COMPMID-2980 (Nightly) armv7a build failures

Change-Id: I8c2a20fc345694d1ad6e0fe63e4f22fb73e6c1df
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2463
Tested-by: Arm Jenkins <bsgcomp@arm.com>

13 files changed, 83 insertions, 81 deletions

diff --git a/arm_compute/core/CPP/Validate.h b/arm_compute/core/CPP/Validate.h
index 1ec41a9f1..f195a31d0 100644
--- a/arm_compute/core/CPP/Validate.h
+++ b/arm_compute/core/CPP/Validate.h
@@ -37,15 +37,15 @@ namespace arm_compute
  *
  * @return Status
  */
-inline arm_compute::Status error_on_unsupported_cpu_fp16(const char *function, const char *file, const int line,
-                                                         const ITensorInfo *tensor_info)
+inline Status error_on_unsupported_cpu_fp16(const char *function, const char *file, const int line,
+                                            const ITensorInfo *tensor_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor_info == nullptr, function, file, line);
 #ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
     ARM_COMPUTE_RETURN_ERROR_ON_LOC_MSG(tensor_info->data_type() == DataType::F16,
                                         function, file, line, "This CPU architecture does not support F16 data type, you need v8.2 or above");
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-    return arm_compute::Status {};
+    return Status {};
 }
 
 /** Return an error if the data type of the passed tensor is FP16 and FP16 support is not compiled in.
@@ -57,12 +57,12 @@ inline arm_compute::Status error_on_unsupported_cpu_fp16(const char *function, c
  *
  * @return Status
  */
-inline arm_compute::Status error_on_unsupported_cpu_fp16(const char *function, const char *file, const int line,
-                                                         const ITensor *tensor)
+inline Status error_on_unsupported_cpu_fp16(const char *function, const char *file, const int line,
+                                            const ITensor *tensor)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_LOC(tensor == nullptr, function, file, line);
     ARM_COMPUTE_RETURN_ON_ERROR(::arm_compute::error_on_unsupported_cpu_fp16(function, file, line, tensor->info()));
-    return arm_compute::Status{};
+    return Status{};
 }
 
 #define ARM_COMPUTE_ERROR_ON_CPU_F16_UNSUPPORTED(tensor) \
diff --git a/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_quantized.hpp b/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_quantized.hpp
index ef3adc4c0..37c1f1bc8 100644
--- a/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_quantized.hpp
+++ b/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_quantized.hpp
@@ -31,31 +31,21 @@
 using namespace neon_convolution_kernels;
 using namespace qasymm8;
 
-template <typename T, typename U = int32_t>
-inline T saturating_doubling_high_mul(const T&, const U&);
-
-template <>
 inline int32x4_t saturating_doubling_high_mul(const int32x4_t& a, const int32x4_t& b)
 {
   return vqrdmulhq_s32(a, b);
 }
 
-template <>
 inline int32x4_t saturating_doubling_high_mul(const int32x4_t& a, const int32_t& b)
 {
   return vqrdmulhq_n_s32(a, b);
 }
 
-template <>
 inline int32_t saturating_doubling_high_mul(const int32_t& a, const int32_t& b)
 {
   return vget_lane_s32(vqrdmulh_n_s32(vdup_n_s32(a), b), 0);
 }
 
-template <typename T, typename U = int32_t>
-inline T rounding_divide_by_exp2(const T& x, const U exponent);
-
-template <>
 inline int32x4_t rounding_divide_by_exp2(const int32x4_t& x, const int32x4_t shift)
 {
   const int32x4_t fixup = vshrq_n_s32(vandq_s32(x, shift), 31);
@@ -63,7 +53,6 @@ inline int32x4_t rounding_divide_by_exp2(const int32x4_t& x, const int32x4_t shi
   return vrshlq_s32(fixed, shift);
 }
 
-template <>
 inline int32x4_t rounding_divide_by_exp2(const int32x4_t& x, const int exponent)
 {
   const int32x4_t shift = vdupq_n_s32(-exponent);
@@ -72,7 +61,6 @@ inline int32x4_t rounding_divide_by_exp2(const int32x4_t& x, const int exponent)
   return vrshlq_s32(fixed, shift);
 }
 
