Google Git
Sign in
android / platform / external / ComputeLibrary / 8938bd3f4 / . / src / core / NEON / kernels / NEGEMMMatrixMultiplyKernel.cpp
blob: 69090825fad77c6cea7892b90953ce19b7239ff6 [file] [log] [blame]
/*
* 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/AccessWindowStatic.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 <arm_neon.h>
#include <cstddef>
#include <cstdint>
#include <tuple>
using namespace arm_compute;
namespace arm_compute
{
class Coordinates;
} // namespace arm_compute
namespace
{
template <bool multiply_alpha>
void vector_matrix_multiply_f16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha)
{
#ifdef ARM_COMPUTE_ENABLE_FP16
const auto width_matrix_b = static_cast<int>(output->info()->dimension(0));
const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()));
const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0));
// The implementation computes 32 elements per iteration
const int window_start_x = 32 * info.thread_id;
const int window_step_x = 32 * info.num_threads;
const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x;
ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, " (window_end_x - window_start_x) must be multiple of window_step_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);
const float16x8_t alpha_f16 = vdupq_n_f16(alpha);
ARM_COMPUTE_UNUSED(alpha_f16);
execute_window_loop(win_out, [&](const Coordinates & id)
{
if(id.x() > width_matrix_b)
{
return;
}
float16x8_t acc0 = vdupq_n_f16(0.f);
float16x8_t acc1 = vdupq_n_f16(0.f);
float16x8_t acc2 = vdupq_n_f16(0.f);
float16x8_t acc3 = vdupq_n_f16(0.f);
auto vec_a = reinterpret_cast<const float16_t *>(ina.ptr());
auto matrix_b = reinterpret_cast<const float16_t *>(inb.ptr());
const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a;
for(; vec_a <= (vec_a_end_addr - 4);)
{
const float16x4_t a0l = vld1_f16(vec_a);
float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride);
float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride);
float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride);
float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride);
float16x8_t b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride);
float16x8_t b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride);
float16x8_t b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride);
float16x8_t b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride);
acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 0));
acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 0));
acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 0));
acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 0));
acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 1));
acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 1));
acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 1));
acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 1));
matrix_b += 2 * in_b_stride;
b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride);
b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride);
b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride);
b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride);
b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride);
b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride);
b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride);
b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride);
acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 2));
acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 2));
acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 2));
acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 2));
acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 3));
acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 3));
acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 3));
acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 3));
vec_a += 4;
matrix_b += 2 * in_b_stride;
}
for(; vec_a < vec_a_end_addr;)
{
const float16_t a0 = *vec_a;
const float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride);
const float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride);
