From c3c352e60050f3deacad767e429a88dc24b31af0 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 18 Mar 2021 10:59:40 +0000 Subject: Add Queue support Queues are responsible for scheduling operators and performing other runtime related activities like for example tuning. Signed-off-by: Georgios Pinitas Change-Id: I0366d9048470d277b8cbf59fa42f95c0ae57c5c9 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5487 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Reviewed-by: Michalis Spyrou Comments-Addressed: Arm Jenkins --- .../CpuDirectConvolutionOutputStageKernel.cpp | 513 +++++++++++++++++++++ .../kernels/CpuDirectConvolutionStageKernel.cpp | 513 --------------------- src/core/cpu/kernels/activation/NEON/fp16.cpp | 217 --------- src/core/cpu/kernels/activation/NEON/fp32.cpp | 212 --------- src/core/cpu/kernels/activation/NEON/qasymm8.cpp | 262 ----------- .../cpu/kernels/activation/NEON/qasymm8_signed.cpp | 261 ----------- src/core/cpu/kernels/activation/NEON/qsymm16.cpp | 138 ------ src/core/cpu/kernels/activation/SVE/fp16.cpp | 130 ------ src/core/cpu/kernels/activation/SVE/fp32.cpp | 131 ------ src/core/cpu/kernels/activation/SVE/qasymm8.cpp | 254 ---------- .../cpu/kernels/activation/SVE/qasymm8_signed.cpp | 253 ---------- src/core/cpu/kernels/activation/SVE/qsymm16.cpp | 120 ----- src/core/cpu/kernels/activation/neon/fp16.cpp | 217 +++++++++ src/core/cpu/kernels/activation/neon/fp32.cpp | 212 +++++++++ src/core/cpu/kernels/activation/neon/qasymm8.cpp | 262 +++++++++++ .../cpu/kernels/activation/neon/qasymm8_signed.cpp | 261 +++++++++++ src/core/cpu/kernels/activation/neon/qsymm16.cpp | 138 ++++++ src/core/cpu/kernels/activation/sve/fp16.cpp | 130 ++++++ src/core/cpu/kernels/activation/sve/fp32.cpp | 131 ++++++ src/core/cpu/kernels/activation/sve/qasymm8.cpp | 254 ++++++++++ .../cpu/kernels/activation/sve/qasymm8_signed.cpp | 253 ++++++++++ src/core/cpu/kernels/activation/sve/qsymm16.cpp | 120 +++++ 22 files changed, 2491 insertions(+), 2491 deletions(-) create mode 100644 src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.cpp delete mode 100644 src/core/cpu/kernels/CpuDirectConvolutionStageKernel.cpp delete mode 100644 src/core/cpu/kernels/activation/NEON/fp16.cpp delete mode 100644 src/core/cpu/kernels/activation/NEON/fp32.cpp delete mode 100644 src/core/cpu/kernels/activation/NEON/qasymm8.cpp delete mode 100644 src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp delete mode 100644 src/core/cpu/kernels/activation/NEON/qsymm16.cpp delete mode 100644 src/core/cpu/kernels/activation/SVE/fp16.cpp delete mode 100644 src/core/cpu/kernels/activation/SVE/fp32.cpp delete mode 100644 src/core/cpu/kernels/activation/SVE/qasymm8.cpp delete mode 100644 src/core/cpu/kernels/activation/SVE/qasymm8_signed.cpp delete mode 100644 src/core/cpu/kernels/activation/SVE/qsymm16.cpp create mode 100644 src/core/cpu/kernels/activation/neon/fp16.cpp create mode 100644 src/core/cpu/kernels/activation/neon/fp32.cpp create mode 100644 src/core/cpu/kernels/activation/neon/qasymm8.cpp create mode 100644 src/core/cpu/kernels/activation/neon/qasymm8_signed.cpp create mode 100644 src/core/cpu/kernels/activation/neon/qsymm16.cpp create mode 100644 src/core/cpu/kernels/activation/sve/fp16.cpp create mode 100644 src/core/cpu/kernels/activation/sve/fp32.cpp create mode 100644 src/core/cpu/kernels/activation/sve/qasymm8.cpp create mode 100644 src/core/cpu/kernels/activation/sve/qasymm8_signed.cpp create mode 100644 src/core/cpu/kernels/activation/sve/qsymm16.cpp (limited to 'src/core/cpu') diff --git a/src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.cpp b/src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.cpp new file mode 100644 index 0000000000..5f7a574e5a --- /dev/null +++ b/src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.cpp @@ -0,0 +1,513 @@ +/* + * Copyright (c) 2017-2021 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 "src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/CPP/Validate.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEFixedPoint.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" + +#include +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +namespace +{ +Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, + const DirectConvolutionLayerOutputStageKernelInfo &info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::UNKNOWN); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::S32, DataType::F32); + + if(bias != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias); + ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(0) != src->dimension(get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::CHANNEL))); + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); + } + + if(src->data_type() == DataType::S32) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst == nullptr, "In-place computation not allowed for quantized output"); + } + + // Checks performed when output is configured + if((dst != nullptr) && (dst->total_size() != 0)) + { + if(is_data_type_float(src->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); + } + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); + } + else if(src->data_type() == DataType::S32) + { + // In case of quantized computation and unconfigured output, the output data type must be provided through DirectConvolutionLayerOutputStageKernelInfo + ARM_COMPUTE_RETURN_ERROR_ON((info.output_data_type != DataType::QASYMM8) && (info.output_data_type != DataType::QASYMM8_SIGNED)); + } + + return Status{}; +} + +template +typename std::enable_if::value, void>::type +output_stage_nchw(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, + int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) +{ + const bool has_bias = bias != nullptr; + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + + ARM_COMPUTE_ERROR_ON(src->info()->data_layout() == DataLayout::UNKNOWN); + ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier); + ARM_COMPUTE_UNUSED(result_shift); + ARM_COMPUTE_UNUSED(result_offset_after_shift); + + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16 / src->info()->element_size(); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, win); + Iterator out(dst, win); + execute_window_loop(win, [&](const Coordinates & id) + { + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Get bias and pointer to input + const auto in_ptr = reinterpret_cast(in.ptr()) + x; + auto v_in = wrapper::vloadq(in_ptr); + + // Accumulate bias + if(has_bias) + { + const auto vb = wrapper::vdup_n(*reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))), ExactTagType{}); + v_in = wrapper::vadd(v_in, vb); + } + + const auto out_ptr = reinterpret_cast(out.ptr()) + x; + wrapper::vstore(out_ptr, v_in); + } + + // Left-overs loop + for(; x < window_end_x; ++x) + { + // Get bias and pointer to input + auto s_in = *(reinterpret_cast(in.ptr()) + x); + + // Accumulate bias + if(has_bias) + { + const auto b = *reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))); + s_in += b; + } + + *(reinterpret_cast(out.ptr()) + x) = s_in; + } + + }, + in, out); +} + +template +typename std::enable_if::value, void>::type +output_stage_nhwc(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, + int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) +{ + const bool has_bias = bias != nullptr; + ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier); + ARM_COMPUTE_UNUSED(result_shift); + ARM_COMPUTE_UNUSED(result_offset_after_shift); + + Window window_bias = window; + window_bias.set(Window::DimX, Window::Dimension(0, 1, 1)); + window_bias.set(Window::DimY, Window::Dimension(0, 0, 0)); + window_bias.set(Window::DimZ, Window::Dimension(0, 0, 0)); + window_bias.set(3, Window::Dimension(0, 0, 0)); + + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16 / src->info()->element_size(); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, win); + Iterator bi(bias, window_bias); + Iterator out(dst, win); + + execute_window_loop(win, [&](const Coordinates &) + { + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Get bias and pointer to input + const auto in_ptr = reinterpret_cast(in.ptr()); + auto v_in = wrapper::vloadq(in_ptr + x); + + // Accumulate bias + if(has_bias) + { + const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; + v_in = wrapper::vadd(v_in, wrapper::vloadq(bias_ptr)); + } + + const auto out_ptr = reinterpret_cast(out.ptr()); + wrapper::vstore(out_ptr + x, v_in); + } + + // Left-overs loop + for(; x < window_end_x; ++x) + { + // Get bias and pointer to input + auto s_in = *(reinterpret_cast(in.ptr()) + x); + + // Accumulate bias + if(has_bias) + { + const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; + s_in += *bias_ptr; + } + + const auto out_ptr = reinterpret_cast(out.ptr()); + *(out_ptr + x) = s_in; + } + }, + in, bi, out); +} + +// Quantized case +template < typename TOut, typename std::enable_if < std::is_same::value || std::is_same::value, int >::type = 0 > +void output_stage_nchw(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, + int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) +{ + const bool has_bias = bias != nullptr; + using VectorType = typename wrapper::traits::neon_bitvector_t; + using TagType = typename wrapper::traits::neon_bitvector_tag_t; + + const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift); + + const VectorType min = wrapper::vdup_n(std::numeric_limits::lowest(), TagType{}); + const VectorType max = wrapper::vdup_n(std::numeric_limits::max(), TagType{}); + + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16 / src->info()->element_size(); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, win); + Iterator out(dst, win); + + execute_window_loop(win, [&](const Coordinates & id) + { + + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Get bias and pointer to input + const auto in_ptr = reinterpret_cast(in.ptr()) + x; + int32x4x4_t v_in = + { + { + wrapper::vloadq(in_ptr), + wrapper::vloadq(in_ptr + 4), + wrapper::vloadq(in_ptr + 8), + wrapper::vloadq(in_ptr + 12) + } + }; + + // Accumulate bias + if(has_bias) + { + const auto vb = wrapper::vdup_n(*reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))), TagType{}); + v_in = + { + { + wrapper::vadd(v_in.val[0], vb), + wrapper::vadd(v_in.val[1], vb), + wrapper::vadd(v_in.val[2], vb), + wrapper::vadd(v_in.val[3], vb) + } + }; + } + + const auto out_ptr = reinterpret_cast(out.ptr()) + x; + wrapper::vstore(out_ptr, finalize_quantization(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, + min, max, false)); + } + + // Left-overs loop + for(; x < window_end_x; ++x) + { + // Get bias and pointer to input + int32_t s_in = *(reinterpret_cast(in.ptr()) + x); + + // Accumulate bias + if(has_bias) + { + const auto b = *reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))); + s_in += b; + } + + const auto out_ptr = reinterpret_cast(out.ptr()) + x; + *out_ptr = finalize_quantization(s_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, + std::numeric_limits::lowest(), std::numeric_limits::max(), false); + } + }, + in, out); +} +template < typename TOut, typename std::enable_if < std::is_same::value || std::is_same::value, int >::type = 0 > +void output_stage_nhwc(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, + int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) +{ + const bool has_bias = bias != nullptr; + using VectorType = typename wrapper::traits::neon_bitvector_t; + using TagType = typename wrapper::traits::neon_bitvector_tag_t; + + const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift); + + const VectorType min = wrapper::vdup_n(std::numeric_limits::lowest(), TagType{}); + const VectorType max = wrapper::vdup_n(std::numeric_limits::max(), TagType{}); + + Window window_bias = window; + window_bias.set(Window::DimX, Window::Dimension(0, 1, 1)); + window_bias.set(Window::DimY, Window::Dimension(0, 0, 0)); + window_bias.set(Window::DimZ, Window::Dimension(0, 0, 0)); + window_bias.set(3, Window::Dimension(0, 0, 0)); + + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16 / src->info()->element_size(); + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, win); + Iterator bi(bias, window_bias); + Iterator out(dst, win); + + execute_window_loop(win, [&](const Coordinates &) + { + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Get bias and pointer to input + const auto in_ptr = reinterpret_cast(in.ptr()) + x; + int32x4x4_t v_in = + { + { + wrapper::vloadq(in_ptr), + wrapper::vloadq(in_ptr + 