diff options
-rw-r--r-- | Android.bp | 5 | ||||
-rw-r--r-- | filelist.json | 9 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEROIAlignLayerKernel.cpp | 382 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEROIAlignLayerKernel.h | 5 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/fp16.cpp | 36 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/fp32.cpp | 34 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/impl.cpp | 316 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/impl.h | 40 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/qasymm8.cpp | 34 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp | 34 | ||||
-rw-r--r-- | src/cpu/kernels/roialign/list.h | 40 |
11 files changed, 616 insertions, 319 deletions
diff --git a/Android.bp b/Android.bp index d472e6250f..0c2f6716ee 100644 --- a/Android.bp +++ b/Android.bp @@ -481,6 +481,11 @@ cc_library_static { "src/cpu/kernels/range/generic/neon/fp32.cpp", "src/cpu/kernels/range/generic/neon/impl.cpp", "src/cpu/kernels/range/generic/neon/integer.cpp", + "src/cpu/kernels/roialign/generic/neon/fp16.cpp", + "src/cpu/kernels/roialign/generic/neon/fp32.cpp", + "src/cpu/kernels/roialign/generic/neon/impl.cpp", + "src/cpu/kernels/roialign/generic/neon/qasymm8.cpp", + "src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp", "src/cpu/kernels/scale/neon/fp16.cpp", "src/cpu/kernels/scale/neon/integer.cpp", "src/cpu/kernels/scale/neon/qasymm8.cpp", diff --git a/filelist.json b/filelist.json index 359c0046b9..f5a544868c 100644 --- a/filelist.json +++ b/filelist.json @@ -1790,7 +1790,14 @@ "common": [ "src/core/NEON/kernels/NEROIAlignLayerKernel.cpp", "src/runtime/NEON/functions/NEROIAlignLayer.cpp" - ] + ], + "neon":{ + "common":["src/cpu/kernels/roialign/generic/neon/impl.cpp"], + "fp32":["src/cpu/kernels/roialign/generic/neon/fp32.cpp"], + "fp16":["src/cpu/kernels/roialign/generic/neon/fp16.cpp"], + "qasymm8":["src/cpu/kernels/roialign/generic/neon/qasymm8.cpp"], + "qasymm8_signed":["src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp"] + } } }, "ROIPool2d": { diff --git a/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp b/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp index ece7e40e31..802aebb526 100644 --- a/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp +++ b/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -30,8 +30,10 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/misc/Utility.h" #include "src/core/CPP/Validate.h" +#include "src/core/common/Registrars.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/cpu/kernels/roialign/list.h" #include <arm_neon.h> @@ -41,6 +43,67 @@ namespace arm_compute { namespace { +struct ROIAlignSelectorData +{ + DataType dt; +}; + +using ROIAlignSelctorPtr = std::add_pointer<bool(const ROIAlignSelectorData &data)>::type; +using ROIAlignUKernelPtr = std::add_pointer<void(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info)>::type; + +struct ROIAlignKernel +{ + const char *name; + const ROIAlignSelctorPtr is_selected; + ROIAlignUKernelPtr ukernel; +}; + +static const ROIAlignKernel available_kernels[] = +{ + { + "fp32_neon_roialign", + [](const ROIAlignSelectorData & data) { return data.dt == DataType::F32; }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_roialign) + }, +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + { + "fp16_neon_roialign", + [](const ROIAlignSelectorData & data) { return data.dt == DataType::F16; }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_roialign) + }, +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(ARM_COMPUTE_ENABLE_NEON) + { + "qu8_neon_roialign", + [](const ROIAlignSelectorData & data) { return data.dt == DataType::QASYMM8; }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qu8_roialign) + }, + { + "qs8_neon_roialign", + [](const ROIAlignSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; }, + REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qs8_roialign) + }, +#endif //defined(ARM_COMPUTE_ENABLE_NEON) +}; + +/** Micro-kernel selector + * + * @param[in] data Selection data passed to help pick the appropriate micro-kernel + * + * @return A matching micro-kernel else nullptr + */ +const ROIAlignKernel *get_implementation(const