diff options
Diffstat (limited to 'src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp')
-rw-r--r-- | src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp | 169 |
1 files changed, 64 insertions, 105 deletions
diff --git a/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp b/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp index 6a41e3a161..451031d696 100644 --- a/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp +++ b/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,26 +28,74 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Window.h" + +#include "src/core/common/Registrars.h" #include "src/core/CPP/Validate.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/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/meanstddevnorm/list.h" namespace arm_compute { namespace { +struct MeanStdDevNormSelectorData +{ + DataType dt; +}; + +using MeanStdDevNormSelctorPtr = std::add_pointer<bool(const MeanStdDevNormSelectorData &data)>::type; +using MeanStdDevNormUKernelPtr = + std::add_pointer<void(ITensor *input, ITensor *output, float epsilon, const Window &window)>::type; + +struct MeanStdDevNormKernel +{ + const char *name; + const MeanStdDevNormSelctorPtr is_selected; + MeanStdDevNormUKernelPtr ukernel; +}; + +static const std::vector<MeanStdDevNormKernel> available_kernels = { + {"fp32_neon_meanstddevnorm", [](const MeanStdDevNormSelectorData &data) { return data.dt == DataType::F32; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_meanstddevnorm)}, +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + {"fp16_neon_meanstddevnorm", [](const MeanStdDevNormSelectorData &data) { return data.dt == DataType::F16; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_meanstddevnorm)}, +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + {"qasymm8_neon_meanstddevnorm", [](const MeanStdDevNormSelectorData &data) { return data.dt == DataType::QASYMM8; }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_meanstddevnorm)}, +}; + +/** Micro-kernel selector + * + * @param[in] data Selection data passed to help pick the appropriate micro-kernel + * + * @return A matching micro-kernel else nullptr + */ +const MeanStdDevNormKernel *get_implementation(const MeanStdDevNormSelectorData &data) +{ + for (const auto &uk : available_kernels) + { + if (uk.is_selected(data)) + { + return &uk; + } + } + return nullptr; +} + Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, float epsilon) { ARM_COMPUTE_UNUSED(epsilon); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8); // Checks performed when output is configured - if((output != nullptr) && (output->total_size() != 0)) + if ((output != nullptr) && (output->total_size() != 0)) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); @@ -57,7 +105,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, f std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) { - if(output != nullptr) + if (output != nullptr) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output auto inizialitation if not yet initialized @@ -67,89 +115,12 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen // This kernel doesn't need padding. A left-over for loop on dimension X, we cannot have any read or write out of memory // For this reason num_elems_processed_per_iteration is set to 1 Window win = calculate_max_window(*input, Steps()); - if(output != nullptr) - { - output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); - } return std::make_pair(Status{}, win); } } // namespace -template <typename ScalarType, int size> -void NEMeanStdDevNormalizationKernel::mean_stddev_normalization(const Window &window) -{ - using ExactTagType = typename wrapper::traits::neon_vector<ScalarType, size>::tag_type; - - // Set build options - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - const int window_step_x = size; - const auto window_start_x = static_cast<int>(window.x().start()); - const auto window_end_x = static_cast<int>(window.x().end()); - - Iterator input(_input, win); - Iterator output(_output, win); - - execute_window_loop(win, [&](const Coordinates &) - { - int x = window_start_x; - auto in_ptr = reinterpret_cast<const ScalarType *>(input.ptr()); - auto out_ptr = reinterpret_cast<ScalarType *>(output.ptr()); - - auto sum_vec = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{}); - auto sum_sq_vec = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{}); - - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - auto data = wrapper::vloadq(in_ptr + x); - sum_vec = wrapper::vadd(sum_vec, data); - sum_sq_vec = wrapper::vadd(sum_sq_vec, wrapper::vmul(data, data)); - } - - auto sum_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_vec), wrapper::vgetlow(sum_vec)); - auto sum_sq_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_sq_vec), wrapper::vgetlow(sum_sq_vec)); - for(int i = 0; i < size / 4; ++i) - { - sum_carry_res = wrapper::vpadd(sum_carry_res, sum_carry_res); - sum_sq_carry_res = wrapper::vpadd(sum_sq_carry_res, sum_sq_carry_res); - } - - auto sum = wrapper::vgetlane(sum_carry_res, 0); - auto sum_sq = wrapper::vgetlane(sum_sq_carry_res, 0); - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - ScalarType data = *(in_ptr + x); - sum += data; - sum_sq += data * data; - } - - ScalarType mean = sum / _input->info()->dimension(0); - ScalarType var = (sum_sq / _input->info()->dimension(0)) - (mean * mean); - ScalarType stddev_inv = 1.f / sqrt(var + _epsilon); - - auto mean_vec = wrapper::vdup_n(mean, ExactTagType{}); - auto stddev_inv_vec = wrapper::vdup_n(stddev_inv, ExactTagType{}); - for(x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x) - { - auto data = wrapper::vloadq(in_ptr + x); - auto res = wrapper::vmul(wrapper::vsub(data, mean_vec), stddev_inv_vec); - // Store results - wrapper::vstore(out_ptr + x, res); - } - for(; x < window_end_x; ++x) - { - *(out_ptr + x) = (*(in_ptr + x) - mean) * stddev_inv; - } - }, - input, output); -} - -NEMeanStdDevNormalizationKernel::NEMeanStdDevNormalizationKernel() - : _input(nullptr), _output(nullptr), _epsilon(1e-8f), _func(nullptr) +NEMeanStdDevNormalizationKernel::NEMeanStdDevNormalizationKernel() : _input(nullptr), _output(nullptr), _epsilon(1e-8f) { } @@ -157,7 +128,8 @@ void NEMeanStdDevNormalizationKernel::configure(ITensor *input, ITensor *output, { ARM_COMPUTE_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_ERROR_THROW_ON(NEMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon)); + ARM_COMPUTE_ERROR_THROW_ON(NEMeanStdDevNormalizationKernel::validate( + input->info(), (output != nullptr) ? output->info() : nullptr, epsilon)); _input = input; _output = (output == nullptr) ? input : output; @@ -167,29 +139,14 @@ void NEMeanStdDevNormalizationKernel::configure(ITensor *input, ITensor *output, auto win_config = validate_and_configure_window(input->info(), (output == nullptr) ? nullptr : output->info()); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICPPKernel::configure(win_config.second); - - // Configure function to run based on different data types - const DataType data_type = input->info()->data_type(); - switch(data_type) - { - case DataType::F32: - _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float, 4>; - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float16_t, 8>; - break; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - default: - ARM_COMPUTE_ERROR("Not Supported"); - break; - } } Status NEMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first); + ARM_COMPUTE_RETURN_ON_ERROR( + validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr) + .first); return Status{}; } @@ -198,8 +155,10 @@ void NEMeanStdDevNormalizationKernel::run(const Window &window, const ThreadInfo ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - (this->*_func)(window); + const auto *uk = get_implementation(MeanStdDevNormSelectorData{_output->info()->data_type()}); + ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr); + + uk->ukernel(_input, _output, _epsilon, window); } } // namespace arm_compute |