diff options
Diffstat (limited to 'src/cpu/kernels/pool2d/neon')
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/fp16.cpp | 659 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/fp32.cpp | 481 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/impl.h | 138 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/list.h | 113 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/nchw/all.cpp | 462 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/qasymm8.cpp | 47 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp | 47 | ||||
| -rw-r--r-- | src/cpu/kernels/pool2d/neon/quantized.h | 832 |
8 files changed, 2779 insertions, 0 deletions
diff --git a/src/cpu/kernels/pool2d/neon/fp16.cpp b/src/cpu/kernels/pool2d/neon/fp16.cpp new file mode 100644 index 0000000000..9d24d79afb --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/fp16.cpp @@ -0,0 +1,659 @@ +/* + * Copyright (c) 2021-2023 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/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/impl.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +namespace arm_compute +{ +namespace cpu +{ +#ifdef ENABLE_NCHW_KERNELS + +namespace +{ +float16x4_t +read_4_boundary_aware_fp16(int srcw, int srch, int pad_l, int pad_t, int x, int y, const float16_t *ptr, float16_t fval) +{ + float16_t vec[4]; + const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t))); + for (int i = 0; i < 4; i++) + { + if (row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l))) + { + vec[i] = *(ptr + i); + } + else + { + vec[i] = fval; + } + } + return wrapper::vload(vec); +} +} // namespace + +void pooling3_fp16_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); + + 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float16_t fp16_min = get_initial_min<half_float::half>(pool_info.use_inf_as_limit); + const float16_t fill_value = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.f; + const unsigned char *const src_top_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); + const unsigned char *const src_middle_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); + const unsigned char *const src_bottom_ptr = + 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 x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + const auto y_val_2 = (id.y() * pool_stride_y) + 2; + float16x4_t top_data = + read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_0, + reinterpret_cast<const float16_t *>(src_top_ptr + in.offset()), fill_value); + float16x4_t middle_data = read_4_boundary_aware_fp16( + src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_1, + reinterpret_cast<const float16_t *>(src_middle_ptr + in.offset()), fill_value); + float16x4_t bottom_data = read_4_boundary_aware_fp16( + src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_2, + reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset()), fill_value); + float16x4_t res = {}; + + // Get power of 2 in case of l2 pooling + if (pool_info.pool_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 (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + pool_info.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(fp16_min, max_data, 3), max_data); + res = vpmax_f16(res, res); + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + res = vsqrt_f16(res); + } + + *(reinterpret_cast<float16_t *>(out.ptr())) = vget_lane_f16(res, 0); + }, + in, out); +} +#endif // ENABLE_NCHW_KERNELS + +void pooling2_f16_maxpool_indices(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 8; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, window_src); + Iterator out(dst0, window_out); + Iterator indices(dst1, window_out); + + 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 int pad_right = src->info()->padding().right; + const int pad_left = src->info()->padding().left; + const int pad_horizontal = pad_right + pad_left; + const int in_stride_y = static_cast<int>(src->info()->strides_in_bytes().y()); + const int in_stride_z = static_cast<int>(src->info()->strides_in_bytes().z()); + + 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_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int in_x0_offset = + (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x1_offset = + (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x2_offset = + (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y + 1 - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x3_offset = + (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y + 1 - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + + int x_off = window_start_x; + for (; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + const auto in_x0_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x0_offset) + x_off; + const auto in_x1_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x1_offset) + x_off; + const auto in_x2_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x2_offset) + x_off; + const auto in_x3_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x3_offset) + x_off; + const auto v_x0 = vld1q_f16(in_x0_ptr); + const auto v_x1 = vld1q_f16(in_x1_ptr); + const auto v_x2 = vld1q_f16(in_x2_ptr); + const auto v_x3 = vld1q_f16(in_x3_ptr); + float16x8_t vres = vmaxq_f16(vmaxq_f16(v_x2, v_x3), vmaxq_f16(v_x0, v_x1)); + // Store result + vst1q_f16(reinterpret_cast<float16_t *>(out.ptr()) + x_off, vres); + + const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, + pool_stride_y, DataLayout::NHWC); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_horizontal; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - + pad_horizontal * src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_horizontal; + const uint32x4_t voffset_x0_0 = {offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3}; + const uint32x4_t voffset_x0_1 = {offset_x0 + 4, offset_x0 + 5, offset_x0 + 6, offset_x0 + 7}; + const uint16x8_t voffset_x0 = vcombine_u16(vmovn_u32(voffset_x0_0), vmovn_u32(voffset_x0_1)); + const uint32x4_t voffset_x1_0 = {offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3}; + const uint32x4_t