Remove special handling of host-memory/device-memory for int32 argume…

…nts in type lists.

The current special case behavior is preserved only for functions and their gradients.
Change: 130100547

RELEASE.md

@@ -7,6 +7,9 @@
   default, simply pass the argument `state_is_tuple=False`.
 * DeviceFactory's AddDevices and CreateDevices functions now return
   a Status instead of void.
+* Int32 elements of list(type) arguments are no longer placed in host memory by
+  default. If necessary, a list(type) argument to a kernel can be placed in host
+  memory using a HostMemory annotation.
 
 # Release 0.10.0
 

tensorflow/core/framework/memory_types.cc

@@ -61,29 +61,6 @@ MemoryType MTypeFromDType(const DataType dtype) {
   return (dtype == DT_INT32) ? HOST_MEMORY : DEVICE_MEMORY;
 }
 
-// Initialize the default memory types for type list arguments from the data
-// types. (The default can be overridden by an explicit HostMemory()
-// declaration.)
-Status SetTypeListMTypesFromDTypes(
-    const NameRangeMap& name_ranges,
-    const protobuf::RepeatedPtrField<OpDef::ArgDef>& args,
-    const DataTypeVector& dtypes, MemoryTypeVector* mtypes) {
-  for (const auto& a : args) {
-    if (!a.type_list_attr().empty()) {
-      auto it = name_ranges.find(a.name());
-      if (it == name_ranges.end()) {
-        return errors::InvalidArgument("Name range for argument ", a.name(),
-                                       " not found.");
-      }
-
-      for (int i = it->second.first; i < it->second.second; ++i) {
-        (*mtypes)[i] = MTypeFromDType(dtypes[i]);
-      }
-    }
-  }
-  return Status::OK();
-}
-
 }  // namespace
 
 Status MemoryTypesForNode(const OpRegistryInterface* op_registry,
@@ -107,12 +84,13 @@ Status MemoryTypesForNode(const OpRegistryInterface* op_registry,
   inp_mtypes->clear();
   out_mtypes->clear();
 
-  if (!status.ok()) {
-    // When there is no kernel def for this op, we can only best-effort derive
-    // the memory type from the data type.  For now, we assume int32 is always
-    // on host memory and other types are always on device memory. We should
-    // do type inference over function body to derive the correct
-    // input/output memory types.
+  // For functions (which have no KernelDef) and their gradients, we can only
+  // best-effort derive the memory type from the data type. For now, we assume
+  // int32 is always on host memory and other types are always on device memory.
+  // TODO(zhifengc,phawkins): We should do type inference over function bodies
+  // to derive the correct input/output memory types. We should also split
+  // host-memory and non host-memory arguments into separate type lists.
+  if (!status.ok() || ndef.op() == "SymbolicGradient") {
     for (const auto& t : inp_dtypes) inp_mtypes->push_back(MTypeFromDType(t));
     for (const auto& t : out_dtypes) out_mtypes->push_back(MTypeFromDType(t));
     return Status::OK();
@@ -127,12 +105,6 @@ Status MemoryTypesForNode(const OpRegistryInterface* op_registry,
   inp_mtypes->resize(GetTotal(inp_names), DEVICE_MEMORY);
   out_mtypes->resize(GetTotal(out_names), DEVICE_MEMORY);
 
-  // For type list arguments, mark int32 arguments as host memory.
-  TF_RETURN_IF_ERROR(SetTypeListMTypesFromDTypes(inp_names, op_def->input_arg(),
-                                                 inp_dtypes, inp_mtypes));
-  TF_RETURN_IF_ERROR(SetTypeListMTypesFromDTypes(
-      out_names, op_def->output_arg(), out_dtypes, out_mtypes));
-
   // Fills in host memory types based on the kernel def.
   const auto& from_proto = kdef->host_memory_arg();
   std::vector<string> host_memory_args(from_proto.begin(), from_proto.end());

tensorflow/core/framework/memory_types_test.cc

@@ -63,11 +63,11 @@ TEST(MemoryTypesForNode, Simple) {
   TF_EXPECT_OK(MemoryTypesForNode(OpRegistry::Global(), DEVICE_CPU, node_def,
                                   &input, &output));
   EXPECT_EQ(MemoryTypeVector({DEVICE_MEMORY, DEVICE_MEMORY, DEVICE_MEMORY,
-                              DEVICE_MEMORY, DEVICE_MEMORY, HOST_MEMORY,
-                              DEVICE_MEMORY, HOST_MEMORY}),
+                              DEVICE_MEMORY, DEVICE_MEMORY, DEVICE_MEMORY,
+                              DEVICE_MEMORY, DEVICE_MEMORY}),
             input);
   EXPECT_EQ(MemoryTypeVector({DEVICE_MEMORY, DEVICE_MEMORY, DEVICE_MEMORY,
-                              HOST_MEMORY, DEVICE_MEMORY, HOST_MEMORY}),
+                              DEVICE_MEMORY, DEVICE_MEMORY, DEVICE_MEMORY}),
             output);
 
   TF_EXPECT_OK(MemoryTypesForNode(OpRegistry::Global(), DEVICE_GPU, node_def,
@@ -77,7 +77,7 @@ TEST(MemoryTypesForNode, Simple) {
                         HOST_MEMORY, HOST_MEMORY, HOST_MEMORY, HOST_MEMORY}),
       input);
   EXPECT_EQ(MemoryTypeVector({HOST_MEMORY, HOST_MEMORY, HOST_MEMORY,
-                              HOST_MEMORY, DEVICE_MEMORY, HOST_MEMORY}),
+                              DEVICE_MEMORY, DEVICE_MEMORY, DEVICE_MEMORY}),
             output);
 }
 

tensorflow/core/kernels/function_ops.cc

@@ -109,12 +109,20 @@ REGISTER_KERNEL_BUILDER(Name("_Retval")
 
 class PassOn : public OpKernel {
  public:
-  explicit PassOn(OpKernelConstruction* ctx) : OpKernel(ctx) {}
-
-  void Compute(OpKernelContext* ctx) override {
+  explicit PassOn(OpKernelConstruction* ctx) : OpKernel(ctx) {
     OP_REQUIRES(ctx, ctx->num_inputs() == ctx->num_outputs(),
                 errors::Internal("#inputs != #outputs : ", ctx->num_inputs(),
                                  " vs. ", ctx->num_outputs()));
+    for (int i = 0; i < ctx->num_inputs(); ++i) {
+      OP_REQUIRES(
+          ctx, input_type(i) == output_type(i),
+          errors::Internal("Input and output types for position ", i,
+                           " do not match: ", DataTypeString(input_type(i)),
+                           " vs. ", DataTypeString(output_type(i))));
+    }
+  }
+
+  void Compute(OpKernelContext* ctx) override {
     for (int i = 0; i < ctx->num_inputs(); ++i) {
       ctx->set_output(i, ctx->input(i));
     }
@@ -140,12 +148,14 @@ REGISTER_GPU_KERNELS(double);
 
 REGISTER_KERNEL_BUILDER(Name("_ListToArray")
                             .Device(DEVICE_GPU)
+                            .HostMemory("input")
                             .HostMemory("output")
                             .TypeConstraint<int32>("T"),
                         PassOn);
 REGISTER_KERNEL_BUILDER(Name("_ArrayToList")
                             .Device(DEVICE_GPU)
                             .HostMemory("input")
+                            .HostMemory("output")
                             .TypeConstraint<int32>("T"),
                         PassOn);