diff options
Diffstat (limited to 'src/cpu/kernels/CpuDirectConv2dKernel.cpp')
| -rw-r--r-- | src/cpu/kernels/CpuDirectConv2dKernel.cpp | 425 |
1 files changed, 29 insertions, 396 deletions
diff --git a/src/cpu/kernels/CpuDirectConv2dKernel.cpp b/src/cpu/kernels/CpuDirectConv2dKernel.cpp index f3560156bd..a4cdddee5e 100644 --- a/src/cpu/kernels/CpuDirectConv2dKernel.cpp +++ b/src/cpu/kernels/CpuDirectConv2dKernel.cpp @@ -22,26 +22,14 @@ * SOFTWARE. */ #include "src/cpu/kernels/CpuDirectConv2dKernel.h" +#include "src/cpu/kernels/directconv2d/list.h" -#include "src/core/NEON/kernels/detail/NEDirectConvolutionDetail.h" -#include "src/core/NEON/wrapper/wrapper.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/IAccessWindow.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/AccessWindowStatic.h" #include "src/core/CPP/Validate.h" -#include "src/core/NEON/NEFixedPoint.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" -#include <algorithm> - using namespace arm_compute::detail; namespace arm_compute @@ -50,8 +38,25 @@ namespace cpu { namespace kernels { -namespace +static const std::vector<CpuDirectConv2dKernel::DirectConv2dKernel> available_kernels = { + { + "neon_fp32_nhwc_directconv2d", + [](const DataTypeDataLayoutISASelectorData & data) { return data.dt == DataType::F32 && data.dl == DataLayout::NHWC; }, + REGISTER_FP32_NEON(arm_compute::cpu::kernels::neon_fp32_nhwc_directconv2d) + }, + { + "neon_fp32_nchw_directconv2d", + [](const DataTypeDataLayoutISASelectorData & data) { return data.dt == DataType::F32 && data.dl == DataLayout::NCHW; }, + REGISTER_FP32_NEON(arm_compute::cpu::kernels::neon_fp32_nchw_directconv2d) + }, + { + "neon_fp16_nchw_directconv2d", + [](const DataTypeDataLayoutISASelectorData & data) { return data.dt == DataType::F16 && data.dl == DataLayout::NCHW && data.isa.fp16; }, + REGISTER_FP16_NEON(arm_compute::cpu::kernels::neon_fp16_nchw_directconv2d) + }, +}; + Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); @@ -99,346 +104,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso return std::make_pair(err, win); } -bool have_zero_x_internal_padding(ITensorInfo *src, const ITensorInfo *weights) -{ - return (src->padding().left == 0 && weights->padding().left == 0 && src->padding().right == 0 && weights->padding().right == 0); -} - -} // namespace - -template <typename T> -void CpuDirectConv2dKernel::convolve_nhwc_optimized(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst) -{ - // This function assumes that input and weights have not padding in channel - - // Declare useful types - using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>; - using vector_type = typename vtype::type; - using tag_type = typename vtype::tag_type; - - // Scalar quantities - const int element_size = src->info()->element_size(); - const int input_stride_w = src->info()->strides_in_bytes().y() / element_size; - const int input_stride_h = src->info()->strides_in_bytes().z() / element_size; - const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size; - const int input_dim_w = src->info()->dimension(1); - const int input_dim_h = src->info()->dimension(2); - - const int output_stride_c = dst->info()->strides_in_bytes().x(); - - const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().y() / element_size; - const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().z() / element_size; - const int kernel_dim_w = weights->info()->dimension(1); - const int kernel_dim_h = weights->info()->dimension(2); - - const int conv_pad_top = _conv_info.pad_top(); - const int conv_pad_left = _conv_info.pad_left(); - const int conv_stride_w = std::get<0>(_conv_info.stride()); - const int conv_stride_h = std::get<1>(_conv_info.stride()); - - // Setup input window for the output iterator - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - // Setup input window for the weights iterator - Window window_w = calculate_max_window(*weights->info(), Steps()); - window_w.