From 572659a0e5dd1086b1c7d16fe331ff73d2acd93a Mon Sep 17 00:00:00 2001 From: Adnan AlSinan Date: Tue, 15 Mar 2022 18:46:42 +0000 Subject: Add CPU Pool3d FP16/32 implementation - Add implementation for the CPU pooling 3d layer. - NDHWC data layout support - Support FP32/FP16. - Add Pool3d to the operator list. - Fix CL Pool3d kernel comments to generate the operator list. Resolves: COMPMID-4671 Signed-off-by: Adnan AlSinan Change-Id: I92478a154beb12541525b648ed3dd5a58c8f27fa Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7311 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins --- src/cpu/kernels/pool3d/list.h | 42 ++++ src/cpu/kernels/pool3d/neon/fp16.cpp | 36 +++ src/cpu/kernels/pool3d/neon/fp32.cpp | 34 +++ src/cpu/kernels/pool3d/neon/impl.cpp | 460 +++++++++++++++++++++++++++++++++++ src/cpu/kernels/pool3d/neon/impl.h | 42 ++++ 5 files changed, 614 insertions(+) create mode 100644 src/cpu/kernels/pool3d/list.h create mode 100644 src/cpu/kernels/pool3d/neon/fp16.cpp create mode 100644 src/cpu/kernels/pool3d/neon/fp32.cpp create mode 100644 src/cpu/kernels/pool3d/neon/impl.cpp create mode 100644 src/cpu/kernels/pool3d/neon/impl.h (limited to 'src/cpu/kernels/pool3d') diff --git a/src/cpu/kernels/pool3d/list.h b/src/cpu/kernels/pool3d/list.h new file mode 100644 index 0000000000..ece780eb0b --- /dev/null +++ b/src/cpu/kernels/pool3d/list.h @@ -0,0 +1,42 @@ +/* + * Copyright (c) 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_POOLING3D_LIST_H +#define SRC_CORE_NEON_KERNELS_POOLING3D_LIST_H + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_POOLING_KERNEL(func_name) \ + void func_name(const ITensor *src0, ITensor *dst0, Pooling3dLayerInfo &, const Window &window) + +DECLARE_POOLING_KERNEL(neon_fp16_pool3d); +DECLARE_POOLING_KERNEL(neon_fp32_pool3d); + +#undef DECLARE_POOLING_KERNEL + +} // namespace cpu +} // namespace arm_compute + +#endif // SRC_CORE_NEON_KERNELS_POOLING3D_LIST_H \ No newline at end of file diff --git a/src/cpu/kernels/pool3d/neon/fp16.cpp b/src/cpu/kernels/pool3d/neon/fp16.cpp new file mode 100644 index 0000000000..b79bcd93b5 --- /dev/null +++ b/src/cpu/kernels/pool3d/neon/fp16.cpp @@ -0,0 +1,36 @@ +/* + * Copyright (c) 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. + */ +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +#include "src/cpu/kernels/pool3d/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_fp16_pool3d(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window) +{ + return poolingMxNxD_fp_neon_ndhwc(src, dst0, pool_info, window); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ \ No newline at end of file diff --git a/src/cpu/kernels/pool3d/neon/fp32.cpp b/src/cpu/kernels/pool3d/neon/fp32.cpp new file mode 100644 index 0000000000..2c06a9d57a --- /dev/null +++ b/src/cpu/kernels/pool3d/neon/fp32.cpp @@ -0,0 +1,34 @@ +/* + * Copyright (c) 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. + */ +#include "src/cpu/kernels/pool3d/neon/impl.h" +namespace arm_compute +{ +namespace cpu +{ +void neon_fp32_pool3d(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window) +{ + return poolingMxNxD_fp_neon_ndhwc(src, dst0, pool_info, window); +} +} // namespace cpu +} // namespace arm_compute \ No newline at end of file diff --git a/src/cpu/kernels/pool3d/neon/impl.cpp b/src/cpu/kernels/pool3d/neon/impl.cpp new file mode 100644 index 0000000000..bb3999b104 --- /dev/null +++ b/src/cpu/kernels/pool3d/neon/impl.cpp @@ -0,0 +1,460 @@ +/* + * Copyright (c) 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. + */ +#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/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/core/helpers/WindowHelpers.h" + +#include "src/cpu/kernels/pool3d/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace +{ +inline float calculate_avg_scale(bool exclude_padding, const Coordinates &id, const int pool_size_x, const int pool_size_y, const int pool_size_z, const int upper_bound_w, + const int upper_bound_h, const int upper_bound_d, const int pad_x, const int pad_y, const int pad_z, const int stride_x, const int stride_y, const int stride_z) +{ + // Based on NDHWC + int start_x = id[1] * stride_x - pad_x; + int start_y = id[2] * stride_y - pad_y; + int start_z = id[3] * stride_z - pad_z; + + const int end_x = std::min(start_x + pool_size_x, upper_bound_w); + const int end_y = std::min(start_y + pool_size_y, upper_bound_h); + const int end_z = std::min(start_z + pool_size_z, upper_bound_d); + if(exclude_padding) + { + start_x = std::max(0, start_x); + start_y = std::max(0, start_y); + start_z = std::max(0, start_z); + } + return 1.f / ((end_y - start_y) * (end_x - start_x) * (end_z - start_z)); +} + + +template +void max_poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window_out, + const int window_start_x, const int window_end_x, const int window_step_x) + +{ + using vtype = wrapper::traits::neon_bitvector; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast(pool_info.stride.width); + int pool_stride_y = static_cast(pool_info.stride.height); + int pool_stride_z = static_cast(pool_info.stride.depth); + + 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_size_z = pool_info.is_global_pooling ? src->info()->tensor_shape()[3] : pool_info.pool_size.depth; + + const int pool_pad_top = static_cast(pool_info.padding.top); + const int pool_pad_left = static_cast(pool_info.padding.left); + const int pool_pad_front = static_cast(pool_info.padding.front); + + const int input_dim_w = src->info()->dimension(1); + const int input_dim_h = src->info()->dimension(2); + const int input_dim_d = src->info()->dimension(3); + + const int y_stride = static_cast(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast(src->info()->strides_in_bytes()[4]); + + const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); + + Iterator out(dst0, window_out); + + vector_type vres; + execute_window_loop(window_out, [&](const Coordinates & id) + { + // Computing the theoretical input starting/ending points + const int in_idx_width = static_cast(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast(id[3]) * pool_stride_z - pool_pad_front; + + const int pool_start_x = std::max(0, -in_idx_width); + const int pool_end_x_t = std::min(input_dim_w + pool_pad_left - in_idx_width, pool_size_x); + const int pool_start_y = std::max(0, -in_idx_height); + const int pool_end_y_t = std::min(input_dim_h + pool_pad_top - in_idx_height, pool_size_y); + + const int pool_start_z = std::max(0, -in_idx_depth); + const int pool_end_z_t = std::min(input_dim_d + pool_pad_front - in_idx_depth, pool_size_z); + + // The end of width to consider in calculation should exclude PAD_X, PAD_Y and PAD_Z + const int pool_end_x = std::min(pool_end_x_t, input_dim_w - in_idx_width); + const int pool_end_y = std::min(pool_end_y_t, input_dim_h - in_idx_height); + const int pool_end_z = std::min(pool_end_z_t, input_dim_d - in_idx_depth); + + const uint8_t *in_ptr_n = in_ptr_start + id[4] * n_stride; + + int x_off = window_start_x; + + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) // C + { + vres = wrapper::vdup_n(static_cast(-std::numeric_limits::infinity()), tag_type()); + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const vector_type data = wrapper::vloadq(reinterpret_cast(in_ptr_x) + x_off); + vres = wrapper::vmax(vres, data); + } + } + } + // Store result + wrapper::vstore(reinterpret_cast(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + T res(0); + res = -std::numeric_limits::infinity(); + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const T data = *(reinterpret_cast(in_ptr_x) + x_off); + res = std::max(res, data); + } + } + } + // Store result + *(reinterpret_cast(out.ptr()) + x_off) = res; + } + }, + out); +} + +template +void avg_poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, + const Window &window_out, const int window_start_x, const int window_end_x, const int window_step_x) +{ + using vtype = wrapper::traits::neon_bitvector; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast(pool_info.stride.width); + int pool_stride_y = static_cast(pool_info.stride.height); + int pool_stride_z = static_cast(pool_info.stride.depth); + + 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_size_z = pool_info.is_global_pooling ? src->info()->tensor_shape()[3] : pool_info.pool_size.depth; + + const int pool_pad_top = static_cast(pool_info.padding.top); + const int pool_pad_bottom = static_cast(pool_info.padding.bottom); + const int pool_pad_left = static_cast(pool_info.padding.left); + const int pool_pad_right = static_cast(pool_info.padding.right); + const int pool_pad_front = static_cast(pool_info.padding.front); + const int pool_pad_back = static_cast(pool_info.padding.back); + + 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 int upper_bound_d = src->info()->dimension(3) + (pool_info.exclude_padding ? 