!2762 Adjust the thread num to compute optimizer of cpu kernel

Merge pull request !2762 from YuJianfeng/master

mindspore/ccsrc/kernel/common_utils.cc

@@ -876,12 +876,13 @@ bool IsWeightBoundary(const AnfNodePtr &node) {
   return false;
 }
 
-void MultiThreadCompute(const MultiThreadComputeFunc &func, MultiThreadComputeParams *params, size_t thread_num,
+void MultiThreadCompute(const MultiThreadComputeFunc &func, MultiThreadComputeParams *params,
                         size_t total_compute_size) {
+  const size_t kThreadNum = 24;
   std::vector<std::thread> threads;
-  threads.reserve(thread_num);
+  threads.reserve(kThreadNum);
   size_t start = 0;
-  size_t once_compute_size = (total_compute_size + thread_num - 1) / thread_num;
+  size_t once_compute_size = (total_compute_size + kThreadNum - 1) / kThreadNum;
   while (start < total_compute_size) {
     size_t end = (start + once_compute_size) > total_compute_size ? total_compute_size : (start + once_compute_size);
     threads.emplace_back(std::thread(func, params, start, end));

mindspore/ccsrc/kernel/common_utils.h

@@ -128,7 +128,7 @@ void GetValidKernelNodes(const FuncGraphPtr &func_graph, std::vector<AnfNodePtr>
 bool GetInputTensorValue(const AnfNodePtr &anf_node, size_t input_idx, nlohmann::json *const node_json);
 void GetGraphRealOutput(const FuncGraphPtr &func_graph, std::vector<std::pair<AnfNodePtr, size_t>> *node_list);
 bool IsWeightBoundary(const AnfNodePtr &node);
-void MultiThreadCompute(const MultiThreadComputeFunc &func, MultiThreadComputeParams *params, size_t thread_num,
+void MultiThreadCompute(const MultiThreadComputeFunc &func, MultiThreadComputeParams *params,
                         size_t total_compute_size);
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/cpu/sparse_apply_adam_cpu_kernel.cc

@@ -155,15 +155,14 @@ bool SparseApplyAdamCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inp
   input_params.v_ = v;
   input_params.beta1_ = beta1;
   input_params.beta2_ = beta2;
-  const size_t kThreadNum = 16;
-  MultiThreadCompute(ComputeMomentum, &input_params, kThreadNum, total_dim_size);
+  MultiThreadCompute(ComputeMomentum, &input_params, total_dim_size);
 
   input_params.m_t_ = m_t;
   input_params.use_nesterov_ = use_nesterov_;
   input_params.sparse_grad_ = unique_sparse_grad;
   input_params.var_first_dim_size_ = var_first_dim_size_;
   input_params.var_outer_dim_size_ = var_outer_dim_size_;
-  MultiThreadCompute(ComputeAdam, &input_params, kThreadNum, unique_sparse_grad.indices_size_);
+  MultiThreadCompute(ComputeAdam, &input_params, unique_sparse_grad.indices_size_);
 
   if (use_nesterov_) {
     input_params.m_ = input_params.m_t_;
@@ -171,7 +170,7 @@ bool SparseApplyAdamCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inp
   input_params.var_ = var;
   input_params.lr_ = lr;
   input_params.epsilon_ = epsilon;
-  MultiThreadCompute(ComputeWeight, &input_params, kThreadNum, total_dim_size);
+  MultiThreadCompute(ComputeWeight, &input_params, total_dim_size);
   return true;
 }
 }  // namespace kernel

mindspore/ccsrc/kernel/cpu/sparse_apply_ftrl_cpu_kernel.cc

@@ -145,8 +145,7 @@ bool SparseApplyFtrlCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inp
   input_params.sparse_grad_ = unique_sparse_grad;
   input_params.var_first_dim_size_ = var_first_dim_size_;
   input_params.var_outer_dim_size_ = var_outer_dim_size_;
-  const size_t kThreadNum = 16;
-  MultiThreadCompute(ComputeFtrl, &input_params, kThreadNum, unique_sparse_grad.indices_size_);
+  MultiThreadCompute(ComputeFtrl, &input_params, unique_sparse_grad.indices_size_);
   return true;
 }
 }  // namespace kernel

mindspore/ccsrc/kernel/cpu/sparse_apply_lazy_adam_cpu_kernel.cc

@@ -139,8 +139,7 @@ bool SparseApplyLazyAdamCPUKernel::Launch(const std::vector<kernel::AddressPtr>
   input_params.sparse_grad_ = unique_sparse_grad;
   input_params.var_first_dim_size_ = var_first_dim_size_;
   input_params.var_outer_dim_size_ = var_outer_dim_size_;
-  const size_t kThreadNum = 16;
-  MultiThreadCompute(ComputeLazyAdam, &input_params, kThreadNum, unique_sparse_grad.indices_size_);
+  MultiThreadCompute(ComputeLazyAdam, &input_params, unique_sparse_grad.indices_size_);
   return true;
 }
 }  // namespace kernel

mindspore/ccsrc/kernel/cpu/sparse_apply_proximal_adagrad_cpu_kernel.cc

@@ -132,8 +132,7 @@ bool SparseApplyProximalAdagradCPUKernel::Launch(const std::vector<kernel::Addre
   input_params.sparse_grad_ = unique_sparse_grad;
   input_params.var_first_dim_size_ = var_first_dim_size_;
   input_params.var_outer_dim_size_ = var_outer_dim_size_;
-  const size_t kThreadNum = 16;
-  MultiThreadCompute(ComputeProximalAdagrad, &input_params, kThreadNum, unique_sparse_grad.indices_size_);
+  MultiThreadCompute(ComputeProximalAdagrad, &input_params, unique_sparse_grad.indices_size_);
   return true;
 }
 }  // namespace kernel