From d28b751cf2ba9fcf4ccf294b31bf9d2ec5dfd8bb Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Fri, 6 Jul 2018 12:59:28 +0100 Subject: COMPMID-1340 - Implementing Winograd Convolution Layer 1x5/5x1 on OpenCL NHWC Change-Id: Id5e0795238f77c049df9c109dafc5ef878c1897d Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139234 Tested-by: Jenkins Reviewed-by: Giorgio Arena Reviewed-by: Anthony Barbier --- .../CL/kernels/CLWinogradFilterTransformKernel.h | 10 +- .../CL/kernels/CLWinogradInputTransformKernel.h | 10 +- .../CL/kernels/CLWinogradOutputTransformKernel.h | 10 +- .../CL/functions/CLWinogradConvolutionLayer.h | 6 +- .../CL/functions/CLWinogradInputTransform.h | 10 +- src/core/CL/CLHelpers.cpp | 4 +- src/core/CL/CLKernelLibrary.cpp | 6 + .../CL/cl_kernels/winograd_filter_transform.cl | 550 ++++++---- src/core/CL/cl_kernels/winograd_input_transform.cl | 1056 +++++++++++--------- .../CL/cl_kernels/winograd_output_transform.cl | 413 +++++--- tests/datasets/WinogradOutputTransformDataset.h | 28 + tests/validation/CL/Winograd.cpp | 18 +- 12 files changed, 1300 insertions(+), 821 deletions(-) diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h index 9d0833d695..62f55fa91e 100644 --- a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h +++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h @@ -54,8 +54,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd filter transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3) - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F32. @@ -71,8 +72,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd filter transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F32. diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h index 410e8ba765..517b348ffb 100644 --- a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h +++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h @@ -52,8 +52,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3) - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input The input tensor to transform. Data types supported: F32 @@ -69,8 +70,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input The input tensor to transform. Data types supported: F32 diff --git a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h index 0798172ba7..bab93de4b0 100644 --- a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h +++ b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h @@ -54,8 +54,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd output transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32. @@ -72,8 +73,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd output transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3) - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32. diff --git a/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h index 683aa79788..a24ae31d41 100644 --- a/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h @@ -59,8 +59,7 @@ public: CLWinogradConvolutionLayer &operator=(CLWinogradConvolutionLayer &&) = default; /** Set the input and output tensors. * - * @note: This function only works with 3x3,3x1,1x3,5x5,5x1 and 1x5 kernels along with unit strides for NCHW data layout - * @note: This function only works with 3x3, 3x1, 1x3 and 5x5 kernels along with unit strides for NHWC data layout + * @note: This function only works with 3x3,3x1,1x3,5x5,5x1 and 1x5 kernels along with unit strides for both NCHW and NHWC data layout * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true * * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], @@ -79,8 +78,7 @@ public: const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradConvolutionLayer * - * @note: This function only works with 3x3,3x1,1x3,5x5,5x1 and 1x5 kernels along with unit strides for NCHW data layout - * @note: This function only works with 3x3 and 5x5 kernels along with unit strides for NHWC data layout + * @note: This function only works with 3x3,3x1,1x3,5x5,5x1 and 1x5 kernels along with unit strides for both NCHW and NHWC data layout * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true * * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], diff --git a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h index 1f89455aee..8ea25a116a 100644 --- a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h +++ b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h @@ -45,8 +45,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input The input tensor to transform. Data types supported: F32 @@ -62,8 +63,9 @@ public: * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) * * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3) - * F(4x4, 5x5) + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * * Strides: only unit strides * * @param[in] input The input tensor to transform. Data types supported: F32 diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp index 3965be76fd..0e83ff20c8 100644 --- a/src/core/CL/CLHelpers.cpp +++ b/src/core/CL/CLHelpers.cpp @@ -172,7 +172,9 @@ bool cl_winograd_convolution_layer_supported(const Size2D &output_tile, const Si WinogradConfiguration(std::pair(1, 4), std::pair(1, 3)), WinogradConfiguration(std::pair(4, 1), std::pair(3, 1)), WinogradConfiguration(std::pair(4, 4), std::pair(3, 3)), - WinogradConfiguration(std::pair(4, 4), std::pair(5, 5)) + WinogradConfiguration(std::pair(4, 4), std::pair(5, 5)), + WinogradConfiguration(std::pair(4, 1), std::pair(5, 1)), + WinogradConfiguration(std::pair(1, 4), std::pair(1, 5)) }; auto p = std::make_pair(std::pair(output_tile.width, output_tile.height), diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 03731ee93c..9a3ebc0021 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -380,6 +380,8 @@ const std::map CLKernelLibrary::_kernel_program_map = { "winograd_filter_transform_1x4_1x3_nhwc", "winograd_filter_transform.cl" }, { "winograd_filter_transform_4x4_3x3_nhwc", "winograd_filter_transform.cl" }, { "winograd_filter_transform_4x4_5x5_nhwc", "winograd_filter_transform.cl" }, + { "winograd_filter_transform_4x1_5x1_nhwc", "winograd_filter_transform.cl" }, + { "winograd_filter_transform_1x4_1x5_nhwc", "winograd_filter_transform.cl" }, { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd_input_transform.cl" }, { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd_input_transform.cl" }, { "winograd_input_transform_2x1_3x1_stepz1_nchw", "winograd_input_transform.cl" }, @@ -396,6 +398,8 @@ const std::map CLKernelLibrary::_kernel_program_map = { "winograd_input_transform_1x4_1x3_stepz1_nhwc", "winograd_input_transform.cl" }, { "winograd_input_transform_4x4_3x3_stepz1_nhwc", "winograd_input_transform.cl" }, { "winograd_input_transform_4x4_5x5_stepz1_nhwc", "winograd_input_transform.cl" }, + { "winograd_input_transform_4x1_5x1_stepz1_nhwc", "winograd_input_transform.cl" }, + { "winograd_input_transform_1x4_1x5_stepz1_nhwc", "winograd_input_transform.cl" }, { "winograd_output_transform_2x2_3x3_nchw", "winograd_output_transform.cl" }, { "winograd_output_transform_2x1_3x1_nchw", "winograd_output_transform.cl" }, { "winograd_output_transform_1x2_1x3_nchw", "winograd_output_transform.cl" }, @@ -409,6 +413,8 @@ const std::map CLKernelLibrary::_kernel_program_map = { "winograd_output_transform_1x4_1x3_nhwc", "winograd_output_transform.cl" }, { "winograd_output_transform_4x4_3x3_nhwc", "winograd_output_transform.cl" }, { "winograd_output_transform_4x4_5x5_nhwc", "winograd_output_transform.cl" }, + { "winograd_output_transform_4x1_5x1_nhwc", "winograd_output_transform.cl" }, + { "winograd_output_transform_1x4_1x5_nhwc", "winograd_output_transform.cl" }, { "YUYV422_to_IYUV_bt709", "color_convert.cl" }, { "YUYV422_to_NV12_bt709", "color_convert.cl" }, { "YUYV422_to_RGB888_bt709", "color_convert.cl" }, diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl index e53da9b278..73da005996 100644 --- a/src/core/CL/cl_kernels/winograd_filter_transform.cl +++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl @@ -442,105 +442,6 @@ __kernel void winograd_filter_transform_4x4_3x3_nhwc( #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } -#if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) -/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1 - * - * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 - * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_filter_transform_4x1_3x1_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_filter_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_stride_w, - src_step_w, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) - -#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) -/** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4 - * - * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 - * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_filter_transform_1x4_1x3_nhwc( - TENSOR4D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_filter_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_stride_w, - src_step_w, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) /** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NCHW and the output tile is 4x4/4x1 or 1x4 * * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 @@ -827,9 +728,11 @@ __kernel void winograd_filter_transform_4x4_5x5_nchw( #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } -/** This OpenCL kernel performs Winograd filter transform 5x5 when the data layout is NHWC and the output tile is 4x4 +/** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NHWC and the output tile is 4x4/4x1 or 1x4 * * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time + * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) @@ -858,12 +761,23 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(float) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w; +#if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Load the values from the input tensor - float w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y)); - float w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y)); - float w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y)); - float w03 = *((__global float *)(src_addr + 0 * src_stride_z + 3 * src_stride_y)); - float w04 = *((__global float *)(src_addr + 0 * src_stride_z + 4 * src_stride_y)); + float w00 = *((__global float *)(src_addr + 0 * src_stride_z)); + float w01 = *((__global float *)(src_addr + 1 * src_stride_z)); + float w02 = *((__global float *)(src_addr + 2 * src_stride_z)); + float w03 = *((__global float *)(src_addr + 3 * src_stride_z)); + float