diff options
Diffstat (limited to 'src/cpu/kernels/boundingboxtransform/generic')
5 files changed, 113 insertions, 75 deletions
diff --git a/src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp b/src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp index 5661479059..dbdec5fb50 100644 --- a/src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp +++ b/src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp @@ -29,7 +29,11 @@ namespace arm_compute { namespace cpu { -void neon_fp16_boundingboxtransform(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window) +void neon_fp16_boundingboxtransform(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window) { return bounding_box_transform<float16_t>(boxes, pred_boxes, deltas, bbinfo, window); } diff --git a/src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp b/src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp index 34ff9224d5..0224b3406a 100644 --- a/src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp +++ b/src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp @@ -26,7 +26,11 @@ namespace arm_compute { namespace cpu { -void neon_fp32_boundingboxtransform(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window) +void neon_fp32_boundingboxtransform(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window) { return bounding_box_transform<float>(boxes, pred_boxes, deltas, bbinfo, window); } diff --git a/src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp b/src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp index b3ffd0a676..5a2939b587 100644 --- a/src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp +++ b/src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp @@ -29,7 +29,11 @@ namespace arm_compute { namespace cpu { -void bounding_box_transform_qsymm16(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window) +void bounding_box_transform_qsymm16(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window) { const size_t num_classes = deltas->info()->tensor_shape()[0] >> 2; @@ -41,7 +45,8 @@ void bounding_box_transform_qsymm16(const ITensor *boxes, ITensor *pred_boxes, c const auto scale_before = bbinfo.scale(); const auto offset = (bbinfo.correct_transform_coords() ? 1.f : 0.f); - auto pred_ptr = reinterpret_cast<uint16_t *>(pred_boxes->buffer() + pred_boxes->info()->offset_first_element_in_bytes()); + auto pred_ptr = + reinterpret_cast<uint16_t *>(pred_boxes->buffer() + pred_boxes->info()->offset_first_element_in_bytes()); auto delta_ptr = reinterpret_cast<uint8_t *>(deltas->buffer() + deltas->info()->offset_first_element_in_bytes()); const auto boxes_qinfo = boxes->info()->quantization_info().uniform(); @@ -49,41 +54,49 @@ void bounding_box_transform_qsymm16(const ITensor *boxes, ITensor *pred_boxes, c const auto pred_qinfo = pred_boxes->info()->quantization_info().uniform(); Iterator box_it(boxes, window); - execute_window_loop(window, [&](const Coordinates & id) - { - const auto ptr = reinterpret_cast<uint16_t *>(box_it.ptr()); - const auto b0 = dequantize_qasymm16(*ptr, boxes_qinfo); - const auto b1 = dequantize_qasymm16(*(ptr + 1), boxes_qinfo); - const auto b2 = dequantize_qasymm16(*(ptr + 2), boxes_qinfo); - const auto b3 = dequantize_qasymm16(*(ptr + 3), boxes_qinfo); - const float width = (b2 / scale_before) - (b0 / scale_before) + 1.f; - const float height = (b3 / scale_before) - (b1 / scale_before) + 1.f; - const float ctr_x = (b0 / scale_before) + 0.5f * width; - const float ctr_y = (b1 / scale_before) + 0.5f * height; - for(size_t j = 0; j < num_classes; ++j) + execute_window_loop( + window, + [&](const Coordinates &id) { - // Extract deltas - const size_t delta_id = id.y() * deltas_width + 4u * j; - const float dx = dequantize_qasymm8(delta_ptr[delta_id], deltas_qinfo) / bbinfo.weights()[0]; - const float dy = dequantize_qasymm8(delta_ptr[delta_id + 1], deltas_qinfo) / bbinfo.weights()[1]; - float dw = dequantize_qasymm8(delta_ptr[delta_id + 2], deltas_qinfo) / bbinfo.weights()[2]; - float dh = dequantize_qasymm8(delta_ptr[delta_id + 3], deltas_qinfo) / bbinfo.weights()[3]; - // Clip dw and dh - dw = std::min(dw, bbinfo.bbox_xform_clip()); - dh = std::min(dh, bbinfo.bbox_xform_clip()); - // Determine the predictions - const float pred_ctr_x = dx * width + ctr_x; - const float pred_ctr_y = dy * height + ctr_y; - const float pred_w = std::exp(dw) * width; - const float pred_h = std::exp(dh) * height; - // Store the prediction into the output tensor - pred_ptr[delta_id] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_x - 0.5f * pred_w, 0.f, img_w - 1.f), pred_qinfo); - pred_ptr[delta_id + 1] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_y - 0.5f * pred_h, 0.f, img_h - 1.f), pred_qinfo); - pred_ptr[delta_id + 2] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_x + 0.5f * pred_w - offset, 0.f, img_w - 1.f), pred_qinfo); - pred_ptr[delta_id + 3] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_y + 0.5f * pred_h - offset, 0.f, img_h - 1.f), pred_qinfo); - } - }, - box_it); + const auto ptr = reinterpret_cast<uint16_t *>(box_it.ptr()); + const auto b0 = dequantize_qasymm16(*ptr, boxes_qinfo); + const auto b1 = dequantize_qasymm16(*(ptr + 1), boxes_qinfo); + const auto b2 = dequantize_qasymm16(*(ptr + 2), boxes_qinfo); + const auto b3 = dequantize_qasymm16(*(ptr + 3), boxes_qinfo); + const float width = (b2 / scale_before) - (b0 / scale_before) + 1.f; + const float height = (b3 / scale_before) - (b1 / scale_before) + 1.f; + const float ctr_x = (b0 / scale_before) + 0.5f * width; + const float ctr_y = (b1 / scale_before) + 0.5f * height; + for (size_t j = 0; j < num_classes; ++j) + { + // Extract deltas + const size_t delta_id = id.y() * deltas_width + 4u * j; + const float dx = dequantize_qasymm8(delta_ptr[delta_id], deltas_qinfo) / bbinfo.weights()[0]; + const float dy = dequantize_qasymm8(delta_ptr[delta_id + 1], deltas_qinfo) / bbinfo.weights()[1]; + float dw = dequantize_qasymm8(delta_ptr[delta_id + 2], deltas_qinfo) / bbinfo.weights()[2]; + float dh = dequantize_qasymm8(delta_ptr[delta_id + 3], deltas_qinfo) / bbinfo.weights()[3]; + // Clip dw and dh + dw = std::min(dw, bbinfo.bbox_xform_clip()); + dh = std::min(dh, bbinfo.bbox_xform_clip()); + // Determine the predictions + const float pred_ctr_x = dx * width + ctr_x; + const float pred_ctr_y = dy * height + ctr_y; + const float pred_w = std::exp(dw) * width; + const float pred_h = std::exp(dh) * height; + // Store the prediction into the output tensor + pred_ptr[delta_id] = quantize_qasymm16( + scale_after * utility::clamp<float>(pred_ctr_x - 0.5f * pred_w, 0.f, img_w - 1.f), pred_qinfo); + pred_ptr[delta_id + 1] = quantize_qasymm16( + scale_after * utility::clamp<float>(pred_ctr_y - 0.5f * pred_h, 0.f, img_h - 1.f), pred_qinfo); + pred_ptr[delta_id + 2] = quantize_qasymm16( + scale_after * utility::clamp<float>(pred_ctr_x + 0.5f * pred_w - offset, 0.f, img_w - 1.f), + pred_qinfo); + pred_ptr[delta_id + 3] = quantize_qasymm16( + scale_after * utility::clamp<float>(pred_ctr_y + 0.5f * pred_h - offset, 0.f, img_h - 1.f), + pred_qinfo); + } + }, + box_it); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/boundingboxtransform/generic/neon/impl.h b/src/cpu/kernels/boundingboxtransform/generic/neon/impl.h index 7f990396df..d8013c6227 100644 --- a/src/cpu/kernels/boundingboxtransform/generic/neon/impl.h +++ b/src/cpu/kernels/boundingboxtransform/generic/neon/impl.h @@ -30,7 +30,11 @@ namespace arm_compute namespace cpu { template <typename T> -void bounding_box_transform(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window) +void bounding_box_transform(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window) { const size_t num_classes = deltas->info()->tensor_shape()[0] >> 2; const size_t deltas_width = deltas->info()->tensor_shape()[0]; @@ -46,44 +50,53 @@ void bounding_box_transform(const ITensor *boxes, ITensor *pred_boxes, const ITe auto delta_ptr = reinterpret_cast<T *>(deltas->buffer() + deltas->info()->offset_first_element_in_bytes()); Iterator box_it(boxes, window); - execute_window_loop(window, [&](const Coordinates & id) - { - const auto ptr = reinterpret_cast<T *>(box_it.ptr()); - const auto b0 = *ptr; - const auto b1 = *(ptr + 1); - const auto b2 = *(ptr + 2); - const auto b3 = *(ptr + 3); - const T width = (b2 / scale_before) - (b0 / scale_before) + T(1.f); - const