-template <>
 inline int32x2_t rounding_divide_by_exp2(const int32x2_t& x, const int exponent)
 {
   const int32x2_t shift = vdup_n_s32(-exponent);
@@ -81,7 +69,6 @@ inline int32x2_t rounding_divide_by_exp2(const int32x2_t& x, const int exponent)
   return vrshl_s32(fixed, shift);
 }
 
-template <>
 inline int32_t rounding_divide_by_exp2(const int32_t& x, const int exponent)
 {
   const int32x2_t xs = vdup_n_s32(x);
diff --git a/arm_compute/core/PixelValue.h b/arm_compute/core/PixelValue.h
index 8c2ab92ad..52b1f654a 100644
--- a/arm_compute/core/PixelValue.h
+++ b/arm_compute/core/PixelValue.h
@@ -103,6 +103,15 @@ public:
                 break;
         }
     }
+    /** Initialize the union with a S8 pixel value
+     *
+     * @param[in] v S8 value.
+     */
+    PixelValue(int8_t v)
+        : PixelValue()
+    {
+        value.s8 = v;
+    }
     /** Initialize the union with a U8 pixel value
      *
      * @param[in] v U8 value.
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index c2813122b..901d080b0 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1876,11 +1876,11 @@ enum class GEMMLowpOutputStageType
 struct GEMMLowpOutputStageInfo
 {
     GEMMLowpOutputStageType type{ GEMMLowpOutputStageType::NONE }; /**< GEMMLowp output stage type */
-    int                     gemmlowp_offset{ 0 };                  /**< GEMMLowp output stage offset used for quantizing to QASYMM8 */
-    int                     gemmlowp_multiplier{ 0 };              /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
-    int                     gemmlowp_shift{ 0 };                   /**< GEMMLowp output stage shift used for quantizing to uint8 */
-    int                     gemmlowp_min_bound{ 0 };               /**< GEMMLowp min value used to saturate down the output result before converting back to QASYMM8 */
-    int                     gemmlowp_max_bound{ 0 };               /**< GEMMLowp max value used to saturate down the output result before converting back to QASYMM8 */
+    int32_t                 gemmlowp_offset{ 0 };                  /**< GEMMLowp output stage offset used for quantizing to QASYMM8 */
+    int32_t                 gemmlowp_multiplier{ 0 };              /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
+    int32_t                 gemmlowp_shift{ 0 };                   /**< GEMMLowp output stage shift used for quantizing to uint8 */
+    int32_t                 gemmlowp_min_bound{ 0 };               /**< GEMMLowp min value used to saturate down the output result before converting back to QASYMM8 */
+    int32_t                 gemmlowp_max_bound{ 0 };               /**< GEMMLowp max value used to saturate down the output result before converting back to QASYMM8 */
     std::vector<int32_t>    gemmlowp_multipliers{};                /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
     std::vector<int32_t>    gemmlowp_shifts{};                     /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
     bool                    is_quantized_per_channel{ false };     /**< GEMMLowp quantized per-channel flag */
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index c11fffec9..18c5471f8 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -557,15 +557,15 @@ inline DataType get_promoted_data_type(DataType dt)
  */
 inline std::tuple<PixelValue, PixelValue> get_min_max(DataType dt)
 {
-    PixelValue min(0);
-    PixelValue max(0);
+    PixelValue min{};
+    PixelValue max{};
     switch(dt)
     {
         case DataType::U8:
         case DataType::QASYMM8:
         {
-            min = PixelValue(std::numeric_limits<uint8_t>::lowest());
-            max = PixelValue(std::numeric_limits<uint8_t>::max());
+            min = PixelValue(static_cast<int32_t>(std::numeric_limits<uint8_t>::lowest()));
+            max = PixelValue(static_cast<int32_t>(std::numeric_limits<uint8_t>::max()));
             break;
         }
         case DataType::S8:
@@ -573,22 +573,22 @@ inline std::tuple<PixelValue, PixelValue> get_min_max(DataType dt)
         case DataType::QASYMM8_SIGNED:
         case DataType::QSYMM8_PER_CHANNEL:
         {
-            min = PixelValue(std::numeric_limits<int8_t>::lowest());
-            max = PixelValue(std::numeric_limits<int8_t>::max());
+            min = PixelValue(static_cast<int32_t>(std::numeric_limits<int8_t>::lowest()));
+            max = PixelValue(static_cast<int32_t>(std::numeric_limits<int8_t>::max()));
             break;
         }
         case DataType::U16:
         case DataType::QASYMM16:
         {
-            min = PixelValue(std::numeric_limits<uint16_t>::lowest());
-            max = PixelValue(std::numeric_limits<uint16_t>::max());
+            min = PixelValue(static_cast<int32_t>(std::numeric_limits<uint16_t>::lowest()));
+            max = PixelValue(static_cast<int32_t>(std::numeric_limits<uint16_t>::max()));
             break;
         }
         case DataType::S16:
         case DataType::QSYMM16:
         {
-            min = PixelValue(std::numeric_limits<int16_t>::lowest());
-            max = PixelValue(std::numeric_limits<int16_t>::max());
+            min = PixelValue(static_cast<int32_t>(std::numeric_limits<int16_t>::lowest()));
+            max = PixelValue(static_cast<int32_t>(std::numeric_limits<int16_t>::max()));
             break;
         }
         case DataType::U32:
diff --git a/arm_compute/core/utils/quantization/AsymmHelpers.h b/arm_compute/core/utils/quantization/AsymmHelpers.h
index f0b077096..1bdc9959c 100644
--- a/arm_compute/core/utils/quantization/AsymmHelpers.h
+++ b/arm_compute/core/utils/quantization/AsymmHelpers.h
@@ -40,7 +40,7 @@ namespace quantization
  *
  * @return a status
  */
-Status calculate_quantized_multiplier(float multiplier, int *quant_multiplier, int *shift);
+Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift);
 /** Calculate quantized representation of multiplier with value less than one.
  *
  * @param[in]  multiplier       Real multiplier.
@@ -49,7 +49,7 @@ Status calculate_quantized_multiplier(float multiplier, int *quant_multiplier, i
  *
  * @return a status
  */
-Status calculate_quantized_multiplier_less_than_one(float multiplier, int *quant_multiplier, int *right_shift);
+Status calculate_quantized_multiplier_less_than_one(float multiplier, int32_t *quant_multiplier, int32_t *right_shift);
 /** Calculate quantized representation of multiplier having value greater than one.
  *
  * @param[in]  multiplier           Real multiplier.
@@ -58,7 +58,7 @@ Status calculate_quantized_multiplier_less_than_one(float multiplier, int *quant
  *
  * @return a status
  */
-Status calculate_quantized_multiplier_greater_than_one(float multiplier, int *quantized_multiplier, int *left_shift);
+Status calculate_quantized_multiplier_greater_than_one(float multiplier, int32_t *quantized_multiplier, int32_t *left_shift);
 
