From 8a9a0fbf6509159cbb8721500517a1497c390f27 Mon Sep 17 00:00:00 2001 From: Anton Vainer Date: Sun, 9 Jan 2022 14:37:12 +0200 Subject: Select kernel decoupling Resolves COMPMID-4614 Signed-off-by: Anton Vainer Change-Id: I19476d43b8e685de2eed973425d5d31b9cdb84ca Signed-off-by: Anton Vainer Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6960 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena --- src/core/NEON/kernels/NESelectKernel.cpp | 260 +++++++++++++------------------ 1 file changed, 111 insertions(+), 149 deletions(-) (limited to 'src/core/NEON/kernels/NESelectKernel.cpp') diff --git a/src/core/NEON/kernels/NESelectKernel.cpp b/src/core/NEON/kernels/NESelectKernel.cpp index 7c988e9fab..b8c9b244ee 100644 --- a/src/core/NEON/kernels/NESelectKernel.cpp +++ b/src/core/NEON/kernels/NESelectKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -34,6 +34,10 @@ #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/core/common/Registrars.h" + +#include "src/cpu/kernels/select/list.h" + #include #include #include @@ -42,125 +46,123 @@ namespace arm_compute { namespace { -template -void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, - const int window_step_x, const int window_start_x, const int window_end_x, const int limit, VectorType (*condition_conversion)(const uint8_t *)) -{ - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator condition(cond, win); - Iterator input1(in1, win); - Iterator input2(in2, win); - Iterator output(out, win); - - execute_window_loop(win, [&](const Coordinates &) - { - auto output_ptr = reinterpret_cast(output.ptr()); - const auto condition_ptr = reinterpret_cast(condition.ptr()); - const auto input1_ptr = reinterpret_cast(input1.ptr()); - const auto input2_ptr = reinterpret_cast(input2.ptr()); - int x = window_start_x; - for(; x <= limit; x += window_step_x) - { - const auto c = (*condition_conversion)(condition_ptr + x); - const auto a = wrapper::vloadq(input1_ptr + x); - const auto b = wrapper::vloadq(input2_ptr + x); - wrapper::vstore(output_ptr + x, wrapper::vbsl(c, a, b)); - } - for(; x < window_end_x; ++x) - { - const auto c = *(condition_ptr + x); - const auto a = *(input1_ptr + x); - const auto b = *(input2_ptr + x); - *(output_ptr + x) = static_cast(c) ? a : b; - } - }, - condition, input1, input2, output); -} - -template -void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +struct SelectKernelSelectorData { - const auto 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()); + DataType dt; + bool is_same_rank; +}; - select_op(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType - { - static const auto zero = wrapper::vdup_n(static_cast(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero); - }); -} +using SelectorPtr = std::add_pointer::type; +using KernelPtr = std::add_pointer::type; -template -void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +struct SelectKernelSelector { - const auto 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 char *name; + const SelectorPtr is_selected; + KernelPtr ukernel; +}; - select_op(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType - { - static const auto zero = wrapper::vdup_n(static_cast(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero); - }); -} - -template -void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +static const SelectKernelSelector available_kernels[] = { - const auto 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()); - - select_op(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType { - static const auto zero = wrapper::vdup_n(static_cast(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero); - }); -} + "neon_s8_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S8 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s8_select_same_rank) + }, + { + "neon_s16_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S16 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s16_select_same_rank) + }, + { + "neon_s32_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S32 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s32_select_same_rank) + }, + { + "neon_u8_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U8 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u8_select_same_rank) + }, + { + "neon_u16_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U16 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u16_select_same_rank) + }, + { + "neon_u32_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U32 && data.is_same_rank == true; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u32_select_same_rank) + }, + { + "neon_s8_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S8 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s8_select_not_same_rank) + }, + { + "neon_s16_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S16 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s16_select_not_same_rank) + }, + { + "neon_s32_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::S32 