From ded5b182675e3166e947a8eb637b5b1e925816ab Mon Sep 17 00:00:00 2001 From: David Svantesson Date: Wed, 2 Aug 2023 14:23:00 +0000 Subject: thread_local _custom_scheduler Resolves ONCPUML-1331 This patch adds an option to make _custom_scheduler thread_local to support usage of multiple schedulers handled outside of ACL. It also adds num_threads() function to Scheduler which reverts to querying CPUInfo if no scheduler has been set. Change-Id: Iff706165d8d091895331a5bb3a76f6cabe048912 Signed-off-by: David Svantesson-Yeung Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10748 Comments-Addressed: Arm Jenkins Reviewed-by: SiCong Li Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../CpuDepthwiseConv2dAssemblyDispatch.cpp | 4 ++-- src/cpu/operators/CpuPool2d.cpp | 4 ++-- src/cpu/operators/CpuWinogradConv2d.cpp | 6 +++--- .../operators/internal/CpuGemmAssemblyDispatch.cpp | 24 ++++++++++------------ 4 files changed, 18 insertions(+), 20 deletions(-) (limited to 'src/cpu/operators') diff --git a/src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp b/src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp index 8d3741de96..8507c59e6b 100644 --- a/src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp +++ b/src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp @@ -60,8 +60,8 @@ void CpuDepthwiseConv2dAssemblyDispatch::configure(const ITensorInfo *src, const ConvolutionInfo &info) { ARM_COMPUTE_LOG_PARAMS(src, weights, bias, dst, info); - const CPUInfo &ci = NEScheduler::get().cpu_info(); - const unsigned int num_threads = NEScheduler::get().num_threads(); + const CPUInfo &ci = CPUInfo::get(); + const unsigned int num_threads = NEScheduler::num_threads(); _pImpl->is_prepared = false; _pImpl->are_weights_const = weights->are_values_constant(); diff --git a/src/cpu/operators/CpuPool2d.cpp b/src/cpu/operators/CpuPool2d.cpp index b72bde6978..d00efd191d 100644 --- a/src/cpu/operators/CpuPool2d.cpp +++ b/src/cpu/operators/CpuPool2d.cpp @@ -69,8 +69,8 @@ void CpuPool2d::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayer if (run_optimised) { - const CPUInfo &ci = NEScheduler::get().cpu_info(); - const unsigned int num_threads = NEScheduler::get().num_threads(); + const CPUInfo &ci = CPUInfo::get(); + const unsigned int num_threads = NEScheduler::num_threads(); auto pooling_wrapper = std::make_unique(); ARM_COMPUTE_ERROR_ON(pooling_wrapper == nullptr); diff --git a/src/cpu/operators/CpuWinogradConv2d.cpp b/src/cpu/operators/CpuWinogradConv2d.cpp index e4bcdc0b64..1fb6d33a61 100644 --- a/src/cpu/operators/CpuWinogradConv2d.cpp +++ b/src/cpu/operators/CpuWinogradConv2d.cpp @@ -103,7 +103,7 @@ bool get_winograd_kernel_implementation(const ITensorInfo Tensor4DShape in_shape{internal_get_shape(src)}; Tensor4DShape out_shape{internal_get_shape(dst)}; Tensor4DShape kernel_shape{internal_get_shape(weights)}; - uint32_t nthreads = NEScheduler::get().num_threads(); + uint32_t nthreads = NEScheduler::num_threads(); // Get configuration arguments for Winograd winograd_cfg.output_rows = 0; winograd_cfg.output_cols = 0; @@ -183,7 +183,7 @@ void CpuWinogradConv2d::configure(const ITensorInfo *src, ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, act_info, enable_fast_math); ARM_COMPUTE_UNUSED(biases); const DataType data_type = src->data_type(); - uint32_t nthreads = NEScheduler::get().num_threads(); + uint32_t nthreads = NEScheduler::num_threads(); _data_layout = src->data_layout(); const Tensor4DShape kernel_shape{internal_get_shape(weights)}; @@ -361,7 +361,7 @@ void CpuWinogradConv2d::run(ITensorPack &tensors) auto output = tensors.get_tensor(ACL_DST); Window win; - const uint32_t nthreads = NEScheduler::get().num_threads(); + const uint32_t nthreads = NEScheduler::num_threads(); // The Winograd transform implementation does fine-grain threading inside the transforms. Just pass thread_id and nthreads. win.set(Window::DimX, Window::Dimension(0, nthreads, 1)); diff --git a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp index 611bc76463..7f851aa755 100644 --- a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp +++ b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp @@ -579,9 +579,8 @@ void Fallback::prepare(ITensorPack &tensors) CpuAuxTensorHandler pretranspose(offset_int_vec(Pretranspose), _pretranspose_info, tensors, false); ARM_COMPUTE_ERROR_ON(pretranspose.get()->buffer() == nullptr); - run_parallel_pretranspose_B_array(_gemm_kernel_asm.get(), pretranspose.get(), - in1_ptr, ldb, multi_stride_b, - NEScheduler::get().num_threads()); + run_parallel_pretranspose_B_array( + _gemm_kernel_asm.get(), pretranspose.get(), in1_ptr, ldb, multi_stride_b, NEScheduler::num_threads()); b->mark_as_unused(); // Note that we don't need to mark b_to_use as unused, as if it's been assigned to pre_pretransposed_b, its memory will be auto-managed by the handler @@ -691,9 +690,8 @@ void Fallback::run(ITensorPack &tensors) } else { - run_parallel_pretranspose_B_array(_gemm_kernel_asm.get(), pretranspose.get(), - b_ptr, ldb, multi_stride_b, - NEScheduler::get().num_threads()); + run_parallel_pretranspose_B_array( + _gemm_kernel_asm.get(), pretranspose.get(), b_ptr, ldb, multi_stride_b, NEScheduler::num_threads()); } } } @@ -707,7 +705,7 @@ void Fallback::run(ITensorPack &tensors) _gemm_kernel_asm->set_working_space(reinterpret_cast(workspace.get()->buffer())); const unsigned int split_dim = scheduling_hint.split_dimension(); const unsigned int window_size = _gemm_kernel_asm->get_window_size().total_size(); - unsigned int num_threads = NEScheduler::get().num_threads(); + unsigned int num_threads = NEScheduler::num_threads(); if (window_size < num_threads) { num_threads = window_size; @@ -756,8 +754,8 @@ void create_arm_gemm(std::unique_ptr &arm_ge const AsmGemmInfo &info) { Params p = extract_parameters(a, b, d, info); - const CPUInfo &ci = NEScheduler::get().cpu_info(); - unsigned int num_threads = NEScheduler::get().num_threads(); + const CPUInfo &ci = CPUInfo::get(); + unsigned int num_threads = NEScheduler::num_threads(); arm_gemm::GemmConfig cfg; cfg.weight_format = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format); @@ -781,8 +779,8 @@ void create_arm_gemm_quant(std::unique_ptr & { ARM_COMPUTE_UNUSED(activation); Params p = extract_parameters(a, b, d, info); - const CPUInfo &ci = NEScheduler::get().cpu_info(); - const unsigned int num_threads = NEScheduler::get().num_threads(); + const CPUInfo &ci = CPUInfo::get(); + const unsigned int num_threads = NEScheduler::num_threads(); arm_gemm::GemmConfig cfg; cfg.weight_format = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format); @@ -836,8 +834,8 @@ Status CpuGemmAssemblyDispatch::has_opt_impl(arm_compute::WeightFormat &expected ARM_COMPUTE_UNUSED(c); arm_gemm::Activation act = assembly_utils::map_to_arm_gemm_activation(info.activation_info); Params p = extract_parameters(a, b, d, info); - const CPUInfo &ci = NEScheduler::get().cpu_info(); - unsigned int num_threads = NEScheduler::get().num_threads(); + const CPUInfo &ci = CPUInfo::get(); + unsigned int num_threads = NEScheduler::num_threads(); arm_gemm::GemmConfig cfg; cfg.weight_format = assembly_utils::map_to_arm_gemm_weight_format(info.weight_format); arm_gemm::WeightFormat arm_gemm_expected_wf = assembly_utils::map_to_arm_gemm_weight_format(expected_weight_format); -- cgit v1.2.1