voffset_x1_1 = {offset_x1 + 4, offset_x1 + 5, offset_x1 + 6, offset_x1 + 7}; + const uint16x8_t voffset_x1 = vcombine_u16(vmovn_u32(voffset_x1_0), vmovn_u32(voffset_x1_1)); + const uint32x4_t voffset_x2_0 = {offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3}; + const uint32x4_t voffset_x2_1 = {offset_x2 + 4, offset_x2 + 5, offset_x2 + 6, offset_x2 + 7}; + const uint16x8_t voffset_x2 = vcombine_u16(vmovn_u32(voffset_x2_0), vmovn_u32(voffset_x2_1)); + const uint32x4_t voffset_x3_0 = {offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3}; + const uint32x4_t voffset_x3_1 = {offset_x3 + 4, offset_x3 + 5, offset_x3 + 6, offset_x3 + 7}; + const uint16x8_t voffset_x3 = vcombine_u16(vmovn_u32(voffset_x3_0), vmovn_u32(voffset_x3_1)); + const uint16x8_t tmp_indices0 = vbslq_u16(vcgeq_f16(v_x0, v_x1), voffset_x0, voffset_x1); + const uint16x8_t tmp_indices1 = vbslq_u16(vcgeq_f16(v_x2, v_x3), voffset_x2, voffset_x3); + const uint16x8_t tmp_indices2 = + vbslq_u16(vcgeq_f16(vmaxq_f16(v_x0, v_x1), vmaxq_f16(v_x2, v_x3)), tmp_indices0, tmp_indices1); + const uint32x4_t tmp_indeces3_0 = vmovl_u16(vget_low_u16(tmp_indices2)); + const uint32x4_t tmp_indeces3_1 = vmovl_u16(vget_high_u16(tmp_indices2)); + // Store indicies + vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, tmp_indeces3_0); + vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr() + 16) + x_off, tmp_indeces3_1); + } + + // Left-overs loop + for (; x_off < window_end_x; ++x_off) + { + const auto x0 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x0_offset) + x_off); + const auto x1 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x1_offset) + x_off); + const auto x2 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x2_offset) + x_off); + const auto x3 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x3_offset) + x_off); + float16_t res = std::max(std::max(x2, x3), std::max(x0, x1)); + + // Store result + *(reinterpret_cast<float16_t *>(out.ptr()) + x_off) = res; + + const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, + pool_stride_y, DataLayout::NHWC); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_horizontal; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - + pad_horizontal * src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_horizontal; + const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; + const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; + const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; + + // Store indices + *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = tmp_idx2; + } + }, + in, out, indices); +} +#ifdef ENABLE_NCHW_KERNELS + +void pooling2_fp16_neon_nchw(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + if (pool_info.pool_type == PoolingType::MAX && dst1) + { + pooling2_nchw_maxpool_indices<float16_t>(src, dst0, dst1, pool_info, window_src, window); + } + else + { + Iterator in(src, window_src); + Iterator out(dst0, 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float16_t fp16_min = get_initial_min<half_float::half>(pool_info.use_inf_as_limit); + const float16_t fill_value = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.0f; + + const unsigned char *const src_top_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); + const unsigned char *const src_bottom_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto in_top_ptr = reinterpret_cast<const float16_t *>(src_top_ptr + in.offset()); + const auto in_bottom_ptr = reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset()); + + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + float16x4_t top_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top, x_val, + y_val_0, in_top_ptr, fill_value); + float16x4_t bottom_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top, x_val, + y_val_1, in_bottom_ptr, fill_value); + float16x4_t res = {}; + + // Get power of 2 in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + top_data = vmul_f16(top_data, top_data); + bottom_data = vmul_f16(bottom_data, bottom_data); + } + + if (pool_info.pool_type != PoolingType::MAX) + { + const float scale = calculate_avg_scale_pool2d( + pool_info.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 (pool_info.pool_type == PoolingType::L2) + { + res = vsqrt_f16(res); + } + + // Store result + *(reinterpret_cast<float16_t *>(out.ptr())) = vget_lane_f16(res, 0); + }, + in, out); + } +} + +void poolingMxN_fp16_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); + + 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float16_t fp16_min = get_initial_min<half_float::half>(pool_info.use_inf_as_limit); + const float16_t fill_value = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.0f; + + execute_window_loop( + window, + [&](const Coordinates &id) + { + float16_t res = 0.0f; + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + const float16_t scale = calculate_avg_scale_pool2d( + 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) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto ptr = reinterpret_cast<const float16_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 int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + float16_t data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr; + + if (pool_info.pool_type == PoolingType::L2) + { + data *= data; + } + + res += data; + } + } + + // Divide by scale + res *= scale; + } + else // if max pooling + { + res = fp16_min; + + for (int y = 0; y < pool_size_y; ++y) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto ptr = reinterpret_cast<const float16_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 int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + float16_t data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr; + res = std::max(res, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast<float16_t *>(out.ptr())) = res; + }, + in, out); +} +#endif // ENABLE_NCHW_KERNELS + +void poolingMxN_fp16_neon_nhwc(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + if (pool_info.pool_size == Size2D(2, 2) && pool_info.pool_type == PoolingType::MAX && dst1) + { + pooling2_f16_maxpool_indices(src, dst0, dst1, pool_info, window_src, window); + } + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 8; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, window_src); + Iterator out(dst0, window_out); + + 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 float16_t min_value = get_initial_min<half_float::half>(pool_info.use_inf_as_limit); + float16x8_t vres; + + 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) + { + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + 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); + 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<const float16_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); + + // Get power of 2 in case of l2 pooling and accumulate + if (pool_info.pool_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(min_value); + + 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<const float16_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 = vmaxq_f16(vres, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_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<float16_t *>(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for (; x_off < window_end_x; ++x_off) + { + float16_t res = 0.0f; + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + const float16_t scale = calculate_avg_scale_pool2d( + 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); + + for (int y = pool_start_y; y < pool_end_y; ++y) + { + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = + *(reinterpret_cast<const float16_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); + + // Get power of 2 in case of l2 pooling and accumulate + if (pool_info.pool_type == PoolingType::L2) + { + res += data * data; + } + else + { + res += data; + } + } + } + + // Divide by scale + res *= scale; + } + else + { + res = min_value; + for (int y = pool_start_y; y < pool_end_y; ++y) + { + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float16_t data = + *(reinterpret_cast<const float16_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); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast<float16_t *>(out.ptr()) + x_off) = res; + } + }, + in, out); +} +} // namespace cpu +} // namespace arm_compute + +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/pool2d/neon/fp32.cpp b/src/cpu/kernels/pool2d/neon/fp32.cpp new file mode 100644 index 0000000000..aaa37863cb --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/fp32.cpp @@ -0,0 +1,481 @@ +/* + * Copyright (c) 2021-2023 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/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +void pooling2_f32_maxpool_indices(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 4; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, window_src); + Iterator out(dst0, window_out); + Iterator indices(dst1, window_out); + + 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; + float res; + + const int pad_right = src->info()->padding().right; + const int pad_left = src->info()->padding().left; + const int pad_horizontal = pad_right + pad_left; + const int in_stride_y = static_cast<int>(src->info()->strides_in_bytes().y()); + const int in_stride_z = static_cast<int>(src->info()->strides_in_bytes().z()); + + 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_start_x = std::max(0, window_src.y().start() + pool_limit_x); + + const int in_x0_offset = + (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x1_offset = + (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x2_offset = + (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y + 1 - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + const int in_x3_offset = + (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + + (pool_start_y + 1 - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); + + int x_off = window_start_x; + for (; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + const auto in_x0_ptr = reinterpret_cast<const float *>(in.ptr() + in_x0_offset); + const auto in_x1_ptr = reinterpret_cast<const float *>(in.ptr() + in_x1_offset); + const auto in_x2_ptr = reinterpret_cast<const float *>(in.ptr() + in_x2_offset); + const auto in_x3_ptr = reinterpret_cast<const float *>(in.ptr() + in_x3_offset); + const auto v_x0 = vld1q_f32(in_x0_ptr + x_off); + const auto v_x1 = vld1q_f32(in_x1_ptr + x_off); + const auto v_x2 = vld1q_f32(in_x2_ptr + x_off); + const auto v_x3 = vld1q_f32(in_x3_ptr + x_off); + vres = vmaxq_f32(vmaxq_f32(v_x2, v_x3), vmaxq_f32(v_x0, v_x1)); + // Store result + vst1q_f32(reinterpret_cast<float *>(out.ptr()) + x_off, vres); + + const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, + pool_stride_y, DataLayout::NHWC); + const uint32_t offset_x0 = offset_base / sizeof(float) + x_off; + const uint32_t offset_x1 = offset_x0 + in_stride_y / sizeof(float) - pad_horizontal; + const uint32_t offset_x2 = + offset_x0 + in_stride_z / sizeof(float) - pad_horizontal * src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = offset_x2 + in_stride_y / sizeof(float) - pad_horizontal; + 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(vcgeq_f32(v_x0, v_x1), voffset_x0, voffset_x1); + const uint32x4_t tmp_indices1 = vbslq_u32(vcgeq_f32(v_x2, v_x3), voffset_x2, voffset_x3); + const uint32x4_t tmp_indices2 = + vbslq_u32(vcgeq_f32(vmaxq_f32(v_x0, v_x1), vmaxq_f32(v_x2, v_x3)), tmp_indices0, tmp_indices1); + + // Store indices + vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, tmp_indices2); + } + + // Left-overs loop + for (; x_off < window_end_x; ++x_off) + { + const auto x0 = *(reinterpret_cast<const float *>(in.ptr() + in_x0_offset) + x_off); + const auto x1 = *(reinterpret_cast<const float *>(in.ptr() + in_x1_offset) + x_off); + const auto x2 = *(reinterpret_cast<const float *>(in.ptr() + in_x2_offset) + x_off); + const auto x3 = *(reinterpret_cast<const float *>(in.ptr() + in_x3_offset) + x_off); + res = std::max(std::max(x2, x3), std::max(x0, x1)); + + // Store result + *(reinterpret_cast<float *>(out.ptr()) + x_off) = res; + + const uint32_t offset_base = offset_no_padding<float>(in.offset(), id, *src->info(), pool_stride_x, + pool_stride_y, DataLayout::NHWC); + const uint32_t offset_x0 = offset_base / sizeof(float) + x_off; + const uint32_t offset_x1 = offset_x0 + in_stride_y / sizeof(float) - pad_horizontal; + const uint32_t offset_x2 = + offset_x0 + in_stride_z / sizeof(float) - pad_horizontal * src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = offset_x2 + in_stride_y / sizeof(float) - pad_horizontal; + const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; + const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; + const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; + + // Store indices + *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = tmp_idx2; + } + }, + in, out, indices); +} +} // namespace + +void poolingMxN_fp32_neon_nhwc_kernel_indices( + const ITensor *src, ITensor *dst0, ITensor *dst1, const PoolingLayerInfo &pool_info, const Window &window) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + constexpr int window_step_x = 4; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator out(dst0, window_out); + Iterator indices(dst1, window_out); + + 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_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 float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + + float32x4_t vres; + uint32x4_t vidx; + + constexpr int idx_width = 1; + constexpr int idx_height = 2; + constexpr int idx_batch = 3; + + const int y_stride = static_cast<int>(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z()); + const int n_stride = static_cast<int>(src->info()->strides_in_bytes()[idx_batch]); + + const int input_dim_w = src->info()->dimension(idx_width); + const int input_dim_h = src->info()->dimension(idx_height); + + const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); + + execute_window_loop( + window_out, + [&](const Coordinates &id) + { + const int idx_width = static_cast<int>(id.y()) * pool_stride_x - pool_pad_left; + const int idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top; + + const int pool_start_x = std::max(0, -idx_width); + const int pool_start_y = std::max(0, -idx_height); + + const int pool_end_x = std::min(pool_size_x, input_dim_w - idx_width); + const int pool_end_y = std::min(pool_size_y, input_dim_h - idx_height); + + const uint8_t *in_ptr_n = in_ptr_start + id[idx_batch] * n_stride; + + const int in_ptr_y_offset = (z_stride * idx_height) + (pool_start_y * z_stride); + const int in_ptr_x_offset = (y_stride * idx_width) + (pool_start_x * y_stride); + + int x_off = window_start_x; + + for (; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + vres = vdupq_n_f32(min_value); + vidx = vdupq_n_u32(0U); + const uint8_t *in_ptr_y = in_ptr_n + in_ptr_y_offset + in_ptr_x_offset; + uint32_t curr_kernel_index = pool_size_x * pool_start_y; + for (int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_x = in_ptr_y + (x_off * sizeof(float)); + curr_kernel_index += pool_start_x; + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(in_ptr_x)); + const uint32x4_t vidx_curr = vdupq_n_u32(curr_kernel_index); + const uint32x4_t idxMask = vcgtq_f32(data, vres); + vidx = vbslq_u32(idxMask, vidx_curr, vidx); + vres = vmaxq_f32(vres, data); + in_ptr_x += y_stride; + curr_kernel_index++; + } + curr_kernel_index += (pool_size_x - pool_end_x); + in_ptr_y += z_stride; + } + // Store result + vst1q_f32(reinterpret_cast<float *>(out.ptr()) + x_off, vres); + vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, vidx); + } + + // Left-overs loop + for (; x_off < window_end_x; ++x_off) + { + float res = min_value; + uint32_t idx = 0U; + const uint8_t *in_ptr_y = in_ptr_n + in_ptr_y_offset + in_ptr_x_offset; + for (int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_x = in_ptr_y + (x_off * sizeof(float)); + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = *(reinterpret_cast<const float *>(in_ptr_x)); + if (data > res) + { + idx = pool_size_x * y + x; + res = data; + } + in_ptr_x += y_stride; + } + in_ptr_y += z_stride; + } + + // Store result + *(reinterpret_cast<float *>(out.ptr()) + x_off) = res; + *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = idx; + } + }, + out, indices); +} + +void poolingMxN_fp32_neon_nhwc(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + if ((pool_info.pool_type == PoolingType::MAX) && pool_info.use_kernel_indices && (dst1 != nullptr)) + { + poolingMxN_fp32_neon_nhwc_kernel_indices(src, dst0, dst1, pool_info, window); + } + else if (pool_info.pool_size == Size2D(2, 2) && pool_info.pool_type == PoolingType::MAX && + !pool_info.pad_stride_info.has_padding() && (dst1 != nullptr)) + { + pooling2_f32_maxpool_indices(src, dst0, dst1, pool_info, window_src, window); + } + else + { + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 4; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in(src, window_src); + Iterator out(dst0, window_out); + + 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 float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + float32x4_t vres; + + 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) + { + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + 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); + 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<const float *>( + 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); + + // Get power of 2 in case of l2 pooling and accumulate + if (pool_info.pool_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(min_value); + 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<const float *>( + 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 = vmaxq_f32(vres, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + float32x4_t l2_res = {static_cast<float>(sqrt(vgetq_lane_f32(vres, 0))), + static_cast<float>(sqrt(vgetq_lane_f32(vres, 1))), + static_cast<float>(sqrt(vgetq_lane_f32(vres, 2))), + static_cast<float>(sqrt(vgetq_lane_f32(vres, 3)))}; + vres = l2_res; + } + + // Store