COMPMID-1710: Collapse window in CLDepthConvertKernel

Change-Id: I16589a2b3beb18e20b56059fdabccc61e26e3944
Reviewed-on: https://review.mlplatform.org/481
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 45 insertions, 23 deletions

diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl
index 7b03273b7..75192e6a9 100644
--- a/src/core/CL/cl_kernels/depth_convert.cl
+++ b/src/core/CL/cl_kernels/depth_convert.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,78 +38,92 @@
 
 /** This function performs a down-scaling depth conversion.
  *
- * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
+ * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
  *
  * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
  * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in_stride_y                       Stride of the source image in Y dimension (in bytes)
  * @param[in]  in_step_y                         in_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in_step_z                         in_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  in_offset_first_element_in_bytes  The offset of the first element in the source image
  * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
  * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
  * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
  * @param[in]  shift                             The integer shift amount value. Supported data types: S32
  */
 __kernel void convert_depth_down(
-    IMAGE_DECLARATION(in),
-    IMAGE_DECLARATION(out),
+    TENSOR3D_DECLARATION(in),
+    TENSOR3D_DECLARATION(out),
     const int shift)
 {
     // Get pixels pointer
-    Image in  = CONVERT_TO_IMAGE_STRUCT(in);
-    Image out = CONVERT_TO_IMAGE_STRUCT(out);
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(in);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
 
     // Load data
-    VEC_DATA_TYPE(DATA_TYPE_IN, 16)
-    in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
+    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in.ptr);
 
 #if defined(IS_DATA_TYPE_FLOAT)
-    vstore16(CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VSTORE(VEC_SIZE)
+    (CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
 #else  /* defined(IS_DATA_TYPE_FLOAT) */
-    vstore16(CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VSTORE(VEC_SIZE)
+    (CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
 #endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
 
 /** This function performs a up-scaling depth conversion.
  *
- * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
+ * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
  *
  * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
  * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in_stride_y                       Stride of the source image in Y dimension (in bytes)
  * @param[in]  in_step_y                         in_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in_step_z                         in_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  in_offset_first_element_in_bytes  The offset of the first element in the source image
  * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
  * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
  * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
  * @param[in]  shift                             The integer shift amount value. Supported data types: S32
  */
 __kernel void convert_depth_up(
-    IMAGE_DECLARATION(in),
-    IMAGE_DECLARATION(out),
+    TENSOR3D_DECLARATION(in),
+    TENSOR3D_DECLARATION(out),
     const int shift)
 {
     // Get pixels pointer
-    Image in  = CONVERT_TO_IMAGE_STRUCT(in);
-    Image out = CONVERT_TO_IMAGE_STRUCT(out);
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(in);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
 
     // Load data
-    VEC_DATA_TYPE(DATA_TYPE_IN, 16)
-    in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
+    in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in.ptr);
 
 #if defined(IS_DATA_TYPE_FLOAT)
-    vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VSTORE(VEC_SIZE)
+    (CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)out.ptr);
 #else  /* defined(IS_DATA_TYPE_FLOAT) */
-    vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VSTORE(VEC_SIZE)
+    (CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
 #endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
index b0c21624e..e188ee92a 100644
--- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -83,8 +83,12 @@ void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *out
     const size_t input_size  = data_size_from_type(input->info()->data_type());
     const size_t output_size = data_size_from_type(output->info()->data_type());
 
+    // Get number of elements to process per iterations
+    const unsigned int num_elems_processed_per_iteration = 16;
+
     // Set build options
     CLBuildOptions build_opts;
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
     // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
@@ -96,12 +100,16 @@ void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *out
     _kernel                       = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set shift arg
-    unsigned int idx = 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
+    unsigned int idx = 2 * num_arguments_per_3D_tensor(); //Skip the input and output parameters
     _kernel.setArg(idx++, shift);
 
     // Configure kernel
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
-    ICLSimple2DKernel::configure(input, output, num_elems_processed_per_iteration);
+    ICLSimple3DKernel::configure(input, output, num_elems_processed_per_iteration);
+
+    // Collapse window
+    const Window &full_window      = window();
+    Window        collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ);
+    ICLKernel::configure_internal(collapsed_window);
 }
 
 Status CLDepthConvertLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)