!28314 add ascend Conv2dBackpropInputBiasaddFusion

Merge pull request !28314 from lyqlola/master

mindspore/ccsrc/backend/optimizer/ascend/ascend_backend_optimization.cc

@@ -68,6 +68,7 @@
 #include "backend/optimizer/ascend/ir_fusion/mul_add_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/mul_addn_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/matmul_biasadd_fusion.h"
+#include "backend/optimizer/ascend/ir_fusion/conv2d_backprop_input_biasadd_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/remove_reshape_pair.h"
 #include "backend/optimizer/ascend/ir_fusion/derelu_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/batchnorm_to_bninfer.h"
@@ -218,6 +219,7 @@ void AddAscendIRFusionPass(PassManager *ir_fusion_pm) {
   ir_fusion_pm->AddPass(std::make_shared<MulAddFusion>());
   ir_fusion_pm->AddPass(std::make_shared<MulAddNFusion>());
   ir_fusion_pm->AddPass(std::make_shared<MatmulBiasaddFusion>());
+  ir_fusion_pm->AddPass(std::make_shared<Conv2dBackpropInputBiasaddFusion>());
   ir_fusion_pm->AddPass(std::make_shared<MatmulAddFusion>());
   ir_fusion_pm->AddPass(std::make_shared<AddnFission>());
   ir_fusion_pm->AddPass(std::make_shared<DereluFusion>());

mindspore/ccsrc/backend/optimizer/ascend/ir_fusion/conv2d_backprop_input_biasadd_fusion.cc

@@ -0,0 +1,62 @@
+/**
+ * Copyright 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.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "backend/optimizer/ascend/ir_fusion/conv2d_backprop_input_biasadd_fusion.h"
+#include <memory>
+#include <vector>
+#include "backend/session/anf_runtime_algorithm.h"
+#include "backend/optimizer/common/helper.h"
+#include "utils/utils.h"
+#include "utils/trace_base.h"
+#include "runtime/device/ascend/lic_manager.h"
+
+namespace mindspore {
+namespace opt {
+const BaseRef Conv2dBackpropInputBiasaddFusion::DefinePattern() const {
+  VectorRef conv2d_bp_input({conv2d_bp_input_var_, x0_, x1_});
+  VectorRef pattern({prim::kPrimBiasAdd, conv2d_bp_input, x2_});
+  return pattern;
+}
+
+const AnfNodePtr Conv2dBackpropInputBiasaddFusion::Process(const FuncGraphPtr &graph, const AnfNodePtr &node,
+                                                           const EquivPtr &equiv) const {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(node);
+
+  auto conv2d_bp_input = GetAnfNodeByVar(equiv, conv2d_bp_input_var_);
+  if (conv2d_bp_input == nullptr || !conv2d_bp_input->isa<CNode>()) {
+    MS_LOG(EXCEPTION) << "Get cnode Conv2DBackpropInput failed!" << trace::DumpSourceLines(conv2d_bp_input);
+  }
+
+  // If there is a side-effect operator in the fusion, do not merge
+  if (!IsStateEquivalent(node, conv2d_bp_input)) {
+    return node;
+  }
+
+  std::vector<AnfNodePtr> inputs;
+  inputs.emplace_back(NewValueNode(std::make_shared<Primitive>(prim::kPrimConv2DTranspose->name())));
+  inputs.emplace_back(GetAnfNodeByVar(equiv, x0_));
+  inputs.emplace_back(GetAnfNodeByVar(equiv, x1_));
+  inputs.emplace_back(GetAnfNodeByVar(equiv, x2_));
+  auto new_node = NewCNode(inputs, graph);
+  MS_EXCEPTION_IF_NULL(new_node);
+  new_node->set_scope(node->scope());
+  new_node->set_abstract(node->abstract());
+
+  AnfAlgo::CopyNodeAttrs(conv2d_bp_input, new_node);
+  return new_node;
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/backend/optimizer/ascend/ir_fusion/conv2d_backprop_input_biasadd_fusion.h

