From c4f2743951473f8d97f5a43767fdbb31a4df967c Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Sun, 11 Sep 2022 15:59:19 +0100 Subject: =?UTF-8?q?Optimize=20Quantized/Integer=20Bilinear=20Scale=20for?= =?UTF-8?q?=20Neon=E2=84=A2?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This patch introduces several performance optimizations regarding the Bilinear Scale operator with REPLICATE Border mode. Changes apply only to NHWC. This patch - Reduces the memory footprint by disabling precomputation of indices and weights when they're not used - Rewrites the kernels for QASYMM8/QASYMM8_SIGNED/U8(Uint8) - Adds S8(Int8) Bilinear Scale for Border mode REPLICATE - Removes Bilinear Scale SVE kernels for Quantized and Integer types and adjust the heuristics to choose the Neon™ implementation - Adds new test cases where the input and output of the Bilinear Scale operator have different quantization scale and offset Resolves: COMPMID-5453, COMPMID-5454 Change-Id: I3d251e76e0c6978fd5a0a1795ec62ab536bec93c Signed-off-by: Gunes Bayir Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8250 Reviewed-by: SiCong Li Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- src/cpu/kernels/scale/neon/qasymm8.cpp | 298 ++++++++++++++++++++++++++------- 1 file changed, 242 insertions(+), 56 deletions(-) (limited to 'src/cpu/kernels/scale/neon/qasymm8.cpp') diff --git a/src/cpu/kernels/scale/neon/qasymm8.cpp b/src/cpu/kernels/scale/neon/qasymm8.cpp index daa157ec3d..778459ae39 100644 --- a/src/cpu/kernels/scale/neon/qasymm8.cpp +++ b/src/cpu/kernels/scale/neon/qasymm8.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,6 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ +#include "src/core/helpers/ScaleHelpers.h" #include "src/cpu/kernels/scale/neon/list.h" namespace arm_compute @@ -32,56 +33,60 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor bool align_corners, const Window &window) { // Data layout is NHWC - // 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); - Window win_off; - win_off.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_off.set(Window::DimY, Window::Dimension(0, 0, 0)); - - // Don't increment in X and Y 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(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) - { - win_off.set(d, Window::Dimension(0, 0, 0)); - } - - Iterator in(src, win_in); - Iterator out(dst, window); - - const int32_t in_dim_w = src->info()->dimension(1); - const int32_t in_dim_h = src->info()->dimension(2); - const int32_t stride_w = src->info()->strides_in_bytes()[1]; - const int32_t stride_h = src->info()->strides_in_bytes()[2]; + const int32_t input_width = src->info()->dimension(1); + const int32_t input_height = src->info()->dimension(2); const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform(); const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform(); + // 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); + 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]; + + // Compute the ratio between source height and destination height + Window win_off; + win_off.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_off.set(Window::DimY, Window::Dimension(0, 0, 0)); + + // Don't increment in X and Y 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(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) + { + win_off.set(d, Window::Dimension(0, 0, 0)); + } + + Iterator in(src, win_in); + Iterator out(dst, window); + const uint8_t const_border_value = static_cast(constant_border_value.get()); execute_window_loop(window, [&](const Coordinates & id) { - const int32_t index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset); + const int32_t index_h = std::floor((id[2] + sampling_offset) * scale_y - sampling_offset); const int32_t index_w = *(reinterpret_cast(offsets->ptr_to_element(Coordinates(id[1], id[2])))); const auto dx_val = *(reinterpret_cast(dx->ptr_to_element(Coordinates(id[1], id[2])))); const auto dy_val = *(reinterpret_cast(dy->ptr_to_element(Coordinates(id[1], id[2])))); const auto pixel_row_ptr = reinterpret_cast(in.ptr()); - const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ? - (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) : + 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 < in_dim_w && 0 <= index_h && index_h < in_dim_h) ? - (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) : + 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 < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ? - (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) : + 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 < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ? - (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) : + 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::dequantize(a00, iq_info); @@ -94,31 +99,205 @@ void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor } else if(border_mode == BorderMode::REPLICATE) { - execute_window_loop(window, [&](const Coordinates & id) + using FloatTagType = typename wrapper::traits::neon_bitvector_tag_t; + using Int32TagType = typename wrapper::traits::neon_bitvector_tag_t; + + const int in_stride_x = src->info()->strides_in_bytes()[1]; + const int in_stride_y = src->info()->strides_in_bytes()[2]; + const int in_stride_b = src->info()->strides_in_bytes()[3]; + const int out_stride_x = dst->info()->strides_in_bytes()[1]; + const int out_stride_y = dst->info()->strides_in_bytes()[2]; + const int out_stride_b = dst->info()->strides_in_bytes()[3]; + const int out_dim_ch = dst->info()->dimension(0); + constexpr int step_cout = 16; + + 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[1].start(); + const int xo_end = window_execution[1].end(); + const int xo_step = window_execution[1].step(); + const int yo_start = window_execution[2].start(); + const int yo_end = window_execution[2].end(); + const int yo_step = window_execution[2].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(); + + const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform(); + + const float32x4_t vscale_in = wrapper::vdup_n(iq_info.scale, FloatTagType{}); + const int32x4_t voffset_in = wrapper::vdup_n(iq_info.offset, Int32TagType{}); // Offsets will be Int32 + + 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); + + const float fp_coord_offset_y = sampling_offset * (scale_y - 1); + const float fp_coord_offset_x = sampling_offset * (scale_x - 1); + + for(int bo = bo_start; bo < bo_end; bo += bo_step) { - const int index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset); - const int32_t index_w = *(reinterpret_cast(offsets->ptr_to_element(Coordinates(id[1], id[2])))); - const auto dx_val = *(reinterpret_cast(dx->ptr_to_element(Coordinates(id[1], id[2])))); - const auto dy_val = *(reinterpret_cast(dy->ptr_to_element(Coordinates(id[1], id[2])))); - const auto pixel_row_ptr = reinterpret_cast(in.ptr()); + const uint8_t *in_ptr = in.ptr() + bo * in_stride_b; + uint8_t *out_ptr = out.ptr() + bo * out_stride_b; - auto clamped_w = utility::clamp(index_w, 0, in_dim_w - 1); - auto clamped_w1 = utility::clamp(index_w + 1, 0, in_dim_w - 1); - auto clamped_h = utility::clamp(index_h, 0, in_dim_h - 1); - auto clamped_h1 = utility::clamp(index_h + 1, 0, in_dim_h - 1); + 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; + // Integer coordinate + const int yi = static_cast(std::floor(yi_f)); + // Weight for the y coordinate + const float a1 = (yi_f - static_cast(yi)); + const float b1 = (1.f - a1); - const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h); - const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h); - const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h); - const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h); + const