COMPMID-1201 - Implementing Winograd Convolution Layer 1x3 and 3x1 kernels on OpenCL

Change-Id: I39667bab49daa4da009694163274a59fd3574c73
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137595
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 49 insertions, 31 deletions

diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
index 779df637f6..e6c713e5e7 100644
--- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -25,7 +25,6 @@
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Helpers.h"
@@ -54,12 +53,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
     const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Winograd filter transform only supports 3x3 and 5x5 kernels");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && output_tile_size != Size2D(4U, 4U), "Winograd filter transform only supports 4x4 output tile for NHWC data layout");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U)
-                                    && output_tile_size != Size2D(4U, 4U),
-                                    "Winograd filter transform only supports 2x2 or 4x4 output tile for 3x3 kernels");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U), "Winograd filter transform only supports 4x4 output tile for 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, input->data_layout()), "Winograd filter transform not supported");
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != kernel_size.width || input->dimension(idx_h) != kernel_size.height);
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
 
@@ -115,6 +109,8 @@ void CLWinogradFilterTransformKernel::configure(const ICLTensor *input, ICLTenso
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL");
+    build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_FILTER_TRANSFORM_VERTICAL");
 
     const Size2D kernel_size      = winograd_info.kernel_size;
     const Size2D output_tile_size = winograd_info.output_tile_size;
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
index 274c9e7c3d..bb484afafb 100644
--- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -30,6 +30,7 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "support/ToolchainSupport.h"
 
@@ -45,12 +46,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        kernel_size      = winograd_info.kernel_size;
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Winograd input transform only supports 3x3 and 5x5 kernels");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && output_tile_size != Size2D(4U, 4U), "Winograd input transform only supports 4x4 output tile for NHWC data layout");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U)
-                                    && output_tile_size != Size2D(4U, 4U),
-                                    "Winograd input transform only supports 2x2 or 4x4 output tile for 3x3 kernels");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U), "Winograd input transform only supports 4x4 output tile for 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, input->data_layout()), "Winograd input transform not supported");
+
     ARM_COMPUTE_UNUSED(conv_info);
     ARM_COMPUTE_UNUSED(output_tile_size);
     ARM_COMPUTE_UNUSED(kernel_size);
@@ -131,8 +128,6 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     const int num_elements_x = input->info()->dimension(idx_w) - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
     const int num_elements_y = input->info()->dimension(idx_h) - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
 
-    _input  = input;
-    _output = output;
     if(input->info()->data_layout() == DataLayout::NCHW)
     {
         // Check if we need to extend the right or bottom border
@@ -145,8 +140,17 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     {
         _border_size = BorderSize(1U, 0U, 1U, 0);
     }
-    _num_tiles_x = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
-    _num_tiles_y = std::ceil(num_elements_y / static_cast<float>(output_tile_size.height));
+
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(input->info()->dimension(idx_w), input->info()->dimension(idx_h)),
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
+
+    _input       = input;
+    _output      = output;
+    _num_tiles_x = num_tiles.width;
+    _num_tiles_y = num_tiles.height;
 
     const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input->info(), winograd_info);
 
@@ -159,6 +163,10 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
     build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+    build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
+    build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+    build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
+    build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
 
     if(input->info()->data_layout() == DataLayout::NHWC)
     {
@@ -169,8 +177,11 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     // Create kernel
     std::string kernel_name = "winograd_input_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string();
 
+    // Get the maximum dimension from the tile size
+    const unsigned int tile_max_dim = std::max(output_tile_size.width, output_tile_size.height);
+
     // Check optimized kernel if output_dims == 2x2
-    if(output_tile_size == Size2D(2U, 2U))
+    if(tile_max_dim == 2)
     {
         _step_z = (_input->info()->dimension(2) % 2) != 0 ? 1 : 2;
     }
@@ -199,6 +210,8 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     _config_id += support::cpp11::to_string(conv_info.pad_left());
     _config_id += "_";
     _config_id += support::cpp11::to_string(conv_info.pad_top());
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(input->info()->data_layout()));
 }
 
 Status CLWinogradInputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
index 980498c4d1..40d5f6588f 100644
--- a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
@@ -55,20 +55,19 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        kernel_size      = winograd_info.kernel_size;
     const Size2D        input_dimensions = winograd_info.input_dimensions;
+    const unsigned int  num_channels     = (winograd_info.kernel_size.width + winograd_info.output_tile_size.width - 1) * (winograd_info.kernel_size.height + winograd_info.output_tile_size.height - 1);
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Only 3x3 and 5x5 kernels are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && output_tile_size != Size2D(4U, 4U), "Only 4x4 output tile supported for NHWC data layout");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size == Size2D(2U, 2U) && input->dimension(2) != 16, "Wrong number of batches");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size == Size2D(4U, 4U) && input->dimension(2) != 36, "Wrong number of batches");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(5U, 5U) && output_tile_size == Size2D(4U, 4U) && input->dimension(2) != 64, "Wrong number of batches");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, winograd_info.output_data_layout), "Winograd output transform not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->dimension(2) != num_channels, "Wrong number of channels");
 
     // Compute number of elements to process in the X and Y direction
-    const int num_elements_x = input_dimensions.width - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
-    const int num_elements_y = input_dimensions.height - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
-    const int num_tiles_x    = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
-    const int num_tiles_y    = std::ceil(num_elements_y / static_cast<float>(output_tile_size.height));
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
 
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != static_cast<unsigned int>((num_tiles_x * num_tiles_y)));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != static_cast<unsigned int>((num_tiles.area())));
 
     if(bias != nullptr)
     {
@@ -150,13 +149,21 @@ void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const IC
     const Size2D        kernel_size      = winograd_info.kernel_size;
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const PadStrideInfo conv_info        = winograd_info.convolution_info;
-    const int           num_elements_x   = input_dimensions.width - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
-    const int           num_tiles_x      = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
+
+    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+    const Size2D num_tiles = compute_winograd_convolution_tiles(input_dimensions,
+                                                                kernel_size,
+                                                                output_tile_size,
+                                                                conv_info);
 
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS"));
-    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles_x));
+    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
+    build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
+    build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+    build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
+    build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");
 
     // Create kernel
     std::string kernel_name = "winograd_output_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(winograd_info.output_data_layout));
@@ -179,6 +186,8 @@ void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const IC
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(winograd_info.output_data_layout));
 }
 
 Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info)