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_s32_select_not_same_rank) + }, + { + "neon_u8_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U8 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u8_select_not_same_rank) + }, + { + "neon_u16_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U16 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u16_select_not_same_rank) + }, + { + "neon_u32_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::U32 && data.is_same_rank == false; }, + REGISTER_INTEGER_NEON(arm_compute::cpu::neon_u32_select_not_same_rank) + }, + { + "neon_f16_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::F16 && data.is_same_rank == true; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_f16_select_same_rank) + }, + { + "neon_f16_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::F16 && data.is_same_rank == false; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_f16_select_not_same_rank) + }, + { + "neon_f32_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::F32 && data.is_same_rank == true; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_f32_select_same_rank) + }, + { + "neon_f32_not_same_rank", + [](const SelectKernelSelectorData & data) { return data.dt == DataType::F32 && data.is_same_rank == false; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_f32_select_not_same_rank) + }, +}; -template -void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +const SelectKernelSelector *get_implementation(const SelectKernelSelectorData &data) { - ARM_COMPUTE_UNUSED(window); - - auto output_ptr = reinterpret_cast(out->buffer()); - const auto condition_ptr = reinterpret_cast(cond->buffer()); - const auto input1_ptr = reinterpret_cast(in1->buffer()); - const auto input2_ptr = reinterpret_cast(in2->buffer()); - - const int outer_size = cond->info()->total_size() / cond->info()->element_size(); - const int inner_size = (in1->info()->total_size() / in1->info()->element_size()) / outer_size; - int offset = 0; - const int step = 16 / in1->info()->element_size(); - - for(int i = 0; i < outer_size; ++i) + for(const auto &uk : available_kernels) { - int x = offset; - const auto input_ptr = static_cast(*(condition_ptr + i)) ? input1_ptr : input2_ptr; - for(; x <= offset + inner_size - step; x += step) - { - wrapper::vstore(output_ptr + x, wrapper::vloadq(input_ptr + x)); - } - if(x <= offset + inner_size - (step / 2)) - { - wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x)); - x += step / 2; - } - for(; x < offset + inner_size; ++x) + if(uk.is_selected(data)) { - *(output_ptr + x) = *(input_ptr + x); + return &uk; } - offset += inner_size; } + return nullptr; } + } // namespace NESelectKernel::NESelectKernel() - : _function(nullptr), _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false) + : /*_function(nullptr), */ _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false) { } @@ -178,51 +180,6 @@ void NESelectKernel::configure(const ITensor *c, const ITensor *x, const ITensor _output = output; _has_same_rank = (c->info()->tensor_shape().num_dimensions() == x->info()->tensor_shape().num_dimensions()); - std::string function_to_call("op_"); - function_to_call += string_from_data_type(x->info()->data_type()); - - static std::map map_function; - - if(_has_same_rank) - { - map_function = - { - { "op_S8", &select_op_8 }, - { "op_S16", &select_op_16 }, - { "op_S32", &select_op_32 }, - { "op_U8", &select_op_8 }, - { "op_U16", &select_op_16 }, - { "op_U32", &select_op_32 }, - { "op_F32", &select_op_32 } - }; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - map_function["op_F16"] = &select_op_16; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - } - else - { - map_function = - { - { "op_S8", &select_op_not_same_rank }, - { "op_S16", &select_op_not_same_rank }, - { "op_S32", &select_op_not_same_rank }, - { "op_U8", &select_op_not_same_rank }, - { "op_U16", &select_op_not_same_rank }, - { "op_U32", &select_op_not_same_rank }, - { "op_F32", &select_op_not_same_rank } - }; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - map_function["op_F16"] = &select_op_not_same_rank; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - } - - auto it = map_function.find(function_to_call); - - if(it != map_function.end()) - { - _function = it->second; - } - Window win = calculate_max_window(*x->info()); INEKernel::configure(win); } @@ -254,7 +211,12 @@ void NESelectKernel::run(const Window &window, const ThreadInfo &info) ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_function == nullptr); - _function(_c, _x, _y, _output, window); + ARM_COMPUTE_ERROR_ON(_output == nullptr); + ARM_COMPUTE_ERROR_ON(_output->info() == nullptr); + + const auto *uk = get_implementation(SelectKernelSelectorData{ _output->info()->data_type(), _has_same_rank }); + ARM_COMPUTE_ERROR_ON(uk == nullptr); + ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr); + uk->ukernel(_c, _x, _y, _output, window); } } // namespace arm_compute -- cgit v1.2.1