diff options
author | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-09-08 14:56:01 +0100 |
---|---|---|
committer | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-09-13 09:02:51 +0000 |
commit | cf219a4be6e9e9637193b5c9aa4f1eedd0a23900 (patch) | |
tree | d0f82767492d8f24bdf6a2641576210e48a37cd7 /src/cpu/kernels/crop/generic/neon/impl.h | |
parent | 3912f47b5849c2c7c4e68ff922869decf22fe303 (diff) | |
download | ComputeLibrary-cf219a4be6e9e9637193b5c9aa4f1eedd0a23900.tar.gz |
Changes in NECropResize to enable fp16 in armv8a multi_isa builds
* Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs
to be moved to an fp16.cpp file to allow compilation with
-march=armv8.2-a+fp16
* fp16.cpp needs to use the template in_bounds_crop_window so it had to be moved from
impl.cpp to impl.h
* Removed the file src/cpu/kernels/crop/generic/neon/impl.cpp
* Partially resolves MLCE-1102
Change-Id: I1953849153e672ff7938f54c877c7498117dcca4
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10282
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/kernels/crop/generic/neon/impl.h')
-rw-r--r-- | src/cpu/kernels/crop/generic/neon/impl.h | 74 |
1 files changed, 72 insertions, 2 deletions
diff --git a/src/cpu/kernels/crop/generic/neon/impl.h b/src/cpu/kernels/crop/generic/neon/impl.h index 50f889705a..a59588be45 100644 --- a/src/cpu/kernels/crop/generic/neon/impl.h +++ b/src/cpu/kernels/crop/generic/neon/impl.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,6 +28,7 @@ #include "arm_compute/core/TensorInfo.h" #include "src/core/NEON/wrapper/wrapper.h" #include "src/core/common/Registrars.h" +#include "src/cpu/kernels/crop/generic/neon/crop_helper.h" namespace arm_compute { @@ -35,7 +36,76 @@ namespace cpu { template <typename T> void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset, - int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped); + int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped) +{ + // Reverse elements if width flipped. + if(is_width_flipped) + { + // Collapse first dimension if possible. + if(input_has_single_channel) + { + int32_t x = output_width_start; + Coordinates negative_offset(input_offset); + negative_offset.set(1, negative_offset[1] - window_step_x + 1); + for(; x <= output_width_limit - window_step_x; x += window_step_x, negative_offset[1] -= window_step_x) + { + auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(negative_offset))); + + in = wrapper::vrev64(in); + in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in)); + + wrapper::vstore(output_ptr + x, in); + } + input_offset[1] = negative_offset[1] + window_step_x - 1; + for(; x < output_width_limit; ++x, --input_offset[1]) + { + *(output_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + } + } + else + { + for(int32_t x = output_width_start; x < output_width_limit; ++x, --input_offset[1]) + { + input_offset.set(0, 0); + int32_t c = 0; + for(; c <= static_cast<int32_t>(input->info()->dimension(0)) - window_step_x; c += window_step_x, input_offset[0] += window_step_x) + { + auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + wrapper::vstore(output_ptr + x * output->info()->dimension(0) + c, in); + } + for(; c < static_cast<int32_t>(input->info()->dimension(0)); ++c, ++input_offset[0]) + { + *(output_ptr + x * output->info()->dimension(0) + c) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + } + } + } + } + else + { + // Use memcpy if the elements don't need converting to float. + if(std::is_same<T, float>::value) + { + memcpy(static_cast<void *>(output_ptr + output_width_start * output->info()->dimension(0)), + reinterpret_cast<const void *>(input->ptr_to_element(input_offset)), + (output_width_limit - output_width_start) * output->info()->dimension(0) * output->info()->element_size()); + } + else + { + int32_t x = 0; + int32_t limit = (output_width_limit - output_width_start) * static_cast<int32_t>(output->info()->dimension(0)); + float *output_start_ptr = output_ptr + output_width_start * output->info()->dimension(0); + for(; x <= limit - window_step_x; x += window_step_x, input_offset[0] += window_step_x) + { + auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + wrapper::vstore(output_start_ptr + x, in); + } + for(; x < limit; ++x, ++input_offset[0]) + { + *(output_start_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + } + } + } +} } // namespace cpu } // namespace arm_compute #endif //SRC_CORE_NEON_KERNELS_CROP_IMPL_H |