w04 = *((__global float *)(src_addr + 4 * src_stride_z)); +#else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) + // Load the values from the input tensor + float w00 = *((__global float *)(src_addr + 0 * src_stride_y)); + float w01 = *((__global float *)(src_addr + 1 * src_stride_y)); + float w02 = *((__global float *)(src_addr + 2 * src_stride_y)); + float w03 = *((__global float *)(src_addr + 3 * src_stride_y)); + float w04 = *((__global float *)(src_addr + 4 * src_stride_y)); +#endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) + +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y)); float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y)); float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y)); @@ -884,128 +798,129 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( float w42 = *((__global float *)(src_addr + 4 * src_stride_z + 2 * src_stride_y)); float w43 = *((__global float *)(src_addr + 4 * src_stride_z + 3 * src_stride_y)); float w44 = *((__global float *)(src_addr + 4 * src_stride_z + 4 * src_stride_y)); - - // Transform the 3x3 tile in a 8x8 tile - float8 out0 = 0.0f; - float8 out1 = 0.0f; - float8 out2 = 0.0f; - float8 out3 = 0.0f; - float8 out4 = 0.0f; - float8 out5 = 0.0f; - float8 out6 = 0.0f; - float8 out7 = 0.0f; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Row 0 - out0.s0 = w00; - out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f; - out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f; - out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f; - out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f; - out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f; - out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f; - out0.s7 = w04; + float8 out0 = 0.0f; + out0.s0 = w00; + out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f; + out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f; + out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f; + out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f; + out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f; + out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f; + out0.s7 = w04; +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) // Row 1 - out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f; - out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; - out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; - out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * - (w04 + w14 + w24 + w34 + w44)) / 405.f; - out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * - (w04 + w14 + w24 + w34 + w44)) / 405.f; - out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) + - (w04 + w14 + w24 + w34 + w44)) / 810.f; - out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) + - (w04 + w14 + w24 + w34 + w44)) / 810.f; - out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f; + float8 out1 = 0.0f; + out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f; + out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; + out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; + out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * + (w04 + w14 + w24 + w34 + w44)) / 405.f; + out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * + (w04 + w14 + w24 + w34 + w44)) / 405.f; + out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) + + (w04 + w14 + w24 + w34 + w44)) / 810.f; + out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) + + (w04 + w14 + w24 + w34 + w44)) / 810.f; + out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f; // Row 2 - out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f; - out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; - out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; - out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * - (w04 - w14 + w24 - w34 + w44)) / 405.f; - out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * - (w04 - w14 + w24 - w34 + w44)) / 405.f; - out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) + - (w04 - w14 + w24 - w34 + w44)) / 810.f; - out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) + - (w04 - w14 + w24 - w34 + w44)) / 810.f; - out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f; + float8 out2 = 0.0f; + out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f; + out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; + out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; + out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * + (w04 - w14 + w24 - w34 + w44)) / 405.f; + out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * + (w04 - w14 + w24 - w34 + w44)) / 405.f; + out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) + + (w04 - w14 + w24 - w34 + w44)) / 810.f; + out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) + + (w04 - w14 + w24 - w34 + w44)) / 810.f; + out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f; // Row 3 - out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f; - out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + - (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; - out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - - (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; - out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f - * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; - out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f - * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; - out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * - (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; - out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * - (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; - out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f; + float8 out3 = 0.0f; + out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f; + out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + + (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; + out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - + (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; + out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; + out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; + out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; + out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; + out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f; // Row 4 - out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f; - out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + - (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; - out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - - (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; - out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f - * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; - out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f - * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; - out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * - (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; - out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * - (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; - out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f; + float8 out4 = 0.0f; + out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f; + out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + + (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; + out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - + (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; + out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; + out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; + out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; + out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; + out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f; // Row 5 - out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f; - out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + - (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; - out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - - (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; - out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f - * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; - out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f - * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; - out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * - (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; - out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * - (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; - out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f; + float8 out5 = 0.0f; + out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f; + out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + + (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; + out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - + (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; + out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; + out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; + out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; + out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; + out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f; // Row 6 - out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f; - out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + - (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; - out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - - (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; - out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f - * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; - out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f - * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; - out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * - (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; - out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * - (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; - out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f; + float8 out6 = 0.0f; + out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f; + out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + + (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; + out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - + (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; + out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; + out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; + out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; + out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; + out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f; // Row 7 - out7.s0 = w40; - out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f; - out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f; - out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f; - out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f; - out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f; - out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f; - out7.s7 = w44; + float8 out7 = 0.0f; + out7.s0 = w40; + out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f; + out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f; + out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f; + out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f; + out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f; + out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f; + out7.s7 = w44; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) int x0 = get_global_id(2); // idx filter int y0 = get_global_id(0); // idx channel @@ -1013,15 +928,17 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( // Get output address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y; - // Store the 64 values across the 64 channels - *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; - *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; - *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; - *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; - *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; - *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; - *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; - *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; + // Store the values across the channels + *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; + *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; + *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; + *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; + *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; + *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; + *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; + *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; + +#if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) *(__global float *)(dst_addr + 8 * dst_stride_z) = out1.s0; *(__global float *)(dst_addr + 9 * dst_stride_z) = out1.s1; *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2; @@ -1078,6 +995,7 @@ __kernel void winograd_filter_transform_4x4_5x5_nhwc( *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5; *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6; *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7; +#endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) } #endif // defined(SRC_DIM_Z) @@ -1225,6 +1143,102 @@ __kernel void winograd_filter_transform_4x1_5x1_nchw( dst_step_z, dst_offset_first_element_in_bytes); } + +/** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_filter_transform_4x1_3x1_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel performs Winograd filter transform 5x1 when the data layout is NHWC and the output tile is 4x1 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_filter_transform_4x1_5x1_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) @@ -1371,4 +1385,100 @@ __kernel void winograd_filter_transform_1x4_1x5_nchw( dst_step_z, dst_offset_first_element_in_bytes); } + +/** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_filter_transform_1x4_1x3_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel performs Winograd filter transform 1x5 when the data layout is NHWC and the output tile is 1x4 + * + * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 + * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_filter_transform_1x4_1x5_nhwc( + TENSOR4D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_filter_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_stride_w, + src_step_w, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) diff --git a/src/core/CL/cl_kernels/winograd_input_transform.cl b/src/core/CL/cl_kernels/winograd_input_transform.cl index 01cbc84ff3..fcd1b3b9ce 100644 --- a/src/core/CL/cl_kernels/winograd_input_transform.cl +++ b/src/core/CL/cl_kernels/winograd_input_transform.cl @@ -23,6 +23,26 @@ */ #include "helpers.h" +#define OUTPUT_ROW_4x4_5x5(out, tmp, comm_fact) \ + ({ \ + comm_fact.s0 = tmp.s2 - 4.25f * tmp.s4 + tmp.s6; \ + comm_fact.s1 = tmp.s1 - 4.25f * tmp.s3 + tmp.s5; \ + comm_fact.s2 = 2.5f * tmp.s3; \ + comm_fact.s3 = 0.5f * tmp.s1 + 2.f * tmp.s5 - comm_fact.s2; \ + comm_fact.s4 = 0.25f * tmp.s2 - 1.25f * tmp.s4 + tmp.s6; \ + comm_fact.s5 = 4.f * tmp.s2 + tmp.s6 - 5.f * tmp.s4; \ + comm_fact.s6 = 2.f * tmp.s1 + 0.5f * tmp.s5 - comm_fact.s2; \ + \ + out.s0 = tmp.s0 - tmp.s6 + 5.25f * tmp.s4 - 5.25f * tmp.s2; \ + out.s1 = comm_fact.s0 + comm_fact.s1; \ + out.s2 = comm_fact.s0 - comm_fact.s1; \ + out.s3 = comm_fact.s3 + comm_fact.s4; \ + out.s4 = comm_fact.s4 - comm_fact.s3; \ + out.s5 = comm_fact.s5 + comm_fact.s6; \ + out.s6 = comm_fact.s5 - comm_fact.s6; \ + out.s7 = tmp.s7 - tmp.s1 + 5.25f * tmp.s3 - 5.25f * tmp.s5; \ + }) + #if defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H) /** This OpenCL kernel computes the input transform when the kernel size is 3x3/3x1 or 1x3 and the output tile is 2x2/2x1 or 1x2 * @@ -936,132 +956,14 @@ __kernel void winograd_input_transform_4x4_3x3_stepz1_nhwc( #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) } -#if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) -/** This OpenCL kernel computes the input transform when the kernel size is 3x1 and the output tile is 4x1 for data layout NHWC +/** This OpenCL kernel computes the input transform when the kernel size is 5x5/5x1 or 1x5 and the output tile is 4x4/4x1 or 1x4 when the data layout is NHWC * * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). + * @note Dimension one of the input tensor (width for NHWC data layout) must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM_1=112) + * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 - * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 - * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source image. Supported data types: F32 - * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_input_transform_4x1_3x1_stepz1_nhwc( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_input_transform_4x4_3x3_stepz1_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) - -#if defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) -/** This OpenCL kernel computes the input transform when the kernel size is 1x3 and the output tile is 1x4 for data layout NHWC - * - * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). - * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). - * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 - * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source image. Supported data types: F32 - * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_input_transform_1x4_1x3_stepz1_nhwc( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - winograd_input_transform_4x4_3x3_stepz1_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes); -} -#endif // defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) - -#endif // defined(SRC_DIM_1) && defined(SRC_DIM_2) - -#define OUTPUT_ROW_4x4_5x5(out, tmp, comm_fact) \ - ({ \ - comm_fact.s0 = tmp.s2 - 4.25f * tmp.s4 + tmp.s6; \ - comm_fact.s1 = tmp.s1 - 4.25f * tmp.s3 + tmp.s5; \ - comm_fact.s2 = 2.5f * tmp.s3; \ - comm_fact.s3 = 0.5f * tmp.s1 + 2.f * tmp.s5 - comm_fact.s2; \ - comm_fact.s4 = 0.25f * tmp.s2 - 1.25f * tmp.s4 + tmp.s6; \ - comm_fact.s5 = 4.f * tmp.s2 + tmp.s6 - 5.f * tmp.s4; \ - comm_fact.s6 = 2.f * tmp.s1 + 0.5f * tmp.s5 - comm_fact.s2; \ - \ - out.s0 = tmp.s0 - tmp.s6 + 5.25f * tmp.s4 - 5.25f * tmp.s2; \ - out.s1 = comm_fact.s0 + comm_fact.s1; \ - out.s2 = comm_fact.s0 - comm_fact.s1; \ - out.s3 = comm_fact.s3 + comm_fact.s4; \ - out.s4 = comm_fact.s4 - comm_fact.s3; \ - out.s5 = comm_fact.s5 + comm_fact.s6; \ - out.s6 = comm_fact.s5 - comm_fact.s6; \ - out.s7 = tmp.s7 - tmp.s1 + 5.25f * tmp.s3 - 5.25f * tmp.s5; \ - }) - -/** This OpenCL kernel computes the input transform when the kernel size is 5x5/5x1 or 1x5 and the output tile is 4x4/4x1 or 1x4 when the data layout is NCHW - * - * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). - * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). - * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2 - * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2 * @note If this kernel is used to perform Winograd input transform 5x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time * @note If this kernel is used to perform Winograd input transform 1x5, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time * @@ -1082,7 +984,7 @@ __kernel void winograd_input_transform_1x4_1x3_stepz1_nhwc( * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor */ -__kernel void winograd_input_transform_4x4_5x5_stepz1_nchw( +__kernel void winograd_input_transform_4x4_5x5_stepz1_nhwc( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst)) { @@ -1091,104 +993,433 @@ __kernel void winograd_input_transform_4x4_5x5_stepz1_nchw( int z = get_global_id(2); // Compute input address - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; - - src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y); + __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float); - // Load input tile #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) - const float8 in_row0 = vload8(0, (__global float *)(src_addr)); -#elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) - const float8 in_row0 = (float8)(*((__global float *)(src_addr + 0 * src_stride_y)), - *((__global float *)(src_addr + 1 * src_stride_y)), - *((__global float *)(src_addr + 2 * src_stride_y)), - *((__global float *)(src_addr + 3 * src_stride_y)), - *((__global float *)(src_addr + 4 * src_stride_y)), - *((__global float *)(src_addr + 5 * src_stride_y)), - *((__global float *)(src_addr + 6 * src_stride_y)), - *((__global float *)(src_addr + 7 * src_stride_y))); -#else // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) - const float8 in_row0 = vload8(0, (__global float *)(src_addr + 0 * src_stride_y)); - const float8 in_row1 = vload8(0, (__global float *)(src_addr + 1 * src_stride_y)); - const float8 in_row2 = vload8(0, (__global float *)(src_addr + 2 * src_stride_y)); - const float8 in_row3 = vload8(0, (__global float *)(src_addr + 3 * src_stride_y)); - const float8 in_row4 = vload8(0, (__global float *)(src_addr + 4 * src_stride_y)); - const float8 in_row5 = vload8(0, (__global float *)(src_addr + 5 * src_stride_y)); - const float8 in_row6 = vload8(0, (__global float *)(src_addr + 6 * src_stride_y)); - const float8 in_row7 = vload8(0, (__global float *)(src_addr + 7 * src_stride_y)); -#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + // Clamp coordinates. This clamp is valid for all rows + int8 y_coord = (int8)(y * OUTPUT_TILE_W) + (int8)(0, 1, 2, 3, 4, 5, 6, 7) - (int8)PAD_LEFT; + y_coord = clamp(y_coord, -1, SRC_DIM_1); + + // Row0 + // We can skip the border clamping along the z dimension as we cannot read out-of-bound in case of 5x1 kernels + int z_coord = z * OUTPUT_TILE_H; + + // Load the input tile + float8 in_row0; + in_row0.s0 = *(__global float *)(src_addr + y_coord.