From 2d7e683e79c8ad328d4930c1f82a46827313faf4 Mon Sep 17 00:00:00 2001 From: George Wort Date: Fri, 22 Feb 2019 16:37:41 +0000 Subject: COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047 Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/561 Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins --- ...GEMMLowpOffsetContributionOutputStageKernel.cpp | 3 +- .../kernels/NEGEMMLowpOffsetContributionKernel.cpp | 11 +- ...GEMMLowpOffsetContributionOutputStageKernel.cpp | 651 +++++++++++++++++++++ ...tizeDownInt32ToUint8ScaleByFixedPointKernel.cpp | 39 +- .../NEON/functions/NEGEMMConvolutionLayer.cpp | 241 ++++---- .../functions/NEGEMMLowpMatrixMultiplyCore.cpp | 215 ++++--- 6 files changed, 903 insertions(+), 257 deletions(-) create mode 100644 src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp (limited to 'src') diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp index 83af0c63eb..8fba342e74 100644 --- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,7 +51,6 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(output_stage.type == GEMMLowpOutputStageType::NONE); - ARM_COMPUTE_RETURN_ERROR_ON(bias == nullptr && a_offset == 0 && b_offset == 0); ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > 255); ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < 0 || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound); diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp index 33a5b4ace3..22939266e5 100644 --- a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -106,20 +106,17 @@ std::pair validate_and_configure_window(ITensorInfo *mm_result, Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal mm_result_access(mm_result, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, - mm_result_access); + window_changed = window_changed || update_window_and_padding(win, mm_result_access); if(a_offset != 0) { AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, - vector_sum_col_access); + window_changed = window_changed || update_window_and_padding(win, vector_sum_col_access); } if(b_offset != 0) { AccessWindowStatic vector_sum_row_access(vector_sum_row, 0, 0, vector_sum_row->dimension(0), 0); // NOLINT - window_changed = window_changed || update_window_and_padding(win, - vector_sum_row_access); + window_changed = window_changed || update_window_and_padding(win, vector_sum_row_access); } Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp new file mode 100644 index 0000000000..ebbea083e3 --- /dev/null +++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp @@ -0,0 +1,651 @@ +/* + * 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 "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/NEAsymm.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +#include +#include +#include +#include + +namespace arm_compute +{ +class Coordinates; + +namespace +{ +inline int32x4x4_t load_results_input(const Iterator &mm_result_it, int32_t x) +{ + return + { + { + vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 0), + vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 4), + vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 8), + vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 12) + } + }; +} + +inline int32x4x4_t load(const int32_t *ptr, int32_t x) +{ + return + { + { + vld1q_s32(ptr + x + 0), + vld1q_s32(ptr + x + 4), + vld1q_s32(ptr + x + 8), + vld1q_s32(ptr + x + 12) + } + }; +} + +inline int32x4x4_t get_a_offset(const int32_t *vector_sum_col_ptr, int32_t a_offset, int32_t x) +{ + int32x4x4_t a_offset_term_s32 = load(vector_sum_col_ptr, x); + + a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset); + a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset); + a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset); + a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset); + return a_offset_term_s32; +} + +inline int32x4_t get_b_offset(const int32_t *vector_sum_row_ptr, int32_t b_offset) +{ + int32x4_t b_offset_term_s32 = vld1q_dup_s32(vector_sum_row_ptr); + b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset); + return b_offset_term_s32; +} + +inline int32x4x4_t get_k_offset(int32_t k_offset) +{ + return + { + { + vdupq_n_s32(k_offset), + vdupq_n_s32(k_offset), + vdupq_n_s32(k_offset), + vdupq_n_s32(k_offset) + } + }; +} + +template +inline uint8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4_t result_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8) +{ + const static int32x4_t zero_s32 = vdupq_n_s32(0); + + // Shift final result (negative value shift right) + in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); + in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); + in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); + in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); + + // Saturate negative values + in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); + in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); + in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); + in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); + + // Convert S32 to S16 + const int16x8x2_t in_s16 = + { + { + vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), + vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) + } + }; + + // Convert S16 to U8 + uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1])); + + if(is_bounded_relu) + { + out_u8 = vmaxq_u8(out_u8, min_u8); + out_u8 = vminq_u8(out_u8, max_u8); + } + + return out_u8; +} + +inline Window get_win_vector_sum(const Window &window) +{ + Window win_vector_sum(window); + win_vector_sum.