From 3a9ecdfdc76abd7f9acdab42a1f7e4c0188d6f48 Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Thu, 17 Aug 2023 13:30:22 +0100 Subject: CpuAdd rework to enable fp16 in armv8a multi_isa builds * Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs to be moved to an fp16.cpp file to allow compilation with -march=armv8.2-a+fp16 * fp16.cpp needs to use the template add_same_neon() so it had to be moved from impl.cpp to impl.h Change-Id: I9e64a3101958fcb9c3d5c8e9b148b498b2bee05f Signed-off-by: Pablo Marquez Tello Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10154 Comments-Addressed: Arm Jenkins Reviewed-by: Viet-Hoa Do Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- src/cpu/kernels/add/generic/neon/impl.cpp | 111 +----------------------------- src/cpu/kernels/add/generic/neon/impl.h | 107 +++++++++++++++++++++++++++- 2 files changed, 105 insertions(+), 113 deletions(-) (limited to 'src/cpu/kernels/add') diff --git a/src/cpu/kernels/add/generic/neon/impl.cpp b/src/cpu/kernels/add/generic/neon/impl.cpp index a1734d7dd6..2dde13544a 100644 --- a/src/cpu/kernels/add/generic/neon/impl.cpp +++ b/src/cpu/kernels/add/generic/neon/impl.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2022 Arm Limited. + * Copyright (c) 2020-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -30,106 +30,6 @@ namespace arm_compute { namespace cpu { -template -void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) -{ - /** SIMD vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; - - // Create input windows - Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); - Window input2_win = window.broadcast_if_dimension_le_one(src1->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(ScalarType); - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->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 ? src1 : src0; - const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; - - // 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 output(dst, win); - - execute_window_loop( - win, [&](const Coordinates &) - { - const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - const ScalarType broadcast_value = *reinterpret_cast(broadcast_input.ptr()); - const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); - - // Compute S 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); - const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v); - 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) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v; - } - }, - broadcast_input, non_broadcast_input, output); - } - 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(src0, input1_win); - Iterator input2(src1, input2_win); - Iterator output(dst, win); - - execute_window_loop( - win, [&](const Coordinates &) - { - const auto input1_ptr = reinterpret_cast(input1.ptr()); - const auto input2_ptr = reinterpret_cast(input2.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto val1 = wrapper::vloadq(input1_ptr + x); - const auto val2 = wrapper::vloadq(input2_ptr + x); - const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2); - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const auto val1 = *(input1_ptr + x); - const auto val2 = *(input2_ptr + x); - *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2; - } - }, - input1, input2, output); - } -} - bool sub_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst) { return add_sub_q8_neon_fixedpoint_possible(src0, src1, dst, false); @@ -745,15 +645,6 @@ void add_sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITens } } -template void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); -template void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); -template void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); -template void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); - -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -template void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); -#endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ - template void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); template void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); diff --git a/src/cpu/kernels/add/generic/neon/impl.h b/src/cpu/kernels/add/generic/neon/impl.h index d544ef5728..fb786c5bc1 100644 --- a/src/cpu/kernels/add/generic/neon/impl.h +++ b/src/cpu/kernels/add/generic/neon/impl.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,15 +23,116 @@ */ #ifndef SRC_CORE_NEON_KERNELS_ADD_IMPL_H #define SRC_CORE_NEON_KERNELS_ADD_IMPL_H +#include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/wrapper.h" + namespace arm_compute { namespace cpu { template -void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); +void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + + // Create input windows + Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(src1->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(ScalarType); + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->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 ? src1 : src0; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; + + // 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 output(dst, win); + + execute_window_loop( + win, [&](const Coordinates &) + { + const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + const ScalarType broadcast_value = *reinterpret_cast(broadcast_input.ptr()); + const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); + + // Compute S 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); + const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v); + 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) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v; + } + }, + broadcast_input, non_broadcast_input, output); + } + 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(src0, input1_win); + Iterator input2(src1, input2_win); + Iterator output(dst, win); + + execute_window_loop( + win, [&](const Coordinates &) + { + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(input2.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto val1 = wrapper::vloadq(input1_ptr + x); + const auto val2 = wrapper::vloadq(input2_ptr + x); + const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2); + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const auto val1 = *(input1_ptr + x); + const auto val2 = *(input2_ptr + x); + *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2; + } + }, + input1, input2, output); + } +} bool add_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst); @@ -50,4 +151,4 @@ template void add_sub_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition); } // namespace cpu } // namespace arm_compute -#endif // SRC_CORE_NEON_KERNELS_ADD_IMPL_H \ No newline at end of file +#endif // SRC_CORE_NEON_KERNELS_ADD_IMPL_H -- cgit v1.2.1