From 0ded4c40578bc78003756d171f2bbe15f6ac72bc Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Tue, 9 Mar 2021 14:15:27 +0000 Subject: Port Arm(R) Neon(TM) Quantization to new API Partially resolves: COMPMID-4193 Change-Id: I91dc964d4308687e76127c305a6bedca796f8ba0 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5246 Reviewed-by: Michalis Spyrou Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- src/core/cpu/kernels/CpuQuantizationKernel.cpp | 271 +++++++++++++++++++++++++ 1 file changed, 271 insertions(+) create mode 100644 src/core/cpu/kernels/CpuQuantizationKernel.cpp (limited to 'src/core/cpu/kernels/CpuQuantizationKernel.cpp') diff --git a/src/core/cpu/kernels/CpuQuantizationKernel.cpp b/src/core/cpu/kernels/CpuQuantizationKernel.cpp new file mode 100644 index 0000000000..9b1e017275 --- /dev/null +++ b/src/core/cpu/kernels/CpuQuantizationKernel.cpp @@ -0,0 +1,271 @@ +/* + * 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/CpuQuantizationKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.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 "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" + +#include "src/core/CPP/Validate.h" + +#include +#include + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +namespace +{ +constexpr auto window_step = 16; + +Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QASYMM16); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); + + return Status{}; +} + +template +inline float32x4x4_t load_value(const T *input_ptr) +{ + using Tx16_t = typename wrapper::traits::neon_vector::type; + return arm_compute::convert_to_float32x4x4(wrapper::vloadq(input_ptr)); +} + +template <> +inline float32x4x4_t load_value(const float *input_ptr) +{ + return { wrapper::vloadq(input_ptr), + wrapper::vloadq(input_ptr + 4), + wrapper::vloadq(input_ptr + 8), + wrapper::vloadq(input_ptr + 12) }; +} +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template <> +inline float32x4x4_t load_value(const float16_t *input_ptr) +{ + return { vcvt_f32_f16(wrapper::vload(input_ptr)), + vcvt_f32_f16(wrapper::vload(input_ptr + 4)), + vcvt_f32_f16(wrapper::vload(input_ptr + 8)), + vcvt_f32_f16(wrapper::vload(input_ptr + 12)) }; +} + +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +template +using vector_type = wrapper::traits::neon_vector_t; + +template +vector_type vquantize_qasymm8(const float32x4x4_t &qv, const UniformQuantizationInfo &qi); + +template <> +vector_type vquantize_qasymm8(const float32x4x4_t &qv, const UniformQuantizationInfo &qi) +{ + return vquantize(qv, qi); +} + +template <> +vector_type vquantize_qasymm8(const float32x4x4_t &qv, const UniformQuantizationInfo &qi) +{ + return vquantize_signed(qv, qi); +} + +} // namespace + +CpuQuantizationKernel::CpuQuantizationKernel() + : _func(nullptr) +{ +} + +void CpuQuantizationKernel::configure(ITensorInfo *src, ITensorInfo *dst) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst)); + + static const std::map quant_map = + { + { "op_QASYMM8_QASYMM8", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_QASYMM8_QASYMM8_SIGNED", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_QASYMM8_QASYMM16", &CpuQuantizationKernel::run_quantize_qasymm16 }, + + { "op_QASYMM8_SIGNED_QASYMM8", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_QASYMM8_SIGNED_QASYMM8_SIGNED", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_QASYMM8_SIGNED_QASYMM16", &CpuQuantizationKernel::run_quantize_qasymm16 }, + + { "op_F32_QASYMM8", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_F32_QASYMM8_SIGNED", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_F32_QASYMM16", &CpuQuantizationKernel::run_quantize_qasymm16 }, + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + { "op_F16_QASYMM8", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_F16_QASYMM8_SIGNED", &CpuQuantizationKernel::run_quantize_qasymm8 }, + { "op_F16_QASYMM16", &CpuQuantizationKernel::run_quantize_qasymm16 }, +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/ + }; + + std::string function_to_call("op_"); + function_to_call += string_from_data_type(src->data_type()) + "_"; + function_to_call += string_from_data_type(dst->data_type()); + + auto it = quant_map.find(function_to_call); + + if(it == quant_map.end()) + { + ARM_COMPUTE_ERROR("Unsupported combination of input and output data types"); + } + _func = it->second; + + // Configure kernel window + Window win_config = calculate_max_window(*src, Steps()); + ICpuKernel::configure(win_config); +} + +Status CpuQuantizationKernel::validate(const ITensorInfo *src, const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst)); + return Status{}; +} + +template +void CpuQuantizationKernel::run_quantize_qasymm8(const ITensor *src, ITensor *dst, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + const UniformQuantizationInfo uqinfo_in = src->info()->quantization_info().uniform(); + UniformQuantizationInfo uqinfo = dst->info()->quantization_info().uniform(); + if(is_data_type_quantized_asymmetric(src->info()->data_type())) + { + uqinfo = compute_requantization_scale_offset(uqinfo_in, uqinfo); + } +#ifdef __aarch64__ + constexpr RoundingPolicy rounding_policy = RoundingPolicy::TO_NEAREST_EVEN; +#else //__aarch64__ + constexpr RoundingPolicy rounding_policy = RoundingPolicy::TO_ZERO; +#endif //__aarch64__ + + // Collapse window and reset first dimension to handle tail calculations manually + 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); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + auto input_ptr = reinterpret_cast(input.ptr()); + auto output_ptr = reinterpret_cast(output.ptr()); + + int x = window_start_x; + for(; x <= (window_end_x - window_step); x += window_step) + { + wrapper::vstore(&output_ptr[x], vquantize_qasymm8(load_value(&input_ptr[x]), uqinfo)); + } + // Compute left-over elements + for(; x < window_end_x; ++x) + { + output_ptr[x] = Qasymm8QuantizationHelper::quantize(input_ptr[x], uqinfo, rounding_policy); + } + }, + input, output); +} + +template +void CpuQuantizationKernel::run_quantize_qasymm16(const ITensor *src, ITensor *dst, const Window &window) +{ + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + const UniformQuantizationInfo uqinfo_in = src->info()->quantization_info().uniform(); + UniformQuantizationInfo uqinfo = dst->info()->quantization_info().uniform(); + if(is_data_type_quantized_asymmetric(src->info()->data_type())) + { + uqinfo = compute_requantization_scale_offset(uqinfo_in, uqinfo); + } +#ifdef __aarch64__ + constexpr RoundingPolicy rounding_policy = RoundingPolicy::TO_NEAREST_EVEN; +#else //__aarch64__ + constexpr RoundingPolicy rounding_policy = RoundingPolicy::TO_ZERO; +#endif //__aarch64__ + + // Collapse window and reset first dimension to handle tail calculations manually + 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); + execute_window_loop(win_collapsed, [&](const Coordinates &) + { + auto input_ptr = reinterpret_cast(input.ptr()); + auto output_ptr = reinterpret_cast(output.ptr()); + + int x = window_start_x; + for(; x <= (window_end_x - window_step); x += window_step) + { + uint16x8x2_t tmp = vquantize_qasymm16(load_value(&input_ptr[x]), uqinfo); + vst1q_u16(&output_ptr[x], tmp.val[0]); + vst1q_u16(&output_ptr[x + 8], tmp.val[1]); + } + // Compute left-over elements + for(; x < window_end_x; ++x) + { + output_ptr[x] = quantize_qasymm16(input_ptr[x], uqinfo, rounding_policy); + } + }, + input, output); +} + +void CpuQuantizationKernel::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); + + const auto src = tensors.get_const_tensor(TensorType::ACL_SRC); + auto dst = tensors.get_tensor(TensorType::ACL_DST); + (this->*_func)(src, dst, window); +} + +const char *CpuQuantizationKernel::name() const +{ + return "CpuQuantizationKernel"; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute \ No newline at end of file -- cgit v1.2.1