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_vector_sum.set(Window::DimZ, Window::Dimension(0, 0, 0)); + return win_vector_sum; +} + +inline Iterator get_vector_sum_col_it(const Window &window, const ITensor *vector_sum_col) +{ + Iterator vector_sum_col_it(vector_sum_col, get_win_vector_sum(window)); + return vector_sum_col_it; +} + +inline Iterator get_vector_sum_row_it(const Window &window, const ITensor *vector_sum_row) +{ + Window win_vector_sum_row = get_win_vector_sum(window); + win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); + Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row); + return vector_sum_row_it; +} + +inline Iterator get_bias_it(const Window &window, const ITensor *bias) +{ + Window win_bias(window); + win_bias.set(Window::DimY, Window::Dimension(0, 1, 1)); + win_bias.set(Window::DimZ, Window::Dimension(0, 1, 1)); + Iterator bias_it(bias, win_bias); + return bias_it; +} + +inline int32x4x4_t add_s32(int32x4x4_t a, int32x4_t b) +{ + return + { + { + vaddq_s32(a.val[0], b), + vaddq_s32(a.val[1], b), + vaddq_s32(a.val[2], b), + vaddq_s32(a.val[3], b) + } + }; +} + +inline int32x4x4_t add_s32(int32x4x4_t a, int32x4x4_t b) +{ + return + { + { + vaddq_s32(a.val[0], b.val[0]), + vaddq_s32(a.val[1], b.val[1]), + vaddq_s32(a.val[2], b.val[2]), + vaddq_s32(a.val[3], b.val[3]) + } + }; +} + +inline int32x4x4_t mul_s32(int32x4x4_t &a, int32_t mul_scalar) +{ + return + { + { + vmulq_n_s32(a.val[0], mul_scalar), + vmulq_n_s32(a.val[1], mul_scalar), + vmulq_n_s32(a.val[2], mul_scalar), + vmulq_n_s32(a.val[3], mul_scalar) + } + }; +} + +template +inline void run_offset_contribution_output_stage_window(const int32_t *vector_sum_col_ptr, const int32_t *vector_sum_row_ptr, const int32_t *bias_ptr, Iterator mm_result_it, Iterator out_it, + const int32x4_t result_offset_s32, const int32x4_t result_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8, + int32_t a_offset, int32_t b_offset, int32_t k_offset, + GEMMLowpOutputStageInfo output_stage, int window_step_x, int window_start_x, int window_end_x) +{ + int32x4x4_t offset_term_s32 = { 0, 0, 0, 0 }; + if(!is_fixed_point) + { + // Combine quantization offset with other offsets. + offset_term_s32 = add_s32(offset_term_s32, result_offset_s32); + } + if(has_a_offset && has_b_offset) + { + offset_term_s32 = add_s32(offset_term_s32, get_k_offset(k_offset)); + } + if(has_b_offset) + { + offset_term_s32 = add_s32(offset_term_s32, get_b_offset(vector_sum_row_ptr, b_offset)); + } + + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + int32x4x4_t in_s32 = load_results_input(mm_result_it, x); + + if(has_a_offset) + { + in_s32 = add_s32(in_s32, get_a_offset(vector_sum_col_ptr, a_offset, x)); + } + if(has_bias) + { + in_s32 = add_s32(in_s32, load(bias_ptr, x)); + } + if(!is_fixed_point || has_b_offset) + { + in_s32 = add_s32(in_s32, offset_term_s32); + } + if(!is_fixed_point) + { + in_s32 = mul_s32(in_s32, output_stage.gemmlowp_multiplier); + } + + if(is_fixed_point) + { + vst1q_u8(out_it.ptr() + x, finalize_quantization(in_s32, output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, result_offset_s32, min_u8, max_u8)); + } + else + { + vst1q_u8(out_it.ptr() + x, finalize_quantization_floating_point(in_s32, result_shift_s32, min_u8, max_u8)); + } + } + // Compute left-over elements + for(; x < window_end_x; ++x) + { + int32_t in_value = *(reinterpret_cast(mm_result_it.ptr()) + x) + wrapper::vgetlane(offset_term_s32.val[0], 0); + + if(has_a_offset) + { + in_value += (*(vector_sum_col_ptr + x) * a_offset); + } + if(has_bias) + { + in_value += *(bias_ptr + x); + } + + if(is_fixed_point) + { + // Finalize and store the result + *(out_it.ptr() + x) = finalize_quantization(in_value, output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, + output_stage.gemmlowp_offset, static_cast(output_stage.gemmlowp_min_bound), static_cast(output_stage.gemmlowp_max_bound)); + } + else + { + // Finalize quantization + in_value = (in_value * output_stage.gemmlowp_multiplier) >> output_stage.gemmlowp_shift; + + // Bound and store the result + if(is_bounded_relu) + { + in_value = static_cast(std::max(output_stage.gemmlowp_min_bound, std::min(output_stage.gemmlowp_max_bound, in_value))); + } + *(out_it.ptr() + x) = static_cast(std::max(0, std::min(255, in_value))); + } + } +} + +template +void run_offset_contribution_output_stage(const Window &window, + const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, + int32_t a_offset, int32_t b_offset, int32_t k_offset, bool slide_vector_sum_col, + GEMMLowpOutputStageInfo output_stage) +{ + const int height_input = is_gemm3d ? mm_result->info()->dimension(1) : 0; + const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; + + const int32x4_t result_offset_s32 = vdupq_n_s32(output_stage.gemmlowp_offset); + const int32x4_t result_shift_s32 = vdupq_n_s32(is_fixed_point ? output_stage.gemmlowp_shift : -output_stage.gemmlowp_shift); + const uint8x16_t min_u8 = vdupq_n_u8(static_cast(output_stage.gemmlowp_min_bound)); + const uint8x16_t max_u8 = vdupq_n_u8(static_cast(output_stage.gemmlowp_max_bound)); + + const int window_step_x = 16; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Window win(window); + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Window collapsed_window = win.collapse_if_possible(win, Window::DimZ); + + Iterator mm_result_it(mm_result, win); + Iterator out_it(output, win); + + if((a_offset != 0) && (b_offset != 0)) + { + ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); + ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); + + Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); + Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); + + const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); + + // Offset in case vector_sum_col is batched + const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0; + + if(bias != nullptr) + { + Iterator bias_it = get_bias_it(collapsed_window, bias); + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); + const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + + id.y() + (id.z() % depth_input) * height_input; + run_offset_contribution_output_stage_window(vector_sum_col_ptr, vector_sum_row_ptr, reinterpret_cast(bias_it.ptr()), mm_result_it, + out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_col_it, vector_sum_row_it, bias_it, mm_result_it, out_it); + } + else + { + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); + const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + + id.y() + (id.z() % depth_input) * height_input; + run_offset_contribution_output_stage_window(vector_sum_col_ptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_col_it, vector_sum_row_it, mm_result_it, out_it); + } + } + else if((a_offset == 0) && (b_offset != 0)) + { + ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); + + Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); + + const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); + + if(bias != nullptr) + { + Iterator bias_it = get_bias_it(collapsed_window, bias); + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + + id.y() + (id.z() % depth_input) * height_input; + run_offset_contribution_output_stage_window(nullptr, vector_sum_row_ptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_row_it, bias_it, mm_result_it, out_it); + } + else + { + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + + id.y() + (id.z() % depth_input) * height_input; + run_offset_contribution_output_stage_window(nullptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_row_it, mm_result_it, out_it); + } + } + else if((a_offset != 0) && (b_offset == 0)) + { + ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); + + Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); + + // Offset in case vector_sum_col is batched + const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0; + + if(bias != nullptr) + { + Iterator bias_it = get_bias_it(collapsed_window, bias); + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); + run_offset_contribution_output_stage_window(vector_sum_col_ptr, nullptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_col_it, bias_it, mm_result_it, out_it); + } + else + { + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + const int batch_id = id.z() / depth_input; + const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); + run_offset_contribution_output_stage_window(vector_sum_col_ptr, nullptr, nullptr, mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + vector_sum_col_it, mm_result_it, out_it); + } + } + else + { + if(bias != nullptr) + { + Iterator bias_it = get_bias_it(collapsed_window, bias); + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + run_offset_contribution_output_stage_window(nullptr, nullptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + bias_it, mm_result_it, out_it); + } + else + { + execute_window_loop(collapsed_window, [&](const Coordinates & id) + { + run_offset_contribution_output_stage_window(nullptr, nullptr, nullptr, mm_result_it, out_it, + result_offset_s32, result_shift_s32, min_u8, max_u8, a_offset, b_offset, k_offset, + output_stage, window_step_x, window_start_x, window_end_x); + }, + mm_result_it, out_it); + } + return; + } +} + +Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, + int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > 255); + ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < 0 || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound); + ARM_COMPUTE_RETURN_ERROR_ON(output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN && output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT); + + if(bias != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0)); + } + + // If a_offset == 0, vector_sum_col can be a nullptr + if(a_offset != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0)); + } + + // If b_offset == 0, vector_sum_row can be a nullptr + if(b_offset != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); + + // Check if input is a 3D reinterpretation + const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x(); + + // Validate input + ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2))); + ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1)); + + TensorShape output_shape = output->tensor_shape(); + if(output_shape.num_dimensions() > 1) + { + const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2; + + TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape(); + vector_sum_row_shape.collapse_from(1); + output_shape.collapse_from(output_batch_idx); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx], + "mm_result tensor must have the same number of batches of output tensor"); + + if(a_offset != 0) + { + TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape(); + vector_sum_col_shape.collapse_from(1); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1], + "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1"); + } + } + } + + if(output->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mm_result, output); + } + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *output) +{ + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output, mm_result->clone()->set_data_type(DataType::QASYMM8)); + + // Configure kernel window + Window win = calculate_max_window(*mm_result, Steps()); + + // Note: This kernel performs 16 elements per iteration. + // However, since we use a left-over for loop, we cannot have any read or write out of memory + // For this reason num_elems_processed_per_iteration is 1 and so update_window_and_padding() can be skipped + Coordinates coord; + coord.set_num_dimensions(output->num_dimensions()); + output->set_valid_region(ValidRegion(coord, output->tensor_shape())); + + return std::make_pair(Status{}, win); +} + +NEGEMMLowpOffsetContributionOutputStageKernel::NEGEMMLowpOffsetContributionOutputStageFunction +get_configured_function(const ITensor *mm_result, const ITensor *vector_sum_row, GEMMLowpOutputStageInfo output_stage) +{ + static std::map map_function = + { + { 0, &run_offset_contribution_output_stage }, + { 1, &run_offset_contribution_output_stage }, + { 2, &run_offset_contribution_output_stage }, + { 3, &run_offset_contribution_output_stage }, + { 4, &run_offset_contribution_output_stage }, + { 5, &run_offset_contribution_output_stage }, + { 6, &run_offset_contribution_output_stage }, + { 7, &run_offset_contribution_output_stage } + }; + + // Check if input is a 3D reinterpretation + const bool reinterpret_as_3d = vector_sum_row != nullptr + && mm_result->info()->num_dimensions() > 1 + && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x(); + + // Check if we need to clamp the result using min and max + const bool is_bounded_relu = ((output_stage.gemmlowp_min_bound != output_stage.gemmlowp_max_bound) + && !