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s1 = *(__global float *)(src_addr + y_coord.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s2 = *(__global float *)(src_addr + y_coord.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s3 = *(__global float *)(src_addr + y_coord.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s4 = *(__global float *)(src_addr + y_coord.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s5 = *(__global float *)(src_addr + y_coord.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s6 = *(__global float *)(src_addr + y_coord.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s7 = *(__global float *)(src_addr + y_coord.s7 * (int)src_stride_y + z_coord * src_stride_z); // Calculate common factors for intermediate tensor - float8 tmp0 = in_row0; float8 comm_fact0 = 0.0f; + float8 tmp0 = in_row0; -#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) - comm_fact0 += in_row2 + in_row6 - 4.25f * in_row4; - tmp0 += -in_row6 + 5.25f * in_row4 - 5.25f * in_row2; - - float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3; - float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6; + float8 out0 = (float8)0.0f; - const float8 tmp1 = comm_fact0 + comm_fact1; - const float8 tmp2 = comm_fact0 - comm_fact1; + OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); - comm_fact0 = 2.5f * in_row3; - comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5; +#elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) + // We can skip the border clamping along the y dimension as we cannot read out-of-bound in case of 1x5 kernels + int y_coord = y * OUTPUT_TILE_W; - const float8 tmp3 = comm_fact1 + comm_fact2; - const float8 tmp4 = comm_fact2 - comm_fact1; + // Row0 + // We can skip the border clamping along the z dimension as we cannot read out-of-bound in case of 5x1 kernels + int8 z_coord = (int8)(z * OUTPUT_TILE_H) + (int8)(0, 1, 2, 3, 4, 5, 6, 7) - (int8)PAD_TOP; + int8 valid_y = select((int8)y_coord, (int8) - 1, z_coord < (int8)0); // If z < 0, set y to -1 + valid_y = select(valid_y, SRC_DIM_1, z_coord >= (int8)SRC_DIM_2); // If z >= SRC_DIM_2, set y to SRC_DIM_2 + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); // Clamp z coordinate - comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5; - comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6; + // Load the input tile + float8 in_row0; + in_row0.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord.s0 * src_stride_z); + in_row0.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord.s1 * src_stride_z); + in_row0.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord.s2 * src_stride_z); + in_row0.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord.s3 * src_stride_z); + in_row0.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord.s4 * src_stride_z); + in_row0.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord.s5 * src_stride_z); + in_row0.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord.s6 * src_stride_z); + in_row0.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord.s7 * src_stride_z); - const float8 tmp5 = comm_fact1 + comm_fact2; - const float8 tmp6 = comm_fact2 - comm_fact1; - const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5; -#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + // Calculate common factors for intermediate tensor + float8 comm_fact0 = 0.0f; + float8 tmp0 = in_row0; - // Calculate output rows (reuse comm_fact0 vector) - float8 out0; + float8 out0 = (float8)0.0f; OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); +#else // defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) + float8 in_row0, in_row1, in_row2, in_row3, in_row4, in_row5, in_row6, in_row7; -#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) - float8 out1, out2, out3, out4, out5, out6, out7; + // Clamp coordinates. This clamp is valid for all rows + int8 y_coord = (int8)(y * OUTPUT_TILE_W) + (int8)(0, 1, 2, 3, 4, 5, 6, 7) - (int8)PAD_LEFT; + y_coord = clamp(y_coord, -1, SRC_DIM_1); - OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0); - OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0); - OUTPUT_ROW_4x4_5x5(out3, tmp3, comm_fact0); - OUTPUT_ROW_4x4_5x5(out4, tmp4, comm_fact0); - OUTPUT_ROW_4x4_5x5(out5, tmp5, comm_fact0); - OUTPUT_ROW_4x4_5x5(out6, tmp6, comm_fact0); - OUTPUT_ROW_4x4_5x5(out7, tmp7, comm_fact0); -#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + // Row0 + int z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 0; + int8 valid_y = select(y_coord, -1, (int8)z_coord < 0); // If z < 0, set y to -1 + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); // If z >= SRC_DIM_2, set y to SRC_DIM_2 + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); // Clamp z coordinate - // Store values across the channels - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y; + // Load the input tile + in_row0.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row0.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0; - *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1; - *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2; - *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3; - *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4; - *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5; - *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6; - *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7; + // Row1 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 1; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); -#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) - *((__global float *)(dst_addr + 8 * dst_stride_z)) = out1.s0; - *((__global float *)(dst_addr + 9 * dst_stride_z)) = out1.s1; - *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2; - *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3; - *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4; - *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5; - *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6; - *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7; - *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0; - *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1; - *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2; - *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3; - *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4; + in_row1.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row1.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row2 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 2; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row2.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row2.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row3 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 3; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row3.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row3.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row4 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 4; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row4.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row4.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row5 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 5; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row5.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row5.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row6 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 6; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row6.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row6.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + // Row7 + z_coord = (z * OUTPUT_TILE_H) - PAD_TOP + 7; + valid_y = select(y_coord, -1, (int8)z_coord < 0); + valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); + z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); + + in_row7.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); + in_row7.