int yi0 = utility::clamp(yi, 0, input_height - 1); + const int yi1 = utility::clamp(yi + 1, 0, input_height - 1); - const float inp00 = Qasymm8QuantizationHelper::dequantize(a00, iq_info); - const float inp01 = Qasymm8QuantizationHelper::dequantize(a01, iq_info); - const float inp10 = Qasymm8QuantizationHelper::dequantize(a10, iq_info); - const float inp11 = Qasymm8QuantizationHelper::dequantize(a11, iq_info); - *reinterpret_cast(out.ptr()) = Qasymm8QuantizationHelper::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info); - }, - in, out); + const uint8_t *in_ptr_yi0 = in_ptr + yi0 * in_stride_y; + const uint8_t *in_ptr_yi1 = in_ptr + yi1 * in_stride_y; + + uint8_t *out_ptr_yo = out_ptr + yo * out_stride_y; + 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; + // Integer coordinate + const int xi = static_cast(std::floor(xi_f)); + // Weight for the x coordinate + const float a = (xi_f - static_cast(xi)); + const float b = (1.f - a); + + const float s00_s = b * b1; + const float s01_s = a * b1; + const float s10_s = b * a1; + const float s11_s = a * a1; + + const auto s00 = wrapper::vdup_n(s00_s, FloatTagType{}); + const auto s01 = wrapper::vdup_n(s01_s, FloatTagType{}); + const auto s10 = wrapper::vdup_n(s10_s, FloatTagType{}); + const auto s11 = wrapper::vdup_n(s11_s, FloatTagType{}); + + const int xi0 = utility::clamp(xi, 0, input_width - 1); + const int xi1 = utility::clamp(xi + 1, 0, input_width - 1); + + const auto in_ptr_xi0_yi0 = in_ptr_yi0 + xi0 * in_stride_x; + const auto in_ptr_xi1_yi0 = in_ptr_yi0 + xi1 * in_stride_x; + const auto in_ptr_xi0_yi1 = in_ptr_yi1 + xi0 * in_stride_x; + const auto in_ptr_xi1_yi1 = in_ptr_yi1 + xi1 * in_stride_x; + + uint8_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) + { + const auto in00 = wrapper::vloadq(in_ptr_xi0_yi0 + cout * sizeof(uint8_t)); + const auto in01 = wrapper::vloadq(in_ptr_xi1_yi0 + cout * sizeof(uint8_t)); + const auto in10 = wrapper::vloadq(in_ptr_xi0_yi1 + cout * sizeof(uint8_t)); + const auto in11 = wrapper::vloadq(in_ptr_xi1_yi1 + cout * sizeof(uint8_t)); + + const uint16x8_t in00_low = wrapper::vmovl(wrapper::vgetlow(in00)); + const uint16x8_t in00_high = wrapper::vmovl(wrapper::vgethigh(in00)); + + const auto in00_0 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in00_low))), voffset_in)), vscale_in); + const auto in00_1 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in00_low))), voffset_in)), vscale_in); + const auto in00_2 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in00_high))), voffset_in)), vscale_in); + const auto in00_3 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in00_high))), voffset_in)), vscale_in); + + const uint16x8_t in01_low = wrapper::vmovl(wrapper::vgetlow(in01)); + const uint16x8_t in01_high = wrapper::vmovl(wrapper::vgethigh(in01)); + + const auto in01_0 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in01_low))), voffset_in)), vscale_in); + const auto in01_1 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in01_low))), voffset_in)), vscale_in); + const auto in01_2 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in01_high))), voffset_in)), vscale_in); + const auto in01_3 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in01_high))), voffset_in)), vscale_in); + + const uint16x8_t in10_low = wrapper::vmovl(wrapper::vgetlow(in10)); + const uint16x8_t in10_high = wrapper::vmovl(wrapper::vgethigh(in10)); + + const auto in10_0 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in10_low))), voffset_in)), vscale_in); + const auto in10_1 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in10_low))), voffset_in)), vscale_in); + const auto in10_2 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in10_high))), voffset_in)), vscale_in); + const auto in10_3 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in10_high))), voffset_in)), vscale_in); + + const