(output_stage.gemmlowp_min_bound == 0 && output_stage.gemmlowp_max_bound == 255)); + + const bool is_fixed_point = output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN; + + // key acts as a bitset, setting the first bit on reinterpret_as_3d, + // the second on is_bounded_relu, and the third on is_fixed_point. + uint8_t key = (reinterpret_as_3d ? 1UL : 0UL) | ((is_bounded_relu ? 1UL : 0UL) << 1) | ((is_fixed_point ? 1UL : 0UL) << 2); + return map_function.find(key)->second; +} +} // namespace + +NEGEMMLowpOffsetContributionOutputStageKernel::NEGEMMLowpOffsetContributionOutputStageKernel() + : _function(nullptr), _vector_sum_col(nullptr), _vector_sum_row(nullptr), _bias(nullptr), _mm_result(nullptr), _output(nullptr), _a_offset(0), _b_offset(0), _k_offset(0), _slide_vector_sum_col(true), + _output_stage(GEMMLowpOutputStageInfo()) + +{ +} + +void NEGEMMLowpOffsetContributionOutputStageKernel::configure(const ITensor *mm_result, const ITensor *vector_sum_col, + const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t k, + int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) +{ + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output); + + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result->info(), + vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, // NOLINT + vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, // NOLINT + bias != nullptr ? bias->info() : nullptr, // NOLINT + output->info(), a_offset, b_offset, output_stage)); // NOLINT + + _vector_sum_col = vector_sum_col; + _vector_sum_row = vector_sum_row; + _bias = bias; + _mm_result = mm_result; + _output = output; + _a_offset = a_offset; + _b_offset = b_offset; + _k_offset = a_offset * b_offset * k; + _output_stage = output_stage; + + // If a_offset == 0, vector_sum_col can be a nullptr + if(a_offset != 0) + { + // Check if vector_sum_col_shape should be slidden or not + // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + _slide_vector_sum_col = vector_sum_col->info()->tensor_shape().num_dimensions() > 1; + } + + // Configure kernel window + auto win_config = validate_and_configure_window(mm_result->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); + + _function = get_configured_function(mm_result, vector_sum_row, output_stage); +} + +Status NEGEMMLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, + const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, + int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(mm_result->clone().get(), output->clone().get()).first); + return Status{}; +} + +void NEGEMMLowpOffsetContributionOutputStageKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + _function(window, _mm_result, _vector_sum_col, _vector_sum_row, _bias, _output, _a_offset, _b_offset, _k_offset, _slide_vector_sum_col, _output_stage); +} + +} // namespace arm_compute \ No newline at end of file diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp index f0ac695b20..d3cfc7a8fa 100644 --- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp @@ -86,37 +86,6 @@ std::pair validate_and_configure_window(ITensorInfo *input, ITen namespace arm_compute { class Coordinates; - -/* Function used by the left-over for loop to perform the quantization */ -template -inline uint8_t finalize_quantization(int32x4_t in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int32x4_t result_offset_after_shift_s32, uint8_t min_u8, uint8_t max_u8) -{ - const static int32x4_t zero_s32 = vdupq_n_s32(0); - const static int32x4_t sat_value_s32 = vdupq_n_s32(255); - - // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar - in_s32 = vqrdmulhq_n_s32(in_s32, result_fixedpoint_multiplier); - - // Round to the nearest division by a power-of-two using result_shift_s32 - in_s32 = rounding_divide_by_pow2(in_s32, result_shift); - - // Add the offset terms - in_s32 = vaddq_s32(in_s32, result_offset_after_shift_s32); - - // Saturate negative values - in_s32 = vmaxq_s32(in_s32, zero_s32); - in_s32 = vminq_s32(in_s32, sat_value_s32); - - auto out_u8 = static_cast(vgetq_lane_s32(in_s32, 0)); - - if(is_bounded_relu) - { - out_u8 = std::max(out_u8, min_u8); - out_u8 = std::min(out_u8, max_u8); - } - - return out_u8; -} } // namespace arm_compute template @@ -188,10 +157,8 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window // Add bias in_value += bias_value; - // Finalize and store the result - *(out.ptr() + x) = finalize_quantization(vdupq_n_s32(in_value), _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, static_cast(_min), - static_cast(_max)); + *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast(_min), static_cast(_max)); } }, in, out, bias); @@ -220,10 +187,10 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window // Compute left-over elements for(; x < window_end_x; ++x) { - const int32x4_t in_s32 = vld1q_dup_s32(reinterpret_cast(in.ptr()) + x); + const int32_t in_value = *(reinterpret_cast(in.ptr()) + x); // Finalize and store the result - *(out.ptr() + x) = finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, static_cast(_min), static_cast(_max)); + *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast(_min), static_cast(_max)); } }, in, out); diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index be7cc2d0e1..b6c37349c1 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -90,16 +90,17 @@ void NEConvolutionLayerReshapeWeights::run() } NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr &memory_manager) - : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _activationlayer_function(), - _add_bias_kernel(), _reshape_layer(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false), - _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false) + : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _col2im_kernel(), _activationlayer_function(), _add_bias_kernel(), + _reshape_layer(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false), _skip_im2col(false), + _skip_col2im(false), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false) { } -void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, int