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); + + float8 comm_fact0 = in_row2 + in_row6 - 4.25f * in_row4; + float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3; + float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6; + + // Calculate intermediate tensor and reuse common factor vectors + const float8 tmp0 = in_row0 - in_row6 + 5.25f * in_row4 - 5.25f * in_row2; + const float8 tmp1 = comm_fact0 + comm_fact1; + const float8 tmp2 = comm_fact0 - comm_fact1; + + comm_fact0 = 2.5f * in_row3; + comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5; + + const float8 tmp3 = comm_fact1 + comm_fact2; + const float8 tmp4 = comm_fact2 - comm_fact1; + + comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5; + comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6; + + const float8 tmp5 = comm_fact1 + comm_fact2; + const float8 tmp6 = comm_fact2 - comm_fact1; + const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5; + + // Calculate output rows (reuse comm_fact0 vector) + float8 out0, out1, out2, out3, out4, out5, out6, out7; + OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); + OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0); + OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0); + OUTPUT_ROW_4x4_5x5(out3, tmp3, comm_fact0); + OUTPUT_ROW_4x4_5x5(out4, tmp4, comm_fact0); + OUTPUT_ROW_4x4_5x5(out5, tmp5, comm_fact0); + OUTPUT_ROW_4x4_5x5(out6, tmp6, comm_fact0); + OUTPUT_ROW_4x4_5x5(out7, tmp7, comm_fact0); +#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + + // Store values across the channels + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(float) + (y + z * (int)NUM_TILES_X) * dst_stride_y; + + *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0; + *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1; + *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2; + *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3; + *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4; + *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5; + *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6; + *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7; + +#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + *((__global float *)(dst_addr + 8 * dst_stride_z)) = out1.s0; + *((__global float *)(dst_addr + 9 * dst_stride_z)) = out1.s1; + *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2; + *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3; + *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4; + *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5; + *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6; + *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7; + *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0; + *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1; + *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2; + *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3; + *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4; + *((__global float *)(dst_addr + 21 * dst_stride_z)) = out2.s5; + *((__global float *)(dst_addr + 22 * dst_stride_z)) = out2.s6; + *((__global float *)(dst_addr + 23 * dst_stride_z)) = out2.s7; + *((__global float *)(dst_addr + 24 * dst_stride_z)) = out3.s0; + *((__global float *)(dst_addr + 25 * dst_stride_z)) = out3.s1; + *((__global float *)(dst_addr + 26 * dst_stride_z)) = out3.s2; + *((__global float *)(dst_addr + 27 * dst_stride_z)) = out3.s3; + *((__global float *)(dst_addr + 28 * dst_stride_z)) = out3.s4; + *((__global float *)(dst_addr + 29 * dst_stride_z)) = out3.s5; + *((__global float *)(dst_addr + 30 * dst_stride_z)) = out3.s6; + *((__global float *)(dst_addr + 31 * dst_stride_z)) = out3.s7; + *((__global float *)(dst_addr + 32 * dst_stride_z)) = out4.s0; + *((__global float *)(dst_addr + 33 * dst_stride_z)) = out4.s1; + *((__global float *)(dst_addr + 34 * dst_stride_z)) = out4.s2; + *((__global float *)(dst_addr + 35 * dst_stride_z)) = out4.s3; + *((__global float *)(dst_addr + 36 * dst_stride_z)) = out4.s4; + *((__global float *)(dst_addr + 37 * dst_stride_z)) = out4.s5; + *((__global float *)(dst_addr + 38 * dst_stride_z)) = out4.s6; + *((__global float *)(dst_addr + 39 * dst_stride_z)) = out4.s7; + *((__global float *)(dst_addr + 40 * dst_stride_z)) = out5.s0; + *((__global float *)(dst_addr + 41 * dst_stride_z)) = out5.s1; + *((__global float *)(dst_addr + 42 * dst_stride_z)) = out5.s2; + *((__global float *)(dst_addr + 43 * dst_stride_z)) = out5.s3; + *((__global float *)(dst_addr + 44 * dst_stride_z)) = out5.s4; + *((__global float *)(dst_addr + 45 * dst_stride_z)) = out5.s5; + *((__global float *)(dst_addr + 46 * dst_stride_z)) = out5.s6; + *((__global float *)(dst_addr + 47 * dst_stride_z)) = out5.s7; + *((__global float *)(dst_addr + 48 * dst_stride_z)) = out6.s0; + *((__global float *)(dst_addr + 49 * dst_stride_z)) = out6.s1; + *((__global float *)(dst_addr + 50 * dst_stride_z)) = out6.s2; + *((__global float *)(dst_addr + 51 * dst_stride_z)) = out6.s3; + *((__global float *)(dst_addr + 52 * dst_stride_z)) = out6.s4; + *((__global float *)(dst_addr + 53 * dst_stride_z)) = out6.s5; + *((__global float *)(dst_addr + 54 * dst_stride_z)) = out6.s6; + *((__global float *)(dst_addr + 55 * dst_stride_z)) = out6.s7; + *((__global float *)(dst_addr + 56 * dst_stride_z)) = out7.s0; + *((__global float *)(dst_addr + 57 * dst_stride_z)) = out7.s1; + *((__global float *)(dst_addr + 58 * dst_stride_z)) = out7.s2; + *((__global float *)(dst_addr + 59 * dst_stride_z)) = out7.s3; + *((__global float *)(dst_addr + 60 * dst_stride_z)) = out7.s4; + *((__global float *)(dst_addr + 61 * dst_stride_z)) = out7.s5; + *((__global float *)(dst_addr + 62 * dst_stride_z)) = out7.s6; + *((__global float *)(dst_addr + 63 * dst_stride_z)) = out7.s7; +#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) +} +#endif // defined(SRC_DIM_1) && defined(SRC_DIM_2) + +/** This OpenCL kernel computes the input transform when the kernel size is 5x5/5x1 or 1x5 and the output tile is 4x4/4x1 or 1x4 when the data layout is NCHW + * + * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2 + * @note If this kernel is used to perform Winograd input transform 5x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * @note If this kernel is used to perform Winograd input transform 1x5, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source image. Supported data types: F32 + * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_input_transform_4x4_5x5_stepz1_nchw( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + int x = get_global_id(0); + int y = get_global_id(1); + int z = get_global_id(2); + + // Compute input address + __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; + + src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y); + + // Load input tile +#if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) + const float8 in_row0 = vload8(0, (__global float *)(src_addr)); +#elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) + const float8 in_row0 = (float8)(*((__global float *)(src_addr + 0 * src_stride_y)), + *((__global float *)(src_addr + 1 * src_stride_y)), + *((__global float *)(src_addr + 2 * src_stride_y)), + *((__global float *)(src_addr + 3 * src_stride_y)), + *((__global float *)(src_addr + 4 * src_stride_y)), + *((__global float *)(src_addr + 5 * src_stride_y)), + *((__global float *)(src_addr + 6 * src_stride_y)), + *((__global float *)(src_addr + 7 * src_stride_y))); +#else // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + const float8 in_row0 = vload8(0, (__global float *)(src_addr + 0 * src_stride_y)); + const float8 in_row1 = vload8(0, (__global float *)(src_addr + 1 * src_stride_y)); + const float8 in_row2 = vload8(0, (__global float *)(src_addr + 2 * src_stride_y)); + const float8 in_row3 = vload8(0, (__global float *)(src_addr + 3 * src_stride_y)); + const float8 in_row4 = vload8(0, (__global float *)(src_addr + 4 * src_stride_y)); + const float8 in_row5 = vload8(0, (__global float *)(src_addr + 5 * src_stride_y)); + const float8 in_row6 = vload8(0, (__global float *)(src_addr + 6 * src_stride_y)); + const float8 in_row7 = vload8(0, (__global float *)(src_addr + 7 * src_stride_y)); +#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + + // Calculate common factors for intermediate tensor + float8 tmp0 = in_row0; + float8 comm_fact0 = 0.0f; + +#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + comm_fact0 += in_row2 + in_row6 - 4.25f * in_row4; + tmp0 += -in_row6 + 5.25f * in_row4 - 5.25f * in_row2; + + float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3; + float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6; + + const float8 tmp1 = comm_fact0 + comm_fact1; + const float8 tmp2 = comm_fact0 - comm_fact1; + + comm_fact0 = 2.5f * in_row3; + comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5; + + const float8 tmp3 = comm_fact1 + comm_fact2; + const float8 tmp4 = comm_fact2 - comm_fact1; + + comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5; + comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6; + + const float8 tmp5 = comm_fact1 + comm_fact2; + const float8 tmp6 = comm_fact2 - comm_fact1; + const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5; +#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + + // Calculate output rows (reuse comm_fact0 vector) + float8 out0; + + OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); + +#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + float8 out1, out2, out3, out4, out5, out6, out7; + + OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0); + OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0); + OUTPUT_ROW_4x4_5x5(out3, tmp3, comm_fact0); + OUTPUT_ROW_4x4_5x5(out4, tmp4, comm_fact0); + OUTPUT_ROW_4x4_5x5(out5, tmp5, comm_fact0); + OUTPUT_ROW_4x4_5x5(out6, tmp6, comm_fact0); + OUTPUT_ROW_4x4_5x5(out7, tmp7, comm_fact0); +#endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + + // Store values across the channels + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y; + + *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0; + *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1; + *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2; + *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3; + *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4; + *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5; + *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6; + *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7; + +#if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) + *((__global float *)(dst_addr + 8 * dst_stride_z)) = out1.s0; + *((__global float *)(dst_addr + 9 * dst_stride_z)) = out1.s1; + *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2; + *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3; + *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4; + *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5; + *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6; + *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7; + *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0; + *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1; + *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2; + *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3; + *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4; *((__global float *)(dst_addr + 21 * dst_stride_z)) = out2.s5; *((__global float *)(dst_addr + 22 * dst_stride_z)) = out2.s6; *((__global float *)(dst_addr + 23 * dst_stride_z)) = out2.s7; @@ -1424,6 +1655,105 @@ __kernel void winograd_input_transform_4x1_5x1_stepz1_nchw( dst_offset_first_element_in_bytes); } +#if defined(SRC_DIM_1) && defined(SRC_DIM_2) +/** This OpenCL kernel computes the input transform when the kernel size is 3x1 and the output tile is 4x1 for data layout NHWC + * + * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note Dimension one of the input tensor (width for NHWC data