From ff1fe3e32e25069fed750cdfe3046b7d8d5a2628 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Sat, 2 Jan 2021 09:58:51 +0000 Subject: Remove padding from direct convolution - OpenCL - Refactor direct convolution for NHWC - Remove old kernels for NHWC - Change the heuristic in CLConvolutionLayer.cpp. The new direct convolution implementation is faster than FFT Resolves COMPMID-3908 Change-Id: Iee15ce7b04e21847b6eaae5c6d3c1b18180e7efc Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4876 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas --- src/core/CL/cl_kernels/direct_convolution3x3.cl | 181 +----------------------- 1 file changed, 1 insertion(+), 180 deletions(-) (limited to 'src/core/CL/cl_kernels/direct_convolution3x3.cl') diff --git a/src/core/CL/cl_kernels/direct_convolution3x3.cl b/src/core/CL/cl_kernels/direct_convolution3x3.cl index da7a1e7410..811df053c4 100644 --- a/src/core/CL/cl_kernels/direct_convolution3x3.cl +++ b/src/core/CL/cl_kernels/direct_convolution3x3.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2018 Arm Limited. + * Copyright (c) 2016-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -66,185 +66,6 @@ acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2)); \ }) -#if defined(DATA_LAYOUT_NHWC) - -#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR)) - -#if STRIDE_X == 1 -#define CONVOLUTION1x3_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x3_STRIDE_NHWC_STRIDE1(acc, row_ptr, weights_ptr) -#elif STRIDE_X == 2 /* STRIDE_X == 1 */ -#define CONVOLUTION1x3_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x3_STRIDE_NHWC_STRIDE2(acc, row_ptr, weights_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X == 2 */ - -#define CONVOLUTION1x3_STRIDE_NHWC_STRIDE1(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, 2) \ - src1 = (VEC_DATA_TYPE(DATA_TYPE, 2))( \ - PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), \ - PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE)); \ - VEC_DATA_TYPE(DATA_TYPE, 3) \ - weights = (VEC_DATA_TYPE(DATA_TYPE, 3))( \ - 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)); \ - acc = ADD_OP(acc, MUL_OP(src0, (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s0)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s1)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s2)); \ - } - -#define CONVOLUTION1x3_STRIDE_NHWC_STRIDE2(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)); \ - DATA_TYPE src1 = PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE); \ - VEC_DATA_TYPE(DATA_TYPE, 3) \ - weights = (VEC_DATA_TYPE(DATA_TYPE, 3))( \ - 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)); \ - \ - acc = ADD_OP(acc, MUL_OP(src0.s02468ACE, (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s0)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s1)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s2)); \ - } - -/** This kernel performs a direct convolution to convolve the low three dimensions. - * - * @note This OpenCL kernel works with stride_x = 1 and 2 - * @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: QS8/QS16/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 Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution3x3_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_PROMOTED, 8) - values0 = 0; - 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) - src_stride_x * id0 + ((id2 * STRIDE_Y) - PAD_TOP) * (int)src_stride_z; - - weights_addr += id0 * weights_stride_w; - - const int coordy = ((id2 * STRIDE_Y) - PAD_TOP); - for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) - { -#if PAD_TOP > 0 - if(coordy < 0) // special case Z = -1 doesn't exists - { - //skip first row and load the two next ones - CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z)); - } - else if(coordy == (SRC_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. - CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z)); - } - else - { - CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z)); - } -#else // PAD_TOP > 0 - CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z)); - CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z)); -#endif // PAD_TOP > 0 - src_addr += src_stride_x; - weights_addr += weights_stride_x; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - values0 = ADD_OP(values0, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0)))); -#endif /* defined(HAS_BIAS) */ - - *((__global DATA_TYPE *)(dst.ptr + 0 * dst_stride_y)) = values0.s0; - *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values0.s1; - *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values0.s2; - *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values0.s3; - *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values0.s4; - *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values0.s5; - *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values0.s6; - *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values0.s7; -} -#endif // defined(DATA_LAYOUT_NHWC) - /** This kernel performs a direct convolution to convolve the low three dimensions. * * @note This OpenCL kernel works with stride_x = 1 and 2 -- cgit v1.2.1