gemm_3d_depth) +void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act_info, int gemm_3d_depth) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights); - ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), output->info(), gemm_3d_depth, _skip_im2col)); + ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output == nullptr ? nullptr : output->info(), act_info, gemm_3d_depth, + _skip_im2col)); const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */); @@ -114,7 +115,40 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); - _mm_gemmlowp.configure(input, weights, nullptr, output, gemm_info); + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input_quantization_info : output->info()->quantization_info(); + + float multiplier = input_quantization_info.scale * weights->info()->quantization_info().scale / output_quant_info.scale; + int output_multiplier, output_shift; + quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + + // Merge activation with output stage + int min_activation = 0; + int max_activation = 0; + + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + if(_is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0) + { + const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); + const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); + + min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int; + max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int; + + _is_activationlayer_enabled = false; + } + + GEMMLowpOutputStageInfo output_info; + output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + output_info.gemmlowp_offset = output_quant_info.offset; + output_info.gemmlowp_multiplier = output_multiplier; + output_info.gemmlowp_shift = output_shift; + output_info.gemmlowp_min_bound = min_activation; + output_info.gemmlowp_max_bound = max_activation; + + _mm_gemmlowp.configure(input, weights, biases, output, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info)); // Revert back QuantizatioInfo as input and weights could be used in other convolution layers input->info()->set_quantization_info(input_quantization_info); @@ -127,9 +161,11 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w } } -Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth, bool skip_im2col) +Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act_info, + int gemm_3d_depth, bool skip_im2col) { - const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); + const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); + const bool is_activation_enabled = act_info.enabled(); const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */); @@ -145,8 +181,39 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens input_qa->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); + const QuantizationInfo output_quant_info = (output->total_size() == 0) ? input_quantization_info : output->quantization_info(); + + float multiplier = input_quantization_info.scale * weights->quantization_info().scale / output_quant_info.scale; + int output_multiplier, output_shift; + quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + + // Merge activation with output stage + int min_activation = 0; + int max_activation = 0; + + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + if(is_activation_enabled && supported_acts.count(act_info.activation()) != 0) + { + const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); + const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); + + min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int; + max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int; + } + + GEMMLowpOutputStageInfo output_info; + output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + output_info.gemmlowp_offset = output_quant_info.offset; + output_info.gemmlowp_multiplier = output_multiplier; + output_info.gemmlowp_shift = output_shift; + output_info.gemmlowp_min_bound = min_activation; + output_info.gemmlowp_max_bound = max_activation; + // Perform validation step on GEMMLowp - return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), nullptr, output, gemm_info); + return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info)); } else { @@ -155,19 +222,18 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens } } -Status NEGEMMConvolutionLayer::validate_gemm3d(DataType data_type, int gemm_3d_depth, bool skip_im2col) +Status NEGEMMConvolutionLayer::validate_gemm3d(const ITensorInfo *input_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col) { - const bool is_quantized = is_data_type_quantized_asymmetric(data_type); - const DataType output_gemm_data_type = is_quantized ? DataType::S32 : data_type; - const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth; - const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U; + const DataType data_type = input_info->data_type(); + const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth; + const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U; // Set dummy tensor shapes for the validation - const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type); + const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info()); const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type); - const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, output_gemm_data_type); + const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info()); - return validate_mm(&dummy_input_info, &dummy_weights_info, &dummy_output_info, gemm_3d_depth, skip_im2col); + return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, gemm_3d_depth, skip_im2col); } void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, @@ -202,9 +268,8 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig _append_bias = (biases != nullptr) && (!