layout) must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM_1=112) + * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) + * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 + * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source image. Supported data types: F32 + * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_input_transform_4x1_3x1_stepz1_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_input_transform_4x4_3x3_stepz1_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel computes the input transform when the kernel size is 5x1 and the output tile is 4x1 for data layout NHWC + * + * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note Dimension one of the input tensor (width for NHWC data layout) must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM_1=112) + * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) + * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 + * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source image. Supported data types: F32 + * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_input_transform_4x1_5x1_stepz1_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_input_transform_4x4_5x5_stepz1_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} +#endif // defined(SRC_DIM_1) && defined(SRC_DIM_2) #endif // defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) #if defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) @@ -1546,11 +1876,58 @@ __kernel void winograd_input_transform_1x2_1x3_stepz2_nchw( * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor */ -__kernel void winograd_input_transform_1x4_1x3_stepz1_nchw( +__kernel void winograd_input_transform_1x4_1x3_stepz1_nchw( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) +{ + winograd_input_transform_4x4_3x3_stepz1_nchw(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); +} + +/** This OpenCL kernel computes the input transform when the kernel size is 1x5 and the output tile is 1x4 + * + * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 + * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source image. Supported data types: F32 + * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_input_transform_1x4_1x5_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst)) { - winograd_input_transform_4x4_3x3_stepz1_nchw(src_ptr, + winograd_input_transform_4x4_5x5_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, @@ -1568,9 +1945,12 @@ __kernel void winograd_input_transform_1x4_1x3_stepz1_nchw( dst_offset_first_element_in_bytes); } -/** This OpenCL kernel computes the input transform when the kernel size is 1x5 and the output tile is 1x4 +#if defined(SRC_DIM_1) && defined(SRC_DIM_2) +/** This OpenCL kernel computes the input transform when the kernel size is 1x3 and the output tile is 1x4 for data layout NHWC * * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note Dimension one of the input tensor (width for NHWC data layout) must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM_1=112) + * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 @@ -1593,11 +1973,11 @@ __kernel void winograd_input_transform_1x4_1x3_stepz1_nchw( * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor */ -__kernel void winograd_input_transform_1x4_1x5_stepz1_nchw( +__kernel void winograd_input_transform_1x4_1x3_stepz1_nhwc( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst)) { - winograd_input_transform_4x4_5x5_stepz1_nchw(src_ptr, + winograd_input_transform_4x4_3x3_stepz1_nhwc(src_ptr, src_stride_x, src_step_x, src_stride_y, @@ -1614,15 +1994,16 @@ __kernel void winograd_input_transform_1x4_1x5_stepz1_nchw( dst_step_z, dst_offset_first_element_in_bytes); } -#endif // defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) -#if defined(SRC_DIM_1) && defined(SRC_DIM_2) -/** This OpenCL kernel computes the input transform when the kernel size is 5x5 and the output tile is 4x4 when the data layout is NHWC +/** This OpenCL kernel computes the input transform when the kernel size is 1x5 and the output tile is 1x4 for data layout NHWC * * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5). + * @note Dimension one of the input tensor (width for NHWC data layout) must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM_1=112) + * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0). - * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 + * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time * * @param[in] src_ptr Pointer to the source image. Supported data types: F32 * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) @@ -1641,246 +2022,27 @@ __kernel void winograd_input_transform_1x4_1x5_stepz1_nchw( * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor */ -__kernel void winograd_input_transform_4x4_5x5_stepz1_nhwc( +__kernel void winograd_input_transform_1x4_1x5_stepz1_nhwc( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst)) { - int x = get_global_id(0); - int y = get_global_id(1); - int z = get_global_id(2); - - // Compute input address - __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float); - - // Clamp coordinates. This clamp is valid for all rows - int8 y_coord = (int8)(y * 4) + (int8)(0, 1, 2, 3, 4, 5, 6, 7) - (int8)PAD_LEFT; - y_coord = clamp(y_coord, -1, SRC_DIM_1); - - // Load 8x8 input tile - float8 in_row0, in_row1, in_row2, in_row3, in_row4, in_row5, in_row6, in_row7; - - // Row0 - int z_coord = (z * 4) - PAD_TOP + 0; - int8 valid_y = select(y_coord, -1, (int8)z_coord < 0); // If z < 0, set y to -1 - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); // If z >= SRC_DIM_2, set y to SRC_DIM_2 - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); // Clamp z coordinate - - in_row0.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row0.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row1 - z_coord = (z * 4) - PAD_TOP + 1; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row1.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row1.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row2 - z_coord = (z * 4) - PAD_TOP + 2; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row2.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row2.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row3 - z_coord = (z * 4) - PAD_TOP + 3; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row3.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row3.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row4 - z_coord = (z * 4) - PAD_TOP + 4; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row4.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row4.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row5 - z_coord = (z * 4) - PAD_TOP + 5; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row5.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row5.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row6 - z_coord = (z * 4) - PAD_TOP + 6; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row6.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row6.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Row7 - z_coord = (z * 4) - PAD_TOP + 7; - valid_y = select(y_coord, -1, (int8)z_coord < 0); - valid_y = select(valid_y, SRC_DIM_1, (int8)z_coord >= SRC_DIM_2); - z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1); - - in_row7.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z); - in_row7.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z); - - // Calculate common factors for intermediate tensor - float8 comm_fact0 = in_row2 + in_row6 - 4.25f * in_row4; - float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3; - float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6; - - // Calculate intermediate tensor and reuse common factor vectors - const float8 tmp0 = in_row0 - in_row6 + 5.25f * in_row4 - 5.25f * in_row2; - const float8 tmp1 = comm_fact0 + comm_fact1; - const float8 tmp2 = comm_fact0 - comm_fact1; - - comm_fact0 = 2.5f * in_row3; - comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5; - - const float8 tmp3 = comm_fact1 + comm_fact2; - const float8 tmp4 = comm_fact2 - comm_fact1; - - comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5; - comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6; - - const float8 tmp5 = comm_fact1 + comm_fact2; - const float8 tmp6 = comm_fact2 - comm_fact1; - const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5; - - // Calculate output rows (reuse comm_fact0 vector) - float8 out0, out1, out2, out3, out4, out5, out6, out7; - - OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); - OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0); - OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0); - OUTPUT_ROW_4x4_5x5(out3, tmp3, comm_fact0); - OUTPUT_ROW_4x4_5x5(out4, tmp4, comm_fact0); - OUTPUT_ROW_4x4_5x5(out5, tmp5, comm_fact0); - OUTPUT_ROW_4x4_5x5(out6, tmp6, comm_fact0); - OUTPUT_ROW_4x4_5x5(out7, tmp7, comm_fact0); - - // Store values across the 64 channels - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(float) + (y + z * (int)NUM_TILES_X) * dst_stride_y; - - *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0; - *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1; - *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2; - *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3; - *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4; - *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5; - *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6; - *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7; - *((__global float *)(dst_addr + 8 * dst_stride_z)) = out1.s0; - *((__global float *)(dst_addr + 9 * dst_stride_z)) = out1.s1; - *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2; - *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3; - *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4; - *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5; - *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6; - *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7; - *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0; - *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1; - *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2; - *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3; - *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4; - *((__global float *)(dst_addr + 21 * dst_stride_z)) = out2.s5; - *((__global float *)(dst_addr + 22 * dst_stride_z)) = out2.s6; - *((__global float *)(dst_addr + 23 * dst_stride_z)) = out2.s7; - *((__global