From 8b8405aec8af17b04205b60094680751abfdc94a Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Fri, 1 Oct 2021 17:48:02 +0100 Subject: Optimize CpuScale NHWC F32/F16 - Rework CpuScaleKernel F32/F16 NHWC - bilinear - Rework CpuScaleKernel F32/F16 NHWC - nearest - Add test to validate the vector computation path Resolves COMPMID-4801, COMPMID-4802 Change-Id: Ie6e4f262a8cce509edd7b8f564c940758625c58a Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6361 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Pablo Marquez Tello --- src/cpu/kernels/CpuScaleKernel.cpp | 4 +- src/cpu/kernels/scale/neon/list.h | 377 +++++++++++++++++++++++++++++-------- 2 files changed, 305 insertions(+), 76 deletions(-) (limited to 'src/cpu') diff --git a/src/cpu/kernels/CpuScaleKernel.cpp b/src/cpu/kernels/CpuScaleKernel.cpp index 1108c7a78e..3063d8f682 100644 --- a/src/cpu/kernels/CpuScaleKernel.cpp +++ b/src/cpu/kernels/CpuScaleKernel.cpp @@ -123,12 +123,12 @@ static const ScaleKernel available_kernels[] = { "neon_u8_scale", [](const ScaleSelectorData & data) { return data.dt == DataType::U8; }, - REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale) + REGISTER_INTEGER_NEON(arm_compute::cpu::u8_neon_scale) }, { "neon_s16_scale", [](const ScaleSelectorData & data) { return data.dt == DataType::S16; }, - REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale) + REGISTER_INTEGER_NEON(arm_compute::cpu::s16_neon_scale) }, #endif /* defined(ARM_COMPUTE_ENABLE_NEON) */ }; diff --git a/src/cpu/kernels/scale/neon/list.h b/src/cpu/kernels/scale/neon/list.h index c91242f5b2..9679f161e7 100644 --- a/src/cpu/kernels/scale/neon/list.h +++ b/src/cpu/kernels/scale/neon/list.h @@ -42,6 +42,8 @@ namespace cpu InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, \ bool align_corners, const Window &window) +DECLARE_SCALE_KERNEL(s16_neon_scale); +DECLARE_SCALE_KERNEL(u8_neon_scale); DECLARE_SCALE_KERNEL(qasymm8_neon_scale); DECLARE_SCALE_KERNEL(qasymm8_signed_neon_scale); @@ -51,43 +53,90 @@ template void nearest_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, float sampling_offset, bool align_corners, const Window &window) { - const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; - const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom; - const size_t in_stride_wc = in_stride_w * in_stride_c; - const size_t in_dim_h = src->info()->dimension(2); + ARM_COMPUTE_UNUSED(offsets); - // Compute the ratio between source height and destination height - const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners); - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - const int window_step_x = 16 / sizeof(T); + // 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); - Window win(window); - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - Iterator out(dst, win); + const int in_stride_y = src->info()->strides_in_bytes()[1]; + const int in_stride_z = src->info()->strides_in_bytes()[2]; + const int in_stride_w = src->info()->strides_in_bytes()[3]; + const int out_stride_y = dst->info()->strides_in_bytes()[1]; + const int out_stride_z = dst->info()->strides_in_bytes()[2]; + const int out_stride_w = dst->info()->strides_in_bytes()[3]; + const int out_dim_ch = dst->info()->dimension(0); + const int step_cout = 16 / sizeof(T); - const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); - const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3]; + Window window_execution = window; + window_execution.set(Window::DimX, Window::Dimension(0, 1, 1)); + Window win_in_out(window); + win_in_out.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_in_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); + Iterator in(src, win_in_out); + Iterator out(dst, win_in_out); - execute_window_loop(win, [&](const Coordinates & id) + const int xo_start = window_execution.y().start(); + const int xo_end = window_execution.y().end(); + const int xo_step = window_execution.y().step(); + const int yo_start = window_execution.z().start(); + const int yo_end = window_execution.z().end(); + const int yo_step = window_execution.z().step(); + const int bo_start = window_execution[3].start(); + const int bo_end = window_execution[3].end(); + const int bo_step = window_execution[3].step(); + + for(int bo = bo_start; bo < bo_end; bo += bo_step) { - const int32_t offset = *reinterpret_cast(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; - const auto in_hi = static_cast(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr)); - const int