From 876be2a0d11874d871860dbd22481f831d6878f6 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Tue, 3 Jul 2018 12:22:09 +0100 Subject: COMPMID-1339 - Implementing Winograd Convolution Layer 1x5 and 5x1 kernels on OpenCL NCHW Change-Id: Ia293cd89651146a0e27e5f7c74ca9c924807e83c Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138707 Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- .../CL/kernels/CLWinogradOutputTransformKernel.h | 22 ++++++++++++++++------ 1 file changed, 16 insertions(+), 6 deletions(-) (limited to 'arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h') diff --git a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h index cd46e9813e..0798172ba7 100644 --- a/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h +++ b/arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h @@ -48,10 +48,15 @@ public: ~CLWinogradOutputTransformKernel() = default; /** Set the input and output tensor. * - * @note Winograd output transform supports the following configurations: - * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5) + * @note Winograd output transform supports the following configurations for NCWH data layout + * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), + * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * 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) * Strides: only unit strides - * Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5) * * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32. * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input @@ -61,10 +66,15 @@ public: void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info); /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradOutputTransformKernel * - * @note Winograd output transform supports the following configurations: - * F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5) + * @note Winograd output transform supports the following configurations for NCWH data layout + * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), + * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * 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) * Strides: only unit strides - * Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3) and F(4x4, 5x5) * * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F32. * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input -- cgit v1.2.1