!15962 modify split cpu kernel

From: @shaoxiangdong
Reviewed-by: @wuxuejian,@c_34
Signed-off-by: @c_34

mindspore/ccsrc/backend/kernel_compiler/cpu/split_cpu_kernel.cc

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2020-2021 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -13,8 +13,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+#include <algorithm>
 #include "backend/kernel_compiler/cpu/split_cpu_kernel.h"
 #include "runtime/device/cpu/cpu_device_address.h"
+#include "common/thread_pool.h"
 
 namespace mindspore {
 namespace kernel {
@@ -29,20 +31,19 @@ void SplitCPUKernel<T>::InitKernel(const CNodePtr &kernel_node) {
 
 template <typename T>
 void SplitCPUKernel<T>::Reshape() {
-  input_size_ = 1;
-  dims_current_after_axis_ = 1;
-  dims_after_axis_ = 1;
-  axis_step_ = input_shape_[axis_] / output_num_;
+  param_ = new SplitParameter();
+  param_->num_split_ = output_num_;
+  param_->split_dim_ = axis_ >= 0 ? axis_ : input_shape_.size() + axis_;
 
-  for (int i = 0; i < SizeToInt(input_shape_.size()); i++) {
-    input_size_ *= input_shape_[i];
-    if (i > axis_) {
-      dims_current_after_axis_ *= input_shape_[i];
-      dims_after_axis_ *= input_shape_[i];
-    }
-    if (i == axis_) {
-      dims_current_after_axis_ *= input_shape_[i];
-    }
+  param_->strides_[input_shape_.size() - 1] = 1;
+  for (int i = input_shape_.size() - 2; i >= 0; i--) {
+    param_->strides_[i] = param_->strides_[i + 1] * input_shape_[i + 1];
+  }
+
+  param_->split_sizes_ = new int[sizeof(int) * param_->num_split_];
+  int split_size = input_shape_[param_->split_dim_] / output_num_;
+  for (int i = 0; i < param_->num_split_; i++) {
+    param_->split_sizes_[i] = split_size;
   }
 }
 
@@ -61,17 +62,42 @@ bool SplitCPUKernel<T>::Launch(const std::vector<kernel::AddressPtr> &inputs,
 }
 
 template <typename T>
-void SplitCPUKernel<T>::LaunchSplit(const T *input, T **output, size_t size) {
+void SplitCPUKernel<T>::LaunchSplit(T *input, T **output, size_t size) {
+  (void)std::transform(input_shape_.begin(), input_shape_.end(), std::back_inserter(input_shape_int_),
+                       [](const int &value) { return static_cast<int>(value); });
+  auto max_thread_num = common::ThreadPool::GetInstance().GetSyncRunThreadNum();
+  const float block_size = 128.0;
+  size_t thread_num = size < block_size * max_thread_num ? std::ceil(size / block_size) : max_thread_num;
+
+  param_->split_count_ = size / (input_shape_[param_->split_dim_] * param_->strides_[param_->split_dim_]);
+  int num_unit = param_->split_count_ * param_->num_split_;
+  int thread_n_stride;
+  if (thread_num != 0) {
+    thread_n_stride = UP_DIV(num_unit, thread_num);
+  }
+
   auto task = [&](size_t start, size_t end) {
-    for (size_t i = start; i < end; i++) {
-      int num = i % dims_current_after_axis_ / dims_after_axis_;
-      int block = num / axis_step_;
-      int block_pos = i / dims_current_after_axis_ * axis_step_ * dims_after_axis_ +
-                      num % axis_step_ * dims_after_axis_ + i % dims_after_axis_;
-      output[block][block_pos] = input[i];
-    }
+    int task_id = start / (size / thread_num);
+    int thread_offset = task_id * thread_n_stride;
+    int num_unit_thread = MSMIN(thread_n_stride, num_unit - task_id * thread_n_stride);
+    DoSplit(input, reinterpret_cast<void **>(output), &input_shape_int_[0], thread_offset, num_unit_thread, param_,
+            sizeof(T));
   };
   CPUKernelUtils::ParallelFor(task, size);
+
+  return;
+}
+
+template <typename T>
+void SplitCPUKernel<T>::FreeTmpBuff() {
+  if (param_->split_sizes_ != nullptr) {
+    delete[] param_->split_sizes_;
+    param_->split_sizes_ = nullptr;
+  }
+  if (param_ != nullptr) {
+    delete param_;
+    param_ = nullptr;
+  }
   return;
 }
 
@@ -86,6 +112,7 @@ void SplitCPUKernel<T>::LaunchKernel(const std::vector<AddressPtr> &inputs,
   }
   size_t size = static_cast<size_t>(inputs[0]->size / sizeof(T));
   LaunchSplit(input, output, size);
+  FreeTmpBuff();
   return;
 }
 

mindspore/ccsrc/backend/kernel_compiler/cpu/split_cpu_kernel.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2020-2021 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -20,6 +20,7 @@
 #include <thread>
 #include "backend/kernel_compiler/cpu/cpu_kernel.h"
 #include "backend/kernel_compiler/cpu/cpu_kernel_factory.h"
+#include "backend/kernel_compiler/cpu/nnacl/base/split_base.h"
 
 namespace mindspore {
 namespace kernel {
@@ -42,7 +43,8 @@ class SplitCPUKernel : public CPUKernel {
  private:
   void CheckParam(const CNodePtr &kernel_node);
   void Reshape();
-  void LaunchSplit(const T *input, T **output, size_t size);
+  void LaunchSplit(T *input, T **output, size_t size);
+  void FreeTmpBuff();
   int64_t axis_;
   int64_t output_num_;
   int64_t axis_step_;
@@ -53,7 +55,11 @@ class SplitCPUKernel : public CPUKernel {
 
   std::vector<std::vector<size_t>> output_shape_list_;
   std::vector<size_t> input_shape_;
+  std::vector<int> input_shape_int_;
   TypeId dtype_{kTypeUnknown};
+
+ protected:
+  SplitParameter *param_ = nullptr;
 };
 
 MS_REG_CPU_KERNEL_T(