4), + wrapper::vloadq(in_ptr + 8), + wrapper::vloadq(in_ptr + 12), + } + }; + + // Accumulate bias + if(has_bias) + { + const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; + + wrapper::vadd(v_in.val[0], wrapper::vloadq(bias_ptr)); + wrapper::vadd(v_in.val[1], wrapper::vloadq(bias_ptr + 4)); + wrapper::vadd(v_in.val[2], wrapper::vloadq(bias_ptr + 8)); + wrapper::vadd(v_in.val[3], wrapper::vloadq(bias_ptr + 12)); + } + + const auto out_ptr = reinterpret_cast(out.ptr()) + x; + wrapper::vstore(out_ptr, finalize_quantization(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, min, max, false)); + } + + // Left-overs loop + for(; x < window_end_x; ++x) + { + // Get bias and pointer to input + const auto in_ptr = reinterpret_cast(in.ptr()) + x; + int32_t s_in = *in_ptr; + + // Accumulate bias + if(has_bias) + { + const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; + s_in += *bias_ptr; + } + + const auto out_ptr = reinterpret_cast(out.ptr()) + x; + *out_ptr = finalize_quantization(s_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, + std::numeric_limits::lowest(), std::numeric_limits::max(), false); + } + }, + in, bi, out); +} +} // namespace + +void CpuDirectConvolutionOutputStageKernel::configure(ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst, + const DirectConvolutionLayerOutputStageKernelInfo &info) +{ + ARM_COMPUTE_UNUSED(bias); + // Perform validation step + ARM_COMPUTE_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, info)); + + _func = nullptr; + _result_fixedpoint_multiplier = info.result_fixedpoint_multiplier; + _result_shift = info.result_shift; + _result_offset_after_shift = info.result_offset_after_shift; + + // Auto-initialize output output if required + if(dst != nullptr) + { + // Work out expected output data type + const DataType output_dt = (src->data_type() == DataType::S32) ? info.output_data_type : DataType::S32; + // Output tensor auto initialization if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_data_type(output_dt)); + } + + Window win = calculate_max_window(*src, Steps()); + + ICpuKernel::configure(win); + + const bool is_qasymm8_signed = (dst != nullptr) ? is_data_type_quantized_asymmetric_signed(dst->data_type()) : false; + + // Set appropriate function + if(src->data_layout() == DataLayout::NCHW) + { + switch(src->data_type()) + { + case DataType::S32: + { + if(is_qasymm8_signed) + { + _func = &output_stage_nchw; + } + else + { + _func = &output_stage_nchw; + } + break; + } +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + { + _func = &output_stage_nchw; + break; + } +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::F32: + { + _func = &output_stage_nchw; + break; + } + default: + { + ARM_COMPUTE_ERROR("Unsupported combination of types among the inputs."); + } + } + } + else + { + switch(src->data_type()) + { + case DataType::S32: + { + if(is_qasymm8_signed) + { + _func = &output_stage_nhwc; + } + else + { + _func = &output_stage_nhwc; + } + break; + } +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + { + _func = &output_stage_nhwc; + break; + } +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::F32: + { + _func = &output_stage_nhwc; + break; + } + default: + { + ARM_COMPUTE_ERROR("Unsupported combination of types among the inputs."); + } + } + } +} + +Status CpuDirectConvolutionOutputStageKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, + const DirectConvolutionLayerOutputStageKernelInfo &info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, info)); + return Status{}; +} + +void CpuDirectConvolutionOutputStageKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + ARM_COMPUTE_ERROR_ON(_func == nullptr); + + auto src = tensors.get_tensor(TensorType::ACL_SRC_0); + auto bias = tensors.get_const_tensor(TensorType::ACL_SRC_1); + auto dst = tensors.get_tensor(TensorType::ACL_DST); + + (*_func)(src, bias, window, dst, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift); +} + +const char *CpuDirectConvolutionOutputStageKernel::name() const +{ + return "CpuDirectConvolutionOutputStageKernel"; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/CpuDirectConvolutionStageKernel.cpp b/src/core/cpu/kernels/CpuDirectConvolutionStageKernel.cpp deleted file mode 100644 index 5f7a574e5a..0000000000 --- a/src/core/cpu/kernels/CpuDirectConvolutionStageKernel.cpp +++ /dev/null @@ -1,513 +0,0 @@ -/* - * Copyright (c) 2017-2021 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 "src/core/cpu/kernels/CpuDirectConvolutionOutputStageKernel.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/utils/misc/Traits.h" -#include "src/core/CPP/Validate.h" -#include "src/core/NEON/NEAsymm.h" -#include "src/core/NEON/NEFixedPoint.h" -#include "src/core/NEON/wrapper/wrapper.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace cpu -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, - const DirectConvolutionLayerOutputStageKernelInfo &info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::S32, DataType::F32); - - if(bias != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias); - ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(0) != src->dimension(get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::CHANNEL))); - ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); - } - - if(src->data_type() == DataType::S32) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst == nullptr, "In-place computation not allowed for quantized output"); - } - - // Checks performed when output is configured - if((dst != nullptr) && (dst->total_size() != 0)) - { - if(is_data_type_float(src->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); - } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); - } - else if(src->data_type() == DataType::S32) - { - // In case of quantized computation and unconfigured output, the output data type must be provided through DirectConvolutionLayerOutputStageKernelInfo - ARM_COMPUTE_RETURN_ERROR_ON((info.output_data_type != DataType::QASYMM8) && (info.output_data_type != DataType::QASYMM8_SIGNED)); - } - - return Status{}; -} - -template -typename std::enable_if::value, void>::type -output_stage_nchw(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, - int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) -{ - const bool has_bias = bias != nullptr; - /** SIMD vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; - - ARM_COMPUTE_ERROR_ON(src->info()->data_layout() == DataLayout::UNKNOWN); - ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier); - ARM_COMPUTE_UNUSED(result_shift); - ARM_COMPUTE_UNUSED(result_offset_after_shift); - - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16 / src->info()->element_size(); - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(src, win); - Iterator out(dst, win); - execute_window_loop(win, [&](const Coordinates & id) - { - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - // Get bias and pointer to input - const auto in_ptr = reinterpret_cast(in.ptr()) + x; - auto v_in = wrapper::vloadq(in_ptr); - - // Accumulate bias - if(has_bias) - { - const auto vb = wrapper::vdup_n(*reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))), ExactTagType{}); - v_in = wrapper::vadd(v_in, vb); - } - - const auto out_ptr = reinterpret_cast(out.ptr()) + x; - wrapper::vstore(out_ptr, v_in); - } - - // Left-overs loop - for(; x < window_end_x; ++x) - { - // Get bias and pointer to input - auto s_in = *(reinterpret_cast(in.ptr()) + x); - - // Accumulate bias - if(has_bias) - { - const auto b = *reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))); - s_in += b; - } - - *(reinterpret_cast(out.ptr()) + x) = s_in; - } - - }, - in, out); -} - -template -typename std::enable_if::value, void>::type -output_stage_nhwc(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, - int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) -{ - const bool has_bias = bias != nullptr; - ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier); - ARM_COMPUTE_UNUSED(result_shift); - ARM_COMPUTE_UNUSED(result_offset_after_shift); - - Window window_bias = window; - window_bias.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_bias.set(Window::DimY, Window::Dimension(0, 0, 0)); - window_bias.set(Window::DimZ, Window::Dimension(0, 0, 0)); - window_bias.set(3, Window::Dimension(0, 0, 0)); - - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16 / src->info()->element_size(); - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(src, win); - Iterator bi(bias, window_bias); - Iterator out(dst, win); - - execute_window_loop(win, [&](const Coordinates &) - { - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - // Get bias and pointer to input - const auto in_ptr = reinterpret_cast(in.ptr()); - auto v_in = wrapper::vloadq(in_ptr + x); - - // Accumulate bias - if(has_bias) - { - const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; - v_in = wrapper::vadd(v_in, wrapper::vloadq(bias_ptr)); - } - - const auto out_ptr = reinterpret_cast(out.ptr()); - wrapper::vstore(out_ptr + x, v_in); - } - - // Left-overs loop - for(; x < window_end_x; ++x) - { - // Get bias and pointer to input - auto s_in = *(reinterpret_cast(in.ptr()) + x); - - // Accumulate bias - if(has_bias) - { - const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; - s_in += *bias_ptr; - } - - const auto out_ptr = reinterpret_cast(out.ptr()); - *(out_ptr + x) = s_in; - } - }, - in, bi, out); -} - -// Quantized case -template < typename TOut, typename std::enable_if < std::is_same::value || std::is_same::value, int >::type = 0 > -void output_stage_nchw(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, - int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) -{ - const bool has_bias = bias != nullptr; - using VectorType = typename wrapper::traits::neon_bitvector_t; - using TagType = typename wrapper::traits::neon_bitvector_tag_t; - - const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift); - - const VectorType min = wrapper::vdup_n(std::numeric_limits::lowest(), TagType{}); - const VectorType max = wrapper::vdup_n(std::numeric_limits::max(), TagType{}); - - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16 / src->info()->element_size(); - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(src, win); - Iterator out(dst, win); - - execute_window_loop(win, [&](const Coordinates & id) - { - - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - // Get bias and pointer to input - const auto in_ptr = reinterpret_cast(in.ptr()) + x; - int32x4x4_t v_in = - { - { - wrapper::vloadq(in_ptr), - wrapper::vloadq(in_ptr + 4), - wrapper::vloadq(in_ptr + 8), - wrapper::vloadq(in_ptr + 12) - } - }; - - // Accumulate bias - if(has_bias) - { - const auto vb = wrapper::vdup_n(*reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))), TagType{}); - v_in = - { - { - wrapper::vadd(v_in.val[0], vb), - wrapper::vadd(v_in.val[1], vb), - wrapper::vadd(v_in.val[2], vb), - wrapper::vadd(v_in.val[3], vb) - } - }; - } - - const auto out_ptr = reinterpret_cast(out.ptr()) + x; - wrapper::vstore(out_ptr, finalize_quantization(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, - min, max, false)); - } - - // Left-overs loop - for(; x < window_end_x; ++x) - { - // Get bias and pointer to input - int32_t s_in = *(reinterpret_cast(in.ptr()) + x); - - // Accumulate bias - if(has_bias) - { - const auto b = *reinterpret_cast(bias->ptr_to_element(Coordinates(id.z()))); - s_in += b; - } - - const auto out_ptr = reinterpret_cast(out.ptr()) + x; - *out_ptr = finalize_quantization(s_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, - std::numeric_limits::lowest(), std::numeric_limits::max(), false); - } - }, - in, out); -} -template < typename TOut, typename std::enable_if < std::is_same::value || std::is_same::value, int >::type = 0 > -void output_stage_nhwc(ITensor *src, const ITensor *bias, const Window &window, ITensor *dst, - int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift) -{ - const bool has_bias = bias != nullptr; - using VectorType = typename wrapper::traits::neon_bitvector_t; - using TagType = typename wrapper::traits::neon_bitvector_tag_t; - - const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift); - - const VectorType min = wrapper::vdup_n(std::numeric_limits::lowest(), TagType{}); - const VectorType max = wrapper::vdup_n(std::numeric_limits::max(), TagType{}); - - Window window_bias = window; - window_bias.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_bias.set(Window::DimY, Window::Dimension(0, 0, 0)); - window_bias.set(Window::DimZ, Window::Dimension(0, 0, 0)); - window_bias.set(3, Window::Dimension(0, 0, 0)); - - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16 / src->info()->element_size(); - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(src, win); - Iterator bi(bias, window_bias); - Iterator out(dst, win); - - execute_window_loop(win, [&](const Coordinates &) - { - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - // Get bias and pointer to input - const auto in_ptr = reinterpret_cast(in.ptr()) + x; - int32x4x4_t v_in = - { - { - wrapper::vloadq(in_ptr), - wrapper::vloadq(in_ptr + 4), - wrapper::vloadq(in_ptr + 8), - wrapper::vloadq(in_ptr + 12), - } - }; - - // Accumulate bias - if(has_bias) - { - const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; - - wrapper::vadd(v_in.val[0], wrapper::vloadq(bias_ptr)); - wrapper::vadd(v_in.val[1], wrapper::vloadq(bias_ptr + 4)); - wrapper::vadd(v_in.val[2], wrapper::vloadq(bias_ptr + 8)); - wrapper::vadd(v_in.val[3], wrapper::vloadq(bias_ptr + 12)); - } - - const auto out_ptr = reinterpret_cast(out.ptr()) + x; - wrapper::vstore(out_ptr, finalize_quantization(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, min, max, false)); - } - - // Left-overs loop - for(; x < window_end_x; ++x) - { - // Get bias and pointer to input - const auto in_ptr = reinterpret_cast(in.ptr()) + x; - int32_t s_in = *in_ptr; - - // Accumulate bias - if(has_bias) - { - const auto bias_ptr = reinterpret_cast(bi.ptr()) + x; - s_in += *bias_ptr; - } - - const auto out_ptr = reinterpret_cast(out.ptr()) + x; - *out_ptr = finalize_quantization(s_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, - std::numeric_limits::lowest(), std::numeric_limits::max(), false); - } - }, - in, bi, out); -} -} // namespace - -void CpuDirectConvolutionOutputStageKernel::configure(ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst, - const DirectConvolutionLayerOutputStageKernelInfo &info) -{ - ARM_COMPUTE_UNUSED(bias); - // Perform validation step - ARM_COMPUTE_ERROR_ON_NULLPTR(src); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, info)); - - _func = nullptr; - _result_fixedpoint_multiplier = info.result_fixedpoint_multiplier; - _result_shift = info.result_shift; - _result_offset_after_shift = info.result_offset_after_shift; - - // Auto-initialize output output if required - if(dst != nullptr) - { - // Work out expected output data type - const DataType output_dt = (src->data_type() == DataType::S32) ? info.output_data_type : DataType::S32; - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*dst, src->clone()->set_data_type(output_dt)); - } - - Window win = calculate_max_window(*src, Steps()); - - ICpuKernel::configure(win); - - const bool is_qasymm8_signed = (dst != nullptr) ? is_data_type_quantized_asymmetric_signed(dst->data_type()) : false; - - // Set appropriate function - if(src->data_layout() == DataLayout::NCHW) - { - switch(src->data_type()) - { - case DataType::S32: - { - if(is_qasymm8_signed) - { - _func = &output_stage_nchw; - } - else - { - _func = &output_stage_nchw; - } - break; - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - { - _func = &output_stage_nchw; - break; - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - { - _func = &output_stage_nchw; - break; - } - default: - { - ARM_COMPUTE_ERROR("Unsupported combination of types among the inputs."); - } - } - } - else - { - switch(src->data_type()) - { - case DataType::S32: - { - if(is_qasymm8_signed) - { - _func = &output_stage_nhwc; - } - else - { - _func = &output_stage_nhwc; - } - break; - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - { - _func = &output_stage_nhwc; - break; - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - { - _func = &output_stage_nhwc; - break; - } - default: - { - ARM_COMPUTE_ERROR("Unsupported combination of types among the inputs."); - } - } - } -} - -Status CpuDirectConvolutionOutputStageKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, - const DirectConvolutionLayerOutputStageKernelInfo &info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, info)); - return Status{}; -} - -void CpuDirectConvolutionOutputStageKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - - auto src = tensors.get_tensor(TensorType::ACL_SRC_0); - auto bias = tensors.get_const_tensor(TensorType::ACL_SRC_1); - auto dst = tensors.get_tensor(TensorType::ACL_DST); - - (*_func)(src, bias, window, dst, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift); -} - -const char *CpuDirectConvolutionOutputStageKernel::name() const -{ - return "CpuDirectConvolutionOutputStageKernel"; -} -} // namespace kernels -} // namespace cpu -} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/NEON/fp16.cpp b/src/core/cpu/kernels/activation/NEON/fp16.cpp deleted file mode 100644 index 6f2d5d8533..0000000000 --- a/src/core/cpu/kernels/activation/NEON/fp16.cpp +++ /dev/null @@ -1,217 +0,0 @@ -/* - * Copyright (c) 2020-2021 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 "src/core/NEON/NEMath.h" - -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Validate.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include -#include - -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) - -namespace arm_compute -{ -namespace cpu -{ -namespace -{ -#ifndef __aarch64__ -inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &mask) -{ - auto int_in = vreinterpretq_u16_f16(in); - return vreinterpretq_f16_u16(wrapper::vand(int_in, mask)); -} -#endif /* __aarch64__ */ -} // namespace - -void fp16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - /** SIMD vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - constexpr int window_step_x = 8; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - // In case of non-aarch64, a small delta value is added to the input - // to prevent NAN values caused by zeros in inputs to SQRT. - // In case of aarh64, we call vsqrt directly, so we don't use delta. -#ifndef __aarch64__ - const auto delta = wrapper::vdup_n(static_cast((1e-7), ExactTagType {})); -#endif /* __aarch64__ */ - - const auto const_1 = wrapper::vdup_n(static_cast(1.f), ExactTagType{}); - const auto const_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - const auto const_6 = wrapper::vdup_n(static_cast(6.f), ExactTagType{}); - const auto const_3 = wrapper::vdup_n(static_cast(3.f), ExactTagType{}); - const auto const_inv_6 = wrapper::vdup_n(static_cast(0.166666667f), ExactTagType{}); - - constexpr float soft_relu_thresh = 12.f; - const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast(soft_relu_thresh), ExactTagType{}); - - const auto va = wrapper::vdup_n(static_cast(act_info.a()), ExactTagType{}); - const auto vb = wrapper::vdup_n(static_cast(act_info.b()), ExactTagType{}); - const auto a = static_cast(act_info.a()); - const auto b = static_cast(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - wrapper::traits::neon_bitvector_t tmp; - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vin = wrapper::vloadq(input_ptr + x); - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = wrapper::vabs(vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = wrapper::vmla(vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = wrapper::vmax(const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: -#ifdef __aarch64__ - tmp = wrapper::vsqrt(vin); -#else /* __aarch64__ */ - { - const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0, ExactTagType{})); - tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); - tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); - } -#endif /* __aarch64__ */ - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = wrapper::vmul(vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - wrapper::vstore(output_ptr + x, tmp); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const float16_t in = *(reinterpret_cast(input_ptr + x)); - float16_t tmp; - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = std::abs(in); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = a * in + b; - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = static_cast(1) / (static_cast(1) + std::exp(-in)); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = std::max(static_cast(0), in); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = std::min(a, std::max(static_cast(0), in)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = std::min(a, std::max(b, in)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = (in > 0) ? in : a * in; - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = (in > soft_relu_thresh) ? in : std::log(static_cast(1) + std::exp(in)); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = (in >= 0) ? in : a * (std::exp(in) - 1); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = std::sqrt(in); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = in * in; - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = a * std::tanh(b * in); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = in; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - *(output_ptr + x) = tmp; - } - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute - -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/core/cpu/kernels/activation/NEON/fp32.cpp b/src/core/cpu/kernels/activation/NEON/fp32.cpp deleted file mode 100644 index 54301d45ad..0000000000 --- a/src/core/cpu/kernels/activation/NEON/fp32.cpp +++ /dev/null @@ -1,212 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/ITensorPack.h" -#include "arm_compute/core/Window.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace cpu -{ -namespace -{ -#ifndef __aarch64__ -inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask) -{ - auto int_in = vreinterpretq_u32_f32(in); - return vreinterpretq_f32_u32(wrapper::vand(int_in, mask)); -} -#endif /* __aarch64__ */ -} // namespace - -void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - /** SIMD vector tag type. */ - using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t; - - constexpr int window_step_x = 4; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - // In case of non-aarch64, a small delta value is added to the input - // to prevent NAN values caused by zeros in inputs to SQRT. - // In case of aarh64, we call vsqrt directly, so we don't use delta. -#ifndef __aarch64__ - const auto delta = wrapper::vdup_n(static_cast(1e-24), ExactTagType {}); -#endif /* __aarch64__ */ - const auto const_1 = wrapper::vdup_n(static_cast(1.f), ExactTagType {}); - const auto const_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - const auto const_6 = wrapper::vdup_n(static_cast(6.f), ExactTagType{}); - const auto const_3 = wrapper::vdup_n(static_cast(3.f), ExactTagType{}); - const auto const_inv_6 = wrapper::vdup_n(static_cast(0.166666667f), ExactTagType{}); - - constexpr float soft_relu_thresh = 12.f; - const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast(soft_relu_thresh), ExactTagType{}); - - const auto va = wrapper::vdup_n(static_cast(act_info.a()), ExactTagType{}); - const auto vb = wrapper::vdup_n(static_cast(act_info.b()), ExactTagType{}); - const auto a = static_cast(act_info.a()); - const auto b = static_cast(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - wrapper::traits::neon_bitvector_t tmp; - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vin = wrapper::vloadq(input_ptr + x); - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = wrapper::vabs(vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = wrapper::vmla(vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = wrapper::vmax(const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: -#ifdef __aarch64__ - tmp = wrapper::vsqrt(vin); -#else /* __aarch64__ */ - { - const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{})); - tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); - tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); - } -#endif /* __aarch64__ */ - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = wrapper::vmul(vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - wrapper::vstore(output_ptr + x, tmp); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const float in = *(reinterpret_cast(input_ptr + x)); - float tmp; - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = std::abs(in); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = a * in + b; - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = static_cast(1) / (static_cast(1) + std::exp(-in)); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = std::max(static_cast(0), in); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = std::min(a, std::max(static_cast(0), in)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = std::min(a, std::max(b, in)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = (in > 0) ? in : a * in; - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = (in > soft_relu_thresh) ? in : std::log(static_cast(1) + std::exp(in)); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = (in >= 0) ? in : a * (std::exp(in) - 1); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = std::sqrt(in); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = in * in; - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = a * std::tanh(b * in); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = in; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - *(output_ptr + x) = tmp; - } - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/NEON/qasymm8.cpp b/src/core/cpu/kernels/activation/NEON/qasymm8.cpp deleted file mode 100644 index a1217435b6..0000000000 --- a/src/core/cpu/kernels/activation/NEON/qasymm8.cpp +++ /dev/null @@ -1,262 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/Window.h" -#include "src/core/NEON/NEAsymm.