const float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride);
const float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride);
acc0 = vaddq_f16(acc0, vmulq_n_f16(b00, a0));
acc1 = vaddq_f16(acc1, vmulq_n_f16(b01, a0));
acc2 = vaddq_f16(acc2, vmulq_n_f16(b02, a0));
acc3 = vaddq_f16(acc3, vmulq_n_f16(b03, a0));
vec_a += 1;
matrix_b += in_b_stride;
}
// Multiply by the weight of matrix product (alpha)
if(multiply_alpha)
{
acc0 = vmulq_f16(acc0, alpha_f16);
acc1 = vmulq_f16(acc1, alpha_f16);
acc2 = vmulq_f16(acc2, alpha_f16);
acc3 = vmulq_f16(acc3, alpha_f16);
}
const auto vec_out = reinterpret_cast<float16_t *>(out.ptr());
vst1q_f16(vec_out + 0, acc0);
vst1q_f16(vec_out + 8, acc1);
vst1q_f16(vec_out + 16, acc2);
vst1q_f16(vec_out + 24, acc3);
},
ina, inb, out);
#else /* ARM_COMPUTE_ENABLE_FP16 */
ARM_COMPUTE_UNUSED(input0);
ARM_COMPUTE_UNUSED(input1);
ARM_COMPUTE_UNUSED(output);
ARM_COMPUTE_UNUSED(window);
ARM_COMPUTE_UNUSED(info);
ARM_COMPUTE_UNUSED(alpha);
ARM_COMPUTE_ERROR("Not implemented");
#endif /* ARM_COMPUTE_ENABLE_FP16 */
}
template <bool multiply_alpha>
void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha)
{
const auto width_matrix_b = static_cast<int>(output->info()->dimension(0));
const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()));
const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0));
// The implementation computes 16 elements per iteration
const int window_start_x = 16 * info.thread_id;
const int window_step_x = 16 * info.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<const float *>(ina.ptr());
auto matrix_b = reinterpret_cast<const float *>(inb.ptr());
#if __arm__
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(vec_a)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + in_b_stride)));
#endif /* __arm__ */
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<const uint8_t *>(vec_a)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 1 * in_b_stride)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 2 * in_b_stride)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 3 * in_b_stride)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 4 * in_b_stride)));
#endif /* __arm__ */
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<float *>(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 <bool multiply_alpha>
void vector_matrix_multiply_qs8(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha)
{
const auto width_matrix_b = static_cast<int>(output->info()->dimension(0));
const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()));
const auto num_elems_vec_a = static_cast<int>(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 * info.thread_id;
const int window_step_x = 32 * info.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<const qint8_t *>(ina.ptr());
auto matrix_b = reinterpret_cast<const qint8_t *>(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(sqcvt_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<qint8_t *>(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 <bool multiply_alpha>
void vector_matrix_multiply_qs16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha)
{
const auto width_matrix_b = static_cast<int>(output->info()->dimension(0));
const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()));
const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0));
const int fixed_point_position = input0->info()->fixed_point_position();
// The implementation computes 16 elements per iteration
const int window_start_x = 16 * info.thread_id;
const int window_step_x = 16 * info.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;
ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, " (window_end_x - window_start_x) must be multiple of window_step_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
qint32x4_t acc00_qs32 = vdupq_n_qs32(0);
qint32x4_t acc01_qs32 = vdupq_n_qs32(0);
qint32x4_t acc02_qs32 = vdupq_n_qs32(0);
qint32x4_t acc03_qs32 = vdupq_n_qs32(0);
auto vec_a = reinterpret_cast<const qint16_t *>(ina.ptr());
auto matrix_b = reinterpret_cast<const qint16_t *>(inb.ptr());
auto vec_a_end_addr = vec_a + num_elems_vec_a;
for(; vec_a <= (vec_a_end_addr - 2);)
{
const qint16x4_t a0 = vld1_dup_qs16(vec_a + 0);