result + vst1q_f32(reinterpret_cast<float *>(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for (; x_off < window_end_x; ++x_off) + { + float res = 0.0f; + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + 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); + + for (int y = pool_start_y; y < pool_end_y; ++y) + { + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = + *(reinterpret_cast<const float *>( + 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); + + // Get power of 2 in case of l2 pooling and accumulate + if (pool_info.pool_type == PoolingType::L2) + { + res += data * data; + } + else + { + res += data; + } + } + } + + // Divide by scale + res *= scale; + } + else + { + res = min_value; + for (int y = pool_start_y; y < pool_end_y; ++y) + { + for (int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = + *(reinterpret_cast<const float *>( + 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); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast<float *>(out.ptr()) + x_off) = res; + } + }, + in, out); + } +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/pool2d/neon/impl.h b/src/cpu/kernels/pool2d/neon/impl.h new file mode 100644 index 0000000000..008cf651e1 --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/impl.h @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2023 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 ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H +#define ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +#include <limits> + +#ifdef ENABLE_NCHW_KERNELS +namespace arm_compute +{ +namespace cpu +{ + +namespace +{ +template <typename T> +auto read_2_boundary_aware_as_f32(int srcw, int srch, int pad_l, int pad_t, int x, int y, const T *ptr, T fval) +{ + T vec[2]; + const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t))); + for (int i = 0; i < 2; i++) + { + if (row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l))) + { + vec[i] = *(ptr + i); + } + else + { + vec[i] = fval; + } + } + float32_t vec_f32[2] = {vec[0], vec[1]}; + return wrapper::vload(vec_f32); +} +} // namespace + +template <typename T> +void pooling2_nchw_maxpool_indices(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + Iterator in(src, window_src); + Iterator out(dst0, window); + Iterator indices(dst1, 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(); + const int src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const uint8_t *const src_top_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); + const uint8_t *const src_bottom_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); + const int pad_left = src->info()->padding().left; + const int pad_right = src->info()->padding().right; + const int in_stride_y = static_cast<int>(src->info()->strides_in_bytes().y()); + const T float_min = get_initial_min<T>(pool_info.use_inf_as_limit); + const T fill_value = (pool_info.pool_type == PoolingType::MAX) ? float_min : 0.f; + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + auto top_data = + read_2_boundary_aware_as_f32(src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_0, + reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value); + auto bottom_data = + read_2_boundary_aware_as_f32(src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_1, + reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value); + + // Calculate max data, compare top first, then bottom, to make sue the first max is recorded. + const float32x2_t max_data_top = vpmax_f32(top_data, top_data); + const float32x2_t max_data_bottom = vpmax_f32(bottom_data, bottom_data); + const float32x2_t max_data = vmax_f32(max_data_top, max_data_bottom); + *(reinterpret_cast<T *>(out.ptr())) = static_cast<T>(vget_lane_f32(max_data, 0)); + + // Calculate max data indice, which will be used in max unpool. + const uint32_t offset_base = + offset_no_padding<T>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NCHW); + const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); + const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; + 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_top = + vbsl_u32(vcge_f32(top_data, vrev64_f32(top_data)), voffset_top, vrev64_u32(voffset_top)); + const uint32x2_t tmp_indices_bottom = + vbsl_u32(vcge_f32(bottom_data, vrev64_f32(bottom_data)), voffset_bottom, vrev64_u32(voffset_bottom)); + *(reinterpret_cast<int *>(indices.ptr())) = vget_lane_u32( + vbsl_u32(vcge_f32(max_data_top, max_data_bottom), tmp_indices_top, tmp_indices_bottom), 0); + }, + in, out, indices); +} + +} // namespace cpu +} // namespace arm_compute + +#endif // ENABLE_NCHW_KERNELS + +#endif // ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H diff --git a/src/cpu/kernels/pool2d/neon/list.h b/src/cpu/kernels/pool2d/neon/list.h new file mode 100644 index 0000000000..5db843d56b --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/list.h @@ -0,0 +1,113 @@ +/* + * Copyright (c) 2021, 2023 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 ACL_SRC_CPU_KERNELS_POOL2D_NEON_LIST_H +#define ACL_SRC_CPU_KERNELS_POOL2D_NEON_LIST_H + +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/pool2d/neon/quantized.h" + +#include <arm_neon.h> + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_POOLING_KERNEL(func_name) \ + void func_name(const ITensor *src0, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &, const Window &window_src, \ + const Window &window) + +DECLARE_POOLING_KERNEL(poolingMxN_qasymm8_neon_nhwc); +DECLARE_POOLING_KERNEL(poolingMxN_qasymm8_signed_neon_nhwc); +DECLARE_POOLING_KERNEL(poolingMxN_fp16_neon_nhwc); +DECLARE_POOLING_KERNEL(poolingMxN_fp32_neon_nhwc); + +#if defined(ENABLE_NCHW_KERNELS) + +#if defined(ENABLE_FP16_KERNELS) +DECLARE_POOLING_KERNEL(pooling2_fp16_neon_nchw); +DECLARE_POOLING_KERNEL(pooling3_fp16_neon_nchw); +DECLARE_POOLING_KERNEL(poolingMxN_fp16_neon_nchw); +#endif /* defined(ENABLE_FP16_KERNELS) */ + +DECLARE_POOLING_KERNEL(pooling2_fp32_neon_nchw); +DECLARE_POOLING_KERNEL(pooling3_fp32_neon_nchw); +DECLARE_POOLING_KERNEL(pooling7_fp32_neon_nchw); +DECLARE_POOLING_KERNEL(poolingMxN_fp32_neon_nchw); +#endif /* defined(ENABLE_NCHW_KERNELS) */ + +#undef DECLARE_POOLING_KERNEL + +template <typename T> +T get_initial_min(bool use_inf_as_limit) +{ + return use_inf_as_limit ? -std::numeric_limits<T>::infinity() : std::numeric_limits<T>::lowest(); +} + +template <typename T> +inline uint32_t offset_no_padding(uint32_t padded_offset, + const Coordinates &id, + const ITensorInfo &info, + int pool_stride_x, + int pool_stride_y, + DataLayout data_layout) +{ + const int pad_left = info.padding().left; + const int pad_right = info.padding().right; + const int pad_top = info.padding().top; + const int pad_bottom = info.padding().bottom; + const int in_stride_y = static_cast<int>(info.strides_in_bytes().y()); + const int in_stride_w = static_cast<int>(info.strides_in_bytes()[3]); + const int pad_horiz = pad_left + pad_right; + const int pad_vert = pad_top + pad_bottom; + + if (data_layout == DataLayout::NCHW) + { + const uint32_t offset_base = + padded_offset - sizeof(T) * pad_horiz * id.y() * pool_stride_y /* subtract padding elems per row */ + - pad_top * sizeof(T) /* top padding */ + - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() - + pad_vert * in_stride_y * id.z() /* for each Z plane there are height*pad_right padding elems */ + - in_stride_w * id[3]; + + return offset_base; + } + else + { + const uint32_t offset_base = padded_offset - + sizeof(T) * pad_horiz * id.y() * pool_stride_x // subtract padding elems per row + - pad_top * sizeof(T) // top padding + - sizeof(T) * pad_horiz * 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]; + + return offset_base; + } +} +} // namespace cpu +} // namespace arm_compute + +#endif // ACL_SRC_CPU_KERNELS_POOL2D_NEON_LIST_H diff --git a/src/cpu/kernels/pool2d/neon/nchw/all.cpp b/src/cpu/kernels/pool2d/neon/nchw/all.cpp new file mode 100644 index 0000000000..0602bea667 --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/nchw/all.cpp @@ -0,0 +1,462 @@ +/* + * Copyright (c) 2021-2023 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/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/impl.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +#include <limits> + +#ifdef ENABLE_NCHW_KERNELS +namespace arm_compute +{ +namespace cpu +{ +#define READ_2_RIGHT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \ + (x == width + pad_left - 1) ? vset_lane_f32(*(ptr), vdup_n_f32(fval), 0) : vld1_f32(ptr) +#define READ_2_LEFT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \ + (x == pad_left - 1) ? vset_lane_f32(*(1 + ptr), vdup_n_f32(fval), 1) \ + : READ_2_RIGHT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) +#define READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \ + ((y < pad_top) || (x < pad_left - 1) || (y >= height + pad_top) || (x > width + pad_left - 1)) \ + ? vdup_n_f32(fval) \ + : READ_2_LEFT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) + +#define READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \ + vcombine_f32(READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval), \ + READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, (x + 2), y, (ptr + 2), fval)) + +float32x4x2_t +read_8_boundary_aware(int height, int width, int pad_left, int pad_top, int x, int y, const float *ptr, float fval) +{ + float32x4x2_t vec; + vec.val[0] = READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval); + vec.val[1] = READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, (x + 4), y, (ptr + 4), fval); + return vec; +} + +void poolingMxN_fp32_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); + + 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + const float fill_value = (pool_info.pool_type == PoolingType::MAX) ? min_value : 0.0f; + + execute_window_loop( + window, + [&](const Coordinates &id) + { + float res = 0.0f; + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + 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) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto ptr = reinterpret_cast<const float *>( + 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 int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + float data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr; + + if (pool_info.pool_type == PoolingType::L2) + { + data *= data; + } + + res += data; + } + } + + // Divide by scale + res *= scale; + } + else // if max pooling + { + res = min_value; + + for (int y = 0; y < pool_size_y; ++y) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto ptr = reinterpret_cast<const float *>( + 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 int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + float data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr; + res = std::max(res, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast<float *>(out.ptr())) = res; + }, + in, out); +} + +void pooling2_fp32_neon_nchw(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + if (pool_info.pool_type == PoolingType::MAX && dst1) + { + pooling2_nchw_maxpool_indices<float>(src, dst0, dst1, pool_info, window_src, window); + } + else + { + Iterator in(src, window_src); + Iterator out(dst0, 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + const float fill_value = (pool_info.pool_type == PoolingType::MAX) ? min_value : 0.0f; + + const uint8_t *const src_top_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); + const uint8_t *const src_bottom_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto in_top_ptr = reinterpret_cast<const float *>(src_top_ptr + in.offset()); + const auto in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + in.offset()); + + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + auto top_data = READ_2_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_0, + in_top_ptr, fill_value); + auto bottom_data = READ_2_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_1, + in_bottom_ptr, fill_value); + float32x2_t res = {}; + float final_res = 0; + + // Get power of 2 in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + top_data = vmul_f32(top_data, top_data); + bottom_data = vmul_f32(bottom_data, bottom_data); + } + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale_pool2d(pool_info.