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace cpu -{ -void qasymm8_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - constexpr int window_step_x = 16; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const qasymm8x16_t va = vdupq_n_u8(quantize_qasymm8(act_info.a(), qi_in)); - const qasymm8x16_t vb = vdupq_n_u8(quantize_qasymm8(act_info.b(), qi_in)); - const qasymm8_t a = quantize_qasymm8(act_info.a(), qi_in); - const qasymm8_t b = quantize_qasymm8(act_info.b(), qi_in); - const qasymm8_t const_0 = quantize_qasymm8(0.f, qi_in); - const qasymm8x16_t vconst_0 = vdupq_n_u8(const_0); - const auto vconst_1 = vdupq_n_f32(1.f); -#ifndef __aarch64__ - const auto vconst_0_f32 = vdupq_n_f32(0); -#endif // __aarch64__ - const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); - const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); - const float a_f32 = act_info.a(); - const float b_f32 = act_info.b(); - const auto const_6_f32 = vdupq_n_f32(6.f); - const auto const_0_f32 = vdupq_n_f32(0.f); - const auto const_3_f32 = vdupq_n_f32(3.f); - const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); - - // Initialise scale/offset for re-quantization - float s = qi_in.scale / qi_out.scale; - float o = -qi_in.offset * s + qi_out.offset; - float32x4_t vs = vdupq_n_f32(s); - float32x4_t vo = vdupq_n_f32(o); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - wrapper::traits::neon_bitvector_t tmp; - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vin = wrapper::vloadq(input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = vmaxq_u8(vconst_0, vin); - // Re-quantize to new output space - tmp = vmlaq_qasymm8(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = vminq_u8(va, vmaxq_u8(vconst_0, vin)); - // Re-quantize to new output space - tmp = vmlaq_qasymm8(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = vminq_u8(va, vmaxq_u8(vb, vin)); - // Re-quantize to new output space - tmp = vmlaq_qasymm8(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), - } - }; - // Re-quantize to new output space - tmp = vquantize(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), - } - }; - // Re-quantize to new output space - tmp = vquantize(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), - wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), - wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), - wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), - } - }; - // Re-quantize to new output space - tmp = vquantize(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - const auto vin_deq = vdequantize(vin, qi_in); - -#ifdef __aarch64__ - const uint32x4x4_t pos_mask = - { - { - wrapper::vcgtz(vin_deq.val[0]), - wrapper::vcgtz(vin_deq.val[1]), - wrapper::vcgtz(vin_deq.val[2]), - wrapper::vcgtz(vin_deq.val[3]), - } - }; -#else // __aarch64__ - const uint32x4x4_t pos_mask = - { - { - wrapper::vcgt(vin_deq.val[0], vconst_0_f32), - wrapper::vcgt(vin_deq.val[1], vconst_0_f32), - wrapper::vcgt(vin_deq.val[2], vconst_0_f32), - wrapper::vcgt(vin_deq.val[3], vconst_0_f32), - } - }; -#endif // __aarch64__ - - const float32x4x4_t tmp_dep = - { - { - wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), - wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), - wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), - wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), - } - }; - - tmp = vquantize(tmp_dep, qi_out); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - wrapper::vstore(output_ptr + x, tmp); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - qasymm8_t in = *(reinterpret_cast(input_ptr + x)); - qasymm8_t tmp = 0; - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - tmp = std::max(const_0, in); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - tmp = std::min(a, std::max(const_0, in)); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - tmp = std::min(a, std::max(b, in)); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - float tmp_f = dequantize_qasymm8(in, qi_in); - tmp_f = 1.f / (1.f + std::exp(-tmp_f)); - tmp = quantize_qasymm8(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - float tmp_f = dequantize_qasymm8(in, qi_in); - tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); - tmp = quantize_qasymm8(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - float tmp_f = dequantize_qasymm8(in, qi_in); - tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); - tmp = quantize_qasymm8(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - float tmp_f = dequantize_qasymm8(in, qi_in); - tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; - tmp = quantize_qasymm8(tmp_f, qi_out); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - *(output_ptr + x) = tmp; - } - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp b/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp deleted file mode 100644 index 8b40bf8e72..0000000000 --- a/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp +++ /dev/null @@ -1,261 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/Window.h" -#include "src/core/NEON/NEAsymm.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace cpu -{ -void qasymm8_signed_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - constexpr int window_step_x = 16; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const qasymm8x16_signed_t va = vdupq_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in)); - const qasymm8x16_signed_t vb = vdupq_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in)); - const qasymm8_signed_t a = quantize_qasymm8_signed(act_info.a(), qi_in); - const qasymm8_signed_t b = quantize_qasymm8_signed(act_info.b(), qi_in); - const qasymm8_signed_t const_0 = quantize_qasymm8_signed(0.f, qi_in); - const qasymm8x16_signed_t vconst_0 = vdupq_n_s8(const_0); - const auto vconst_1 = vdupq_n_f32(1.f); -#ifndef __aarch64__ - const auto vconst_0_f32 = vdupq_n_f32(1.f); -#endif // __aarch64__ - const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); - const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); - const float a_f32 = act_info.a(); - const float b_f32 = act_info.b(); - const auto const_6_f32 = vdupq_n_f32(6.f); - const auto const_0_f32 = vdupq_n_f32(0.f); - const auto const_3_f32 = vdupq_n_f32(3.f); - const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); - - // Initialise scale/offset for re-quantization - float s = qi_in.scale / qi_out.scale; - float o = -qi_in.offset * s + qi_out.offset; - float32x4_t vs = vdupq_n_f32(s); - float32x4_t vo = vdupq_n_f32(o); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - wrapper::traits::neon_bitvector_t tmp; - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vin = wrapper::vloadq(input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = vmaxq_s8(vconst_0, vin); - // Re-quantize to new output space - tmp = vmlaq_qasymm8_signed(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = vminq_s8(va, vmaxq_s8(vconst_0, vin)); - // Re-quantize to new output space - tmp = vmlaq_qasymm8_signed(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = vminq_s8(va, vmaxq_s8(vb, vin)); - // Re-quantize to new output space - tmp = vmlaq_qasymm8_signed(tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), - } - }; - // Re-quantize to new output space - tmp = vquantize_signed(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), - } - }; - // Re-quantize to new output space - tmp = vquantize_signed(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - // De-quantize - const auto vin_deq = vdequantize(vin, qi_in); - // Perform activation - const float32x4x4_t tmp_dep = - { - { - wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), - wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), - wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), - wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), - } - }; - // Re-quantize to new output space - tmp = vquantize_signed(tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - const auto vin_deq = vdequantize(vin, qi_in); - -#ifdef __aarch64__ - const uint32x4x4_t pos_mask = - { - { - wrapper::vcgtz(vin_deq.val[0]), - wrapper::vcgtz(vin_deq.val[1]), - wrapper::vcgtz(vin_deq.val[2]), - wrapper::vcgtz(vin_deq.val[3]), - } - }; -#else // __aarch64__ - const uint32x4x4_t pos_mask = - { - { - wrapper::vcgt(vin_deq.val[0], vconst_0_f32), - wrapper::vcgt(vin_deq.val[1], vconst_0_f32), - wrapper::vcgt(vin_deq.val[2], vconst_0_f32), - wrapper::vcgt(vin_deq.val[3], vconst_0_f32), - } - }; -#endif // __aarch64__ - - const float32x4x4_t tmp_dep = - { - { - wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), - wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), - wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), - wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), - } - }; - - tmp = vquantize_signed(tmp_dep, qi_out); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - wrapper::vstore(output_ptr + x, tmp); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - qasymm8_signed_t in = *(reinterpret_cast(input_ptr + x)); - qasymm8_signed_t tmp = 0; - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - tmp = std::max(const_0, in); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - tmp = std::min(a, std::max(const_0, in)); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - tmp = std::min(a, std::max(b, in)); - tmp = utility::clamp(tmp * s + o); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - float tmp_f = dequantize_qasymm8_signed(in, qi_in); - tmp_f = 1.f / (1.f + std::exp(-tmp_f)); - tmp = quantize_qasymm8_signed(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - float tmp_f = dequantize_qasymm8_signed(in, qi_in); - tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); - tmp = quantize_qasymm8_signed(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - float tmp_f = dequantize_qasymm8_signed(in, qi_in); - tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); - tmp = quantize_qasymm8_signed(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - float tmp_f = dequantize_qasymm8_signed(in, qi_in); - tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; - tmp = quantize_qasymm8_signed(tmp_f, qi_out); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - *(output_ptr + x) = tmp; - } - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/NEON/qsymm16.cpp b/src/core/cpu/kernels/activation/NEON/qsymm16.cpp deleted file mode 100644 index 54b41820f2..0000000000 --- a/src/core/cpu/kernels/activation/NEON/qsymm16.cpp +++ /dev/null @@ -1,138 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/ITensorPack.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/experimental/Types.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/NEON/NESymm.