_is_quantized); _is_activationlayer_enabled = act_info.enabled(); - const ITensor *gemm_input_to_use = input; - ITensor *gemm_output_to_use = output; - ITensor *gemm_output_staged_to_use = output; + const ITensor *gemm_input_to_use = input; + ITensor *gemm_output_to_use = output; // Get convolved dimensions unsigned int conv_w = 0; @@ -219,7 +284,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig // Check if GEMM3D is supported if(data_layout == DataLayout::NHWC) { - _skip_col2im = bool(validate_gemm3d(input->info()->data_type(), conv_h, true)); + _skip_col2im = bool(validate_gemm3d(input->info(), act_info, conv_h, true)); // If not supported, we need to perform im2col and col2im (or reshape layer) if(!_skip_col2im) { @@ -262,26 +327,17 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig } // Create temporary GEMM output tensor in case we cannot skip col2im - if(!_skip_col2im || _is_quantized) + if(!_skip_col2im) { - // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. - const DataType gemm_data_type = _is_quantized ? DataType::S32 : data_type; - TensorShape shape_gemm; + TensorShape shape_gemm; - if(_is_quantized && _skip_col2im) - { - shape_gemm = output->info()->tensor_shape(); - } - else - { - // Calculate GEMM output shape - shape_gemm = _im2col_output.info()->tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, conv_w * conv_h); - } + // Calculate GEMM output shape + shape_gemm = _im2col_output.info()->tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, conv_w * conv_h); // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type); + TensorInfo info_gemm(shape_gemm, 1, data_type); info_gemm.set_quantization_info(output->info()->quantization_info()).set_data_layout(input->info()->data_layout()); _gemm_output.allocator()->init(info_gemm); _memory_group.manage(&_gemm_output); @@ -293,62 +349,24 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig // Configure GEMM // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0; - configure_mm(gemm_input_to_use, &_weights_reshaped, gemm_output_to_use, gemm_3d_depth); + configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, gemm_3d_depth); if(!_skip_im2col) { _im2col_output.allocator()->allocate(); } - // Configure output stage for quantized case - if(_is_quantized) - { - const QuantizationInfo input_quant_info = input->info()->quantization_info(); - const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input_quant_info : output->info()->quantization_info(); - - float multiplier = input_quant_info.scale * weights->info()->quantization_info().scale / output_quant_info.scale; - int output_multiplier, output_shift; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - - if(!_skip_col2im) - { - _memory_group.manage(&_tmp_output); - gemm_output_staged_to_use = &_tmp_output; - } - - // Merge activation with output stage - int min_activation = 0; - int max_activation = 0; - - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - if(_is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0) - { - const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); - const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); - - min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int; - max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int; - - _is_activationlayer_enabled = false; - } - - _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset, min_activation, max_activation); - } - if(!_skip_col2im) { if(_data_layout == DataLayout::NCHW) { // Configure col2im - _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, Size2D(conv_w, conv_h)); + _col2im_kernel.configure(gemm_output_to_use, output, Size2D(conv_w, conv_h)); } else { // Configure reshape layer - _reshape_layer.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output); + _reshape_layer.configure(gemm_output_to_use, output); } } @@ -395,10 +413,9 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI const unsigned int kernel_height = weights->dimension(idx_height); TensorInfo im2col_reshaped_info, info_gemm, tmp_info, weights_reshaped_info; - const ITensorInfo *gemm_input_to_use = input; - const ITensorInfo *gemm_output_to_use = output; - const ITensorInfo *gemm_output_staged_to_use = output; - const ITensorInfo *weights_to_use = weights; + const ITensorInfo *gemm_input_to_use = input; + const ITensorInfo *gemm_output_to_use = output; + const ITensorInfo *weights_to_use = weights; const bool is_quantized = is_data_type_quantized_asymmetric(data_type); const bool append_bias = (biases != nullptr) && (!is_quantized); @@ -420,7 +437,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI bool skip_col2im = false; if(data_layout == DataLayout::NHWC) { - skip_col2im = bool(validate_gemm3d(input->data_type(), conv_h, true)); + skip_col2im = bool(validate_gemm3d(input, act_info, conv_h, true)); // If not supported, we need to perform im2col and col2im (or reshape layer) if(!skip_col2im) { @@ -431,7 +448,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI if(skip_col2im) { // If not supported, we need to perform im2col and col2im (or reshape layer) - if(!bool(validate_gemm3d(input->data_type(), conv_h, skip_im2col))) + if(!bool(validate_gemm3d(input, act_info, conv_h, skip_im2col))) { skip_im2col = false; skip_col2im = false; @@ -495,68 +512,25 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI } // Create temporary GEMM output tensor in case we cannot skip col2im - const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type; if(!skip_col2im) { TensorShape shape_gemm = gemm_input_to_use->tensor_shape(); shape_gemm.set(0, mat_weights_cols); shape_gemm.set(1, conv_w * conv_h); - info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type); + info_gemm = TensorInfo(shape_gemm, 1, data_type); } else { - info_gemm = TensorInfo(output->tensor_shape(), 1, gemm_data_type); + info_gemm = TensorInfo(output->tensor_shape(), 1, data_type); } info_gemm.set_quantization_info(output->quantization_info()).set_data_layout(input->data_layout()); gemm_output_to_use = &info_gemm; - - ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, gemm_output_to_use, skip_col2im ? conv_h : 0, skip_im2col)); - - if(is_quantized) - { - const QuantizationInfo