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 (pool_info.pool_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast<float *>(out.ptr())) = final_res; + }, + in, out); + } +} + +void pooling3_fp32_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); + + 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + const float fill_value = (pool_info.pool_type == PoolingType::MAX) ? min_value : 0.0f; + + const uint8_t *const src_top_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); + const uint8_t *const src_middle_ptr = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); + const uint8_t *const src_bottom_ptr = + 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 in_top_ptr = reinterpret_cast<const float *>(src_top_ptr + in.offset()); + const auto in_middle_ptr = reinterpret_cast<const float *>(src_middle_ptr + in.offset()); + const auto in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + in.offset()); + + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + const auto y_val_2 = (id.y() * pool_stride_y) + 2; + auto top_data = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_0, in_top_ptr, + fill_value); + auto middle_data = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_1, + in_middle_ptr, fill_value); + auto bottom_data = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_2, + in_bottom_ptr, fill_value); + + float32x2_t res = {}; + float final_res = 0; + + // Get power of 2 in case of l2 pooling + if (pool_info.pool_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 (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale_pool2d(pool_info.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(min_value, 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 (pool_info.pool_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast<float *>(out.ptr())) = final_res; + }, + in, out); +} + +void pooling7_fp32_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); + + 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 src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom); + const float min_value = get_initial_min<float>(pool_info.use_inf_as_limit); + const float fill_value = (pool_info.pool_type == PoolingType::MAX) ? min_value : 0.0f; + + std::array<const uint8_t *, pool_size> src_ptrs{{}}; + for (int i = 0; i < pool_size; ++i) + { + src_ptrs[i] = + src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + i)); + } + + execute_window_loop( + window, + [&](const Coordinates &id) + { + auto in_ptr = reinterpret_cast<const float *>(src_ptrs[0] + in.offset()); + + auto x_val = id.x() * pool_stride_x; + auto y_val = id.y() * pool_stride_y; + float32x4x2_t data = + read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, in_ptr, fill_value); + + float32x2_t res = {}; + float final_res = 0.f; + + if (pool_info.pool_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale_pool2d(pool_info.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); + + // Get power of 2 in case of l2 pooling + if (pool_info.pool_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) + { + in_ptr = reinterpret_cast<const float *>(src_ptrs[i] + in.offset()); + + x_val = id.x() * pool_stride_x; + y_val = (id.y() * pool_stride_y) + i; + data = read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, in_ptr, + fill_value); + // Get power of 2 in case of l2 pooling + if (pool_info.pool_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 + { + for (int i = 1; i < pool_size; ++i) + { + in_ptr = reinterpret_cast<const float *>(src_ptrs[i] + in.offset()); + + x_val = id.x() * pool_stride_x; + y_val = (id.y() * pool_stride_y) + i; + float32x4x2_t temp = read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, + in_ptr, fill_value); + data = vmax2q_f32(data, temp); + } + res = vpmax_f32(vget_high_f32(vsetq_lane_f32(min_value, data.val[1], 3)), vget_low_f32(data.val[1])); + res = vpmax_f32(res, vpmax_f32(vget_high_f32(data.val[0]), vget_low_f32(data.val[0]))); + res = vpmax_f32(res, res); + } + final_res = vget_lane_f32(res, 0); + + // Calculate square-root in case of l2 pooling + if (pool_info.pool_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast<float *>(out.ptr())) = final_res; + }, + in, out); +} +} // namespace cpu +} // namespace arm_compute + +#endif // ENABLE_NCHW_KERNELS diff --git a/src/cpu/kernels/pool2d/neon/qasymm8.cpp b/src/cpu/kernels/pool2d/neon/qasymm8.cpp new file mode 100644 index 0000000000..44675b5394 --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/qasymm8.cpp @@ -0,0 +1,47 @@ +/* + * 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +namespace arm_compute +{ +namespace cpu +{ +void poolingMxN_qasymm8_neon_nhwc(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + poolingMxN_q8_neon_nhwc<uint8_t>(src, dst0, dst1, pool_info, window_src, window); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp b/src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp new file mode 100644 index 0000000000..d434323e89 --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp @@ -0,0 +1,47 @@ +/* + * 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/cpu/kernels/pool2d/neon/list.h" + +namespace arm_compute +{ +namespace cpu +{ +void poolingMxN_qasymm8_signed_neon_nhwc(const ITensor *src, + ITensor *dst0, + ITensor *dst1, + PoolingLayerInfo &pool_info, + const Window &window_src, + const Window &window) +{ + poolingMxN_q8_neon_nhwc<int8_t>(src, dst0, dst1, pool_info, window_src, window); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/pool2d/neon/quantized.h b/src/cpu/kernels/pool2d/neon/quantized.h new file mode 100644 index 0000000000..38f1b2f1f9 --- /dev/null +++ b/src/cpu/kernels/pool2d/neon/quantized.h @@ -0,0 +1,832 @@ +/* + * Copyright (c) 2021-2022 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/helpers/PoolingHelpers.