[oneDNN ] conv2d fwd&bwd optimization (#27871)

paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc

@@ -211,22 +211,8 @@ class ConvMKLDNNHandlerT
        * ('any') which lets a primitive (convolution in this case) choose
        * the memory format preferred for best performance
        */
-      // TODO(jczaja): This is workaround to make grad op UT's numerical
-      // gradient computation proper as this op is called directly without
-      // fetch op following it , so numercial grad is computed (in python)
-      // using block formats which will give wrong results
-      const std::string data_format = ctx.Attr<std::string>("data_format");
-      auto chosen_memory_format =
-          is_test ? MKLDNNMemoryFormat::any
-                  : platform::data_format_to_memory_format(data_format);
-
-      // Check the format for user's special output
-      if (chosen_memory_format != MKLDNNMemoryFormat::any) {
-        if (is_conv3d) {
-          chosen_memory_format = platform::MKLDNNFormatForSize(
-              src_tz.size(), chosen_memory_format);
-        }
-      }
+      auto chosen_memory_format = MKLDNNMemoryFormat::any;
+
       auto data_type = mkldnn::memory::data_type::f32;
       if (ctx.Attr<std::string>("mkldnn_data_type") == "bfloat16" ||
           std::is_same<T_out, platform::bfloat16>::value)
@@ -351,14 +337,16 @@ class ConvMKLDNNHandlerT
 
   std::shared_ptr<mkldnn::memory> AcquireResidualMemory(
       const framework::Tensor* residual_param) {
-    const T* residual_data = residual_param->data<T>();
+    void* residual_data =
+        residual_param->type() == framework::DataTypeTrait<T_out>::DataType()
+            ? to_void_cast<T_out>(residual_param->data<T_out>())
+            : to_void_cast<T>(residual_param->data<T>());
     auto user_residual_md = platform::MKLDNNMemDesc(
         framework::vectorize(residual_param->dims()),
         framework::ToMKLDNNDataType(residual_param->type()),
         residual_param->format());
 
-    return this->AcquireMemoryFromPrimitive(user_residual_md,
-                                            to_void_cast<T>(residual_data),
+    return this->AcquireMemoryFromPrimitive(user_residual_md, residual_data,
                                             "@user_residual_data_mem_p");
   }
 
@@ -973,22 +961,8 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
      * the memory format preferred for best performance
      */
 
-    // TODO(jczaja): Once GRAD NHWC is working then format 'any'
-    // should be used exclusively. But till forward pass enforce
-    // NCHW for training we need to have NCHW here as well
-    // to avoid performance degradation in relu_grad and pool2d_grad
-    std::string data_format = ctx.Attr<std::string>("data_format");
-    auto chosen_memory_format =
-        platform::data_format_to_memory_format(data_format);
-
+    auto chosen_memory_format = MKLDNNMemoryFormat::any;
     weights_format = MKLDNNMemoryFormat::any;
-    // Check the format for user's special output
-    if (chosen_memory_format != MKLDNNMemoryFormat::any) {
-      if (is_conv3d) {
-        chosen_memory_format =
-            platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format);
-      }
-    }
 
     auto src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
@@ -1055,8 +1029,11 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
       const size_t size = handler.GetDiffWeightsMemorySize();
       filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace(), size);
 
+      // For convoluition with groups write filter grad into
+      // oneDNN buffer and then we reorder it into filter_grad tensor
       auto diff_weights_memory_p =
-          handler.AcquireDiffWeightsMemoryFromWeightsPrimitive(
+          g > 1 ? handler.AcquireDiffWeightsMemoryFromWeightsPrimitive()
+                : handler.AcquireDiffWeightsMemoryFromWeightsPrimitive(
                       reinterpret_cast<void*>(filter_grad_data));
 
       auto conv_bwd_weights_p = handler.AcquireConvolutionBackwardWeights();
@@ -1072,8 +1049,43 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
       // in OneDNN groups in convolution are treated as separate dimension
       // which is not the case in paddlepaddle
       auto filter_fmt = GetMKLDNNFormat(*diff_weights_memory_p);
-      filter_grad->set_format(platform::MKLDNNFormatForSize(
-          g > 1 ? weights_tz.size() - 1 : weights_tz.size(), filter_fmt));
+
+      // For convolution with groups convert from blocked to NCHW
+      // otherwise there will be problems in next operators working on this data
+      if (g > 1) {
+        memory::data_type in_type =
+            framework::ToMKLDNNDataType(filter_grad->type());
+        // for 3d conv with groups (six dimensional data reorder to goidhw)
+        // for 2d conv with groups (five dimensional data reorder to goihw)
+        mkldnn::memory::format_tag out_format =
+            weights_tz.size() == 6 ? mkldnn::memory::format_tag::goidhw
+                                   : mkldnn::memory::format_tag::goihw;
+        const std::string key =
+            platform::CreateKey(weights_tz, filter_fmt, out_format, in_type);
+
+        platform::ReorderMKLDNNHandler handler(weights_tz, filter_grad->type(),
+                                               in_type, dev_ctx, mkldnn_engine,
+                                               key);
+        auto reorder_dst_memory_p =
+            handler.AcquireDstMemory(filter_grad, out_format, ctx.GetPlace());
+
+        auto reorder_p =
+            handler.AcquireReorder(reorder_dst_memory_p, diff_weights_memory_p);
+
+        reorder_p->execute(astream, *diff_weights_memory_p,
+                           *reorder_dst_memory_p);
+        astream.wait();
+
+        // So here we have a data in goihw , which can be interpreted as OIHW
+        // (OIDHW for conv3d)
+        // because filter_grad shape is set for OIHW (OIDHW for conv3d)
+        mkldnn::memory::format_tag target_format =
+            weights_tz.size() == 6 ? mkldnn::memory::format_tag::oidhw
+                                   : mkldnn::memory::format_tag::oihw;
+        filter_grad->set_format(target_format);
+      } else {
+        filter_grad->set_format(filter_fmt);
+      }
     }
     if (input_grad) {
       auto weights_memory_p = handler.AcquireWeightsMemoryFromDataPrimitive(

