COMPMID-1554 Implementing Space to Batch on OpenCL - NHWC

Change-Id: Ifa37a6758f79d0a6ca771dcfb4c55a5d96b452d0
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148892
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>

3 files changed, 155 insertions, 17 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 1bc1036bed..a7a95b8b94 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -374,8 +374,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "softmax_layer_max_shift_exp_sum_quantized_serial", "softmax_layer_quantized.cl" },
     { "softmax_layer_max_shift_exp_sum_quantized_parallel", "softmax_layer_quantized.cl" },
     { "softmax_layer_max_shift_exp_sum_serial", "softmax_layer.cl" },
-    { "space_to_batch", "space_to_batch.cl" },
-    { "space_to_batch_static", "space_to_batch.cl" },
+    { "space_to_batch_nchw", "space_to_batch.cl" },
+    { "space_to_batch_static_nchw", "space_to_batch.cl" },
+    { "space_to_batch_nhwc", "space_to_batch.cl" },
+    { "space_to_batch_static_nhwc", "space_to_batch.cl" },
     { "softmax_layer_max_shift_exp_sum_parallel", "softmax_layer.cl" },
     { "strided_slice", "slice_ops.cl" },
     { "suppress_non_maximum", "canny.cl" },
diff --git a/src/core/CL/cl_kernels/space_to_batch.cl b/src/core/CL/cl_kernels/space_to_batch.cl
index 1343695ed1..1d641cc0f4 100644
--- a/src/core/CL/cl_kernels/space_to_batch.cl
+++ b/src/core/CL/cl_kernels/space_to_batch.cl
@@ -59,7 +59,7 @@
  * @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 image
  */
-__kernel void space_to_batch(
+__kernel void space_to_batch_nchw(
     TENSOR4D_DECLARATION(input),
     IMAGE_DECLARATION(paddings),
     VECTOR_DECLARATION(block_shape),
@@ -92,6 +92,74 @@ __kernel void space_to_batch(
         *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w));
     }
 }
+/** Calculate the space to batch conversion. (NHWC)
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The block shape tensor rank must be passed at compile time using -DBLOCK_SHAPE_DIM. e.g. -DBLOCK_SHAPE_DIM=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 image 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 image
+ * @param[in]  paddings_ptr                              Pointer to the second source image. Supported data types: S32
+ * @param[in]  paddings_stride_x                         Stride of the paddinds tensor in X dimension (in bytes)
+ * @param[in]  paddings_step_x                           paddings_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  paddings_stride_y                         Stride of the paddinds tensor in Y dimension (in bytes)
+ * @param[in]  paddings_step_y                           paddings_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  paddingse_offset_first_element_in_bytes   The offset of the first element in the second source image
+ * @param[in]  block_shape_ptr                           Pointer to the block shape tensor. Supported data types: S32
+ * @param[in]  block_shape_stride_x                      Stride of the block shape 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 block shape 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]  block_shape_offset_first_element_in_bytes The offset of the first element in the block shapetensor
+ * @param[in]  batch_id                                  The output 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 destination 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 image
+ */
+__kernel void space_to_batch_nhwc(
+    TENSOR4D_DECLARATION(input),
+    IMAGE_DECLARATION(paddings),
+    VECTOR_DECLARATION(block_shape),
+    const int batch_id,
+    TENSOR3D_DECLARATION(output))
+{
+    Tensor4D in    = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+    Image    pad   = CONVERT_TO_IMAGE_STRUCT_NO_STEP(paddings);
+    Vector   block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape);
+    Tensor3D out   = CONVERT_TO_TENSOR3D_STRUCT(output);
+
+    const int PAD_LEFT_X  = *((__global int *)offset(&pad, 0, 0));
+    const int PAD_RIGHT_X = *((__global int *)offset(&pad, 1, 0));
+    const int PAD_LEFT_Y  = *((__global int *)offset(&pad, 0, 1));
+    const int PAD_RIGHT_Y = *((__global int *)offset(&pad, 1, 1));
+
+    int block_x = *((__global int *)vector_offset(&block, 0));
+    int block_y = *((__global int *)vector_offset(&block, 1));
+
+    const int out_x = get_global_id(1);
+    const int out_y = get_global_id(2);
+    const int z     = get_global_id(0);
+
+    if((out_x >= PAD_LEFT_X && out_x < WIDTH_OUT - PAD_RIGHT_X) && (out_y >= PAD_LEFT_Y && out_y < HEIGHT_OUT - PAD_RIGHT_Y))
+    {
+        const int r                      = (BATCH_SIZE / (block_x * block_y));
+        const int w                      = batch_id % r;
+        const int in_x                   = (out_x - PAD_LEFT_X) * block_x + (batch_id / r) % block_x;
+        const int in_y                   = (out_y - PAD_LEFT_Y) * block_y + (batch_id / r) / block_x;
+        *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w));
+    }
+}
 #endif // defined(BATCH_SIZE) && defined(DATA_TYPE)
 
 #if defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y)
@@ -124,7 +192,7 @@ __kernel void space_to_batch(
  * @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 image
  */
-__kernel void space_to_batch_static(
+__kernel void space_to_batch_static_nchw(
     TENSOR4D_DECLARATION(input),
     const int batch_id,
     TENSOR3D_DECLARATION(output))
@@ -148,4 +216,57 @@ __kernel void space_to_batch_static(
         *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, w));
     }
 }
+/** Calculate the space to batch conversion. (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
+ * @note The starting pad value of x must be passed at compile time using -DPAD_LEFT_X. e.g. -DPAD_LEFT_X=2
+ * @note The ending pad value of x must be passed at compile time using -DPAD_RIGHT_X. e.g. -DPAD_RIGHT_X=2
+ * @note The starting pad value of y must be passed at compile time using -DPAD_LEFT_Y. e.g. -DPAD_LEFT_Y=2
+ * @note The ending pad value of  y must be passed at compile time using -DPAD_RIGHT_Y. e.g. -DPAD_RIGHT_X=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 image 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 image
+ * @param[in]  batch_id                             The output 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 image
+ */
+__kernel void space_to_batch_static_nhwc(
+    TENSOR4D_DECLARATION(input),
+    const int batch_id,
+    TENSOR3D_DECLARATION(output))
+{
+    Tensor4D in  = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
+
+    int block_x = BLOCK_SHAPE_X;
+    int block_y = *((__global int *)vector_offset(&block, 1));
+
+    const int out_x = get_global_id(1);
+    const int out_y = get_global_id(2);
+    const int z     = get_global_id(0);
+
+    if((out_x >= PAD_LEFT_X && out_x < WIDTH_OUT - PAD_RIGHT_X) && (out_y >= PAD_LEFT_Y && out_y < HEIGHT_OUT - PAD_RIGHT_Y))
+    {
+        const int r                      = (BATCH_SIZE / (block_x * block_y));
+        const int w                      = batch_id % r;
+        const int in_x                   = (out_x - PAD_LEFT_X) * block_x + (batch_id / r) % block_x;
+        const int in_y                   = (out_y - PAD_LEFT_Y) * block_y + (batch_id / r) / block_x;
+        *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w));
+    }
+}
 #endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y)
diff --git a/src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp b/src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp
index cda6e96806..9e4010e6c6 100644
--- a/src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp
+++ b/src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp
@@ -58,11 +58,16 @@ 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(output->tensor_shape()[0] < padding_left.x() + padding_right.y());
-        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[0] / block_shape_x != (output->tensor_shape()[0] - padding_left.x() - padding_right.y()));
-        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[1] / block_shape_y != (output->tensor_shape()[1] - padding_left.x() - padding_right.y()));
-        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(output->tensor_shape()[idx_width] < padding_left.x() + padding_right.y());
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_width] / block_shape_x != (output->tensor_shape()[idx_width] - padding_left.x() - padding_right.y()));
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_height] / block_shape_y != (output->tensor_shape()[idx_height] - padding_left.x() - padding_right.y()));
+        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_MISMATCHING_DATA_TYPES(input, output);
     }
 
@@ -85,13 +90,18 @@ void CLSpaceToBatchLayerKernel::configure(const ICLTensor *input, const ICLTenso
     _paddings    = paddings;
     _output      = output;
 
+    const DataLayout data_layout = input->info()->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_batch   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+
     // 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("-DWIDTH_OUT=" + support::cpp11::to_string(output->info()->dimension(0)));
-    build_opts.add_option("-DHEIGHT_OUT=" + support::cpp11::to_string(output->info()->dimension(1)));
-    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(output->info()->dimension(3)));
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("space_to_batch", build_opts.options()));
+    build_opts.add_option("-DWIDTH_OUT=" + support::cpp11::to_string(output->info()->dimension(idx_width)));
+    build_opts.add_option("-DHEIGHT_OUT=" + support::cpp11::to_string(output->info()->dimension(idx_height)));
+    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(output->info()->dimension(idx_batch)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("space_to_batch_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
 
     // Configure kernel window
     Window win = calculate_max_window(*output->info(), Steps());
@@ -111,19 +121,24 @@ void CLSpaceToBatchLayerKernel::configure(const ICLTensor *input, const int bloc
     _input  = input;
     _output = output;
 
+    const DataLayout data_layout = input->info()->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_batch   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+
     // 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("-DWIDTH_OUT=" + support::cpp11::to_string(output->info()->dimension(0)));
-    build_opts.add_option("-DHEIGHT_OUT=" + support::cpp11::to_string(output->info()->dimension(1)));
-    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(output->info()->dimension(3)));
+    build_opts.add_option("-DWIDTH_OUT=" + support::cpp11::to_string(output->info()->dimension(idx_width)));
+    build_opts.add_option("-DHEIGHT_OUT=" + support::cpp11::to_string(output->info()->dimension(idx_height)));
+    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(output->info()->dimension(idx_batch)));
     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("-DPAD_START_X=" + support::cpp11::to_string(padding_left.x()));
     build_opts.add_option("-DPAD_END_X=" + support::cpp11::to_string(padding_right.x()));
     build_opts.add_option("-DPAD_START_Y=" + support::cpp11::to_string(padding_left.y()));
     build_opts.add_option("-DPAD_END_Y=" + support::cpp11::to_string(padding_right.y()));
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("space_to_batch_static", build_opts.options()));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("space_to_batch_static_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
 
     // Configure kernel window
     Window win = calculate_max_window(*output->info(), Steps());