diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-04-11 15:59:10 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:49:37 +0000 |
commit | e52a3000d2c13bc1b66ca66b3d12b6b836982394 (patch) | |
tree | 70e8ef5ba216762604f84228805aac9bd65747b6 /src/core/CL/kernels/CLWinogradInputTransformKernel.cpp | |
parent | dd03870b63784abe499761da2b26b209b33f2db2 (diff) | |
download | ComputeLibrary-e52a3000d2c13bc1b66ca66b3d12b6b836982394.tar.gz |
COMPMID-1026 - Add support for 4x4 output tile in CLWinogradConvolutionLayer
The performance achieved can be found at the following confluence page:
https://confluence.arm.com/display/MLENG/GEMM-based+convolution+vs+Winograd-based+convolution+on+OpenCL
Change-Id: I4b690cfdd4eb4ff0cd17b14fdd49ccaa1d1dc85c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127729
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLWinogradInputTransformKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLWinogradInputTransformKernel.cpp | 19 |
1 files changed, 7 insertions, 12 deletions
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp index a47590d20f..febd22b04e 100644 --- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp +++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp @@ -47,7 +47,9 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c 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(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U), "Winograd input transform only supports 2x2 output tile for 3x3 kernels"); + 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_UNUSED(conv_info); ARM_COMPUTE_UNUSED(output_tile_size); @@ -111,7 +113,6 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor const int num_elements_y = input->info()->dimension(1) - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom(); // Check if we need to extend the right or bottom border - // FIXME: This actually is not needed. Added just for validating the result; const unsigned int extra_border_right = ((num_elements_x % output_tile_size.width) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.width - 1); const unsigned int extra_border_bottom = ((num_elements_y % output_tile_size.height) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.height - 1); @@ -137,19 +138,13 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor std::string kernel_name = "winograd_input_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string(); // Check optimized kernel if output_dims == 2x2 - if(output_tile_size.width == 2 && output_tile_size.height == 2) + if(output_tile_size == Size2D(2U, 2U)) { - if((_input->info()->dimension(2) % 2) != 0) - { - _step_z = 1; - } - else - { - _step_z = 2; - _lws_hint = cl::NDRange(1, 1, 8); - } + _step_z = (_input->info()->dimension(2) % 2) != 0 ? 1 : 2; } + _lws_hint = cl::NDRange(1, 1, 8); + // Append stepz and data layout kernel_name += "_stepz"; kernel_name += support::cpp11::to_string(_step_z); |