!1803 Optimized MixedPrecisionCast function

Merge pull request !1803 from Kang/master

mindspore/ccsrc/operator/ops.cc

@@ -242,6 +242,7 @@ const PrimitivePtr kPrimIs_ = std::make_shared<Primitive>("is_");
 const PrimitivePtr kPrimIsNot = std::make_shared<Primitive>("is_not");
 const PrimitivePtr kPrimInDict = std::make_shared<Primitive>("in_dict");
 const PrimitivePtr kPrimNotInDict = std::make_shared<Primitive>("not_in_dict");
+const PrimitivePtr kPrimMixedPrecisionCast = std::make_shared<Primitive>("mixed_precision_cast");
 
 // Comm ops
 const PrimitivePtr kPrimMirror = std::make_shared<Primitive>("_MirrorOperator");

mindspore/ccsrc/operator/ops.h

@@ -251,6 +251,7 @@ extern const PrimitivePtr kPrimIs_;
 extern const PrimitivePtr kPrimIsNot;
 extern const PrimitivePtr kPrimInDict;
 extern const PrimitivePtr kPrimNotInDict;
+extern const PrimitivePtr kPrimMixedPrecisionCast;
 
 // Comm ops
 extern const PrimitivePtr kPrimMirror;

mindspore/ccsrc/pipeline/parse/parse.cc

@@ -67,7 +67,7 @@ AnfNodePtr GetMixedPrecisionCastHelp(const FuncGraphPtr &func_graph, const AnfNo
   } else {
     return param;
   }
-  auto cast_helper = prim::GetPythonOps("_mp_cast_helper", "mindspore.ops.composite.base");
+  auto cast_helper = prim::kPrimMixedPrecisionCast;
   auto cast = func_graph->NewCNode({NewValueNode(cast_helper), NewValueNode(dst_type), param});
   return cast;
 }

mindspore/ccsrc/pipeline/static_analysis/prim.cc

@@ -147,9 +147,6 @@ EvalResultPtr StandardPrimEvaluator::EvalPrim(const AnalysisEnginePtr &engine, c
 EvalResultPtr DoSignatureEvaluator::Run(AnalysisEnginePtr engine, const ConfigPtrList &args_conf_list,
                                         AnfNodeConfigPtr out_conf) {
   AbstractBasePtrList args_spec_list;
-  if (!prim_->isa<prim::DoSignaturePrimitive>()) {
-    MS_LOG(EXCEPTION) << "Primitive should be DoSignature, but " << prim_->ToString();
-  }
   if (out_conf->node() == nullptr || !out_conf->node()->isa<CNode>()) {
     MS_LOG(EXCEPTION) << "Node of out_conf should be CNode";
   }
@@ -221,9 +218,6 @@ EvalResultPtr UnpackGraphEvaluator::Run(AnalysisEnginePtr engine, const ConfigPt
   if (out_conf->node() == nullptr || !out_conf->node()->isa<CNode>()) {
     MS_LOG(EXCEPTION) << "Node of out_conf should be CNode";
   }
-  if (!prim_->isa<prim::UnpackGraphPrimitive>()) {
-    MS_LOG(EXCEPTION) << "Primitive should be UnpackGraphPrimitive, but got " << prim_->ToString();
-  }
 
   auto unpack_graph = prim_->cast<prim::UnpackGraphPrimitivePtr>();
   auto out_node = out_conf->node()->cast<CNodePtr>();
@@ -267,6 +261,63 @@ EvalResultPtr UnpackGraphEvaluator::Run(AnalysisEnginePtr engine, const ConfigPt
   return engine->ForwardConfig(out_conf, fn_conf);
 }
 
