From 8fce496a715929372b3c448a233713d87d65f768 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Wed, 1 Sep 2021 14:05:00 +0100 Subject: Remove padding from ClPool2dKernel NCHW - Simplify NCHW kernel structure by removing old optimized paths - Merge quantized with fp kernels Resolve COMPMID-4722 Signed-off-by: Giorgio Arena Change-Id: I79016b119619aed6a6193295601cd6517f14b88c Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6183 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- SConscript | 2 - src/core/CL/cl_kernels/common/pooling_layer.cl | 390 --------------------- src/core/CL/cl_kernels/nchw/pooling_layer.cl | 328 ++++++++--------- .../CL/cl_kernels/nchw/pooling_layer_quantized.cl | 142 -------- src/gpu/cl/ClKernelLibrary.cpp | 16 +- src/gpu/cl/kernels/ClPool2dKernel.cpp | 255 +++----------- src/gpu/cl/kernels/ClPool2dKernel.h | 2 - src/gpu/cl/operators/ClPool2d.cpp | 52 +-- src/gpu/cl/operators/ClPool2d.h | 8 - 9 files changed, 197 insertions(+), 998 deletions(-) delete mode 100644 src/core/CL/cl_kernels/common/pooling_layer.cl delete mode 100644 src/core/CL/cl_kernels/nchw/pooling_layer_quantized.cl diff --git a/SConscript b/SConscript index 62ff266d34..df8f33a917 100644 --- a/SConscript +++ b/SConscript @@ -291,7 +291,6 @@ if env['opencl'] and env['embed_kernels']: 'src/core/CL/cl_kernels/common/quantization_layer.cl', 'src/core/CL/cl_kernels/common/range.cl', 'src/core/CL/cl_kernels/common/reduction_operation.cl', - 'src/core/CL/cl_kernels/common/pooling_layer.cl', 'src/core/CL/cl_kernels/common/reshape_layer.cl', 'src/core/CL/cl_kernels/common/convolution_layer.cl', 'src/core/CL/cl_kernels/common/reverse.cl', @@ -322,7 +321,6 @@ if env['opencl'] and env['embed_kernels']: 'src/core/CL/cl_kernels/nchw/normalize_planar_yuv_layer.cl', 'src/core/CL/cl_kernels/nchw/normalize_planar_yuv_layer_quantized.cl', 'src/core/CL/cl_kernels/nchw/pooling_layer.cl', - 'src/core/CL/cl_kernels/nchw/pooling_layer_quantized.cl', 'src/core/CL/cl_kernels/nchw/prior_box_layer.cl', 'src/core/CL/cl_kernels/nchw/remap.cl', 'src/core/CL/cl_kernels/nchw/reorg_layer.cl', diff --git a/src/core/CL/cl_kernels/common/pooling_layer.cl b/src/core/CL/cl_kernels/common/pooling_layer.cl deleted file mode 100644 index 5122f2c251..0000000000 --- a/src/core/CL/cl_kernels/common/pooling_layer.cl +++ /dev/null @@ -1,390 +0,0 @@ -/* - * Copyright (c) 2017-2021 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" -#include "repeat.h" -#include "tile_helpers.h" - -#if defined(POOL_AVG) || defined(POOL_L2) -#define POOL_OP(x, y) ((x) + (y)) -#else /* defined(POOL_AVG) || defined(POOL_L2) */ -#define POOL_OP(x, y) (fmax((x), (y))) -#endif /* defined(POOL_AVG) || defined(POOL_L2) */ - -#if defined(POOL_L2) -#define POW2_OP(x, vec_size) ((x) * (x)) -#else /* defined(POOL_L2) */ -#define POW2_OP(x, vec_size) (x) -#endif /* defined(POOL_L2) */ - -#define DIV_OP(x, y) (x * (1.f / y)) -#define SQRT_OP(x) sqrt((x)) - -#if STRIDE_X == 1 -#define POOLING3x3(res, input, output) POOLING3x3_STRIDE1(res, input, output) -#elif STRIDE_X == 2 /* STRIDE_X == 1 */ -#define POOLING3x3(res, input, output) POOLING3x3_STRIDE2(res, input, output) -#elif STRIDE_X == 3 /* STRIDE_X not equals 1 or 2 */ -#define POOLING3x3(res, input, output) POOLING3x3_STRIDE3(res, input, output) -#endif /* STRIDE_X == 3 */ - -#if defined(FP_MIXED_PRECISION) -#define CONVERT_TO_ACC_DATA_TYPE(x, n) CONVERT(x, VEC_DATA_TYPE(ACC_DATA_TYPE, n)) -#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) \ - CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n) -#else /* defined(FP_MIXED_PRECISION) */ -#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) vload##n(offset, ptr) -#endif /* defined(FP_MIXED_PRECISION) */ - -#define POOLING3x3_STRIDE1(res, input, output) \ - ({ \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ - data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 4); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ - data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 4); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ - data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 4); \ - data00 = POW2_OP(data00, 4); \ - data01 = POW2_OP(data01, 2); \ - data10 = POW2_OP(data10, 4); \ - data11 = POW2_OP(data11, 2); \ - data20 = POW2_OP(data20, 4); \ - data21 = POW2_OP(data21, 2); \ - \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01212323); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data01.s0, data00.s3, data01.s01); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01212323); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data11.s0, data10.s3, data11.s01); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01212323); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data21.s0, data20.s3, data21.s01); \ - \ - values00 = POOL_OP(values00, values10); \ - values01 = POOL_OP(values01, values11); \ - values00 = POOL_OP(values00, values20); \ - values01 = POOL_OP(values01, values21); \ - \ - res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \ - }) - -#define POOLING3x3_STRIDE2(res, input, output) \ - ({ \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - ACC_DATA_TYPE data01 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - ACC_DATA_TYPE data11 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - ACC_DATA_TYPE data21 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8)); \ - data00 = POW2_OP(data00, 8); \ - data01 = POW2_OP(data01, 1); \ - data10 = POW2_OP(data10, 8); \ - data11 = POW2_OP(data11, 1); \ - data20 = POW2_OP(data20, 8); \ - data21 = POW2_OP(data21, 1); \ - \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01223445); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s667, data01); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01223445); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data10.s667, data11); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01223445); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data20.s667, data21); \ - \ - values00 = POOL_OP(values00, values10); \ - values01 = POOL_OP(values01, values11); \ - values00 = POOL_OP(values00, values20); \ - values01 = POOL_OP(values01, values21); \ - \ - res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \ - }) - -#define POOLING3x3_STRIDE3(res, input, output) \ - ({ \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ - data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ - data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \ - data00 = POW2_OP(data00, 8); \ - data01 = POW2_OP(data01, 4); \ - data10 = POW2_OP(data10, 8); \ - data11 = POW2_OP(data11, 4); \ - data20 = POW2_OP(data20, 8); \ - data21 = POW2_OP(data21, 4); \ - \ - data00 = POOL_OP(data00, data10); \ - data01 = POOL_OP(data01, data11); \ - data00 = POOL_OP(data00, data20); \ - data01 = POOL_OP(data01, data21); \ - \ - res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s036, data01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s147, data01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s25, data01.s03)); \ - }) - -ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) -{ - int start_x = get_global_id(0) * stride_x - pad_x; - int start_y = get_global_id(1) * stride_y - pad_y; - const int end_x = min(start_x + pool_size_x, upper_bound_w); - const int end_y = min(start_y + pool_size_y, upper_bound_h); -#if defined(EXCLUDE_PADDING) - start_x = max(0, start_x); - start_y = max(0, start_y); -#endif /* defined(EXCLUDE_PADDING) */ - return ((end_y - start_y) * (end_x - start_x)); -} - -/** Performs a pooling function of pool size equal to 2. - * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; - * @note In case of average pooling the following information must be passed at compile time: - * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. - * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) - * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension - * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void pooling_layer_2( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - - // Load data - VEC_DATA_TYPE(ACC_DATA_TYPE, 2) - data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); - VEC_DATA_TYPE(ACC_DATA_TYPE, 2) - data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); - -#if defined(POOL_L2) - // Raise to power of 2 for L2 Pooling - data0 = POW2_OP(data0, 2); - data1 = POW2_OP(data1, 2); -#endif /* defined(POOL_L2) */ - - // Perform calculations - data0 = POOL_OP(data0, data1); - ACC_DATA_TYPE res = POOL_OP(data0.s0, data0.s1); - -#if defined(POOL_AVG) || defined(POOL_L2) - // Divide by pool region in case of average or l2 pooling - res = DIV_OP(res, calculate_avg_scale(2, 2, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y)); -#endif /* defined(POOL_AVG) || defined(POOL_L2) */ - -#if defined(POOL_L2) - // Take square root of the result in L2 pooling - res = SQRT_OP(res); -#endif /* defined(POOL_L2) */ - - // Store result - *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; -} - -/** Performs a pooling function of pool size equal to 3 - * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; - * @note In case of average pooling the following information must be passed at compile time: - * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. - * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) - * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension - * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void pooling_layer_3( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - - // Load data - VEC_DATA_TYPE(ACC_DATA_TYPE, 3) - data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); - VEC_DATA_TYPE(ACC_DATA_TYPE, 3) - data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); - VEC_DATA_TYPE(ACC_DATA_TYPE, 3) - data2 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); - -#if defined(POOL_L2) - // Raise to power of 2 for L2 Pooling - data0 = POW2_OP(data0, 3); - data1 = POW2_OP(data1, 3); - data2 = POW2_OP(data2, 3); -#endif /* defined(POOL_L2) */ - - // Perform calculations - data0 = POOL_OP(data0, data1); - data0 = POOL_OP(data0, data2); - ACC_DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2); - -#if defined(POOL_AVG) || defined(POOL_L2) - // Divide by pool region in case of average pooling - res = DIV_OP(res, calculate_avg_scale(3, 3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y)); -#endif /* defined(POOL_AVG) || defined(POOL_L2) */ - -#if defined(POOL_L2) - // Take square root of the result in L2 pooling - res = SQRT_OP(res); -#endif /* defined(POOL_L2) */ - - // Store result - *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; -} - -#if defined(POOLING3x3) - -#define CONVERT_OP(data_type) convert_##data_type##4 -#define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type) - -VEC_DATA_TYPE(ACC_DATA_TYPE, 4) -calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) -{ - int4 start_x = ((int4)get_global_id(0) * 4 + (int4)(0, 1, 2, 3)) * (int4)stride_x - (int4)pad_x; - int start_y = get_global_id(1) * stride_y - pad_y; - const int4 end_x = min(start_x + (int4)pool_size, (int4)upper_bound_w); - const int end_y = min(start_y + pool_size, upper_bound_h); -#if defined(EXCLUDE_PADDING) - start_x = max((int4)0, start_x); - start_y = max(0, start_y); -#endif /* defined(EXCLUDE_PADDING) */ - return (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(ACC_DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x)); -} - -/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3 - * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; - * @note In case of average pooling the following information must be passed at compile time: - * -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed. - * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) - * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension - * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - */ -__kernel void pooling_layer_optimized_3( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - - VEC_DATA_TYPE(ACC_DATA_TYPE, 4) - res; - - // Perform pooling 3x3 for 4 output elements - POOLING3x3(res, input, output); - -#if defined(POOL_AVG) || defined(POOL_L2) - // Divide by pool region in case of average pooling - res *= calculate_avg_scale4(3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y); -#endif /* defined(POOL_AVG) || defined(POOL_L2) */ - -#if defined(POOL_L2) - // Take square root of the result in L2 pooling - res = SQRT_OP(res); -#endif /* defined(POOL_L2) */ - - vstore4(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 4)), 0, (__global DATA_TYPE *)output.ptr); -} -#endif // defined(POOLING3x3) diff --git a/src/core/CL/cl_kernels/nchw/pooling_layer.cl b/src/core/CL/cl_kernels/nchw/pooling_layer.cl index 790ddb381a..15ad116289 100644 --- a/src/core/CL/cl_kernels/nchw/pooling_layer.cl +++ b/src/core/CL/cl_kernels/nchw/pooling_layer.cl @@ -22,13 +22,15 @@ * SOFTWARE. */ #include "helpers.h" -#include "repeat.h" -#include "tile_helpers.h" #if defined(POOL_AVG) || defined(POOL_L2) #define POOL_OP(x, y) ((x) + (y)) #else /* defined(POOL_AVG) || defined(POOL_L2) */ +#if defined(QUANTIZED) +#define POOL_OP(x, y) (max((x), (y))) +#else // defined(QUANTIZED) #define POOL_OP(x, y) (fmax((x), (y))) +#endif // defined(QUANTIZED) #endif /* defined(POOL_AVG) || defined(POOL_L2) */ #if defined(POOL_L2) @@ -40,13 +42,12 @@ #define DIV_OP(x, y) (x * (1.f / y)) #define SQRT_OP(x) sqrt((x)) -#if defined(FP_MIXED_PRECISION) +#if defined(FP_MIXED_PRECISION) || defined(QUANTIZED) #define CONVERT_TO_ACC_DATA_TYPE(x, n) CONVERT(x, VEC_DATA_TYPE(ACC_DATA_TYPE, n)) -#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) \ - CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n) -#else /* defined(FP_MIXED_PRECISION) */ +#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n) +#else /* defined(FP_MIXED_PRECISION) || defined(QUANTIZED)*/ #define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) vload##n(offset, ptr) -#endif /* defined(FP_MIXED_PRECISION) */ +#endif /* defined(FP_MIXED_PRECISION) || defined(QUANTIZED)*/ ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) @@ -66,7 +67,7 @@ ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, /** Performs a pooling function of pool size equal to N (NCHW) * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32/QASYMM8; * @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13; * @note In case of average pooling the following information must be passed at compile time: * -DPOOL_AVG must be provided otherwise max pooling will be performed. @@ -75,59 +76,93 @@ ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension * @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0 * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32/QASYMM8 + * @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] dst_stride_z Stride of the source 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 */ __kernel void pooling_layer_MxN_nchw( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst)) { - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + int id0 = get_global_id(0); + int id1 = get_global_id(1); + int id2 = get_global_id(2); + + int x_coords = (id0 * STRIDE_X) - PAD_X; + int y_coords = (id1 * STRIDE_Y) - PAD_Y; + + __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + y_coords * (int)src_stride_y + id2 * src_stride_z; VEC_DATA_TYPE(ACC_DATA_TYPE, 8) vdata = INITIAL_VALUE; ACC_DATA_TYPE sdata = INITIAL_VALUE; + const int end_x = min((int)POOL_SIZE_X, (int)(SRC_WIDTH - x_coords)); + const int end_y = min((int)POOL_SIZE_Y, (int)(SRC_HEIGHT - y_coords)); + // Load data - for(int y = 0; y < POOL_SIZE_Y; y++) + for(int y = 0; y < end_y; ++y) { - int x = 0; - for(; x <= ((int)POOL_SIZE_X - 8); x += 8) + if((y_coords + y) >= 0) { - VEC_DATA_TYPE(ACC_DATA_TYPE, 8) - data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); + int x = 0; + for(; x <= (end_x - 8); x += 8) + { + int8 src_x = (int8)(x_coords + x) + VEC_OFFS(int, 8); +#if defined(POOL_AVG) || defined(POOL_L2) + SELECT_VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + cond_x = CONVERT(src_x < 0, SELECT_VEC_DATA_TYPE(ACC_DATA_TYPE, 8)); + src_x = clamp(src_x, (int8)0, (int8)(SRC_WIDTH - 1)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + data0 = select(VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)(src_addr + src_x.s0 * sizeof(DATA_TYPE) + y * src_stride_y)), (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))0, REVERSE(cond_x, 8)); +#else // defined(POOL_AVG) || defined(POOL_L2) + src_x = clamp(src_x, 0, SRC_WIDTH - 1); + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)(src_addr + src_x.s0 * sizeof(DATA_TYPE) + y * src_stride_y)); +#endif // defined(POOL_AVG) || defined(POOL_L2 + #if defined(POOL_L2) - // Raise to power of 2 for L2 Pooling - data0 *= data0; + // Raise to power of 2 for L2 Pooling + data0 *= data0; #endif /* defined(POOL_L2) */ - vdata = POOL_OP(vdata, data0); - } - // Leftover - for(; x < (int)POOL_SIZE_X; ++x) - { - ACC_DATA_TYPE data0 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0))); + vdata = POOL_OP(vdata, data0); + } + + // Leftover + for(; x < end_x; ++x) + { + int src_x = x_coords + x; +#if defined(POOL_AVG) || defined(POOL_L2) + SELECT_DATA_TYPE(ACC_DATA_TYPE) + cond_x = (src_x < 0); + src_x = clamp(src_x, 0, SRC_WIDTH - 1); + ACC_DATA_TYPE data0 = select((ACC_DATA_TYPE)(*((__global DATA_TYPE *)(src_addr + src_x * sizeof(DATA_TYPE) + y * src_stride_y))), (ACC_DATA_TYPE)0, cond_x); +#else // defined(POOL_AVG) || defined(POOL_L2) + src_x = clamp(src_x, 0, SRC_WIDTH - 1); + ACC_DATA_TYPE data0 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)(src_addr + src_x * sizeof(DATA_TYPE) + y * src_stride_y))); +#endif // defined(POOL_AVG) || defined(POOL_L2) + #if defined(POOL_L2) - // Raise to power of 2 for L2 Pooling - data0 *= data0; + // Raise to power of 2 for L2 Pooling + data0 *= data0; #endif /* defined(POOL_L2) */ - sdata = POOL_OP(sdata, data0); + + sdata = POOL_OP(sdata, data0); + } } } @@ -144,76 +179,61 @@ __kernel void pooling_layer_MxN_nchw( res = DIV_OP(res, calculate_avg_scale(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y)); #endif /* defined(POOL_AVG) || defined(POOL_L2) */ +#if defined(QUANTIZED) + + DATA_TYPE result_q8 = CONVERT(res, DATA_TYPE); + +#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) + + const float result_f32 = convert_float(result_q8); + const float input_offset = (float)OFFSET_IN1; + const float input_scale = (float)SCALE_IN1; + const float scale_out = (float)SCALE_OUT; + const float offset_out = (float)OFFSET_OUT; + const float in_f32 = (result_f32 - input_offset) * input_scale; + const float out_f32 = in_f32 / scale_out + offset_out; + result_q8 = CONVERT_SAT(convert_int_rte(out_f32), DATA_TYPE); + +#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ + + *(__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + id0 * sizeof(DATA_TYPE) + id1 * dst_stride_y + id2 * dst_stride_z) = result_q8; + +#else // defined(QUANTIZED) + #if defined(POOL_L2) // Take square root of the result in L2 pooling res = SQRT_OP(res); #endif /* defined(POOL_L2) */ // Store result - *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; + *(__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + id0 * sizeof(DATA_TYPE) + id1 * dst_stride_y + id2 * dst_stride_z) = (DATA_TYPE)res; +#endif // defined(QUANTIZED) } #endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) -#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) - -inline void offset_no_padding_nchw(const Tensor3D *input, uint *offset_top, uint *offset_bottom) -{ - const int pad_horiz = PAD_TENSOR_LEFT + PAD_TENSOR_RIGHT; - const int pad_vert = PAD_TENSOR_TOP + PAD_TENSOR_BOTTOM; - - const int x = get_global_id(0) * STRIDE_X; - const int y = get_global_id(1) * STRIDE_Y; - const int z = get_global_id(2); - - //x axis: width, y axis: height, z axis: component - const uint padded_offset = input->offset_first_element_in_bytes - + x * input->stride_x - + y * input->stride_y - + z * input->stride_z; - - const uint offset_base = padded_offset - - y * pad_horiz * sizeof(DATA_TYPE) /* Horizontal padding for each row */ - - PAD_TENSOR_TOP * input->stride_y /* top padding */ - - z * MAX_HEIGHT * pad_horiz * sizeof(DATA_TYPE) - z * pad_vert * input->stride_y /* Z plane padding */ - - PAD_TENSOR_LEFT * sizeof(DATA_TYPE); - -#if defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT) - *offset_top = (uint)((offset_base / sizeof(DATA_TYPE)) % (TENSOR_CHANNEL * TENSOR_WIDTH * TENSOR_HEIGHT)); -#else /* defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT) */ - *offset_top = (uint)(offset_base / sizeof(DATA_TYPE)); -#endif /* defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT) */ - - *offset_bottom = *offset_top + input->stride_y / sizeof(DATA_TYPE) - pad_horiz; - - return; -} - -#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) - /** Performs a MAX pooling of pool size equal to 2, and record max value indices for NCHW. * * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F32 * @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13; * @note Tensors width and height must be passed at compile time using -DMAX_WIDTH and -DMAX_HEIGHT * @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * @note Tensor padding values must be passed at compile time using PAD_TENSOR_LEFT, PAD_TENSOR_RIGHT, PAD_TENSOR_TOP and PAD_TENSOR_BOTTOM * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @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 Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the source 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] indices_ptr Pointer to the indices tensor. Supported data types: U32 * @param[in] indices_stride_x Stride of the indices tensor in X dimension (in bytes) * @param[in] indices_step_x indices_stride_x * number of elements along X processed per workitem(in bytes) @@ -223,109 +243,43 @@ inline void offset_no_padding_nchw(const Tensor3D *input, uint *offset_top, uint * @param[in] indices_step_z indices_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] indices_offset_first_element_in_bytes The offset of the first element in the indices tensor */ -__kernel void pooling_layer_2_nchw_indices_fp32( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output), +__kernel void pooling_layer_2_nchw_indices( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), TENSOR3D_DECLARATION(indices)) { - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - Tensor3D indices = CONVERT_TO_TENSOR3D_STRUCT(indices); - - // Load data - float2 data0 = VLOAD(2)(0, (__global float *)tensor3D_offset(&input, 0, 0, 0)); - float2 data1 = VLOAD(2)(0, (__global float *)tensor3D_offset(&input, 0, 1, 0)); - - // Perform calculations - float data0_max = POOL_OP(data0.s0, data0.s1); - float data1_max = POOL_OP(data1.s0, data1.s1); - float res = POOL_OP(data0_max, data1_max); - // Store result - *(__global float *)output.ptr = res; + int id0 = get_global_id(0); + int id1 = get_global_id(1); + int id2 = get_global_id(2); -#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) + int2 x_coords = clamp((int2)((id0 * STRIDE_X) - PAD_X), (int2)0, (int2)(SRC_WIDTH - 1)); + int2 y_coords = clamp((int2)((id1 * STRIDE_Y) - PAD_Y) + VEC_OFFS(int, 2), (int2)0, (int2)(SRC_HEIGHT - 1)); - uint offset_top = 0; - uint offset_bottom = 0; - - offset_no_padding_nchw(&input, &offset_top, &offset_bottom); - - uint index0 = select(offset_top + 1, offset_top, isgreaterequal(data0.s0, data0.s1)); - uint index1 = select(offset_bottom + 1, offset_bottom, isgreaterequal(data1.s0, data1.s1)); - uint index = select(index1, index0, isgreaterequal(data0_max, data1_max)); - - *(__global uint *)indices.ptr = index; - -#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) -} - -/** Performs a MAX pooling of pool size equal to 2, and record max value indices for NCHW. - * - * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F16 - * @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13; - * @note Tensors width and height must be passed at compile time using -DMAX_WIDTH and -DMAX_HEIGHT - * @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * @note Tensor padding values must be passed at compile time using PAD_TENSOR_LEFT, PAD_TENSOR_RIGHT, PAD_TENSOR_TOP and PAD_TENSOR_BOTTOM - * - * @param[in] input_ptr Pointer to the source tensor. Supported data types: F16 - * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] indices_ptr Pointer to the indices tensor. Supported data types: U32 - * @param[in] indices_stride_x Stride of the indices tensor in X dimension (in bytes) - * @param[in] indices_step_x indices_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] indices_stride_y Stride of the indices tensor in Y dimension (in bytes) - * @param[in] indices_step_y indices_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] indices_stride_z Stride of the indices tensor in Z dimension (in bytes) - * @param[in] indices_step_z indices_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] indices_offset_first_element_in_bytes The offset of the first element in the indices tensor - */ -__kernel void pooling_layer_2_nchw_indices_fp16( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output), - TENSOR3D_DECLARATION(indices)) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - Tensor3D indices = CONVERT_TO_TENSOR3D_STRUCT(indices); + __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + id2 * src_stride_z; // Load data - half2 data0 = VLOAD(2)(0, (__global half *)tensor3D_offset(&input, 0, 0, 0)); - half2 data1 = VLOAD(2)(0, (__global half *)tensor3D_offset(&input, 0, 1, 0)); + VEC_DATA_TYPE(DATA_TYPE, 2) + data0 = VLOAD(2)(0, (__global DATA_TYPE *)(src_addr + x_coords.s0 * sizeof(DATA_TYPE) + y_coords.s0 * (int)src_stride_y)); + VEC_DATA_TYPE(DATA_TYPE, 2) + data1 = VLOAD(2)(0, (__global DATA_TYPE *)(src_addr + x_coords.s1 * sizeof(DATA_TYPE) + y_coords.s1 * (int)src_stride_y)); // Perform calculations - half data0_max = POOL_OP(data0.s0, data0.s1); - half data1_max = POOL_OP(data1.s0, data1.s1); - half res = POOL_OP(data0_max, data1_max); + DATA_TYPE data0_max = POOL_OP(data0.s0, data0.s1); + DATA_TYPE data1_max = POOL_OP(data1.s0, data1.s1); + DATA_TYPE res = POOL_OP(data0_max, data1_max); // Store result - *(__global half *)output.ptr = res; - -#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) + *(__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + id0 * sizeof(DATA_TYPE) + id1 * dst_stride_y + id2 * dst_stride_z) = res; - uint offset_top = 0; - uint offset_bottom = 0; +#if defined(SRC_BATCH) - offset_no_padding_nchw(&input, &offset_top, &offset_bottom); + uint offset_top = (x_coords.s0 + y_coords.s0 * SRC_WIDTH + id2 * (SRC_WIDTH * SRC_HEIGHT)) % SRC_BATCH; + uint offset_bottom = offset_top + SRC_WIDTH; uint index0 = select(offset_top + 1, offset_top, isgreaterequal(data0.s0, data0.s1)); uint index1 = select(offset_bottom + 1, offset_bottom, isgreaterequal(data1.s0, data1.s1)); uint index = select(index1, index0, isgreaterequal(data0_max, data1_max)); - *(__global uint *)indices.ptr = index; + *(__global uint *)(indices_ptr + indices_offset_first_element_in_bytes + id0 * sizeof(uint) + id1 * indices_stride_y + id2 * indices_stride_z) = index; -#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM) +#endif // defined(SRC_BATCH) } \ No newline at end of file diff --git a/src/core/CL/cl_kernels/nchw/pooling_layer_quantized.cl b/src/core/CL/cl_kernels/nchw/pooling_layer_quantized.cl deleted file mode 100644 index 1440ef3ed1..0000000000 --- a/src/core/CL/cl_kernels/nchw/pooling_layer_quantized.cl +++ /dev/null @@ -1,142 +0,0 @@ -/* - * Copyright (c) 2017-2021 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" - -#if defined(DATA_TYPE) && defined(INITIAL_VALUE) -#define VEC_TYPE(VEC_SIZE) VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - -#if defined(POOL_AVG) -#define POOL_OP(x, y) ((x) + (y)) -#else /* defined(POOL_AVG) */ -#define POOL_OP(x, y) (max((x), (y))) -#endif /* defined(POOL_AVG) */ - -#define DIV_OP(x, y) (x * (1.f / y)) - -#if defined(POOL_L2) -#error "L2 pooling is not supported" -#endif /* defined(POOL_L2) */ - -int calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) -{ - int start_x = get_global_id(0) * stride_x - pad_x; - int start_y = get_global_id(1) * stride_y - pad_y; - const int end_x = min(start_x + pool_size_x, upper_bound_w); - const int end_y = min(start_y + pool_size_y, upper_bound_h); -#if defined(EXCLUDE_PADDING) - start_x = max(0, start_x); - start_y = max(0, start_y); -#endif /* defined(EXCLUDE_PADDING) */ - return ((end_y - start_y) * (end_x - start_x)); -} - -/** Performs a pooling function of pool size equal to N (NCHW) - * - * @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13; - * @note In case of average pooling the following information must be passed at compile time: - * -DPOOL_AVG must be provided otherwise max pooling will be performed. - * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) - * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions - * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension - * @note Input data type must be passed at compile time using -DDAT_TYPE=type, e.g. -DDATA_TYPE=uchar - * @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0 - * - * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED - * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image - */ -__kernel void pooling_layer_MxN_quantized_nchw( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) -{ - // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - - int8 vdata = INITIAL_VALUE; - int sdata = INITIAL_VALUE; - - // Load data - for(int y = 0; y < POOL_SIZE_Y; y++) - { - int x = 0; - for(; x <= ((int)POOL_SIZE_X - 8); x += 8) - { - VEC_TYPE(8) - data = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); - int8 data0 = convert_int8(data); - vdata = POOL_OP(vdata, data0); - } - - // Leftover - for(; x < (int)POOL_SIZE_X; ++x) - { - DATA_TYPE data = *((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); - int data0 = convert_int(data); - sdata = POOL_OP(sdata, data0); - } - } - - // Reduce result - int4 reduce4 = POOL_OP(vdata.s0123, vdata.s4567); - int2 reduce2 = POOL_OP(reduce4.s01, reduce4.s23); - int res = POOL_OP(reduce2.s0, reduce2.s1); - res = POOL_OP(res, sdata); - -#if defined(POOL_AVG) - res = round(DIV_OP(res, calculate_avg_scale(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y))); -#endif /* defined(POOL_AVG) */ - - DATA_TYPE result_q8 = CONVERT(res, DATA_TYPE); - -#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) - - const float result_f32 = convert_float(result_q8); - const float input_offset = (float)OFFSET_IN1; - const float input_scale = (float)SCALE_IN1; - const float scale_out = (float)SCALE_OUT; - const float offset_out = (float)OFFSET_OUT; - const float in_f32 = (result_f32 - input_offset) * input_scale; - const float out_f32 = in_f32 / scale_out + offset_out; - result_q8 = CONVERT_SAT(convert_int_rte(out_f32), DATA_TYPE); - -#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ - - *(__global DATA_TYPE *)output.ptr = result_q8; -} -#endif // defined(DATA_TYPE) && defined(INITIAL_VALUE) \ No newline at end of file diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp index 5cd969e7f2..c05bb96753 100644 --- a/src/gpu/cl/ClKernelLibrary.cpp +++ b/src/gpu/cl/ClKernelLibrary.cpp @@ -328,10 +328,6 @@ const std::map ClKernelLibrary::_kernel_program_map = { "pixelwise_mul_float", "common/pixelwise_mul_float.cl" }, { "pixelwise_mul_int", "common/pixelwise_mul_int.cl" }, { "pixelwise_mul_quantized", "common/pixelwise_mul_int.cl" }, - { "pooling_layer_2", "common/pooling_layer.cl" }, - { "pooling_layer_3", "common/pooling_layer.cl" }, - { "pooling_layer_optimized_3", "common/pooling_layer.cl" }, - { "pooling_layer_7", "common/pooling_layer.cl" }, { "qlstm_layer_normalization", "common/qlstm_layer_normalization.cl" }, { "quantization_layer", "common/quantization_layer.cl" }, { "range", "common/range.cl" }, @@ -385,9 +381,7 @@ const std::map ClKernelLibrary::_kernel_program_map = { "normalize_planar_yuv_layer_nchw", "nchw/normalize_planar_yuv_layer.cl" }, { "normalize_planar_yuv_layer_q8_nchw", "nchw/normalize_planar_yuv_layer_quantized.cl" }, { "pooling_layer_MxN_nchw", "nchw/pooling_layer.cl" }, - { "pooling_layer_2_nchw_indices_fp32", "nchw/pooling_layer.cl" }, - { "pooling_layer_2_nchw_indices_fp16", "nchw/pooling_layer.cl" }, - { "pooling_layer_MxN_quantized_nchw", "nchw/pooling_layer_quantized.cl" }, + { "pooling_layer_2_nchw_indices", "nchw/pooling_layer.cl" }, { "prior_box_layer_nchw", "nchw/prior_box_layer.cl" }, { "remap_nearest_neighbour_nchw", "nchw/remap.cl" }, { "remap_bilinear_nchw", "nchw/remap.cl" }, @@ -666,10 +660,6 @@ const std::map ClKernelLibrary::_program_source_map = { "common/pixelwise_mul_int.cl", #include "./cl_kernels/common/pixelwise_mul_int.clembed" - }, - { - "common/pooling_layer.cl", -#include "./cl_kernels/common/pooling_layer.clembed" }, { "common/qlstm_layer_normalization.cl", @@ -803,10 +793,6 @@ const std::map ClKernelLibrary::_program_source_map = { "nchw/pooling_layer.cl", #include "./cl_kernels/nchw/pooling_layer.clembed" - }, - { - "nchw/pooling_layer_quantized.cl", -#include "./cl_kernels/nchw/pooling_layer_quantized.clembed" }, { "nchw/prior_box_layer.cl", diff --git a/src/gpu/cl/kernels/ClPool2dKernel.cpp b/src/gpu/cl/kernels/ClPool2dKernel.cpp index 04f2b142bd..5e53799f30 100644 --- a/src/gpu/cl/kernels/ClPool2dKernel.cpp +++ b/src/gpu/cl/kernels/ClPool2dKernel.cpp @@ -23,18 +23,13 @@ */ #include "src/gpu/cl/kernels/ClPool2dKernel.h" -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "support/Cast.h" -#include "support/StringSupport.h" namespace arm_compute { @@ -46,19 +41,6 @@ using namespace arm_compute::misc::shape_calculator; namespace { -// Internal window config info -using ClPoolingConfig = std::pair; //num_elems_processed_per_iteration, border_size - -void auto_init(const ITensorInfo *src, ITensorInfo *dst, ITensorInfo *indices, PoolingLayerInfo pool_info) -{ - TensorShape out_shape = compute_pool_shape(*src, pool_info); - auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape)); - if(indices) - { - auto_init_if_empty(*indices, src->clone()->set_tensor_shape(out_shape).set_data_type(DataType::U32)); - } -} - Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); @@ -104,102 +86,6 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const return Status{}; } - -std::tuple validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - - // Get data layout - const DataLayout data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - int pool_stride_x = 0; - int pool_stride_y = 0; - unsigned int pooled_w = 0; - unsigned int pooled_h = 0; - int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; - int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - const int pool_pad_right = pad_stride_info.pad_right(); - const int pool_pad_top = pad_stride_info.pad_top(); - const int pool_pad_left = pad_stride_info.pad_left(); - const int pool_pad_bottom = pad_stride_info.pad_bottom(); - BorderSize border_size = BorderSize(); - - auto_init(src, dst, indices, pool_info); - pooled_w = dst->tensor_shape()[idx_width]; - pooled_h = dst->tensor_shape()[idx_height]; - - const DataType data_type = src->data_type(); - - const int src_width = src->dimension(idx_width); - const int src_height = src->dimension(idx_height); - - unsigned int num_elems_processed_per_iteration = 0; - bool window_changed = false; - Window win{}; - switch(data_layout) - { - case DataLayout::NCHW: - { - // Initialize border size - border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left); - // Change the number of elements processed per iteration - // for pooling 3x3 with stride less equal than 3 - const bool can_optimize = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3) && !is_data_type_quantized(data_type); - num_elems_processed_per_iteration = can_optimize ? 4 : 1; - const unsigned int num_elems_read_per_iteration = (num_elems_processed_per_iteration - 1) * pool_stride_x + pool_size_x; - - // Number of iterations in X dimension - const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; - - // Upper limit for the number of right/bottom border elements that are accessed - const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - src_width; - const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - src_height; - - border_size.right = std::max(upper_bound_w, pool_pad_right); - border_size.bottom = std::max(upper_bound_h, pool_pad_bottom); - - win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - - AccessWindowRectangle src_access(src, -pool_pad_left, -pool_pad_top, num_elems_read_per_iteration, pool_size_y, - pool_stride_x, pool_stride_y); - AccessWindowHorizontal dst_access(dst, 0, num_elems_processed_per_iteration); - - // Update indices window - if(indices) - { - AccessWindowHorizontal indices_access(indices, 0, num_elems_processed_per_iteration); - window_changed = update_window_and_padding(win, src_access, dst_access, indices_access); - indices_access.set_valid_region(win, ValidRegion(Coordinates(), indices->tensor_shape())); - } - else - { - window_changed = update_window_and_padding(win, src_access, dst_access); - } - - dst_access.set_valid_region(win, ValidRegion(Coordinates(), dst->tensor_shape())); - break; - } - case DataLayout::NHWC: - { - const size_t vec_size = dst->data_type() == DataType::F32 ? 2 : 4; - - // Initialize border size - border_size = BorderSize(); - num_elems_processed_per_iteration = adjust_vec_size(vec_size, dst->dimension(0)); - win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - break; - } - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_tuple(err, win, ClPoolingConfig(num_elems_processed_per_iteration, border_size)); -} } // namespace ClPool2dKernel::ClPool2dKernel() @@ -207,20 +93,27 @@ ClPool2dKernel::ClPool2dKernel() _type = CLKernelType::POOL; } -BorderSize ClPool2dKernel::border_size() const -{ - return _border_size; -} - void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, indices)); auto padding_info = get_padding_info({ src, dst, indices }); + // Auto init if empty + TensorShape out_shape = compute_pool_shape(*src, pool_info); + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape)); + if(indices) + { + auto_init_if_empty(*indices, src->clone()->set_tensor_shape(out_shape).set_data_type(DataType::U32)); + } + // Set instance variables - _pool_info = pool_info; - _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; + _pool_info = pool_info; + _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; + _num_elems_processed_per_iteration = (_data_layout == DataLayout::NCHW) ? 