From cf219a4be6e9e9637193b5c9aa4f1eedd0a23900 Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Fri, 8 Sep 2023 14:56:01 +0100 Subject: 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 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10282 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins --- src/cpu/kernels/crop/generic/neon/impl.cpp | 127 ----------------------------- src/cpu/kernels/crop/generic/neon/impl.h | 74 ++++++++++++++++- 2 files changed, 72 insertions(+), 129 deletions(-) delete mode 100644 src/cpu/kernels/crop/generic/neon/impl.cpp (limited to 'src/cpu/kernels') diff --git a/src/cpu/kernels/crop/generic/neon/impl.cpp b/src/cpu/kernels/crop/generic/neon/impl.cpp deleted file mode 100644 index 95ab804940..0000000000 --- a/src/cpu/kernels/crop/generic/neon/impl.cpp +++ /dev/null @@ -1,127 +0,0 @@ -/* - * Copyright (c) 2018-2021 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/cpu/kernels/crop/generic/neon/impl.h" -#include "arm_compute/core/Helpers.h" -#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 -{ -namespace cpu -{ -template -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) -{ - // 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(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(*reinterpret_cast(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(input->info()->dimension(0)) - window_step_x; c += window_step_x, input_offset[0] += window_step_x) - { - auto in = load_as_f32(reinterpret_cast(input->ptr_to_element(input_offset))); - wrapper::vstore(output_ptr + x * output->info()->dimension(0) + c, in); - } - for(; c < static_cast(input->info()->dimension(0)); ++c, ++input_offset[0]) - { - *(output_ptr + x * output->info()->dimension(0) + c) = static_cast(*reinterpret_cast(input->ptr_to_element(input_offset))); - } - } - } - } - else - { - // Use memcpy if the elements don't need converting to float. - if(std::is_same::value) - { - memcpy(static_cast(output_ptr + output_width_start * output->info()->dimension(0)), - reinterpret_cast(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(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(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(*reinterpret_cast(input->ptr_to_element(input_offset))); - } - } - } -} - -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -template 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); -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ -template 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); -template 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); -template 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); -template 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); -template 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); -template 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); -template 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); -} -} // namespace arm_compute 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 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(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(*reinterpret_cast(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(input->info()->dimension(0)) - window_step_x; c += window_step_x, input_offset[0] += window_step_x) + { + auto in = load_as_f32(reinterpret_cast(input->ptr_to_element(input_offset))); + wrapper::vstore(output_ptr + x * output->info()->dimension(0) + c, in); + } + for(; c < static_cast(input->info()->dimension(0)); ++c, ++input_offset[0]) + { + *(output_ptr + x * output->info()->dimension(0) + c) = static_cast(*reinterpret_cast(input->ptr_to_element(input_offset))); + } + } + } + } + else + { + // Use memcpy if the elements don't need converting to float. + if(std::is_same::value) + { + memcpy(static_cast(output_ptr + output_width_start * output->info()->dimension(0)), + reinterpret_cast(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(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(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(*reinterpret_cast(input->ptr_to_element(input_offset))); + } + } + } +} } // namespace cpu } // namespace arm_compute #endif //SRC_CORE_NEON_KERNELS_CROP_IMPL_H -- cgit v1.2.1