 /** Calculate quantized representation of per-channel multipliers with value less than one.
  *
diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
index a0d45afd2..a9a3183c5 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
@@ -528,8 +528,8 @@ void NEDepthwiseConvolutionLayerNativeKernel::configure(const ITensor *input, co
 
         for(size_t i = 0; i < weights_scale.size(); ++i)
         {
-            int         out_mult   = 0;
-            int         out_shift  = 0;
+            int32_t     out_mult   = 0;
+            int32_t     out_shift  = 0;
             const float multiplier = input_scale * weights_scale.at(i) / output_scale;
             ARM_COMPUTE_ERROR_ON(multiplier > 1.f);
             arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &out_mult, &out_shift);
diff --git a/src/core/utils/quantization/AsymmHelpers.cpp b/src/core/utils/quantization/AsymmHelpers.cpp
index 619666244..11241e83a 100644
--- a/src/core/utils/quantization/AsymmHelpers.cpp
+++ b/src/core/utils/quantization/AsymmHelpers.cpp
@@ -35,7 +35,7 @@ namespace quantization
 constexpr int64_t fixed_point_one_Q0 = (1LL << 31);
 constexpr float   epsilon            = 0.00001f;
 
-Status calculate_quantized_multiplier(float multiplier, int *quant_multiplier, int *shift)
+Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift)
 {
     if(multiplier >= 1.f)
     {
@@ -49,9 +49,9 @@ Status calculate_quantized_multiplier(float multiplier, int *quant_multiplier, i
     }
 }
 
