diff options
Diffstat (limited to 'src/cpu/kernels/scale/neon/qasymm8_signed.cpp')
-rw-r--r-- | src/cpu/kernels/scale/neon/qasymm8_signed.cpp | 206 |
1 files changed, 135 insertions, 71 deletions
diff --git a/src/cpu/kernels/scale/neon/qasymm8_signed.cpp b/src/cpu/kernels/scale/neon/qasymm8_signed.cpp index cd63dfba63..5a885178a7 100644 --- a/src/cpu/kernels/scale/neon/qasymm8_signed.cpp +++ b/src/cpu/kernels/scale/neon/qasymm8_signed.cpp @@ -28,9 +28,16 @@ namespace arm_compute { namespace { -void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, - BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, - bool align_corners, const Window &window) +void qasymm8_signed_neon_scale_bilinear(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + const ITensor *dx, + const ITensor *dy, + BorderMode border_mode, + PixelValue constant_border_value, + float sampling_offset, + bool align_corners, + const Window &window) { // Data layout is NHWC const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform(); @@ -40,10 +47,12 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const const int32_t input_height = src->info()->dimension(2); // Compute the ratio between source and destination dimensions - const float scale_x = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), align_corners); - const float scale_y = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); + const float scale_x = + scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), align_corners); + const float scale_y = + scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); - if(border_mode == BorderMode::CONSTANT) + if (border_mode == BorderMode::CONSTANT) { const int32_t in_stride_y = src->info()->strides_in_bytes()[1]; const int32_t in_stride_z = src->info()->strides_in_bytes()[2]; @@ -58,7 +67,7 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const win_in.set(1, Window::Dimension(0, 0, 0)); win_in.set(2, Window::Dimension(0, 0, 0)); - for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d) + for (size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d) { win_off.set(d, Window::Dimension(0, 0, 0)); } @@ -67,36 +76,41 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const Iterator out(dst, window); const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>()); - execute_window_loop(window, [&](const Coordinates & id) - { - const int32_t index_h = std::floor((id[2] + sampling_offset) * scale_y - sampling_offset); - const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2])))); - const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2])))); - const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2])))); - const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr()); - - const auto a00 = (0 <= index_w && index_w < input_width && 0 <= index_h && index_h < input_height) ? - (*(pixel_row_ptr + index_w * in_stride_y + index_h * in_stride_z)) : - const_border_value; - const auto a01 = (-1 <= index_w && index_w + 1 < input_width && 0 <= index_h && index_h < input_height) ? - (*(pixel_row_ptr + (index_w + 1) * in_stride_y + index_h * in_stride_z)) : - const_border_value; - const auto a10 = (0 <= index_w && index_w < input_width && -1 <= index_h && index_h < input_height - 1) ? - (*(pixel_row_ptr + index_w * in_stride_y + (index_h + 1) * in_stride_z)) : - const_border_value; - const auto a11 = (-1 <= index_w && index_w < input_width - 1 && -1 <= index_h && index_h < input_height - 1) ? - (*(pixel_row_ptr + (index_w + 1) * in_stride_y + (index_h + 1) * in_stride_z)) : - const_border_value; - - const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info); - const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info); - const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info); - const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info); - *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info); - }, - in, out); + execute_window_loop( + window, + [&](const Coordinates &id) + { + const int32_t index_h = std::floor((id[2] + sampling_offset) * scale_y - sampling_offset); + const int32_t index_w = + *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2])))); + const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2])))); + const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2])))); + const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr()); + + const auto a00 = (0 <= index_w && index_w < input_width && 0 <= index_h && index_h < input_height) + ? (*(pixel_row_ptr + index_w * in_stride_y + index_h * in_stride_z)) + : const_border_value; + const auto a01 = (-1 <= index_w && index_w + 1 < input_width && 0 <= index_h && index_h < input_height) + ? (*(pixel_row_ptr + (index_w + 1) * in_stride_y + index_h * in_stride_z)) + : const_border_value; + const auto a10 = (0 <= index_w && index_w < input_width && -1 <= index_h && index_h < input_height - 1) + ? (*(pixel_row_ptr + index_w * in_stride_y + (index_h + 1) * in_stride_z)) + : const_border_value; + const auto a11 = + (-1 <= index_w && index_w < input_width - 1 && -1 <= index_h && index_h < input_height - 1) + ? (*(pixel_row_ptr + (index_w + 1) * in_stride_y + (index_h + 1) * in_stride_z)) + : const_border_value; + + const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info); + const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info); + const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info); + const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info); + *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize( + scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info); + }, + in, out); } - else if(border_mode == BorderMode::REPLICATE) + else if (border_mode == BorderMode::REPLICATE) { using FloatTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>; using Int32TagType = typename wrapper::traits::neon_bitvector_tag_t<int32_t, wrapper::traits::BitWidth::W128>; @@ -140,12 +154,12 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const const float32x4_t invvscale_o = wrapper::vdup_n(1.f / oq_info.scale, FloatTagType{}); const float32x4_t voffset_o = vdupq_n_f32(oq_info.offset); - for(int bo = bo_start; bo < bo_end; bo += bo_step) + for (int bo = bo_start; bo < bo_end; bo += bo_step) { const int8_t *in_ptr = reinterpret_cast<int8_t *>(in.ptr() + bo * in_stride_b); int8_t *out_ptr = reinterpret_cast<int8_t *>(out.ptr() + bo * out_stride_b); - for(int yo = yo_start; yo < yo_end; yo += yo_step) + for (int yo = yo_start; yo < yo_end; yo += yo_step) { // Floating-point coordinate const float yi_f = yo * scale_y + fp_coord_offset_y; @@ -162,7 +176,7 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const const int8_t *in_ptr_yi1 = in_ptr + yi1 * in_stride_y; int8_t *out_ptr_yo = out_ptr + yo * out_stride_y; - for(int xo = xo_start; xo < xo_end; xo += xo_step) + for (int xo = xo_start; xo < xo_end; xo += xo_step) { // Floating-point coordinate const float xi_f = xo * scale_x + fp_coord_offset_x; @@ -193,7 +207,7 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const int8_t *out_ptr_xo_yo = out_ptr_yo + xo * out_stride_x; int cout = 0; - for(; cout <= (out_dim_ch - step_cout); cout += step_cout) + for (; cout <= (out_dim_ch - step_cout); cout += step_cout) { const auto in00 = wrapper::vloadq(in_ptr_xi0_yi0 + cout * sizeof(int8_t)); const auto in01 = wrapper::vloadq(in_ptr_xi1_yi0 + cout * sizeof(int8_t)); @@ -203,34 +217,70 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const const int16x8_t in00_low = wrapper::vmovl(wrapper::vgetlow(in00)); const int16x8_t in00_high = wrapper::vmovl(wrapper::vgethigh(in00)); - const auto in00_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in00_low)), voffset_in)), vscale_in); - const auto in00_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in00_low)), voffset_in)), vscale_in); - const auto in00_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in00_high)), voffset_in)), vscale_in); - const auto in00_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in00_high)), voffset_in)), vscale_in); + const auto in00_0 = wrapper::vmul( + wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in00_low)), voffset_in)), + vscale_in); + const auto in00_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgethigh(in00_low)), voffset_in)), + vscale_in); + const auto in00_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgetlow(in00_high)), voffset_in)), + vscale_in); + const auto in00_3 = + wrapper::vmul(wrapper::vcvt<float>( + wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in00_high)), voffset_in)), + vscale_in); const int16x8_t in01_low = wrapper::vmovl(wrapper::vgetlow(in01)); const int16x8_t in01_high = wrapper::vmovl(wrapper::vgethigh(in01)); - const auto in01_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in01_low)), voffset_in)), vscale_in); - const auto in01_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in01_low)), voffset_in)), vscale_in); - const auto in01_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in01_high)), voffset_in)), vscale_in); - const auto in01_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in01_high)), voffset_in)), vscale_in); + const auto in01_0 = wrapper::vmul( + wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in01_low)), voffset_in)), + vscale_in); + const auto in01_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgethigh(in01_low)), voffset_in)), + vscale_in); + const auto in01_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgetlow(in01_high)), voffset_in)), + vscale_in); + const auto in01_3 = + wrapper::vmul(wrapper::vcvt<float>( + wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in01_high)), voffset_in)), + vscale_in); const int16x8_t in10_low = wrapper::vmovl(wrapper::vgetlow(in10)); const int16x8_t in10_high = wrapper::vmovl(wrapper::vgethigh(in10)); - const auto in10_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in10_low)), voffset_in)), vscale_in); - const auto in10_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in10_low)), voffset_in)), vscale_in); - const auto in10_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in10_high)), voffset_in)), vscale_in); - const auto in10_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in10_high)), voffset_in)), vscale_in); + const auto in10_0 = wrapper::vmul( + wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in10_low)), voffset_in)), + vscale_in); + const auto in10_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgethigh(in10_low)), voffset_in)), + vscale_in); + const auto in10_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgetlow(in10_high)), voffset_in)), + vscale_in); + const auto in10_3 = + wrapper::vmul(wrapper::vcvt<float>( + wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in10_high)), voffset_in)), + vscale_in); const int16x8_t in11_low = wrapper::vmovl(wrapper::vgetlow(in11)); const int16x8_t in11_high = wrapper::vmovl(wrapper::vgethigh(in11)); - const auto in11_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in11_low)), voffset_in)), vscale_in); - const auto in11_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in11_low)), voffset_in)), vscale_in); - const auto in11_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in11_high)), voffset_in)), vscale_in); - const auto in11_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in11_high)), voffset_in)), vscale_in); + const auto in11_0 = wrapper::vmul( + wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in11_low)), voffset_in)), + vscale_in); + const auto in11_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgethigh(in11_low)), voffset_in)), + vscale_in); + const auto in11_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub( + wrapper::vmovl(wrapper::vgetlow(in11_high)), voffset_in)), + vscale_in); + const auto in11_3 = + wrapper::vmul(wrapper::vcvt<float>( + wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in11_high)), voffset_in)), + vscale_in); auto out_0 = wrapper::vmul(in00_0, s00); out_0 = wrapper::vmla(out_0, in01_0, s01); @@ -263,14 +313,16 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const const auto out_2_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_2, invvscale_o)); const auto out_3_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_3, invvscale_o)); #endif // defined(__aarch64__) && !defined(BARE_METAL) - const auto low_part = wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_0_int), wrapper::vqmovn(out_1_int))); - const auto high_part = wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_2_int), wrapper::vqmovn(out_3_int))); - const auto out = wrapper::vcombine(low_part, high_part); + const auto low_part = + wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_0_int), wrapper::vqmovn(out_1_int))); + const auto high_part = + wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_2_int), wrapper::vqmovn(out_3_int))); + const auto out = wrapper::vcombine(low_part, high_part); wrapper::vstore(out_ptr_xo_yo + cout * sizeof(int8_t), out); } - for(; cout < out_dim_ch; ++cout) + for (; cout < out_dim_ch; ++cout) { const int8_t in00 = *(in_ptr_xi0_yi0 + cout * sizeof(int8_t)); const int8_t in01 = *(in_ptr_xi1_yi0 + cout * sizeof(int8_t)); @@ -291,7 +343,8 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const #if defined(__aarch64__) && !defined(BARE_METAL) *(out_ptr_xo_yo + cout * sizeof(int8_t)) = quantize_qasymm8_signed(out, oq_info); #else // defined(__aarch64__) && !defined(BARE_METAL) - *(out_ptr_xo_yo + cout * sizeof(int8_t)) = quantize_qasymm8_signed(out, oq_info, RoundingPolicy::TO_ZERO); + *(out_ptr_xo_yo + cout * sizeof(int8_t)) = + quantize_qasymm8_signed(out, oq_info, RoundingPolicy::TO_ZERO); #endif // defined(__aarch64__) && !defined(BARE_METAL) } } @@ -303,28 +356,39 @@ void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ARM_COMPUTE_ERROR("Not implemented"); } } -} +} // namespace namespace cpu { -void qasymm8_signed_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, - InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, - bool align_corners, const Window &window) +void qasymm8_signed_neon_scale(const ITensor *src, + ITensor *dst, + const ITensor *offsets, + const ITensor *dx, + const ITensor *dy, + InterpolationPolicy policy, + BorderMode border_mode, + PixelValue constant_border_value, + float sampling_offset, + bool align_corners, + const Window &window) { - if(policy == InterpolationPolicy::BILINEAR) + if (policy == InterpolationPolicy::BILINEAR) { - if(src->info()->quantization_info() == dst->info()->quantization_info() && border_mode == BorderMode::REPLICATE) + if (src->info()->quantization_info() == dst->info()->quantization_info() && + border_mode == BorderMode::REPLICATE) { - s8_neon_scale(src, dst, offsets, dx, dy, policy, border_mode, constant_border_value, sampling_offset, align_corners, window); + s8_neon_scale(src, dst, offsets, dx, dy, policy, border_mode, constant_border_value, sampling_offset, + align_corners, window); } else { - qasymm8_signed_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window); + qasymm8_signed_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, + sampling_offset, align_corners, window); } } - else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) + else if (policy == InterpolationPolicy::NEAREST_NEIGHBOR) { nearest_neon_scale<int8_t>(src, dst, offsets, sampling_offset, align_corners, window); } } } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute |