diff options
Diffstat (limited to 'src/cpu/kernels/genproposals/generic/neon/impl.cpp')
| -rw-r--r-- | src/cpu/kernels/genproposals/generic/neon/impl.cpp | 43 |
1 files changed, 24 insertions, 19 deletions
diff --git a/src/cpu/kernels/genproposals/generic/neon/impl.cpp b/src/cpu/kernels/genproposals/generic/neon/impl.cpp index 9224e32a94..8cb76f3afb 100644 --- a/src/cpu/kernels/genproposals/generic/neon/impl.cpp +++ b/src/cpu/kernels/genproposals/generic/neon/impl.cpp @@ -28,7 +28,10 @@ class ITensor; class Window; namespace cpu { -void compute_all_anchors_qasymm16(const ITensor *anchors, ITensor *all_anchors, ComputeAnchorsInfo anchors_info, const Window &window) +void compute_all_anchors_qasymm16(const ITensor *anchors, + ITensor *all_anchors, + ComputeAnchorsInfo anchors_info, + const Window &window) { Iterator all_anchors_it(all_anchors, window); Iterator anchors_it(all_anchors, window); @@ -39,28 +42,30 @@ void compute_all_anchors_qasymm16(const ITensor *anchors, ITensor *all_anchors, const UniformQuantizationInfo qinfo = anchors->info()->quantization_info().uniform(); - execute_window_loop(window, [&](const Coordinates & id) - { - const size_t anchor_offset = id.y() % num_anchors; + execute_window_loop( + window, + [&](const Coordinates &id) + { + const size_t anchor_offset = id.y() % num_anchors; - 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 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; + 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; - 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; + 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); + *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); } } // namespace cpu } // namespace arm_compute |