Add metadata-only inputs. (#5635)

* Assign a stream to CPU inputs with GPU (non-metadata) inputs * Add Metadata input device - this declares that the input is used for metadata (shape, dtype, etc) access only * Don't synchronize metadata inputs in executor. * Don't prolong the lifetime of metadata-only inputs. * Add InputDevice specifier to random number generators shape_like input. --------- Signed-off-by: Michał Zientkiewicz <[email protected]>
NVIDIA · Sep 18, 2024 · 94f02ad · 94f02ad
1 parent 8904209
commit 94f02ad
Show file tree

Hide file tree

Showing 15 changed files with 170 additions and 34 deletions.
diff --git a/dali/operators/generic/cast.cc b/dali/operators/generic/cast.cc
@@ -76,7 +76,7 @@ DALI_SCHEMA(Cast)
 DALI_SCHEMA(CastLike)
     .DocStr("Cast the first tensor to the type of the second tensor.")
     .NumInput(2)
-    .InputDevice(1, InputDevice::Any)
+    .InputDevice(1, InputDevice::Metadata)
     .NumOutput(1)
     .AllowSequences()
     .SupportVolumetric();

diff --git a/dali/operators/generic/constant_value.cc b/dali/operators/generic/constant_value.cc
@@ -84,7 +84,7 @@ DALI_SCHEMA(FullLike)
     .DocStr(R"code(Returns new data with the same shape and type as the input data, filled with a `fill_value`.)code")
     .NumInput(2)
     .InputDox(0, "data_like", "TensorList", R"code(The input data value to copy the shape and type from.)code")
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .InputDox(1, "fill_value", "TensorList", R"code(The fill value.)code")
     .NumOutput(1);
 DALI_REGISTER_OPERATOR(FullLike, FullLike<CPUBackend>, CPU);
@@ -102,7 +102,7 @@ DALI_SCHEMA(ZerosLike)
     .DocStr(R"code(Returns new data with the same shape and type as the input array, filled with zeros.)code")
     .NumInput(1)
     .InputDox(0, "data_like", "TensorList", R"code(The input data value to copy the shape and type from.)code")
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalTypeArg("dtype", R"code(Overrides the output data type.)code", DALI_INT32);
 DALI_REGISTER_OPERATOR(ZerosLike, ZerosLike<CPUBackend>, CPU);
@@ -120,7 +120,7 @@ DALI_SCHEMA(OnesLike)
     .DocStr(R"code(Returns new data with the same shape and type as the input array, filled with ones.)code")
     .NumInput(1)
     .InputDox(0, "data_like", "TensorList", R"code(The input data value to copy the shape and type from.)code")
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalTypeArg("dtype", R"code(Overrides the output data type.)code", DALI_INT32);
 DALI_REGISTER_OPERATOR(OnesLike, OnesLike<CPUBackend>, CPU);

diff --git a/dali/operators/generic/shapes.cc b/dali/operators/generic/shapes.cc
@@ -19,7 +19,7 @@ namespace dali {
 DALI_SCHEMA(Shapes)
     .DocStr(R"code(Returns the shapes of inputs.)code")
     .NumInput(1)
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AllowSequences()
     .SupportVolumetric()

diff --git a/dali/operators/random/beta_distribution_cpu.cc b/dali/operators/random/beta_distribution_cpu.cc
@@ -38,6 +38,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg("alpha", R"code(The alpha parameter, a positive ``float32`` scalar.)code", 1.0f,
                     true)

diff --git a/dali/operators/random/choice_cpu.cc b/dali/operators/random/choice_cpu.cc
@@ -42,6 +42,7 @@ that is: :meth:`nvidia.dali.types.DALIDataType`, :meth:`nvidia.dali.types.DALIIm
               "Otherwise ``__a`` is treated as 1D array of input samples.")
     .InputDox(1, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(1, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg<std::vector<float>>("p",
                                         "Distribution of the probabilities. "

diff --git a/dali/operators/random/coin_flip_cpu.cc b/dali/operators/random/coin_flip_cpu.cc
@@ -30,6 +30,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg<float>("probability",
       R"code(Probability of value 1.)code",
@@ -51,6 +52,7 @@ sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddParent("random__CoinFlip")
     .Deprecate("random__CoinFlip");  // Deprecated in 0.30

diff --git a/dali/operators/random/normal_distribution_cpu.cc b/dali/operators/random/normal_distribution_cpu.cc
@@ -28,7 +28,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg<float>("mean",
       R"code(Mean of the distribution.)code",
@@ -51,7 +51,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
-    .InputDevice(0, InputDevice::Any)
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddParent("random__Normal")
     .Deprecate("random__Normal");  // Deprecated in 0.30

diff --git a/dali/operators/random/uniform_distribution_cpu.cc b/dali/operators/random/uniform_distribution_cpu.cc
@@ -32,6 +32,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg("range",
       R"code(Range ``[min, max)`` of a continuous uniform distribution.
@@ -67,6 +68,7 @@ a single value per sample is generated.
     .NumInput(0, 1)
     .InputDox(0, "shape_like", "TensorList",
               "Shape of this input will be used to infer the shape of the output, if provided.")
+    .InputDevice(0, InputDevice::Metadata)
     .NumOutput(1)
     .AddOptionalArg("range",
       R"code(Range ``[min, max)`` of a continuous uniform distribution.