input_quant_info = input->quantization_info(); - const QuantizationInfo output_quant_info = (output->total_size() == 0) ? input_quant_info : output->quantization_info(); - const float multiplier = input_quant_info.scale * weights_to_use->quantization_info().scale / output_quant_info.scale; - int output_multiplier, output_shift; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - - if(!skip_col2im) - { - tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8); - tmp_info.set_quantization_info(output->quantization_info()).set_data_layout(data_layout); - gemm_output_staged_to_use = &tmp_info; - } - - // Merge activation with output stage - int min_activation = 0; - int max_activation = 0; - - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - - if(is_activation_enabled && supported_acts.count(act_info.activation()) != 0) - { - const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP); - const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP); - - min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int; - max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int; - - is_activation_enabled = false; - } - - // Validate output stage for quantized case - NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use, min_activation, max_activation); - } + ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, skip_col2im ? conv_h : 0, skip_im2col)); // Validate Col2Im/ReshapeLayer if(!skip_col2im && (data_layout == DataLayout::NCHW)) { - ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, - output, - Size2D(conv_w, conv_h))); + ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h))); } //Validate Activation Layer @@ -586,9 +560,6 @@ void NEGEMMConvolutionLayer::run() { // Run gemmlowp _mm_gemmlowp.run(); - - // Run output stage - _gemmlowp_output_stage.run(); } else { diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index 5286f113a5..85e49fd265 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -42,8 +42,8 @@ using namespace arm_compute::misc::shape_calculator; NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr memory_manager) : _memory_group(memory_manager), _asm_glue(memory_manager), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), - _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _original_b(nullptr), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false), - _dot_product_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false) + _offset_contribution_kernel(), _offset_contribution_output_stage_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _mm_result_s32(), _original_b(nullptr), _a_offset(0), _b_offset(0), + _run_vector_matrix_multiplication(false), _dot_product_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false), _fuse_output_stage(false) { } @@ -53,6 +53,9 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ARM_COMPUTE_UNUSED(c); ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info)); + const ITensor *matrix_a = a; + const ITensor *matrix_b = b; + // Clear state _mtx_a_reshape_kernel = nullptr; _mtx_b_reshape_kernel = nullptr; @@ -65,6 +68,18 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, _is_prepared = false; _original_b = b; + // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage + if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) + { + _fuse_output_stage = true; + + _memory_group.manage(&_mm_result_s32); + + TensorInfo info_mm_result_s32(output->info()->tensor_shape(), 1, DataType::S32); + + _mm_result_s32.allocator()->init(info_mm_result_s32); + } + #ifdef __aarch64__ switch(a->info()->data_type()) { @@ -72,7 +87,7 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, case DataType::U8: case DataType::S8: { - _asm_glue.configure(a, b, output, 1.f, 0.f, _reshape_b_only_on_first_run); + _asm_glue.configure(a, b, _fuse_output_stage ? &_mm_result_s32 : output, 1.f, 0.f, _reshape_b_only_on_first_run); _dot_product_path = _asm_glue.is_configured(); break; } @@ -83,51 +98,35 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, } } #endif /* __aarch64__ */ - if(!_dot_product_path) + if(!(_dot_product_path || _run_vector_matrix_multiplication)) { - if(_run_vector_matrix_multiplication) + matrix_a = &_tmp_a; + matrix_b = &_tmp_b; + + // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] + TensorInfo a_info(compute_interleaved_shape(*a->info()), 1, a->info()->data_type()); + // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] + TensorInfo b_info(compute_transpose1xW_shape(*b->info()), 1, b->info()->data_type()); + _tmp_a.allocator()->init(a_info); + _tmp_b.allocator()->init(b_info); + _memory_group.manage(&_tmp_a); + if(!_reshape_b_only_on_first_run) { - // Configure matrix multiply kernel - { - auto k = arm_compute::support::cpp14::make_unique(); - k->configure(a, b, output); - _mm_kernel = std::move(k); - } + _memory_group.manage(&_tmp_b); } - else - { - // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] - TensorInfo info_a = a->info()->clone()->set_tensor_shape(compute_interleaved_shape(*a->info())).set_is_resizable(true); - // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] - TensorInfo info_b = b->info()->clone()->set_tensor_shape(compute_transpose1xW_shape(*b->info())).set_is_resizable(true); - _tmp_a.allocator()->init(info_a); - _tmp_b.allocator()->init(info_b); - _memory_group.manage(&_tmp_a); - if(!_reshape_b_only_on_first_run) - { - _memory_group.manage(&_tmp_b); - } - // Configure interleave kernel - { - auto k = arm_compute::support::cpp14::make_unique(); - k->configure(a, &_tmp_a); - _mtx_a_reshape_kernel = std::move(k); - } - - // Configure transpose kernel - { - auto k = arm_compute::support::cpp14::make_unique(); - k->configure(b, &_tmp_b); - _mtx_b_reshape_kernel = std::move(k); - } + // Configure interleave kernel + { + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(a, &_tmp_a); + _mtx_a_reshape_kernel = std::move(k); + } - // Configure matrix multiply kernel - { - auto k = arm_compute::support::cpp14::make_unique(); - k->configure(&_tmp_a, &_tmp_b, output); - _mm_kernel = std::move(k); - } + // Configure transpose kernel + { + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(b, &_tmp_b); + _mtx_b_reshape_kernel = std::move(k); } } @@ -158,8 +157,33 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false); } - // Configure offset contribution kernel - _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset); + if(_fuse_output_stage) + { + // Configure matrix multiply kernel + if(!