diff options
Diffstat (limited to 'arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h')
| -rw-r--r-- | arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h | 140 |
1 files changed, 131 insertions, 9 deletions
diff --git a/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h b/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h index e46c33018b..ba3223f66d 100644 --- a/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h +++ b/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,14 +24,16 @@ #ifndef __ARM_COMPUTE_NEGEMMINTERLEAVEDPREPAREBWRAPPERKERNEL_H__ #define __ARM_COMPUTE_NEGEMMINTERLEAVEDPREPAREBWRAPPERKERNEL_H__ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/INEKernel.h" #include "arm_compute/core/NEON/kernels/assembly/Helpers.h" #include "arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" namespace arm_compute { -class ITensor; - /** Unit of work for @ref NEGEMMInterleavedPrepareBWrapperKernel to process */ struct PrepareBWorkload { @@ -56,6 +58,84 @@ struct PrepareBWorkload unsigned int _kmax; /**< Last value to process along the K dimension. */ }; +namespace detail +{ +// Call the lambda function for each workload generated by the passed window. +template <typename strategy, bool use_buffer_manager, typename Lambda> +void for_each_element_in_window(const Window &window, const ITensor *b, ITensor *transformed_b, unsigned int N, unsigned int K, Lambda &&lambda) +{ + unsigned int wl_index = 0; + unsigned int num_buffers = 0, reshaped_block_size = 0; + + if(use_buffer_manager) + { + num_buffers = transformed_b->info()->tensor_shape()[1]; + reshaped_block_size = transformed_b->info()->strides_in_bytes().y(); + } + + unsigned int offset_transformed_b = transformed_b->info()->offset_first_element_in_bytes(); + execute_window_loop(window, [&](const Coordinates & coordinates) + { + const unsigned int x0 = coordinates.x(); + const unsigned int k0 = coordinates.y(); + const unsigned int multi = coordinates.z(); + + const unsigned int offset_b = b->info()->offset_element_in_bytes(Coordinates(0, 0, multi)); + const unsigned int xmax = std::min(x0 + window.x().step(), N); + const unsigned int kmax = std::min(k0 + window.y().step(), K); + + /* Figure out the size of each block. */ + unsigned int x_size = (xmax - x0); + unsigned int k_size = (kmax - k0); + + /* Round sizes up as needed. */ + x_size = ceil_to_multiple(x_size, strategy::out_width()); + k_size = ceil_to_multiple(k_size, strategy::k_unroll()); + + lambda(PrepareBWorkload(offset_b, offset_transformed_b, x0, xmax, k0, kmax)); + + //Each workload represents one block: + if(use_buffer_manager) + { + // Rotate through the BufferManager's buffers: + wl_index++; + offset_transformed_b = (wl_index % num_buffers) * reshaped_block_size; + } + else + { + offset_transformed_b += (x_size * k_size * sizeof(typename strategy::operand_type)); + } + }); +} + +// Calculate the size of transformed_b: +template <typename strategy> +unsigned int get_B_pretransposed_array_size(unsigned int N, unsigned int K, const BlockSizes &bs, unsigned int multis) +{ + // How many full blocks do N / K contain ? + size_t num_full_k = K / bs.k_block; + size_t num_full_x = N / bs.x_block; + + ARM_COMPUTE_ERROR_ON(bs.x_block % strategy::out_width() != 0); + ARM_COMPUTE_ERROR_ON(bs.k_block % strategy::k_unroll() != 0); + + size_t normal_x_size = bs.x_block; + size_t normal_k_size = bs.k_block; + + // Round up the leftovers to be a multiple of the strategy processing size: + size_t left_over_x_size = ceil_to_multiple(N % bs.x_block, strategy::out_width()); + size_t left_over_k_size = ceil_to_multiple(K % bs.k_block, strategy::k_unroll()); + + // Calculate the total size of the buffer: + size_t total = num_full_k * normal_k_size * (num_full_x * normal_x_size + left_over_x_size); + total += left_over_k_size * (left_over_x_size + num_full_x * normal_x_size); + + total *= multis; + + return total; +} +} // namespace detail + /** Common interface for the templated wrappers around the B reshape NEON assembly implementations */ class NEGEMMInterleavedPrepareBWrapperKernel : public INEKernel { @@ -93,7 +173,7 @@ public: /** Equivalent to arm_gemm::GemmInterleaved's strategy::transforms::PrepareB() but using Compute Library types. */ -template <typename To, bool use_dot = false> +template <typename strategy> class NEGEMMInterleavedPrepareBWrapperKernelTemplate : public NEGEMMInterleavedPrepareBWrapperKernel { public: @@ -105,13 +185,55 @@ public: * @param[in] ci CPU information * @param[in] params M, N, K sizes. */ - void configure(const ITensor *b, ITensor *transformed_b, bool transpose_b, const CPUInfo &ci, const INEGEMMWrapperKernel::Params ¶ms); + void configure(const ITensor *b, ITensor *transformed_b, bool transpose_b, const CPUInfo &ci, const INEGEMMWrapperKernel::Params ¶ms) + { + const unsigned int multis = b->info()->tensor_shape().z(); + _Nsize = b->info()->tensor_shape().x(); + _Ksize = b->info()->tensor_shape().y(); + _b = b; + _transformed_b = transformed_b; + _transpose_b = transpose_b; + + _block_sizes = calculate_block_sizes<strategy>(ci, params.M, params.N, params.K); + + auto_init_if_empty(*transformed_b->info(), b->info()->clone()->set_tensor_shape(TensorShape{ detail::get_B_pretransposed_array_size<strategy>(_Nsize, _Ksize, _block_sizes, multis) })); + + Window window; + window.set(Window::DimX, Window::Dimension(0, ceil_to_multiple(_Nsize, _block_sizes.x_block), _block_sizes.x_block)); + window.set(Window::DimY, Window::Dimension(0, ceil_to_multiple(_Ksize, _block_sizes.k_block), _block_sizes.k_block)); + window.set(Window::DimZ, Window::Dimension(0, multis)); + + INEKernel::configure(window); + } // Inherited methods overridden: - void transform(const PrepareBWorkload &wl, const ThreadInfo &info) override; - void create_workloads(std::vector<PrepareBWorkload> &workloads) override; - void run(const Window &window, const ThreadInfo &info) override; - BlockSizes block_sizes() const override; + void transform(const PrepareBWorkload &wl, const ThreadInfo &info) override + { + strategy strat(info.cpu_info); + strat.transforms.PrepareB(reinterpret_cast<typename strategy::operand_type *>(_transformed_b->buffer() + wl._offset_transformed_b), + reinterpret_cast<typename strategy::operand_type *>(_b->buffer() + wl._offset_b), + _b->info()->strides_in_bytes().y() / sizeof(typename strategy::operand_type), + wl._x0, wl._xmax, wl._k0, wl._kmax, _transpose_b); + } + void create_workloads(std::vector<PrepareBWorkload> &workloads) override + { + detail::for_each_element_in_window<strategy, true>(window(), _b, _transformed_b, _Nsize, _Ksize, [&workloads](PrepareBWorkload && wl) + { + workloads.push_back(std::move(wl)); + }); + } + void run(const Window &window, const ThreadInfo &info) override + { + ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(window, INEKernel::window()); + detail::for_each_element_in_window<strategy, false>(window, _b, _transformed_b, _Nsize, _Ksize, [&](PrepareBWorkload && wl) + { + this->transform(wl, info); + }); + } + BlockSizes block_sizes() const override + { + return _block_sizes; + } private: const ITensor *_b |