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimY, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimZ, Window::Dimension(0, 1, 1)); - - Iterator out(dst, window_out); - Iterator wei(weights, window_w); - - constexpr int num_elems_read_per_iteration = 16 / sizeof(T); - /* - * This implementation parallelize the full WC plane of input and weights by - * treating them as series of elements. So for example, a 3x3 weights and - * floating point vector operations of 4 elements per time, the first 3 - * channel elements of the first row would be taken and additionally the first - * element of the second row. The 9 elements in each single WC weight plane - * would require 2 4-element vector operations and a last single element operation. - * - * This works since when we create the input vector to multiply with the weights, - * the exact required elements are loaded in the same order. Therefore the - * multiplication works on the correct input/weight elements. - */ - execute_window_loop(window_out, [&](const Coordinates & id) - { - /* - * In here we create theoretical indexes which then we validate for both - * inputs and weights. - * As a reminder, this loop take each output point in NHW, C is treated - * in the weights loop. - */ - // We are computing the theoretical starting input starting points - const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; - const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; - const int in_w_end_t = in_w_start_t + kernel_dim_w; - const int in_h_end_t = in_h_start_t + kernel_dim_h; - - // We are computing the valid initial and ending input points by checking the borders - const int in_w_start = std::max(in_w_start_t, 0); - const int in_h_start = std::max(in_h_start_t, 0); - const int in_w_end = std::min(in_w_end_t, input_dim_w); - const int in_h_end = std::min(in_h_end_t, input_dim_h); - - // We use the input points to select the valid weight points to use - const int index_wc_start = (in_w_start - in_w_start_t) * kernel_stride_w; - const int index_h_start = in_h_start - in_h_start_t; - const int index_wc_end = (kernel_dim_w - (in_w_end_t - in_w_end)) * kernel_stride_w; - const int index_h_end = kernel_dim_h - (in_h_end_t - in_h_end); - - execute_window_loop(window_w, [&](const Coordinates & id_w) - { - /* - * This is the loop in the weights, and it goes along N (the batches) - * As a reminder, the batches of the weights are translated into the - * channels of the output - */ - const T *in_ptr_row = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) - + id[3] * input_stride_n + in_w_start * input_stride_w + in_h_start * input_stride_h; - const T *weights_ptr_row = reinterpret_cast<const T *>(wei.ptr()) + index_h_start * kernel_stride_h; - uint8_t *out_ptr = out.ptr() + id_w[3] * output_stride_c; - - T out_temp = static_cast<T>(0); - for(int index_h = index_h_start; index_h < index_h_end; ++index_h, in_ptr_row += input_stride_h, weights_ptr_row += kernel_stride_h) - { - const T *in_ptr_mover = in_ptr_row; - int index_wc = index_wc_start; - vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); - for(; index_wc <= index_wc_end - num_elems_read_per_iteration; index_wc += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration) - { - const auto src_vec = wrapper::vloadq(in_ptr_mover); - const auto w_vec = wrapper::vloadq(weights_ptr_row + index_wc); - out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); - } - out_temp += vreduce(out_temp_vec); - for(; index_wc < index_wc_end; ++index_wc, ++in_ptr_mover) - { - const auto src_val = *(in_ptr_mover); - const auto w_val = *(weights_ptr_row + index_wc); - out_temp += src_val * w_val; - } - } - *(reinterpret_cast<T *>(out_ptr)) = out_temp; - }, - wei); - }, - out); -} - -template <typename T> -void CpuDirectConv2dKernel::convolve_nhwc(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst) -{ - // Declare useful types - using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>; - using vector_type = typename vtype::type; - using tag_type = typename vtype::tag_type; - - // Scalar quantities - const int element_size = src->info()->element_size(); - const int input_stride_w = src->info()->strides_in_bytes().y() / element_size; - const int input_stride_h = src->info()->strides_in_bytes().z() / element_size; - const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size; - const int input_dim_w = src->info()->dimension(1); - const int input_dim_h = src->info()->dimension(2); - - const int output_stride_c = dst->info()->strides_in_bytes().x(); - - const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().y() / element_size; - const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().z() / element_size; - const int kernel_dim_w = weights->info()->dimension(1); - const int kernel_dim_h = weights->info()->dimension(2); - - const int conv_pad_top = _conv_info.pad_top(); - const int conv_pad_left = _conv_info.pad_left(); - const int conv_stride_w = std::get<0>(_conv_info.stride()); - const int conv_stride_h = std::get<1>(_conv_info.stride()); - - // Setup input window for the output iterator - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - // Setup input window for the weights iterator - Window window_w = calculate_max_window(*weights->info(), Steps()); - window_w.