diff options
Diffstat (limited to 'src/core/cpu/kernels/pooling/neon/quantized.h')
| -rw-r--r-- | src/core/cpu/kernels/pooling/neon/quantized.h | 863 |
1 files changed, 863 insertions, 0 deletions
diff --git a/src/core/cpu/kernels/pooling/neon/quantized.h b/src/core/cpu/kernels/pooling/neon/quantized.h new file mode 100644 index 0000000000..81fb777ef2 --- /dev/null +++ b/src/core/cpu/kernels/pooling/neon/quantized.h @@ -0,0 +1,863 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef SRC_CORE_NEON_KERNELS_QUANTIZED_H +#define SRC_CORE_NEON_KERNELS_QUANTIZED_H + +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEFixedPoint.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include <arm_neon.h> + +namespace arm_compute +{ +namespace cpu +{ +template <typename T> +inline typename std::enable_if<std::is_same<T, int8_t>::value, int8_t>::type +quantize(float val, const UniformQuantizationInfo &info) +{ + return quantize_qasymm8_signed(val, info); +} + +template <typename T> +inline typename std::enable_if<std::is_same<T, uint8_t>::value, uint8_t>::type +quantize(float val, const UniformQuantizationInfo &info) +{ + return quantize_qasymm8(val, info); +} + +template <typename T> +inline T vcvtq_q32_f32(float32x4_t values); + +template <> +inline uint32x4_t vcvtq_q32_f32(float32x4_t values) +{ + return vcvtq_u32_f32(values); +} + +template <> +inline int32x4_t vcvtq_q32_f32(float32x4_t values) +{ + return vcvtq_s32_f32(values); +} + +template <typename T> +inline float32x4_t vcvtq_f32_q32(T values); + +template <> +inline float32x4_t vcvtq_f32_q32(uint32x4_t values) +{ + return vcvtq_f32_u32(values); +} + +template <> +inline float32x4_t vcvtq_f32_q32(int32x4_t values) +{ + return vcvtq_f32_s32(values); +} + +template <typename Tout> +inline Tout vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset); + +template <> +inline uint8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) +{ + const float new_scale = quant_rescale / scale_pooling; + return vquantize(acc, UniformQuantizationInfo(new_scale, new_offset)); +} + +template <> +inline int8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) +{ + const float new_scale = quant_rescale / scale_pooling; + return vquantize_signed(acc, UniformQuantizationInfo(new_scale, new_offset)); +} + +template <typename Tin, typename Tout> +inline Tout vrequantize_pooling(Tin vec1, Tin vec2, const UniformQuantizationInfo &requant_qinfo); + +template <> +inline uint8x16_t vrequantize_pooling(uint8x8_t vec1, uint8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x4_t acc = + { + { + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec1))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec1))))), + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec2))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec2))))), + } + }; + return vquantize(acc, requant_qinfo); +} + +template <> +inline int8x16_t vrequantize_pooling(int8x8_t vec1, int8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x4_t acc = + { + { + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec1))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec1))))), + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec2))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec2))))), + } + }; + return vquantize_signed(acc, requant_qinfo); +} + +template <typename T> +inline T vrequantize_pooling(T &vec, const UniformQuantizationInfo &requant_qinfo); + +template <> +inline uint8x8_t vrequantize_pooling(uint8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x2_t acc = + { + { + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec))))), + } + }; + return vquantize(acc, requant_qinfo); +} + +template <> +inline int8x8_t vrequantize_pooling(int8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x2_t acc = + { + { + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec))))), + } + }; + return vquantize_signed(acc, requant_qinfo); +} + +inline float calculate_avg_scale(bool exclude_padding, DataLayout data_layout, const Coordinates &id, const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) +{ + const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + + int start_x = id[idx_width] * stride_x - pad_x; + int start_y = id[idx_height] * stride_y - pad_y; + + const int end_x = std::min(start_x + pool_size_x, upper_bound_w); + const int end_y = std::min(start_y + pool_size_y, upper_bound_h); + if(exclude_padding) + { + start_x = std::max(0, start_x); + start_y = std::max(0, start_y); + } + return 1.f / ((end_y - start_y) * (end_x - start_x)); +} + +template <typename T> +void poolingMxN_q8_neon_nhwc(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) +{ + ARM_COMPUTE_UNUSED(dst1); + + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16; + const int window_half_step_x = window_step_x / 2; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, window_src); + Iterator