Merge branch 'develop' into pten/add_standrad_suffix_name_set

Author: chenwhql

.gitignore

@@ -46,6 +46,7 @@ tools/__pycache__
 # This file is automatically generated.
 # TODO(zhiqiang) Move this file to build directory.
 paddle/infrt/dialect/pd_ops.td
+paddle/infrt/dialect/pd_ops_info.h
 .lit_test_times.txt
 paddle/infrt/tests/dialect/Output
 paddle/infrt/tests/lit.cfg.py

paddle/fluid/distributed/ps/coordinator/README.md

@@ -43,10 +43,19 @@ static void ParseArgs(const OpKernelInfo& op_kernel_info,
  auto& attribute_defs = OpKernelInfoHelper::GetAttributeDefs(op_kernel_info);
 
  for (auto& input : input_defs) {
- args_def->AppendInput(input.backend, input.layout, input.dtype);
+ auto type_index =
+ input.is_vector
+ ? std::type_index(typeid(const std::vector<pten::DenseTensor>&))
+ : std::type_index(typeid(const pten::DenseTensor&));
+ args_def->AppendInput(input.backend, input.layout, input.dtype, type_index);
  }
  for (auto& output : output_defs) {
- args_def->AppendOutput(output.backend, output.layout, output.dtype);
+ auto type_index =
+ output.is_vector
+ ? std::type_index(typeid(const std::vector<pten::DenseTensor>&))
+ : std::type_index(typeid(const pten::DenseTensor&));
+ args_def->AppendOutput(output.backend, output.layout, output.dtype,
+ type_index);
  }
  for (auto& attr : attribute_defs) {
  args_def->AppendAttribute(attr.type_index);

paddle/fluid/framework/custom_kernel.cc

@@ -73,7 +73,7 @@ void FakeDot(const paddle::CPUContext& dev_ctx, const paddle::Tensor& x,
  assert(fake_attr_float == 2);
  assert(fake_attr_double == 3);
  assert(fake_attr_int64 == 4);
- assert(fake_attr_f16 == 5);
+ assert(fake_attr_f16 == pten::dtype::float16(5));
  assert(fake_attr_dtype == pten::DataType::UINT32);
  assert(fake_attr_int64_vec.size() == 0);
  assert(fake_attr_int_vec.size() == 0);

paddle/fluid/framework/custom_kernel_test.cc

@@ -63,9 +63,13 @@ void TransformData(const OpKernelType &expected_kernel_type,
  out.ShareDataWith(input_tensor);
  // For NHWC data we need reshape of tensors as MKL-DNN
  // is expecting NHWC dims description order
- platform::MatchShapeToLayout(&out, lin, lout);
- paddle::platform::MKLDNNDeviceContext::tls().set_cur_paddle_data_layout(
- lin);
+ if (lin == DataLayout::kNHWC) {
+ platform::MatchShapeToLayout(&out, lin, lout);
+ // We register only NHWC assuming that model is consistent e.g. either
+ // NHWC or NCHW
+ paddle::platform::MKLDNNDeviceContext::tls()
+ .set_cur_paddle_data_layout(lin);
+ }
  out.set_layout(DataLayout::kMKLDNN);
  out.set_format(out_format);
  } else {

paddle/fluid/framework/data_transform.cc

@@ -27,6 +27,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+extern std::string DataTypeToString(const proto::VarType::Type type);
+extern size_t SizeOfType(proto::VarType::Type type);
+
 template <typename T>
 struct IsComplex : public std::false_type {};
 
@@ -63,6 +66,13 @@ struct DataTypeTrait<void> {
  _ForEachDataTypeHelper_(callback, ::paddle::platform::complex<double>, \
  COMPLEX128);
 
+#define _ForEachIntDataType_(callback) \
+ _ForEachDataTypeHelper_(callback, int, INT32); \
+ _ForEachDataTypeHelper_(callback, int64_t, INT64); \
+ _ForEachDataTypeHelper_(callback, uint8_t, UINT8); \
+ _ForEachDataTypeHelper_(callback, int16_t, INT16); \
+ _ForEachDataTypeHelper_(callback, int8_t, INT8);
+
 #define _ForEachDataTypeSmall_(callback) \
  _ForEachDataTypeHelper_(callback, float, FP32); \
  _ForEachDataTypeHelper_(callback, double, FP64); \
@@ -138,6 +148,24 @@ inline void VisitDataTypeSmall(proto::VarType::Type type, Visitor visitor) {
 #undef VisitDataTypeCallbackSmall
 }
 