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimY, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimZ, Window::Dimension(0, 1, 1)); - - Iterator out(dst, window_out); - Iterator wei(weights, window_w); - - constexpr int num_elems_read_per_iteration = 16 / sizeof(T); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - // We are computing the theoretical starting input starting points - const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; - const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; - const int in_w_end_t = in_w_start_t + kernel_dim_w; - const int in_h_end_t = in_h_start_t + kernel_dim_h; - - // We are computing the valid initial and ending input points by checking the borders - const int in_w_start = std::max(in_w_start_t, 0); - const int in_h_start = std::max(in_h_start_t, 0); - const int in_w_end = std::min(in_w_end_t, input_dim_w); - const int in_h_end = std::min(in_h_end_t, input_dim_h); - - // We use the input points to select the valid weight points to use - const int wei_w_start = in_w_start - in_w_start_t; - const int wei_h_start = in_h_start - in_h_start_t; - const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end); - const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); - - const int index_c_end = weights->info()->dimension(0); - const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[3] * input_stride_n; - - execute_window_loop(window_w, [&](const Coordinates & id_w) - { - const T *const weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); - uint8_t *out_ptr = out.ptr() + id_w[3] * output_stride_c; - - T out_temp = static_cast<T>(0); - for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h) - { - const T *const in_ptr_row = in_ptr_start + index_in_h * input_stride_h; - const T *const weights_ptr_row = weights_ptr_start + index_wei_h * kernel_stride_h; - for(int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; ++index_wei_w, ++index_in_w) - { - const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w; - const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w; - int index_c = 0; - vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); - for(; index_c <= index_c_end - num_elems_read_per_iteration; index_c += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration, weights_ptr_mover += num_elems_read_per_iteration) - { - const auto src_vec = wrapper::vloadq(in_ptr_mover); - const auto w_vec = wrapper::vloadq(weights_ptr_mover); - out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); - } - out_temp += vreduce(out_temp_vec); - for(; index_c < index_c_end; ++index_c, ++in_ptr_mover, ++weights_ptr_mover) - { - const auto src_val = *(in_ptr_mover); - const auto w_val = *(weights_ptr_mover); - out_temp += src_val * w_val; - } - } - } - *(reinterpret_cast<T *>(out_ptr)) = out_temp; - }, - wei); - }, - out); -} - -template <typename T> -void CpuDirectConv2dKernel::convolve_nchw(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst) -{ - // Declare useful types - using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>; - using vector_type = typename vtype::type; - using tag_type = typename vtype::tag_type; - - // Scalar quantities - const int element_size = src->info()->element_size(); - const int input_stride_w = src->info()->strides_in_bytes()[0] / element_size; - const int input_stride_h = src->info()->strides_in_bytes()[1] / element_size; - const int input_stride_c = src->info()->strides_in_bytes()[2] / element_size; - const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size; - - const int input_dim_w = src->info()->dimension(0); - const int input_dim_h = src->info()->dimension(1); - - const int output_stride_c = dst->info()->strides_in_bytes()[2]; - - const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().x() / element_size; - const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().y() / element_size; - const unsigned int kernel_stride_c = weights->info()->strides_in_bytes().z() / element_size; - - const int kernel_dim_w = weights->info()->dimension(0); - const int kernel_dim_h = weights->info()->dimension(1); - - const int conv_pad_top = _conv_info.pad_top(); - const int conv_pad_left = _conv_info.pad_left(); - const int conv_stride_w = std::get<0>(_conv_info.stride()); - const int conv_stride_h = std::get<1>(_conv_info.stride()); - - // Setup input window for the output iterator - Window window_out = window; - window_out.set(Window::DimZ, Window::Dimension(0, 1, 1)); - - // Setup input window for the weights iterator - Window window_w = calculate_max_window(*weights->info(), Steps()); - window_w.