diff options
Diffstat (limited to 'src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp')
-rw-r--r-- | src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp | 255 |
1 files changed, 87 insertions, 168 deletions
diff --git a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp index 9900446218..eea57a17d3 100644 --- a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp +++ b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,18 +21,23 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h" +#include "src/core/NEON/kernels/NEL2NormalizeLayerKernel.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/NEMath.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" -#include "arm_compute/core/NEON/wrapper/wrapper.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/core/common/Registrars.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/NEMath.h" +#include "src/cpu/kernels/l2normlayer/list.h" + #include <arm_neon.h> #include <cmath> @@ -42,108 +47,68 @@ namespace { constexpr int max_input_tensor_dim = 3; -template <typename T, int S> -void l2_normalize_X(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window) +struct L2NormalizeLayerSelectorData { - /** NEON vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; - - Window window_sum(window); - window_sum.set(Window::DimX, Window::Dimension(0, 0, 0)); - - Window in_slice = window.first_slice_window_1D(); - Window sum_slice = window_sum.first_slice_window_1D(); - - do - { - Iterator input_it(in, in_slice); - Iterator sum_it(sum, sum_slice); - Iterator output_it(out, in_slice); + DataType dt; + unsigned int actual_axis; + cpuinfo::CpuIsaInfo isa; +}; - const auto sum_value = *reinterpret_cast<const T *>(sum_it.ptr()); - const auto vec_normalize_value = wrapper::vdup_n(static_cast<T>(1.f / std::sqrt(std::max(sum_value, static_cast<T>(epsilon)))), ExactTagType{}); +using L2NormalizeLayerKernelSelctorPtr = std::add_pointer<bool(const L2NormalizeLayerSelectorData &data)>::type; - execute_window_loop(in_slice, [&](const Coordinates &) - { - const auto in_ptr = reinterpret_cast<const T *>(input_it.ptr()); - const auto out_ptr = reinterpret_cast<T *>(output_it.ptr()); - - wrapper::vstore(out_ptr, wrapper::vmul(wrapper::vloadq(in_ptr), vec_normalize_value)); - }, - input_it, output_it); - } - while(window.slide_window_slice_1D(in_slice) && window.slide_window_slice_1D(sum_slice)); -} +using L2NormalizeLayerPtr = std::add_pointer<void( + const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window, size_t axis)>::type; -template <typename T, int S> -void l2_normalize_Y(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window) +struct L2NormalizeLayerKernel { - /** NEON vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; - - Window window_sum(window); - window_sum.set(Window::DimY, Window::Dimension(0, 0, 0)); - - Window in_slice = window.first_slice_window_2D(); - Window sum_slice = window_sum.first_slice_window_2D(); - - do + const char *name; + const L2NormalizeLayerKernelSelctorPtr is_selected; + L2NormalizeLayerPtr ukernel; +}; + +static const L2NormalizeLayerKernel available_kernels[] = { + {"fp32_neon_l2normalize_x", + [](const L2NormalizeLayerSelectorData &data) + { return data.dt == DataType::F32 && data.actual_axis == Window::DimX; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_l2_normalize_x)}, + {"fp32_neon_l2normalize_yz", + [](const L2NormalizeLayerSelectorData &data) + { return data.dt == DataType::F32 && data.actual_axis != Window::DimX; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_l2_normalize_yz)}, { - Iterator input_it(in, in_slice); - Iterator sum_it(sum, sum_slice); - Iterator output_it(out, in_slice); - - auto eps = wrapper::vdup_n(static_cast<T>(epsilon), ExactTagType{}); - - execute_window_loop(in_slice, [&](const Coordinates &) - { - const auto in_ptr = reinterpret_cast<const T *>(input_it.ptr()); - const auto sum_ptr = reinterpret_cast<const T *>(sum_it.ptr()); - const auto out_ptr = reinterpret_cast<T *>(output_it.ptr()); - - const auto vec_normalize_value = wrapper::vinvsqrt(wrapper::vmax(wrapper::vloadq(sum_ptr), eps)); - wrapper::vstore(out_ptr, wrapper::vmul(wrapper::vloadq(in_ptr), vec_normalize_value)); - }, - input_it, sum_it, output_it); - } - while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice)); -} - -template <typename T, int S> -void l2_normalize_Z(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window) + "fp16_neon_l2normalize_x", + [](const L2NormalizeLayerSelectorData &data) + { return data.dt == DataType::F16 && data.isa.fp16 && data.actual_axis == Window::DimX; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_l2_normalize_x), + }, + { + "fp16_neon_l2normalize_yz", + [](const L2NormalizeLayerSelectorData &data) + { return data.dt == DataType::F16 && data.isa.fp16 && data.actual_axis != Window::DimX; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_l2_normalize_yz), + }, +}; + +/** Micro-kernel selector + * + * @param[in] data Selection data passed to help pick the appropriate micro-kernel + * + * @return A matching micro-kernel else nullptr + */ +const L2NormalizeLayerKernel *get_implementation(const L2NormalizeLayerSelectorData &data) { - /** NEON vector tag type. */ - using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; - - Window window_sum(window); - window_sum.