1 : ((dst->data_type() == DataType::F32) ? 2 : 4); + _num_elems_processed_per_iteration = adjust_vec_size(_num_elems_processed_per_iteration, dst->dimension(0)); + int pool_stride_x = 0; int pool_stride_y = 0; const PoolingType pool_type = pool_info.pool_type; @@ -233,61 +126,47 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; const bool exclude_padding = pool_info.exclude_padding; std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - const int pool_pad_top = pad_stride_info.pad_top(); - const int pool_pad_left = pad_stride_info.pad_left(); + const int pool_pad_top = pad_stride_info.pad_top(); + const int pool_pad_left = pad_stride_info.pad_left(); + const DataType data_type = src->data_type(); // Set build options CLBuildOptions build_opts; - const DataType data_type = src->data_type(); - - // Configure kernel window - auto win_config = validate_and_configure_window(src, dst, pool_info, indices); - - ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); - ICLKernel::configure_internal(std::get<1>(win_config)); - - ClPoolingConfig pooling_config = std::get<2>(win_config); - _num_elems_processed_per_iteration = pooling_config.first; - _border_size = pooling_config.second; - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration)); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); + build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type)); + build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x)); + build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y)); + build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left)); + build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top)); + build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x)); + build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y)); + build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width))); + build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height))); + build_opts.add_option("-DMAX_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width) + (exclude_padding ? 0 : pool_pad_left))); + build_opts.add_option("-DMAX_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height) + (exclude_padding ? 0 : pool_pad_top))); // Tensor paddings are used to calculate the indicies for MAX pooling if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type)) { - build_opts.add_option("-DPAD_TENSOR_LEFT=" + support::cpp11::to_string(src->padding().left)); - build_opts.add_option("-DPAD_TENSOR_RIGHT=" + support::cpp11::to_string(src->padding().right)); - build_opts.add_option("-DPAD_TENSOR_TOP=" + support::cpp11::to_string(src->padding().top)); - build_opts.add_option("-DPAD_TENSOR_BOTTOM=" + support::cpp11::to_string(src->padding().bottom)); - build_opts.add_option("-DTENSOR_CHANNEL=" + support::cpp11::to_string(src->dimension(idx_channel))); - build_opts.add_option("-DTENSOR_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width))); - build_opts.add_option("-DTENSOR_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height))); + build_opts.add_option("-DSRC_BATCH=" + support::cpp11::to_string(src->tensor_shape().total_size_lower(3))); } - if(is_data_type_quantized_asymmetric(data_type) && src->quantization_info() != dst->quantization_info()) + if(is_data_type_quantized_asymmetric(data_type)) { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Check dst dimensions - auto_init(src, dst, indices, pool_info); + build_opts.add_option("-DQUANTIZED"); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, indices)); + if(src->quantization_info() != dst->quantization_info()) + { + const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); - build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type)); - build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x)); - build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y)); - build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left)); - build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top)); - build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x)); - build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y)); + build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); + build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); + build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); + build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); + } + } // Set the initial value for the pooling operation accordingly with the data type if(pool_type == PoolingType::MAX) @@ -309,9 +188,6 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI build_opts.add_option("-DINITIAL_VALUE=0"); } - build_opts.add_option("-DMAX_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width) + (exclude_padding ? 0 : pool_pad_left))); - build_opts.add_option("-DMAX_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height) + (exclude_padding ? 0 : pool_pad_top))); - // Create kernel switch(_data_layout) { @@ -319,7 +195,7 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI { const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision; const auto use_wider_accumulator = use_fp_mixed_precision && (pool_type != PoolingType::MAX); - const auto acc_data_type = get_cl_type_from_data_type(use_wider_accumulator ? DataType::F32 : data_type); + const auto acc_data_type = get_cl_type_from_data_type(use_wider_accumulator ? DataType::F32 : (is_data_type_quantized(data_type) ? DataType::S32 : data_type)); build_opts.add_option("-DACC_DATA_TYPE=" + acc_data_type); build_opts.add_option_if(use_wider_accumulator, "-DFP_MIXED_PRECISION"); @@ -328,33 +204,15 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING"); } - if((pool_size_x == 3) && (pool_size_y == 3) && !is_data_type_quantized_asymmetric(data_type)) - { - // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where - // each thread computes 4 dst elements - const bool is_pool3x3_stride_le3 = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3); - - std::string kernel_name = ((is_pool3x3_stride_le3) ? "pooling_layer_optimized_" : "pooling_layer_") - + support::cpp11::to_string(pool_size_x); - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - else if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type)) + if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type)) { // For max pooling with pool2x2, store indicies which will