out(dst0, window_out); + + using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; + using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; + using q16_t = typename wrapper::traits::promote_t<T>; + using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; + using q32_t = typename wrapper::traits::promote_t<q16_t>; + using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type; + + const int pool_size_x = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.width; + const int pool_size_y = pool_info.is_global_pooling ? src->info()->tensor_shape().z() : pool_info.pool_size.height; + const int pool_pad_right = pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = pool_info.pad_stride_info.pad_bottom(); + + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); + const int upper_bound_w = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src->info()->dimension(2) + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + + const float32x4_t half_scale_v = vdupq_n_f32(0.5f); + const UniformQuantizationInfo src_qinfo = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo dst_qinfo = dst0->info()->quantization_info().uniform(); + + const float quant_rescale = dst_qinfo.scale / src_qinfo.scale; + // "new_offset" doesn't have to consider the "half_scale_v" in its computation + // With a requantization performed in a single step there won't be uncertainties introduced + const int32_t new_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / quant_rescale); + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + if(pool_info.pool_type != PoolingType::MAX) + { + q32x4_t vres1 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres2 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres3 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres4 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); + + // Calculate scale + const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + // Perform pooling + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x16_t data = wrapper::vloadq(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().z())) + x_off); + + const q16x8_t data_q16 = wrapper::vmovl(wrapper::vgetlow(data)); + const q16x8_t data2_q16 = wrapper::vmovl(wrapper::vgethigh(data)); + vres1 = wrapper::vadd(vres1, wrapper::vmovl(wrapper::vgetlow(data_q16))); + vres2 = wrapper::vadd(vres2, wrapper::vmovl(wrapper::vgethigh(data_q16))); + vres3 = wrapper::vadd(vres3, wrapper::vmovl(wrapper::vgetlow(data2_q16))); + vres4 = wrapper::vadd(vres4, wrapper::vmovl(wrapper::vgethigh(data2_q16))); + } + } + + if(src_qinfo != dst_qinfo) + { + const float32x4x4_t vres = + { + { + vcvtq_f32_q32(vres1), + vcvtq_f32_q32(vres2), + vcvtq_f32_q32(vres3), + vcvtq_f32_q32(vres4), + } + }; + const auto requantized_dst = vrequantize_pooling_with_scale<q8x16_t>(vres, quant_rescale, scale, new_offset); + // Store result + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, wrapper::vgetlow(requantized_dst)); + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off + 8, wrapper::vgethigh(requantized_dst)); + } + else + { + const float32x4_t scale_v = vdupq_n_f32(scale); + // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero + vres1 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v)); + vres2 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v)); + vres3 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v)); + vres4 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres4), scale_v)); + + const q8x8_t res1 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres1), wrapper::vmovn(vres2))); + const q8x8_t res2 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres3), wrapper::vmovn(vres4))); + // Store result + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, res1); + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off + 8, res2); + } + } + else + { + q8x16_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_128_tag{}); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x16_t data = wrapper::vloadq(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().z())) + x_off); + vres = wrapper::vmax(vres, data); + } + } + + // Store result + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(vres), wrapper::vgethigh(vres), + requant_qinfo) : + vres); + } + } + + if(pool_info.pool_type == PoolingType::MAX) + { + for(; x_off <= (window_end_x - window_half_step_x); x_off += window_half_step_x) + { + q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{}); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().z())) + x_off); + vres = wrapper::vmax(vres, data); + } + } + + // Store result + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, + (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t>(vres, requant_qinfo) : vres); + } + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + if(pool_info.pool_type != PoolingType::MAX) + { + q32_t res = static_cast<q32_t>(0.f); + + // Calculate scale + const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + // Perform pooling + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().z())) + x_off); + res += data; + } + } + + if(src_qinfo != dst_qinfo) + { + const float res_f = static_cast<float>(res); + const float new_scale = quant_rescale / scale; + const auto requantized_dst = quantize<T>(res_f, UniformQuantizationInfo(new_scale, new_offset)); + + // Store result + *(reinterpret_cast<T *>(out.ptr()) + x_off) = requantized_dst; + } + else + { + // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero + res = static_cast<T>(0.5f + static_cast<float>(res) * scale); + + // Store result + *(reinterpret_cast<T *>(out.ptr()) + x_off) = res; + } + } + else + { + T res = std::numeric_limits<T>::min(); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().z())) + x_off); + res = std::max(res, data); + } + } + + // Store result + if(src_qinfo != dst_qinfo) + { + const float res_f = static_cast<float>(res); + *(reinterpret_cast<T *>(out.ptr()) + x_off) = quantize<T>(res_f, requant_qinfo); + } + else + { + *(reinterpret_cast<T *>(out.ptr()) + x_off) = res; + } + } + } + + }, + in, out); +} + +#if defined(ENABLE_NCHW_KERNELS) +template <typename T, typename TVec> +inline void scale_vector_q16x8(bool exclude_padding, TVec &v, const Coordinates &id, int id_offset, int step, + const int pool_size, const int upper_bound_w, const int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) +{ + int start_x = (id.x() + id_offset) * stride_x - pad_x; + int start_y = id.y() * stride_y - pad_y; + const int end_y = std::min(start_y + pool_size, upper_bound_h); + if(exclude_padding) + { + start_y = std::max(0, start_y); + } + + std::array<T, 8> elems = + { + { + wrapper::vgetlane(v, 0), + wrapper::vgetlane(v, 1), + wrapper::vgetlane(v, 2), + wrapper::vgetlane(v, 3), + wrapper::vgetlane(v, 4), + wrapper::vgetlane(v, 5), + wrapper::vgetlane(v, 6), + wrapper::vgetlane(v, 7), + } + }; + + for(auto &el : elems) + { + int c_start_x = start_x; + const int end_x = std::min(c_start_x + pool_size, upper_bound_w); + if(exclude_padding) + { + c_start_x = std::max(0, c_start_x); + } + float scale = 1.f / ((end_y - start_y) * (end_x - c_start_x)); + el *= scale; + start_x += step * stride_x; + } + + v = wrapper::vsetlane(elems[0], v, 0); + v = wrapper::vsetlane(elems[1], v, 1); + v = wrapper::vsetlane(elems[2], v, 2); + v = wrapper::vsetlane(elems[3], v, 3); + v = wrapper::vsetlane(elems[4], v, 4); + v = wrapper::vsetlane(elems[5], v, 5); + v = wrapper::vsetlane(elems[6], v, 6); + v = wrapper::vsetlane(elems[7], v, 7); +} + +template <typename T> +void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) +{ + ARM_COMPUTE_UNUSED(dst1); + Iterator in(src, window_src); + Iterator out(dst0, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; + using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; + using q8x8x2_t = typename std::conditional<std::is_same<T, uint8_t>::value, uint8x8x2_t, int8x8x2_t>::type; + using q16_t = typename wrapper::traits::promote_t<T>; + using q16x4_t = typename wrapper::traits::neon_vector<q16_t, 4>::type; + using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; + using q16x8x2_t = typename wrapper::traits::neon_vector<q16_t, 16>::type; + + constexpr int pool_size = 2; + int pool_stride_x = 0; + int pool_stride_y = 0; + const int pool_pad_right = pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = pool_info.pad_stride_info.pad_bottom(); + std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); + const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + + const T *const src_top_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)))); + const T *const src_bottom_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1))); + + const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; + + const UniformQuantizationInfo src_qinfo = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo dst_qinfo = dst0->info()->quantization_info().uniform(); + const bool have_different_qinfo = src_qinfo != dst_qinfo; + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto top_data = wrapper::vloadq(src_top_ptr + in.offset()); + const auto bottom_data = wrapper::vloadq(src_bottom_ptr + in.offset()); + q8x8_t lower_res = {}; + q8x8_t upper_res = {}; + + if(pool_info.pool_type != PoolingType::MAX) + { + const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; + const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; + + // Add rows + const q16x8x2_t vrsum = + { + { + wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), + wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), + } + }; + + // Pair-wise add row data + const q16x4_t vpsum_1 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[0]), wrapper::vgethigh(vrsum.val[0])); + const