offset_row = in_hi * in_stride_wc; - int32_t x = window_start_x; - const T *in_ptr = reinterpret_cast(in_ptr_start + in_stride_bytes_hwc * id[3]); + const uint8_t *in_ptr_base = in.ptr() + bo * in_stride_w; + uint8_t *out_ptr_base = out.ptr() + bo * out_stride_w; - for(; x <= window_end_x - window_step_x; x += window_step_x) - { - wrapper::vstore(reinterpret_cast(out.ptr()) + x, - wrapper::vloadq(in_ptr + offset + offset_row + x)); - } - for(; x < window_end_x; ++x) + for(int yo = yo_start; yo < yo_end; yo += yo_step) { - *(reinterpret_cast(out.ptr()) + x) = *(in_ptr + offset + offset_row + x); + // Floating-point coordinate + float yi_f = ((yo + sampling_offset) * scale_y); + int yi = 0; + if(align_corners) + { + yi = utils::rounding::round_half_away_from_zero(yi_f); + } + else + { + yi = static_cast(std::floor(yi_f)); + } + + for(int xo = xo_start; xo < xo_end; xo += xo_step) + { + // Floating-point coordinate + float xi_f = ((xo + sampling_offset) * scale_x); + int xi = 0; + if(align_corners) + { + xi = utils::rounding::round_half_away_from_zero(xi_f); + } + else + { + xi = static_cast(std::floor(xi_f)); + } + + const uint8_t *in_ptr = in_ptr_base + xi * in_stride_y + yi * in_stride_z; + uint8_t *out_ptr = out_ptr_base + xo * out_stride_y + yo * out_stride_z; + + int cout = 0; + for(; cout <= (out_dim_ch - step_cout); cout += step_cout) + { + auto out0 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T))); + wrapper::vstore(reinterpret_cast(out_ptr + cout * sizeof(T)), out0); + } + + for(; cout < out_dim_ch; ++cout) + { + auto out0 = *(reinterpret_cast(in_ptr + cout * sizeof(T))); + *(reinterpret_cast(out_ptr + cout * sizeof(T))) = out0; + } + } } - }, - out); + } } template @@ -95,21 +144,43 @@ void bilinear_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offset BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window) { - // Compute the ratio between source height and destination height - const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); + ARM_COMPUTE_UNUSED(offsets); + ARM_COMPUTE_UNUSED(dx); + ARM_COMPUTE_UNUSED(dy); + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + + // 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); - Iterator out(dst, window); - const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; + const int in_stride_y = src->info()->strides_in_bytes()[1]; + const int in_stride_z = src->info()->strides_in_bytes()[2]; + const int in_stride_w = src->info()->strides_in_bytes()[3]; + const int out_stride_y = dst->info()->strides_in_bytes()[1]; + const int out_stride_z = dst->info()->strides_in_bytes()[2]; + const int out_stride_w = dst->info()->strides_in_bytes()[3]; const int in_dim_w = src->info()->dimension(1); const int in_dim_h = src->info()->dimension(2); - const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom); + const int out_dim_ch = dst->info()->dimension(0); + const int step_cout = 16 / sizeof(T); - // Don't increment in Y and Z direction for the input tensor - // A pointer to the start of this plane is needed as base for the precomputed offsets - Window win_in(window); - win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); - win_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); - Iterator in(src, win_in); + Window window_execution = window; + window_execution.set(Window::DimX, Window::Dimension(0, 1, 1)); + Window win_in_out(window); + win_in_out.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_in_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); + Iterator in(src, win_in_out); + Iterator out(dst, win_in_out); + + const int xo_start = window_execution.y().start(); + const int xo_end = window_execution.y().end(); + const int xo_step = window_execution.y().step(); + const int yo_start = window_execution.z().start(); + const int yo_end = window_execution.z().end(); + const int yo_step = window_execution.z().step(); + const int bo_start = window_execution[3].start(); + const int bo_end = window_execution[3].end(); + const int bo_step = window_execution[3].step(); if(border_mode == BorderMode::CONSTANT) { @@ -119,45 +190,203 @@ void bilinear_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offset using ConstType = T; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const T const_border_value = static_cast(constant_border_value.get()); - execute_window_loop(window, [&](const Coordinates & id) + + for(int bo = bo_start; bo < bo_end; bo += bo_step) { - const auto offset = *reinterpret_cast(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dx_val = *reinterpret_cast(dx->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dy_val = *reinterpret_cast(dy->ptr_to_element(Coordinates(id.y(), id.z()))); - const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); - const T *in_ptr = reinterpret_cast(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc; - - const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value; - const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value; - const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value; - const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value; - - *reinterpret_cast(out.ptr()) = static_cast(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); - }, - in, out); + const uint8_t *in_ptr_base = in.ptr() + bo * in_stride_w; + uint8_t *out_ptr_base = out.ptr() + bo * out_stride_w; + + for(int yo = yo_start; yo < yo_end; yo += yo_step) + { + // Floating-point coordinate + const float yi_f = ((yo + sampling_offset) * scale_y - sampling_offset); + // Integer coordinate + const auto yi = static_cast(std::floor(yi_f)); + // Weight for the y coordinate + const auto a1 = (yi_f - static_cast(yi)); + const auto b1 = (1.f - a1); + + for(int xo = xo_start; xo < xo_end; xo += xo_step) + { + // Floating-point coordinate + const float xi_f = ((xo + sampling_offset) * scale_x - sampling_offset); + // Integer coordinate + const auto xi = static_cast(std::floor(xi_f)); + // Weight for the x coordinate + const auto a = (xi_f - static_cast(xi)); + const auto b = (1.f - a); + + const auto s00_s = static_cast(b * b1); + const auto s01_s = static_cast(a * b1); + const auto s10_s = static_cast(b * a1); + const auto s11_s = static_cast(a * a1); + + const uint8_t *in_ptr = in_ptr_base + xi * in_stride_y + yi * in_stride_z; + uint8_t *out_ptr = out_ptr_base + xo * out_stride_y + yo * out_stride_z; + + int cout = 0; + for(; cout <= (out_dim_ch - step_cout); cout += step_cout) + { + auto in00 = wrapper::vdup_n(static_cast(const_border_value), ExactTagType{}); + auto in01 = wrapper::vdup_n(static_cast(const_border_value), ExactTagType{}); + auto in10 = wrapper::vdup_n(static_cast(const_border_value), ExactTagType{}); + auto in11 = wrapper::vdup_n(static_cast(const_border_value), ExactTagType{}); + if((yi >= 0) && (yi < in_dim_h)) + { + if((xi >= 0) && (xi < in_dim_w)) + { + in00 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T))); + } + if(((xi + 1) >= 0) && ((xi + 1) < in_dim_w)) + { + in01 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_y)); + } + } + if(((yi + 1) >= 0) && ((yi + 1) < in_dim_h)) + { + if((xi >= 0) && (xi < in_dim_w)) + { + in10 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_z)); + } + if(((xi + 1) >= 0) && ((xi + 1) < in_dim_w)) + { + in11 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_y + in_stride_z)); + } + } + + const auto s00 = wrapper::vdup_n(s00_s, ExactTagType{}); + const auto s01 = wrapper::vdup_n(s01_s, ExactTagType{}); + const auto s10 = wrapper::vdup_n(s10_s, ExactTagType{}); + const auto s11 = wrapper::vdup_n(s11_s, ExactTagType{}); + auto out0 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + out0 = wrapper::vmla(out0, in00, s00); + out0 = wrapper::vmla(out0, in01, s01); + out0 = wrapper::vmla(out0, in10, s10); + out0 = wrapper::vmla(out0, in11, s11); + wrapper::vstore(reinterpret_cast(out_ptr + cout * sizeof(T)), out0); + } + + for(; cout < out_dim_ch; ++cout) + { + auto in00 = static_cast(const_border_value); + auto in01 = static_cast(const_border_value); + auto in10 = static_cast(const_border_value); + auto in11 = static_cast(const_border_value); + if((yi >= 0) && (yi < in_dim_h)) + { + if((xi >= 0) && (xi < in_dim_w)) + { + in00 = *(reinterpret_cast(in_ptr + cout * sizeof(T))); + } + if(((xi + 1) >= 0) && ((xi + 1) < in_dim_w)) + { + in01 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_y)); + } + } + if(((yi + 1) >= 0) && ((yi + 1) < in_dim_h)) + { + if((xi >= 0) && (xi < in_dim_w)) + { + in10 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_z)); + } + if(((xi + 1) >= 0) && ((xi + 1) < in_dim_w)) + { + in11 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + in_stride_y + in_stride_z)); + } + } + auto out0 = static_cast(0); + out0 += in00 * s00_s; + out0 += in01 * s01_s; + out0 += in10 * s10_s; + out0 += in11 * s11_s; + *(reinterpret_cast(out_ptr + cout * sizeof(T))) = out0; + } + } + } + } } else if(border_mode == BorderMode::REPLICATE) { - execute_window_loop(window, [&](const Coordinates & id) + for(int bo = bo_start; bo < bo_end; bo += bo_step) { - const auto offset = *reinterpret_cast(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dx_val = *reinterpret_cast(dx->ptr_to_element(Coordinates(id.y(), id.z()))); - const auto dy_val = *reinterpret_cast(dy->ptr_to_element(Coordinates(id.y(), id.z()))); - const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); - - auto clamped_w = utility::clamp(offset, 0, in_dim_w - 1); - auto clamped_w1 = utility::clamp(offset + 1, 0, in_dim_w - 1); - auto clamped_h = utility::clamp(in_hi, 0, in_dim_h - 1); - auto clamped_h1 = utility::clamp(in_hi + 1, 0, in_dim_h - 1); - - const auto a00 = *(reinterpret_cast(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc); - const auto a01 = *(reinterpret_cast(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc); - const auto a10 = *(reinterpret_cast(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc); - const auto a11 = *(reinterpret_cast(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc); - - *reinterpret_cast(out.ptr()) = static_cast(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); - }, - in, out); + const uint8_t *in_ptr = in.ptr() + bo * in_stride_w; + uint8_t *out_ptr = out.ptr() + bo * out_stride_w; + + for(int yo = yo_start; yo < yo_end; yo += yo_step) + { + // Floating-point coordinate + const float yi_f = ((yo + sampling_offset) * scale_y - sampling_offset); + // Integer coordinate + const auto yi = static_cast(std::floor(yi_f)); + // Weight for the y coordinate + const auto a1 = (yi_f - static_cast(yi)); + const auto b1 = (1.f - a1); + + const auto yi0 = utility::clamp(yi, 0, in_dim_h - 1); + const auto yi1 = utility::clamp(yi + 1, 0, in_dim_h - 1); + + for(int xo = xo_start; xo < xo_end; xo += xo_step) + { + // Floating-point coordinate + const float xi_f = ((xo + sampling_offset) * scale_x - sampling_offset); + // Integer coordinate + const auto xi = static_cast(std::floor(xi_f)); + // Weight for the x coordinate + const auto a = (xi_f - static_cast(xi)); + const auto b = (1.f - a); + + const auto s00_s = static_cast(b * b1); + const auto s01_s = static_cast(a * b1); + const auto s10_s = static_cast(b * a1); + const auto s11_s = static_cast(a * a1); + + const auto xi0 = utility::clamp(xi, 0, in_dim_w - 1); + const auto xi1 = utility::clamp(xi + 1, 0, in_dim_w - 1); + + int cout = 0; + for(; cout <= (out_dim_ch - step_cout); cout += step_cout) + { + auto in00 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + auto in01 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + auto in10 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + auto in11 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + in00 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi0) * in_stride_y + (yi0) * in_stride_z)); + in01 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi1) * in_stride_y + (yi0) * in_stride_z)); + in10 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi0) * in_stride_y + (yi1) * in_stride_z)); + in11 = wrapper::vloadq(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi1) * in_stride_y + (yi1) * in_stride_z)); + + const auto s00 = wrapper::vdup_n(s00_s, ExactTagType{}); + const auto s01 = wrapper::vdup_n(s01_s, ExactTagType{}); + const auto s10 = wrapper::vdup_n(s10_s, ExactTagType{}); + const auto s11 = wrapper::vdup_n(s11_s, ExactTagType{}); + auto out0 = wrapper::vdup_n(static_cast(0), ExactTagType{}); + out0 = wrapper::vmla(out0, in00, s00); + out0 = wrapper::vmla(out0, in01, s01); + out0 = wrapper::vmla(out0, in10, s10); + out0 = wrapper::vmla(out0, in11, s11); + wrapper::vstore(reinterpret_cast(out_ptr + cout * sizeof(T) + xo * out_stride_y + yo * out_stride_z), out0); + } + + for(; cout < out_dim_ch; ++cout) + { + auto in00 = static_cast(0); + auto in01 = static_cast(0); + auto in10 = static_cast(0); + auto in11 = static_cast(0); + in00 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi0) * in_stride_y + (yi0) * in_stride_z)); + in01 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi1) * in_stride_y + (yi0) * in_stride_z)); + in10 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi0) * in_stride_y + (yi1) * in_stride_z)); + in11 = *(reinterpret_cast(in_ptr + cout * sizeof(T) + (xi1) * in_stride_y + (yi1) * in_stride_z)); + auto out0 = static_cast(0); + out0 += in00 * s00_s; + out0 += in01 * s01_s; + out0 += in10 * s10_s; + out0 += in11 * s11_s; + *(reinterpret_cast(out_ptr + cout * sizeof(T) + xo * out_stride_y + yo * out_stride_z)) = out0; + } + } + } + } } else { -- cgit v1.2.1