T height = (b3 / scale_before) - (b1 / scale_before) + T(1.f); - const T ctr_x = (b0 / scale_before) + T(0.5f) * width; - const T ctr_y = (b1 / scale_before) + T(0.5f) * height; - for(size_t j = 0; j < num_classes; ++j) + execute_window_loop( + window, + [&](const Coordinates &id) { - // Extract deltas - const size_t delta_id = id.y() * deltas_width + 4u * j; - const T dx = delta_ptr[delta_id] / T(bbinfo.weights()[0]); - const T dy = delta_ptr[delta_id + 1] / T(bbinfo.weights()[1]); - T dw = delta_ptr[delta_id + 2] / T(bbinfo.weights()[2]); - T dh = delta_ptr[delta_id + 3] / T(bbinfo.weights()[3]); - // Clip dw and dh - dw = std::min(dw, T(bbinfo.bbox_xform_clip())); - dh = std::min(dh, T(bbinfo.bbox_xform_clip())); - // Determine the predictions - const T pred_ctr_x = dx * width + ctr_x; - const T pred_ctr_y = dy * height + ctr_y; - const T pred_w = std::exp(dw) * width; - const T pred_h = std::exp(dh) * height; - // Store the prediction into the output tensor - pred_ptr[delta_id] = scale_after * utility::clamp<T>(pred_ctr_x - T(0.5f) * pred_w, T(0), T(img_w - 1)); - pred_ptr[delta_id + 1] = scale_after * utility::clamp<T>(pred_ctr_y - T(0.5f) * pred_h, T(0), T(img_h - 1)); - pred_ptr[delta_id + 2] = scale_after * utility::clamp<T>(pred_ctr_x + T(0.5f) * pred_w - offset, T(0), T(img_w - 1)); - pred_ptr[delta_id + 3] = scale_after * utility::clamp<T>(pred_ctr_y + T(0.5f) * pred_h - offset, T(0), T(img_h - 1)); - } - }, - box_it); + const auto ptr = reinterpret_cast<T *>(box_it.ptr()); + const auto b0 = *ptr; + const auto b1 = *(ptr + 1); + const auto b2 = *(ptr + 2); + const auto b3 = *(ptr + 3); + const T width = (b2 / scale_before) - (b0 / scale_before) + T(1.f); + const T height = (b3 / scale_before) - (b1 / scale_before) + T(1.f); + const T ctr_x = (b0 / scale_before) + T(0.5f) * width; + const T ctr_y = (b1 / scale_before) + T(0.5f) * height; + for (size_t j = 0; j < num_classes; ++j) + { + // Extract deltas + const size_t delta_id = id.y() * deltas_width + 4u * j; + const T dx = delta_ptr[delta_id] / T(bbinfo.weights()[0]); + const T dy = delta_ptr[delta_id + 1] / T(bbinfo.weights()[1]); + T dw = delta_ptr[delta_id + 2] / T(bbinfo.weights()[2]); + T dh = delta_ptr[delta_id + 3] / T(bbinfo.weights()[3]); + // Clip dw and dh + dw = std::min(dw, T(bbinfo.bbox_xform_clip())); + dh = std::min(dh, T(bbinfo.bbox_xform_clip())); + // Determine the predictions + const T pred_ctr_x = dx * width + ctr_x; + const T pred_ctr_y = dy * height + ctr_y; + const T pred_w = std::exp(dw) * width; + const T pred_h = std::exp(dh) * height; + // Store the prediction into the output tensor + pred_ptr[delta_id] = scale_after * utility::clamp<T>(pred_ctr_x - T(0.5f) * pred_w, T(0), T(img_w - 1)); + pred_ptr[delta_id + 1] = + scale_after * utility::clamp<T>(pred_ctr_y - T(0.5f) * pred_h, T(0), T(img_h - 1)); + pred_ptr[delta_id + 2] = + scale_after * utility::clamp<T>(pred_ctr_x + T(0.5f) * pred_w - offset, T(0), T(img_w - 1)); + pred_ptr[delta_id + 3] = + scale_after * utility::clamp<T>(pred_ctr_y + T(0.5f) * pred_h - offset, T(0), T(img_h - 1)); + } + }, + box_it); } -void bounding_box_transform_qsymm16(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window); +void bounding_box_transform_qsymm16(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window); } // namespace cpu } // namespace arm_compute #endif //define SRC_CORE_SVE_KERNELS_BOUNDINGBOXTRANFORM_IMPL_H diff --git a/src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp b/src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp index b27c187df3..64ef815195 100644 --- a/src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp +++ b/src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp @@ -26,7 +26,11 @@ namespace arm_compute { namespace cpu { -void neon_qu16_boundingboxtransform(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, BoundingBoxTransformInfo bbinfo, const Window &window) +void neon_qu16_boundingboxtransform(const ITensor *boxes, + ITensor *pred_boxes, + const ITensor *deltas, + BoundingBoxTransformInfo bbinfo, + const Window &window) { return bounding_box_transform_qsymm16(boxes, pred_boxes, deltas, bbinfo, window); } |