const qint16x4_t a1 = vld1_dup_qs16(vec_a + 1);
const qint16x4_t b00 = vld1_qs16(matrix_b + 0 + 0 * in_b_stride);
const qint16x4_t b01 = vld1_qs16(matrix_b + 4 + 0 * in_b_stride);
const qint16x4_t b02 = vld1_qs16(matrix_b + 8 + 0 * in_b_stride);
const qint16x4_t b03 = vld1_qs16(matrix_b + 12 + 0 * in_b_stride);
const qint16x4_t b10 = vld1_qs16(matrix_b + 0 + 1 * in_b_stride);
const qint16x4_t b11 = vld1_qs16(matrix_b + 4 + 1 * in_b_stride);
const qint16x4_t b12 = vld1_qs16(matrix_b + 8 + 1 * in_b_stride);
const qint16x4_t b13 = vld1_qs16(matrix_b + 12 + 1 * in_b_stride);
// First accumulation
acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position);
acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position);
acc02_qs32 = vqmlal_qs16(acc02_qs32, b02, a0, fixed_point_position);
acc03_qs32 = vqmlal_qs16(acc03_qs32, b03, a0, fixed_point_position);
// Second accumulation
acc00_qs32 = vqmlal_qs16(acc00_qs32, b10, a1, fixed_point_position);
acc01_qs32 = vqmlal_qs16(acc01_qs32, b11, a1, fixed_point_position);
acc02_qs32 = vqmlal_qs16(acc02_qs32, b12, a1, fixed_point_position);
acc03_qs32 = vqmlal_qs16(acc03_qs32, b13, a1, fixed_point_position);
vec_a += 2;
matrix_b += 2 * in_b_stride;
}
for(; vec_a < vec_a_end_addr;)
{
const qint16x4_t a0 = vld1_dup_qs16(vec_a);
const qint16x4_t b00 = vld1_qs16(matrix_b + 0);
const qint16x4_t b01 = vld1_qs16(matrix_b + 4);
const qint16x4_t b02 = vld1_qs16(matrix_b + 8);
const qint16x4_t b03 = vld1_qs16(matrix_b + 12);
acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position);
acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position);
acc02_qs32 = vqmlal_qs16(acc02_qs32, b02, a0, fixed_point_position);
acc03_qs32 = vqmlal_qs16(acc03_qs32, b03, a0, fixed_point_position);
vec_a += 1;
matrix_b += in_b_stride;
}
// Convert back to qint16x4_t and saturate
qint16x4_t acc00_qs16 = vqmovn_qs32(acc00_qs32);
qint16x4_t acc01_qs16 = vqmovn_qs32(acc01_qs32);
qint16x4_t acc02_qs16 = vqmovn_qs32(acc02_qs32);
qint16x4_t acc03_qs16 = vqmovn_qs32(acc03_qs32);
// Multiply by the weight of the matrix product (alpha)
if(multiply_alpha)
{
const qint16x4_t alpha_qs16 = vdup_n_qs16(sqcvt_qs16_f32(alpha, fixed_point_position));
acc00_qs16 = vqmul_qs16(acc00_qs16, alpha_qs16, fixed_point_position);
acc01_qs16 = vqmul_qs16(acc01_qs16, alpha_qs16, fixed_point_position);
acc02_qs16 = vqmul_qs16(acc02_qs16, alpha_qs16, fixed_point_position);
acc03_qs16 = vqmul_qs16(acc03_qs16, alpha_qs16, fixed_point_position);
}
const auto mtx_out0 = reinterpret_cast<qint16_t *>(out.ptr());
// Store 16x4 output elements
vst1_qs16(mtx_out0 + 0, acc00_qs16);
vst1_qs16(mtx_out0 + 4, acc01_qs16);
vst1_qs16(mtx_out0 + 8, acc02_qs16);
vst1_qs16(mtx_out0 + 12, acc03_qs16);
},
ina, inb, out);
}
template <bool multiply_alpha>
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<const float *>(ina.ptr());
auto mtx_b0 = reinterpret_cast<const float *>(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<const uint8_t *>(mtx_a0)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0)));
asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1)));
#endif /* __arm__ */
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<const uint8_t *>(mtx_a0)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1)));
#endif /* __arm__ */
// 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<const uint8_t *>(mtx_a0)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0)));
asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1)));
#endif /* __arm__ */
// 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<const uint8_t *>(mtx_a0)));
asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0)));
asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1)));
#endif /* __arm__ */
// 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<float *>(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 <bool multiply_alpha>
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());
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 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);
const float16x8_t alpha_f16 = vdupq_n_f16(alpha);
execute_window_loop(window, [&](const Coordinates & id)
{
const auto *mtx_a0 = reinterpret_cast<const float16_t *>(ina.ptr());
const auto *mtx_b0 = reinterpret_cast<const float16_t *>(inb.ptr());
auto *mtx_out = reinterpret_cast<float16_t *>(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.
*/
const float16_t *mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x;
for(; mtx_b0 <= (mtx_b0_end_addr - 32);)
{
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)));
mtx_a0 += 16;
mtx_b0 += 32;
}
for(; mtx_b0 < mtx_b0_end_addr;)
{
const float16x4_t p00 = vld1_f16(mtx_a0);
const float16x8_t q00 = vld1q_f16(mtx_b0);
c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vget_lane_f16(p00, 0)));
c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vget_lane_f16(p00, 1)));
c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vget_lane_f16(p00, 2)));
c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vget_lane_f16(p00, 3)));
mtx_a0 += 4;
mtx_b0 += 8;
}
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_ENABLE_FP16 */
ARM_COMPUTE_UNUSED(input0);
ARM_COMPUTE_UNUSED(input1);
ARM_COMPUTE_UNUSED(output);
ARM_COMPUTE_UNUSED(window);
ARM_COMPUTE_UNUSED(alpha);
ARM_COMPUTE_ERROR("Not implemented");
#endif /* ARM_COMPUTE_ENABLE_FP16 */
}
template <bool multiply_alpha>
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(sqcvt_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<const qint8_t *>(ina.ptr());
auto mtx_b0 = reinterpret_cast<const qint8_t *>(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<const uint8_t *>(mtx_a0)));
asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0)));
asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1)));
#endif /* __arm__ */
// 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<qint8_t *>(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);
}
template <bool multiply_alpha>
void matrix_matrix_multiply_qs16(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 qint16x4_t alpha_qs16 = vdup_n_qs16(sqcvt_qs16_f32(alpha, fixed_point_position));
ARM_COMPUTE_UNUSED(alpha_qs16);
// 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
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, 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 8x4 elements per iteration
// All the values needed for computing a single 8x4 block will be read from consecutive memory positions
execute_window_loop(window, [&](const Coordinates & id)
{
auto mtx_a0 = reinterpret_cast<const qint16_t *>(ina.ptr());
auto mtx_b0 = reinterpret_cast<const qint16_t *>(inb.ptr());
auto mtx_b1 = mtx_b0 + in_b_stride;
qint32x4_t acc00_qs32 = vdupq_n_qs32(0);
qint32x4_t acc10_qs32 = vdupq_n_qs32(0);
qint32x4_t acc20_qs32 = vdupq_n_qs32(0);
qint32x4_t acc30_qs32 = vdupq_n_qs32(0);
qint32x4_t acc01_qs32 = vdupq_n_qs32(0);
qint32x4_t acc11_qs32 = vdupq_n_qs32(0);
qint32x4_t acc21_qs32 = vdupq_n_qs32(0);
qint32x4_t acc31_qs32 = vdupq_n_qs32(0);
// This for loop performs 1 accumulation
for(int k = 0; k <= (num_elems_matrix_b_x - 8); k += 8)
{
const qint16x4_t a0 = vld1_dup_qs16(mtx_a0 + 0);
const qint16x4_t a1 = vld1_dup_qs16(mtx_a0 + 1);
const qint16x4_t a2 = vld1_dup_qs16(mtx_a0 + 2);
const qint16x4_t a3 = vld1_dup_qs16(mtx_a0 + 3);
const qint16x4_t b00 = vld1_qs16(mtx_b0 + 0);
const qint16x4_t b01 = vld1_qs16(mtx_b0 + 4);
acc00_qs32 = vqmlal_qs16(acc00_qs32, b00, a0, fixed_point_position);
acc10_qs32 = vqmlal_qs16(acc10_qs32, b00, a1, fixed_point_position);
acc20_qs32 = vqmlal_qs16(acc20_qs32, b00, a2, fixed_point_position);
acc30_qs32 = vqmlal_qs16(acc30_qs32, b00, a3, fixed_point_position);
acc01_qs32 = vqmlal_qs16(acc01_qs32, b01, a0, fixed_point_position);
acc11_qs32 = vqmlal_qs16(acc11_qs32, b01, a1, fixed_point_position);
acc21_qs32 = vqmlal_qs16(acc21_qs32, b01, a2, fixed_point_position);
acc31_qs32 = vqmlal_qs16(acc31_qs32, b01, a3, fixed_point_position);
mtx_a0 += 4;
mtx_b0 += 8;
mtx_b1 += 8;
}
// Convert back to qint16x4_t and saturate
qint16x4_t acc00_qs16 = vqmovn_qs32(acc00_qs32);
qint16x4_t acc10_qs16 = vqmovn_qs32(acc10_qs32);
qint16x4_t acc20_qs16 = vqmovn_qs32(acc20_qs32);
qint16x4_t acc30_qs16 = vqmovn_qs32(acc30_qs32);
qint16x4_t acc01_qs16 = vqmovn_qs32(acc01_qs32);
qint16x4_t acc11_qs16 = vqmovn_qs32(acc11_qs32);
qint16x4_t acc21_qs16 = vqmovn_qs32(acc21_qs32);
qint16x4_t acc31_qs16 = vqmovn_qs32(acc31_qs32);