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace cpu -{ -void qsymm16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - constexpr int window_step_x = 8; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const auto vconst_1 = vdupq_n_f32(1.f); - const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); - const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); - const float a_f32 = act_info.a(); - const float b_f32 = act_info.b(); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - wrapper::traits::neon_bitvector_t tmp; - ARM_COMPUTE_UNUSED(tmp); - - // Compute S elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vin = wrapper::vloadq(input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = vdequantize_int16(vin, qi_in.scale); - // Perform activation - const float32x4x2_t tmp_dep = - { - { - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), - wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), - } - }; - // Re-quantize to new output space - tmp = vquantize_int16(tmp_dep, qi_out.scale); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = vdequantize_int16(vin, qi_in.scale); - // Perform activation - const float32x4x2_t tmp_dep = - { - { - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), - wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), - } - }; - // Re-quantize to new output space - tmp = vquantize_int16(tmp_dep, qi_out.scale); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - wrapper::vstore(output_ptr + x, tmp); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - qsymm16_t in = *(reinterpret_cast(input_ptr + x)); - qsymm16_t tmp = 0; - if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - float tmp_f = dequantize_qsymm16(in, qi_in.scale); - tmp_f = 1.f / (1.f + std::exp(-tmp_f)); - tmp = quantize_qsymm16(tmp_f, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - float tmp_f = dequantize_qsymm16(in, qi_in.scale); - tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); - tmp = quantize_qsymm16(tmp_f, qi_out); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - *(output_ptr + x) = tmp; - } - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/SVE/fp16.cpp b/src/core/cpu/kernels/activation/SVE/fp16.cpp deleted file mode 100644 index bf31fd7d93..0000000000 --- a/src/core/cpu/kernels/activation/SVE/fp16.cpp +++ /dev/null @@ -1,130 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/ITensorPack.h" -#include "arm_compute/core/Window.h" - -#include -#include - -#if defined(__ARM_FEATURE_SVE) -#include "src/core/NEON/SVEMath.h" -#include - -namespace arm_compute -{ -namespace cpu -{ -void fp16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const auto const_1 = svdup_n_f16(1.f); - const auto const_0 = svdup_n_f16(0.f); - const auto const_6 = svdup_n_f16(6.f); - const auto const_3 = svdup_n_f16(3.f); - const auto const_inv_6 = svdup_n_f16(0.166666667f); - - const auto va = svdup_n_f16(act_info.a()); - const auto vb = svdup_n_f16(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - svfloat16_t tmp; - - int x = window_start_x; - svbool_t pg = svwhilelt_b16(x, window_end_x); - do - { - const auto vin = svld1_f16(pg, input_ptr + x); - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = svabs_f16_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = svmla_f16_z(pg, vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = svmax_f16_z(pg, const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = svadd_f16_z(pg, svmul_f16_z(pg, svmin_f16_z(pg, vin, const_0), va), svmax_f16_z(pg, vin, const_0)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = svlog_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, vin))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = svsel_f16(svcmpgt_f16(pg, vin, const_0), vin, svmul_f16_z(pg, va, svsub_f16_z(pg, svexp_f16_z(pg, vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = svsqrt_f16_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = svmul_f16_z(pg, vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = svmul_f16_z(pg, va, svtanh_f16_z(pg, svmul_f16_z(pg, vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = svmul_f16_z(pg, vin, svmul_f16_z(pg, const_inv_6, svmin_f16_z(pg, const_6, svmax_f16_z(pg, const_0, svadd_f16_z(pg, vin, const_3))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - svst1_f16(pg, output_ptr + x, tmp); - - x += svcnth(); - pg = svwhilelt_b16(x, window_end_x); - - } - while(svptest_any(svptrue_b16(), pg)); - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute -#endif // __ARM_FEATURE_SVE \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/SVE/fp32.cpp b/src/core/cpu/kernels/activation/SVE/fp32.cpp deleted file mode 100644 index 75f9f8a4c3..0000000000 --- a/src/core/cpu/kernels/activation/SVE/fp32.cpp +++ /dev/null @@ -1,131 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/ITensorPack.h" -#include "arm_compute/core/Window.h" -#include "src/core/NEON/SVEMath.h" - -#include -#include - -#if defined(__ARM_FEATURE_SVE) -#include - -namespace arm_compute -{ -namespace cpu -{ -void fp32_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const auto const_1 = svdup_n_f32(1.f); - const auto const_0 = svdup_n_f32(0.f); - const auto const_6 = svdup_n_f32(6.f); - const auto const_3 = svdup_n_f32(3.f); - const auto const_inv_6 = svdup_n_f32(0.166666667f); - - const auto va = svdup_n_f32(act_info.a()); - const auto vb = svdup_n_f32(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - svfloat32_t tmp; - - // Compute S elements per iteration - int x = window_start_x; - svbool_t pg = svwhilelt_b32(x, window_end_x); - do - { - const auto vin = svld1_f32(pg, input_ptr + x); - switch(act) - { - case ActivationLayerInfo::ActivationFunction::ABS: - tmp = svabs_f32_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::LINEAR: - tmp = svmla_f32_z(pg, vb, va, vin); - break; - case ActivationLayerInfo::ActivationFunction::LOGISTIC: - tmp = svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, vin)))); - break; - case ActivationLayerInfo::ActivationFunction::RELU: - tmp = svmax_f32_z(pg, const_0, vin); - break; - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: - tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, vin)); - break; - case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - tmp = svadd_f32_z(pg, svmul_f32_z(pg, svmin_f32_z(pg, vin, const_0), va), svmax_f32_z(pg, vin, const_0)); - break; - case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - tmp = svlog_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, vin))); - break; - case ActivationLayerInfo::ActivationFunction::ELU: - tmp = svsel_f32(svcmpgt_f32(pg, vin, const_0), vin, svmul_f32_z(pg, va, svsub_f32_z(pg, svexp_f32_z(pg, vin), const_1))); - break; - case ActivationLayerInfo::ActivationFunction::SQRT: - tmp = svsqrt_f32_z(pg, vin); - break; - case ActivationLayerInfo::ActivationFunction::SQUARE: - tmp = svmul_f32_z(pg, vin, vin); - break; - case ActivationLayerInfo::ActivationFunction::TANH: - tmp = svmul_f32_z(pg, va, svtanh_f32_z(pg, svmul_f32_z(pg, vb, vin))); - break; - case ActivationLayerInfo::ActivationFunction::IDENTITY: - tmp = vin; - break; - case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - tmp = svmul_f32_z(pg, vin, svmul_f32_z(pg, const_inv_6, svmin_f32_z(pg, const_6, svmax_f32_z(pg, const_0, svadd_f32_z(pg, vin, const_3))))); - break; - default: - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - svst1_f32(pg, output_ptr + x, tmp); - - x += svcntw(); - pg = svwhilelt_b32(x, window_end_x); - - } - while(svptest_any(svptrue_b32(), pg)); - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute -#endif // __ARM_FEATURE_SVE \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/SVE/qasymm8.cpp b/src/core/cpu/kernels/activation/SVE/qasymm8.cpp deleted file mode 100644 index 228b4ae530..0000000000 --- a/src/core/cpu/kernels/activation/SVE/qasymm8.cpp +++ /dev/null @@ -1,254 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/Window.h" - -#include -#include - -#if defined(__ARM_FEATURE_SVE2) -#include "src/core/NEON/SVEAsymm.h" -#include "src/core/NEON/SVEMath.h" -#include - -namespace arm_compute -{ -namespace cpu -{ -void qasymm8_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const auto va = svdup_n_u8(quantize_qasymm8(act_info.a(), qi_in)); - const auto vb = svdup_n_u8(quantize_qasymm8(act_info.b(), qi_in)); - const auto const_0 = quantize_qasymm8(0.f, qi_in); - const auto vconst_0 = svdup_n_u8(const_0); - const auto vconst_1 = svdup_n_f32(1.f); - const auto va_f32 = svdup_n_f32(act_info.a()); - const auto vb_f32 = svdup_n_f32(act_info.b()); - const auto const_6_f32 = svdup_n_f32(6.f); - const auto const_0_f32 = svdup_n_f32(0.f); - const auto const_3_f32 = svdup_n_f32(3.f); - const auto const_inv_6_f32 = svdup_n_f32(0.166666667f); - - // Initialise scale/offset for re-quantization - bool requant = true; - if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) - { - requant = false; - } - float s = qi_in.scale / qi_out.scale; - float o = -qi_in.offset * s + qi_out.offset; - auto vs = svdup_n_f32(s); - auto vo = svdup_n_f32(o); - - // Initialise scale/offset for re-quantization with int32_t - const auto voffset_in = svdup_n_s32(qi_in.offset); - int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - const auto vs_s32 = svdup_n_s32(s_s32); - const auto vo_s32 = svdup_n_s32(o_s32); - - // Initialise scale/offset for re-quantization for leaky relu - int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); - const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - svuint8_t tmp; - - int x = window_start_x; - svbool_t pg = svwhilelt_b8(x, window_end_x); - do - { - const auto vin = svld1_u8(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = svmax_u8_z(pg, vconst_0, vin); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); - // Re-quantize to new output space - tmp = svmla_qasymm8_z(pg, tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - svbool_t p0, p1, p2, p3; - svint32x4_t tmp_dep; - - // Expand to int32 - const svint32x4_t vin_s32 = - { - { { - svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin))), - } - } - }; - - // Compare elements to input offset - if(qi_in.scale >= 0) - { - p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); - } - else - { - p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); - } - - // Multiply negative elements and requantize if necessary - if(requant) - { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); - } - else - { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); - } - - // Convert uint32 vectors to uint16 vectors (with saturation) - const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); - const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); - - // convert uint16 vectors to uint8 vectors (with saturation) - tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - - svst1_u8(pg, output_ptr + x, tmp); - - x += svcntb(); - pg = svwhilelt_b8(x, window_end_x); - - } - while(svptest_any(svptrue_b8(), pg)); - - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute -#endif /* defined(__ARM_FEATURE_SVE2) */ \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/SVE/qasymm8_signed.cpp b/src/core/cpu/kernels/activation/SVE/qasymm8_signed.cpp deleted file mode 100644 index 989f825eb9..0000000000 --- a/src/core/cpu/kernels/activation/SVE/qasymm8_signed.cpp +++ /dev/null @@ -1,253 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/Window.