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> +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_pool2d( + 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_pool2d( + 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> +auto load16_boundary_aware( + int srcw, int srch, int pad_l, int pad_r, int pad_t, int pad_b, int x, int y, const T *ptr, T fval) +{ + ARM_COMPUTE_UNUSED(pad_b, pad_r); + T vec[16]; + //handle reading a row out of the tensor + const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t))); + for (int i = 0; i < 16; i++) + { + if (row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l))) + { + vec[i] = *(ptr + i); + } + else + { + vec[i] = fval; + } + } + return wrapper::vloadq(vec); +} + +template <typename T, typename V, bool deinterleave> +inline void write16_boundary_aware(int x, int dst_w, const V &lower, const V &upper, T *ptr) +{ + if (deinterleave) + { + for (int i = 0; i < 8 && (i * 2 + x) < dst_w; ++i) + { + *(ptr + i * 2) = lower[i]; + } + for (int i = 0; i < 8 && (i * 2 + x + 1) < dst_w; ++i) + { + *(ptr + 1 + i * 2) = upper[i]; + } + } + else + { + for (int i = 0; i < 8 && (i + x) < dst_w; ++i) + { + *(ptr + i) = lower[i]; + } + for (int i = 0; i < 8 && (i + x + 8) < dst_w; ++i) + { + *(ptr + i + 8) = upper[i]; + } + } +} + +template <typename T, typename V> +inline void write8_boundary_aware(int x, int dst_w, const V &v, T *ptr) +{ + for (int i = 0; i < 8 && (i + x) < dst_w; ++i) + { + *(ptr + i) = v[i]; + } +} + +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); + + /** SIMD 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 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); + const int src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const int dst_w = dst0->info()->dimension(0); + + const T fill_value = (pool_info.pool_type == PoolingType::MAX) ? std::numeric_limits<T>::min() : T(0); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + + auto top_data = + load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom, x_val, + y_val_0, reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value); + auto bottom_data = + load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom, x_val, + y_val_1, reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value); + + 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); + } + auto out_ptr = reinterpret_cast<T *>(out.ptr()); + // Store result + if (pool_stride_x == 1) + { + write16_boundary_aware<T, q8x8_t, true>(id.x(), dst_w, lower_res, upper_res, out_ptr); + } + else + { + write8_boundary_aware<T, q8x8_t>(id.x(), dst_w, lower_res, out_ptr); + } + }, + 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); + + /** SIMD 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))); + + const int src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const T fill_value = (pool_info.pool_type == PoolingType::AVG) ? T(0) : std::numeric_limits<T>::min(); + const int dst_w = dst0->info()->dimension(0); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto x_val = id.x() * pool_stride_x; + const auto y_val_0 = id.y() * pool_stride_y; + const auto y_val_1 = (id.y() * pool_stride_y) + 1; + const auto y_val_2 = (id.y() * pool_stride_y) + 2; + + auto top_data = + load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom, x_val, + y_val_0, reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value); + auto middle_data = + load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom, x_val, + y_val_1, reinterpret_cast<const T *>(src_middle_ptr + in.offset()), fill_value); + auto bottom_data = + load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom, x_val, + y_val_2, reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value); + + 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); + } + write16_boundary_aware<T, q8x8_t, false>(id.x(), dst_w, wrapper::vgetlow(fqres), + wrapper::vgethigh(fqres), reinterpret_cast<T *>(out.ptr())); + } + else + { + if (src_qinfo != dst_qinfo) + { + fres = vrequantize_pooling<q8x8_t>(fres, requant_qinfo); + } + write8_boundary_aware<T, q8x8_t>(id.x(), dst_w, fres, reinterpret_cast<T *>(out.ptr())); + } + }, + 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); + + /** SIMD vector types */ + using q16_t = typename wrapper::traits::promote_t<T>; + using q32_t = typename wrapper::traits::promote_t<q16_t>; + + 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(); + const int src_w = src->info()->dimension(0); + const int src_h = src->info()->dimension(1); + const T fill_value = (pool_info.pool_type == PoolingType::AVG) ? T(0) : std::numeric_limits<T>::min(); + const int stridex_in_bytes = static_cast<int>(src->info()->strides_in_bytes().x()); + const int stridey_in_bytes = static_cast<int>(src->info()->strides_in_bytes().y()); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + T res = std::numeric_limits<T>::min(); + + if (pool_info.pool_type != PoolingType::MAX) + { + q32_t sres = 0; + + // Calculate scale + const float scale = calculate_avg_scale_pool2d( + 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) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto in_ptr = reinterpret_cast<const T *>( + in.ptr() + (x - pool_pad_left) * stridex_in_bytes + (y - pool_pad_top) * stridey_in_bytes); + + const int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + const T data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *in_ptr; + sres += data; + } + } + // Divide by scale + res = static_cast<T>(support::cpp11::round(sres * scale)); + } + else + { + for (int y = 0; y < pool_size_y; ++y) + { + for (int x = 0; x < pool_size_x; ++x) + { + const auto in_ptr = reinterpret_cast<const T *>( + in.ptr() + (x - pool_pad_left) * stridex_in_bytes + (y - pool_pad_top) * stridey_in_bytes); + + const int idx = x + id.x() * pool_stride_x - pool_pad_left; + const int idy = y + id.y() * pool_stride_y - pool_pad_top; + const T data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *in_ptr; + res = std::max(res, data); + } + } + } + // 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 |