diff options
Diffstat (limited to 'src/cpu/kernels/crop')
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/crop_helper.h | 86 | ||||
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/fp16.cpp | 47 | ||||
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/fp32.cpp | 45 | ||||
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/impl.h | 124 | ||||
| -rw-r--r-- | src/cpu/kernels/crop/generic/neon/integer.cpp | 116 | ||||
| -rw-r--r-- | src/cpu/kernels/crop/list.h | 56 |
6 files changed, 474 insertions, 0 deletions
diff --git a/src/cpu/kernels/crop/generic/neon/crop_helper.h b/src/cpu/kernels/crop/generic/neon/crop_helper.h new file mode 100644 index 0000000000..8fb7ad2087 --- /dev/null +++ b/src/cpu/kernels/crop/generic/neon/crop_helper.h @@ -0,0 +1,86 @@ +/* + * Copyright (c) 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. + */ +#ifndef SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H +#define SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H + +#include "src/core/NEON/wrapper/wrapper.h" + +namespace arm_compute +{ +namespace cpu +{ +template <typename T> +inline float32x4_t load_as_f32(T *ptr) +{ + ARM_COMPUTE_UNUSED(ptr); + ARM_COMPUTE_ERROR("Type not supported."); +} + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +template <> +inline float32x4_t load_as_f32(float16_t *ptr) +{ + return vcvt_f32_f16(wrapper::vload(ptr)); +} +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ + +template <> +inline float32x4_t load_as_f32(float *ptr) +{ + return wrapper::vloadq(ptr); +} + +template <> +inline float32x4_t load_as_f32(int32_t *ptr) +{ + return vcvtq_f32_s32(wrapper::vloadq(ptr)); +} + +template <> +inline float32x4_t load_as_f32(uint32_t *ptr) +{ + return vcvtq_f32_u32(wrapper::vloadq(ptr)); +} + +template <> +inline float32x4_t load_as_f32(int16_t *ptr) +{ + return vcvtq_f32_s32(vmovl_s16(wrapper::vload(ptr))); +} + +template <> +inline float32x4_t load_as_f32(uint16_t *ptr) +{ + return vcvtq_f32_u32(vmovl_u16(wrapper::vload(ptr))); +} + +template <> +inline float32x4_t load_as_f32(uint8_t *ptr) +{ + return vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(wrapper::vload(ptr))))); +} +} // namespace cpu +} // namespace arm_compute + +#endif //SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H diff --git a/src/cpu/kernels/crop/generic/neon/fp16.cpp b/src/cpu/kernels/crop/generic/neon/fp16.cpp new file mode 100644 index 0000000000..3739c9d4e0 --- /dev/null +++ b/src/cpu/kernels/crop/generic/neon/fp16.cpp @@ -0,0 +1,47 @@ +/* + * Copyright (c) 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. + */ +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +#include "src/cpu/kernels/crop/generic/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +void fp16_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) +{ + return in_bounds_crop_window<float16_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/crop/generic/neon/fp32.cpp b/src/cpu/kernels/crop/generic/neon/fp32.cpp new file mode 100644 index 0000000000..f665c3652c --- /dev/null +++ b/src/cpu/kernels/crop/generic/neon/fp32.cpp @@ -0,0 +1,45 @@ +/* + * Copyright (c) 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" + +namespace arm_compute +{ +namespace cpu +{ +void fp32_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) +{ + return in_bounds_crop_window<float32_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/crop/generic/neon/impl.h b/src/cpu/kernels/crop/generic/neon/impl.h new file mode 100644 index 0000000000..b90ba9ddbf --- /dev/null +++ b/src/cpu/kernels/crop/generic/neon/impl.h @@ -0,0 +1,124 @@ +/* + * Copyright (c) 2018-2021, 2023 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. + */ +#ifndef SRC_CORE_NEON_KERNELS_CROP_IMPL_H +#define SRC_CORE_NEON_KERNELS_CROP_IMPL_H + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.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 +{ +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) +{ + // 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 diff --git a/src/cpu/kernels/crop/generic/neon/integer.cpp b/src/cpu/kernels/crop/generic/neon/integer.cpp new file mode 100644 index 0000000000..602434f54f --- /dev/null +++ b/src/cpu/kernels/crop/generic/neon/integer.cpp @@ -0,0 +1,116 @@ +/* + * Copyright (c) 2021-2022 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 "src/cpu/kernels/crop/list.h" + +namespace arm_compute +{ +namespace cpu +{ +void u8_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) +{ + return in_bounds_crop_window<uint8_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} + +void u16_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) +{ + return in_bounds_crop_window<uint16_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} + +void u32_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) +{ + return in_bounds_crop_window<uint32_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} + +void s8_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) +{ + return in_bounds_crop_window<int8_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} + +void s16_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) +{ + return in_bounds_crop_window<int16_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} + +void s32_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) +{ + return in_bounds_crop_window<int32_t>(input, output, output_ptr, input_offset, window_step_x, output_width_start, + output_width_limit, input_has_single_channel, is_width_flipped); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/crop/list.h b/src/cpu/kernels/crop/list.h new file mode 100644 index 0000000000..9cb7726203 --- /dev/null +++ b/src/cpu/kernels/crop/list.h @@ -0,0 +1,56 @@ +/* + * Copyright (c) 2021-2022 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. + */ +#ifndef SRC_CORE_NEON_KERNELS_CROP_LIST_H +#define SRC_CORE_NEON_KERNELS_CROP_LIST_H + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" + +#include "src/core/common/Registrars.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/crop/generic/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_CROP_KERNEL(func_name) \ + void func_name(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) + +DECLARE_CROP_KERNEL(fp16_in_bounds_crop_window); +DECLARE_CROP_KERNEL(fp32_in_bounds_crop_window); +DECLARE_CROP_KERNEL(s8_in_bounds_crop_window); +DECLARE_CROP_KERNEL(s16_in_bounds_crop_window); +DECLARE_CROP_KERNEL(s32_in_bounds_crop_window); +DECLARE_CROP_KERNEL(u8_in_bounds_crop_window); +DECLARE_CROP_KERNEL(u16_in_bounds_crop_window); +DECLARE_CROP_KERNEL(u32_in_bounds_crop_window); + +#undef DECLARE_CROP_KERNEL + +} // namespace cpu +} // namespace arm_compute +#endif //SRC_CORE_NEON_KERNELS_CROP_LIST_H |