float *)(dst_addr + 24 * dst_stride_z)) = out3.s0; - *((__global float *)(dst_addr + 25 * dst_stride_z)) = out3.s1; - *((__global float *)(dst_addr + 26 * dst_stride_z)) = out3.s2; - *((__global float *)(dst_addr + 27 * dst_stride_z)) = out3.s3; - *((__global float *)(dst_addr + 28 * dst_stride_z)) = out3.s4; - *((__global float *)(dst_addr + 29 * dst_stride_z)) = out3.s5; - *((__global float *)(dst_addr + 30 * dst_stride_z)) = out3.s6; - *((__global float *)(dst_addr + 31 * dst_stride_z)) = out3.s7; - *((__global float *)(dst_addr + 32 * dst_stride_z)) = out4.s0; - *((__global float *)(dst_addr + 33 * dst_stride_z)) = out4.s1; - *((__global float *)(dst_addr + 34 * dst_stride_z)) = out4.s2; - *((__global float *)(dst_addr + 35 * dst_stride_z)) = out4.s3; - *((__global float *)(dst_addr + 36 * dst_stride_z)) = out4.s4; - *((__global float *)(dst_addr + 37 * dst_stride_z)) = out4.s5; - *((__global float *)(dst_addr + 38 * dst_stride_z)) = out4.s6; - *((__global float *)(dst_addr + 39 * dst_stride_z)) = out4.s7; - *((__global float *)(dst_addr + 40 * dst_stride_z)) = out5.s0; - *((__global float *)(dst_addr + 41 * dst_stride_z)) = out5.s1; - *((__global float *)(dst_addr + 42 * dst_stride_z)) = out5.s2; - *((__global float *)(dst_addr + 43 * dst_stride_z)) = out5.s3; - *((__global float *)(dst_addr + 44 * dst_stride_z)) = out5.s4; - *((__global float *)(dst_addr + 45 * dst_stride_z)) = out5.s5; - *((__global float *)(dst_addr + 46 * dst_stride_z)) = out5.s6; - *((__global float *)(dst_addr + 47 * dst_stride_z)) = out5.s7; - *((__global float *)(dst_addr + 48 * dst_stride_z)) = out6.s0; - *((__global float *)(dst_addr + 49 * dst_stride_z)) = out6.s1; - *((__global float *)(dst_addr + 50 * dst_stride_z)) = out6.s2; - *((__global float *)(dst_addr + 51 * dst_stride_z)) = out6.s3; - *((__global float *)(dst_addr + 52 * dst_stride_z)) = out6.s4; - *((__global float *)(dst_addr + 53 * dst_stride_z)) = out6.s5; - *((__global float *)(dst_addr + 54 * dst_stride_z)) = out6.s6; - *((__global float *)(dst_addr + 55 * dst_stride_z)) = out6.s7; - *((__global float *)(dst_addr + 56 * dst_stride_z)) = out7.s0; - *((__global float *)(dst_addr + 57 * dst_stride_z)) = out7.s1; - *((__global float *)(dst_addr + 58 * dst_stride_z)) = out7.s2; - *((__global float *)(dst_addr + 59 * dst_stride_z)) = out7.s3; - *((__global float *)(dst_addr + 60 * dst_stride_z)) = out7.s4; - *((__global float *)(dst_addr + 61 * dst_stride_z)) = out7.s5; - *((__global float *)(dst_addr + 62 * dst_stride_z)) = out7.s6; - *((__global float *)(dst_addr + 63 * dst_stride_z)) = out7.s7; + winograd_input_transform_4x4_5x5_stepz1_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes); } #endif // defined(SRC_DIM_1) && defined(SRC_DIM_2) +#endif // defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) #endif // defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H) \ No newline at end of file diff --git a/src/core/CL/cl_kernels/winograd_output_transform.cl b/src/core/CL/cl_kernels/winograd_output_transform.cl index 61f0f61db7..bb159a7f4c 100644 --- a/src/core/CL/cl_kernels/winograd_output_transform.cl +++ b/src/core/CL/cl_kernels/winograd_output_transform.cl @@ -577,124 +577,6 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc( #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) } -#if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) -/** This OpenCL kernel performs Winograd output transform when the output tile is 4x1, the filter size 3x1 and the data layout is NHWC - * - * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 - * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 - * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 - * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_output_transform_4x1_3x1_nhwc( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(bias), -#endif // defined(HAS_BIAS) - int dst_size) -{ - winograd_output_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes, -#if defined(HAS_BIAS) - bias_ptr, - bias_stride_x, - bias_step_x, - bias_offset_first_element_in_bytes, -#endif // defined(HAS_BIAS) - dst_size); -} -#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) - -#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) -/** This OpenCL kernel performs Winograd output transform when the output tile is 1x4, the filter size 1x3 and the data layout is NHWC - * - * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 - * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 - * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 - * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void winograd_output_transform_1x4_1x3_nhwc( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), -#if defined(HAS_BIAS) - VECTOR_DECLARATION(bias), -#endif // defined(HAS_BIAS) - int dst_size) -{ - winograd_output_transform_4x4_3x3_nhwc(src_ptr, - src_stride_x, - src_step_x, - src_stride_y, - src_step_y, - src_stride_z, - src_step_z, - src_offset_first_element_in_bytes, - dst_ptr, - dst_stride_x, - dst_step_x, - dst_stride_y, - dst_step_y, - dst_stride_z, - dst_step_z, - dst_offset_first_element_in_bytes, -#if defined(HAS_BIAS) - bias_ptr, - bias_stride_x, - bias_step_x, - bias_offset_first_element_in_bytes, -#endif // defined(HAS_BIAS) - dst_size); -} -#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) - #define COMPUTE_TMP_COL(col, d0, d1, d2, d3, d4, d5, d6, d7, comm_fact) \ ({ \ comm_fact.s0 = d1 + d2; \ @@ -910,9 +792,13 @@ __kernel void winograd_output_transform_4x4_5x5_nchw( #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) } -/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 5x5 and the data layout is NHWC +/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4/4x1 or 1x4, the filter size 5x5/5x1 or 1x5 and the data layout is NHWC * * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 + * @note If this kernel is used to perform Winograd output transform 5x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * @note If this kernel is used to perform Winograd output transform 1x5, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) @@ -939,12 +825,17 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( #endif // defined(HAS_BIAS) int dst_size) { - // Each thread stores a 4x4 tile + // Each thread stores a 4x4/4x1 or 1x4 tile Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0); - // Load the values across the 64 channels to compose the 8x8 input tile + int y_in = get_global_id(1); + int x_out = get_global_id(0); + int y_out = (y_in % NUM_TILES_X) * OUTPUT_TILE_W; + int z_out = (y_in / NUM_TILES_X) * OUTPUT_TILE_H; + + // Load the values across the channels to compose the input tile float d00 = *((__global float *)(src_addr + 0 * src_stride_z)); float d01 = *((__global float *)(src_addr + 1 * src_stride_z)); float d02 = *((__global float *)(src_addr + 2 * src_stride_z)); @@ -954,6 +845,47 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( float d06 = *((__global float *)(src_addr + 6 * src_stride_z)); float d07 = *((__global float *)(src_addr + 7 * src_stride_z)); +#if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) + // Compute out00, out01, out02 and out03 + float out00 = d00 + d01 + d02 + d03 + d04 + 8.0f * d05 + 8.0f * d06; + float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04 + 4.0f * d05 - 4.0f * d06; + float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04 + 2.0f * d05 + 2.0f * d06; + float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05 - d06 + d07; + +#if defined(HAS_BIAS) + // Add bias + Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias); + + float b = (float) * ((__global float *)(vector_offset(&bias, z_out))); + + out00 += (float)b; + out01 += (float)b; + out02 += (float)b; + out03 += (float)b; +#endif // defined(HAS_BIAS) + + // Store the output tile +#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) + // Get output address + int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z); + offset = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding). + + *(__global float *)(dst_ptr + offset.s0) = out00; + *(__global float *)(dst_ptr + offset.s1) = out01; + *(__global float *)(dst_ptr + offset.s2) = out02; + *(__global float *)(dst_ptr + offset.s3) = out03; +#else // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) + // Get output address + int offset = dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z; + + *(__global float *)(dst_ptr + 0 * dst_stride_y + offset) = out00; + *(__global float *)(dst_ptr + 1 * dst_stride_y + offset) = out01; + *(__global float *)(dst_ptr + 2 * dst_stride_y + offset) = out02; + *(__global float *)(dst_ptr + 3 * dst_stride_y + offset) = out03; +#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) + +#else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) + float d10 = *((__global float *)(src_addr + 8 * src_stride_z)); float d11 = *((__global float *)(src_addr + 9 * src_stride_z)); float d12 = *((__global float *)(src_addr + 10 * src_stride_z)); @@ -1030,7 +962,7 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( COMPUTE_TMP_COL(tmp_col6, d06, d16, d26, d36, d46, d56, d66, d76, comm_fact0); COMPUTE_TMP_COL(tmp_col7, d07, d17, d27, d37, d47, d57, d67, d77, comm_fact0); - // Compute the 4x4 output tile + // Compute the output tile comm_fact0 = tmp_col1 + tmp_col2; comm_fact1 = tmp_col3 + tmp_col4; comm_fact2 = tmp_col5 + tmp_col6; @@ -1045,11 +977,6 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( float4 out_col1 = comm_fact0 + 2.f * comm_fact1 + 4.f * comm_fact2; float4 out_col3 = comm_fact0 + 8.f * comm_fact1 + comm_fact2 + tmp_col7; - int y_in = get_global_id(1); - int x_out = get_global_id(0); - int y_out = (y_in % NUM_TILES_X) * 4; - int z_out = (y_in / NUM_TILES_X) * 4; - #if defined(HAS_BIAS) // Add bias Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias); @@ -1061,13 +988,12 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( out_col2 += (float4)b; out_col3 += (float4)b; #endif // defined(HAS_BIAS) - // Get output address int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z); offset = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding). int4 mult_y = min(dst_size - offset, 1); // If out of bound, we don't want to increase dst_stride_y, so we set the multiplier to 0. It will be 1 otherwise. - // Store the 4x4 output tile + // Store the output tile *(__global float *)(dst_ptr + mult_y.s0 * 0 * dst_stride_y + offset.s0) = out_col0.s0; *(__global float *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset.s0) = out_col1.s0; *(__global