diff options
Diffstat (limited to 'src/cpu/kernels')
| -rw-r--r-- | src/cpu/kernels/CpuPool3dKernel.cpp | 176 | ||||
| -rw-r--r-- | src/cpu/kernels/CpuPool3dKernel.h | 90 | ||||
| -rw-r--r-- | src/cpu/kernels/pool3d/list.h | 42 | ||||
| -rw-r--r-- | src/cpu/kernels/pool3d/neon/fp16.cpp | 36 | ||||
| -rw-r--r-- | src/cpu/kernels/pool3d/neon/fp32.cpp | 34 | ||||
| -rw-r--r-- | src/cpu/kernels/pool3d/neon/impl.cpp | 460 | ||||
| -rw-r--r-- | src/cpu/kernels/pool3d/neon/impl.h | 42 |
7 files changed, 880 insertions, 0 deletions
diff --git a/src/cpu/kernels/CpuPool3dKernel.cpp b/src/cpu/kernels/CpuPool3dKernel.cpp new file mode 100644 index 0000000000..3321967d2f --- /dev/null +++ b/src/cpu/kernels/CpuPool3dKernel.cpp @@ -0,0 +1,176 @@ +/* + * 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/CpuPool3dKernel.h" + +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "src/core/CPP/Validate.h" +#include "src/core/common/Registrars.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/cpu/kernels/pool3d/list.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +namespace +{ +using namespace misc::shape_calculator; + +static const std::vector<CpuPool3dKernel::Pooling3dKernel> available_kernels = +{ + { + "neon_fp16_ndhwc_poolMxNxD", + [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F16 && data.isa.fp16); }, + REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_pool3d) + }, + + { + "neon_fp32_ndhwc_poolMxNxD", + [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F32); }, + REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_pool3d) + } +}; + +Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_layout() != DataLayout::NDHWC, "Only NDHWC layout supported"); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + + const auto data_layout = src->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH); + + const bool is_global_pooling = pool_info.is_global_pooling; + const unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; + const unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; + const unsigned int pool_size_z = is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth; + + const unsigned int stride_x = pool_info.stride.x(); + const unsigned int stride_y = pool_info.stride.y(); + const unsigned int stride_z = pool_info.stride.z(); + + ARM_COMPUTE_RETURN_ERROR_ON((pool_size_x == 0) || (pool_size_y == 0) || (pool_size_z == 0)); + ARM_COMPUTE_RETURN_ERROR_ON((stride_x == 0) || (stride_y == 0) || (stride_z == 0)); + + int output_width = 0; + int output_height = 0; + int output_depth = 0; + + std::tie(output_width, output_height, output_depth) = scaled_3d_dimensions_signed(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height], src->tensor_shape()[idx_depth], + pool_size_x, pool_size_y, pool_size_z, pool_info); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_width < 1 || output_height < 1 || output_depth < 1), "Calculated output dimension size is invalid"); + + if(dst->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst); + TensorInfo out_info(TensorInfo(compute_pool3d_shape(src->tensor_shape(), pool_info), 1, dst->data_type(), DataLayout::NDHWC)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &out_info); + } + + const auto *uk = CpuPool3dKernel::get_implementation(DataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_isa() }); + ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr); + + return Status{}; +} +} //namespace + +void CpuPool3dKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const Pooling3dLayerInfo &pool_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + + // Perform validation step + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info)); + + // dst auto inizialitation if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_pool3d_shape(src->tensor_shape(), pool_info))); + + // Get data layout + const auto data_layout = src->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH); + + // Update pool size in case of global pooling + const bool is_global_pooling = pool_info.is_global_pooling; + const Size3D pool_size( + is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width, + is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height, + is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth); + + const auto *uk = CpuPool3dKernel::get_implementation(DataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_isa() }); + ARM_COMPUTE_ERROR_ON(uk == nullptr); + + // Set instance variables + _pool_info = pool_info; + _run_method = uk->ukernel; + _name = std::string("CpuPool3dKernel").append("/").append(uk->name); + + // Configure kernel window + Window win = calculate_max_window(*dst, Steps()); + ICpuKernel::configure(win); +} + +Status CpuPool3dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info)); + + return Status{}; +} + +void CpuPool3dKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + ARM_COMPUTE_ERROR_ON(_run_method == nullptr); + + const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + ITensor *dst = tensors.get_tensor(TensorType::ACL_DST_0); + + _run_method(src, dst, _pool_info, window); +} + +const char *CpuPool3dKernel::name() const +{ + return _name.c_str(); +} + +const std::vector<CpuPool3dKernel::Pooling3dKernel> &CpuPool3dKernel::get_available_kernels() +{ + return available_kernels; +} + +} // namespace kernels +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/cpu/kernels/CpuPool3dKernel.h b/src/cpu/kernels/CpuPool3dKernel.h new file mode 100644 index 0000000000..f762cfca9a --- /dev/null +++ b/src/cpu/kernels/CpuPool3dKernel.h @@ -0,0 +1,90 @@ +/* + * 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 ARM_COMPUTE_CPU_POOL3D_KERNEL_H +#define ARM_COMPUTE_CPU_POOL3D_KERNEL_H + +#include "arm_compute/core/Types.h" +#include "src/core/common/Macros.h" +#include "src/cpu/ICpuKernel.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +/** Interface for the kernel to perform Pooling 3D. */ +class CpuPool3dKernel : public ICpuKernel<CpuPool3dKernel> +{ +private: + /* Template function for Pooling 3D NDHWC */ + using Pooling3dKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, Pooling3dLayerInfo &, const Window &)>::type; + +public: + CpuPool3dKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuPool3dKernel); + /** Set the src, dst tensor and pooling info. + * + * Valid data type configurations: + * |src |dst | + * |:--------------|:--------------| + * |F16 |F16 | + * |F32 |F32 | + * + * @param[in] src Source tensor info. Data types supported: F16/F32. + * @param[out] dst Destination tensor info. Data types supported: Same as @p src. + * @param[in] pool_info Contains pooling operation information described in @ref Pooling3dLayerInfo. + */ + void configure(const ITensorInfo *src, ITensorInfo *dst, const Pooling3dLayerInfo &pool_info); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to CpuPool3dKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + const char *name() const override; + + struct Pooling3dKernel + { + const char *name; + const DataTypeISASelectorPtr is_selected; + Pooling3dKernelPtr ukernel; + }; + + static const std::vector<Pooling3dKernel> &get_available_kernels(); + +private: + Pooling3dLayerInfo _pool_info{}; + Pooling3dKernelPtr _run_method{ nullptr }; + std::string _name{}; +}; + +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /*ARM_COMPUTE_CPU_POOL3D_KERNEL_H */