0 : pool_pad_back); + + const int input_dim_w = src->info()->dimension(1); + const int input_dim_h = src->info()->dimension(2); + const int input_dim_d = src->info()->dimension(3); + + const int y_stride = static_cast(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast(src->info()->strides_in_bytes()[4]); + + const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); + + Iterator out(dst0, window_out); + + vector_type vres; + execute_window_loop(window_out, [&](const Coordinates & id) + { + // Computing the theoretical input starting/ending points + const int in_idx_width = static_cast(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast(id[3]) * pool_stride_z - pool_pad_front; + + const int pool_start_x = std::max(0, -in_idx_width); + const int pool_end_x_t = std::min(input_dim_w + pool_pad_left - in_idx_width, pool_size_x); + const int pool_start_y = std::max(0, -in_idx_height); + const int pool_end_y_t = std::min(input_dim_h + pool_pad_top - in_idx_height, pool_size_y); + + const int pool_start_z = std::max(0, -in_idx_depth); + const int pool_end_z_t = std::min(input_dim_d + pool_pad_front - in_idx_depth, pool_size_z); + + // The end of width to consider in calculation should exclude PAD_X, PAD_Y and PAD_Z + const int pool_end_x = std::min(pool_end_x_t, input_dim_w - in_idx_width); + const int pool_end_y = std::min(pool_end_y_t, input_dim_h - in_idx_height); + const int pool_end_z = std::min(pool_end_z_t, input_dim_d - in_idx_depth); + + const uint8_t *in_ptr_n = in_ptr_start + id[4] * n_stride; + + // Calculate scale + const float scale = calculate_avg_scale(pool_info.exclude_padding, id, pool_size_x, pool_size_y, pool_size_z, upper_bound_w, upper_bound_h, upper_bound_d, pool_pad_left, + pool_pad_top, pool_pad_front, pool_stride_x, + pool_stride_y, pool_stride_z); + const vector_type scale_v = wrapper::vdup_n(static_cast(scale), tag_type()); + + int x_off = window_start_x; + + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) // C + { + // Perform pooling + vres = wrapper::vdup_n(static_cast(0.0f), tag_type()); + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const vector_type data = wrapper::vloadq(reinterpret_cast(in_ptr_x) + x_off); + vres = wrapper::vadd(vres, data); + } + } + } + + // Divide by scale + vres = wrapper::vmul(vres, scale_v); + + // Store result + wrapper::vstore(reinterpret_cast(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + T res(0); + + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const T data = *(reinterpret_cast(in_ptr_x) + x_off); + res += data; + } + } + } + + // Divide by scale + res *= scale; + + // Store result + *(reinterpret_cast(out.ptr()) + x_off) = res; + } + }, + out); +} + +template +void l2_poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, + const Window &window_out, const int window_start_x, const int window_end_x, const int window_step_x) +{ + using vtype = wrapper::traits::neon_bitvector; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast(pool_info.stride.width); + int pool_stride_y = static_cast(pool_info.stride.height); + int pool_stride_z = static_cast(pool_info.stride.depth); + + 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_size_z = pool_info.is_global_pooling ? src->info()->tensor_shape()[3] : pool_info.pool_size.depth; + + const int pool_pad_top = static_cast(pool_info.padding.top); + const int pool_pad_bottom = static_cast(pool_info.padding.bottom); + const int pool_pad_left = static_cast(pool_info.padding.left); + const int pool_pad_right = static_cast(pool_info.padding.right); + const int pool_pad_front = static_cast(pool_info.padding.front); + const int pool_pad_back = static_cast(pool_info.padding.back); + + 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 int upper_bound_d = src->info()->dimension(3) + (pool_info.exclude_padding ? 