ROIAlignSelectorData &data) +{ + for(const auto &uk : available_kernels) + { + if(uk.is_selected(data)) + { + return &uk; + } + } + return nullptr; +} + Status validate_arguments(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output); @@ -110,328 +173,19 @@ Status NEROIAlignLayerKernel::validate(const ITensorInfo *input, const ITensorIn return Status{}; } -/** Average pooling over an aligned window */ -template <typename input_data_type> -inline input_data_type roi_align_1x1(const ITensor *input, - unsigned int roi_batch, - float region_start_x, - float bin_size_x, - int grid_size_x, - float region_end_x, - float region_start_y, - float bin_size_y, - int grid_size_y, - float region_end_y, - int pz) -{ - if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) - { - return input_data_type(0); - } - else - { - const DataLayout data_layout = input->info()->data_layout(); - float avg = 0; - // Iterate through the aligned pooling region - for(int iy = 0; iy < grid_size_y; ++iy) - { - for(int ix = 0; ix < grid_size_x; ++ix) - { - // Align the window in the middle of every bin - float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y); - float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x); - - // Interpolation in the [0,0] [0,1] [1,0] [1,1] square - const int y_low = y; - const int x_low = x; - const int y_high = y_low + 1; - const int x_high = x_low + 1; - - const float ly = y - y_low; - const float lx = x - x_low; - const float hy = 1. - ly; - const float hx = 1. - lx; - - const float w1 = hy * hx; - const float w2 = hy * lx; - const float w3 = ly * hx; - const float w4 = ly * lx; - if(data_layout == DataLayout::NCHW) - { - const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))); - const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))); - const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))); - const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - else - { - const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))); - const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))); - const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))); - const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - } - } - - avg /= grid_size_x * grid_size_y; - return input_data_type(avg); - } -} - -/** Average pooling over an aligned window */ -template <typename input_data_type> -inline input_data_type roi_align_1x1_qasymm8(const ITensor *input, - unsigned int roi_batch, - float region_start_x, - float bin_size_x, - int grid_size_x, - float region_end_x, - float region_start_y, - float bin_size_y, - int grid_size_y, - float region_end_y, - int pz, - const QuantizationInfo &out_qinfo) -{ - if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) - { - return input_data_type(out_qinfo.uniform().offset); - } - else - { - float avg = 0; - const UniformQuantizationInfo input_qinfo = input->info()->quantization_info().uniform(); - const bool is_qasymm_signed = is_data_type_quantized_asymmetric_signed(input->info()->data_type()); - const DataLayout data_layout = input->info()->data_layout(); - - // Iterate through the aligned pooling region - for(int iy = 0; iy < grid_size_y; ++iy) - { - for(int ix = 0; ix < grid_size_x; ++ix) - { - // Align the window in the middle of every bin - float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y); - float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x); - - // Interpolation in the [0,0] [0,1] [1,0] [1,1] square - const int y_low = y; - const int x_low = x; - const int y_high = y_low + 1; - const int x_high = x_low + 1; - - const float ly = y - y_low; - const float lx = x - x_low; - const float hy = 1. - ly; - const float hx = 1. - lx; - - const float w1 = hy * hx; - const float w2 = hy * lx; - const float w3 = ly * hx; - const float w4 = ly * lx; - - if(data_layout == DataLayout::NCHW) - { - if(is_qasymm_signed) - { - float data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo); - float data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo); - float data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo); - float data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - else - { - float data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo); - float