/* * Copyright (c) 2017-2020 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. */ #include "arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h" #include "arm_compute/core/AccessWindowStatic.h" #include "arm_compute/core/CPP/Validate.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/NEAsymm.h" #include "arm_compute/core/NEON/NEFixedPoint.h" #include "arm_compute/core/NEON/NEMath.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "support/ToolchainSupport.h" #include "arm_compute/core/NEON/wrapper/wrapper.h" #include #include #include #include #include #include #include namespace arm_compute { using namespace misc::shape_calculator; namespace { 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 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 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); } Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, unsigned int &pooled_w, unsigned int pooled_h, const ITensorInfo *indices, Size2D pool_size) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); int pool_stride_x = 0; int pool_stride_y = 0; PoolingType pool_type = pool_info.pool_type; const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); if(indices) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method"); } ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_quantized(input->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(input->data_type()) && !pool_info.exclude_padding && (pool_info.pool_type == PoolingType::AVG) && pool_info.pad_stride_info.has_padding() && (input->data_layout() == DataLayout::NHWC), "exclude_padding equal false is not supported for AVG Pooling with padding on quantized types"); if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) || (output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); if(indices) { ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2"); ARM_COMPUTE_RETURN_ERROR_ON((indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) || (indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); } } return Status{}; } Status validate_arguments_pool_info(const unsigned int pool_size_x, const unsigned int pool_size_y) { ARM_COMPUTE_RETURN_ERROR_ON(pool_size_x == 0); ARM_COMPUTE_RETURN_ERROR_ON(pool_size_y == 0); return Status{}; } std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *indices, const PoolingLayerInfo &pool_info, unsigned int &num_elems_processed_per_iteration, BorderSize &border_size, unsigned int pooled_w, unsigned int pooled_h, int pool_size_x, int pool_size_y) { // Output auto inizialitation if not yet initialized auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_pool_shape(*input, pool_info))); if(indices) { // Indices auto inizialitation if not yet initialized auto_init_if_empty(*indices, (input->clone()->set_tensor_shape(compute_pool_shape(*input, pool_info))) .set_data_type(DataType::U32) /* we store the offset to the element */); } const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout; unsigned int num_elems_read_per_iteration = 0; unsigned int num_elems_horizontal_window = 0; int pool_stride_x = 0; int pool_stride_y = 0; const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const int input_width = input->dimension(idx_width); const int input_height = input->dimension(idx_height); const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); const int pool_pad_right = pad_stride_info.pad_right(); const int pool_pad_top = pad_stride_info.pad_top(); const int pool_pad_left = pad_stride_info.pad_left(); const int pool_pad_bottom = pad_stride_info.pad_bottom(); const bool is_square = pool_size_x == pool_size_y; // Check output dimensions std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(idx_width), input->dimension(idx_height), pool_size_x, pool_size_y, pad_stride_info); //If it's not squared and optimized will be executed the MxN num_elems_read_per_iteration = 1; num_elems_processed_per_iteration = 1; num_elems_horizontal_window = 1; const bool is_nhwc = data_layout == DataLayout::NHWC; if(is_square) { switch(input->data_type()) { case DataType::QASYMM8: case DataType::QASYMM8_SIGNED: if(is_nhwc) { num_elems_processed_per_iteration = 16; break; } switch(pool_size_x) { case 2: num_elems_read_per_iteration = 16; num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; break; case 3: num_elems_read_per_iteration = 16; num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; break; default: break; } break; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: if(is_nhwc) { num_elems_processed_per_iteration = 8; break; } switch(pool_size_x) { case 2: case 3: num_elems_read_per_iteration = 4; num_elems_processed_per_iteration = 1; num_elems_horizontal_window = 1; break; default: break; } break; #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::F32: if(is_nhwc) { num_elems_processed_per_iteration = 4; break; } switch(pool_size_x) { case 2: num_elems_read_per_iteration = 2; break; case 3: num_elems_read_per_iteration = 4; // We use vload4 for pooling3 break; case 7: num_elems_read_per_iteration = 8; // We use vload8 for pooling7 break; default: break; } num_elems_processed_per_iteration = 1; num_elems_horizontal_window = 1; break; default: ARM_COMPUTE_ERROR("Element size not supported"); break; } } else { if(is_nhwc) { num_elems_processed_per_iteration = 16 / input->element_size(); } } bool window_changed = false; Window win{}; if(data_layout == DataLayout::NCHW) { // Number of iterations in X dimension const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; // Upper limit for the number of right/bottom border elements that are accessed