,7 @@
 #include "Builders.h"
 
 #include "kernels/DepthwiseConv2D.h"
+#include <luci/Plan/CircleNodeExecutionPlan.h>
 
 namespace luci_interpreter
 {
@@ -43,7 +44,26 @@ std::unique_ptr<Kernel> build_kernel_CircleDepthwiseConv2D(const luci::CircleNod
   params.dilation_width_factor = node->dilation()->w();
   params.activation = node->fusedActivationFunction();
 
-  return std::make_unique<kernels::DepthwiseConv2D>(input, filter, bias, output, params);
+  // It is unknown what data will be stored in scratchpad tensor,
+  // using UINT8 as a most general option
+  auto scratchpad = std::make_unique<Tensor>(DataType::U8, Shape({}), AffineQuantization{}, "");
+  scratchpad->set_observable(false);
+  scratchpad->set_data_buffer(nullptr);
+  // If node has execution plan then read memory offsets for scratchpad temporary tensor
+  // from the beginning of shared memory buffer.
+  // Used in Static Memory Manager.
+  // TODO move tensors offset initialization to one place
+  if (luci::has_execution_plan(node))
+  {
+    const auto execution_plan = luci::get_execution_plan(node);
+    // Check whether the offset for the current CircleConv2D temporary was found.
+    if (execution_plan.offsets().size() > 1)
+      // If this is true, then we keep this offset in scratchpad.
+      scratchpad->set_offset(execution_plan.offsets().at(1));
+  }
+  Tensor *tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad));
+
+  return std::make_unique<kernels::DepthwiseConv2D>(input, filter, bias, output, tmp, params);
 }
 
 } // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/Dequantize.cpp b/compiler/luci-interpreter/src/loader/nodes/Dequantize.cpp
new file mode 100644
index 000000000..4aae56469
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/Dequantize.cpp
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/Dequantize.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleDequantize(const luci::CircleNode *circle_node,
+                                                      KernelBuilderHelper &helper)
+{
+  const auto *node = dynamic_cast<const luci::CircleDequantize *>(circle_node);
+  if (node == nullptr)
+    throw std::runtime_error("wrong builder for operation");
+
+  const Tensor *input = helper.getInputTensor(node->input());
+  Tensor *output = helper.getOutputTensor(node);
+
+  return std::make_unique<kernels::Dequantize>(input, output);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/ExpandDims.cpp b/compiler/luci-interpreter/src/loader/nodes/ExpandDims.cpp
new file mode 100644
index 000000000..9840c34e5
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/ExpandDims.cpp
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/ExpandDims.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleExpandDims(const luci::CircleNode *circle_node,
+                                                      KernelBuilderHelper &helper)
+{
+  const auto *node = loco::must_cast<const luci::CircleExpandDims *>(circle_node);
+  assert(node->arity() == 2);
+
+  const Tensor *input = helper.getInputTensor(node->input());
+  const Tensor *axis = helper.getInputTensor(node->axis());
+  Tensor *output = helper.getOutputTensor(node);
+
+  return std::make_unique<kernels::ExpandDims>(input, axis, output);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/FullyConnected.cpp b/compiler/luci-interpreter/src/loader/nodes/FullyConnected.cpp
index 2917598fc..0b8ac44bd 100644
--- a/compiler/luci-interpreter/src/loader/nodes/FullyConnected.cpp
+++ b/compiler/luci-interpreter/src/loader/nodes/FullyConnected.cpp
@@ -36,6 +36,7 @@ std::unique_ptr<Kernel> build_kernel_CircleFullyConnected(const luci::CircleNode
 
   FullyConnectedParams params{};
   params.activation = node->fusedActivationFunction();
+  params.keep_num_dims = node->keep_num_dims();
 
   return std::make_unique<kernels::FullyConnected>(input, weights, bias, output, params);
 }
diff --git a/compiler/luci-interpreter/src/loader/nodes/Gather.cpp b/compiler/luci-interpreter/src/loader/nodes/Gather.cpp
new file mode 100644
index 000000000..9df9775c5
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/Gather.cpp
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/Gather.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleGather(const luci::CircleNode *circle_node,
+                                                  KernelBuilderHelper &helper)
+{
+  const auto *node = dynamic_cast<const luci::CircleGather *>(circle_node);
+  if (node == nullptr)
+    throw std::runtime_error("wrong builder for operation");
+  assert(node->arity() == 2);
+
+  const Tensor *params = helper.getInputTensor(node->params());
+  const Tensor *indices = helper.getInputTensor(node->indices());
+  Tensor *output = helper.getOutputTensor(node);
+
+  GatherParams gparams{};
+  gparams.axis = node->axis();
+  // TODO support batch_dims
+  gparams.batch_dims = 0;
+
+  return std::make_unique<kernels::Gather>(params, indices, output, gparams);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/OneHot.cpp b/compiler/luci-interpreter/src/loader/nodes/OneHot.cpp
new file mode 100644
index 000000000..a40160945
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/OneHot.cpp
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2021 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/OneHot.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleOneHot(const luci::CircleNode *circle_node,
+                                                  KernelBuilderHelper &helper)
+{
+  const auto *node = loco::must_cast<const luci::CircleOneHot *>(circle_node);
+  assert(node->arity() == 4);
+
+  const Tensor *indices = helper.getInputTensor(node->indices());
+  const Tensor *depth = helper.getInputTensor(node->depth());
+  const Tensor *on_value = helper.getInputTensor(node->on_value());
+  const Tensor *off_value = helper.getInputTensor(node->off_value());
+  Tensor *output = helper.getOutputTensor(node);
+
+  OneHotParams params{};
+  params.axis = node->axis();
+
+  return std::make_unique<kernels::OneHot>(indices, depth, on_value, off_value, output, params);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/Quantize.cpp b/compiler/luci-interpreter/src/loader/nodes/Quantize.cpp
new file mode 100644
index 000000000..fd9836345
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/Quantize.cpp
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/Quantize.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleQuantize(const luci::CircleNode *circle_node,
+                                                    KernelBuilderHelper &helper)
+{
+  const auto *node = dynamic_cast<const luci::CircleQuantize *>(circle_node);
+  if (node == nullptr)
+    throw std::runtime_error("wrong builder for operation");
+
+  const Tensor *input = helper.getInputTensor(node->input());
+  Tensor *output = helper.getOutputTensor(node);
+
+  return std::make_unique<kernels::Quantize>(input, output);
+}
+
+} // namespace luci_interpreter
diff --git a/compiler/luci-interpreter/src/loader/nodes/SVDF.cpp b/compiler/luci-interpreter/src/loader/nodes/SVDF.cpp
new file mode 100644
index 000000000..89528d5ee
--- /dev/null
+++ b/compiler/luci-interpreter/src/loader/nodes/SVDF.cpp
@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2022 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Builders.h"
+
+#include "kernels/SVDF.h"
+
+namespace luci_interpreter
+{
+
+std::unique_ptr<Kernel> build_kernel_CircleSVDF(const luci::CircleNode *circle_node,
+                                                KernelBuilderHelper &helper)
+{
+  const auto *node = dynamic_cast<const luci::CircleSVDF *>(circle_node);
+  if (node == nullptr)
+    throw std::runtime_error("wrong builder for operation");
+
+  const Tensor *input = helper.getInputTensor(node->input());
+  const Tensor *feature = helper.getInputTensor(node->weight_feature());
+  const Tensor *time = helper.getInputTensor(node->weight_time());
+  const Tensor *bias = helper.getOptionalInputTensor(node->bias());
+  const Tensor *input_activation_state = helper.getInputTensor(node->input_activation_state());
+  Tensor *output = helper.getOutputTensor(node);
+
+  auto scratchpad_tensor = std::make_unique<Tensor>(input_activation_state->element_type(),
+                                                    Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  DataType data_type = input->element_type() == DataType::S8 ? DataType::S32 : DataType::FLOAT32;
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_1 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  if (data_type == DataType::FLOAT32 &&
+      (feature->element_type() == DataType::S8 || feature->element_type() == DataType::U8))
+  {
+    data_type = feature->element_type();
+  }
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_2 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  data_type = DataType::FLOAT32;
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_3 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_4 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_5 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  scratchpad_tensor = std::make_unique<Tensor>(data_type, Shape({}), AffineQuantization{}, "");
+  scratchpad_tensor->set_observable(false);
+  scratchpad_tensor->set_data_buffer(nullptr);
+  Tensor *tmp_6 = helper.getRuntimeGraph(node->graph())->addTensor(std::move(scratchpad_tensor));
+
+  SVDFParams params{};
+  params.activation = node->fusedActivationFunction();
+  params.svdf_rank = node->svdf_rank();
+  params.asymmetric_quantize_inputs = node->asymmetric_quantize_inputs();
+
+  return std::make_unique<kernels::SVDF>(input, feature, time, bias, input_activation_state, output,
+                                         tmp, tmp_1, tmp_2, tmp_3, tmp_4, tmp_5, tmp_6, params);
+}
+
+} // namespace luci_interpreter