data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo); - float data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo); - float data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - } - else - { - if(is_qasymm_signed) - { - const auto data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo); - const auto data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo); - const auto data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo); - const auto data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - else - { - const auto data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo); - const auto data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo); - const auto data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo); - const auto data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo); - avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; - } - } - } - } - - avg /= grid_size_x * grid_size_y; - - input_data_type res = 0; - if(is_qasymm_signed) - { - res = quantize_qasymm8_signed(avg, out_qinfo); - } - else - { - res = quantize_qasymm8(avg, out_qinfo); - } - return res; - } -} - -inline float compute_region_coordinate(int p, float bin_size, float roi_anchor, float max_value) -{ - const float region_start = p * bin_size + roi_anchor; - return utility::clamp(region_start, 0.0f, max_value); -} - void NEROIAlignLayerKernel::run(const Window &window, const ThreadInfo &info) { const DataLayout data_layout = _input->info()->data_layout(); if(data_layout == DataLayout::NCHW || data_layout == DataLayout::NHWC) { - switch(_input->info()->data_type()) - { - case DataType::QASYMM8: - { - NEROIAlignLayerKernel::internal_run<uint8_t, uint16_t>(window, info); - break; - } - case DataType::QASYMM8_SIGNED: - { - NEROIAlignLayerKernel::internal_run<int8_t, uint16_t>(window, info); - break; - } - case DataType::F32: - { - NEROIAlignLayerKernel::internal_run<float>(window, info); - break; - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - { - NEROIAlignLayerKernel::internal_run<float16_t>(window, info); - break; - } -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - default: - { - ARM_COMPUTE_ERROR("DataType not supported"); - break; - } - } + const auto *uk = get_implementation(ROIAlignSelectorData{ _input->info()->data_type() }); + ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr); + + uk->ukernel(_input, _output, _rois, _pool_info, window, info); } else { ARM_COMPUTE_ERROR("Invalid layout"); } } - -template <typename input_data_type, typename roi_data_type> -void NEROIAlignLayerKernel::internal_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); - - const DataLayout data_layout = _input->info()->data_layout(); - const size_t values_per_roi = _rois->info()->dimension(0); - - const int roi_list_start = window.x().start(); - const int roi_list_end = window.x().end(); - - const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const unsigned int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); - - const int input_width = _input->info()->dimension(idx_width); - const int input_height = _input->info()->dimension(idx_height); - const int input_chanels = _input->info()->dimension(idx_depth); - const int pooled_w = _pool_info.pooled_width(); - const int pooled_h = _pool_info.pooled_height(); - - const DataType data_type = _input->info()->data_type(); - const bool is_qasymm = is_data_type_quantized_asymmetric(data_type); - - const auto *rois_ptr = reinterpret_cast<const roi_data_type *>(_rois->buffer()); - const QuantizationInfo &rois_qinfo = _rois->info()->quantization_info(); - for(int roi_indx = roi_list_start; roi_indx < roi_list_end; ++roi_indx) - { - const unsigned int roi_batch = rois_ptr[values_per_roi * roi_indx]; - - roi_data_type qx1 = rois_ptr[values_per_roi * roi_indx + 1]; - roi_data_type qy1 = rois_ptr[values_per_roi * roi_indx + 2]; - roi_data_type qx2 = rois_ptr[values_per_roi * roi_indx + 3]; - roi_data_type qy2 = rois_ptr[values_per_roi * roi_indx + 4]; - float x1(qx1); - float x2(qx2); - float y1(qy1); - float y2(qy2); - if(is_qasymm) - { - x1 = dequantize_qasymm16(qx1, rois_qinfo); - x2 = dequantize_qasymm16(qx2, rois_qinfo); - y1 = dequantize_qasymm16(qy1, rois_qinfo); - y2 = dequantize_qasymm16(qy2, rois_qinfo); - } - const float roi_anchor_x = x1 * _pool_info.spatial_scale(); - const float roi_anchor_y = y1 * _pool_info.spatial_scale(); - const float roi_dims_x = std::max((x2 - x1) * _pool_info.spatial_scale(), 1.0f); - const float roi_dims_y = std::max((y2 - y1) * _pool_info.spatial_scale(), 1.0f); - float bin_size_x = roi_dims_x / _pool_info.pooled_width(); - float bin_size_y = roi_dims_y / _pool_info.pooled_height(); - - // Iterate through all feature maps - for(int ch = 0; ch < input_chanels; ++ch) - { - // Iterate through all output pixels - for(int py = 0; py < pooled_h; ++py) - { - for(int px = 0; px < pooled_w; ++px) - { - const float region_start_x = compute_region_coordinate(px, bin_size_x, roi_anchor_x, input_width); - const float region_start_y = compute_region_coordinate(py, bin_size_y, roi_anchor_y, input_height); - const float region_end_x = compute_region_coordinate(px + 1, bin_size_x, roi_anchor_x, input_width); - const float region_end_y = compute_region_coordinate(py + 1, bin_size_y, roi_anchor_y, input_height); - const int roi_bin_grid_x = (_pool_info.sampling_ratio() > 0) ? _pool_info.sampling_ratio() : int(ceil(bin_size_x)); - const int roi_bin_grid_y = (_pool_info.sampling_ratio() > 0) ? _pool_info.sampling_ratio() : int(ceil(bin_size_y)); - input_data_type out_val(0); - if(is_qasymm) - { - out_val = roi_align_1x1_qasymm8<input_data_type>( - _input, roi_batch, region_start_x, bin_size_x, - roi_bin_grid_x, region_end_x, region_start_y, bin_size_y, - roi_bin_grid_y, region_end_y, ch, _output->info()->quantization_info()); - } - else - { - out_val = roi_align_1x1<input_data_type>( - _input, roi_batch, region_start_x, bin_size_x, - roi_bin_grid_x, region_end_x, region_start_y, bin_size_y, - roi_bin_grid_y, region_end_y, ch); - } - - if(data_layout == DataLayout::NCHW) - { - auto out_ptr = reinterpret_cast<input_data_type *>(_output->ptr_to_element(Coordinates(px, py, ch, roi_indx))); - *out_ptr = out_val; - } - else - { - auto out_ptr = reinterpret_cast<input_data_type *>(_output->ptr_to_element(Coordinates(ch, px, py, roi_indx))); - *out_ptr = out_val; - } - } - } - } - } -} } // namespace arm_compute diff --git a/src/core/NEON/kernels/NEROIAlignLayerKernel.h b/src/core/NEON/kernels/NEROIAlignLayerKernel.h index fa31a879b7..48a3de7285 100644 --- a/src/core/NEON/kernels/NEROIAlignLayerKernel.h +++ b/src/core/NEON/kernels/NEROIAlignLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -89,9 +89,6 @@ public: void run(const Window &window, const ThreadInfo &info) override; private: - template <typename input_data_type, typename roi_data_type = input_data_type> - void internal_run(const Window &window, const ThreadInfo &info); - const ITensor *_input; ITensor *_output; const ITensor *_rois; diff --git a/src/cpu/kernels/roialign/generic/neon/fp16.cpp b/src/cpu/kernels/roialign/generic/neon/fp16.cpp new file mode 100644 index 0000000000..6e585a4dfa --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/fp16.cpp @@ -0,0 +1,36 @@ +/* + * Copyright (c) 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. + */ +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +#include "src/cpu/kernels/roialign/generic/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_fp16_roialign(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +{ + return roi_align<float16_t, float16_t>(input, output, rois, pool_info, window, info); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/roialign/generic/neon/fp32.cpp b/src/cpu/kernels/roialign/generic/neon/fp32.cpp new file mode 100644 index 0000000000..51355aaef0 --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/fp32.cpp @@ -0,0 +1,34 @@ +/* + * Copyright (c) 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/roialign/generic/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_fp32_roialign(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +{ + return roi_align<float, float>(input, output, rois, pool_info, window, info); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/roialign/generic/neon/impl.cpp b/src/cpu/kernels/roialign/generic/neon/impl.cpp new file mode 100644 index 0000000000..a4502e703d --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/impl.cpp @@ -0,0 +1,316 @@ +/* + * Copyright (c) 2019-2022 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/roialign/generic/neon/impl.h" +#include "src/core/NEON/INEKernel.h" +namespace arm_compute +{ +namespace cpu +{ +/** Average pooling over an aligned window */ +template <typename input_data_type> +inline input_data_type roi_align_1x1(const ITensor *input, + unsigned int roi_batch, + float region_start_x, + float bin_size_x, + int grid_size_x, + float region_end_x, + float region_start_y, + float bin_size_y, + int grid_size_y, + float region_end_y, + int pz) +{ + if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) + { + return input_data_type(0); + } + else + { + const DataLayout data_layout = input->info()->data_layout(); + float avg = 0; + // Iterate through the aligned pooling region + for(int iy = 0; iy < grid_size_y; ++iy) + { + for(int ix = 0; ix < grid_size_x; ++ix) + { + // Align the window in the middle of every bin + float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y); + float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x); + + // Interpolation in the [0,0] [0,1] [1,0] [1,1] square + const int y_low = y; + const int x_low = x; + const int y_high = y_low + 1; + const int x_high = x_low + 1; + + const float ly = y - y_low; + const float lx = x - x_low; + const float hy = 1. - ly; + const float hx = 1. - lx; + + const float w1 = hy * hx; + const float w2 = hy * lx; + const float w3 = ly * hx; + const float w4 = ly * lx; + if(data_layout == DataLayout::NCHW) + { + const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))); + const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))); + const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))); + const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + else + { + const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))); + const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))); + const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))); + const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + } + } + + avg /= grid_size_x * grid_size_y; + return input_data_type(avg); + } +} + +/** Average pooling over an aligned window */ +template <typename input_data_type> +inline input_data_type roi_align_1x1_qasymm8(const ITensor *input, + unsigned int roi_batch, + float region_start_x, + float bin_size_x, + int grid_size_x, + float region_end_x, + float region_start_y, + float bin_size_y, + int grid_size_y, + float region_end_y, + int pz, + const QuantizationInfo &out_qinfo) +{ + if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) + { + return input_data_type(out_qinfo.uniform().offset); + } + else + { + float avg = 0; + const UniformQuantizationInfo input_qinfo = input->info()->quantization_info().uniform(); + const bool is_qasymm_signed = is_data_type_quantized_asymmetric_signed(input->info()->data_type()); + const DataLayout data_layout = input->info()->data_layout(); + + // Iterate through the aligned pooling region + for(int iy = 0; iy < grid_size_y; ++iy) + { + for(int ix = 0; ix < grid_size_x; ++ix) + { + // Align the window in the middle of every bin + float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y); + float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x); + + // Interpolation in the [0,0] [0,1] [1,0] [1,1] square + const int y_low = y; + const int x_low = x; + const int y_high = y_low + 1; + const int x_high = x_low + 1; + + const float ly = y - y_low; + const float lx = x - x_low; + const float hy = 1. - ly; + const float hx = 1. - lx; + + const float w1 = hy * hx; + const float w2 = hy * lx; + const float w3 = ly * hx; + const float w4 = ly * lx; + + if(data_layout == DataLayout::NCHW) + { + if(is_qasymm_signed) + { + float data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo); + float data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo); + float data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo); + float data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + else + { + float data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo); + float data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo); + float