From 154bc1c3e6a0182e2130c7966af3944ee6ca20b3 Mon Sep 17 00:00:00 2001 From: giuros01 Date: Tue, 26 Mar 2019 17:44:40 +0000 Subject: COMPMID-1973: Implement FFTConvolutionLayer on NEON Change-Id: I2e667c0411bda0164a616ffe44473a78de6752c9 Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/1066 Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins --- .../kernels/NEPixelWiseMultiplicationKernel.cpp | 137 +++++++++++++++++++++ 1 file changed, 137 insertions(+) (limited to 'src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp') diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp index b565300906..fa16484cd3 100644 --- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp +++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp @@ -30,6 +30,7 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/NEAsymm.h" #include "arm_compute/core/NEON/NEFixedPoint.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Validate.h" @@ -353,6 +354,35 @@ void mul_F32_F32_F32_n(const void *__restrict input1_ptr, const void *__restrict vst4q_f32(output, result); } +void c_mul_F32_F32_F32_n(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr) +{ + const auto input1 = static_cast(input1_ptr); + const auto input2 = static_cast(input2_ptr); + const auto output = static_cast(output_ptr); + + const float32x4_t a = wrapper::vloadq(input1); + float32x4_t b = wrapper::vloadq(input2); + + using ExactTagType = typename wrapper::traits::neon_vector::tag_type; + + const float32x4_t mask = { -1.0f, 1.0f, -1.0f, 1.0f }; + const float32x2_t tmp00 = wrapper::vdup_n(wrapper::vgetlane(a, 0), ExactTagType{}); + const float32x2_t tmp01 = wrapper::vdup_n(wrapper::vgetlane(a, 1), ExactTagType{}); + const float32x2_t tmp10 = wrapper::vdup_n(wrapper::vgetlane(a, 2), ExactTagType{}); + const float32x2_t tmp11 = wrapper::vdup_n(wrapper::vgetlane(a, 3), ExactTagType{}); + + const float32x4_t tmp0 = wrapper::vcombine(tmp00, tmp10); + const float32x4_t tmp1 = wrapper::vcombine(tmp01, tmp11); + + float32x4_t res = wrapper::vmul(tmp0, b); + + b = wrapper::vrev64(b); + b = wrapper::vmul(b, mask); + + res = wrapper::vmla(res, tmp1, b); + wrapper::vstore(output, res); +} + void mul_F16_F16_F16_n(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, float scale) { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC @@ -665,4 +695,111 @@ BorderSize NEPixelWiseMultiplicationKernel::border_size() const const unsigned int border = std::min(num_elems_processed_per_iteration - 1U, replicateSize); return BorderSize{ 0, border, 0, 0 }; } + +namespace +{ +constexpr unsigned int num_elems_processed_per_iteration_complex = 2; + +Status validate_arguments_complex(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 2, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 2, DataType::F32); + + const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); + + // Validate in case of configured output + if(output->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 2, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); + } + + return Status{}; +} + +std::pair validate_and_configure_window_complex(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output) +{ + const std::pair broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2); + const TensorShape &out_shape = broadcast_pair.first; + const ValidRegion &valid_region = broadcast_pair.second; + + // Auto initialize output if not initialized + const TensorInfo out_info(out_shape, input1->num_channels(), input1->data_type()); + auto_init_if_empty(*output, out_info); + + Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration_complex)); + Window win_input1 = win.broadcast_if_dimension_le_one(*input1); + Window win_input2 = win.broadcast_if_dimension_le_one(*input2); + + AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration_complex); + AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_complex); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_complex); + + bool window_changed = update_window_and_padding(win_input1, input1_access) + || update_window_and_padding(win_input2, input2_access) + || update_window_and_padding(win, output_access); + + output_access.set_valid_region(win, valid_region); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace + +NEComplexPixelWiseMultiplicationKernel::NEComplexPixelWiseMultiplicationKernel() + : _input1(nullptr), _input2(nullptr), _output(nullptr) +{ +} + +void NEComplexPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITensor *input2, ITensor *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1->info(), input2->info(), output->info())); + + // Configure kernel window + auto win_config = validate_and_configure_window_complex(input1->info(), input2->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + + _input1 = input1; + _input2 = input2; + _output = output; + + // Create kernel + INEKernel::configure(win_config.second); +} + +Status NEComplexPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_complex(input1, input2, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_complex(input1->clone().get(), input2->clone().get(), output->clone().get()).first); + + return Status{}; +} + +void NEComplexPixelWiseMultiplicationKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + Iterator input1(_input1, window.broadcast_if_dimension_le_one(_input1->info()->tensor_shape())); + Iterator input2(_input2, window.broadcast_if_dimension_le_one(_input2->info()->tensor_shape())); + Iterator output(_output, window); + + execute_window_loop(window, [&](const Coordinates &) + { + c_mul_F32_F32_F32_n(input1.ptr(), input2.ptr(), output.ptr()); + }, + input1, input2, output); +} + +BorderSize NEComplexPixelWiseMultiplicationKernel::border_size() const +{ + const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int border = std::min(num_elems_processed_per_iteration_complex - 1U, replicateSize); + return { 0, border, 0, 0 }; +} } // namespace arm_compute -- cgit v1.2.1