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include -#include - -#if defined(__ARM_FEATURE_SVE2) -#include "src/core/NEON/SVEAsymm.h" -#include "src/core/NEON/SVEMath.h" -#include - -namespace arm_compute -{ -namespace cpu -{ -void qasymm8_signed_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const auto va = svdup_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in)); - const auto vb = svdup_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in)); - const auto const_0 = quantize_qasymm8_signed(0.f, qi_in); - const auto vconst_0 = svdup_n_s8(const_0); - const auto vconst_1 = svdup_n_f32(1.f); - const auto va_f32 = svdup_n_f32(act_info.a()); - const auto vb_f32 = svdup_n_f32(act_info.b()); - const auto const_6_f32 = svdup_n_f32(6.f); - const auto const_0_f32 = svdup_n_f32(0.f); - const auto const_3_f32 = svdup_n_f32(3.f); - const auto const_inv_6_f32 = svdup_n_f32(0.166666667f); - - // Initialise scale/offset for re-quantization - bool requant = true; - if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) - { - requant = false; - } - float s = qi_in.scale / qi_out.scale; - float o = -qi_in.offset * s + qi_out.offset; - auto vs = svdup_n_f32(s); - auto vo = svdup_n_f32(o); - - // Initialise scale/offset for re-quantization with int32_t - const auto voffset_in = svdup_n_s32(qi_in.offset); - int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - const auto vs_s32 = svdup_n_s32(s_s32); - const auto vo_s32 = svdup_n_s32(o_s32); - - // Initialise scale/offset for re-quantization for leaky relu - int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); - const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - svint8_t tmp; - - int x = window_start_x; - svbool_t pg = svwhilelt_b8(x, window_end_x); - do - { - const auto vin = svld1_s8(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = svmax_s8_z(pg, vconst_0, vin); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = - { - { { - svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))), - svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_signed_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - svbool_t p0, p1, p2, p3; - svint32x4_t tmp_dep; - - // Expand to int32 - const svint32x4_t vin_s32 = - { - { { - svmovlb_s32(svmovlb_s16(vin)), - svmovlt_s32(svmovlb_s16(vin)), - svmovlb_s32(svmovlt_s16(vin)), - svmovlt_s32(svmovlt_s16(vin)), - } - } - }; - - // Compare elements to input offset - if(qi_in.scale >= 0) - { - p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); - } - else - { - p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); - } - - // Multiply negative elements and requantize if necessary - if(requant) - { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); - } - else - { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); - } - - // Convert uint32 vectors to uint16 vectors (with saturation) - const auto v_low_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); - const auto v_high_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); - - // convert uint16 vectors to uint8 vectors (with saturation) - tmp = svqxtnt_s16(svqxtnb_s16(v_low_s16), v_high_s16); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - - svst1_s8(pg, output_ptr + x, tmp); - - x += svcntb(); - pg = svwhilelt_b8(x, window_end_x); - - } - while(svptest_any(svptrue_b8(), pg)); - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute -#endif /* defined(__ARM_FEATURE_SVE2) */ diff --git a/src/core/cpu/kernels/activation/SVE/qsymm16.cpp b/src/core/cpu/kernels/activation/SVE/qsymm16.cpp deleted file mode 100644 index 66974875da..0000000000 --- a/src/core/cpu/kernels/activation/SVE/qsymm16.cpp +++ /dev/null @@ -1,120 +0,0 @@ -/* - * Copyright (c) 2020-2021 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/Helpers.h" -#include "arm_compute/core/ITensorPack.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/experimental/Types.h" - -#include -#include - -#if defined(__ARM_FEATURE_SVE2) -#include "src/core/NEON/SVEMath.h" -#include "src/core/NEON/SVESymm.h" -#include - -namespace arm_compute -{ -namespace cpu -{ -void qsymm16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) -{ - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const ActivationLayerInfo::ActivationFunction act = act_info.activation(); - - Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); - win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(src, win_collapsed); - Iterator output(dst, win_collapsed); - - const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); - const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); - const auto vconst_1 = svdup_n_f32(1.f); - const auto va_f32 = svdup_n_f32(act_info.a()); - const auto vb_f32 = svdup_n_f32(act_info.b()); - - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - const auto output_ptr = reinterpret_cast(output.ptr()); - - svint16_t tmp; - - int x = window_start_x; - svbool_t pg = svwhilelt_b16(x, window_end_x); - do - { - const auto vin = svld1_s16(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); - // Perform activation - const svfloat32x2_t tmp_dep = - { - { { - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1))))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) - { - // De-quantize - auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); - // Perform activation - const svfloat32x2_t tmp_dep = - { - { { - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32))), - } - } - }; - // Re-quantize to new output space - tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - - svst1_s16(pg, output_ptr + x, tmp); - - x += svcnth(); - pg = svwhilelt_b16(x, window_end_x); - - } - while(svptest_any(svptrue_b16(), pg)); - }, - input, output); -} -} // namespace cpu -} // namespace arm_compute -#endif /* defined(__ARM_FEATURE_SVE2) */ diff --git a/src/core/cpu/kernels/activation/neon/fp16.cpp b/src/core/cpu/kernels/activation/neon/fp16.cpp new file mode 100644 index 0000000000..6f2d5d8533 --- /dev/null +++ b/src/core/cpu/kernels/activation/neon/fp16.cpp @@ -0,0 +1,217 @@ +/* + * Copyright (c) 2020-2021 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 "src/core/NEON/NEMath.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Validate.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include +#include + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +#ifndef __aarch64__ +inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &mask) +{ + auto int_in = vreinterpretq_u16_f16(in); + return vreinterpretq_f16_u16(wrapper::vand(int_in, mask)); +} +#endif /* __aarch64__ */ +} // namespace + +void fp16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + constexpr int window_step_x = 8; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + // In case of non-aarch64, a small delta value is added to the input + // to prevent NAN values caused by zeros in inputs to SQRT. + // In case of aarh64, we call vsqrt directly, so we don't use delta. +#ifndef __aarch64__ + const auto delta = wrapper::vdup_n(static_cast((1e-7), ExactTagType {})); +#endif /* __aarch64__ */ + + const auto const_1 = wrapper::vdup_n(static_cast(1.f), ExactTagType{}); + const auto const_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + const auto const_6 = wrapper::vdup_n(static_cast(6.f), ExactTagType{}); + const auto const_3 = wrapper::vdup_n(static_cast(3.f), ExactTagType{}); + const auto const_inv_6 = wrapper::vdup_n(static_cast(0.166666667f), ExactTagType{}); + + constexpr float soft_relu_thresh = 12.f; + const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast(soft_relu_thresh), ExactTagType{}); + + const auto va = wrapper::vdup_n(static_cast(act_info.a()), ExactTagType{}); + const auto vb = wrapper::vdup_n(static_cast(act_info.b()), ExactTagType{}); + const auto a = static_cast(act_info.a()); + const auto b = static_cast(act_info.b()); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + wrapper::traits::neon_bitvector_t tmp; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = wrapper::vabs(vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = wrapper::vmla(vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = wrapper::vmax(const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: +#ifdef __aarch64__ + tmp = wrapper::vsqrt(vin); +#else /* __aarch64__ */ + { + const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0, ExactTagType{})); + tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); + tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); + } +#endif /* __aarch64__ */ + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = wrapper::vmul(vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float16_t in = *(reinterpret_cast(input_ptr + x)); + float16_t tmp; + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = std::abs(in); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = a * in + b; + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = static_cast(1) / (static_cast(1) + std::exp(-in)); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = std::max(static_cast(0), in); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = std::min(a, std::max(static_cast(0), in)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = std::min(a, std::max(b, in)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = (in > 0) ? in : a * in; + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = (in > soft_relu_thresh) ? in : std::log(static_cast(1) + std::exp(in)); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = (in >= 0) ? in : a * (std::exp(in) - 1); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = std::sqrt(in); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = in * in; + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = a * std::tanh(b * in); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = in; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute + +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/core/cpu/kernels/activation/neon/fp32.cpp b/src/core/cpu/kernels/activation/neon/fp32.cpp new file mode 100644 index 0000000000..54301d45ad --- /dev/null +++ b/src/core/cpu/kernels/activation/neon/fp32.cpp @@ -0,0 +1,212 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +#ifndef __aarch64__ +inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask) +{ + auto int_in = vreinterpretq_u32_f32(in); + return vreinterpretq_f32_u32(wrapper::vand(int_in, mask)); +} +#endif /* __aarch64__ */ +} // namespace + +void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t; + + constexpr int window_step_x = 4; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + // In case of non-aarch64, a small delta value is added to the input + // to prevent NAN values caused by zeros in inputs to SQRT. + // In case of aarh64, we call vsqrt directly, so we don't use delta. +#ifndef __aarch64__ + const auto delta = wrapper::vdup_n(static_cast(1e-24), ExactTagType {}); +#endif /* __aarch64__ */ + const auto const_1 = wrapper::vdup_n(static_cast(1.f), ExactTagType {}); + const auto const_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + const auto const_6 = wrapper::vdup_n(static_cast(6.f), ExactTagType{}); + const auto const_3 = wrapper::vdup_n(static_cast(3.f), ExactTagType{}); + const auto const_inv_6 = wrapper::vdup_n(static_cast(0.166666667f), ExactTagType{}); + + constexpr float soft_relu_thresh = 12.f; + const auto vsoft_relu_thresh = wrapper::vdup_n(static_cast(soft_relu_thresh), ExactTagType{}); + + const auto va = wrapper::vdup_n(static_cast(act_info.a()), ExactTagType{}); + const auto vb = wrapper::vdup_n(static_cast(act_info.b()), ExactTagType{}); + const auto a = static_cast(act_info.a()); + const auto b = static_cast(act_info.b()); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + wrapper::traits::neon_bitvector_t tmp; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = wrapper::vabs(vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = wrapper::vmla(vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = wrapper::vmax(const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = wrapper::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: +#ifdef __aarch64__ + tmp = wrapper::vsqrt(vin); +#else /* __aarch64__ */ + { + const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{})); + tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); + tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); + } +#endif /* __aarch64__ */ + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = wrapper::vmul(vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + const float in = *(reinterpret_cast(input_ptr + x)); + float tmp; + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = std::abs(in); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = a * in + b; + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = static_cast(1) / (static_cast(1) + std::exp(-in)); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = std::max(static_cast(0), in); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = std::min(a, std::max(static_cast(0), in)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = std::min(a, std::max(b, in)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = (in > 0) ? in : a * in; + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = (in > soft_relu_thresh) ? in : std::log(static_cast(1) + std::exp(in)); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = (in >= 0) ? in : a * (std::exp(in) - 1); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = std::sqrt(in); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = in * in; + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = a * std::tanh(b * in); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = in; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/neon/qasymm8.cpp b/src/core/cpu/kernels/activation/neon/qasymm8.cpp new file mode 100644 index 0000000000..a1217435b6 --- /dev/null +++ b/src/core/cpu/kernels/activation/neon/qasymm8.cpp @@ -0,0 +1,262 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/Window.