fused select and greater op to improve bert perfermance on GPU

akg

@@ -1 +1 @@
-Subproject commit d237aa7d8e9d3fb709bda9f30205b02129bc2b59
+Subproject commit 4d897c23fc41c5f7013efd0c517796233671518a

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

@@ -44,10 +44,11 @@ bool IsBasicOp(const AnfNodePtr &node, bool is_before_kernel_select) {
   }
 #elif ENABLE_GPU
   std::vector<PrimitivePtr> fusable_basic_ops = {
-    prim::kPrimAbs,     prim::kPrimRound, prim::kPrimNeg,        prim::kPrimExp,      prim::kPrimTensorAdd,
-    prim::kPrimRealDiv, prim::kPrimMul,   prim::kPrimMinimum,    prim::kPrimMaximum,  prim::kPrimLog,
-    prim::kPrimPow,     prim::kPrimSub,   prim::kPrimRsqrt,      prim::kPrimSqrt,     prim::kPrimCast,
-    prim::kPrimAddN,    prim::kPrimEqual, prim::kPrimReciprocal, prim::KPrimTransData};
+    prim::kPrimAbs,     prim::kPrimRound, prim::kPrimNeg,        prim::kPrimExp,       prim::kPrimTensorAdd,
+    prim::kPrimRealDiv, prim::kPrimMul,   prim::kPrimMinimum,    prim::kPrimMaximum,   prim::kPrimLog,
+    prim::kPrimPow,     prim::kPrimSub,   prim::kPrimRsqrt,      prim::kPrimSqrt,      prim::kPrimCast,
+    prim::kPrimAddN,    prim::kPrimEqual, prim::kPrimReciprocal, prim::KPrimTransData, prim::kPrimSelect,
+    prim::kPrimGreater};
 #else
   std::vector<PrimitivePtr> fusable_basic_ops;
 #endif

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

@@ -50,10 +50,11 @@ bool IsBasicFuseOp(const AnfNodePtr &node, bool is_before_kernel_select) {
   }
 #elif ENABLE_GPU
   std::vector<PrimitivePtr> basic_ops = {
-    prim::kPrimAbs,     prim::kPrimRound, prim::kPrimNeg,        prim::kPrimExp,      prim::kPrimTensorAdd,
-    prim::kPrimRealDiv, prim::kPrimMul,   prim::kPrimMinimum,    prim::kPrimMaximum,  prim::kPrimLog,
-    prim::kPrimPow,     prim::kPrimSub,   prim::kPrimRsqrt,      prim::kPrimSqrt,     prim::kPrimCast,
-    prim::kPrimAddN,    prim::kPrimEqual, prim::kPrimReciprocal, prim::KPrimTransData};
+    prim::kPrimAbs,     prim::kPrimRound, prim::kPrimNeg,        prim::kPrimExp,       prim::kPrimTensorAdd,
+    prim::kPrimRealDiv, prim::kPrimMul,   prim::kPrimMinimum,    prim::kPrimMaximum,   prim::kPrimLog,
+    prim::kPrimPow,     prim::kPrimSub,   prim::kPrimRsqrt,      prim::kPrimSqrt,      prim::kPrimCast,
+    prim::kPrimAddN,    prim::kPrimEqual, prim::kPrimReciprocal, prim::KPrimTransData, prim::kPrimSelect,
+    prim::kPrimGreater};
 #else
   std::vector<PrimitivePtr> basic_ops;
 #endif

mindspore/ops/_op_impl/akg/gpu/__init__.py

@@ -52,5 +52,7 @@ from .squeeze import _squeeze_akg
 from .squeeze_grad import _squeeze_grad_akg
 from .sub import _sub_akg
 from .tile import _tile_akg
+from .select import _select_akg
+from .greater import _greater_akg
 
 # Please insert op register in lexicographical order of the filename.

mindspore/ops/_op_impl/akg/gpu/greater.py

@@ -0,0 +1,34 @@
+# Copyright 2020 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.
+# ============================================================================
+
+"""Greater op"""
+from mindspore.ops.op_info_register import op_info_register, AkgGpuRegOp, DataType as DT
+
+op_info = AkgGpuRegOp("Greater") \
+    .fusion_type("ELEMWISE") \
+    .input(0, "x") \
+    .input(1, "y") \
+    .output(0, "output") \
+    .dtype_format(DT.F16_Default, DT.F16_Default, DT.BOOL_Default) \
+    .dtype_format(DT.F32_Default, DT.F32_Default, DT.BOOL_Default) \
+    .dtype_format(DT.F16_5HD, DT.F16_5HD, DT.BOOL_5HD) \
+    .dtype_format(DT.F32_5HD, DT.F32_5HD, DT.BOOL_5HD) \
+    .get_op_info()
+
+
+@op_info_register(op_info)
+def _greater_akg():
+    """Greater Akg register"""
+    return

mindspore/ops/_op_impl/akg/gpu/select.py

@@ -0,0 +1,37 @@
+# Copyright 2020 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.
+# ============================================================================
+
+"""Select op"""
+from mindspore.ops.op_info_register import op_info_register, AkgGpuRegOp, DataType as DT
+
+op_info = AkgGpuRegOp("Select") \
+    .fusion_type("ELEMWISE") \
+    .input(0, "condition") \
+    .input(1, "x") \
+    .input(2, "y") \
+    .output(0, "output") \
+    .dtype_format(DT.BOOL_Default, DT.F16_Default, DT.F16_Default, DT.F16_Default) \
+    .dtype_format(DT.BOOL_Default, DT.F32_Default, DT.F32_Default, DT.F32_Default) \
+    .dtype_format(DT.BOOL_Default, DT.I32_Default, DT.I32_Default, DT.I32_Default) \
+    .dtype_format(DT.BOOL_5HD, DT.F16_5HD, DT.F16_5HD, DT.F16_5HD) \
+    .dtype_format(DT.BOOL_5HD, DT.F32_5HD, DT.F32_5HD, DT.F32_5HD) \
+    .dtype_format(DT.BOOL_5HD, DT.I32_5HD, DT.I32_5HD, DT.I32_5HD) \
+    .get_op_info()
+
+
+@op_info_register(op_info)
+def _select_akg():
+    """Select Akg register"""
+    return