data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo); + float data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + } + else + { + if(is_qasymm_signed) + { + const auto data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo); + const auto data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo); + const auto data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo); + const auto data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + else + { + const auto data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo); + const auto data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo); + const auto data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo); + const auto data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo); + avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4; + } + } + } + } + + avg /= grid_size_x * grid_size_y; + + input_data_type res = 0; + if(is_qasymm_signed) + { + res = quantize_qasymm8_signed(avg, out_qinfo); + } + else + { + res = quantize_qasymm8(avg, out_qinfo); + } + return res; + } +} +inline float compute_region_coordinate(int p, float bin_size, float roi_anchor, float max_value) +{ + const float region_start = p * bin_size + roi_anchor; + return utility::clamp(region_start, 0.0f, max_value); +} + +template <typename input_data_type, typename roi_data_type> +void roi_align(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + + const DataLayout data_layout = input->info()->data_layout(); + const size_t values_per_roi = rois->info()->dimension(0); + + const int roi_list_start = window.x().start(); + const int roi_list_end = window.x().end(); + + const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const unsigned int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + + const int input_width = input->info()->dimension(idx_width); + const int input_height = input->info()->dimension(idx_height); + const int input_chanels = input->info()->dimension(idx_depth); + const int pooled_w = pool_info.pooled_width(); + const int pooled_h = pool_info.pooled_height(); + + const DataType data_type = input->info()->data_type(); + const bool is_qasymm = is_data_type_quantized_asymmetric(data_type); + + const auto *rois_ptr = reinterpret_cast<const roi_data_type *>(rois->buffer()); + const QuantizationInfo &rois_qinfo = rois->info()->quantization_info(); + for(int roi_indx = roi_list_start; roi_indx < roi_list_end; ++roi_indx) + { + const unsigned int roi_batch = rois_ptr[values_per_roi * roi_indx]; + + roi_data_type qx1 = rois_ptr[values_per_roi * roi_indx + 1]; + roi_data_type qy1 = rois_ptr[values_per_roi * roi_indx + 2]; + roi_data_type qx2 = rois_ptr[values_per_roi * roi_indx + 3]; + roi_data_type qy2 = rois_ptr[values_per_roi * roi_indx + 4]; + float x1(qx1); + float x2(qx2); + float y1(qy1); + float y2(qy2); + if(is_qasymm) + { + x1 = dequantize_qasymm16(qx1, rois_qinfo); + x2 = dequantize_qasymm16(qx2, rois_qinfo); + y1 = dequantize_qasymm16(qy1, rois_qinfo); + y2 = dequantize_qasymm16(qy2, rois_qinfo); + } + const float roi_anchor_x = x1 * pool_info.spatial_scale(); + const float roi_anchor_y = y1 * pool_info.spatial_scale(); + const float roi_dims_x = std::max((x2 - x1) * pool_info.spatial_scale(), 1.0f); + const float roi_dims_y = std::max((y2 - y1) * pool_info.spatial_scale(), 1.0f); + float bin_size_x = roi_dims_x / pool_info.pooled_width(); + float bin_size_y = roi_dims_y / pool_info.pooled_height(); + + // Iterate through all feature maps + for(int ch = 0; ch < input_chanels; ++ch) + { + // Iterate through all output pixels + for(int py = 0; py < pooled_h; ++py) + { + for(int px = 0; px < pooled_w; ++px) + { + const float region_start_x = compute_region_coordinate(px, bin_size_x, roi_anchor_x, input_width); + const float region_start_y = compute_region_coordinate(py, bin_size_y, roi_anchor_y, input_height); + const float region_end_x = compute_region_coordinate(px + 1, bin_size_x, roi_anchor_x, input_width); + const float region_end_y = compute_region_coordinate(py + 1, bin_size_y, roi_anchor_y, input_height); + const int roi_bin_grid_x = (pool_info.sampling_ratio() > 0) ? pool_info.sampling_ratio() : int(ceil(bin_size_x)); + const int roi_bin_grid_y = (pool_info.sampling_ratio() > 