const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - input_width; const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - input_height; border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left); border_size.right = std::max(upper_bound_w, pool_pad_right); border_size.bottom = std::max(upper_bound_h, pool_pad_bottom); TensorShape output_shape{ input->tensor_shape() }; output_shape.set(0, pooled_w); output_shape.set(1, pooled_h); TensorInfo output_info(input->clone()->set_tensor_shape(output_shape)); win = calculate_max_window(output_info, Steps(num_elems_processed_per_iteration)); AccessWindowStatic input_access(input, -pool_pad_left, -pool_pad_top, input_width + border_size.right, input_height + border_size.bottom); AccessWindowHorizontal output_access(output, 0, num_elems_horizontal_window); if(indices) { AccessWindowHorizontal indices_access(indices, 0, num_elems_horizontal_window); window_changed = update_window_and_padding(win, input_access, output_access, indices_access); } else { window_changed = update_window_and_padding(win, input_access, output_access); } output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); } else { TensorShape output_shape{ input->tensor_shape() }; output_shape.set(1, pooled_w); output_shape.set(2, pooled_h); TensorInfo output_info(input->clone()->set_tensor_shape(output_shape)); win = calculate_max_window(output_info, Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); if(indices) { AccessWindowHorizontal indices_access(indices, 0, num_elems_processed_per_iteration); window_changed = update_window_and_padding(win, input_access, output_access, indices_access); } else { window_changed = update_window_and_padding(win, input_access, output_access); } output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); } Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; return std::make_pair(err, win); } template 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 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 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 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 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); } } // namespace NEPoolingLayerKernel::NEPoolingLayerKernel() : _func(nullptr), _input(nullptr), _output(nullptr), _indices(nullptr), _pool_info(), _data_layout(DataLayout::UNKNOWN), _num_elems_processed_per_iteration(0), _border_size(0), _is_square(false) { } BorderSize NEPoolingLayerKernel::border_size() const { return _border_size; } void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; const bool is_global_pooling = pool_info.is_global_pooling; const int pool_stride_x = pad_stride_info.stride().first; // Get data layout const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : pool_info.data_layout; const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); // Update pool size in case of global pooling const Size2D pool_size( is_global_pooling ? input->info()->dimension(idx_width) : pool_info.pool_size.width, is_global_pooling ? input->info()->dimension(idx_height) : pool_info.pool_size.height); // Validate pool info before calling scaled_dimensions ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_pool_info(pool_size.x(), pool_size.y())); // Check output dimensions unsigned int pooled_w; unsigned int pooled_h; std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(idx_width), input->info()->dimension(idx_height), pool_size.x(), pool_size.y(), pad_stride_info); // Perform validation step ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info, pooled_w, pooled_h, (indices) ? indices->info() : nullptr, pool_size)); // Set instance variables _input = input; _output = output; _indices = indices; _pool_info = pool_info; _data_layout = input->info()->data_layout(); _is_square = (pool_size.x() == pool_size.y()); // Get data type const DataType data_type = input->info()->data_type(); const bool is_nchw = _data_layout == DataLayout::NCHW; if(data_type == DataType::QASYMM8) { if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling2_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling3_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } else { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } } else if(data_type == DataType::QASYMM8_SIGNED) { if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling2_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling3_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } else { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; } } } else if(data_type == DataType::F16) { if(_is_square) { switch(pool_size.x()) { case 2: { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling2_f16_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f16_nhwc; } } break; case 3: { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling3_f16_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f16_nhwc; } } break; default: { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_f16_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f16_nhwc; } break; } break; } } else { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_f16_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f16_nhwc; } } } else if(data_type == DataType::F32) { if(_is_square) { switch(pool_size.x()) { case 2: { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling2_f32_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; } break; } case 3: { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling3_f32_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; } break; } case 7: { if(is_nchw) { _func = &NEPoolingLayerKernel::pooling7_f32_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; } break; } default: { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_f32_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; } break; } } } else { if(is_nchw) { _func = &NEPoolingLayerKernel::poolingMxN_f32_nchw; } else { _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; } } } // Configure kernel window auto win_config = validate_and_configure_window(input->info(), output->info(), (indices) ? indices->info() : nullptr, pool_info, _num_elems_processed_per_iteration, _border_size, pooled_w, pooled_h, pool_size.x(), pool_size.y()); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); INEKernel::configure(win_config.second); } template void NEPoolingLayerKernel::pooling2_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); /** NEON vector types */ using q8x8_t = typename wrapper::traits::neon_vector::type; using q8x16_t = typename wrapper::traits::neon_vector::type; using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; using q16_t = typename wrapper::traits::promote_t; using q16x4_t = typename wrapper::traits::neon_vector::type; using q16x8_t = typename wrapper::traits::neon_vector::type; using q16x8x2_t = typename wrapper::traits::neon_vector::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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const T *const input_top_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); const T *const input_bottom_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; const UniformQuantizationInfo input_qinfo = _input->info()->quantization_info().uniform(); const UniformQuantizationInfo output_qinfo = _output->info()->quantization_info().uniform(); const bool have_different_qinfo = input_qinfo != output_qinfo; const float requant_scale = output_qinfo.scale / input_qinfo.scale; const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_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(input_top_ptr + input.offset()); const auto bottom_data = wrapper::vloadq(input_bottom_ptr + input.offset()); q8x8_t lower_res = {}; q8x8_t upper_res = {}; if(pooling_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(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(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_output = vrequantize_pooling(lower_res, upper_res, requant_qinfo); lower_res = wrapper::vgetlow(requantized_output); upper_res = wrapper::vgethigh(requantized_output); } // Store result if(pool_stride_x == 1) { const q8x8x2_t res = { { lower_res, upper_res } }; wrapper::vstore(reinterpret_cast(output.ptr()), res); } else { wrapper::vstore(reinterpret_cast(output.ptr()), lower_res); } }, input, output); } void NEPoolingLayerKernel::pooling3_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { ARM_COMPUTE_UNUSED(pooling_type); ARM_COMPUTE_UNUSED(exclude_padding); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC Iterator input(_input, window_input); Iterator output(_output, window); constexpr const 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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); execute_window_loop(window, [&](const Coordinates & id) { float16x4_t top_data = vld1_f16(reinterpret_cast(input_top_ptr + input.offset())); float16x4_t middle_data = vld1_f16(reinterpret_cast(input_middle_ptr + input.offset())); float16x4_t bottom_data = vld1_f16(reinterpret_cast(input_bottom_ptr + input.offset())); float16x4_t res = {}; // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { top_data = vmul_f16(top_data, top_data); middle_data = vmul_f16(middle_data, middle_data); bottom_data = vmul_f16(bottom_data, bottom_data); } if(pooling_type != PoolingType::MAX) { // Calculate scale const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); const float16x4_t scale_v = vdup_n_f16(scale); // Perform pooling const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); res = vmul_f16(vpadd_f16(res, res), scale_v); } else { const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); res = vpmax_f16(vset_lane_f16(-std::numeric_limits::max(), max_data, 3), max_data); res = vpmax_f16(res, res); } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { res = vinv_f16(vinvsqrt_f16(res)); } *(reinterpret_cast(output.ptr())) = vget_lane_f16(res, 0); }, input, output); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ ARM_COMPUTE_UNUSED(window_input); ARM_COMPUTE_UNUSED(window); ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } void NEPoolingLayerKernel::pooling2_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { ARM_COMPUTE_UNUSED(pooling_type); ARM_COMPUTE_UNUSED(exclude_padding); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC Iterator input(_input, window_input); Iterator output(_output, window); constexpr int pool_size = 2; 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, pool_stride_y = 0; std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); execute_window_loop(window, [&](const Coordinates & id) { float16x4_t top_data = vld1_f16(reinterpret_cast(input_top_ptr + input.offset())); float16x4_t bottom_data = vld1_f16(reinterpret_cast(input_bottom_ptr + input.offset())); float16x4_t res = {}; // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { top_data = vmul_f16(top_data, top_data); bottom_data = vmul_f16(bottom_data, bottom_data); } if(pooling_type != PoolingType::MAX) { const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); const float16x4_t scale_v = vdup_n_f16(scale); const float16x4_t sum_data = vadd_f16(top_data, bottom_data); res = vmul_f16(vpadd_f16(sum_data, sum_data), scale_v); } else { const float16x4_t max_data = vmax_f16(top_data, bottom_data); res = vpmax_f16(max_data, max_data); } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { res = vinv_f16(vinvsqrt_f16(res)); } // Store result *(reinterpret_cast(output.ptr())) = vget_lane_f16(res, 0); }, input, output); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ ARM_COMPUTE_UNUSED(window_input); ARM_COMPUTE_UNUSED(window); ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } template void NEPoolingLayerKernel::pooling3_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); /** NEON vector types */ using q8x8_t = typename wrapper::traits::neon_vector::type; using q8x16_t = typename wrapper::traits::neon_vector::type; using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; using q16_t = typename wrapper::traits::promote_t; using q16x8_t = typename wrapper::traits::neon_vector::type; using q16x8x2_t = typename wrapper::traits::neon_vector::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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const UniformQuantizationInfo &input_qinfo = _input->info()->quantization_info().uniform(); const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform(); const float requant_scale = output_qinfo.scale / input_qinfo.scale; const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / requant_scale); const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); const T *const input_top_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); const T *const input_middle_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); const T *const input_bottom_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2))); execute_window_loop(window, [&](const Coordinates & id) { const auto top_data = wrapper::vloadq(input_top_ptr + input.offset()); const auto middle_data = wrapper::vloadq(input_middle_ptr + input.offset()); const auto bottom_data = wrapper::vloadq(input_bottom_ptr + input.offset()); q8x8_t fres = {}; q8x16_t fqres = {}; if(pooling_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(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(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(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(input_qinfo != output_qinfo) { fqres = vrequantize_pooling(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo); } wrapper::vstore(reinterpret_cast(output.ptr()), fqres); } else { if(input_qinfo != output_qinfo) { fres = vrequantize_pooling(fres, requant_qinfo); } wrapper::vstore(reinterpret_cast(output.ptr()), fres); } }, input, output); } void NEPoolingLayerKernel::poolingMxN_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { ARM_COMPUTE_UNUSED(pooling_type); ARM_COMPUTE_UNUSED(exclude_padding); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC Iterator input(_input, window_input); Iterator output(_output, window); const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); execute_window_loop(window, [&](const Coordinates & id) { float16_t res = 0.0f; float16x8_t vres = vdupq_n_f16(0.0f); if(pooling_type != PoolingType::MAX) { // Calculate scale const float scale = calculate_avg_scale(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 float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); // Get power of 2 in case of l2 pooling and accumulate if(pooling_type == PoolingType::L2) { vres = vaddq_f16(vres, vmulq_f16(data, data)); } else { vres = vaddq_f16(vres, data); } } // Leftover for loop for(; x < pool_size_x; ++x) { float16_t data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { data *= data; } res += data; } } // Reduction float16x4_t tmp = vpadd_f16(vget_high_f16(vres), vget_low_f16(vres)); res += vget_lane_f16(tmp, 0); res += vget_lane_f16(tmp, 1); res += vget_lane_f16(tmp, 2); res += vget_lane_f16(tmp, 3); // Divide by scale res *= scale; } else { float16x8_t vres = vdupq_n_f16(std::numeric_limits::lowest()); res = std::numeric_limits::lowest(); for(int y = 0; y < pool_size_y; ++y) { int x = 0; for(; x <= (pool_size_x - 8); x += 8) { const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); vres = vmaxq_f16(vres, data); } // Leftover for loop for(; x < pool_size_x; ++x) { const float16_t data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); res = std::max(res, data); } } float16x4_t tmp = vpmax_f16(vget_high_f16(vres), vget_low_f16(vres)); res = std::max(res, vget_lane_f16(tmp, 0)); res = std::max(res, vget_lane_f16(tmp, 1)); res = std::max(res, vget_lane_f16(tmp, 2)); res = std::max(res, vget_lane_f16(tmp, 3)); } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { res = std::sqrt(res); } // Store result *(reinterpret_cast(output.ptr())) = res; }, input, output); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ ARM_COMPUTE_UNUSED(window_input); ARM_COMPUTE_UNUSED(window); ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } void NEPoolingLayerKernel::poolingMxN_f16_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { ARM_COMPUTE_UNUSED(pooling_type); ARM_COMPUTE_UNUSED(exclude_padding); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC Iterator input(_input, window_input); Iterator output(_output, window); const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); float16x8_t vres; execute_window_loop(window, [&](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_input.z().start() + pool_limit_y); const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); if(pooling_type != PoolingType::MAX) { // Calculate scale const float scale = calculate_avg_scale(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); const float16x8_t scale_v = vdupq_n_f16(scale); // Perform pooling vres = vdupq_n_f16(0.0f); for(int y = pool_start_y; y < pool_end_y; ++y) { for(int x = pool_start_x; x < pool_end_x; ++x) { const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().z()))); // Get power of 2 in case of l2 pooling and accumulate if(pooling_type == PoolingType::L2) { vres = vaddq_f16(vres, vmulq_f16(data, data)); } else { vres = vaddq_f16(vres, data); } } } // Divide by scale vres = vmulq_f16(vres, scale_v); } else { vres = vdupq_n_f16(std::numeric_limits::lowest()); for(int y = pool_start_y; y < pool_end_y; ++y) { for(int x = pool_start_x; x < pool_end_x; ++x) { const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().z()))); vres = vmaxq_f16(vres, data); } } } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { float16x8_t sqrt_reciprocal = vrsqrteq_f16(vres); vres = vmulq_f16(vres, vmulq_f16(vrsqrtsq_f16(vmulq_f16(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal)); } // Store result vst1q_f16(reinterpret_cast(output.ptr()), vres); }, input, output); #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ ARM_COMPUTE_UNUSED(window_input); ARM_COMPUTE_UNUSED(window); ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } void NEPoolingLayerKernel::poolingMxN_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); execute_window_loop(window, [&](const Coordinates & id) { float res = 0.0f; if(pooling_type != PoolingType::MAX) { // Calculate scale const float scale = calculate_avg_scale(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 float32x4_t vres = vdupq_n_f32(0.0f); for(int y = 0; y < pool_size_y; ++y) { int x = 0; for(; x <= (pool_size_x - 4); x += 4) { const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().y()))); // Get power of 2 in case of l2 pooling and accumulate if(pooling_type == PoolingType::L2) { vres = vmlaq_f32(vres, data, data); } else { vres = vaddq_f32(vres, data); } } // Leftover for loop for(; x < pool_size_x; ++x) { float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().y()))); // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { data *= data; } res += data; } } #if defined(__aarch64__) // Reduction operation available on 64 bit architectures only res += vaddvq_f32(vres); #else // __aarch64__ // Reduction float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); tmp = vpadd_f32(tmp, tmp); res += vget_lane_f32(tmp, 0); #endif // __aarch64__ // Divide by scale res *= scale; } else { float32x4_t vres = vdupq_n_f32(std::numeric_limits::lowest()); res = std::numeric_limits::lowest(); for(int y = 0; y < pool_size_y; ++y) { int x = 0; for(; x <= (pool_size_x - 4); x += 4) { const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().y()))); vres = vmaxq_f32(vres, data); } // Leftover for loop for(; x < pool_size_x; ++x) { const float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().y()))); res = std::max(res, data); } } #if defined(__aarch64__) // Reduction operation available on 64 bit architectures only res = std::max(vmaxvq_f32(vres), res); #else // __aarch64__ float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); tmp = vpmax_f32(tmp, tmp); res = std::max(res, vget_lane_f32(tmp, 0)); #endif // __aarch64__ } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { res = std::sqrt(res); } // Store result *(reinterpret_cast(output.ptr())) = res; }, input, output); } void NEPoolingLayerKernel::pooling2_f32_nchw_maxpool_indices(const Window &window_input, const Window &window) { Iterator input(_input, window_input); Iterator output(_output, window); Iterator indices(_indices, window); int final_index = 0; const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); 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 uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); const Strides &input_strides = _input->info()->strides_in_bytes(); const auto in_stridew = input_strides[1]; execute_window_loop(window, [&](const Coordinates &) { const auto input_offset_top = input_top_ptr + input.offset(); const auto input_offset_bottom = input_bottom_ptr + input.offset(); const auto in_top_ptr = reinterpret_cast(input_offset_top); const auto in_bottom_ptr = reinterpret_cast(input_offset_bottom); float32x2_t top_data = vld1_f32(in_top_ptr); float32x2_t bottom_data = vld1_f32(in_bottom_ptr); float32x2_t res = {}; float final_res = 0; const float32x2_t max_data = vmax_f32(top_data, bottom_data); res = vpmax_f32(max_data, max_data); final_res = vget_lane_f32(res, 0); // Store result *(reinterpret_cast(output.ptr())) = final_res; const uint32_t offset_top = (uint32_t)(input.offset() / sizeof(float)); const uint32_t offset_bottom = (uint32_t)offset_top + (in_stridew / sizeof(float)); const uint32x2_t voffset_top = { offset_top, offset_top + 1u }; const uint32x2_t voffset_bottom = { offset_bottom, offset_bottom + 1u }; const uint32x2_t tmp_indices = vbsl_u32(vcgt_f32(top_data, bottom_data), voffset_top, voffset_bottom); final_index = vget_lane_u32(vbsl_u32(vcgt_f32(max_data, vrev64_f32(max_data)), tmp_indices, vrev64_u32(tmp_indices)), 0); *(reinterpret_cast(indices.ptr())) = final_index; }, input, output, indices); } void NEPoolingLayerKernel::pooling2_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { if(pooling_type == PoolingType::MAX && _indices) { pooling2_f32_nchw_maxpool_indices(window_input, window); } else { Iterator input(_input, window_input); Iterator output(_output, window); constexpr int pool_size = 2; 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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); execute_window_loop(window, [&](const Coordinates & id) { const auto in_top_ptr = reinterpret_cast(input_top_ptr + input.offset()); const auto in_bottom_ptr = reinterpret_cast(input_bottom_ptr + input.offset()); float32x2_t top_data = vld1_f32(in_top_ptr); float32x2_t bottom_data = vld1_f32(in_bottom_ptr); float32x2_t res = {}; float final_res = 0; // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { top_data = vmul_f32(top_data, top_data); bottom_data = vmul_f32(bottom_data, bottom_data); } if(pooling_type != PoolingType::MAX) { // Calculate scale float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); const float32x2_t scale_v = vdup_n_f32(scale); // Perform pooling const float32x2_t sum_data = vadd_f32(top_data, bottom_data); res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); } else { const float32x2_t max_data = vmax_f32(top_data, bottom_data); res = vpmax_f32(max_data, max_data); } final_res = vget_lane_f32(res, 0); // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { final_res = sqrt(final_res); } // Store result *(reinterpret_cast(output.ptr())) = final_res; }, input, output); } } void NEPoolingLayerKernel::pooling3_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); constexpr const 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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); const uint8_t *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); execute_window_loop(window, [&](const Coordinates & id) { float32x4_t top_data = vld1q_f32(reinterpret_cast(input_top_ptr + input.offset())); float32x4_t middle_data = vld1q_f32(reinterpret_cast(input_middle_ptr + input.offset())); float32x4_t bottom_data = vld1q_f32(reinterpret_cast(input_bottom_ptr + input.offset())); float32x2_t res = {}; float final_res = 0; // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { top_data = vmulq_f32(top_data, top_data); middle_data = vmulq_f32(middle_data, middle_data); bottom_data = vmulq_f32(bottom_data, bottom_data); } if(pooling_type != PoolingType::MAX) { // Calculate scale float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); const float32x2_t scale_v = vdup_n_f32(scale); // Perform pooling const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); res = vmul_f32(vpadd_f32(res, res), scale_v); } else { const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data, 3)), vget_low_f32(max_data)); res = vpmax_f32(res, res); } final_res = vget_lane_f32(res, 0); // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { final_res = sqrt(final_res); } // Store result *(reinterpret_cast(output.ptr())) = final_res; }, input, output); } void NEPoolingLayerKernel::pooling7_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); constexpr const int pool_size = 7; 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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); std::array input_ptrs{ {} }; for(int i = 0; i < pool_size; ++i) { input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + i)); } execute_window_loop(window, [&](const Coordinates & id) { float32x2_t res = {}; float final_res = 0.f; if(pooling_type != PoolingType::MAX) { // Calculate scale float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); const float32x2_t scale_v = vdup_n_f32(scale); // Perform pooling float32x4x2_t data = vld2q_f32(reinterpret_cast(input_ptrs[0] + input.offset())); // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { data.val[0] = vmulq_f32(data.val[0], data.val[0]); data.val[1] = vmulq_f32(data.val[1], data.val[1]); } float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); for(int i = 1; i < pool_size; ++i) { data = vld2q_f32(reinterpret_cast(input_ptrs[i] + input.offset())); // Get power of 2 in case of l2 pooling if(pooling_type == PoolingType::L2) { data.val[0] = vmulq_f32(data.val[0], data.val[0]); data.val[1] = vmulq_f32(data.val[1], data.val[1]); } sum_data = vaddq_f32(sum_data, data.val[0]); sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); } res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); res = vmul_f32(vpadd_f32(res, res), scale_v); } else { float32x4x2_t max_data = vld2q_f32(reinterpret_cast(input_ptrs[0] + input.offset())); for(int i = 1; i < pool_size; ++i) { const float32x4x2_t data = vld2q_f32(reinterpret_cast(input_ptrs[i] + input.offset())); max_data = vmax2q_f32(max_data, data); } res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1])); res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); res = vpmax_f32(res, res); } final_res = vget_lane_f32(res, 0); // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { final_res = sqrt(final_res); } // Store result *(reinterpret_cast(output.ptr())) = final_res; }, input, output); } void NEPoolingLayerKernel::poolingMxN_f32_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) { pooling2_f32_nhwc_maxpool_indices(window_input, window); } else { Iterator input(_input, window_input); Iterator output(_output, window); const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); float32x4_t vres; execute_window_loop(window, [&](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_input.z().start() + pool_limit_y); const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); if(pooling_type != PoolingType::MAX) { // Calculate scale const float scale = calculate_avg_scale(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); const float32x4_t scale_v = vdupq_n_f32(scale); // Perform pooling vres = vdupq_n_f32(0.0f); for(int y = pool_start_y; y < pool_end_y; ++y) { for(int x = pool_start_x; x < pool_end_x; ++x) { const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()))); // Get power of 2 in case of l2 pooling and accumulate if(pooling_type == PoolingType::L2) { vres = vmlaq_f32(vres, data, data); } else { vres = vaddq_f32(vres, data); } } } // Divide by scale vres = vmulq_f32(vres, scale_v); } else { vres = vdupq_n_f32(std::numeric_limits::lowest()); for(int y = pool_start_y; y < pool_end_y; ++y) { for(int x = pool_start_x; x < pool_end_x; ++x) { const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()))); vres = vmaxq_f32(vres, data); } } } // Calculate square-root