From 6ff3b19ee6120edf015fad8caab2991faa3070af Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Mon, 4 Sep 2017 18:44:23 +0100 Subject: COMPMID-344 Updated doxygen Change-Id: I32f7b84daa560e460b77216add529c8fa8b327ae --- .../NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp | 1168 ++++++++++++++++++++ 1 file changed, 1168 insertions(+) create mode 100644 src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp (limited to 'src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp') diff --git a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp new file mode 100644 index 0000000000..dcfbb13081 --- /dev/null +++ b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp @@ -0,0 +1,1168 @@ +/* + * Copyright (c) 2017 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/NEGEMMMatrixMultiplyKernel.h" + +#include "arm_compute/core/AccessWindowTranspose.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/NEFixedPoint.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 + +using namespace arm_compute; + +namespace arm_compute +{ +class Coordinates; +} // namespace arm_compute + +namespace +{ +template +void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) +{ + const auto width_matrix_b = static_cast(output->info()->dimension(0)); + const auto in_b_stride = static_cast(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); + const auto num_elems_vec_a = static_cast(input0->info()->dimension(0)); + + // The implementation computes 16 elements per iteration + const int window_start_x = 16 * window.thread_id(); + const int window_step_x = 16 * window.num_threads(); + // Make sure (window_end_x - window_start_x) is a multiple of window_step_x + const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; + + Window win_out(window); + win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Window win_b; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(input1->info()->num_dimensions() >= 3) + { + win_b = window; + } + win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Iterator ina(input0, win_a); + Iterator inb(input1, win_b); + Iterator out(output, win_out); + + execute_window_loop(win_out, [&](const Coordinates & id) + { + if(id.x() > width_matrix_b) + { + return; + } + + float32x4_t acc0 = vdupq_n_f32(0.f); + float32x4_t acc1 = vdupq_n_f32(0.f); + float32x4_t acc2 = vdupq_n_f32(0.f); + float32x4_t acc3 = vdupq_n_f32(0.f); + + auto vec_a = reinterpret_cast(ina.ptr()); + auto matrix_b = reinterpret_cast(inb.ptr()); + +#if __arm__ + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b + in_b_stride))); +#endif + + auto vec_a_end_addr = vec_a + num_elems_vec_a; + for(; vec_a <= (vec_a_end_addr - 4);) + { + float32x2_t a0l = vld1_f32(vec_a); + + float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + + float32x4_t b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); + float32x4_t b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); + float32x4_t b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); + float32x4_t b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); + +#if __arm__ + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 1 * in_b_stride))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 2 * in_b_stride))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 3 * in_b_stride))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 4 * in_b_stride))); +#endif + + acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); + acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); + acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); + acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); + + acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); + acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); + acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); + acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); + + vec_a += 2; + matrix_b += 2 * in_b_stride; + + a0l = vld1_f32(vec_a); + + b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + + b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); + b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); + b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); + b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); + + acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); + acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); + acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); + acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); + + acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); + acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); + acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); + acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); + + vec_a += 2; + matrix_b += 2 * in_b_stride; + } + + for(; vec_a < vec_a_end_addr;) + { + const float a0 = *vec_a; + + const float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + const float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + const float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + const float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + + acc0 = vmlaq_n_f32(acc0, b00, a0); + acc1 = vmlaq_n_f32(acc1, b01, a0); + acc2 = vmlaq_n_f32(acc2, b02, a0); + acc3 = vmlaq_n_f32(acc3, b03, a0); + + vec_a += 1; + matrix_b += in_b_stride; + } + + // Multiply by the weight of matrix product (alpha) + if(multiply_alpha) + { + const float32x4_t alpha_f32 = vdupq_n_f32(alpha); + acc0 = vmulq_f32(acc0, alpha_f32); + acc1 = vmulq_f32(acc1, alpha_f32); + acc2 = vmulq_f32(acc2, alpha_f32); + acc3 = vmulq_f32(acc3, alpha_f32); + } + + const auto vec_out = reinterpret_cast(out.ptr()); + + vst1q_f32(vec_out + 0, acc0); + vst1q_f32(vec_out + 4, acc1); + vst1q_f32(vec_out + 8, acc2); + vst1q_f32(vec_out + 12, acc3); + }, + ina, inb, out); +} + +template +void vector_matrix_multiply_qs8(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) +{ + const auto width_matrix_b = static_cast(output->info()->dimension(0)); + const auto in_b_stride = static_cast(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); + const auto num_elems_vec_a = static_cast(input0->info()->dimension(0)); + const int fixed_point_position = input0->info()->fixed_point_position(); + + // The implementation computes 32 elements per iteration + const int window_start_x = 32 * window.thread_id(); + const int window_step_x = 32 * window.num_threads(); + // Make sure (window_end_x - window_start_x) is a multiple of window_step_x + const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; + + Window win_out(window); + win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Window win_b; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(input1->info()->num_dimensions() >= 3) + { + win_b = window; + } + win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Iterator ina(input0, win_a); + Iterator inb(input1, win_b); + Iterator out(output, win_out); + + execute_window_loop(win_out, [&](const Coordinates & id) + { + if(id.x() > width_matrix_b) + { + return; + } + + // Reset accumulators + qint16x8_t acc00_qs16 = vdupq_n_qs16(0); + qint16x8_t acc01_qs16 = vdupq_n_qs16(0); + qint16x8_t acc02_qs16 = vdupq_n_qs16(0); + qint16x8_t acc03_qs16 = vdupq_n_qs16(0); + + auto vec_a = reinterpret_cast(ina.ptr()); + auto matrix_b = reinterpret_cast(inb.ptr()); + + auto vec_a_end_addr = vec_a + num_elems_vec_a; + for(; vec_a <= (vec_a_end_addr - 2);) + { + const qint8x8_t a0 = vld1_dup_qs8(vec_a + 0); + const qint8x8_t a1 = vld1_dup_qs8(vec_a + 1); + + const qint8x8_t b00 = vld1_qs8(matrix_b + 0 + 0 * in_b_stride); + const qint8x8_t b01 = vld1_qs8(matrix_b + 8 + 0 * in_b_stride); + const qint8x8_t b02 = vld1_qs8(matrix_b + 16 + 0 * in_b_stride); + const qint8x8_t b03 = vld1_qs8(matrix_b + 24 + 0 * in_b_stride); + const qint8x8_t b10 = vld1_qs8(matrix_b + 0 + 1 * in_b_stride); + const qint8x8_t b11 = vld1_qs8(matrix_b + 8 + 1 * in_b_stride); + const qint8x8_t b12 = vld1_qs8(matrix_b + 16 + 1 * in_b_stride); + const qint8x8_t b13 = vld1_qs8(matrix_b + 24 + 1 * in_b_stride); + + // First accumulation + acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); + acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b02, a0, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b03, a0, fixed_point_position); + + // Second accumulation + acc00_qs16 = vqmlal_qs8(acc00_qs16, b10, a1, fixed_point_position); + acc01_qs16 = vqmlal_qs8(acc01_qs16, b11, a1, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b12, a1, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b13, a1, fixed_point_position); + + vec_a += 2; + matrix_b += 2 * in_b_stride; + } + + for(; vec_a < vec_a_end_addr;) + { + const qint8x8_t a0 = vld1_dup_qs8(vec_a); + + const qint8x8_t b00 = vld1_qs8(matrix_b + 0); + const qint8x8_t b01 = vld1_qs8(matrix_b + 8); + const qint8x8_t b02 = vld1_qs8(matrix_b + 16); + const qint8x8_t b03 = vld1_qs8(matrix_b + 24); + + acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); + acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b02, a0, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b03, a0, fixed_point_position); + + vec_a += 1; + matrix_b += in_b_stride; + } + + // Convert back to qint8x8_t and saturate + qint8x8_t acc00_qs8 = vqmovn_qs16(acc00_qs16); + qint8x8_t acc01_qs8 = vqmovn_qs16(acc01_qs16); + qint8x8_t acc02_qs8 = vqmovn_qs16(acc02_qs16); + qint8x8_t acc03_qs8 = vqmovn_qs16(acc03_qs16); + + // Multiply by the weight of the matrix product (alpha) + if(multiply_alpha) + { + const qint8x8_t alpha_qs8 = vdup_n_qs8(scvt_qs8_f32(alpha, fixed_point_position)); + acc00_qs8 = vqmul_qs8(acc00_qs8, alpha_qs8, fixed_point_position); + acc01_qs8 = vqmul_qs8(acc01_qs8, alpha_qs8, fixed_point_position); + acc02_qs8 = vqmul_qs8(acc02_qs8, alpha_qs8, fixed_point_position); + acc03_qs8 = vqmul_qs8(acc03_qs8, alpha_qs8, fixed_point_position); + } + + const auto mtx_out0 = reinterpret_cast(out.ptr()); + + // Store 8x4 output elements + vst1_qs8(mtx_out0 + 0, acc00_qs8); + vst1_qs8(mtx_out0 + 8, acc01_qs8); + vst1_qs8(mtx_out0 + 16, acc02_qs8); + vst1_qs8(mtx_out0 + 24, acc03_qs8); + }, + ina, inb, out); +} + +template +void matrix_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) +{ + const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); + const size_t out_stride1 = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); + const size_t out_stride2 = out_stride1 * 2; + const size_t out_stride3 = out_stride1 * 3; + const int num_elems_matrix_b_x = input1->info()->dimension(0); + + // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); + + Window win_b; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(input1->info()->num_dimensions() >= 3) + { + win_b = window; + } + // Set step_x and step_y for matrix B. Scale by a factor of 4 the X range as the input transposed matrix A has 4 times less the cols of the output matrix + // The step along the x direction is 2 times the in_b_stride because for each iteration we compute 2 blocks of size 4x4 + win_b.set(Window::DimX, Window::Dimension(window.x().start() / 4, window.x().end() / 4, 2 * in_b_stride)); + win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Iterator ina(input0, win_a); + Iterator inb(input1, win_b); + Iterator out(output, window); + + // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW + // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration + // All the values needed for computing a single 4x4 block will be read from consecutive memory positions + execute_window_loop(window, [&](const Coordinates & id) + { + auto mtx_a0 = reinterpret_cast(ina.ptr()); + auto mtx_b0 = reinterpret_cast(inb.ptr()); + auto mtx_b1 = mtx_b0 + in_b_stride; + + float32x4_t acc00 = vdupq_n_f32(0.f); + float32x4_t acc10 = vdupq_n_f32(0.f); + float32x4_t acc20 = vdupq_n_f32(0.f); + float32x4_t acc30 = vdupq_n_f32(0.f); + + float32x4_t acc01 = vdupq_n_f32(0.f); + float32x4_t acc11 = vdupq_n_f32(0.f); + float32x4_t acc21 = vdupq_n_f32(0.f); + float32x4_t acc31 = vdupq_n_f32(0.f); + +#if __arm__ + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b1))); +#endif + + auto mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; + for(; mtx_b0 <= (mtx_b0_end_addr - 32);) + { + float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); + float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); + float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); + float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); + + float32x4_t b00 = vld1q_f32(mtx_b0); + float32x4_t b10 = vld1q_f32(mtx_b1); + float32x4_t b01 = vld1q_f32(mtx_b0 + 4); + float32x4_t b11 = vld1q_f32(mtx_b1 + 4); + +#if __arm__ + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); +#endif + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + float32x4_t a4 = vld1q_dup_f32(mtx_a0 + 4); + float32x4_t a5 = vld1q_dup_f32(mtx_a0 + 5); + float32x4_t a6 = vld1q_dup_f32(mtx_a0 + 6); + float32x4_t a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); + +#if __arm__ + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); +#endif + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + } + + for(; mtx_b0 < mtx_b0_end_addr;) + { + float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); + float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); + float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); + float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); + float32x4_t b00 = vld1q_f32(mtx_b0); + float32x4_t b10 = vld1q_f32(mtx_b1); + +#if __arm__ + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b1))); +#endif + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + mtx_a0 += 4; + mtx_b0 += 4; + mtx_b1 += 4; + } + + // Multiply by the weight of matrix product (alpha) + if(multiply_alpha) + { + const float32x4_t alpha_f32 = vdupq_n_f32(alpha); + acc00 = vmulq_f32(acc00, alpha_f32); + acc10 = vmulq_f32(acc10, alpha_f32); + acc20 = vmulq_f32(acc20, alpha_f32); + acc30 = vmulq_f32(acc30, alpha_f32); + acc01 = vmulq_f32(acc01, alpha_f32); + acc11 = vmulq_f32(acc11, alpha_f32); + acc21 = vmulq_f32(acc21, alpha_f32); + acc31 = vmulq_f32(acc31, alpha_f32); + } + + const auto mtx_out0 = reinterpret_cast(out.ptr()); + const auto mtx_out1 = mtx_out0 + 4; + + // Store the 4 blocks + vst1q_f32(mtx_out0, acc00); + vst1q_f32(mtx_out1, acc01); + vst1q_f32(mtx_out0 + out_stride1, acc10); + vst1q_f32(mtx_out1 + out_stride1, acc11); + vst1q_f32(mtx_out0 + out_stride2, acc20); + vst1q_f32(mtx_out1 + out_stride2, acc21); + vst1q_f32(mtx_out0 + out_stride3, acc30); + vst1q_f32(mtx_out1 + out_stride3, acc31); + }, + ina, inb, out); +} + +template +void matrix_matrix_multiply_f16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) +{ +#ifdef ARM_COMPUTE_ENABLE_FP16 + const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); + const size_t out_stride = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); + + // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); + + Window win_b; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(input1->info()->num_dimensions() >= 3) + { + win_b = window; + } + // Set step_x and step_y for matrix B. Scale by a factor of 8 the X range as the input transposed matrix A has 8 times less the cols of the output matrix + win_b.set(Window::DimX, Window::Dimension(window.x().start() / 8, window.x().end() / 8, in_b_stride)); + win_b.set(Window::DimY, Window::Dimension(0, 1, 0)); + + Iterator ina(input0, win_a); + Iterator inb(input1, win_b); + Iterator out(output, window); + + // Number of iterations of inner loop. Since 8 is the number of accumulations per loop, num_it = (width_mtx_b / 4) / 8 + const size_t num_it = ((input1->info()->dimension(0)) >> 2) >> 3; + + const float16x8_t alpha_f16 = vdupq_n_f16(alpha); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto *mtx_a0 = reinterpret_cast(ina.ptr()); + const auto *mtx_b0 = reinterpret_cast(inb.ptr()); + auto *mtx_out = reinterpret_cast(out.ptr()); + float16x8x4_t c = + { + { + vdupq_n_f16(0.f), + vdupq_n_f16(0.f), + vdupq_n_f16(0.f), + vdupq_n_f16(0.f) + } + }; + + /* + This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values) + |a00 a01 a02 a03 | a04 a05 a06 a07| + |a10 a11 a12 a13 | a14 a15 a16 a17| + |a20 a21 a22 a23 | a24 a25 a26 a27| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 | a40 a50 a60 a70 | ... + |a30 a31 a32 a33 | a34 a35 a36 a37| | a04 a14 a24 a34 || a05 a15 a25 a35 || a06 a15 a26 a36 || a07 a17 a27 a37 | a44 a54 a64 a74 | ... + |a40 a41 a42 a43 | a44 a45 a46 a47| + |a50 a51 a52 a53 | a54 a55 a56 a57| + |a60 a61 a62 a63 | a64 a65 a66 a67| + |a70 a71 a72 a73 | a74 a75 a76 a77| + + After this operation, the output matrix will have the following shape: [ height * 4, width / 4 ] + + B Matrix has been transposed as shown below + + |b00 b01 b02 b03 b04 b05 b06 b07| + |b10 b11 b12 b13 b14 b15 b16 b17| + |b20 b21 b22 b23 b24 b25 b26 b27| + |b30 b31 b32 b33 b34 b35 b36 b37| + -------------------> + + |b00 b01 b02 b03 b04 