q16x4_t vpsum_2 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[1]), wrapper::vgethigh(vrsum.val[1])); + + q16x8_t res_lower = wrapper::vcombine(vpsum_1, vpsum_2); + + // Scale lower result + scale_vector_q16x8<q16_t, q16x8_t>(pool_info.exclude_padding, res_lower, id, 0, scale_step_x, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + lower_res = wrapper::vmovn(res_lower); + + // Compute upper result for stride_x == 1 + if(pool_stride_x == 1) + { + // Shifted row sum + const q16x8x2_t vrsum_shifted = + { + { + wrapper::vext_1(vrsum.val[0], vrsum.val[1]), + wrapper::vext_1(vrsum.val[1], vrsum.val[1]) + } + }; + + // Pair-wise add shifted row + q16x8_t res_upper = wrapper::vcombine( + wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[0]), wrapper::vgethigh(vrsum_shifted.val[0])), + wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[1]), wrapper::vgethigh(vrsum_shifted.val[1]))); + + // Scale upper result + scale_vector_q16x8<q16_t, q16x8_t>(pool_info.exclude_padding, res_upper, id, 1, 2, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + upper_res = wrapper::vmovn(res_upper); + } + } + else + { + const q8x16_t max_data = wrapper::vmax(top_data, bottom_data); + lower_res = wrapper::vpmax(wrapper::vgetlow(max_data), wrapper::vgethigh(max_data)); + if(pool_stride_x == 1) + { + const q8x16_t max_data_shifted = wrapper::vext_1(max_data, max_data); + upper_res = wrapper::vpmax(wrapper::vgetlow(max_data_shifted), wrapper::vgethigh(max_data_shifted)); + } + } + + if(have_different_qinfo) + { + const auto requantized_dst = vrequantize_pooling<q8x8_t, q8x16_t>(lower_res, upper_res, requant_qinfo); + lower_res = wrapper::vgetlow(requantized_dst); + upper_res = wrapper::vgethigh(requantized_dst); + } + + // Store result + if(pool_stride_x == 1) + { + const q8x8x2_t res = { { lower_res, upper_res } }; + wrapper::vstore(reinterpret_cast<T *>(out.ptr()), res); + } + else + { + wrapper::vstore(reinterpret_cast<T *>(out.ptr()), lower_res); + } + }, + in, out); +} + +template <typename T> +void pooling3_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) +{ + ARM_COMPUTE_UNUSED(dst1); + Iterator in(src, window_src); + Iterator out(dst0, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; + using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; + using q8x8x2_t = typename std::conditional<std::is_same<T, uint8_t>::value, uint8x8x2_t, int8x8x2_t>::type; + using q16_t = typename wrapper::traits::promote_t<T>; + using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; + using q16x8x2_t = typename wrapper::traits::neon_vector<q16_t, 16>::type; + + constexpr int pool_size = 3; + const int pool_pad_right = pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); + const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + + const UniformQuantizationInfo &src_qinfo = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo &dst_qinfo = dst0->info()->quantization_info().uniform(); + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + const T *const src_top_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)))); + const T *const src_middle_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1))); + const T *const src_bottom_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2))); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto top_data = wrapper::vloadq(src_top_ptr + in.offset()); + const auto middle_data = wrapper::vloadq(src_middle_ptr + in.offset()); + const auto bottom_data = wrapper::vloadq(src_bottom_ptr + in.offset()); + q8x8_t fres = {}; + q8x16_t fqres = {}; + + if(pool_info.pool_type == PoolingType::AVG) + { + // Convert data to u16 + const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; + const q16x8x2_t middle_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(middle_data)), wrapper::vmovl(wrapper::vgethigh(middle_data)) } }; + const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; + + // Calculate row sums + const q16x8x2_t vrsum = + { + { + wrapper::vadd(wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), middle_data_q16.val[0]), + wrapper::vadd(wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), middle_data_q16.val[1]), + } + }; + const q16x8x2_t vrsum_shifted_1 = + { + { + wrapper::vext_1(vrsum.val[0], vrsum.val[1]), + wrapper::vext_1(vrsum.val[1], vrsum.val[1]) + } + }; + const q16x8x2_t vrsum_shifted_2 = + { + { + wrapper::vext_2(vrsum.val[0], vrsum.val[1]), + wrapper::vext_2(vrsum.val[1], vrsum.val[1]) + } + }; + // Calculate final sum + q16x8x2_t final_sum = + { + { + wrapper::vadd(wrapper::vadd(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), + wrapper::vadd(wrapper::vadd(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), + } + }; + if(pool_stride_x == 2) + { + q16x8_t res = + { + wrapper::vgetlane(final_sum.val[0], 0), + wrapper::vgetlane(final_sum.val[0], 