@@ -0,0 +1,46 @@
+/**
+ * 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.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FUSION_CONV2D_BACKPROP_INPUT_BIASADD_FUSION_H_
+#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FUSION_CONV2D_BACKPROP_INPUT_BIASADD_FUSION_H_
+
+#include <memory>
+#include "backend/optimizer/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+class Conv2dBackpropInputBiasaddFusion : public PatternProcessPass {
+ public:
+  explicit Conv2dBackpropInputBiasaddFusion(bool multigraph = true)
+      : PatternProcessPass("conv2d_backprop_input_biasadd_fusion", multigraph) {
+    x0_ = std::make_shared<Var>();
+    x1_ = std::make_shared<Var>();
+    x2_ = std::make_shared<Var>();
+    conv2d_bp_input_var_ = std::make_shared<Var>(std::make_shared<Primitive>(prim::kPrimConv2DBackpropInput->name()));
+  }
+  ~Conv2dBackpropInputBiasaddFusion() override = default;
+  const BaseRef DefinePattern() const override;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+
+ private:
+  VarPtr x0_;
+  VarPtr x1_;
+  VarPtr x2_;
+  VarPtr conv2d_bp_input_var_;
+};
+}  // namespace opt
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FUSION_CONV2D_BACKPROP_INPUT_BIASADD_FUSION_H_

mindspore/python/mindspore/ops/_op_impl/tbe/__init__.py

@@ -15,7 +15,7 @@
 
 """tbe ops"""
 from .celu import _celu_tbe
-from .cummin import  _cummin_tbe
+from .cummin import _cummin_tbe
 from .abs import _abs_tbe
 from .abs_ds import _abs_ds_tbe
 from .inplace_add import _inplace_add_tbe
@@ -89,6 +89,7 @@ from .conv2d_backprop_input import _conv2d_backprop_input_tbe
 from .conv2d_ds import _conv2d_ds_tbe
 from .conv2d_backprop_filter_ds import _conv2d_backprop_filter_ds_tbe
 from .conv2d_backprop_input_ds import _conv2d_backprop_input_ds_tbe
+from .conv2d_transpose import _conv2d_transpose_tbe
 from .confusion_mul_grad import _confusion_mul_grad_tbe
 from .dropout_do_mask import _dropout_do_mask_tbe
 from .dropout_do_mask_ds import _dropout_do_mask_ds_tbe

mindspore/python/mindspore/ops/_op_impl/tbe/conv2d_transpose.py

@@ -0,0 +1,48 @@
+# Copyright 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.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""Conv2DTranspose op"""
+from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
+
+conv2d_transpose_op_info = TBERegOp("Conv2DTranspose") \
+    .fusion_type("CONVOLUTION") \
+    .async_flag(False) \
+    .binfile_name("conv2d_transpose_d.so") \
+    .compute_cost(10) \
+    .kernel_name("conv2d_transpose_d") \
+    .partial_flag(True) \
+    .attr("input_sizes", "required", "listInt", "all") \
+    .attr("stride", "required", "listInt", "all") \
+    .attr("pad_list", "required", "listInt", "all") \
+    .attr("dilation", "optional", "listInt", "all", "1,1,1,1") \
+    .attr("groups", "optional", "int", "all", "1") \
+    .attr("format", "optional", "str", "all", "NHWC") \
+    .attr("output_padding", "optional", "listInt", "all", "0,0,0,0") \
+    .attr("offset_x", "optional", "int", "all", "0") \
+    .input(0, "x", False, "required", "all") \
+    .input(1, "filter", False, "required", "all") \
+    .input(2, "bias", False, "optional", "all") \
+    .input(3, "offset_w", False, "optional", "all") \
+    .output(0, "y", True, "required", "all") \
+    .dtype_format(DataType.F16_5HD, DataType.F16_FracZ, DataType.F16_Default, DataType.I8_Default, DataType.F16_5HD) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_FracZ, DataType.F32_Default, DataType.I8_Default, DataType.F32_5HD) \
+    .dtype_format(DataType.I8_5HD, DataType.I8_FracZ, DataType.I32_Default, DataType.I8_Default, DataType.I32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(conv2d_transpose_op_info)
+def _conv2d_transpose_tbe():
+    """Conv2DTranspose TBE register"""
+    return