b05 b06 b07||b10 b11 b12 b13 b14 b15 b16 b17||b20 b21 b22 b23 b24 b25 b26 b27||b30 b31 b32 b33 b34 b35 b36 b37| + + c.val[0][0] = a00*b00 + a01*b10 + a02*b20 + a03*b30 + c.val[0][1] = a00*b01 + a01*b11 + a02*b21 + a03*b31 + + The size of the output tensor's XY-plane must be the following shape [ width * 8, height / 8 ]. All other dimensions must have the same size. + */ + for(size_t k = num_it; k > 0; mtx_a0 += 16, mtx_b0 += 32, --k) + { + const float16x8_t p00 = vld1q_f16(mtx_a0); + const float16x8_t p02 = vld1q_f16(mtx_a0 + 8); + const float16x8_t q00 = vld1q_f16(mtx_b0); + const float16x8_t q02 = vld1q_f16(mtx_b0 + 8); + const float16x8_t q04 = vld1q_f16(mtx_b0 + 16); + const float16x8_t q06 = vld1q_f16(mtx_b0 + 24); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vgetq_lane_f16(p00, 0))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vgetq_lane_f16(p00, 1))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vgetq_lane_f16(p00, 2))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vgetq_lane_f16(p00, 3))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q02, vgetq_lane_f16(p00, 4))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q02, vgetq_lane_f16(p00, 5))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q02, vgetq_lane_f16(p00, 6))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q02, vgetq_lane_f16(p00, 7))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q04, vgetq_lane_f16(p02, 0))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q04, vgetq_lane_f16(p02, 1))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q04, vgetq_lane_f16(p02, 2))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q04, vgetq_lane_f16(p02, 3))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q06, vgetq_lane_f16(p02, 4))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q06, vgetq_lane_f16(p02, 5))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q06, vgetq_lane_f16(p02, 6))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q06, vgetq_lane_f16(p02, 7))); + } + + if(multiply_alpha) + { + c.val[0] = vmulq_f16(c.val[0], alpha_f16); + c.val[1] = vmulq_f16(c.val[1], alpha_f16); + c.val[2] = vmulq_f16(c.val[2], alpha_f16); + c.val[3] = vmulq_f16(c.val[3], alpha_f16); + } + + vst1q_f16(mtx_out + 0 * out_stride, c.val[0]); + vst1q_f16(mtx_out + 1 * out_stride, c.val[1]); + vst1q_f16(mtx_out + 2 * out_stride, c.val[2]); + vst1q_f16(mtx_out + 3 * out_stride, c.val[3]); + }, + ina, inb, out); +#else + ARM_COMPUTE_ERROR("Not implemented"); +#endif +} + +template +void matrix_matrix_multiply_qs8(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) +{ + const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); + const size_t out_stride1 = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); + const size_t out_stride2 = out_stride1 * 2; + const size_t out_stride3 = out_stride1 * 3; + const int num_elems_matrix_b_x = input1->info()->dimension(0); + const int fixed_point_position = input0->info()->fixed_point_position(); + const qint8x8_t alpha_qs8 = vdup_n_qs8(scvt_qs8_f32(alpha, fixed_point_position)); + ARM_COMPUTE_UNUSED(alpha_qs8); + + // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); + + Window win_b; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(input1->info()->num_dimensions() >= 3) + { + win_b = window; + } + // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the cols of the output matrix + // The step along the x direction is 2 times the in_b_stride because for each iteration we compute 2 blocks of size 16x4 + win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, 2 * in_b_stride)); + win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Iterator ina(input0, win_a); + Iterator inb(input1, win_b); + Iterator out(output, window); + + // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW + // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration + // All the values needed for computing a single 32x4 block will be read from consecutive memory positions + execute_window_loop(window, [&](const Coordinates & id) + { + auto mtx_a0 = reinterpret_cast(ina.ptr()); + auto mtx_b0 = reinterpret_cast(inb.ptr()); + auto mtx_b1 = mtx_b0 + in_b_stride; + + qint16x8_t acc00_qs16 = vdupq_n_qs16(0); + qint16x8_t