COMPMID-1549 Implementing Batch to Space on OpenCL - NHWC

Change-Id: If7ae0a8b6255a10711365068d9fb153c71f09818
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/147751
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 126 insertions, 14 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index fa164542e4..8f5e81bae9 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -153,8 +153,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
     { "arithmetic_div", "arithmetic_op.cl" },
-    { "batch_to_space", "batch_to_space.cl" },
-    { "batch_to_space_static", "batch_to_space.cl" },
+    { "batch_to_space_nchw", "batch_to_space.cl" },
+    { "batch_to_space_static_nchw", "batch_to_space.cl" },
+    { "batch_to_space_nhwc", "batch_to_space.cl" },
+    { "batch_to_space_static_nhwc", "batch_to_space.cl" },
     { "batchnormalization_layer_nchw", "batchnormalization_layer.cl" },
     { "batchnormalization_layer_nhwc", "batchnormalization_layer.cl" },
     { "bitwise_or", "bitwise_op.cl" },
diff --git a/src/core/CL/cl_kernels/batch_to_space.cl b/src/core/CL/cl_kernels/batch_to_space.cl
index 3043c2cf17..8506fc3709 100644
--- a/src/core/CL/cl_kernels/batch_to_space.cl
+++ b/src/core/CL/cl_kernels/batch_to_space.cl
@@ -24,7 +24,7 @@
 #include "helpers.h"
 
 #if defined(DATA_TYPE) && defined(BATCH_SIZE)
-/** Batch to space transformation.
+/** Batch to space transformation. (NCHW)
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
@@ -54,7 +54,7 @@
  * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
-__kernel void batch_to_space(
+__kernel void batch_to_space_nchw(
     TENSOR3D_DECLARATION(input),
     const int batch_id,
     VECTOR_DECLARATION(block_shape),
@@ -78,10 +78,64 @@ __kernel void batch_to_space(
 
     *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr);
 }
+/** Batch to space transformation. (NHWC)
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[in]  batch_id                             The input tensor batch id
+ * @param[in]  block_shape_ptr                      Pointer to the source tensor. Supported data types: S32
+ * @param[in]  block_shape_stride_x                 Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  block_shape_step_x                   block_shape_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  block_shape_stride_y                 Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  block_shape_step_y                   block_shape_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void batch_to_space_nhwc(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    VECTOR_DECLARATION(block_shape),
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in    = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out   = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+    Vector   block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape);
+
+    const int block_x = *((__global int *)vector_offset(&block, 0));
+    const int block_y = *((__global int *)vector_offset(&block, 1));
+
+    const int r = (BATCH_SIZE / (block_x * block_y));
+    const int x = get_global_id(1);
+    const int y = get_global_id(2);
+    const int z = get_global_id(0);
+    const int w = batch_id % r;
+
+    const int out_x = x * block_x + (batch_id / r) % block_x;
+    const int out_y = y * block_y + (batch_id / r) / block_x;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr);
+}
 #endif // defined(DATA_TYPE) && defined(BATCH_SIZE)
 
 #if defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y)
-/** Batch to space transformation.
+/** Batch to space transformation. (NCHW)
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
  * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
@@ -106,7 +160,7 @@ __kernel void batch_to_space(
  * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
-__kernel void batch_to_space_static(
+__kernel void batch_to_space_static_nchw(
     TENSOR3D_DECLARATION(input),
     const int batch_id,
     TENSOR4D_DECLARATION(output))
@@ -128,4 +182,51 @@ __kernel void batch_to_space_static(
 
     *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr);
 }
+/** Batch to space transformation. (NHWC)
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
+ * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2
+ * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[in]  batch_id                             The input tensor batch id
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void batch_to_space_static_nhwc(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    const int block_x = BLOCK_SHAPE_X;
+    const int block_y = BLOCK_SHAPE_Y;
+
+    const int r = (BATCH_SIZE / (block_x * block_y));
+    const int x = get_global_id(1);
+    const int y = get_global_id(2);
+    const int z = get_global_id(0);
+    const int w = batch_id % r;
+
+    const int out_x = x * block_x + (batch_id / r) % block_x;
+    const int out_y = y * block_y + (batch_id / r) / block_x;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr);
+}
 #endif // defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp b/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp
index e08d6f6ec5..8f56f66845 100644
--- a/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp
+++ b/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp
@@ -58,10 +58,15 @@ Status validate_arguments_static(const ITensorInfo *input, const int block_shape
     // Validate output if initialized
     if(output->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[0] != (block_shape_x * output->tensor_shape()[0]));
-        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[1] != (block_shape_x * output->tensor_shape()[1]));
-        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[2] != output->tensor_shape()[2]);
-        ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[3] % (block_shape_x * block_shape_y) != 0);
+        const DataLayout data_layout = input->data_layout();
+        const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+        const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+        const int        idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+        const int        idx_batch   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_width] != (block_shape_x * output->tensor_shape()[idx_width]));
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_height] != (block_shape_x * output->tensor_shape()[idx_height]));
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channel] != output->tensor_shape()[idx_channel]);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_batch] % (block_shape_x * block_shape_y) != 0);
         ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -84,12 +89,14 @@ void CLBatchToSpaceLayerKernel::configure(const ICLTensor *input, const ICLTenso
     _block_shape = block_shape;
     _output      = output;
 
+    const int idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(input->info()->dimension(3)));
-    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(0)));
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space", build_opts.options()));
+    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(idx_width)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
 
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps());
@@ -108,14 +115,16 @@ void CLBatchToSpaceLayerKernel::configure(const ICLTensor *input, const int32_t
     _input  = input;
     _output = output;
 
+    const int idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(input->info()->dimension(3)));
     build_opts.add_option("-DBLOCK_SHAPE_X=" + support::cpp11::to_string(block_shape_x));
     build_opts.add_option("-DBLOCK_SHAPE_Y=" + support::cpp11::to_string(block_shape_y));
-    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(0)));
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space_static", build_opts.options()));
+    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(idx_width)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space_static_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
 
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps());