From 932b561159cd6a8c9230bbd0343790c85755846e Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 3 May 2018 13:44:35 +0100 Subject: COMPMID-959: Perform pretranspose if allowed on NEON assembly Change-Id: I281699ce7270aec1317c47b5a13799954cf6c9e8 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/130010 Tested-by: Jenkins Reviewed-by: Pablo Tello Reviewed-by: Anthony Barbier --- arm_compute/runtime/NEON/AssemblyHelper.h | 48 +++++++++++++++------- arm_compute/runtime/NEON/functions/NEGEMM.h | 1 + .../NEON/functions/NEGEMMConvolutionLayer.h | 1 + .../NEGEMMLowpAssemblyMatrixMultiplyCore.h | 1 + .../NEON/functions/NEGEMMLowpMatrixMultiplyCore.h | 1 + src/runtime/NEON/functions/NEGEMM.cpp | 5 ++- .../NEON/functions/NEGEMMConvolutionLayer.cpp | 10 ++++- .../NEGEMMLowpAssemblyMatrixMultiplyCore.cpp | 6 +-- .../functions/NEGEMMLowpMatrixMultiplyCore.cpp | 8 ++-- .../NEON/functions/NEWinogradConvolutionLayer.cpp | 2 +- 10 files changed, 57 insertions(+), 26 deletions(-) diff --git a/arm_compute/runtime/NEON/AssemblyHelper.h b/arm_compute/runtime/NEON/AssemblyHelper.h index 2b4f35f2e1..ee09ef531e 100644 --- a/arm_compute/runtime/NEON/AssemblyHelper.h +++ b/arm_compute/runtime/NEON/AssemblyHelper.h @@ -51,7 +51,7 @@ public: using TypeResult = TypeOutput; /** Default constructor. */ AssemblyKernelGlue() - : _gemm_kernel_asm(nullptr), _optimised_kernel(nullptr), _a(nullptr), _b(nullptr), _d(nullptr) + : _gemm_kernel_asm(nullptr), _optimised_kernel(nullptr), _a(nullptr), _b(nullptr), _d(nullptr), _pretranspose(nullptr) { } /** Assembly Gemm */ @@ -72,6 +72,8 @@ public: const ITensor *_b; /** Output */ ITensor *_d; + /** Pre-transpose tensor */ + ITensor *_pretranspose; /** Configures the arrays pointers and strides in the assembly kernel and executes the assembly kernel. * The call to set_arrays is needed to deal with the input sizes containing batches (dims > 2) @@ -94,6 +96,12 @@ public: auto out_ptr = reinterpret_cast(_d->buffer()); _gemm_kernel_asm->set_arrays(in0_ptr, lda, batch_stride_a, multi_stride_a, in1_ptr, ldb, multi_stride_b, out_ptr, ldd, batch_stride_d, multi_stride_d); + if(_gemm_kernel_asm->B_pretranspose_required()) + { + ARM_COMPUTE_ERROR_ON(_pretranspose == nullptr || _pretranspose->buffer() == nullptr); + _gemm_kernel_asm->pretranspose_B_array(reinterpret_cast(_pretranspose->buffer()), in1_ptr, ldb, multi_stride_b); + } + NEScheduler::get().schedule(_optimised_kernel.get(), Window::DimX); } }; @@ -113,8 +121,9 @@ using AssemblyKernelGlueS8S32 = AssemblyKernelGlue; * @param[in] alignment Workspace memory alignment. * @param[in] num_threads Number of workspace threads. */ -inline void allocate_workspace(size_t workspace_size, Tensor &workspace, MemoryGroup &memory_group, size_t alignment, unsigned int num_threads) +inline void allocate_workspace(size_t workspace_size, Tensor &workspace, MemoryGroup *memory_group, size_t alignment, unsigned int num_threads) { + ARM_COMPUTE_UNUSED(memory_group); ARM_COMPUTE_ERROR_ON_MSG(workspace_size == 0, "size cannot be 0"); workspace.allocator()->init(TensorInfo(TensorShape{ (workspace_size + alignment - 1) * num_threads }, 1, DataType::S8)); workspace.allocator()->allocate(); @@ -122,20 +131,22 @@ inline void allocate_workspace(size_t workspace_size, Tensor &workspace, MemoryG /** Create a wrapper kernel. * - * @param[in] a Input tensor A. - * @param[in] b Input tensor B. - * @param[out] d Output tensor. - * @param[in] alpha Alpha value. - * @param[in] beta Beta value. - * @param[out] workspace Workspace tensor - * @param[in] memory_group Tensor memory group. - * @param[out] asm_glue Assembly glue kernel. + * @param[in] a Input tensor A. + * @param[in] b Input tensor B. + * @param[out] d Output tensor. + * @param[in] alpha Alpha value. + * @param[in] beta Beta value. + * @param[in] pretranspose_hint Pre-transpose hint in case matrix b should be pre-transposed + * @param[out] workspace Workspace tensor + * @param[out] B_pretranspose Tensor to hold the pre-transposed B + * @param[in] memory_group Tensor memory group. + * @param[out] asm_glue Assembly glue kernel. * * @return the wrapper kernel. */ template -inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, - Tensor &workspace, MemoryGroup &memory_group, T &asm_glue) +inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint, + Tensor &workspace, Tensor &B_pretranspose, MemoryGroup &memory_group, T &asm_glue) { const CPUInfo &ci = NEScheduler::get().cpu_info(); const int M = d->info()->tensor_shape().y(); @@ -147,7 +158,7 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d // unique_ptr to a Gemm object std::unique_ptr - asm_gemm(arm_gemm::gemm(ci, M, N, K, batches, multis, false, false, alpha, beta, num_threads, false)); + asm_gemm(arm_gemm::gemm(ci, M, N, K, batches, multis, false, false, alpha, beta, num_threads, pretranspose_hint)); // arm_compute wrapper for the Gemm object (see above) std::unique_ptr> acl_gemm_wrapper = support::cpp14::make_unique>(); @@ -159,7 +170,7 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d { // Allocate workspace const unsigned int alignment = 4096; - allocate_workspace(workspace_size, workspace, memory_group, alignment, num_threads); + allocate_workspace(workspace_size, workspace, &memory_group, alignment, num_threads); ARM_COMPUTE_ERROR_ON_NULLPTR(workspace.buffer()); asm_gemm->set_working_space(reinterpret_cast(workspace.buffer())); } @@ -175,6 +186,15 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d } } + // Check for pre-transposed support + if(asm_gemm->B_pretranspose_required()) + { + const size_t B_pretranspose_size = asm_gemm->get_B_pretransposed_array_size(); + allocate_workspace(B_pretranspose_size, B_pretranspose, nullptr, 1, 1); + ARM_COMPUTE_ERROR_ON_NULLPTR(B_pretranspose.buffer()); + asm_glue._pretranspose = &B_pretranspose; + } + asm_glue._gemm_kernel_asm = std::move(asm_gemm); asm_glue._optimised_kernel = std::move(acl_gemm_wrapper); // We need to setup the ptrs in the run() method diff --git a/arm_compute/runtime/NEON/functions/NEGEMM.h b/arm_compute/runtime/NEON/functions/NEGEMM.h index 5279995be4..e2263c2307 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMM.h +++ b/arm_compute/runtime/NEON/functions/NEGEMM.h @@ -83,6 +83,7 @@ private: Tensor _tmp_a; Tensor _tmp_b; Tensor _workspace; + Tensor _B_pretransposed; bool _run_vector_matrix_multiplication; bool _run_addition; bool _is_first_run; diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h index 24e23f133a..752693188c 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h @@ -185,6 +185,7 @@ private: Tensor _gemm_output; Tensor _tmp_output; Tensor _workspace; + Tensor _B_pretransposed; bool _append_bias; bool _is_fully_connected_convolution; diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h index f09c94e726..11ca1bc313 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h @@ -67,6 +67,7 @@ private: Tensor _tmp_a; Tensor _tmp_b; Tensor _workspace; + Tensor _B_pretransposed; }; } #endif /*__ARM_COMPUTE_NEGEMMLOWPASSEMBLYMATRIXMULTIPLYCORE_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h index 3c9fb0ea5f..adcddb8263 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h @@ -102,6 +102,7 @@ private: Tensor _tmp_a; Tensor _tmp_b; Tensor _workspace; + Tensor _B_pretranspose; int32_t _a_offset; int32_t _b_offset; bool _run_vector_matrix_multiplication; diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp index e0859be93e..9168ed4327 100644 --- a/src/runtime/NEON/functions/NEGEMM.cpp +++ b/src/runtime/NEON/functions/NEGEMM.cpp @@ -39,7 +39,7 @@ namespace arm_compute { NEGEMM::NEGEMM(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(), _ma_kernel(), _tmp_a(), _tmp_b(), _workspace(), + : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(), _ma_kernel(), _tmp_a(), _tmp_b(), _workspace(), _B_pretransposed(), _run_vector_matrix_multiplication(false), _run_addition(false), _is_first_run(true), _reshape_b_only_on_first_run(false) { } @@ -66,7 +66,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); _run_vector_matrix_multiplication = a->info()->dimension(1) < 2; - const bool run_optimised = a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f) && setup_assembly_kernel(a, b, d, alpha, beta, _workspace, _memory_group, _asm_glue); + const bool run_optimised = a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f) + && setup_assembly_kernel(a, b, d, alpha, beta, _reshape_b_only_on_first_run, _workspace, _B_pretransposed, _memory_group, _asm_glue); // Check if the first input tensor is a vector. // If so, all the kernels for reshaping the tensors can be skipped diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index 3c48d691ed..1ffeaf227d 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -212,7 +212,8 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr &memory_manager) : _asm_glue(), _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _output_col2im_kernel(), _activationlayer_function(), _original_weights(nullptr), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), - _workspace(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false), _is_activationlayer_enabled(false) + _workspace(), _B_pretransposed(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false), + _is_activationlayer_enabled(false) { } @@ -365,7 +366,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig // Configure matrix multiply if(run_optimised) { - if(!setup_assembly_kernel(&_input_im2col_reshaped, weights, &_gemm_output, 1.f, 0.f, _workspace, _memory_group, _asm_glue)) + if(!setup_assembly_kernel(&_input_im2col_reshaped, weights, &_gemm_output, 1.f, 0.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue)) { ARM_COMPUTE_ERROR("setup_assembly_kernel failed."); } @@ -559,6 +560,11 @@ void NEGEMMConvolutionLayer::run() if(_asm_glue._optimised_kernel != nullptr) { _asm_glue.run(); + // Release weights in case buffer is pretransposed + if(_B_pretransposed.buffer() != nullptr && _weights_reshaped.is_used()) + { + _weights_reshaped.allocator()->free(); + } } else { diff --git a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp index 27dd6c51d7..bd81bf202f 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp @@ -39,7 +39,7 @@ using namespace arm_compute; NEGEMMLowpAssemblyMatrixMultiplyCore::NEGEMMLowpAssemblyMatrixMultiplyCore(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _asm_glue_unsigned(), _asm_glue_signed(), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _tmp_a(), _tmp_b(), - _workspace() + _workspace(), _B_pretransposed() { } @@ -58,13 +58,13 @@ void NEGEMMLowpAssemblyMatrixMultiplyCore::configure(const ITensor *a, const ITe { case DataType::S8: { - run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed); + run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue_signed); break; } case DataType::QASYMM8: case DataType::U8: { - run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned); + run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue_unsigned); break; } default: diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index cbec73fc31..30dd289326 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -42,8 +42,8 @@ using namespace arm_compute::misc::shape_calculator; NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _asm_glue_unsigned(), _asm_glue_signed(), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), - _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _workspace(), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false), - _dot_product_path(false), _is_first_run(true), _reshape_b_only_on_first_run(false) + _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _workspace(), _B_pretranspose(), _a_offset(0), _b_offset(0), + _run_vector_matrix_multiplication(false), _dot_product_path(false), _is_first_run(true), _reshape_b_only_on_first_run(false) { } @@ -62,13 +62,13 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, { case DataType::S8: { - _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed); + _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, true, _workspace, _B_pretranspose, _memory_group, _asm_glue_signed); break; } case DataType::QASYMM8: case DataType::U8: { - _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned); + _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, true, _workspace, _B_pretranspose, _memory_group, _asm_glue_unsigned); break; } default: diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp index a1256ac8cb..f4640fb0b6 100644 --- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp @@ -227,7 +227,7 @@ void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor * if(workspace_size > 0) { const unsigned int alignment = 4096; - allocate_workspace(workspace_size, _workspace, _memory_group, alignment, 1); + allocate_workspace(workspace_size, _workspace, &_memory_group, alignment, 1); _arm_gemm->set_working_space(reinterpret_cast(_workspace.buffer())); } -- cgit v1.2.1