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +void qasymm8_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + constexpr int window_step_x = 16; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const qasymm8x16_t va = vdupq_n_u8(quantize_qasymm8(act_info.a(), qi_in)); + const qasymm8x16_t vb = vdupq_n_u8(quantize_qasymm8(act_info.b(), qi_in)); + const qasymm8_t a = quantize_qasymm8(act_info.a(), qi_in); + const qasymm8_t b = quantize_qasymm8(act_info.b(), qi_in); + const qasymm8_t const_0 = quantize_qasymm8(0.f, qi_in); + const qasymm8x16_t vconst_0 = vdupq_n_u8(const_0); + const auto vconst_1 = vdupq_n_f32(1.f); +#ifndef __aarch64__ + const auto vconst_0_f32 = vdupq_n_f32(0); +#endif // __aarch64__ + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + const auto const_6_f32 = vdupq_n_f32(6.f); + const auto const_0_f32 = vdupq_n_f32(0.f); + const auto const_3_f32 = vdupq_n_f32(3.f); + const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); + + // Initialise scale/offset for re-quantization + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + float32x4_t vs = vdupq_n_f32(s); + float32x4_t vo = vdupq_n_f32(o); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + wrapper::traits::neon_bitvector_t tmp; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = vmaxq_u8(vconst_0, vin); + // Re-quantize to new output space + tmp = vmlaq_qasymm8(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = vminq_u8(va, vmaxq_u8(vconst_0, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = vminq_u8(va, vmaxq_u8(vb, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), + } + }; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), + } + }; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), + wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), + wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), + wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), + } + }; + // Re-quantize to new output space + tmp = vquantize(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + const auto vin_deq = vdequantize(vin, qi_in); + +#ifdef __aarch64__ + const uint32x4x4_t pos_mask = + { + { + wrapper::vcgtz(vin_deq.val[0]), + wrapper::vcgtz(vin_deq.val[1]), + wrapper::vcgtz(vin_deq.val[2]), + wrapper::vcgtz(vin_deq.val[3]), + } + }; +#else // __aarch64__ + const uint32x4x4_t pos_mask = + { + { + wrapper::vcgt(vin_deq.val[0], vconst_0_f32), + wrapper::vcgt(vin_deq.val[1], vconst_0_f32), + wrapper::vcgt(vin_deq.val[2], vconst_0_f32), + wrapper::vcgt(vin_deq.val[3], vconst_0_f32), + } + }; +#endif // __aarch64__ + + const float32x4x4_t tmp_dep = + { + { + wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), + wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), + wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), + wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), + } + }; + + tmp = vquantize(tmp_dep, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + qasymm8_t in = *(reinterpret_cast(input_ptr + x)); + qasymm8_t tmp = 0; + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + tmp = std::max(const_0, in); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + tmp = std::min(a, std::max(const_0, in)); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + tmp = std::min(a, std::max(b, in)); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + float tmp_f = dequantize_qasymm8(in, qi_in); + tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; + tmp = quantize_qasymm8(tmp_f, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/neon/qasymm8_signed.cpp b/src/core/cpu/kernels/activation/neon/qasymm8_signed.cpp new file mode 100644 index 0000000000..8b40bf8e72 --- /dev/null +++ b/src/core/cpu/kernels/activation/neon/qasymm8_signed.cpp @@ -0,0 +1,261 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/Window.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +void qasymm8_signed_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + constexpr int window_step_x = 16; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const qasymm8x16_signed_t va = vdupq_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in)); + const qasymm8x16_signed_t vb = vdupq_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in)); + const qasymm8_signed_t a = quantize_qasymm8_signed(act_info.a(), qi_in); + const qasymm8_signed_t b = quantize_qasymm8_signed(act_info.b(), qi_in); + const qasymm8_signed_t const_0 = quantize_qasymm8_signed(0.f, qi_in); + const qasymm8x16_signed_t vconst_0 = vdupq_n_s8(const_0); + const auto vconst_1 = vdupq_n_f32(1.f); +#ifndef __aarch64__ + const auto vconst_0_f32 = vdupq_n_f32(1.f); +#endif // __aarch64__ + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + const auto const_6_f32 = vdupq_n_f32(6.f); + const auto const_0_f32 = vdupq_n_f32(0.f); + const auto const_3_f32 = vdupq_n_f32(3.f); + const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f); + + // Initialise scale/offset for re-quantization + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + float32x4_t vs = vdupq_n_f32(s); + float32x4_t vo = vdupq_n_f32(o); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + wrapper::traits::neon_bitvector_t tmp; + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = vmaxq_s8(vconst_0, vin); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = vminq_s8(va, vmaxq_s8(vconst_0, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = vminq_s8(va, vmaxq_s8(vb, vin)); + // Re-quantize to new output space + tmp = vmlaq_qasymm8_signed(tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))), + } + }; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))), + } + }; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = vdequantize(vin, qi_in); + // Perform activation + const float32x4x4_t tmp_dep = + { + { + wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))), + wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))), + wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))), + wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))), + } + }; + // Re-quantize to new output space + tmp = vquantize_signed(tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + const auto vin_deq = vdequantize(vin, qi_in); + +#ifdef __aarch64__ + const uint32x4x4_t pos_mask = + { + { + wrapper::vcgtz(vin_deq.val[0]), + wrapper::vcgtz(vin_deq.val[1]), + wrapper::vcgtz(vin_deq.val[2]), + wrapper::vcgtz(vin_deq.val[3]), + } + }; +#else // __aarch64__ + const uint32x4x4_t pos_mask = + { + { + wrapper::vcgt(vin_deq.val[0], vconst_0_f32), + wrapper::vcgt(vin_deq.val[1], vconst_0_f32), + wrapper::vcgt(vin_deq.val[2], vconst_0_f32), + wrapper::vcgt(vin_deq.val[3], vconst_0_f32), + } + }; +#endif // __aarch64__ + + const float32x4x4_t tmp_dep = + { + { + wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])), + wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])), + wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])), + wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])), + } + }; + + tmp = vquantize_signed(tmp_dep, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + qasymm8_signed_t in = *(reinterpret_cast(input_ptr + x)); + qasymm8_signed_t tmp = 0; + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + tmp = std::max(const_0, in); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + tmp = std::min(a, std::max(const_0, in)); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + tmp = std::min(a, std::max(b, in)); + tmp = utility::clamp(tmp * s + o); + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f); + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + float tmp_f = dequantize_qasymm8_signed(in, qi_in); + tmp_f = tmp_f > 0 ? tmp_f : tmp_f * a_f32; + tmp = quantize_qasymm8_signed(tmp_f, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/neon/qsymm16.cpp b/src/core/cpu/kernels/activation/neon/qsymm16.cpp new file mode 100644 index 0000000000..54b41820f2 --- /dev/null +++ b/src/core/cpu/kernels/activation/neon/qsymm16.cpp @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/experimental/Types.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/NESymm.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +void qsymm16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + constexpr int window_step_x = 8; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const auto vconst_1 = vdupq_n_f32(1.f); + const float32x4_t va_f32 = vdupq_n_f32(act_info.a()); + const float32x4_t vb_f32 = vdupq_n_f32(act_info.b()); + const float a_f32 = act_info.a(); + const float b_f32 = act_info.b(); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + wrapper::traits::neon_bitvector_t tmp; + ARM_COMPUTE_UNUSED(tmp); + + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin = wrapper::vloadq(input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = vdequantize_int16(vin, qi_in.scale); + // Perform activation + const float32x4x2_t tmp_dep = + { + { + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))), + wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))), + } + }; + // Re-quantize to new output space + tmp = vquantize_int16(tmp_dep, qi_out.scale); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = vdequantize_int16(vin, qi_in.scale); + // Perform activation + const float32x4x2_t tmp_dep = + { + { + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))), + wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))), + } + }; + // Re-quantize to new output space + tmp = vquantize_int16(tmp_dep, qi_out.scale); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + wrapper::vstore(output_ptr + x, tmp); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + qsymm16_t in = *(reinterpret_cast(input_ptr + x)); + qsymm16_t tmp = 0; + if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + float tmp_f = dequantize_qsymm16(in, qi_in.scale); + tmp_f = 1.f / (1.f + std::exp(-tmp_f)); + tmp = quantize_qsymm16(tmp_f, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + float tmp_f = dequantize_qsymm16(in, qi_in.scale); + tmp_f = a_f32 * std::tanh(b_f32 * tmp_f); + tmp = quantize_qsymm16(tmp_f, qi_out); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + *(output_ptr + x) = tmp; + } + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/activation/sve/fp16.cpp b/src/core/cpu/kernels/activation/sve/fp16.cpp new file mode 100644 index 0000000000..bf31fd7d93 --- /dev/null +++ b/src/core/cpu/kernels/activation/sve/fp16.cpp @@ -0,0 +1,130 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" + +#include +#include + +#if defined(__ARM_FEATURE_SVE) +#include "src/core/NEON/SVEMath.h" +#include + +namespace arm_compute +{ +namespace cpu +{ +void fp16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto const_1 = svdup_n_f16(1.f); + const auto const_0 = svdup_n_f16(0.f); + const auto const_6 = svdup_n_f16(6.f); + const auto const_3 = svdup_n_f16(3.f); + const auto const_inv_6 = svdup_n_f16(0.166666667f); + + const auto va = svdup_n_f16(act_info.a()); + const auto vb = svdup_n_f16(act_info.b()); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + svfloat16_t tmp; + + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do + { + const auto vin = svld1_f16(pg, input_ptr + x); + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = svabs_f16_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = svmla_f16_z(pg, vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = svmax_f16_z(pg, const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = svadd_f16_z(pg, svmul_f16_z(pg, svmin_f16_z(pg, vin, const_0), va), svmax_f16_z(pg, vin, const_0)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = svlog_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, vin))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = svsel_f16(svcmpgt_f16(pg, vin, const_0), vin, svmul_f16_z(pg, va, svsub_f16_z(pg, svexp_f16_z(pg, vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = svsqrt_f16_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = svmul_f16_z(pg, vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = svmul_f16_z(pg, va, svtanh_f16_z(pg, svmul_f16_z(pg, vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = svmul_f16_z(pg, vin, svmul_f16_z(pg, const_inv_6, svmin_f16_z(pg, const_6, svmax_f16_z(pg, const_0, svadd_f16_z(pg, vin, const_3))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + svst1_f16(pg, output_ptr + x, tmp); + + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); + + } + while(svptest_any(svptrue_b16(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif // __ARM_FEATURE_SVE \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/sve/fp32.cpp b/src/core/cpu/kernels/activation/sve/fp32.cpp new file mode 100644 index 0000000000..75f9f8a4c3 --- /dev/null +++ b/src/core/cpu/kernels/activation/sve/fp32.cpp @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "src/core/NEON/SVEMath.h" + +#include +#include + +#if defined(__ARM_FEATURE_SVE) +#include + +namespace