0) ? pool_info.sampling_ratio() : int(ceil(bin_size_y)); + input_data_type out_val(0); + if(is_qasymm) + { + out_val = roi_align_1x1_qasymm8<input_data_type>( + input, roi_batch, region_start_x, bin_size_x, + roi_bin_grid_x, region_end_x, region_start_y, bin_size_y, + roi_bin_grid_y, region_end_y, ch, output->info()->quantization_info()); + } + else + { + out_val = roi_align_1x1<input_data_type>( + input, roi_batch, region_start_x, bin_size_x, + roi_bin_grid_x, region_end_x, region_start_y, bin_size_y, + roi_bin_grid_y, region_end_y, ch); + } + + if(data_layout == DataLayout::NCHW) + { + auto out_ptr = reinterpret_cast<input_data_type *>(output->ptr_to_element(Coordinates(px, py, ch, roi_indx))); + *out_ptr = out_val; + } + else + { + auto out_ptr = reinterpret_cast<input_data_type *>(output->ptr_to_element(Coordinates(ch, px, py, roi_indx))); + *out_ptr = out_val; + } + } + } + } + } +} +template void roi_align<float, float>(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info); +template void roi_align<uint8_t, uint16_t>(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info); +template void roi_align<int8_t, uint16_t>(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info); +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +template void roi_align<float16_t, float16_t>(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info); +#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/roialign/generic/neon/impl.h b/src/cpu/kernels/roialign/generic/neon/impl.h new file mode 100644 index 0000000000..ff96b8edab --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/impl.h @@ -0,0 +1,40 @@ +/* + * Copyright (c) 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_SVE_KERNELS_BOUNDINGBOXTRANFORM_IMPL_H +#define SRC_CORE_SVE_KERNELS_BOUNDINGBOXTRANFORM_IMPL_H +#include "arm_compute/core/CPP/CPPTypes.h" +#include "arm_compute/core/Helpers.h" + +namespace arm_compute +{ +class ITensor; +class Window; +namespace cpu +{ +template <typename input_data_type, typename roi_data_type> +void roi_align(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info); + +} // namespace cpu +} // namespace arm_compute +#endif //define SRC_CORE_SVE_KERNELS_BOUNDINGBOXTRANFORM_IMPL_H diff --git a/src/cpu/kernels/roialign/generic/neon/qasymm8.cpp b/src/cpu/kernels/roialign/generic/neon/qasymm8.cpp new file mode 100644 index 0000000000..d6bd9a95ce --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/qasymm8.cpp @@ -0,0 +1,34 @@ +/* + * Copyright (c) 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/roialign/generic/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_qu8_roialign(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +{ + return roi_align<uint8_t, uint16_t>(input, output, rois, pool_info, window, info); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp b/src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp new file mode 100644 index 0000000000..a839581aff --- /dev/null +++ b/src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp @@ -0,0 +1,34 @@ +/* + * Copyright (c) 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/roialign/generic/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_qs8_roialign(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +{ + return roi_align<int8_t, uint16_t>(input, output, rois, pool_info, window, info); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/roialign/list.h b/src/cpu/kernels/roialign/list.h new file mode 100644 index 0000000000..1c71b02488 --- /dev/null +++ b/src/cpu/kernels/roialign/list.h @@ -0,0 +1,40 @@ +/* + * Copyright (c) 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_ROIALIGN_LIST_H +#define SRC_CORE_NEON_KERNELS_ROIALIGN_LIST_H +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_ROIALIGN_KERNEL(func_name) \ + void func_name(const ITensor *input, ITensor *output, const ITensor *rois, \ + ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info) +DECLARE_ROIALIGN_KERNEL(neon_fp32_roialign); +DECLARE_ROIALIGN_KERNEL(neon_fp16_roialign); +DECLARE_ROIALIGN_KERNEL(neon_qu8_roialign); +DECLARE_ROIALIGN_KERNEL(neon_qs8_roialign); +#undef DECLARE_ROIALIGN_KERNEL +} // namespace cpu +} // namespace arm_compute +#endif /* SRC_CORE_NEON_KERNELS_ROIALIGN_LIST_H */ |