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimY, Window::Dimension(0, 1, 1)); - window_w.set(Window::DimZ, Window::Dimension(0, 1, 1)); - - Iterator out(dst, window_out); - Iterator wei(weights, window_w); - - constexpr int num_elems_read_per_iteration = 16 / sizeof(T); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - // We are computing the theoretical starting input starting points - const int in_w_start_t = static_cast<int>(id.x()) * conv_stride_w - conv_pad_left; - const int in_h_start_t = static_cast<int>(id.y()) * conv_stride_h - conv_pad_top; - const int in_w_end_t = in_w_start_t + kernel_dim_w; - const int in_h_end_t = in_h_start_t + kernel_dim_h; - - // We are computing the valid initial and ending input points by checking the borders - const int in_w_start = std::max(in_w_start_t, 0); - const int in_h_start = std::max(in_h_start_t, 0); - const int in_w_end = std::min(in_w_end_t, input_dim_w); - const int in_h_end = std::min(in_h_end_t, input_dim_h); - - // We use the input points to select the valid weight points to use - const int wei_w_start = in_w_start - in_w_start_t; - const int wei_h_start = in_h_start - in_h_start_t; - const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); - - const int index_c_end = weights->info()->dimension(2); - const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[3] * input_stride_n; - execute_window_loop(window_w, [&](const Coordinates & id_w) - { - const T *const weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); - uint8_t *out_ptr = out.ptr() + id_w[3] * output_stride_c; - T out_temp = static_cast<T>(0); - - for(int index_wei_c = 0, index_in_c = 0; index_wei_c < index_c_end; ++index_wei_c, ++index_in_c) - { - const T *const in_ptr_row_0 = in_ptr_start + index_in_c * input_stride_c; - const T *const weights_ptr_row_0 = weights_ptr_start + index_wei_c * kernel_stride_c; - for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h) - { - const T *in_ptr_row = in_ptr_row_0 + index_in_h * input_stride_h; - const T *weights_ptr_row = weights_ptr_row_0 + index_wei_h * kernel_stride_h; - int index_w = in_w_start; - int index_wei_w = wei_w_start; - vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); - for(; index_w <= ((in_w_end - num_elems_read_per_iteration)); index_w += num_elems_read_per_iteration, index_wei_w += num_elems_read_per_iteration) - { - const auto src_vec = wrapper::vloadq(in_ptr_row + index_w * input_stride_w); - const auto w_vec = wrapper::vloadq(weights_ptr_row + index_wei_w * kernel_stride_w); - out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); - } - out_temp += vreduce(out_temp_vec); - for(; index_w < in_w_end; ++index_w, ++index_wei_w) - { - const auto src_val = *(in_ptr_row + index_w * input_stride_w); - const auto w_val = *(weights_ptr_row + index_wei_w * kernel_stride_w); - out_temp += src_val * w_val; - } - } - } - *(reinterpret_cast<T *>(out_ptr)) = out_temp; - - }, - wei); - }, - out); -} - void CpuDirectConv2dKernel::configure(ITensorInfo *src, ITensorInfo *weights, ITensorInfo *dst, const PadStrideInfo &conv_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); @@ -484,53 +149,21 @@ void CpuDirectConv2dKernel::run_op(ITensorPack &tensors, const Window &window, c auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); auto dst = tensors.get_tensor(TensorType::ACL_DST); - if(_data_layout == DataLayout::NCHW) - { - switch(src->info()->data_type()) - { -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - { - convolve_nchw<float16_t>(window, src, weights, dst); - break; - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - { - convolve_nchw<float>(window, src, weights, dst); - break; - } - default: - ARM_COMPUTE_ERROR("Data type not supported"); - break; - } - } - else - { - switch(src->info()->data_type()) - { - case DataType::F32: - { - if(have_zero_x_internal_padding(src->info(), weights->info())) - { - convolve_nhwc_optimized<float>(window, src, weights, dst); - } - else - { - convolve_nhwc<float>(window, src, weights, dst); - } - break; - } - default: - ARM_COMPUTE_ERROR("Data type not supported"); - break; - } - } + const auto *uk = CpuDirectConv2dKernel::get_implementation(DataTypeDataLayoutISASelectorData{ src->info()->data_type(), _data_layout, CPUInfo::get().get_isa() }); + ARM_COMPUTE_ERROR_ON(uk == nullptr); + + uk->ukernel(window, src, weights, dst, _conv_info); } const char *CpuDirectConv2dKernel::name() const { return "CpuDirectConvolutionLayerKernel"; } + +const std::vector<CpuDirectConv2dKernel::DirectConv2dKernel> &CpuDirectConv2dKernel::get_available_kernels() +{ + return available_kernels; +} + } // namespace kernels } // namespace cpu } // namespace arm_compute |