acc10_qs16 = vdupq_n_qs16(0); + qint16x8_t acc20_qs16 = vdupq_n_qs16(0); + qint16x8_t acc30_qs16 = vdupq_n_qs16(0); + + qint16x8_t acc01_qs16 = vdupq_n_qs16(0); + qint16x8_t acc11_qs16 = vdupq_n_qs16(0); + qint16x8_t acc21_qs16 = vdupq_n_qs16(0); + qint16x8_t acc31_qs16 = vdupq_n_qs16(0); + + qint16x8_t acc02_qs16 = vdupq_n_qs16(0); + qint16x8_t acc12_qs16 = vdupq_n_qs16(0); + qint16x8_t acc22_qs16 = vdupq_n_qs16(0); + qint16x8_t acc32_qs16 = vdupq_n_qs16(0); + + qint16x8_t acc03_qs16 = vdupq_n_qs16(0); + qint16x8_t acc13_qs16 = vdupq_n_qs16(0); + qint16x8_t acc23_qs16 = vdupq_n_qs16(0); + qint16x8_t acc33_qs16 = vdupq_n_qs16(0); + + int k = 0; + // This for loop performs 2 accumulations + for(; k <= (num_elems_matrix_b_x - 32); k += 32) + { + const qint8x8_t a0 = vld1_dup_qs8(mtx_a0 + 0); + const qint8x8_t a1 = vld1_dup_qs8(mtx_a0 + 1); + const qint8x8_t a2 = vld1_dup_qs8(mtx_a0 + 2); + const qint8x8_t a3 = vld1_dup_qs8(mtx_a0 + 3); + const qint8x8_t a4 = vld1_dup_qs8(mtx_a0 + 4); + const qint8x8_t a5 = vld1_dup_qs8(mtx_a0 + 5); + const qint8x8_t a6 = vld1_dup_qs8(mtx_a0 + 6); + const qint8x8_t a7 = vld1_dup_qs8(mtx_a0 + 7); + + const qint8x8_t b00 = vld1_qs8(mtx_b0 + 0); + const qint8x8_t b01 = vld1_qs8(mtx_b0 + 8); + const qint8x8_t b10 = vld1_qs8(mtx_b1 + 0); + const qint8x8_t b11 = vld1_qs8(mtx_b1 + 8); + + // First accumulation + acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); + acc10_qs16 = vqmlal_qs8(acc10_qs16, b00, a1, fixed_point_position); + acc20_qs16 = vqmlal_qs8(acc20_qs16, b00, a2, fixed_point_position); + acc30_qs16 = vqmlal_qs8(acc30_qs16, b00, a3, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b10, a0, fixed_point_position); + acc12_qs16 = vqmlal_qs8(acc12_qs16, b10, a1, fixed_point_position); + acc22_qs16 = vqmlal_qs8(acc22_qs16, b10, a2, fixed_point_position); + acc32_qs16 = vqmlal_qs8(acc32_qs16, b10, a3, fixed_point_position); + + const qint8x8_t b02 = vld1_qs8(mtx_b0 + 16); + const qint8x8_t b03 = vld1_qs8(mtx_b0 + 24); + const qint8x8_t b12 = vld1_qs8(mtx_b1 + 16); + const qint8x8_t b13 = vld1_qs8(mtx_b1 + 24); + + acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); + acc11_qs16 = vqmlal_qs8(acc11_qs16, b01, a1, fixed_point_position); + acc21_qs16 = vqmlal_qs8(acc21_qs16, b01, a2, fixed_point_position); + acc31_qs16 = vqmlal_qs8(acc31_qs16, b01, a3, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b11, a0, fixed_point_position); + acc13_qs16 = vqmlal_qs8(acc13_qs16, b11, a1, fixed_point_position); + acc23_qs16 = vqmlal_qs8(acc23_qs16, b11, a2, fixed_point_position); + acc33_qs16 = vqmlal_qs8(acc33_qs16, b11, a3, fixed_point_position); + +#if __arm__ + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b1))); +#endif + + // Second accumulation + acc00_qs16 = vqmlal_qs8(acc00_qs16, b02, a4, fixed_point_position); + acc10_qs16 = vqmlal_qs8(acc10_qs16, b02, a5, fixed_point_position); + acc20_qs16 = vqmlal_qs8(acc20_qs16, b02, a6, fixed_point_position); + acc30_qs16 = vqmlal_qs8(acc30_qs16, b02, a7, fixed_point_position); + acc01_qs16 = vqmlal_qs8(acc01_qs16, b03, a4, fixed_point_position); + acc11_qs16 = vqmlal_qs8(acc11_qs16, b03, a5, fixed_point_position); + acc21_qs16 = vqmlal_qs8(acc21_qs16, b03, a6, fixed_point_position); + acc31_qs16 = vqmlal_qs8(acc31_qs16, b03, a7, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b12, a4, fixed_point_position); + acc12_qs16 = vqmlal_qs8(acc12_qs16, b12, a5, fixed_point_position); + acc22_qs16 = vqmlal_qs8(acc22_qs16, b12, a6, fixed_point_position); + acc32_qs16 = vqmlal_qs8(acc32_qs16, b12, a7, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b13, a4, fixed_point_position); + acc13_qs16 = vqmlal_qs8(acc13_qs16, b13, a5, fixed_point_position); + acc23_qs16 = vqmlal_qs8(acc23_qs16, b13, a6, fixed_point_position); + acc33_qs16 = vqmlal_qs8(acc33_qs16, b13, a7, fixed_point_position); + + mtx_a0 += 8; + mtx_b0 += 32; + mtx_b1 += 32; + } + + // This for loop performs the left over accumulations + for(; k < num_elems_matrix_b_x; k += 16) + { + const qint8x8_t a0 = vld1_dup_qs8(mtx_a0 + 0); + const qint8x8_t a1 = vld1_dup_qs8(mtx_a0 + 1); + const qint8x8_t a2 = vld1_dup_qs8(mtx_a0 + 2); + const qint8x8_t a3 = vld1_dup_qs8(mtx_a0 + 3); + + const