paddle/fluid/platform/mkldnn_helper.h

@@ -289,6 +289,10 @@ inline mkldnn::memory::format_tag GetMKLDNNFormat(
             strides[3] >= strides[4] && strides[4] >= strides[1]) {
           return mkldnn::memory::format_tag::Acdeb16a;
         }
+        if (strides[0] >= strides[1] && strides[1] >= strides[2] &&
+            strides[2] >= strides[3] && strides[3] >= strides[4]) {
+          return mkldnn::memory::format_tag::Abcde16a;
+        }
       } else if (inner_blks[0] == 16 && inner_idxs[0] == 1) {
         if (strides[0] >= strides[1] && strides[1] >= strides[2] &&
             strides[2] >= strides[3] && strides[3] >= strides[4]) {

paddle/fluid/platform/mkldnn_reuse.h

@@ -346,6 +346,18 @@ class MKLDNNHandler {
     return mem_p;
   }
 
+  std::shared_ptr<mkldnn::memory> AcquireMemoryFromPrimitive(
+      mkldnn::memory::desc md, const std::string& suffix) {
+    const auto local_key = key_ + suffix;
+    auto mem_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    if (mem_p == nullptr) {
+      mem_p = std::make_shared<mkldnn::memory>(md, engine_);
+      dev_ctx_.SetBlob(local_key, mem_p);
+    }
+    return mem_p;
+  }
+
   // This incarnation of AcquireMemory can call user function eg. custom reorder
   // or preprocessing routine if needed
   std::shared_ptr<mkldnn::memory> AcquireMemory(
@@ -1199,6 +1211,12 @@ class ConvMKLDNNTemplateHandler : public MKLDNNHandler {
         conv_bwd_weights_pd_->diff_weights_desc(), ptr, "@diff_weights_mem_p");
   }
 
+  std::shared_ptr<mkldnn::memory> AcquireDiffWeightsMemoryFromWeightsPrimitive(
+      void) {
+    return this->AcquireMemoryFromPrimitive(
+        conv_bwd_weights_pd_->diff_weights_desc(), "@diff_weights_mem_p");
+  }
+
   std::shared_ptr<mkldnn::memory> AcquireDiffDstMemoryFromDataPrimitive(
       const std::shared_ptr<mkldnn::memory> user_memory_p,
       std::vector<mkldnn::primitive>& pipeline) {  // NOLINT

python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_bf16_mkldnn_op.py

@@ -216,4 +216,6 @@ class TestWithInput1x1Filter1x1(TestConv2DBf16Op):
 
 
 if __name__ == '__main__':
+    from paddle import enable_static
+    enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_mkldnn_op.py

@@ -233,4 +233,6 @@ class TestMKLDNNDilations(TestConv2DMKLDNNOp):
 
 
 if __name__ == '__main__':
+    from paddle import enable_static
+    enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/mkldnn/test_fusion_gru_bf16_mkldnn_op.py

@@ -110,4 +110,6 @@ class TestFusionGRUINT8MKLDNNOp3(TestFusionGRUBF16MKLDNNOp):
 
 
 if __name__ == "__main__":
+    from paddle import enable_static
+    enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/op_test.py

@@ -1320,6 +1320,13 @@ class OpTest(unittest.TestCase):
         cache_list = None
         if hasattr(self, "cache_name_list"):
             cache_list = self.cache_name_list
+
+        # oneDNN numeric gradient should use CPU kernel
+        use_onednn = False
+        if "use_mkldnn" in op_attrs and op_attrs["use_mkldnn"] == True:
+            op_attrs["use_mkldnn"] = False
+            use_onednn = True
+
         self.op = create_op(
             self.scope,
             self.op_type,
@@ -1328,6 +1335,9 @@ class OpTest(unittest.TestCase):
             op_attrs,
             cache_list=cache_list)
 
+        if use_onednn:
+            op_attrs["use_mkldnn"] = True
+
         if no_grad_set is None:
             no_grad_set = set()
         else: