diff options
author | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-11-20 14:20:01 +0000 |
---|---|---|
committer | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-11-27 15:28:19 +0000 |
commit | 568aab689c6f0d293a99ea554786a22c76be18b4 (patch) | |
tree | af141ee56d3d34e5c7cb88a530499f53ebc67364 /src/cpu/kernels/CpuMulKernel.cpp | |
parent | ded5b182675e3166e947a8eb637b5b1e925816ab (diff) | |
download | ComputeLibrary-568aab689c6f0d293a99ea554786a22c76be18b4.tar.gz |
CpuMul changes to enable fp16 in armv8a multi_isa builds
* Moved fp16 and fp32 to their corresponding files
src/cpu/kernels/mul/generic/neon/fp16.cpp and
src/cpu/kernels/mul/generic/neon/fp32.cpp
* Changes in filelist.json: added a new fp16.cpp file for the float16_t kernels
* Partially resolves MLCE-1102
Change-Id: I88f24cf034c11b55ff84644b182ba76c7cb94296
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10778
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Diffstat (limited to 'src/cpu/kernels/CpuMulKernel.cpp')
-rw-r--r-- | src/cpu/kernels/CpuMulKernel.cpp | 221 |
1 files changed, 5 insertions, 216 deletions
diff --git a/src/cpu/kernels/CpuMulKernel.cpp b/src/cpu/kernels/CpuMulKernel.cpp index ba086e3ac6..8001482154 100644 --- a/src/cpu/kernels/CpuMulKernel.cpp +++ b/src/cpu/kernels/CpuMulKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2022 Arm Limited. + * Copyright (c) 2016-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,12 +26,14 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/TensorInfo.h" +#include "src/core/common/Registrars.h" #include "src/core/CPP/Validate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/NEON/NEAsymm.h" #include "src/core/NEON/NESymm.h" #include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/mul/generic/neon/list.h" #include <arm_neon.h> @@ -1170,108 +1172,6 @@ void mul_S32_S32_S32(const ITensor *src1, const ITensor *src2, ITensor *out, con } } -void mul_F32_F32_F32(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale) -{ - // Create input windows - Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); - Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); - - // Clear X Dimension on execution window as we handle manually - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - constexpr int window_step_x = 16 / sizeof(float); - const auto window_start_x = static_cast<int>(window.x().start()); - const auto window_end_x = static_cast<int>(window.x().end()); - const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x(); - - using ExactTagType = typename wrapper::traits::neon_vector<float, window_step_x>::tag_type; - - if (is_broadcast_across_x) - { - const bool is_broadcast_input_2 = input2_win.x().step() == 0; - Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; - Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; - const ITensor *broadcast_tensor = is_broadcast_input_2 ? src2 : src1; - const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1; - - // Clear X Dimension on execution window as we handle manually - non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator broadcast_input(broadcast_tensor, broadcast_win); - Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); - Iterator dst(out, win); - - execute_window_loop( - win, - [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast<const float *>(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast<float *>(dst.ptr()); - - const float broadcast_value = *reinterpret_cast<const float *>(broadcast_input.ptr()); - const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); - const auto scale_vec = wrapper::vdup_n(scale, ExactTagType{}); - - // Compute window_step_x elements per iteration - int x = window_start_x; - for (; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); - auto res = wrapper::vmul(wrapper::vmul(broadcast_value_vec, non_broadcast_v), scale_vec); - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for (; x < window_end_x; ++x) - { - const auto non_broadcast_v = *(non_broadcast_input_ptr + x); - *(output_ptr + x) = broadcast_value * non_broadcast_v * scale; - } - }, - broadcast_input, non_broadcast_input, dst); - } - else - { - // Clear X Dimension on execution window as we handle manually - input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input1(src1, input1_win); - Iterator input2(src2, input2_win); - Iterator dst(out, win); - - execute_window_loop( - win, - [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast<const float *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const float *>(input2.ptr()); - const auto output_ptr = reinterpret_cast<float *>(dst.ptr()); - - // Compute window_step_x elements per iteration - int x = window_start_x; - for (; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto ta1 = wrapper::vloadq(input1_ptr + x); - const auto ta2 = wrapper::vloadq(input2_ptr + x); - const auto scale_vec = wrapper::vdup_n(scale, ExactTagType{}); - const auto res = wrapper::vmul(wrapper::vmul(ta1, ta2), scale_vec); - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for (; x < window_end_x; ++x) - { - const auto ta1 = *(input1_ptr + x); - const auto ta2 = *(input2_ptr + x); - *(output_ptr + x) = ta1 * ta2 * scale; - } - }, - input1, input2, dst); - } -} - void c_mul_F32_F32_F32_n(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window) { // Create input windows @@ -1409,115 +1309,6 @@ void c_mul_F32_F32_F32_n(const ITensor *src1, const ITensor *src2, ITensor *out, } } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -void mul_F16_F16_F16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale) -{ - // Create input windows - Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); - Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape()); - - // Clear X Dimension on execution window as we handle manually - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - constexpr int window_step_x = 16; - const auto window_start_x = static_cast<int>(window.x().start()); - const auto window_end_x = static_cast<int>(window.x().end()); - const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x(); - if (is_broadcast_across_x) - { - const bool is_broadcast_input_2 = input2_win.x().step() == 0; - Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; - Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; - const ITensor *broadcast_tensor = is_broadcast_input_2 ? src2 : src1; - const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1; - // Clear X Dimension on execution window as we handle manually - non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - Iterator broadcast_input(broadcast_tensor, broadcast_win); - Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); - Iterator dst(out, win); - execute_window_loop( - win, - [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast<const float16_t *>(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast<float16_t *>(dst.ptr()); - const auto broadcast_value = *reinterpret_cast<const float16_t *>(broadcast_input.ptr()); - const float16x8x2_t broadcast_value_vec = {{ - vdupq_n_f16(broadcast_value), - vdupq_n_f16(broadcast_value), - }}; - const auto scale_vec = vdupq_n_f16(scale); - // Compute window_step_x elements per iteration - int x = window_start_x; - for (; x <= (window_end_x - window_step_x); x += window_step_x) - { - const float16x8x2_t non_broadcast_v = {{ - vld1q_f16(non_broadcast_input_ptr + x), - vld1q_f16(non_broadcast_input_ptr + x + 8), - }}; - const float16x8x2_t result = {{ - vmulq_f16(vmulq_f16(broadcast_value_vec.val[0], non_broadcast_v.val[0]), scale_vec), - vmulq_f16(vmulq_f16(broadcast_value_vec.val[1], non_broadcast_v.val[1]), scale_vec), - }}; - vst1q_f16(output_ptr + x, result.val[0]); - vst1q_f16(output_ptr + x + 8, result.val[1]); - } - // Compute left-over elements - for (; x < window_end_x; ++x) - { - const auto non_broadcast_v = *(non_broadcast_input_ptr + x); - *(output_ptr + x) = broadcast_value * non_broadcast_v * scale; - } - }, - broadcast_input, non_broadcast_input, dst); - } - else - { - input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - Iterator input1(src1, input1_win); - Iterator input2(src2, input2_win); - Iterator dst(out, win); - execute_window_loop( - win, - [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast<const float16_t *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const float16_t *>(input2.ptr()); - const auto output_ptr = reinterpret_cast<float16_t *>(dst.ptr()); - // Compute window_step_x elements per iteration - int x = window_start_x; - for (; x <= (window_end_x - window_step_x); x += window_step_x) - { - const float16x8x2_t ta1 = {{ - vld1q_f16(input1_ptr + x), - vld1q_f16(input1_ptr + x + 8), - }}; - const float16x8x2_t ta2 = {{ - vld1q_f16(input2_ptr + x), - vld1q_f16(input2_ptr + x + 8), - }}; - const float16x8_t scale_vec = vdupq_n_f16(scale); - const float16x8x2_t result = {{ - vmulq_f16(vmulq_f16(ta1.val[0], ta2.val[0]), scale_vec), - vmulq_f16(vmulq_f16(ta1.val[1], ta2.val[1]), scale_vec), - }}; - vst1q_f16(output_ptr + x, result.val[0]); - vst1q_f16(output_ptr + x + 8, result.val[1]); - } - // Compute left-over elements - for (; x < window_end_x; ++x) - { - const auto ta1 = *(input1_ptr + x); - const auto ta2 = *(input2_ptr + x); - *(output_ptr + x) = ta1 * ta2 * scale; - } - }, - input1, input2, dst); - } -} -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - template <bool is_scale255, bool is_sat> void mul_U8_U8_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, int n) { @@ -1857,13 +1648,11 @@ void CpuMulKernel::configure(ITensorInfo *src1, } } break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: - _func_float = &mul_F16_F16_F16; + _func_float = REGISTER_FP16_NEON(cpu::mul_F16_F16_F16); break; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::F32: - _func_float = &mul_F32_F32_F32; + _func_float = REGISTER_FP32_NEON(cpu::mul_F32_F32_F32); break; default: ARM_COMPUTE_ERROR("You called with the wrong img formats"); |