qint8x8_t b00 = vld1_qs8(mtx_b0 + 0); + const qint8x8_t b01 = vld1_qs8(mtx_b0 + 8); + const qint8x8_t b10 = vld1_qs8(mtx_b1 + 0); + const qint8x8_t b11 = vld1_qs8(mtx_b1 + 8); + + acc00_qs16 = vqmlal_qs8(acc00_qs16, b00, a0, fixed_point_position); + acc10_qs16 = vqmlal_qs8(acc10_qs16, b00, a1, fixed_point_position); + acc20_qs16 = vqmlal_qs8(acc20_qs16, b00, a2, fixed_point_position); + acc30_qs16 = vqmlal_qs8(acc30_qs16, b00, a3, fixed_point_position); + acc01_qs16 = vqmlal_qs8(acc01_qs16, b01, a0, fixed_point_position); + acc11_qs16 = vqmlal_qs8(acc11_qs16, b01, a1, fixed_point_position); + acc21_qs16 = vqmlal_qs8(acc21_qs16, b01, a2, fixed_point_position); + acc31_qs16 = vqmlal_qs8(acc31_qs16, b01, a3, fixed_point_position); + acc02_qs16 = vqmlal_qs8(acc02_qs16, b10, a0, fixed_point_position); + acc12_qs16 = vqmlal_qs8(acc12_qs16, b10, a1, fixed_point_position); + acc22_qs16 = vqmlal_qs8(acc22_qs16, b10, a2, fixed_point_position); + acc32_qs16 = vqmlal_qs8(acc32_qs16, b10, a3, fixed_point_position); + acc03_qs16 = vqmlal_qs8(acc03_qs16, b11, a0, fixed_point_position); + acc13_qs16 = vqmlal_qs8(acc13_qs16, b11, a1, fixed_point_position); + acc23_qs16 = vqmlal_qs8(acc23_qs16, b11, a2, fixed_point_position); + acc33_qs16 = vqmlal_qs8(acc33_qs16, b11, a3, fixed_point_position); + + mtx_a0 += 4; + mtx_b0 += 16; + mtx_b1 += 16; + } + + // Convert back to qint8x8_t and saturate + qint8x8_t acc00_qs8 = vqmovn_qs16(acc00_qs16); + qint8x8_t acc10_qs8 = vqmovn_qs16(acc10_qs16); + qint8x8_t acc20_qs8 = vqmovn_qs16(acc20_qs16); + qint8x8_t acc30_qs8 = vqmovn_qs16(acc30_qs16); + + qint8x8_t acc01_qs8 = vqmovn_qs16(acc01_qs16); + qint8x8_t acc11_qs8 = vqmovn_qs16(acc11_qs16); + qint8x8_t acc21_qs8 = vqmovn_qs16(acc21_qs16); + qint8x8_t acc31_qs8 = vqmovn_qs16(acc31_qs16); + + qint8x8_t acc02_qs8 = vqmovn_qs16(acc02_qs16); + qint8x8_t acc12_qs8 = vqmovn_qs16(acc12_qs16); + qint8x8_t acc22_qs8 = vqmovn_qs16(acc22_qs16); + qint8x8_t acc32_qs8 = vqmovn_qs16(acc32_qs16); + + qint8x8_t acc03_qs8 = vqmovn_qs16(acc03_qs16); + qint8x8_t acc13_qs8 = vqmovn_qs16(acc13_qs16); + qint8x8_t acc23_qs8 = vqmovn_qs16(acc23_qs16); + qint8x8_t acc33_qs8 = vqmovn_qs16(acc33_qs16); + + // Multiply by the weight of the matrix product (alpha) + if(multiply_alpha) + { + acc00_qs8 = vqmul_qs8(acc00_qs8, alpha_qs8, fixed_point_position); + acc10_qs8 = vqmul_qs8(acc10_qs8, alpha_qs8, fixed_point_position); + acc20_qs8 = vqmul_qs8(acc20_qs8, alpha_qs8, fixed_point_position); + acc30_qs8 = vqmul_qs8(acc30_qs8, alpha_qs8, fixed_point_position); + acc01_qs8 = vqmul_qs8(acc01_qs8, alpha_qs8, fixed_point_position); + acc11_qs8 = vqmul_qs8(acc11_qs8, alpha_qs8, fixed_point_position); + acc21_qs8 = vqmul_qs8(acc21_qs8, alpha_qs8, fixed_point_position); + acc31_qs8 = vqmul_qs8(acc31_qs8, alpha_qs8, fixed_point_position); + acc02_qs8 = vqmul_qs8(acc02_qs8, alpha_qs8, fixed_point_position); + acc12_qs8 = vqmul_qs8(acc12_qs8, alpha_qs8, fixed_point_position); + acc22_qs8 = vqmul_qs8(acc22_qs8, alpha_qs8, fixed_point_position); + acc32_qs8 = vqmul_qs8(acc32_qs8, alpha_qs8, fixed_point_position); + acc03_qs8 = vqmul_qs8(acc03_qs8, alpha_qs8, fixed_point_position); + acc13_qs8 = vqmul_qs8(acc13_qs8, alpha_qs8, fixed_point_position); + acc23_qs8 = vqmul_qs8(acc23_qs8, alpha_qs8, fixed_point_position); + acc33_qs8 = vqmul_qs8(acc33_qs8, alpha_qs8, fixed_point_position); + } + + const auto mtx_out0 = reinterpret_cast(out.ptr()); + + // Store 32x4 output elements + vst1_qs8(mtx_out0 + 0, acc00_qs8); + vst1_qs8(mtx_out0 + 8, acc01_qs8); + vst1_qs8(mtx_out0 + 16, acc02_qs8); + vst1_qs8(mtx_out0 + 24, acc03_qs8); + vst1_qs8(mtx_out0 + out_stride1 + 0, acc10_qs8); + vst1_qs8(mtx_out0 + out_stride1 + 8, acc11_qs8); + vst1_qs8(mtx_out0 + out_stride1 + 16, acc12_qs8); + vst1_qs8(mtx_out0 + out_stride1 + 24, acc13_qs8); + vst1_qs8(mtx_out0 + out_stride2 + 0, acc20_qs8); + vst1_qs8(mtx_out0 + out_stride2 + 8, acc21_qs8); + vst1_qs8(mtx_out0 + out_stride2 + 16, acc22_qs8); + vst1_qs8(mtx_out0 + out_stride2 + 24, acc23_qs8); + vst1_qs8(mtx_out0 + out_stride3 + 0, acc30_qs8); + vst1_qs8(mtx_out0 + out_stride3 + 8, acc31_qs8); + vst1_qs8(mtx_out0 + out_stride3 + 16, acc32_qs8); + vst1_qs8(mtx_out0 + out_stride3 + 24, acc33_qs8); + }, + ina, inb, out); +} + +} // namespace + +NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel() + : _input0(nullptr), _input1(nullptr), _output(nullptr), _alpha(1.0f) +{ +} + +void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::F16, DataType::F32, DataType::QS8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32, DataType::QS8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32, DataType::QS8); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); + + if(output->info()->dimension(1) == 1) + { + ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); + } + + _input0 = input0; + _input1 = input1; + _output = output; + _alpha = alpha; + + unsigned int num_elems_processed_per_iteration_x = 0; + const unsigned int num_elems_processed_per_iteration_y = 4; + + // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication + if((output->info()->dimension(1) == 1)) + { + switch(input0->info()->data_type()) + { + case DataType::F32: + { + num_elems_processed_per_iteration_x = 16; + break; + } + case DataType::QS8: + { + num_elems_processed_per_iteration_x = 32; + break; + } + default: + { + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + + // Configure kernel window + Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x)); + + AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration_x); + + update_window_and_padding(win, + AccessWindowHorizontal(input0->info(), 0, num_elems_processed_per_iteration_x), + AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration_x), + output_access); + + Coordinates coord; + coord.set_num_dimensions(output->info()->num_dimensions()); + output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); + + INEKernel::configure(win); + } + else + { + switch(input0->info()->data_type()) + { + case DataType::F32: + { + num_elems_processed_per_iteration_x = 8; + break; + } + case DataType::QS8: + { + num_elems_processed_per_iteration_x = 32; + break; + } + case DataType::F16: + { +#ifdef ARM_COMPUTE_ENABLE_FP16 + num_elems_processed_per_iteration_x = 8; + break; +#endif + } + default: + { + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + + // Configure kernel window + Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + + AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); + + update_window_and_padding(win, + AccessWindowRectangle(input0->info(), 0, 0, 4, 1, 1.f, 0.25f), + AccessWindowTranspose(input1->info(), 0, 0, 4, 1, 0.f, 0.25f), + output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); + + INEKernel::configure(win); + } +} + +void NEGEMMMatrixMultiplyKernel::run(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + bool multiply_alpha = std::abs(1.0f - _alpha) > 0.00001f; + + // Check if the output tensor is a vector and the data type is F32. If so,the kernel runs the vector-matrix multiplication + if((_output->info()->dimension(1) == 1)) + { + switch(_input0->info()->data_type()) + { + case DataType::F32: + { + multiply_alpha ? vector_matrix_multiply_f32(_input0, _input1, _output, window, _alpha) : + vector_matrix_multiply_f32(_input0, _input1, _output, window, _alpha); + break; + } + case DataType::QS8: + { + multiply_alpha ? vector_matrix_multiply_qs8(_input0, _input1, _output, window, _alpha) : + vector_matrix_multiply_qs8(_input0, _input1, _output, window, _alpha); + break; + } + default: + { + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + } + else + { + switch(_input0->info()->data_type()) + { + case DataType::F32: + { + multiply_alpha ? matrix_matrix_multiply_f32(_input0, _input1, _output, window, _alpha) : + matrix_matrix_multiply_f32(_input0, _input1, _output, window, _alpha); + break; + } + case DataType::QS8: + { + multiply_alpha ? matrix_matrix_multiply_qs8(_input0, _input1, _output, window, _alpha) : + matrix_matrix_multiply_qs8(_input0, _input1, _output, window, _alpha); + break; + } + case DataType::F16: + { +#ifdef ARM_COMPUTE_ENABLE_FP16 + multiply_alpha ? matrix_matrix_multiply_f16(_input0, _input1, _output, window, _alpha) : + matrix_matrix_multiply_f16(_input0, _input1, _output, window, _alpha); + break; +#endif + } + default: + { + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + } + } +} -- cgit v1.2.1