diff options
author | ramelg01 <ramy.elgammal@arm.com> | 2022-04-08 03:52:28 +0100 |
---|---|---|
committer | Ramy Elgammal <ramy.elgammal@arm.com> | 2022-04-25 15:35:59 +0000 |
commit | c827e99fc46521f43719b0c2d1b6f05d66abf68c (patch) | |
tree | 31df1002673b2a4c4aae66608ad85b1ad6517050 /src/core/NEON/kernels/assembly | |
parent | 0a3948394e7e77344201b8732e9c20fcb5fa9a38 (diff) | |
download | ComputeLibrary-c827e99fc46521f43719b0c2d1b6f05d66abf68c.tar.gz |
Update Neon™ pooling kernel
- Reduce duplication and simplify overall structure.
- Improve multi-threaded performance by sharing more data
in lower-level caches.
Partially Resolves: COMPMID-5054
Signed-off-by: Ramy Elgammal<ramy.elgammal@arm.com>
Change-Id: I5f4dc50913401d5c1cbfc10b866fae9490cbc4d7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7404
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/assembly')
-rw-r--r-- | src/core/NEON/kernels/assembly/pool_common.hpp | 22 | ||||
-rw-r--r-- | src/core/NEON/kernels/assembly/pooling.hpp | 126 |
2 files changed, 130 insertions, 18 deletions
diff --git a/src/core/NEON/kernels/assembly/pool_common.hpp b/src/core/NEON/kernels/assembly/pool_common.hpp index b6a0a0abed..599e18ac59 100644 --- a/src/core/NEON/kernels/assembly/pool_common.hpp +++ b/src/core/NEON/kernels/assembly/pool_common.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,8 +23,9 @@ */ #pragma once - -#include "common.hpp" +#ifdef CYCLE_PROFILING +#include "profiler.hpp" +#endif namespace arm_conv { @@ -53,6 +54,11 @@ struct PoolingStride unsigned int rows, cols; }; +struct PaddingValues +{ + unsigned int left, top, right, bottom; +}; + class IPoolingCommon { public: @@ -60,6 +66,7 @@ public: // Determine the amount of working space required. virtual size_t get_working_size(unsigned int num_threads) const = 0; + virtual size_t get_working_size(unsigned int num_threads, unsigned int n_channels) const = 0; // Execute pooling over the specified area of memory. virtual void execute( @@ -103,14 +110,5 @@ public: unsigned int num_threads) const = 0; }; -struct Nothing -{ -}; - -template <typename TInput, typename TOutput, class OutputStage = Nothing> -class PoolingCommon : public IPoolingCommon -{ -}; - } // namespace pooling } // namespace arm_conv diff --git a/src/core/NEON/kernels/assembly/pooling.hpp b/src/core/NEON/kernels/assembly/pooling.hpp index 2325bd08ca..1b47853eaf 100644 --- a/src/core/NEON/kernels/assembly/pooling.hpp +++ b/src/core/NEON/kernels/assembly/pooling.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -27,8 +27,6 @@ #include "arm_gemm_local.hpp" #include "pool_common.hpp" -#include <memory> - namespace arm_conv { namespace pooling @@ -89,6 +87,10 @@ struct PoolingArgs } }; +struct Nothing +{ +}; + struct Requantize32 { int32_t input_offset = 0; @@ -106,12 +108,124 @@ struct Requantize32 } }; -template <typename TInput, typename TOutput, class OutputStage = Nothing> -using UniquePoolingCommon = std::unique_ptr<PoolingCommon<TInput, TOutput, OutputStage>>; +template <typename TInput, typename TOutput> +class PoolingCommon : public IPoolingCommon +{ +protected: + const PoolingArgs m_args; + +public: + PoolingCommon(const PoolingArgs &args) + : m_args(args) + { + } + PoolingCommon(PoolingCommon &) = delete; + PoolingCommon &operator=(PoolingCommon &) = delete; + + size_t get_working_size(unsigned int, unsigned int) const override = 0; + size_t get_working_size(unsigned int n_threads) const override + { + return this->get_working_size(n_threads, m_args.n_channels); + } + + // Execute pooling over the specified area of memory. + void execute( + const void *const input, + void *const output, + void *working_space, + unsigned int thread_id, + unsigned int num_threads) const override + { + this->execute( + input, + m_args.n_channels, + m_args.n_channels * m_args.input_cols, + m_args.n_channels * m_args.input_cols * m_args.input_rows, + output, + m_args.n_channels, + m_args.n_channels * m_args.output_cols, + m_args.n_channels * m_args.output_cols * m_args.output_rows, + working_space, + thread_id, num_threads); + } + + void execute( + const void *const input, + size_t ld_input_col, + size_t ld_input_row, + size_t ld_input_batch, + void *const output, + size_t ld_output_col, + size_t ld_output_row, + size_t ld_output_batch, + void *working_space, + unsigned int thread_id, + unsigned int num_threads) const override + { + this->execute( + m_args.n_batches, m_args.input_rows, m_args.input_cols, m_args.n_channels, + input, ld_input_col, ld_input_row, ld_input_batch, + m_args.padding, m_args.output_rows, m_args.output_cols, + output, ld_output_col, ld_output_row, ld_output_batch, + working_space, thread_id, num_threads); + } + + void execute( + unsigned int batches, + unsigned int height, + unsigned int width, + unsigned int channels, + const void *const input, + size_t ld_input_col, + size_t ld_input_row, + size_t ld_input_batch, + const PaddingValues &padding, + unsigned int output_height, + unsigned int output_width, + void *const output, + size_t ld_output_col, + size_t ld_output_row, + size_t ld_output_batch, + void *working_space, + unsigned int thread_id, + unsigned int num_threads) const override + { + this->execute_internal( + batches, height, width, channels, padding, + input, ld_input_col, ld_input_row, ld_input_batch, + output_height, output_width, + output, ld_output_col, ld_output_row, ld_output_batch, + working_space, thread_id, num_threads); + } + +protected: + virtual void execute_internal( + unsigned int batches, + unsigned int height, + unsigned int width, + unsigned int channels, + const PaddingValues &, + const void *const input, + size_t ld_input_col, + size_t ld_input_row, + size_t ld_input_batch, + unsigned int output_height, + unsigned int output_width, + void *const output, + size_t ld_output_col, + size_t ld_output_row, + size_t ld_output_batch, + void *working_space, + unsigned int thread_id, + unsigned int num_threads) const = 0; +}; + +template <typename TInput, typename TOutput> +using UniquePoolingCommon = std::unique_ptr<PoolingCommon<TInput, TOutput>>; // Get a pooling engine template <typename TInput, typename TOutput = TInput, class OutputStage = Nothing> -UniquePoolingCommon<TInput, TOutput, OutputStage> pooling(const PoolingArgs &, const OutputStage & = {}); +UniquePoolingCommon<TInput, TOutput> pooling(const PoolingArgs &, const OutputStage & = {}); } // namespace pooling } // namespace arm_conv |