From 45091736a9276919ececee0cba106228246341f8 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 13 May 2019 17:41:01 +0100 Subject: COMPMID-2184: Implement direct convolution 9x9 (NHWC) on OpenCL Change-Id: I8aa929e7e72d2d1ccee07ee2ed9618c15084ae9d Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/1274 Comments-Addressed: Arm Jenkins Reviewed-by: Michele Di Giorgio Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas --- .../CL/kernels/CLDirectConvolutionLayerKernel.h | 3 +- src/core/CL/CLKernelLibrary.cpp | 5 + src/core/CL/cl_kernels/direct_convolution9x9.cl | 444 +++++++++++++++++++++ .../CL/kernels/CLDirectConvolutionLayerKernel.cpp | 54 ++- tests/validation/CL/DirectConvolutionLayer.cpp | 59 ++- 5 files changed, 546 insertions(+), 19 deletions(-) create mode 100644 src/core/CL/cl_kernels/direct_convolution9x9.cl diff --git a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h index bd37e35334..081b01aad3 100644 --- a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -54,6 +54,7 @@ public: * 1x1 convolution with stride_x = 1/2/3, stride_y = 1/2/3 * 3x3 convolution with stride_x = 1/2, stride_y = 1/2 * 5x5 convolution with stride_x = 1/2, stride_y = 1/2 + * 9x9 convolution with stride_x = 1/2, stride_y = 1/2, data_layout=NHWC * * @param[in] input The input tensor to convolve. 3 lower dimensions represent a single input [width, height, IFM], * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F16/F32. diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 253da40077..c0875bebcd 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -248,6 +248,7 @@ const std::map CLKernelLibrary::_kernel_program_map = { "direct_convolution5x5_nhwc", "direct_convolution5x5.cl" }, { "direct_convolution5x5_f32_bifrost", "direct_convolution5x5.cl" }, { "direct_convolution_1x1_3x3_5x5_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" }, + { "direct_convolution9x9_nhwc", "direct_convolution9x9.cl" }, { "elementwise_operation_ADD", "elementwise_operation.cl" }, { "elementwise_operation_SUB", "elementwise_operation.cl" }, { "elementwise_operation_MAX", "elementwise_operation.cl" }, @@ -709,6 +710,10 @@ const std::map CLKernelLibrary::_program_source_map = { "direct_convolution_1x1_3x3_5x5_quantized.cl", #include "./cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.clembed" + }, + { + "direct_convolution9x9.cl", +#include "./cl_kernels/direct_convolution9x9.clembed" }, { "elementwise_operation.cl", diff --git a/src/core/CL/cl_kernels/direct_convolution9x9.cl b/src/core/CL/cl_kernels/direct_convolution9x9.cl new file mode 100644 index 0000000000..22f437f78e --- /dev/null +++ b/src/core/CL/cl_kernels/direct_convolution9x9.cl @@ -0,0 +1,444 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +#undef CONVERT_SAT + +#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(DATA_LAYOUT_NHWC) + +#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR)) + +#define CONVOLUTION1x9_STRIDE1_NHWC(acc, row_ptr, weights_ptr) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + src0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ + PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE)); \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + src1 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ + PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE)); \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ + PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 5 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 6 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 7 * weights_stride_y, DATA_TYPE)); \ + DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 8 * weights_stride_y, DATA_TYPE); \ + acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s4567, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s4; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s567, src1.s0123, src1.s4) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s5; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s67, src1.s012, src1.s345) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s6; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s7, src1.s0123, src1.s456) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s7; \ + acc += src1 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ + }) + +#define CONVOLUTION1x9_STRIDE2_NHWC(acc, row_ptr, weights_ptr) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 16) \ + src0 = (VEC_DATA_TYPE(DATA_TYPE, 16))( \ + PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE)); \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + src1 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ + PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 17 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 18 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 19 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 20 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 21 * src_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(row_ptr + 22 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 23 * src_stride_y, DATA_TYPE)); \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \ + PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 5 * weights_stride_y, DATA_TYPE), \ + PTR_TO_VALUE(weights_ptr + 6 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 7 * weights_stride_y, DATA_TYPE)); \ + DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 8 * weights_stride_y, DATA_TYPE); \ + acc += src0.s02468ACE * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468A, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s4; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s579, src0.sBDF, src1.s13) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s5; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s68A, src0.sCE, src1.s024) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s6; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s79B, src0.sDF, src1.s135) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s7; \ + acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s8AC, src0.sE, src1.s0246) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ + }) + +#if defined(VEC_SIZE) +#define VFMA(acc, w, src0, src1, src2, src3, src4, src5, src6, src7) \ + ({ \ + acc##0 = fma(src0, w, acc##0); \ + acc##1 = fma(src1, w, acc##1); \ + acc##2 = fma(src2, w, acc##2); \ + acc##3 = fma(src3, w, acc##3); \ + acc##4 = fma(src4, w, acc##4); \ + acc##5 = fma(src5, w, acc##5); \ + acc##6 = fma(src6, w, acc##6); \ + acc##7 = fma(src7, w, acc##7); \ + }) + +#define CONVOLUTION1x9_STRIDE1_NHWC_BIFROST(acc, row_ptr, weights_ptr) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)row_ptr); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 2 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 3 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 4 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 5 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 6 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 7 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 8 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 9 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 10 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 11 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 12 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 13 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 14 * src_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + src15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(row_ptr + 15 * src_stride_y)); \ + \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 0 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 1 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 2 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 3 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 4 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 5 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 6 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 7 * weights_stride_y)); \ + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) \ + w8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights_ptr + 8 * weights_stride_y)); \ + \ + VFMA(acc, w0, src0, src1, src2, src3, src4, src5, src6, src7); \ + VFMA(acc, w1, src1, src2, src3, src4, src5, src6, src7, src8); \ + VFMA(acc, w2, src2, src3, src4, src5, src6, src7, src8, src9); \ + VFMA(acc, w3, src3, src4, src5, src6, src7, src8, src9, src10); \ + VFMA(acc, w4, src4, src5, src6, src7, src8, src9, src10, src11); \ + VFMA(acc, w5, src5, src6, src7, src8, src9, src10, src11, src12); \ + VFMA(acc, w6, src6, src7, src8, src9, src10, src11, src12, src13); \ + VFMA(acc, w7, src7, src8, src9, src10, src11, src12, src13, src14); \ + VFMA(acc, w8, src8, src9, src10, src11, src12, src13, src14, src15); \ + }) + +#if VEC_SIZE == 4 +#define REDUCE(out, vec) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 2) \ + tmp1 = vec.s01 + vec.s23; \ + out = tmp1.s0 + tmp1.s1; \ + }) +#else // VEC_SIZE == 4 +#error("Not supported") +#endif // VEC_SIZE == 4 + +#if STRIDE_X == 1 +#define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE1_NHWC_BIFROST(acc, row_ptr, weights_ptr) +#else // STRIDE_X == 1 +#error "Not supported" +#endif // STRIDE_X == 1 + +#else // defined(VEC_SIZE) + +#if STRIDE_X == 1 +#define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE1_NHWC(acc, row_ptr, weights_ptr) +#elif STRIDE_X == 2 // STRIDE_X == 1 +#define CONVOLUTION1x9_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x9_STRIDE2_NHWC(acc, row_ptr, weights_ptr) +#else // STRIDE_X == 1 +#error "STRIDE_X larger than 2 is not supported" +#endif // STRIDE_X == 1 + +#endif // defined(VEC_SIZE) + +//#if defined(VEC_SIZE) +/** This kernel performs a direct convolution to convolve the low three dimensions in a tensor with the NHWC data layout + * + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH + * @note If biases are used then -DHAS_BIAS has to be passed at compile time + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/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 Z 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 Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr + * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) + * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) + * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) + * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) + * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor + * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Same as @p src_ptr + * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) + * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor + * @param[in] weights_stride_w (Optional) Stride of the weights tensor in the 4th dimension + */ +__kernel void direct_convolution9x9_nhwc( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), + TENSOR3D_DECLARATION(weights), +#ifdef HAS_BIAS + VECTOR_DECLARATION(biases), +#endif /* defined(HAS_BIAS) */ + unsigned int weights_stride_w) +{ + Image src = CONVERT_TO_IMAGE_STRUCT(src); + Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); + Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); + + VEC_DATA_TYPE(DATA_TYPE, 8) + values = 0; + +#if defined(VEC_SIZE) + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values0 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values1 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values2 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values3 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values4 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values5 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values6 = 0; + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values7 = 0; +#define STEP_X (VEC_SIZE) +#else // defined(VEC_SIZE) +#define STEP_X (1) +#endif // defined(VEC_SIZE) + + const int id0 = get_global_id(0); + const int id1 = get_global_id(1); + const int id2 = get_global_id(2); + + __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0); + __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0) + ((id2 * STRIDE_Y) - PAD_TOP) * (int)src_stride_z; + +#if(PAD_TOP == 1) + const int coordy = id2 - PAD_TOP; + for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) + { + if(coordy < 0) // special case Z = -1 doesn't exists + { + //skip first row and load the two next ones + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + } + else if(coordy == (DST_HEIGHT - PAD_TOP - 1)) + { + // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the + // Z axis has no padding at all. + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + } + else + { + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + } + src_addr += STEP_X * sizeof(DATA_TYPE); + weights_addr += STEP_X * sizeof(DATA_TYPE); + } +#elif(PAD_TOP == 2) // PAD_TOP == 1 + const int coordy = id2 * STRIDE_Y; + for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) + { + if(coordy == 0) // special case Z = -2 doesn't exists + { + //skip first row and load the two next ones + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + } + else if(coordy == 1) // special case Z = -1 doesn't exists + { + //skip first row and load the two next ones + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + } + else if(coordy == (SRC_HEIGHT - 5)) + { + // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the + // Z axis has no padding at all. + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + } + else if(coordy == (SRC_HEIGHT - 6)) + { + // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the + // Z axis has no padding at all. + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + } + else + { + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + } + src_addr += STEP_X * sizeof(DATA_TYPE); + weights_addr += STEP_X * sizeof(DATA_TYPE); + } + +#else // PAD_TOP == 1 + for(volatile int d = 0; d < WEIGHTS_DEPTH; d += STEP_X) + { + CONVOLUTION1x9_NHWC(values, src_addr, weights_addr); + CONVOLUTION1x9_NHWC(values, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); + CONVOLUTION1x9_NHWC(values, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); + src_addr += STEP_X * sizeof(DATA_TYPE); + weights_addr += STEP_X * sizeof(DATA_TYPE); + } +#endif // PAD_TOP == 1 + +#if defined(VEC_SIZE) + REDUCE(values.s0, values0); + REDUCE(values.s1, values1); + REDUCE(values.s2, values2); + REDUCE(values.s3, values3); + REDUCE(values.s4, values4); + REDUCE(values.s5, values5); + REDUCE(values.s6, values6); + REDUCE(values.s7, values7); +#endif // defined(VEC_SIZE) + +#if defined(HAS_BIAS) + Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); + values += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0))); +#endif // defined(HAS_BIAS) + + *((__global DATA_TYPE *)(dst.ptr + 0 * dst_stride_y)) = values.s0; + *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values.s1; + *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values.s2; + *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values.s3; + *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values.s4; + *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values.s5; + *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values.s6; + *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values.s7; +#undef STEP_X +} +#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(DATA_LAYOUT_NHWC) diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp index 3e158a52ff..b878a2121f 100644 --- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp +++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp @@ -54,14 +54,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != weights->dimension(height_idx), "Weights should have same width and height"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5, - "Kernel sizes other than 1x1, 3x3 or 5x5 are not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5 && weights->dimension(width_idx) != 9, + "Kernel sizes other than 1x1, 3x3, 5x5 or 9x9 are not supported"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx), "Weights feature map dimension should match the respective input's one"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional"); ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 1) && std::get<0>(conv_info.stride()) > 3, "Strides larger than 3 not supported for 1x1 convolution."); ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 3 || weights->dimension(width_idx) == 5) && std::get<0>(conv_info.stride()) > 2, "Strides larger than 2 not supported for 3x3 convolution."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 9) && data_layout == DataLayout::NCHW, "Only NHWC layout is supported for 9x9 convolution."); if(biases != nullptr) { @@ -103,6 +104,19 @@ inline bool can_run_optimized_kernel_for_bifrost(GPUTarget gpu_target, unsigned && (data_layout == DataLayout::NCHW); } +inline bool can_run_optimized_kernel_for_bifrost_nhwc(GPUTarget gpu_target, unsigned int conv_stride_x, unsigned int conv_stride_y, unsigned int kernel_size, + DataType data_type, DataLayout data_layout) +{ + return gpu_target_is_in(gpu_target, + GPUTarget::G71, GPUTarget::G72, GPUTarget::G76, + GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, + GPUTarget::G52, GPUTarget::G52LIT) + && (kernel_size == 9) + && (conv_stride_x == 1) && (conv_stride_y == 1) + && (data_type == DataType::F32) + && (data_layout == DataLayout::NHWC); +} + inline void setup_num_elems(unsigned int &num_elems_read_per_iteration_x, unsigned int &num_elems_read_per_iteration_y, unsigned int &num_elems_written_per_iteration_x, unsigned int &num_elems_written_per_iteration_y, unsigned int kernel_size, const PadStrideInfo &conv_info, const GPUTarget target, ITensorInfo *input) @@ -149,7 +163,7 @@ inline void setup_num_elems(unsigned int &num_elems_read_per_iteration_x, unsign } } } - else + else if(data_layout == DataLayout::NCHW) { num_elems_read_per_iteration_y = kernel_size; num_elems_written_per_iteration_x = 8; @@ -215,11 +229,17 @@ inline void setup_num_elems(unsigned int &num_elems_read_per_iteration_x, unsign ARM_COMPUTE_ERROR("Invalid direct convolution size"); } } - - if(data_layout == DataLayout::NHWC) + else // data_layout == NHWC { + const bool run_optimized_bifrost_nhwc = can_run_optimized_kernel_for_bifrost_nhwc(target, conv_stride_x, conv_stride_y, kernel_size, data_type, data_layout); + num_elems_written_per_iteration_x = 1; - num_elems_read_per_iteration_x = 1; + + if(run_optimized_bifrost_nhwc) + { + num_elems_read_per_iteration_x = 4; + } + switch(kernel_size) { case 1: @@ -267,6 +287,21 @@ inline void setup_num_elems(unsigned int &num_elems_read_per_iteration_x, unsign ARM_COMPUTE_ERROR("Invalid convolution stride X"); } break; + case 9: + switch(conv_stride_x) + { + case 1: + num_elems_read_per_iteration_y = 16; + num_elems_written_per_iteration_y = 8; + break; + case 2: + num_elems_read_per_iteration_y = 24; + num_elems_written_per_iteration_y = 8; + break; + default: + ARM_COMPUTE_ERROR("Invalid convolution stride X"); + } + break; default: ARM_COMPUTE_ERROR("Not implemented."); break; @@ -429,6 +464,7 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL build_options.add_option(std::string("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x))); if(data_layout == DataLayout::NHWC) { + const bool run_optimized_for_bifrost_nhwc = can_run_optimized_kernel_for_bifrost_nhwc(gpu_target, _conv_stride_x, _conv_stride_y, kernel_size, data_type, data_layout); build_options.add_option(std::string("-DDATA_LAYOUT_NHWC=1")); build_options.add_option(std::string("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(height_idx)))); build_options.add_option(std::string("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(width_idx)))); @@ -437,6 +473,12 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL build_options.add_option(std::string("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()))); build_options.add_option(std::string("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()))); build_options.add_option(std::string("-DSTRIDE_Y=" + support::cpp11::to_string(_conv_stride_y))); + if(run_optimized_for_bifrost_nhwc) + { + const unsigned int num_elems_read_per_iteration_x = 4; + _border_size.right = num_elems_read_per_iteration_x; + build_options.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_read_per_iteration_x)); + } } build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(data_type))); // Create kernel diff --git a/tests/validation/CL/DirectConvolutionLayer.cpp b/tests/validation/CL/DirectConvolutionLayer.cpp index 437d5bac8c..6c46374b54 100644 --- a/tests/validation/CL/DirectConvolutionLayer.cpp +++ b/tests/validation/CL/DirectConvolutionLayer.cpp @@ -49,21 +49,28 @@ RelativeTolerance tolerance_fp32(0.02f); /**< Tolerance f constexpr float tolerance_num = 0.07f; /**< Tolerance number */ constexpr AbsoluteTolerance tolerance_qasymm8(1); /**< Tolerance for quantized tests */ -const auto data_strides = combine(framework::dataset::make("StrideX", 1, 3), framework::dataset::make("StrideY", 1, 3)); -const auto data_strides_small = combine(framework::dataset::make("StrideX", 1), framework::dataset::make("StrideY", 1)); -const auto data_ksize_one = combine(framework::dataset::make("PadX", 0, 1), combine(framework::dataset::make("PadY", 