!9428 [GraphKernel] Modify cast to accelerate amp performance when active graph kernel

From: @tronzhang
Reviewed-by: @ryanww,@gaoxiong1
Signed-off-by: @gaoxiong1

mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_cse.cc

@@ -16,6 +16,7 @@
 
 #include "backend/optimizer/graph_kernel/graph_kernel_cse.h"
 
+#include <algorithm>
 #include <memory>
 #include <string>
 #include <utility>
@@ -26,13 +27,15 @@
 namespace mindspore {
 namespace opt {
 namespace {
-bool IsCNodePrimitveEqual(const CNodePtr &main, const CNodePtr &node) {
+bool IsCNodePrimitveEqual(const CNodePtr &main, const CNodePtr &node, const std::vector<PrimitivePtr> &black_list) {
   auto main_primitive = AnfAlgo::GetCNodePrimitive(main);
   auto node_primitive = AnfAlgo::GetCNodePrimitive(node);
   if (main_primitive != nullptr && node_primitive != nullptr) {
     // Some ops such as Reshape is not real op, cse these type will not get gain. And for ops fusion, keep these op
     // alone can prevent some redundant output case (input -> reshape -> output).
-    if (main_primitive->name() != node_primitive->name() || IsPrimitiveCNode(node, prim::kPrimReshape)) {
+    if (main_primitive->name() != node_primitive->name() ||
+        std::any_of(black_list.begin(), black_list.end(),
+                    [&node](const PrimitivePtr &prim) { return IsPrimitiveCNode(node, prim); })) {
       return false;
     }
 
@@ -125,12 +128,12 @@ bool GraphKernelBackendCSE::CheckEqualCnodeInputs(const AnfNodePtr &main, const
       return false;
     }
   }
-  return IsCNodePrimitveEqual(c_main, c_node);
+  return IsCNodePrimitveEqual(c_main, c_node, black_list_);
 }
 
 bool GraphKernelCSE::Run(const FuncGraphPtr &func_graph) {
   MS_EXCEPTION_IF_NULL(func_graph);
-  auto graphkernel_backend_cse = std::make_shared<GraphKernelBackendCSE>();
+  auto graphkernel_backend_cse = std::make_shared<GraphKernelBackendCSE>(black_list_);
   return graphkernel_backend_cse->Cse(func_graph, func_graph->manager());
 }
 }  // namespace opt

mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_cse.h

@@ -13,27 +13,35 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_CSE_H_
-#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_CSE_H_
+#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_GRAPH_KERNEL_CSE_H_
+#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_GRAPH_KERNEL_CSE_H_
 
+#include <vector>
 #include "backend/optimizer/pass/common_subexpression_elimination.h"
 
 namespace mindspore {
 namespace opt {
 class GraphKernelCSE : public Pass {
  public:
-  GraphKernelCSE() : Pass("graph_kernel_cse") {}
+  explicit GraphKernelCSE(const std::vector<PrimitivePtr> &black_list = {})
+      : Pass("graph_kernel_cse"), black_list_(black_list) {}
   ~GraphKernelCSE() override = default;
   bool Run(const FuncGraphPtr &func_graph) override;
+
+ private:
+  std::vector<PrimitivePtr> black_list_;
 };
 
 class GraphKernelBackendCSE : public BackendCSE {
  public:
-  GraphKernelBackendCSE() = default;
+  explicit GraphKernelBackendCSE(const std::vector<PrimitivePtr> &black_list = {}) : black_list_(black_list) {}
   ~GraphKernelBackendCSE() override = default;
   bool CheckEqualKernelBuildInfo(const AnfNodePtr &main, const AnfNodePtr &node) const override;
   bool CheckEqualCnodeInputs(const AnfNodePtr &main, const AnfNodePtr &node) const override;
+
+ private:
+  std::vector<PrimitivePtr> black_list_;
 };
 }  // namespace opt
 }  // namespace mindspore
-#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_CSE_H_
+#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GRAPH_KERNEL_GRAPH_KERNEL_CSE_H_

mindspore/ccsrc/backend/optimizer/graph_kernel/shape_ops_splitter.cc

@@ -34,7 +34,7 @@ namespace mindspore {
 namespace opt {
 namespace {
 bool IsMultiUserShapeOps(AnfNodePtr node, const FuncGraphManagerPtr &mng) {
-  std::vector<PrimitivePtr> shape_ops = {prim::kPrimReshape};
+  std::vector<PrimitivePtr> shape_ops = {prim::kPrimReshape, prim::kPrimCast};
   auto &users = mng->node_users();
   return std::any_of(shape_ops.begin(), shape_ops.end(),
                      [&node](const PrimitivePtr &prim) { return IsPrimitiveCNode(node, prim); }) &&

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -120,7 +120,9 @@ void GPUSession::Optimize(const std::shared_ptr<KernelGraph> &kernel_graph) {
   pm->AddPass(std::make_shared<opt::AdamFusion>());
   pm->AddPass(std::make_shared<opt::ApplyMomentumWeightDecayScaleFusion>());
   pm->AddPass(std::make_shared<opt::ApplyMomentumScaleFusion>());
-  pm->AddPass(std::make_shared<opt::CastAllFusion>("cast_all"));
+  if (!(context_ptr->get_param<bool>(MS_CTX_ENABLE_GRAPH_KERNEL))) {
+    pm->AddPass(std::make_shared<opt::CastAllFusion>("cast_all"));
+  }
   pm->AddPass(std::make_shared<opt::CombineMomentumFusion>("combine_momentum"));
   pm->AddPass(std::make_shared<opt::ReplaceMomentumCastFusion>());
   pm->AddPass(std::make_shared<opt::ReplaceAddNFusion>());
@@ -165,15 +167,17 @@ void GPUSession::GraphKernelOptimize(const std::shared_ptr<KernelGraph> &kernel_
   }
   auto optimizer = std::make_shared<opt::GraphOptimizer>();
   auto pm = std::make_shared<opt::PassManager>("graph_kernel_pm");
+  std::vector<PrimitivePtr> black_list = {prim::kPrimReshape, prim::kPrimCast};
   pm->AddPass(std::make_shared<opt::GraphKernelExpander>());
   pm->AddPass(std::make_shared<opt::ShapeOpsSplitter>());
   pm->AddPass(std::make_shared<opt::BasicOpsFusion>());
   pm->AddPass(std::make_shared<opt::CompositeOpsFusion>());
-  pm->AddPass(std::make_shared<opt::GraphKernelCSE>());
+  pm->AddPass(std::make_shared<opt::GraphKernelCSE>(black_list));
   pm->AddPass(std::make_shared<opt::ArithmeticSimplify>());
-  pm->AddPass(std::make_shared<opt::GraphKernelCSE>());
+  pm->AddPass(std::make_shared<opt::GraphKernelCSE>(black_list));
   pm->AddPass(std::make_shared<opt::TensorPromotion>());
   pm->AddPass(std::make_shared<opt::GraphKernelSplitter>());
+  pm->AddPass(std::make_shared<opt::GraphKernelCSE>());
   // After Simplify and Splitter, a lot of redundant getitem/maketuple
   // will be exposed, use GetitemTuple Pass to delete them.
   pm->AddPass(std::make_shared<opt::GetitemTuple>());