-Status calculate_quantized_multiplier_less_than_one(float multiplier,
-                                                    int *quant_multiplier,
-                                                    int *right_shift)
+Status calculate_quantized_multiplier_less_than_one(float    multiplier,
+                                                    int32_t *quant_multiplier,
+                                                    int32_t *right_shift)
 {
     ARM_COMPUTE_RETURN_ERROR_ON(quant_multiplier == nullptr);
     ARM_COMPUTE_RETURN_ERROR_ON(right_shift == nullptr);
@@ -64,9 +64,10 @@ Status calculate_quantized_multiplier_less_than_one(float multiplier,
         return Status{};
     }
 
-    const double q = std::frexp(multiplier, right_shift);
-    *right_shift *= -1;
-    auto q_fixed = static_cast<int64_t>(support::cpp11::round(q * fixed_point_one_Q0));
+    int          shift_exp = 0;
+    const double q         = std::frexp(multiplier, &shift_exp);
+    *right_shift           = -1 * shift_exp;
+    auto q_fixed           = static_cast<int64_t>(support::cpp11::round(q * fixed_point_one_Q0));
     ARM_COMPUTE_RETURN_ERROR_ON(q_fixed > fixed_point_one_Q0);
     if(q_fixed == fixed_point_one_Q0)
     {
@@ -80,15 +81,18 @@ Status calculate_quantized_multiplier_less_than_one(float multiplier,
     return Status{};
 }
 
-Status calculate_quantized_multiplier_greater_than_one(float multiplier,
-                                                       int *quantized_multiplier,
-                                                       int *left_shift)
+Status calculate_quantized_multiplier_greater_than_one(float    multiplier,
+                                                       int32_t *quantized_multiplier,
+                                                       int32_t *left_shift)
 {
     ARM_COMPUTE_RETURN_ERROR_ON(quantized_multiplier == nullptr);
     ARM_COMPUTE_RETURN_ERROR_ON(left_shift == nullptr);
     ARM_COMPUTE_RETURN_ERROR_ON(multiplier < 1.f);
-    const double q       = std::frexp(multiplier, left_shift);
-    auto         q_fixed = static_cast<int64_t>(support::cpp11::round(q * fixed_point_one_Q0));
+
+    int          shift_exp = 0;
+    const double q         = std::frexp(multiplier, &shift_exp);
+    *left_shift            = shift_exp;
+    auto q_fixed           = static_cast<int64_t>(support::cpp11::round(q * fixed_point_one_Q0));
     ARM_COMPUTE_RETURN_ERROR_ON(q_fixed > fixed_point_one_Q0);
     if(q_fixed == fixed_point_one_Q0)
     {
@@ -125,8 +129,8 @@ arm_compute::Status calculate_quantized_multipliers_less_than_one(const Quantiza
     for(unsigned int i = 0; i < size; ++i)
     {
         const float multiplier       = i_scale * w_scales[i] / o_scale;
-        int         quant_multiplier = 0;
-        int         quant_shift      = 0;
+        int32_t     quant_multiplier = 0;
+        int32_t     quant_shift      = 0;
         ARM_COMPUTE_RETURN_ON_ERROR(calculate_quantized_multiplier_less_than_one(multiplier, &quant_multiplier, &quant_shift));
         quant_multipliers[i] = quant_multiplier;
         quant_shifts[i]      = quant_shift;
@@ -181,8 +185,8 @@ void compute_quantized_multipliers_and_shifts(const ITensorInfo *input,
 
     for(unsigned int i = 0; i < num_filters; ++i)
     {
-        int         output_multiplier = 0;
-        int         output_shift      = 0;
+        int32_t     output_multiplier = 0;
+        int32_t     output_shift      = 0;
         const float multiplier        = iq_info.scale * wq_info.scale()[i] / oq_info.scale;
         calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 5e47dd56a..ca4fe732a 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -202,9 +202,9 @@ void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::
         const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
         const UniformQuantizationInfo oq_info = (output->info()->total_size() == 0) ? iq_info : output->info()->quantization_info().uniform();
 
-        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-        int   output_multiplier;
-        int   output_shift;
+        float   multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
+        int32_t output_multiplier;
+        int32_t output_shift;
         quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
         _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset);
         _accumulator.allocator()->allocate();
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index ee622f469..b3b90f859 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -255,9 +255,9 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
         const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
         const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
 
-        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-        int   output_multiplier;
-        int   output_shift;
+        float   multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
+        int32_t output_multiplier;
+        int32_t output_shift;
         quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
         _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, oq_info.offset);
         _gemmlowp_output.allocator()->allocate();
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index bb9620b29..0507c6b2b 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -33,7 +33,8 @@
 #include <set>
 #include <tuple>
 