float *)(dst_ptr + mult_y.s0 * 2 * dst_stride_y + offset.s0) = out_col2.s0; @@ -1084,6 +1010,7 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc( *(__global float *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset.s3) = out_col1.s3; *(__global float *)(dst_ptr + mult_y.s0 * 2 * dst_stride_y + offset.s3) = out_col2.s3; *(__global float *)(dst_ptr + mult_y.s0 * 3 * dst_stride_y + offset.s3) = out_col3.s3; +#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) } #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) @@ -1263,6 +1190,120 @@ __kernel void winograd_output_transform_4x1_5x1_nchw( #endif // defined(HAS_BIAS) ); } + +/** This OpenCL kernel performs Winograd output transform when the output tile is 4x1, the filter size 3x1 and the data layout is NHWC + * + * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 + * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_output_transform_4x1_3x1_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), +#if defined(HAS_BIAS) + VECTOR_DECLARATION(bias), +#endif // defined(HAS_BIAS) + int dst_size) +{ + winograd_output_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes, +#if defined(HAS_BIAS) + bias_ptr, + bias_stride_x, + bias_step_x, + bias_offset_first_element_in_bytes, +#endif // defined(HAS_BIAS) + dst_size); +} + +/** This OpenCL kernel performs Winograd output transform when the output tile is 4x1, the filter size 5x1 and the data layout is NHWC + * + * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1 + * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_output_transform_4x1_5x1_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), +#if defined(HAS_BIAS) + VECTOR_DECLARATION(bias), +#endif // defined(HAS_BIAS) + int dst_size) +{ + winograd_output_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes, +#if defined(HAS_BIAS) + bias_ptr, + bias_stride_x, + bias_step_x, + bias_offset_first_element_in_bytes, +#endif // defined(HAS_BIAS) + dst_size); +} #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) @@ -1442,5 +1483,119 @@ __kernel void winograd_output_transform_1x4_1x5_nchw( #endif // defined(HAS_BIAS) ); } + +/** This OpenCL kernel performs Winograd output transform when the output tile is 1x4, the filter size 1x3 and the data layout is NHWC + * + * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 + * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_output_transform_1x4_1x3_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), +#if defined(HAS_BIAS) + VECTOR_DECLARATION(bias), +#endif // defined(HAS_BIAS) + int dst_size) +{ + winograd_output_transform_4x4_3x3_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes, +#if defined(HAS_BIAS) + bias_ptr, + bias_stride_x, + bias_step_x, + bias_offset_first_element_in_bytes, +#endif // defined(HAS_BIAS) + dst_size); +} + +/** This OpenCL kernel performs Winograd output transform when the output tile is 1x4, the filter size 1x5 and the data layout is NHWC + * + * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16 + * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1 + * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4 + * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void winograd_output_transform_1x4_1x5_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), +#if defined(HAS_BIAS) + VECTOR_DECLARATION(bias), +#endif // defined(HAS_BIAS) + int dst_size) +{ + winograd_output_transform_4x4_5x5_nhwc(src_ptr, + src_stride_x, + src_step_x, + src_stride_y, + src_step_y, + src_stride_z, + src_step_z, + src_offset_first_element_in_bytes, + dst_ptr, + dst_stride_x, + dst_step_x, + dst_stride_y, + dst_step_y, + dst_stride_z, + dst_step_z, + dst_offset_first_element_in_bytes, +#if defined(HAS_BIAS) + bias_ptr, + bias_stride_x, + bias_step_x, + bias_offset_first_element_in_bytes, +#endif // defined(HAS_BIAS) + dst_size); +} #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL) #endif // defined(NUM_TILES_X) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H) diff --git a/tests/datasets/WinogradOutputTransformDataset.h b/tests/datasets/WinogradOutputTransformDataset.h index 4085e91854..617a67de07 100644 --- a/tests/datasets/WinogradOutputTransformDataset.h +++ b/tests/datasets/WinogradOutputTransformDataset.h @@ -206,6 +206,22 @@ public: add_config(TensorShape(7U, 2U, 64U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); add_config(TensorShape(24U, 9U, 64U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NHWC)); add_config(TensorShape(7U, 2U, 64U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + + // (4x1, 5x1) + add_config(TensorShape(13U, 6U, 8U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(7U, 22U, 8U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(5U, 462U, 8U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(53U, 33U), PadStrideInfo(1, 1, 2, 0), DataLayout::NHWC)); + add_config(TensorShape(7U, 10U, 8U, 3U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(24U, 42U, 8U, 2U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC)); + add_config(TensorShape(7U, 20U, 8U, 5U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(8U, 10U), PadStrideInfo(1, 1, 2, 0), DataLayout::NHWC)); + + // (1x4, 1x5) + add_config(TensorShape(13U, 7U, 8U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(7U, 20U, 8U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(5U, 477U, 8U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(53U, 33U), PadStrideInfo(1, 1, 0, 2), DataLayout::NHWC)); + add_config(TensorShape(7U, 16U, 8U, 3U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(24U, 42U, 8U, 2U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(14U, 14U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC)); + add_config(TensorShape(7U, 24U, 8U, 5U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 2), DataLayout::NHWC)); } }; @@ -316,6 +332,18 @@ public: add_config(TensorShape(13U, 182U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC)); add_config(TensorShape(32U, 756U, 64U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC)); add_config(TensorShape(13U, 182U, 64U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC)); + + // (4x1, 5x1) + add_config(TensorShape(32U, 3136U, 8U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(112U, 112U), PadStrideInfo(1, 1, 2, 0), DataLayout::NHWC)); + add_config(TensorShape(13U, 784U, 8U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC)); + add_config(TensorShape(32U, 3024U, 8U, 2U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(112U, 112U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(13U, 784U, 8U, 5U), WinogradInfo(Size2D(4U, 1U), Size2D(5U, 1U), Size2D(56U, 56U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC)); + + // (1x4, 1x5) + add_config(TensorShape(32U, 3136U, 8U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(112U, 112U), PadStrideInfo(1, 1, 0, 2), DataLayout::NHWC)); + add_config(TensorShape(13U, 784U, 8U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC)); + add_config(TensorShape(32U, 3024U, 8U, 2U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(112U, 112U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC)); + add_config(TensorShape(13U, 784U, 8U, 5U), WinogradInfo(Size2D(1U, 4U), Size2D(1U, 5U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC)); } }; } // namespace datasets diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp index c39cb4e790..004d7c98a0 100644 --- a/tests/validation/CL/Winograd.cpp +++ b/tests/validation/CL/Winograd.cpp @@ -72,7 +72,9 @@ const auto SmallWinogradInputTransformDatasetNCHW = const auto SmallWinogradInputTransformDatasetNHWC = framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(), framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x1_3x1(), framework::dataset::concat(datasets::SmallWinogradInputTransformDataset1x4_1x3(), - datasets::SmallWinogradInputTransformDataset4x4_5x5()))); + framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_5x5(), + framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x1_5x1(), + datasets::SmallWinogradInputTransformDataset1x4_1x5()))))); const auto LargeWinogradInputTransformDatasetNCHW = framework::dataset::concat(datasets::LargeWinogradInputTransformDataset2x2_3x3(), @@ -87,7 +89,9 @@ const auto LargeWinogradInputTransformDatasetNCHW = const auto LargeWinogradInputTransformDatasetNHWC = framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(), - datasets::LargeWinogradInputTransformDataset4x4_5x5()); + framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_5x5(), + framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x1_5x1(), + datasets::LargeWinogradInputTransformDataset1x4_1x5()))); // Filter transform const auto SmallWinogradFilterTransformDatasetNCHW = @@ -102,7 +106,9 @@ const auto SmallWinogradFilterTransformDatasetNHWC = framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })), framework::dataset::concat(combine(datasets::Small3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })), framework::dataset::concat(combine(datasets::Small1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })), - combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) }))))); + framework::dataset::concat(combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })), + framework::dataset::concat(combine(datasets::Small5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })), + combine(datasets::Small1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) }))))))); const auto LargeWinogradFilterTransformDatasetNCHW = framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })), @@ -114,7 +120,11 @@ const auto LargeWinogradFilterTransformDatasetNCHW = const auto LargeWinogradFilterTransformDatasetNHWC = framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })), - combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) }))); + framework::dataset::concat(combine(datasets::Large3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })), + framework::dataset::concat(combine(datasets::Large1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) })), + framework::dataset::concat(combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })), + framework::dataset::concat(combine(datasets::Large5x1Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 1U) })), + combine(datasets::Large1x5Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 4U) }))))))); // Output transform const auto SmallWinogradOutputTransformDatasetNCHW = datasets::SmallWinogradOutputTransformDatasetNCHW(); -- cgit v1.2.1