diff options
Diffstat (limited to 'src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp | 103 |
1 files changed, 85 insertions, 18 deletions
diff --git a/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp index 4a585b70fd..ba5ca78955 100644 --- a/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp +++ b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp @@ -30,6 +30,8 @@ #include "arm_compute/core/Utils.h" #include "arm_compute/core/Window.h" +#include <arm_neon.h> + namespace arm_compute { namespace @@ -39,7 +41,7 @@ Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anc ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(anchors); ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi()); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::QSYMM16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2); if(all_anchors->total_size() > 0) { @@ -50,6 +52,11 @@ Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anc ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi()); ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors); + + if(is_data_type_quantized(anchors->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(anchors, all_anchors); + } } return Status{}; } @@ -74,7 +81,7 @@ void NEComputeAllAnchorsKernel::configure(const ITensor *anchors, ITensor *all_a // Initialize the output if empty const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors); - auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type); + auto_init_if_empty(*all_anchors->info(), TensorInfo(output_shape, 1, data_type, anchors->info()->quantization_info())); // Set instance variables _anchors = anchors; @@ -92,12 +99,9 @@ Status NEComputeAllAnchorsKernel::validate(const ITensorInfo *anchors, const ITe return Status{}; } -void NEComputeAllAnchorsKernel::run(const Window &window, const ThreadInfo &info) +template <> +void NEComputeAllAnchorsKernel::internal_run<int16_t>(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); - Iterator all_anchors_it(_all_anchors, window); Iterator anchors_it(_all_anchors, window); @@ -105,27 +109,90 @@ void NEComputeAllAnchorsKernel::run(const Window &window, const ThreadInfo &info const float stride = 1.f / _anchors_info.spatial_scale(); const size_t feat_width = _anchors_info.feat_width(); + const UniformQuantizationInfo qinfo = _anchors->info()->quantization_info().uniform(); + execute_window_loop(window, [&](const Coordinates & id) { const size_t anchor_offset = id.y() % num_anchors; - const auto out_anchor_ptr = reinterpret_cast<float *>(all_anchors_it.ptr()); - const auto anchor_ptr = reinterpret_cast<float *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset))); - - *out_anchor_ptr = *anchor_ptr; - *(1 + out_anchor_ptr) = *(1 + anchor_ptr); - *(2 + out_anchor_ptr) = *(2 + anchor_ptr); - *(3 + out_anchor_ptr) = *(3 + anchor_ptr); + const auto out_anchor_ptr = reinterpret_cast<int16_t *>(all_anchors_it.ptr()); + const auto anchor_ptr = reinterpret_cast<int16_t *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset))); const size_t shift_idy = id.y() / num_anchors; const float shiftx = (shift_idy % feat_width) * stride; const float shifty = (shift_idy / feat_width) * stride; - *out_anchor_ptr += shiftx; - *(out_anchor_ptr + 1) += shifty; - *(out_anchor_ptr + 2) += shiftx; - *(out_anchor_ptr + 3) += shifty; + const float new_anchor_x1 = dequantize_qsymm16(*anchor_ptr, qinfo.scale) + shiftx; + const float new_anchor_y1 = dequantize_qsymm16(*(1 + anchor_ptr), qinfo.scale) + shifty; + const float new_anchor_x2 = dequantize_qsymm16(*(2 + anchor_ptr), qinfo.scale) + shiftx; + const float new_anchor_y2 = dequantize_qsymm16(*(3 + anchor_ptr), qinfo.scale) + shifty; + + *out_anchor_ptr = quantize_qsymm16(new_anchor_x1, qinfo.scale); + *(out_anchor_ptr + 1) = quantize_qsymm16(new_anchor_y1, qinfo.scale); + *(out_anchor_ptr + 2) = quantize_qsymm16(new_anchor_x2, qinfo.scale); + *(out_anchor_ptr + 3) = quantize_qsymm16(new_anchor_y2, qinfo.scale); + }, + all_anchors_it); +} + +template <typename T> +void NEComputeAllAnchorsKernel::internal_run(const Window &window, const ThreadInfo &info) +{ + Iterator all_anchors_it(_all_anchors, window); + Iterator anchors_it(_all_anchors, window); + + const size_t num_anchors = _anchors->info()->dimension(1); + const T stride = 1.f / _anchors_info.spatial_scale(); + const size_t feat_width = _anchors_info.feat_width(); + + execute_window_loop(window, [&](const Coordinates & id) + { + const size_t anchor_offset = id.y() % num_anchors; + + const auto out_anchor_ptr = reinterpret_cast<T *>(all_anchors_it.ptr()); + const auto anchor_ptr = reinterpret_cast<T *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset))); + + const size_t shift_idy = id.y() / num_anchors; + const T shiftx = (shift_idy % feat_width) * stride; + const T shifty = (shift_idy / feat_width) * stride; + + *out_anchor_ptr = *anchor_ptr + shiftx; + *(out_anchor_ptr + 1) = *(1 + anchor_ptr) + shifty; + *(out_anchor_ptr + 2) = *(2 + anchor_ptr) + shiftx; + *(out_anchor_ptr + 3) = *(3 + anchor_ptr) + shifty; }, all_anchors_it); } + +void NEComputeAllAnchorsKernel::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); + + switch(_anchors->info()->data_type()) + { + case DataType::QSYMM16: + { + internal_run<int16_t>(window, info); + break; + } +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + { + internal_run<float16_t>(window, info); + break; + } +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F32: + { + internal_run<float>(window, info); + break; + } + default: + { + ARM_COMPUTE_ERROR("Data type not supported"); + } + } +} } // namespace arm_compute |