0, 1), framework::dataset::make("KernelSize", 1))); -const auto data_ksize_one_small = combine(framework::dataset::make("PadX", 0), combine(framework::dataset::make("PadY", 0), framework::dataset::make("KernelSize", 1))); -const auto data_ksize_three = combine(framework::dataset::make("PadX", 0, 2), combine(framework::dataset::make("PadY", 0, 2), framework::dataset::make("KernelSize", 3))); -const auto data_ksize_five = combine(framework::dataset::make("PadX", 0, 3), combine(framework::dataset::make("PadY", 0, 3), framework::dataset::make("KernelSize", 5))); -const auto data_all_kernels = concat(concat(data_ksize_one, data_ksize_three), data_ksize_five); +const auto data_strides = combine(framework::dataset::make("StrideX", 1, 3), framework::dataset::make("StrideY", 1, 3)); +const auto data_strides_small = combine(framework::dataset::make("StrideX", 1), framework::dataset::make("StrideY", 1)); +const auto data_ksize_one = combine(framework::dataset::make("PadX", 0, 1), combine(framework::dataset::make("PadY", 0, 1), framework::dataset::make("KernelSize", 1))); +const auto data_ksize_one_small = combine(framework::dataset::make("PadX", 0), combine(framework::dataset::make("PadY", 0), framework::dataset::make("KernelSize", 1))); +const auto data_ksize_three = combine(framework::dataset::make("PadX", 0, 2), combine(framework::dataset::make("PadY", 0, 2), framework::dataset::make("KernelSize", 3))); +const auto data_ksize_five = combine(framework::dataset::make("PadX", 0, 3), combine(framework::dataset::make("PadY", 0, 3), framework::dataset::make("KernelSize", 5))); +const auto data_ksize_nine = combine(framework::dataset::make("PadX", 0, 3), combine(framework::dataset::make("PadY", 0, 3), framework::dataset::make("KernelSize", 9))); +const auto data_ksize_nine_small = combine(framework::dataset::make("PadX", 0, 1), combine(framework::dataset::make("PadY", 0, 1), framework::dataset::make("KernelSize", 9))); -const auto data = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides, data_all_kernels)); -const auto data_small = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides_small, data_ksize_one_small)); +const auto data_all_kernels = concat(concat(data_ksize_one, data_ksize_three), data_ksize_five); + +const auto data = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides, data_all_kernels)); +const auto data9x9 = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides, data_ksize_nine)); +const auto data_small = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides_small, data_ksize_one_small)); +const auto data_small9x9 = combine(datasets::SmallDirectConvolutionShapes(), combine(data_strides_small, data_ksize_nine_small)); /** Direct convolution nightly data set. */ -const auto data_nightly = combine(data, framework::dataset::make("NumKernels", { 1, 4 })); +const auto data_nightly = combine(data, framework::dataset::make("NumKernels", { 1, 4 })); +const auto data_nightly_9x9 = combine(data9x9, framework::dataset::make("NumKernels", { 1, 4 })); /** Direct convolution precommit data set. */ -const auto data_precommit = combine(data_small, framework::dataset::make("NumKernels", { 1 })); +const auto data_precommit = combine(data_small, framework::dataset::make("NumKernels", { 1 })); +const auto data_precommit_9x9 = combine(data_small9x9, framework::dataset::make("NumKernels", { 1 })); /** Activation function Dataset*/ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", @@ -92,7 +99,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( }), framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 3U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(9U, 9U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(11U, 11U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(5U, 3U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U, 3U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), @@ -172,6 +179,20 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerFixture, framewor // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num); } +FIXTURE_DATA_TEST_CASE(RunLarge9x9, CLDirectConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(data_nightly_9x9, framework::dataset::make("DataType", + DataType::F16)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NHWC }))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit_9x9, framework::dataset::make("DataType", + DataType::F16)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NHWC }))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num); +} TEST_SUITE_END() // FP16 TEST_SUITE(FP32) @@ -188,6 +209,20 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerFixture, framewo { validate(CLAccessor(_target), _reference, tolerance_fp32); } +FIXTURE_DATA_TEST_CASE(RunLarge9x9, CLDirectConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(data_nightly_9x9, framework::dataset::make("DataType", + DataType::F32)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NHWC }))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp32); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit_9x9, framework::dataset::make("DataType", + DataType::F32)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NHWC }))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp32); +} TEST_SUITE_END() // FP32 TEST_SUITE(FP32_CustomDataset) -- cgit v1.2.1