Added DNNL cache management for DyGraph (#25624)

* Added DNNL cache management for DyGraph

* move FLAGS_use_mkldnn to more general CMakeLists, getu use of the flag in ClearGradients

* missing file

* Fixes after review

* Bringing back original idea of place for 'use_mkldnn' flag to be accessible from platform nad imperative.

* Removed duplicate and added docs

* Fixes for CI

paddle/fluid/framework/executor.cc

@@ -37,9 +37,12 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 DECLARE_bool(benchmark);
-DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
+DECLARE_bool(use_mkldnn);
 
 namespace paddle {
 namespace framework {
@@ -83,14 +86,7 @@ Executor::~Executor() {
 #ifdef PADDLE_WITH_MKLDNN
   // Clear mkl-dnn cache,
   // this is needed to have mkl-dnn unit tests working
-  if (platform::is_cpu_place(place_)) {
+  ClearMKLDNNCache(place_);
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    platform::MKLDNNDeviceContext* dev_ctx =
-        (platform::MKLDNNDeviceContext*)pool.Get(place_);
-    dev_ctx->ResetBlobMap();
-    platform::MKLDNNDeviceContext::tls().set_cur_paddle_data_layout(
-        paddle::framework::DataLayout::kNCHW);
-  }
 #endif
 }
 

paddle/fluid/framework/ir/pass.cc

@@ -19,6 +19,9 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/platform/device_context.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace framework {
@@ -57,10 +60,7 @@ Graph* Pass::Apply(Graph* graph) const {
 #ifdef PADDLE_WITH_MKLDNN
   // Clear mkl-dnn cache,
   // Passes can change params, tensors, so caching need to be discarded
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  ClearMKLDNNCache(paddle::platform::CPUPlace());
-  platform::MKLDNNDeviceContext* dev_ctx =
-      (platform::MKLDNNDeviceContext*)pool.Get(paddle::platform::CPUPlace());
-  dev_ctx->ResetBlobMap();
 #endif
   return graph;
 }

paddle/fluid/framework/naive_executor.cc

@@ -25,6 +25,9 @@
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/string/pretty_log.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace framework {
@@ -122,14 +125,7 @@ NaiveExecutor::~NaiveExecutor() {
 #ifdef PADDLE_WITH_MKLDNN
   // Clear mkl-dnn cache,
   // this is needed to have mkl-dnn unit tests working
-  if (platform::is_cpu_place(place_)) {
+  ClearMKLDNNCache(place_);
-    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    platform::MKLDNNDeviceContext *dev_ctx =
-        (platform::MKLDNNDeviceContext *)pool.Get(place_);
-    dev_ctx->ResetBlobMap();
-    platform::MKLDNNDeviceContext::tls().set_cur_paddle_data_layout(
-        paddle::framework::DataLayout::kNCHW);
-  }
 #endif
 }
 

paddle/fluid/imperative/layer.cc

@@ -28,6 +28,11 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/profiler.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+DECLARE_bool(use_mkldnn);
 
 namespace paddle {
 namespace imperative {
@@ -192,6 +197,9 @@ void VarBase::ClearGradient() {
       auto* grad_t =
           grad_var_->MutableVar()->GetMutable<framework::SelectedRows>();
       if (grad_t->mutable_value()->IsInitialized()) {
+#ifdef PADDLE_WITH_MKLDNN
+        if (FLAGS_use_mkldnn) ClearMKLDNNCache(grad_t->place());
+#endif
         grad_t->mutable_rows()->clear();
         grad_t->mutable_value()->clear();
       }
@@ -202,6 +210,9 @@ void VarBase::ClearGradient() {
         auto* dev_ctx =
             platform::DeviceContextPool::Instance().Get(grad_t->place());
         operators::math::set_constant(*dev_ctx, grad_t, 0.0);
+#ifdef PADDLE_WITH_MKLDNN
+        if (FLAGS_use_mkldnn) ClearMKLDNNCache(grad_t->place());
+#endif
       }
     }
   }

paddle/fluid/platform/flags.cc

@@ -473,3 +473,13 @@ DEFINE_double(local_exe_sub_scope_limit, 256.0,  // MBytes
               "each CUDAPlace. If you don't need to limit the memory, "
               "you should set FLAGS_local_exe_sub_scope_limit=-1. "
               "The default value is 256 MBytes.");
+
+/**
+ * MKLDNN related FLAG
+ * Name: use_mkldnn
+ * Since Version:
+ * Value Range: bool, default=false
+ * Example:
+ * Note:
+ */
+DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");

paddle/fluid/platform/mkldnn_helper.h

@@ -117,6 +117,18 @@ inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) {
   return use_mkldnn && platform::is_cpu_place(ctx.GetPlace());
 }
 
+inline void ClearMKLDNNCache(const platform::Place& place) {
+  // Clear mkl-dnn cache,
+  if (platform::is_cpu_place(place)) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::MKLDNNDeviceContext* dev_ctx =
+        (platform::MKLDNNDeviceContext*)pool.Get(place);
+    dev_ctx->ResetBlobMap();
+    platform::MKLDNNDeviceContext::tls().set_cur_paddle_data_layout(
+        paddle::framework::DataLayout::kNCHW);
+  }
+}
+
 template <typename Type>
 mkldnn::memory::data_type MKLDNNGetDataType() {
   return mkldnn::memory::data_type::undef;