arm_compute +{ +namespace cpu +{ +void fp32_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto const_1 = svdup_n_f32(1.f); + const auto const_0 = svdup_n_f32(0.f); + const auto const_6 = svdup_n_f32(6.f); + const auto const_3 = svdup_n_f32(3.f); + const auto const_inv_6 = svdup_n_f32(0.166666667f); + + const auto va = svdup_n_f32(act_info.a()); + const auto vb = svdup_n_f32(act_info.b()); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + svfloat32_t tmp; + + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b32(x, window_end_x); + do + { + const auto vin = svld1_f32(pg, input_ptr + x); + switch(act) + { + case ActivationLayerInfo::ActivationFunction::ABS: + tmp = svabs_f32_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::LINEAR: + tmp = svmla_f32_z(pg, vb, va, vin); + break; + case ActivationLayerInfo::ActivationFunction::LOGISTIC: + tmp = svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, vin)))); + break; + case ActivationLayerInfo::ActivationFunction::RELU: + tmp = svmax_f32_z(pg, const_0, vin); + break; + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: + tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, vin)); + break; + case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: + tmp = svadd_f32_z(pg, svmul_f32_z(pg, svmin_f32_z(pg, vin, const_0), va), svmax_f32_z(pg, vin, const_0)); + break; + case ActivationLayerInfo::ActivationFunction::SOFT_RELU: + tmp = svlog_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, vin))); + break; + case ActivationLayerInfo::ActivationFunction::ELU: + tmp = svsel_f32(svcmpgt_f32(pg, vin, const_0), vin, svmul_f32_z(pg, va, svsub_f32_z(pg, svexp_f32_z(pg, vin), const_1))); + break; + case ActivationLayerInfo::ActivationFunction::SQRT: + tmp = svsqrt_f32_z(pg, vin); + break; + case ActivationLayerInfo::ActivationFunction::SQUARE: + tmp = svmul_f32_z(pg, vin, vin); + break; + case ActivationLayerInfo::ActivationFunction::TANH: + tmp = svmul_f32_z(pg, va, svtanh_f32_z(pg, svmul_f32_z(pg, vb, vin))); + break; + case ActivationLayerInfo::ActivationFunction::IDENTITY: + tmp = vin; + break; + case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + tmp = svmul_f32_z(pg, vin, svmul_f32_z(pg, const_inv_6, svmin_f32_z(pg, const_6, svmax_f32_z(pg, const_0, svadd_f32_z(pg, vin, const_3))))); + break; + default: + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + svst1_f32(pg, output_ptr + x, tmp); + + x += svcntw(); + pg = svwhilelt_b32(x, window_end_x); + + } + while(svptest_any(svptrue_b32(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif // __ARM_FEATURE_SVE \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/sve/qasymm8.cpp b/src/core/cpu/kernels/activation/sve/qasymm8.cpp new file mode 100644 index 0000000000..228b4ae530 --- /dev/null +++ b/src/core/cpu/kernels/activation/sve/qasymm8.cpp @@ -0,0 +1,254 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/Window.h" + +#include +#include + +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEAsymm.h" +#include "src/core/NEON/SVEMath.h" +#include + +namespace arm_compute +{ +namespace cpu +{ +void qasymm8_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const auto va = svdup_n_u8(quantize_qasymm8(act_info.a(), qi_in)); + const auto vb = svdup_n_u8(quantize_qasymm8(act_info.b(), qi_in)); + const auto const_0 = quantize_qasymm8(0.f, qi_in); + const auto vconst_0 = svdup_n_u8(const_0); + const auto vconst_1 = svdup_n_f32(1.f); + const auto va_f32 = svdup_n_f32(act_info.a()); + const auto vb_f32 = svdup_n_f32(act_info.b()); + const auto const_6_f32 = svdup_n_f32(6.f); + const auto const_0_f32 = svdup_n_f32(0.f); + const auto const_3_f32 = svdup_n_f32(3.f); + const auto const_inv_6_f32 = svdup_n_f32(0.166666667f); + + // Initialise scale/offset for re-quantization + bool requant = true; + if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) + { + requant = false; + } + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + auto vs = svdup_n_f32(s); + auto vo = svdup_n_f32(o); + + // Initialise scale/offset for re-quantization with int32_t + const auto voffset_in = svdup_n_s32(qi_in.offset); + int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + const auto vs_s32 = svdup_n_s32(s_s32); + const auto vo_s32 = svdup_n_s32(o_s32); + + // Initialise scale/offset for re-quantization for leaky relu + int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), + arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); + const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + svuint8_t tmp; + + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do + { + const auto vin = svld1_u8(pg, input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = svmax_u8_z(pg, vconst_0, vin); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); + // Re-quantize to new output space + tmp = svmla_qasymm8_z(pg, tmp, vs, vo); + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + svbool_t p0, p1, p2, p3; + svint32x4_t tmp_dep; + + // Expand to int32 + const svint32x4_t vin_s32 = + { + { { + svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin))), + } + } + }; + + // Compare elements to input offset + if(qi_in.scale >= 0) + { + p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + else + { + p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + + // Multiply negative elements and requantize if necessary + if(requant) + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); + } + else + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + } + + // Convert uint32 vectors to uint16 vectors (with saturation) + const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); + const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + + // convert uint16 vectors to uint8 vectors (with saturation) + tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + + svst1_u8(pg, output_ptr + x, tmp); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + + } + while(svptest_any(svptrue_b8(), pg)); + + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */ \ No newline at end of file diff --git a/src/core/cpu/kernels/activation/sve/qasymm8_signed.cpp b/src/core/cpu/kernels/activation/sve/qasymm8_signed.cpp new file mode 100644 index 0000000000..989f825eb9 --- /dev/null +++ b/src/core/cpu/kernels/activation/sve/qasymm8_signed.cpp @@ -0,0 +1,253 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/Window.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include +#include + +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEAsymm.h" +#include "src/core/NEON/SVEMath.h" +#include + +namespace arm_compute +{ +namespace cpu +{ +void qasymm8_signed_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const auto va = svdup_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in)); + const auto vb = svdup_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in)); + const auto const_0 = quantize_qasymm8_signed(0.f, qi_in); + const auto vconst_0 = svdup_n_s8(const_0); + const auto vconst_1 = svdup_n_f32(1.f); + const auto va_f32 = svdup_n_f32(act_info.a()); + const auto vb_f32 = svdup_n_f32(act_info.b()); + const auto const_6_f32 = svdup_n_f32(6.f); + const auto const_0_f32 = svdup_n_f32(0.f); + const auto const_3_f32 = svdup_n_f32(3.f); + const auto const_inv_6_f32 = svdup_n_f32(0.166666667f); + + // Initialise scale/offset for re-quantization + bool requant = true; + if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) + { + requant = false; + } + float s = qi_in.scale / qi_out.scale; + float o = -qi_in.offset * s + qi_out.offset; + auto vs = svdup_n_f32(s); + auto vo = svdup_n_f32(o); + + // Initialise scale/offset for re-quantization with int32_t + const auto voffset_in = svdup_n_s32(qi_in.offset); + int32_t s_s32 = round(s * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_s32 = round(o * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + const auto vs_s32 = svdup_n_s32(s_s32); + const auto vo_s32 = svdup_n_s32(o_s32); + + // Initialise scale/offset for re-quantization for leaky relu + int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), + arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); + const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + svint8_t tmp; + + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do + { + const auto vin = svld1_s8(pg, input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::RELU) + { + // Perform activation + tmp = svmax_s8_z(pg, vconst_0, vin); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; + } + else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + // Perform activation + tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; + } + else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp; + } + else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = + { + { { + svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))), + svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_signed_z(pg, tmp_dep, qi_out); + } + else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + svbool_t p0, p1, p2, p3; + svint32x4_t tmp_dep; + + // Expand to int32 + const svint32x4_t vin_s32 = + { + { { + svmovlb_s32(svmovlb_s16(vin)), + svmovlt_s32(svmovlb_s16(vin)), + svmovlb_s32(svmovlt_s16(vin)), + svmovlt_s32(svmovlt_s16(vin)), + } + } + }; + + // Compare elements to input offset + if(qi_in.scale >= 0) + { + p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + else + { + p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + + // Multiply negative elements and requantize if necessary + if(requant) + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), + svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); + } + else + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + } + + // Convert uint32 vectors to uint16 vectors (with saturation) + const auto v_low_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); + const auto v_high_s16 = svqxtnt_s32(svqxtnb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + + // convert uint16 vectors to uint8 vectors (with saturation) + tmp = svqxtnt_s16(svqxtnb_s16(v_low_s16), v_high_s16); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + + svst1_s8(pg, output_ptr + x, tmp); + + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); + + } + while(svptest_any(svptrue_b8(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */ diff --git a/src/core/cpu/kernels/activation/sve/qsymm16.cpp b/src/core/cpu/kernels/activation/sve/qsymm16.cpp new file mode 100644 index 0000000000..66974875da --- /dev/null +++ b/src/core/cpu/kernels/activation/sve/qsymm16.cpp @@ -0,0 +1,120 @@ +/* + * Copyright (c) 2020-2021 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/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/experimental/Types.h" + +#include +#include + +#if defined(__ARM_FEATURE_SVE2) +#include "src/core/NEON/SVEMath.h" +#include "src/core/NEON/SVESymm.h" +#include + +namespace arm_compute +{ +namespace cpu +{ +void qsymm16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const ActivationLayerInfo::ActivationFunction act = act_info.activation(); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const UniformQuantizationInfo qi_in = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo qi_out = dst->info()->quantization_info().uniform(); + const auto vconst_1 = svdup_n_f32(1.f); + const auto va_f32 = svdup_n_f32(act_info.a()); + const auto vb_f32 = svdup_n_f32(act_info.b()); + + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + svint16_t tmp; + + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do + { + const auto vin = svld1_s16(pg, input_ptr + x); + if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); + // Perform activation + const svfloat32x2_t tmp_dep = + { + { { + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))), + svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1))))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); + } + else if(act == ActivationLayerInfo::ActivationFunction::TANH) + { + // De-quantize + auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale); + // Perform activation + const svfloat32x2_t tmp_dep = + { + { { + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32))), + } + } + }; + // Re-quantize to new output space + tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale); + } + else + { + ARM_COMPUTE_ERROR("Unsupported activation function"); + } + + svst1_s16(pg, output_ptr + x, tmp); + + x += svcnth(); + pg = svwhilelt_b16(x, window_end_x); + + } + while(svptest_any(svptrue_b16(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_SVE2) */ -- cgit v1.2.1