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - Window in_slice = window.first_slice_window_3D(); - Window sum_slice = window_sum.first_slice_window_3D(); - - do + for (const auto &uk : available_kernels) { - Iterator input_it(in, in_slice); - Iterator sum_it(sum, sum_slice); - Iterator output_it(out, in_slice); - - auto eps = wrapper::vdup_n(static_cast<T>(epsilon), ExactTagType{}); - - execute_window_loop(in_slice, [&](const Coordinates &) + if (uk.is_selected(data)) { - const auto in_ptr = reinterpret_cast<const T *>(input_it.ptr()); - const auto sum_ptr = reinterpret_cast<const T *>(sum_it.ptr()); - const auto out_ptr = reinterpret_cast<T *>(output_it.ptr()); - - const auto vec_normalize_value = wrapper::vinvsqrt(wrapper::vmax(wrapper::vloadq(sum_ptr), eps)); - wrapper::vstore(out_ptr, wrapper::vmul(wrapper::vloadq(in_ptr), vec_normalize_value)); - }, - input_it, sum_it, output_it); + return &uk; + } } - while(window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice)); + return nullptr; } -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) +Status +validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_UNUSED(epsilon); @@ -152,14 +117,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, + "Actual normalization axis greater than max number of dimensions"); // Reduce shape on axis TensorShape sum_shape = input->tensor_shape(); sum_shape.set(actual_axis, 1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape); - if(output->total_size() != 0) + if (output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); @@ -170,27 +136,16 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons return Status{}; } -std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *sum, ITensorInfo *output, int axis) +std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) { - const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); - const unsigned int num_elems_processed_per_iteration = 16 / data_size_from_type(input->data_type()); - const unsigned int num_elems_processed_per_iteration_sum = (actual_axis == 0) ? 1 : num_elems_processed_per_iteration; - - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); + Window win = calculate_max_window(*input, Steps()); // Output auto initialization if not yet initialized auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type()); - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal sum_access(sum, 0, num_elems_processed_per_iteration_sum); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + // NEL2NormalizeLayerKernel doesn't need padding so update_window_and_padding() can be skipped - bool window_changed = update_window_and_padding(win, input_access, sum_access, output_access); - output_access.set_valid_region(win, input->valid_region()); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - - return std::make_tuple(err, win); + return std::make_tuple(Status{}, win); } } // namespace @@ -199,28 +154,31 @@ NEL2NormalizeLayerKernel::NEL2NormalizeLayerKernel() { } -void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon) +void NEL2NormalizeLayerKernel::configure( + const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon)); - _input = input; - _sum = sum; - _output = output; - _actual_axis = wrap_around(axis, max_input_tensor_dim); - _epsilon = epsilon; + _input = input; + _sum = sum; + _output = output; + _actual_axis = wrap_around(axis, max_input_tensor_dim); + _epsilon = epsilon; // Configure kernel window - auto win_config = validate_and_configure_window(_input->info(), _sum->info(), _output->info(), axis); + auto win_config = validate_and_configure_window(_input->info(), _output->info()); ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); INEKernel::configure(std::get<1>(win_config)); } -Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) +Status NEL2NormalizeLayerKernel::validate( + const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), sum->clone().get(), output->clone().get(), axis))); + ARM_COMPUTE_RETURN_ON_ERROR( + std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get()))); return Status{}; } @@ -231,55 +189,16 @@ void NEL2NormalizeLayerKernel::run(const Window &window, const ThreadInfo &info) ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - switch(_actual_axis) + if (_actual_axis > 2) { - case 0: - switch(_input->info()->data_type()) - { - case DataType::F32: - l2_normalize_X<float, 4>(_input, _sum, _output, _epsilon, window); - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - l2_normalize_X<float16_t, 8>(_input, _sum, _output, _epsilon, window); - break; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - break; - case 1: - switch(_input->info()->data_type()) - { - case DataType::F32: - l2_normalize_Y<float, 4>(_input, _sum, _output, _epsilon, window); - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - l2_normalize_Y<float16_t, 8>(_input, _sum, _output, _epsilon, window); -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - break; - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - break; - case 2: - switch(_input->info()->data_type()) - { - case DataType::F32: - l2_normalize_Z<float, 4>(_input, _sum, _output, _epsilon, window); - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - l2_normalize_Z<float16_t, 8>(_input, _sum, _output, _epsilon, window); - break; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - break; - default: - ARM_COMPUTE_ERROR("Unsupported normalization axis"); + ARM_COMPUTE_ERROR("Unsupported normalization axis"); } + + const auto *uk = get_implementation( + L2NormalizeLayerSelectorData{_output->info()->data_type(), _actual_axis, CPUInfo::get().get_isa()}); + ARM_COMPUTE_ERROR_ON(uk == nullptr); + ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr); + + uk->ukernel(_input, _sum, _output, _epsilon, window, _actual_axis); } } // namespace arm_compute |