+template <typename Visitor>
+inline void VisitIntDataType(proto::VarType::Type type, Visitor visitor) {
+#define VisitIntDataTypeCallback(cpp_type, proto_type) \
+ do { \
+ if (type == proto_type) { \
+ visitor.template apply<cpp_type>(); \
+ return; \
+ } \
+ } while (0)
+
+ _ForEachIntDataType_(VisitIntDataTypeCallback);
+
+ PADDLE_THROW(platform::errors::Unimplemented(
+ "Expected integral data type, but got %s", DataTypeToString(type)));
+
+#undef VisitIntDataTypeCallback
+}
+
 template <typename Visitor>
 inline void VisitDataTypeTiny(proto::VarType::Type type, Visitor visitor) {
 #define VisitDataTypeCallbackTiny(cpp_type, proto_type) \
@@ -166,8 +194,6 @@ inline void VisitDataTypeForHIP(proto::VarType::Type type, Visitor visitor) {
 #undef VisitDataTypeCallbackHIP
 }
 
-extern std::string DataTypeToString(const proto::VarType::Type type);
-extern size_t SizeOfType(proto::VarType::Type type);
 inline std::ostream& operator<<(std::ostream& out,
  const proto::VarType::Type& type) {
  out << DataTypeToString(type);

paddle/fluid/framework/data_type.h

@@ -268,6 +268,9 @@ bool FuseOptimizerOpPass::HasVarDepsBetweenOps(
 
 bool FuseOptimizerOpPass::OpWithKernelSupportCPUAndGPU(
  const std::string &op_type) const {
+ if (op_type == "c_sync_calc_stream" || op_type == "c_sync_comm_stream") {
+ return true;
+ }
  auto &all_kernels = OperatorWithKernel::AllOpKernels();
  auto it = all_kernels.find(op_type);
  // skip op not has kernel

paddle/fluid/framework/ir/fuse_optimizer_ops_pass/fuse_optimizer_op_pass.cc

@@ -1592,11 +1592,8 @@ PDNode *patterns::Transpose::operator()() {
  ->AsOutput()
  ->assert_is_op_output("transpose2", "Out");
 
- auto next_op = pattern->NewNode(next_op_repr())->assert_is_op();
-
  prev_op->LinksTo({transpose_in});
  transpose_op->LinksFrom({transpose_in}).LinksTo({transpose_out});
- next_op->LinksFrom({transpose_out});
  return transpose_out;
 }
 
@@ -1613,11 +1610,8 @@ PDNode *patterns::Reshape::operator()() {
  ->AsOutput()
  ->assert_is_op_output("reshape2", "Out");
 
- auto next_op = pattern->NewNode(next_op_repr())->assert_is_op();
-
  prev_op->LinksTo({reshape_in});
  reshape_op->LinksFrom({reshape_in}).LinksTo({reshape_out});
- next_op->LinksFrom({reshape_out});
  return reshape_out;
 }
 
@@ -1633,11 +1627,8 @@ PDNode *patterns::Slice::operator()() {
  ->AsOutput()
  ->assert_is_op_output("slice", "Out");
 
- auto next_op = pattern->NewNode(next_op_repr())->assert_is_op();
-
  prev_op->LinksTo({slice_in});
  slice_op->LinksFrom({slice_in}).LinksTo({slice_out});
- next_op->LinksFrom({slice_out});
  return slice_out;
 }
 
@@ -1658,12 +1649,9 @@ PDNode *patterns::NearestInterp::operator()() {
  ->assert_is_ops_output({"nearest_interp", "nearest_interp_v2"},
  "Out");
 
- auto next_op = pattern->NewNode(next_op_repr())->assert_is_op();
-
  prev_op->LinksTo({nearest_interp_in});
  nearest_interp_op->LinksFrom({nearest_interp_in})
  .LinksTo({nearest_interp_out});
- next_op->LinksFrom({nearest_interp_out});
  return nearest_interp_out;
 }
 

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -963,7 +963,6 @@ struct Transpose : public PatternBase {
  PATTERN_DECL_NODE(transpose_in);
  PATTERN_DECL_NODE(transpose_op);
  PATTERN_DECL_NODE(transpose_out);
- PATTERN_DECL_NODE(next_op);
 };
 
 // Reshape op
@@ -978,7 +977,6 @@ struct Reshape : public PatternBase {
  PATTERN_DECL_NODE(reshape_in);
  PATTERN_DECL_NODE(reshape_op);
  PATTERN_DECL_NODE(reshape_out);
- PATTERN_DECL_NODE(next_op);
 };
 // Slice op
 // Forward pass for slice.
@@ -992,7 +990,6 @@ struct Slice : public PatternBase {
  PATTERN_DECL_NODE(slice_in);
  PATTERN_DECL_NODE(slice_op);
  PATTERN_DECL_NODE(slice_out);
- PATTERN_DECL_NODE(next_op);
 };
 
 // Nearest Interp op
@@ -1007,7 +1004,6 @@ struct NearestInterp : public PatternBase {
  PATTERN_DECL_NODE(nearest_interp_in);
  PATTERN_DECL_NODE(nearest_interp_op);
  PATTERN_DECL_NODE(nearest_interp_out);
- PATTERN_DECL_NODE(next_op);
 };
 
 // Matmul op