in case of l2 pooling if(pooling_type == PoolingType::L2) { float32x4_t l2_res = { static_cast(sqrt(vgetq_lane_f32(vres, 0))), static_cast(sqrt(vgetq_lane_f32(vres, 1))), static_cast(sqrt(vgetq_lane_f32(vres, 2))), static_cast(sqrt(vgetq_lane_f32(vres, 3))) }; vres = l2_res; } // Store result vst1q_f32(reinterpret_cast(output.ptr()), vres); }, input, output); } } void NEPoolingLayerKernel::pooling2_f32_nhwc_maxpool_indices(const Window &window_input, const Window &window) { Iterator input(_input, window_input); Iterator output(_output, window); Iterator indices(_indices, window); const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); int pool_stride_x = 0; int pool_stride_y = 0; std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); float32x4_t vres; const int pad_right = _input->info()->padding().right; const int pad_top = _input->info()->padding().top; const int in_stride_y = static_cast(_input->info()->strides_in_bytes().y()); const int in_stride_z = static_cast(_input->info()->strides_in_bytes().z()); const int in_stride_w = static_cast(_input->info()->strides_in_bytes()[3]); execute_window_loop(window, [&](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_input.z().start() + pool_limit_y); const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()); const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()); const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()); const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()); const auto in_x0_ptr = reinterpret_cast(input.ptr() + in_x0_offset); const auto in_x1_ptr = reinterpret_cast(input.ptr() + in_x1_offset); const auto in_x2_ptr = reinterpret_cast(input.ptr() + in_x2_offset); const auto in_x3_ptr = reinterpret_cast(input.ptr() + in_x3_offset); const auto v_x0 = vld1q_f32(in_x0_ptr); const auto v_x1 = vld1q_f32(in_x1_ptr); const auto v_x2 = vld1q_f32(in_x2_ptr); const auto v_x3 = vld1q_f32(in_x3_ptr); vres = vmaxq_f32(vmaxq_f32(v_x2, v_x3), vmaxq_f32(v_x0, v_x1)); // Store result vst1q_f32(reinterpret_cast(output.ptr()), vres); const uint32_t offset_base = input.offset() - sizeof(float) * pad_right * id.y() * pool_stride_x /* subtract padding elems per row */ - pad_top * sizeof(float) /* top padding */ - sizeof(float) * pad_right * _input->info()->tensor_shape()[1] * id.z() * pool_stride_y /* for each Z plane there are width*pad_right padding elems */ - in_stride_w * id[3] + _input->info()->tensor_shape()[0] * sizeof(float) * id[3]; const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float); const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _input->info()->tensor_shape()[1]; const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; const uint32x4_t voffset_x0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; const uint32x4_t voffset_x1 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; const uint32x4_t voffset_x2 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; const uint32x4_t voffset_x3 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; const uint32x4_t tmp_indices0 = vbslq_u32(vcgtq_f32(v_x0, v_x1), voffset_x0, voffset_x1); const uint32x4_t tmp_indices1 = vbslq_u32(vcgtq_f32(v_x2, v_x3), voffset_x2, voffset_x3); const uint32x4_t tmp_indices2 = vbslq_u32(vcgtq_f32(vmaxq_f32(v_x0, v_x1), vmaxq_f32(v_x2, v_x3)), tmp_indices0, tmp_indices1); vst1q_u32(reinterpret_cast(indices.ptr()), tmp_indices2); }, input, output, indices); } template void NEPoolingLayerKernel::poolingMxN_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); /** NEON vector types */ using q8x8_t = typename wrapper::traits::neon_vector::type; using q16_t = typename wrapper::traits::promote_t; using q16x8_t = typename wrapper::traits::neon_vector::type; using q32_t = typename wrapper::traits::promote_t; using q32x4_t = typename wrapper::traits::neon_vector::type; const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); const UniformQuantizationInfo &input_qinfo = _input->info()->quantization_info().uniform(); const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform(); execute_window_loop(window, [&](const Coordinates & id) { T res = std::numeric_limits::min(); if(pooling_type != PoolingType::MAX) { q32x4_t vres = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); q32_t sres = 0; // Calculate scale const float scale = calculate_avg_scale(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(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->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(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->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(support::cpp11::round(sres * scale)); } else { q8x8_t vres = wrapper::vdup_n(std::numeric_limits::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(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().y()))); vres = wrapper::vmax(vres, data); } // Leftover for loop for(; x < pool_size_x; ++x) { const T data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast (_input->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 = (input_qinfo != output_qinfo) ? Qasymm8QuantizationHelper::quantize(Qasymm8QuantizationHelper::dequantize(res, input_qinfo), output_qinfo) : res; *(reinterpret_cast(output.ptr())) = res; }, input, output); } template void NEPoolingLayerKernel::poolingMxN_q8_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) { Iterator input(_input, window_input); Iterator output(_output, window); using q8x8_t = typename wrapper::traits::neon_vector::type; using q8x16_t = typename wrapper::traits::neon_vector::type; using q16_t = typename wrapper::traits::promote_t; using q16x8_t = typename wrapper::traits::neon_vector::type; using q32_t = typename wrapper::traits::promote_t; using q32x4_t = typename wrapper::traits::neon_vector::type; const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; const int pool_size_y = _pool_info.is_global_pooling ? _input->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 = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); const float32x4_t half_scale_v = vdupq_n_f32(0.5f); const UniformQuantizationInfo input_qinfo = _input->info()->quantization_info().uniform(); const UniformQuantizationInfo output_qinfo = _output->info()->quantization_info().uniform(); const float quant_rescale = output_qinfo.scale / input_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 = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / quant_rescale); const float requant_scale = output_qinfo.scale / input_qinfo.scale; const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / requant_scale); const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); execute_window_loop(window, [&](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_input.z().start() + pool_limit_y); const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); if(pooling_type != PoolingType::MAX) { q32x4_t vres1 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); q32x4_t vres2 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); q32x4_t vres3 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); q32x4_t vres4 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); // Calculate scale const float scale = calculate_avg_scale(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(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()))); 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(input_qinfo != output_qinfo) { const float32x4x4_t vres = { { vcvtq_f32_q32(vres1), vcvtq_f32_q32(vres2), vcvtq_f32_q32(vres3), vcvtq_f32_q32(vres4), } }; const auto requantized_output = vrequantize_pooling_with_scale(vres, quant_rescale, scale, new_offset); // Store result wrapper::vstore(reinterpret_cast(output.ptr()), wrapper::vgetlow(requantized_output)); wrapper::vstore(reinterpret_cast(output.ptr()) + 8, wrapper::vgethigh(requantized_output)); } 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(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v)); vres2 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v)); vres3 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v)); vres4 = vcvtq_q32_f32(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(output.ptr()), res1); wrapper::vstore(reinterpret_cast(output.ptr()) + 8, res2); } } else { q8x16_t vres = wrapper::vdup_n(std::numeric_limits::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(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast (_input->info()->strides_in_bytes().z()))); vres = wrapper::vmax(vres, data); } } // Store result wrapper::vstore(reinterpret_cast(output.ptr()), (input_qinfo != output_qinfo) ? vrequantize_pooling(wrapper::vgetlow(vres), wrapper::vgethigh(vres), requant_qinfo) : vres); } }, input, output); } Status NEPoolingLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); unsigned int pooled_w = 0; unsigned int pooled_h = 0; unsigned int num_elems_processed_per_iteration = 0; BorderSize border_size(0); const bool is_global_pooling = pool_info.is_global_pooling; unsigned int pool_size_x = 0; unsigned int pool_size_y = 0; // Get data layout const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout; const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); pool_size_x = is_global_pooling ? input->dimension(idx_width) : pool_info.pool_size.width; pool_size_y = is_global_pooling ? input->dimension(idx_height) : pool_info.pool_size.height; // Validate pool info before calling scaled_dimensions ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_pool_info(pool_size_x, pool_size_y)); // Check output dimensions std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(idx_width), input->dimension(idx_height), pool_size_x, pool_size_y, pool_info.pad_stride_info); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, pool_info, pooled_w, pooled_h, indices, Size2D(pool_size_x, pool_size_y))); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration, border_size, pooled_w, pooled_h, pool_size_x, pool_size_y) .first); return Status{}; } void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); ARM_COMPUTE_ERROR_ON(_func == nullptr); const unsigned int pool_stride_x = _pool_info.pad_stride_info.stride().first; const unsigned int pool_stride_y = _pool_info.pad_stride_info.stride().second; const unsigned int pool_size = _pool_info.pool_size.width; const bool exclude_padding = _pool_info.exclude_padding; Window window_input(window); if(_data_layout == DataLayout::NCHW) { // Set step for input in x and y direction for the input unsigned int window_x_inc = 0; switch(_input->info()->data_type()) { case DataType::QASYMM8: case DataType::QASYMM8_SIGNED: { window_x_inc = pool_stride_x; if((pool_size == 2 || pool_size == 3) && pool_stride_x < 3) { window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; } break; } case DataType::F16: case DataType::F32: { window_x_inc = pool_stride_x; break; } default: { ARM_COMPUTE_ERROR("Not supported"); } } window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); } else { window_input.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), _num_elems_processed_per_iteration)); window_input.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), pool_stride_x)); window_input.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), pool_stride_y)); } // Run function (this->*_func)(window_input, window, _pool_info.pool_type, exclude_padding); } } // namespace arm_compute