// Multiply by the weight of the matrix product (alpha)
if(multiply_alpha)
{
acc00_qs16 = vqmul_qs16(acc00_qs16, alpha_qs16, fixed_point_position);
acc10_qs16 = vqmul_qs16(acc10_qs16, alpha_qs16, fixed_point_position);
acc20_qs16 = vqmul_qs16(acc20_qs16, alpha_qs16, fixed_point_position);
acc30_qs16 = vqmul_qs16(acc30_qs16, alpha_qs16, fixed_point_position);
acc01_qs16 = vqmul_qs16(acc01_qs16, alpha_qs16, fixed_point_position);
acc11_qs16 = vqmul_qs16(acc11_qs16, alpha_qs16, fixed_point_position);
acc21_qs16 = vqmul_qs16(acc21_qs16, alpha_qs16, fixed_point_position);
acc31_qs16 = vqmul_qs16(acc31_qs16, alpha_qs16, fixed_point_position);
}
const auto mtx_out0 = reinterpret_cast<qint16_t *>(out.ptr());
// Store 8x4 output elements
vst1_qs16(mtx_out0 + 0, acc00_qs16);
vst1_qs16(mtx_out0 + 4, acc01_qs16);
vst1_qs16(mtx_out0 + out_stride1 + 0, acc10_qs16);
vst1_qs16(mtx_out0 + out_stride1 + 4, acc11_qs16);
vst1_qs16(mtx_out0 + out_stride2 + 0, acc20_qs16);
vst1_qs16(mtx_out0 + out_stride2 + 4, acc21_qs16);
vst1_qs16(mtx_out0 + out_stride3 + 0, acc30_qs16);
vst1_qs16(mtx_out0 + out_stride3 + 4, acc31_qs16);
},
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, DataType::QS16);
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;
}
case DataType::QS16:
{
num_elems_processed_per_iteration_x = 16;
break;
}
#ifdef ARM_COMPUTE_ENABLE_FP16
case DataType::F16:
{
num_elems_processed_per_iteration_x = 32;
break;
}
#endif /* ARM_COMPUTE_ENABLE_FP16 */
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,
AccessWindowStatic(input0->info(), 0, 0, input0->info()->tensor_shape().x(), 1),
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::QS16:
{
num_elems_processed_per_iteration_x = 8;
break;
}
#ifdef ARM_COMPUTE_ENABLE_FP16
case DataType::F16:
{
num_elems_processed_per_iteration_x = 8;
break;
}
#endif /* ARM_COMPUTE_ENABLE_FP16 */
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),
AccessWindowStatic(input1->info(), 0, 0, input1->info()->tensor_shape().x(), ceil_to_multiple(input1->info()->tensor_shape().y(), 4)),
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, const ThreadInfo &info)
{
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. 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<true>(_input0, _input1, _output, window, info, _alpha) :
vector_matrix_multiply_f32<false>(_input0, _input1, _output, window, info, _alpha);
break;
}
case DataType::QS8:
{
multiply_alpha ? vector_matrix_multiply_qs8<true>(_input0, _input1, _output, window, info, _alpha) :
vector_matrix_multiply_qs8<false>(_input0, _input1, _output, window, info, _alpha);
break;
}
case DataType::QS16:
{
multiply_alpha ? vector_matrix_multiply_qs16<true>(_input0, _input1, _output, window, info, _alpha) :
vector_matrix_multiply_qs16<false>(_input0, _input1, _output, window, info, _alpha);
break;
}
#ifdef ARM_COMPUTE_ENABLE_FP16
case DataType::F16:
{
multiply_alpha ? vector_matrix_multiply_f16<true>(_input0, _input1, _output, window, info, _alpha) :
vector_matrix_multiply_f16<false>(_input0, _input1, _output, window, info, _alpha);
break;
}
#endif /* ARM_COMPUTE_ENABLE_FP16 */
default:
{
ARM_COMPUTE_ERROR("Data type not supported");
break;
}
}
}
else
{
switch(_input0->info()->data_type())
{
case DataType::F32:
{
multiply_alpha ? matrix_matrix_multiply_f32<true>(_input0, _input1, _output, window, _alpha) :
matrix_matrix_multiply_f32<false>(_input0, _input1, _output, window, _alpha);
break;
}
case DataType::QS8:
{
multiply_alpha ? matrix_matrix_multiply_qs8<true>(_input0, _input1, _output, window, _alpha) :
matrix_matrix_multiply_qs8<false>(_input0, _input1, _output, window, _alpha);
break;
}
case DataType::QS16:
{
multiply_alpha ? matrix_matrix_multiply_qs16<true>(_input0, _input1, _output, window, _alpha) :
matrix_matrix_multiply_qs16<false>(_input0, _input1, _output, window, _alpha);
break;
}
#ifdef ARM_COMPUTE_ENABLE_FP16
case DataType::F16:
{
multiply_alpha ? matrix_matrix_multiply_f16<true>(_input0, _input1, _output, window, _alpha) :
matrix_matrix_multiply_f16<false>(_input0, _input1, _output, window, _alpha);
break;
}
#endif /* ARM_COMPUTE_ENABLE_FP16 */
default:
{
ARM_COMPUTE_ERROR("Data type not supported");
break;
}
}
}
}