uint16x8_t in11_low = wrapper::vmovl(wrapper::vgetlow(in11)); + const uint16x8_t in11_high = wrapper::vmovl(wrapper::vgethigh(in11)); + + const auto in11_0 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in11_low))), voffset_in)), vscale_in); + const auto in11_1 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(in11_low))), voffset_in)), vscale_in); + const auto in11_2 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(in11_high))), voffset_in)), vscale_in); + const auto in11_3 = wrapper::vmul(wrapper::vcvt(wrapper::vsub(wrapper::vreinterpret(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); + out_0 = wrapper::vmla(out_0, in10_0, s10); + out_0 = wrapper::vmla(out_0, in11_0, s11); + + auto out_1 = wrapper::vmul(in00_1, s00); + out_1 = wrapper::vmla(out_1, in01_1, s01); + out_1 = wrapper::vmla(out_1, in10_1, s10); + out_1 = wrapper::vmla(out_1, in11_1, s11); + + auto out_2 = wrapper::vmul(in00_2, s00); + out_2 = wrapper::vmla(out_2, in01_2, s01); + out_2 = wrapper::vmla(out_2, in10_2, s10); + out_2 = wrapper::vmla(out_2, in11_2, s11); + + auto out_3 = wrapper::vmul(in00_3, s00); + out_3 = wrapper::vmla(out_3, in01_3, s01); + out_3 = wrapper::vmla(out_3, in10_3, s10); + out_3 = wrapper::vmla(out_3, in11_3, s11); + +#if defined(__aarch64__) && !defined(BARE_METAL) + const auto out_0_int = wrapper::vcvta(wrapper::vmla(voffset_o, out_0, invvscale_o)); + const auto out_1_int = wrapper::vcvta(wrapper::vmla(voffset_o, out_1, invvscale_o)); + const auto out_2_int = wrapper::vcvta(wrapper::vmla(voffset_o, out_2, invvscale_o)); + const auto out_3_int = wrapper::vcvta(wrapper::vmla(voffset_o, out_3, invvscale_o)); +#else // defined(__aarch64__) && !defined(BARE_METAL) + const auto out_0_int = wrapper::vcvt(wrapper::vmla(voffset_o, out_0, invvscale_o)); + const auto out_1_int = wrapper::vcvt(wrapper::vmla(voffset_o, out_1, invvscale_o)); + const auto out_2_int = wrapper::vcvt(wrapper::vmla(voffset_o, out_2, invvscale_o)); + const auto out_3_int = wrapper::vcvt(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); + + wrapper::vstore(out_ptr_xo_yo + cout * sizeof(uint8_t), out); + } + + for(; cout < out_dim_ch; ++cout) + { + const uint8_t in00 = *(in_ptr_xi0_yi0 + cout * sizeof(uint8_t)); + const uint8_t in01 = *(in_ptr_xi1_yi0 + cout * sizeof(uint8_t)); + const uint8_t in10 = *(in_ptr_xi0_yi1 + cout * sizeof(uint8_t)); + const uint8_t in11 = *(in_ptr_xi1_yi1 + cout * sizeof(uint8_t)); + + const float in00_f = (static_cast(in00) - iq_info.offset) * iq_info.scale; + const float in01_f = (static_cast(in01) - iq_info.offset) * iq_info.scale; + const float in10_f = (static_cast(in10) - iq_info.offset) * iq_info.scale; + const float in11_f = (static_cast(in11) - iq_info.offset) * iq_info.scale; + + float out = in00_f * s00_s; + out += in01_f * s01_s; + out += in10_f * s10_s; + out += in11_f * s11_s; + + // Rounding modes of vector and scalar loops should match +#if defined(__aarch64__) && !defined(BARE_METAL) + *(out_ptr_xo_yo + cout * sizeof(uint8_t)) = quantize_qasymm8(out, oq_info); +#else // defined(__aarch64__) && !defined(BARE_METAL) + *(out_ptr_xo_yo + cout * sizeof(uint8_t)) = quantize_qasymm8(out, oq_info, RoundingPolicy::TO_ZERO); +#endif // defined(__aarch64__) && !defined(BARE_METAL) + } + } + } + } } else { @@ -134,7 +313,14 @@ void qasymm8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets { if(policy == InterpolationPolicy::BILINEAR) { - qasymm8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window); + if(src->info()->quantization_info() == dst->info()->quantization_info()) + { + u8_neon_scale(src, dst, offsets, dx, dy, policy, border_mode, constant_border_value, sampling_offset, align_corners, window); + } + else + { + qasymm8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window); + } } else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) { -- cgit v1.2.1