0 : pool_pad_back); + + const int input_dim_w = src->info()->dimension(1); + const int input_dim_h = src->info()->dimension(2); + const int input_dim_d = src->info()->dimension(3); + + const int y_stride = static_cast(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast(src->info()->strides_in_bytes()[4]); + + const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); + + Iterator out(dst0, window_out); + + vector_type vres; + execute_window_loop(window_out, [&](const Coordinates & id) + { + // Computing the theoretical input starting/ending points + const int in_idx_width = static_cast(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast(id[3]) * pool_stride_z - pool_pad_front; + + const int pool_start_x = std::max(0, -in_idx_width); + const int pool_end_x_t = std::min(input_dim_w + pool_pad_left - in_idx_width, pool_size_x); + const int pool_start_y = std::max(0, -in_idx_height); + const int pool_end_y_t = std::min(input_dim_h + pool_pad_top - in_idx_height, pool_size_y); + + const int pool_start_z = std::max(0, -in_idx_depth); + const int pool_end_z_t = std::min(input_dim_d + pool_pad_front - in_idx_depth, pool_size_z); + + // The end of width to consider in calculation should exclude PAD_X, PAD_Y and PAD_Z + const int pool_end_x = std::min(pool_end_x_t, input_dim_w - in_idx_width); + const int pool_end_y = std::min(pool_end_y_t, input_dim_h - in_idx_height); + const int pool_end_z = std::min(pool_end_z_t, input_dim_d - in_idx_depth); + + const uint8_t *in_ptr_n = in_ptr_start + id[4] * n_stride; + + // Calculate scale + const float scale = calculate_avg_scale(pool_info.exclude_padding, id, pool_size_x, pool_size_y, pool_size_z, upper_bound_w, upper_bound_h, upper_bound_d, pool_pad_left, + pool_pad_top, pool_pad_front, pool_stride_x, + pool_stride_y, pool_stride_z); + + int x_off = window_start_x; + + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) // C + { + // Perform pooling + vres = wrapper::vdup_n(static_cast(0.0f), tag_type()); + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const vector_type data = wrapper::vloadq(reinterpret_cast(in_ptr_x) + x_off); + vres = wrapper::vmla(vres, data, data); + } + } + } + + const vector_type scale_v = wrapper::vdup_n(static_cast(scale), tag_type()); + + // Divide by scale + vres = wrapper::vmul(vres, scale_v); + + // Calculate square-root + vres = wrapper::vinv(wrapper::vinvsqrt(vres)); + + // Store result + wrapper::vstore(reinterpret_cast(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + T res(0); + + for(int z = pool_start_z; z < pool_end_z; ++z) + { + const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride; + for(int y = pool_start_y; y < pool_end_y; ++y) + { + const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride; + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride; + const T data = *(reinterpret_cast(in_ptr_x) + x_off); + res += data * data; + } + } + } + + // Divide by scale + res *= scale; + + // Square root + res = std::sqrt(res); + + // Store result + *(reinterpret_cast(out.ptr()) + x_off) = res; + } + }, + out); +} +} // namespace + +template +void poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &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 = 16 / sizeof(T); + Window window_out = window; + + // Needed to handle loop left-over + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + switch(pool_info.pool_type) + { + case PoolingType::MAX: + max_poolingMxNxD_fp_neon_ndhwc(src, dst0, pool_info, window_out, window_start_x, window_end_x, window_step_x); + break; + case PoolingType::AVG: + avg_poolingMxNxD_fp_neon_ndhwc(src, dst0, pool_info, window_out, window_start_x, window_end_x, window_step_x); + break; + case PoolingType::L2: + l2_poolingMxNxD_fp_neon_ndhwc(src, dst0, pool_info, window_out, window_start_x, window_end_x, window_step_x); + break; + default: + ARM_COMPUTE_ERROR("Pool operation not supported"); + } +} + +template void poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window); +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +template void poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window); +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/pool3d/neon/impl.h b/src/cpu/kernels/pool3d/neon/impl.h new file mode 100644 index 0000000000..829a9bd192 --- /dev/null +++ b/src/cpu/kernels/pool3d/neon/impl.h @@ -0,0 +1,42 @@ +/* + * Copyright (c) 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_POOLING_3D_LAYER_IMPL_H +#define SRC_CORE_POOLING_3D_LAYER_IMPL_H + +#include "arm_compute/core/Helpers.h" + +namespace arm_compute +{ +// Forward declarations +class ITensor; +class Window; +struct Pooling3dLayerInfo; +namespace cpu +{ +template +void poolingMxNxD_fp_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window); + +} // namespace cpu +} // namespace arm_compute +#endif //define SRC_CORE_POOLING_3D_LAYER_IMPL_H -- cgit v1.2.1