be used in max unpooling - if(data_type == DataType::F32) - { - std::string kernel_name = "pooling_layer_2_nchw_indices_fp32"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - else if(data_type == DataType::F16) - { - std::string kernel_name = "pooling_layer_2_nchw_indices_fp16"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } + std::string kernel_name = "pooling_layer_2_nchw_indices"; + _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); } else // Run general case { - std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nchw" : "pooling_layer_MxN_nchw"; + std::string kernel_name = "pooling_layer_MxN_nchw"; _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); } break; @@ -405,6 +263,10 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI ARM_COMPUTE_ERROR("Not implemented"); } + // Configure kernel window + Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration)); + ICLKernel::configure_internal(win); + // Set config_id for enabling LWS tuning _config_id = "pooling_layer_"; _config_id += lower_string(string_from_data_type(data_type)); @@ -419,14 +281,12 @@ void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorI _config_id += "_"; _config_id += lower_string(string_from_data_layout(src->data_layout())); - ARM_COMPUTE_ERROR_ON(src->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info)); + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } Status ClPool2dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info, indices)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(src->clone().get(), dst->clone().get(), pool_info))); - return Status{}; } @@ -453,18 +313,9 @@ void ClPool2dKernel::run_op(ITensorPack &tensors, const Window &window, cl::Comm Window slice = window_collapsed.first_slice_window_3D(); do { - // Upsample src by pool size - Window in_slice(slice); - in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start() - _pool_info.pad_stride_info.pad_left(), - (in_slice.x().end() - _pool_info.pad_stride_info.pad_left()) * pool_stride_x, - pool_stride_x * _num_elems_processed_per_iteration)); - in_slice.set(Window::DimY, Window::Dimension(in_slice.y().start() - _pool_info.pad_stride_info.pad_top(), - (in_slice.y().end() - _pool_info.pad_stride_info.pad_top()) * pool_stride_y, - pool_stride_y)); - // Set srcs unsigned int idx = 0; - add_3D_tensor_argument(idx, src, in_slice); + add_3D_tensor_argument(idx, src, slice); add_3D_tensor_argument(idx, dst, slice); if(indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_size == Size2D(2, 2))) { diff --git a/src/gpu/cl/kernels/ClPool2dKernel.h b/src/gpu/cl/kernels/ClPool2dKernel.h index 61d204dc68..f5bb0687e8 100644 --- a/src/gpu/cl/kernels/ClPool2dKernel.h +++ b/src/gpu/cl/kernels/ClPool2dKernel.h @@ -61,12 +61,10 @@ public: // Inherited methods overridden: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - BorderSize border_size() const override; public: PoolingLayerInfo _pool_info{}; DataLayout _data_layout{ DataLayout::UNKNOWN }; - BorderSize _border_size{ 0 }; unsigned int _num_elems_processed_per_iteration{ 1 }; }; } // namespace kernels diff --git a/src/gpu/cl/operators/ClPool2d.cpp b/src/gpu/cl/operators/ClPool2d.cpp index fdadd199fc..a5b18a2340 100644 --- a/src/gpu/cl/operators/ClPool2d.cpp +++ b/src/gpu/cl/operators/ClPool2d.cpp @@ -25,7 +25,6 @@ #include "arm_compute/runtime/CL/CLScheduler.h" -#include "src/core/CL/kernels/CLFillBorderKernel.h" #include "src/gpu/cl/ClCompileContext.h" #include "src/gpu/cl/kernels/ClPool2dKernel.h" @@ -40,62 +39,15 @@ void ClPool2d::configure(const ClCompileContext &compile_context, ITensorInfo *s auto k = std::make_unique(); k->set_target(CLScheduler::get().target()); k->configure(compile_context, src, dst, info, indices); - _pooling = std::move(k); - - const DataType data_type = src->data_type(); - - // Configure border depending on operation required (quantize border in case of asymmetric data_type) - BorderMode border_mode{}; - PixelValue pixel_value(0.f); - if(is_data_type_quantized_asymmetric(data_type) && !info.exclude_padding) - { - pixel_value = PixelValue(0, data_type, src->quantization_info()); - } - - // Data layout - const auto data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout; - - switch(data_layout) - { - case DataLayout::NCHW: - border_mode = (PoolingType::MAX == info.pool_type) ? BorderMode::REPLICATE : BorderMode::CONSTANT; - break; - case DataLayout::NHWC: - border_mode = BorderMode::CONSTANT; - if(PoolingType::MAX == info.pool_type) - { - if(is_data_type_quantized(data_type)) - { - std::tie(pixel_value, std::ignore) = get_min_max(data_type); - } - else - { - pixel_value = PixelValue(std::numeric_limits::lowest()); - } - } - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } - auto b = std::make_unique(); - b->configure(compile_context, src, _pooling->border_size(), border_mode, pixel_value); - _border_handler = std::move(b); + _kernel = std::move(k); // Tune kernels - CLScheduler::get().tune_kernel_static(*_pooling); + CLScheduler::get().tune_kernel_static(*_kernel); } Status ClPool2d::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info, const ITensorInfo *indices) { return kernels::ClPool2dKernel::validate(src, dst, info, indices); } - -void ClPool2d::run(ITensorPack &tensors) -{ - ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided"); - - CLScheduler::get().enqueue_op(*_border_handler.get(), tensors, false); - CLScheduler::get().enqueue_op(*_pooling.get(), tensors, false); -} } // namespace opencl } // namespace arm_compute diff --git a/src/gpu/cl/operators/ClPool2d.h b/src/gpu/cl/operators/ClPool2d.h index a041053bb3..f353ba262e 100644 --- a/src/gpu/cl/operators/ClPool2d.h +++ b/src/gpu/cl/operators/ClPool2d.h @@ -35,7 +35,6 @@ namespace opencl { /** Basic function to simulate a pooling layer with the specified pooling operation. This function calls the following OpenCL kernels: * - * -# @ref CLFillBorderKernel (executed if padding size is different from zero) * -# @ref opencl::ClPool2d */ class ClPool2d : public IClOperator @@ -59,13 +58,6 @@ public: * @return a status */ static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info, const ITensorInfo *indices = nullptr); - - // Inherited method overridden - void run(ITensorPack &tensors) override; - -private: - std::unique_ptr _pooling{ nullptr }; - std::unique_ptr _border_handler{ nullptr }; }; } // namespace opencl } // namespace arm_compute -- cgit v1.2.1