-using namespace arm_compute;
+namespace arm_compute
+{
 using namespace arm_compute::misc::shape_calculator;
 
 NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights()
@@ -131,11 +132,11 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
         }
 
         // Merge activation with output stage
-        PixelValue type_min = 0;
-        PixelValue type_max = 0;
+        PixelValue type_min{};
+        PixelValue type_max{};
         std::tie(type_min, type_max) = get_min_max(data_type);
-        int min_activation = type_min.get<int>();
-        int max_activation = type_max.get<int>();
+        int32_t min_activation = type_min.get<int32_t>();
+        int32_t max_activation = type_max.get<int32_t>();
 
         if(supported_acts.count(act_info.activation()) != 0)
         {
@@ -190,11 +191,11 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         const UniformQuantizationInfo uoqinfo = oqinfo.uniform();
 
         // Merge activation with output stage
-        PixelValue type_min = 0;
-        PixelValue type_max = 0;
+        PixelValue type_min{};
+        PixelValue type_max{};
         std::tie(type_min, type_max) = get_min_max(data_type);
-        int min_activation = type_min.get<int>();
-        int max_activation = type_max.get<int>();
+        int32_t min_activation = type_min.get<int32_t>();
+        int32_t max_activation = type_max.get<int32_t>();
 
         const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
                                                                                    ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
@@ -595,3 +596,4 @@ void NEGEMMConvolutionLayer::prepare()
         _is_prepared = true;
     }
 }
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
index cfd996b53..cdfc03540 100644
--- a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
+++ b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
@@ -136,8 +136,8 @@ void NELSTMLayerQuantized::configure(const ITensor *input,
     _output_lowp.allocator()->init(TensorInfo(_output_highp.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_3));
 
     const float multiplier        = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
-    int         output_multiplier = 0;
-    int         output_shift      = 0;
+    int32_t     output_multiplier = 0;
+    int32_t     output_shift      = 0;
     quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
     _memory_group.manage(&_output_lowp);
@@ -342,8 +342,8 @@ Status NELSTMLayerQuantized::validate(const ITensorInfo *input,
     const TensorInfo output_lowp(output_highp.tensor_shape(), 1, DataType::QSYMM16, qsymm_3);
 
     const float multiplier        = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
-    int         output_multiplier = 0;
-    int         output_shift      = 0;
+    int32_t     output_multiplier = 0;
+    int32_t     output_shift      = 0;
     ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
 
     // _output_stage
diff --git a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
index c564e22d4..3235eee19 100644
--- a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
@@ -263,8 +263,8 @@ std::unique_ptr<depthwise::IDepthwiseConvolution> create_convolver(const ITensor
 
         // Calculate rescale parameters
         const float fmultipler  = iqinfo.scale * wqinfo.scale / oqinfo.scale;
-        int         qmultiplier = 0;
-        int         qshift      = 0;
+        int32_t     qmultiplier = 0;
+        int32_t     qshift      = 0;
         quantization::calculate_quantized_multiplier_less_than_one(fmultipler, &qmultiplier, &qshift);
         qasymm8::QAsymm8RescaleParams rescale_params(qshift, qmultiplier, fmultipler);
 
@@ -285,15 +285,15 @@ std::unique_ptr<depthwise::IDepthwiseConvolution> create_convolver(const ITensor
         const qasymm8::QAsymm8Params         oqinfo{ static_cast<uint8_t>(output_qinfo.offset), output_qinfo.scale };
 
         // Calculate rescale parameters
-        std::vector<float> fmultipliers;
-        std::vector<int>   qmultipliers;
-        std::vector<int>   qshifts;
+        std::vector<float>   fmultipliers;
+        std::vector<int32_t> qmultipliers;
+        std::vector<int32_t> qshifts;
 
         for(auto const s : wqinfo.scales)
         {
             const float fmultipler  = iqinfo.scale * s / oqinfo.scale;
-            int         qmultiplier = 0;
-            int         qshift      = 0;
+            int32_t     qmultiplier = 0;
+            int32_t     qshift      = 0;
             quantization::calculate_quantized_multiplier_less_than_one(fmultipler, &qmultiplier, &qshift);
             fmultipliers.push_back(fmultipler);
             qmultipliers.push_back(qmultiplier);