_dot_product_path) + { + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(matrix_a, matrix_b, &_mm_result_s32); + _mm_kernel = std::move(k); + } + + _offset_contribution_output_stage_kernel.configure(&_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, output, a->info()->dimension(0), + _a_offset, _b_offset, gemm_info.gemmlowp_output_stage()); + + _mm_result_s32.allocator()->allocate(); + } + else + { + // Configure matrix multiply kernel + if(!_dot_product_path) + { + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(matrix_a, matrix_b, output); + _mm_kernel = std::move(k); + } + // Configure offset contribution kernel + _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset); + } // Allocate tensors if(!_dot_product_path && !_run_vector_matrix_multiplication) @@ -185,43 +209,53 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::NONE, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore for output S32"); ARM_COMPUTE_RETURN_ERROR_ON_MSG((a)->dimension(0) != (b)->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); + const ITensorInfo *matrix_a_info = a; + const ITensorInfo *matrix_b_info = b; + + TensorInfo tmp_a_info{}; + TensorInfo tmp_b_info{}; + TensorInfo mm_result_s32_info{}; + int32_t a_offset = a->quantization_info().offset; int32_t b_offset = b->quantization_info().offset; const bool reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); + bool fuse_output_stage = gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE; + if(fuse_output_stage) + { + auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(output->tensor_shape()).set_data_type(DataType::S32)); + } + // Check if we need to run the optimized assembly kernel - const bool run_optimised = bool(NEGEMMAssemblyDispatch::validate(a, b, output, 1.f, 0.f, reshape_b_only_on_first_run)); + const bool run_optimised = bool(NEGEMMAssemblyDispatch::validate(a, b, fuse_output_stage ? &mm_result_s32_info : output, 1.f, 0.f, reshape_b_only_on_first_run)); if(run_optimised) { - if(output->total_size() != 0) + ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0)); + if(gemm_info.depth_output_gemm3d() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0)); - if(gemm_info.depth_output_gemm3d() != 0) + if(gemm_info.reinterpret_input_as_3d()) { - if(gemm_info.reinterpret_input_as_3d()) - { - ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); - ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2)); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2)); - } + ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2)); } else { - ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2)); } } + else + { + ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); + } } else { @@ -231,6 +265,9 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso const bool run_vector_matrix_multiplication = a->dimension(1) < 2; if(!run_vector_matrix_multiplication) { + matrix_a_info = &tmp_a_info; + matrix_b_info = &tmp_b_info; + // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] TensorShape shape_tmp_a = a->tensor_shape(); shape_tmp_a.set(0, a->dimension(0) * 4); @@ -241,16 +278,12 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso shape_tmp_b.set(0, b->dimension(1) * 16); shape_tmp_b.set(1, std::ceil(b->dimension(0) / 16.f)); - TensorInfo info_a = a->clone()->set_tensor_shape(shape_tmp_a).set_is_resizable(true); - TensorInfo info_b = b->clone()->set_tensor_shape(shape_tmp_b).set_is_resizable(true); + // Validate interleave kernel + auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(shape_tmp_a)); + auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(shape_tmp_b)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &info_a)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &info_b)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output)); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(a, b, output)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &tmp_a_info)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &tmp_b_info)); } } @@ -274,12 +307,32 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, a->dimension(0), false)); } - // Validate offset contribution kernel - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionKernel::validate(output, - a_offset == 0 ? nullptr : &info_vector_sum_col, - b_offset == 0 ? nullptr : &info_vector_sum_row, - a_offset, b_offset)); + if(fuse_output_stage) + { + if(!run_optimised) + { + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info)); + } + // Validate offset contribution kernel + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info, + a_offset == 0 ? nullptr : &info_vector_sum_col, + b_offset == 0 ? nullptr : &info_vector_sum_row, + c, output, a_offset, b_offset, + gemm_info.gemmlowp_output_stage())); + } + else + { + if(!run_optimised) + { + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output)); + } + // Validate offset contribution kernel + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionKernel::validate(output, + a_offset == 0 ? nullptr : &info_vector_sum_col, + b_offset == 0 ? nullptr : &info_vector_sum_row, + a_offset, b_offset)); + } return Status{}; } @@ -321,8 +374,16 @@ void NEGEMMLowpMatrixMultiplyCore::run() NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); } - // Run offset contribution kernel - NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY); + if(_fuse_output_stage) + { + // Run offset contribution kernel + NEScheduler::get().schedule(&_offset_contribution_output_stage_kernel, Window::DimY); + } + else + { + // Run offset contribution kernel + NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY); + } _memory_group.release(); } -- cgit v1.2.1