\ No newline at end of file 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<float16_t>(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<float>(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 <typename T> +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<T, wrapper::traits::BitWidth::W128>; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast<int>(pool_info.stride.width); + int pool_stride_y = static_cast<int>(pool_info.stride.height); + int pool_stride_z = static_cast<int>(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<int>(pool_info.padding.top); + const int pool_pad_left = static_cast<int>(pool_info.padding.left); + const int pool_pad_front = static_cast<int>(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<int>(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast<int>(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast<int>(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<int>(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast<int>(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<T>(-std::numeric_limits<float>::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<const T *>(in_ptr_x) + x_off); + vres = wrapper::vmax(vres, data); + } + } + } + // Store result + wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + T res(0); + res = -std::numeric_limits<float>::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<const T *>(in_ptr_x) + x_off); + res = std::max(res, data); + } + } + } + // Store result + *(reinterpret_cast<T *>(out.ptr()) + x_off) = res; + } + }, + out); +} + +template <typename T> +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<T, wrapper::traits::BitWidth::W128>; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast<int>(pool_info.stride.width); + int pool_stride_y = static_cast<int>(pool_info.stride.height); + int pool_stride_z = static_cast<int>(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<int>(pool_info.padding.top); + const int pool_pad_bottom = static_cast<int>(pool_info.padding.bottom); + const int pool_pad_left = static_cast<int>(pool_info.padding.left); + const int pool_pad_right = static_cast<int>(pool_info.padding.right); + const int pool_pad_front = static_cast<int>(pool_info.padding.front); + const int pool_pad_back = static_cast<int>(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<int>(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast<int>(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast<int>(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<int>(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast<int>(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<T>(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<T>(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<const T *>(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<T *>(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<const T *>(in_ptr_x) + x_off); + res += data; + } + } + } + + // Divide by scale + res *= scale; + + // Store result + *(reinterpret_cast<T *>(out.ptr()) + x_off) = res; + } + }, + out); +} + +template <typename T> +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<T, wrapper::traits::BitWidth::W128>; + using vector_type = typename vtype::type; + using tag_type = typename vtype::tag_type; + + int pool_stride_x = static_cast<int>(pool_info.stride.width); + int pool_stride_y = static_cast<int>(pool_info.stride.height); + int pool_stride_z = static_cast<int>(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<int>(pool_info.padding.top); + const int pool_pad_bottom = static_cast<int>(pool_info.padding.bottom); + const int pool_pad_left = static_cast<int>(pool_info.padding.left); + const int pool_pad_right = static_cast<int>(pool_info.padding.right); + const int pool_pad_front = static_cast<int>(pool_info.padding.front); + const int pool_pad_back = static_cast<int>(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<int>(src->info()->strides_in_bytes().y()); + const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z()); + const int w_stride = static_cast<int>(src->info()->strides_in_bytes()[3]); + const int n_stride = static_cast<int>(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<int>(id.y()) * pool_stride_x - pool_pad_left; + const int in_idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top; + const int in_idx_depth = static_cast<int>(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<T>(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<const T *>(in_ptr_x) + x_off); + vres = wrapper::vmla(vres, data, data); + } + } + } + + const vector_type scale_v = wrapper::vdup_n(static_cast<T>(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<T *>(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<const T *>(in_ptr_x) + x_off); + res += data * data; + } + } + } + + // Divide by scale + res *= scale; + + // Square root + res = std::sqrt(res); + + // Store result + *(reinterpret_cast<T *>(out.ptr()) + x_off) = res; + } + }, + out); +} +} // namespace + +template <typename T> +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<T>(src, dst0, pool_info, window_out, window_start_x, window_end_x, window_step_x); + break; + case PoolingType::AVG: + avg_poolingMxNxD_fp_neon_ndhwc<T>(src, dst0, pool_info, window_out, window_start_x, window_end_x, window_step_x); + break; + case PoolingType::L2: + l2_poolingMxNxD_fp_neon_ndhwc<T>(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<float>(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<float16_t>(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 <typename T> +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 |