diff --git a/dali/pipeline/executor/executor2/exec_graph.h b/dali/pipeline/executor/executor2/exec_graph.h
@@ -90,13 +90,16 @@ struct ExecEdge {
   /** The index of the input in OpSpec. It matches the edge's index in consumer->inputs. */
   int consumer_input_idx = 0;
   StorageDevice device = {};
+  /** The input passes only the metadata, skipping stream synchronization. */
+  bool metadata = false;
 
   constexpr bool operator==(const ExecEdge &other) const {
     return producer == other.producer &&
            consumer == other.consumer &&
            producer_output_idx == other.producer_output_idx &&
            consumer_input_idx == other.consumer_input_idx &&
-           device == other.device;
+           device == other.device &&
+           metadata == other.metadata;
   }
 
   constexpr bool operator!=(const ExecEdge &other) const {

diff --git a/dali/pipeline/executor/executor2/exec_graph_lowering.cc b/dali/pipeline/executor/executor2/exec_graph_lowering.cc
@@ -50,7 +50,16 @@ void ExecGraph::Lower(const graph::OpGraph &def) {
       for (auto &consumer : out->consumers) {
         auto *exec_con = def2exec[consumer.op];
         assert(exec_con != nullptr);
-        Link(&exec_node, o, exec_con, consumer.idx)->device = dev;
+        auto *edge = Link(&exec_node, o, exec_con, consumer.idx);
+        edge->device = dev;
+        if (consumer.op) {
+          auto &consumer_spec = consumer.op->spec;
+          auto &schema = consumer_spec.GetSchemaOrDefault();
+          if (edge->consumer_input_idx < schema.MaxNumInput()) {  // only regular inputs
+            if (schema.GetInputDevice(edge->consumer_input_idx) == InputDevice::Metadata)
+              edge->metadata = true;
+          }
+        }
       }
       exec_node.outputs[o].device = dev;
     }

diff --git a/dali/pipeline/executor/executor2/exec_node_task.cc b/dali/pipeline/executor/executor2/exec_node_task.cc
@@ -278,16 +278,13 @@ void OpTask::SetWorkspaceInputs() {
   auto process_input = [&](int i, auto backend) {
     using Backend = decltype(backend);
     const auto &inp = TaskInput<Backend>(ti);
-    // If the output order of the operator is `host` then we don't wait for GPU
-    // inputs - they can't be accessed directly on host and the operator will
-    // have to issue some sort of synchronization if and when necessary.
-    // This optimization is essential to avoid oversynchronization
-    // when the operator needs to access the metadata only (e.g. getting the shape).
-    if ((order.is_device() || std::is_same_v<Backend, CPUBackend>) /*see comment above */ &&
-        inp.event && inp.order != order)
+    bool is_meta = node_->inputs[i]->metadata;
+    // metadata-only inputs don't need to be synchronized
+    if (!is_meta && inp.event && inp.order != order)
       events.insert(inp.event);
 
-    if (inp.order == order) {  // use the input directly
+    // metadata-only inputs don't need a proper stream
+    if (inp.order == order || is_meta) {  // use the input directly
       ws_->SetInput(i, inp.data);
     } else {  // create another TL and set its order (and layout, while we're at it)
       auto tl = std::make_shared<TensorList<Backend>>();
@@ -477,7 +474,7 @@ tasking::SharedTask ExecNodeTask::CreateTask(ExecNode *node, const WorkspacePara
   }
 }
 