+AnfNodePtr MixedPrecisionCastHelper(AnfNodePtr source_node, AbstractBasePtr node_type, AnfNodePtr target_type,
+                                    FuncGraphPtr func_graph) {
+  AnfNodePtr target_node = source_node;
+  if (node_type->isa<AbstractTensor>()) {
+    auto x = node_type->cast<AbstractTensorPtr>();
+    if (x->element()->BuildType()->isa<Float>()) {
+      auto cast = prim::GetPythonOps("cast", "mindspore.ops.functional");
+      MS_EXCEPTION_IF_NULL(cast);
+      target_node = func_graph->NewCNode({NewValueNode(cast), source_node, target_type});
+    }
+  } else if (node_type->isa<AbstractTuple>()) {
+    auto x = node_type->cast<AbstractTuplePtr>();
+    auto &items = x->elements();
+    std::size_t size = items.size();
+    std::vector<AnfNodePtr> nodes;
+    nodes.emplace_back(NewValueNode(prim::kPrimMakeTuple));
+    for (int i = 0; i < SizeToInt(size); i++) {
+      AnfNodePtr tuple_node =
+        func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), source_node, NewValueNode(i)});
+      AnfNodePtr node = MixedPrecisionCastHelper(tuple_node, items[i], target_type, func_graph);
+      nodes.emplace_back(node);
+    }
+    target_node = func_graph->NewCNode(nodes);
+  }
+  return target_node;
+}
+
+EvalResultPtr MixedPrecisionCastEvaluator::Run(AnalysisEnginePtr engine, const ConfigPtrList &args_conf_list,
+                                               AnfNodeConfigPtr out_conf) {
+  AbstractBasePtrList args_spec_list;
+  if (out_conf->node() == nullptr || !out_conf->node()->isa<CNode>()) {
+    MS_LOG(EXCEPTION) << "Node of out_conf should be CNode";
+  }
+  auto out_node = out_conf->node()->cast<CNodePtr>();
+  const auto &out_node_inputs = out_node->inputs();
+  if (out_node->inputs().size() == 0 || (out_node_inputs.size() - 1) != args_conf_list.size()) {
+    MS_LOG(EXCEPTION) << "MixedPrecisionCast"
+                      << " args size should equal to inputs size minus 1, but args size " << args_conf_list.size()
+                      << ", inputs size " << out_node_inputs.size();
+  }
+  AnfNodePtrList args_inputs{out_node_inputs.begin() + 1, out_node_inputs.end()};
+  (void)std::transform(args_conf_list.begin(), args_conf_list.end(), std::back_inserter(args_spec_list),
+                       [](const ConfigPtr &ref) -> AbstractBasePtr { return ref->GetEvaluatedValue()->abstract(); });
+
+  ScopePtr scope = kDefaultScope;
+  if (out_conf != nullptr) {
+    scope = out_conf->node()->scope();
+  }
+  ScopeGuard scope_guard(scope);
+
+  FuncGraphPtr func_graph = out_conf->node()->func_graph();
+  AnfNodePtr new_node = MixedPrecisionCastHelper(out_node_inputs[2], args_spec_list[1], out_node_inputs[1], func_graph);
+  AnfNodeConfigPtr fn_conf = engine->MakeConfig(new_node, out_conf->context());
+
+  return engine->ForwardConfig(out_conf, fn_conf);
+}
+
 namespace {
 py::object BuildValue(const ValuePtr &value_ptr) {
   if (value_ptr == nullptr) {

mindspore/ccsrc/pipeline/static_analysis/prim.h

@@ -102,6 +102,22 @@ class UnpackGraphEvaluator : public Evaluator {
   PrimitivePtr prim_;
 };
 
+class MixedPrecisionCastEvaluator : public Evaluator {
+ public:
+  explicit MixedPrecisionCastEvaluator(const PrimitivePtr primitive)
+      : Evaluator("MixedPrecisionCastEvaluator"), prim_(primitive) {}
+  ~MixedPrecisionCastEvaluator() override = default;
+  EvalResultPtr Run(AnalysisEnginePtr engine, const ConfigPtrList &argrefs,
+                    AnfNodeConfigPtr out_config = nullptr) override;
+
+  EvalResultPtr Eval(AnalysisEnginePtr, const AbstractBasePtrList &) override {
+    MS_LOG(EXCEPTION) << "Eval() should not be called, Run() method should be called";
+  }
+
+ private:
+  PrimitivePtr prim_;
+};
+
 bool IsInWhiteList(PrimitivePtr primitive);
 StandardPrimitiveEvalImpl GetPrimitiveInferImpl(const PrimitivePtr &primitive);
 