2), + wrapper::vgetlane(final_sum.val[0], 4), + wrapper::vgetlane(final_sum.val[0], 6), + wrapper::vgetlane(final_sum.val[1], 0), + wrapper::vgetlane(final_sum.val[1], 2), + wrapper::vgetlane(final_sum.val[1], 4), + wrapper::vgetlane(final_sum.val[1], 6), + }; + + scale_vector_q16x8<q16_t, q16x8_t>(pool_info.exclude_padding, res, id, 0, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + fres = wrapper::vmovn(res); + } + else + { + // Scale lower result + scale_vector_q16x8<q16_t, q16x8_t>(pool_info.exclude_padding, final_sum.val[0], id, 0, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + // Scale lower result + scale_vector_q16x8<q16_t, q16x8_t>(pool_info.exclude_padding, final_sum.val[1], id, 8, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + fqres = wrapper::vcombine(wrapper::vmovn(final_sum.val[0]), wrapper::vmovn(final_sum.val[1])); + } + } + else + { + const q8x16_t max_data = wrapper::vmax(wrapper::vmax(top_data, bottom_data), middle_data); + const q8x16_t max_data_shift1 = wrapper::vext_1(max_data, max_data); + const q8x16_t max_data_shift2 = wrapper::vext_2(max_data, max_data); + const q8x16_t final_max = wrapper::vmax(wrapper::vmax(max_data, max_data_shift1), max_data_shift2); + + if(pool_stride_x == 2) + { + const q8x8x2_t table = { { wrapper::vgetlow(final_max), wrapper::vgethigh(final_max) } }; + static const q8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; + fres = wrapper::vtbl(table, lookup_val); + } + else + { + fqres = final_max; + } + } + + // Store result + if(pool_stride_x == 1) + { + if(src_qinfo != dst_qinfo) + { + fqres = vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo); + } + wrapper::vstore(reinterpret_cast<T *>(out.ptr()), fqres); + } + else + { + if(src_qinfo != dst_qinfo) + { + fres = vrequantize_pooling<q8x8_t>(fres, requant_qinfo); + } + wrapper::vstore(reinterpret_cast<T *>(out.ptr()), fres); + } + }, + in, out); +} + +template <typename T> +void poolingMxN_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) +{ + ARM_COMPUTE_UNUSED(dst1); + Iterator in(src, window_src); + Iterator out(dst0, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; + using q16_t = typename wrapper::traits::promote_t<T>; + using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; + using q32_t = typename wrapper::traits::promote_t<q16_t>; + using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type; + + const int pool_size_x = pool_info.is_global_pooling ? src->info()->tensor_shape().x() : pool_info.pool_size.width; + const int pool_size_y = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.height; + const int pool_pad_right = pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); + const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + + const UniformQuantizationInfo &src_qinfo = src->info()->quantization_info().uniform(); + const UniformQuantizationInfo &dst_qinfo = dst0->info()->quantization_info().uniform(); + + execute_window_loop(window, [&](const Coordinates & id) + { + T res = std::numeric_limits<T>::min(); + + if(pool_info.pool_type != PoolingType::MAX) + { + q32x4_t vres = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); + q32_t sres = 0; + + // Calculate scale + const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + // Perform pooling + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().y()))); + + const q16x8_t data_q16 = wrapper::vmovl(data); + vres = wrapper::vadd(vres, wrapper::vaddl(wrapper::vgethigh(data_q16), wrapper::vgetlow(data_q16))); + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().y()))); + sres += data; + } + } + + // Reduction + const auto tmp = wrapper::vpadd(wrapper::vgethigh(vres), wrapper::vgetlow(vres)); + sres += wrapper::vgetlane(tmp, 0) + wrapper::vgetlane(tmp, 1); + + // Divide by scale + res = static_cast<T>(support::cpp11::round(sres * scale)); + } + else + { + q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{}); + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().y()))); + vres = wrapper::vmax(vres, data); + } + // Leftover for loop + for(; x < pool_size_x; ++x) + { + const T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> + (src->info()->strides_in_bytes().y()))); + res = std::max(res, data); + } + } + + // Reduce max + vres = wrapper::vpmax(vres, vres); + vres = wrapper::vpmax(vres, vres); + vres = wrapper::vpmax(vres, vres); + + // Get max value + res = std::max(res, wrapper::vgetlane(vres, 0)); + } + // Store result + res = (src_qinfo != dst_qinfo) ? Qasymm8QuantizationHelper<T>::quantize(Qasymm8QuantizationHelper<T>::dequantize(res, src_qinfo), dst_qinfo) : res; + *(reinterpret_cast<T *>(out.ptr())) = res; + }, + in, out); +} +#endif /* defined(ENABLE_NCHW_KERNELS) */ +} // namespace cpu +} // namespace arm_compute + +#endif // SRC_CORE_NEON_KERNELS_QUANTIZED_H
\ No newline at end of file |