-void ClearWorkspacePayload(Workspace &ws) {
+void ClearWorkspacePayload(Workspace &ws, ExecNode &node) {
   auto event = ws.has_event() ? ws.event() : nullptr;
   for (int i = 0; i < ws.NumInput(); i++) {
     // TODO(michalz): Some smarter deletion management
@@ -492,14 +489,16 @@ void ClearWorkspacePayload(Workspace &ws) {
     if (ws.InputIsType<CPUBackend>(i)) {
       if (auto &pinp = ws.InputPtr<CPUBackend>(i)) {
         auto &inp = *pinp;
-        if (inp.is_pinned() && event && inp.order() != ws.output_order())
+        if (event &&
+            !node.inputs[i]->metadata &&
+            inp.is_pinned() && inp.order() != ws.output_order())
           inp.order().wait(event);
         ws.SetInput<CPUBackend>(i, nullptr);
       }
     } else if (ws.InputIsType<GPUBackend>(i)) {
       if (auto &pinp = ws.InputPtr<GPUBackend>(i)) {
           auto &inp = *pinp;
-        if (event && inp.order() != ws.output_order())
+        if (event && !node.inputs[i]->metadata && inp.order() != ws.output_order())
           inp.order().wait(event);
         ws.SetInput<GPUBackend>(i, nullptr);
       }
@@ -525,7 +524,7 @@ void ClearWorkspacePayload(Workspace &ws) {
 
 void ExecNodeTask::ClearWorkspace() {
   assert(ws_);
-  ClearWorkspacePayload(*ws_);
+  ClearWorkspacePayload(*ws_, *node_);
 }
 
 }  // namespace exec2

diff --git a/dali/pipeline/executor/executor2/stream_assignment.h b/dali/pipeline/executor/executor2/stream_assignment.h
@@ -35,15 +35,15 @@ template <StreamPolicy policy>
 class StreamAssignment;
 
 inline bool NeedsStream(const ExecNode *node) {
-  if (node->is_pipeline_output) {
-    for (auto &pipe_out : node->inputs) {
-      if (pipe_out->device == StorageDevice::GPU)
+  if (node->is_pipeline_output || node->backend == OpType::CPU) {
+    for (auto &input : node->inputs) {
+      if (input->device == StorageDevice::GPU && !input->metadata)
         return true;
     }
+    return false;
   } else {
-    return node->backend != OpType::CPU;
+    return true;
   }
-  return false;
 }
 
 inline OpType NodeType(const ExecNode *node) {
@@ -117,6 +117,12 @@ class StreamAssignment<StreamPolicy::PerBackend> {
         if (has_gpu_)
           return;  // we already have both, nothing more can happen
         break;
+      case OpType::CPU:
+        if (NeedsStream(&node)) {  // treat CPU nodes with GPU inputs as GPU
+          has_gpu_ = true;
+          if (has_mixed_)
+            return;
+        }
       default:
         break;
       }
@@ -128,11 +134,14 @@ class StreamAssignment<StreamPolicy::PerBackend> {
    * If the node is a Mixed node, it gets stream index 0.
    * If the node is a GPU node it gets stream index 1 if there are any mixed nodes, otherwise
    * the only stream is the GPU stream and the returned index is 0.
+   * CPU nodes get GPU stream if they need one (i.e. they have a GPU input)
    */
   std::optional<int> operator[](const ExecNode *node) const {
     switch (NodeType(node)) {
     case OpType::CPU:
-      return std::nullopt;
+      if (!NeedsStream(node))
+        return std::nullopt;
+      // fall-through to GPU
     case OpType::GPU:
       return has_mixed_ ? 1 : 0;
     case OpType::MIXED:

diff --git a/dali/pipeline/executor/executor2/stream_assignment_test.cc b/dali/pipeline/executor/executor2/stream_assignment_test.cc
@@ -48,6 +48,31 @@ DALI_REGISTER_OPERATOR(StreamAssignmentDummyOp, StreamAssignmentDummyOp<CPUBacke
 DALI_REGISTER_OPERATOR(StreamAssignmentDummyOp, StreamAssignmentDummyOp<MixedBackend>, Mixed);
 DALI_REGISTER_OPERATOR(StreamAssignmentDummyOp, StreamAssignmentDummyOp<GPUBackend>, GPU);
 
+
+template <typename Backend>
+class StreamAssignmentMetaOp : public Operator<Backend> {
+ public:
+  using Operator<Backend>::Operator;
+  USE_OPERATOR_MEMBERS();
+
+  void RunImpl(Workspace &ws) override {}
+  bool SetupImpl(std::vector<OutputDesc> &output_desc, const Workspace &ws) override {
+    return false;
+  }
+};
+
+DALI_SCHEMA(StreamAssignmentMetaOp)
+  .NumInput(0, 999)
+  .InputDevice(0, 999, InputDevice::Metadata)
+  .NumOutput(0)
+  .AdditionalOutputsFn([](const OpSpec &spec) {
+    return spec.NumOutput();
+  });
+
+DALI_REGISTER_OPERATOR(StreamAssignmentMetaOp, StreamAssignmentMetaOp<CPUBackend>, CPU);
+DALI_REGISTER_OPERATOR(StreamAssignmentMetaOp, StreamAssignmentMetaOp<MixedBackend>, Mixed);
+DALI_REGISTER_OPERATOR(StreamAssignmentMetaOp, StreamAssignmentMetaOp<GPUBackend>, GPU);
+
 namespace exec2 {
 
 namespace {
@@ -71,6 +96,27 @@ OpSpec SpecMixed() {
   return SpecDev("mixed");
 }
 