mindspore/ccsrc/pipeline/static_analysis/static_analysis.cc

@@ -308,6 +308,10 @@ EvaluatorPtr GetPrimEvaluator(const PrimitivePtr &prim, const AnalysisEnginePtr
     evaluator = std::make_shared<UnpackGraphEvaluator>(prim);
     return evaluator;
   }
+  if (prim->name() == prim::kPrimMixedPrecisionCast->name()) {
+    evaluator = std::make_shared<MixedPrecisionCastEvaluator>(prim);
+    return evaluator;
+  }
   if (prim->HasPyEvaluator()) {
     auto prim_py = dyn_cast<PrimitivePy>(prim);
     if (prim_py != nullptr) {

mindspore/nn/wrap/cell_wrapper.py

@@ -21,7 +21,6 @@ from ...common.parameter import Parameter, ParameterTuple
 from ...ops import composite as C
 from ...ops import functional as F
 from ...ops import operations as P
-from ...ops.composite.base import _mp_cast_helper
 from ...ops.operations.comm_ops import _VirtualDataset
 from ..cell import Cell
 from .grad_reducer import DistributedGradReducer
@@ -345,7 +344,7 @@ class WithEvalCell(Cell):
     def construct(self, data, label):
         outputs = self._network(data)
         if self.add_cast_fp32:
-            label = _mp_cast_helper(mstype.float32, label)
+            label = F.mixed_precision_cast(mstype.float32, label)
             outputs = F.cast(outputs, mstype.float32)
         loss = self._loss_fn(outputs, label)
         return loss, outputs, label

mindspore/ops/composite/base.py

@@ -24,7 +24,6 @@ from ..._c_expression import EnvInstance_, GradOperation_, HyperMap_, MultitypeF
 from ...common import dtype as mstype
 from ...common.api import ms_function, _pynative_exec
 from .. import functional as F
-from .. import operations as P
 from ...common.parameter import Parameter
 
 
@@ -297,32 +296,3 @@ env_get = MultitypeFuncGraph("env_get")
 def _tensor_env_get(env, parameter):
     """Used to get env."""
     return F.env_getitem(env, F.ref_to_embed(parameter), F.zeros_like(parameter))
-
-
-_mp_cast_helper = MultitypeFuncGraph('mixed_precision_cast_helper')
-
-
-@_mp_cast_helper.register("TypeType", "Number")
-@core
-def _mixed_precision_cast_helper_1(type_, x):
-    """if x is float cast to type."""
-    # type_ is place holder
-    return x
-
-
-@_mp_cast_helper.register("TypeType", "Tensor")
-@core
-def _mixed_precision_cast_helper_2(type_, x):
-    """if x is float cast to type."""
-    if F.issubclass_(F.dtype(x), mstype.float_):
-        return P.Cast()(x, type_)
-    return x
-
-@_mp_cast_helper.register("TypeType", "Tuple")
-@core
-def _mixed_precision_cast_helper_3(type_, x):
-    """if x is a tuple"""
-    t = ()
-    for item in x:
-        t = t + (_mp_cast_helper(type_, item),)
-    return t

mindspore/ops/functional.py

@@ -126,6 +126,7 @@ is_ = Primitive("is_")
 is_not = Primitive("is_not")
 in_dict = Primitive("in_dict")
 not_in_dict = Primitive("not_in_dict")
+mixed_precision_cast = Primitive("mixed_precision_cast")
 broadcast_gradient_args = Primitive('BroadcastGradientArgs')
 dot = Primitive('dot')
 array_reduce = Primitive('array_reduce')

mindspore/train/amp.py

@@ -21,7 +21,6 @@ from .._checkparam import Rel
 from ..common import dtype as mstype
 from ..nn.wrap.cell_wrapper import _VirtualDatasetCell
 from ..ops import functional as F
-from ..ops.composite.base import _mp_cast_helper
 from ..parallel._utils import _get_parallel_mode
 from .loss_scale_manager import DynamicLossScaleManager, LossScaleManager
 from .parallel_utils import ParallelMode
@@ -98,7 +97,7 @@ def _add_loss_network(network, loss_fn, cast_model_type):
 
         def construct(self, data, label):
             out = self._backbone(data)
-            label = _mp_cast_helper(mstype.float32, label)
+            label = F.mixed_precision_cast(mstype.float32, label)
             return self._loss_fn(F.cast(out, mstype.float32), label)
 
     validator.check_value_type('loss_fn', loss_fn, nn.Cell, None)