diff options
Diffstat (limited to 'src/cpu/kernels/crop/generic/neon/impl.h')
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/impl.h | 49 |
1 files changed, 31 insertions, 18 deletions
diff --git a/src/cpu/kernels/crop/generic/neon/impl.h b/src/cpu/kernels/crop/generic/neon/impl.h index a59588be45..b90ba9ddbf 100644 --- a/src/cpu/kernels/crop/generic/neon/impl.h +++ b/src/cpu/kernels/crop/generic/neon/impl.h @@ -26,8 +26,9 @@ #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/core/NEON/wrapper/wrapper.h" #include "src/cpu/kernels/crop/generic/neon/crop_helper.h" namespace arm_compute @@ -35,19 +36,26 @@ namespace arm_compute 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) +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) + if (is_width_flipped) { // Collapse first dimension if possible. - if(input_has_single_channel) + 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) + 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))); @@ -57,25 +65,27 @@ void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *o 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]) + 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]) + 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) + 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]) + 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))); + *(output_ptr + x * output->info()->dimension(0) + c) = + static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); } } } @@ -83,25 +93,28 @@ void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *o else { // Use memcpy if the elements don't need converting to float. - if(std::is_same<T, float>::value) + 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()); + (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)); + 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) + 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]) + for (; x < limit; ++x, ++input_offset[0]) { - *(output_start_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); + *(output_start_ptr + x) = + static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset))); } } } |