+
+OpSpec SpecMetaDev(const std::string &device) {
+  return OpSpec("StreamAssignmentMetaOp")
+    .AddArg("device", device)
+    .AddArg("num_threads", 1)
+    .AddArg("max_batch_size", 1);
+}
+
+OpSpec SpecMetaGPU() {
+  return SpecMetaDev("gpu");
+}
+
+OpSpec SpecMetaCPU() {
+  return SpecMetaDev("cpu");
+}
+
+OpSpec SpecMetaMixed() {
+  return SpecMetaDev("mixed");
+}
+
+
 auto MakeNodeMap(const ExecGraph &graph) {
   std::map<std::string_view, const ExecNode *, std::less<>> map;
   for (auto &n : graph.Nodes())
@@ -122,6 +168,38 @@ TEST(Exec2Test, StreamAssignment_Single_CPUMixedGPU) {
   EXPECT_EQ(assignment[map["c"]], 0);
 }
 
+template <StreamPolicy policy>
+void TestGPU2CPUAssignment() {
+      graph::OpGraph::Builder b;
+  b.Add("a",
+        SpecGPU()
+        .AddOutput("a->b", "gpu")
+        .AddOutput("a->c", "gpu"));
+  b.Add("b",
+        SpecCPU()
+        .AddInput("a->b", "gpu")
+        .AddOutput("b->out", "cpu"));
+  b.Add("c",
+        SpecMetaCPU()
+        .AddInput("a->c", "gpu")
+        .AddOutput("c->out", "cpu"));
+  b.AddOutput("b->out_cpu");
+  b.AddOutput("c->out_cpu");
+  auto g = std::move(b).GetGraph(true);
+  ExecGraph eg;
+  eg.Lower(g);
+
+  StreamAssignment<policy> assignment(eg);
+  auto map = MakeNodeMap(eg);
+  EXPECT_EQ(assignment[map["a"]], 0);
+  EXPECT_EQ(assignment[map["b"]], 0);  // CPU operator with GPU input
+  EXPECT_EQ(assignment[map["c"]], std::nullopt);  // metadata-only
+}
+
+TEST(Exec2Test, StreamAssignment_Single_GPU2CPU) {
+  TestGPU2CPUAssignment<StreamPolicy::Single>();
+}
+
 
 TEST(Exec2Test, StreamAssignment_PerBackend_OnlyCPU) {
   graph::OpGraph::Builder b;
@@ -194,6 +272,13 @@ TEST(Exec2Test, StreamAssignment_PerBackend_CPUMixedGPU) {
 }
 
 
+TEST(Exec2Test, StreamAssignment_PerBackend_GPU2CPU) {
+  TestGPU2CPUAssignment<StreamPolicy::PerBackend>();
+}
+
+TEST(Exec2Test, StreamAssignment_OperOperator_GPU2CPU) {
+  TestGPU2CPUAssignment<StreamPolicy::PerOperator>();
+}
 
 TEST(Exec2Test, StreamAssignment_PerOperator_1) {
   ExecGraph eg;
@@ -272,7 +357,7 @@ TEST(Exec2Test, StreamAssignment_PerOperator_2) {
         SpecGPU()
         .AddOutput("i->j", "gpu"));
   b.Add("j",
-        SpecCPU()
+        SpecMetaCPU()
         .AddInput("i->j", "gpu")
         .AddOutput("j->h", "cpu"));
   b.Add("b",
@@ -320,15 +405,15 @@ TEST(Exec2Test, StreamAssignment_PerOperator_2) {
   StreamAssignment<StreamPolicy::PerOperator> assignment(eg);
   auto map = MakeNodeMap(eg);
   EXPECT_EQ(assignment[map["a"]], 0);
-  EXPECT_EQ(assignment[map["b"]], std::nullopt);
+  EXPECT_EQ(assignment[map["b"]], 0);  // CPU operator with a GPU input needs a stream
   EXPECT_EQ(assignment[map["c"]], 0);
   EXPECT_EQ(assignment[map["d"]], 0);
   EXPECT_EQ(assignment[map["e"]], 1);
   EXPECT_EQ(assignment[map["f"]], 0);
   EXPECT_EQ(assignment[map["g"]], 0);
   EXPECT_EQ(assignment[map["h"]], 0);
   EXPECT_EQ(assignment[map["i"]], 2);
-  EXPECT_EQ(assignment[map["j"]], std::nullopt);
+  EXPECT_EQ(assignment[map["j"]], std::nullopt);  // metadata only
   EXPECT_EQ(assignment[map["k"]], 3);
 }