!2045 optimize cpu ftrl

Merge pull request !2045 from kisnwang/optimize-cpu-ftrl

mindspore/ccsrc/kernel/common_utils.cc

@@ -18,6 +18,7 @@
 #include <unordered_map>
 #include <map>
 #include <iostream>
+#include <utility>
 #include <fstream>
 #include "nlohmann/json.hpp"
 #include "session/anf_runtime_algorithm.h"
@@ -583,5 +584,55 @@ void DeduplicateIndexedSlices(const SparseGradient &origin_sparse_grad, SparseGr
   }
   unique_grad->indices_size_ = unique_indices_size;
 }
+
+void ReduceSparseGradient(const SparseGradient &origin_sparse_grad, SparseGradient *unique_grad, size_t first_dim,
+                          size_t outer_dim) {
+  MS_EXCEPTION_IF_NULL(origin_sparse_grad.value_);
+  MS_EXCEPTION_IF_NULL(origin_sparse_grad.indices_);
+  MS_EXCEPTION_IF_NULL(unique_grad);
+  MS_EXCEPTION_IF_NULL(unique_grad->value_);
+  MS_EXCEPTION_IF_NULL(unique_grad->indices_);
+  size_t unique_indices_size = 0;
+  std::vector<std::pair<int, size_t>> sorted_indices;
+  sorted_indices.reserve(origin_sparse_grad.indices_size_);
+  for (size_t i = 0; i < origin_sparse_grad.indices_size_; ++i) {
+    int index = origin_sparse_grad.indices_[i];
+    if (index < 0 || IntToSize(index) >= first_dim) {
+      continue;
+    }
+    sorted_indices.emplace_back(std::pair<int, size_t>(index, i * outer_dim));
+  }
+  std::sort(
+    sorted_indices.begin(), sorted_indices.end(),
+    [](const std::pair<int, size_t> &left, const std::pair<int, size_t> &right) { return left.first < right.first; });
+
+  int last_index = 0;
+  size_t indices_size = sorted_indices.size();
+  size_t start_index = 0;
+  size_t end_index = outer_dim;
+  size_t dst_len = indices_size * outer_dim;
+  for (size_t i = 0; i < indices_size; ++i) {
+    int index = sorted_indices[i].first;
+    if (i == 0 || last_index != index) {
+      if (i > 0 && last_index != index) {
+        unique_indices_size++;
+        start_index += outer_dim;
+        end_index += outer_dim;
+      }
+      unique_grad->indices_[unique_indices_size] = index;
+      auto ret_code = memcpy_s(unique_grad->value_ + start_index, dst_len - start_index,
+                               origin_sparse_grad.value_ + sorted_indices[i].second, outer_dim);
+      if (ret_code != EOK) {
+        MS_LOG(EXCEPTION) << "Failed to copy data!";
+      }
+    } else {
+      for (size_t j = start_index, k = sorted_indices[i].second; j < end_index; ++j, ++k) {
+        unique_grad->value_[j] += origin_sparse_grad.value_[k];
+      }
+    }
+    last_index = index;
+  }
+  unique_grad->indices_size_ = unique_indices_size;
+}
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/common_utils.h

@@ -92,6 +92,8 @@ bool IsSameShape(const std::vector<size_t> &shape_a, const std::vector<size_t> &
 int Sign(float x);
 void DeduplicateIndexedSlices(const SparseGradient &origin_sparse_grad, SparseGradient *unique_grad, size_t first_dim,
                               size_t outer_dim);
+void ReduceSparseGradient(const SparseGradient &origin_sparse_grad, SparseGradient *unique_grad, size_t first_dim,
+                          size_t outer_dim);
 }  // namespace kernel
 }  // namespace mindspore
 

mindspore/ccsrc/kernel/cpu/cpu_kernel.cc

@@ -37,8 +37,8 @@ void CPUKernel::InitInputOutputSize(const CNodePtr &kernel_node) {
 }
 
 void CPUKernel::Init(const CNodePtr &kernel_node) {
-  InitInputOutputSize(kernel_node);
   InitKernel(kernel_node);
+  InitInputOutputSize(kernel_node);
 }
 
 void CPUKernelUtils::ExpandDimsTo4(std::vector<size_t> *shape) {

mindspore/ccsrc/kernel/cpu/sparse_apply_ftrl_cpu_kernel.cc

@@ -23,6 +23,13 @@ namespace {
 constexpr size_t kSparseApplyFtrlInputSize = 5;
 }  // namespace
 
+void SparseApplyFtrlCPUKernel::InitInputOutputSize(const CNodePtr &kernel_node) {
+  CPUKernel::InitInputOutputSize(kernel_node);
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  workspace_size_list_.emplace_back(indices_size_ * var_outer_dim_size_ * sizeof(float));
+  workspace_size_list_.emplace_back(indices_size_ * sizeof(int));
+}
+
 void SparseApplyFtrlCPUKernel::InitKernel(const CNodePtr &kernel_node) {
   MS_EXCEPTION_IF_NULL(kernel_node);
   std::vector<size_t> var_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
@@ -72,7 +79,7 @@ void SparseApplyFtrlCPUKernel::InitKernel(const CNodePtr &kernel_node) {
 }
 
 bool SparseApplyFtrlCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inputs,
-                                      const std::vector<kernel::AddressPtr> & /*workspace*/,
+                                      const std::vector<kernel::AddressPtr> &workspace,
                                       const std::vector<kernel::AddressPtr> & /*outputs*/) {
   if (inputs.size() < kSparseApplyFtrlInputSize) {
     MS_LOG(EXCEPTION) << "error input output size!";
@@ -83,14 +90,11 @@ bool SparseApplyFtrlCPUKernel::Launch(const std::vector<kernel::AddressPtr> &inp
   auto linear = reinterpret_cast<float *>(inputs[2]->addr);
   auto grad = reinterpret_cast<float *>(inputs[3]->addr);
   auto indices = reinterpret_cast<int *>(inputs[4]->addr);
-
-  std::vector<float> new_grad;
-  new_grad.reserve(indices_size_ * var_outer_dim_size_);
-  std::vector<int> new_indices;
-  new_indices.reserve(indices_size_);
-  SparseGradient unique_sparse_grad({new_grad.data(), new_indices.data(), indices_size_});
-  DeduplicateIndexedSlices(SparseGradient({grad, indices, indices_size_}), &unique_sparse_grad, var_first_dim_size_,
-                           var_outer_dim_size_);
+  auto new_grad = reinterpret_cast<float *>(workspace[0]->addr);
+  auto new_indices = reinterpret_cast<int *>(workspace[1]->addr);
+  SparseGradient unique_sparse_grad({new_grad, new_indices, indices_size_});
+  ReduceSparseGradient(SparseGradient({grad, indices, indices_size_}), &unique_sparse_grad, var_first_dim_size_,
+                       var_outer_dim_size_);
 
   for (size_t i = 0; i < unique_sparse_grad.indices_size_; ++i) {
     int index = unique_sparse_grad.indices_[i];

mindspore/ccsrc/kernel/cpu/sparse_apply_ftrl_cpu_kernel.h

@@ -28,7 +28,7 @@ class SparseApplyFtrlCPUKernel : public CPUKernel {
   ~SparseApplyFtrlCPUKernel() override = default;
 
   void InitKernel(const CNodePtr &kernel_node) override;
-
+  void InitInputOutputSize(const CNodePtr &kernel_node) override;
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
               const std::vector<AddressPtr> &outputs) override;