Support Offline Tensors through ONNXIFI layer

Summary:
Previous import was b2ec1a8041879b7be98d81387a14cae895f952f4

Included changes:
- **[97fe555](houseroad/foxi@97fe555)**: Add deferred weight reader pointer when initializing the graph (#15) <Yinghai Lu>
- **[ba2faf7](houseroad/foxi@ba2faf7)**: Add status and timeout to events (#14) <Jack Montgomery>

Reviewed By: ipiszy

Differential Revision: D18231697

fbshipit-source-id: 7566e2438d2b57f0feaadcd51f55a03552adeab9

Author: Yinghai Lu

include/glow/Importer/CommonOperatorLoader.h

@@ -85,6 +85,11 @@ class CommonOperatorLoader : public ProtobufLoader {
       dims.push_back(in.shape[i]);
     }
 
+    // TODO(garret): create cached placeholder if in.isOffline is 1
+    if (in.isOffline) {
+      RETURN_ERR(strFormat("Cannot handle offline tensor %s yet", in.name));
+    }
+
     // Load unquantized tensor.
     if (in.quantizationParams == 0) {
       if (in.dataType == ONNXIFI_DATATYPE_FLOAT32) {

lib/Onnxifi/Base.h

@@ -127,10 +127,12 @@ class Graph {
                          EventPtr outputEvent, onnxTraceEventList *traceEvents);
 
   /// Init Glow graph based on the ONNX model \p onnxModel and
-  /// static trained weights \p weightDescriptors.
-  virtual onnxStatus
-  initGraph(const void *onnxModel, size_t onnxModelSize, uint32_t weightCount,
-            const onnxTensorDescriptorV1 *weightDescriptors) = 0;
+  /// static trained weights \p weightDescriptors. Weights can be read in later
+  /// by a \p deferedBlobReader.
+  virtual onnxStatus initGraph(const void *onnxModel, size_t onnxModelSize,
+                               uint32_t weightCount,
+                               const onnxTensorDescriptorV1 *weightDescriptors,
+                               void *deferedBlobReader) = 0;
 
   virtual onnxStatus run(std::unique_ptr<ExecutionContext> ctx,
                          EventPtr outputEvent,

lib/Onnxifi/HostManagerOnnxifi.cpp

@@ -112,10 +112,9 @@ onnxStatus HostManagerBackend::removeNetwork(const Graph *graph) {
   return ONNXIFI_STATUS_SUCCESS;
 }
 
-onnxStatus
-HostManagerGraph::initGraph(const void *onnxModel, size_t onnxModelSize,
-                            uint32_t weightCount,
-                            const onnxTensorDescriptorV1 *weightDescriptors) {
+onnxStatus HostManagerGraph::initGraph(
+    const void *onnxModel, size_t onnxModelSize, uint32_t weightCount,
+    const onnxTensorDescriptorV1 *weightDescriptors, void * /* unused */) {
 
   netName_ = strFormat("onnxifi_function_%lu", makeUniqueGraphId());
 

lib/Onnxifi/HostManagerOnnxifi.h

@@ -58,10 +58,12 @@ class HostManagerGraph : public Graph {
   static size_t makeUniqueGraphId();
 
   /// Init Glow graph based on the ONNX model \p onnxModel and
-  /// static trained weights \p weightDescriptors.
-  onnxStatus
-  initGraph(const void *onnxModel, size_t onnxModelSize, uint32_t weightCount,
-            const onnxTensorDescriptorV1 *weightDescriptors) override;
+  /// static trained weights \p weightDescriptors. Weights can be read in later
+  /// by a \p deferedBlobReader.
+  onnxStatus initGraph(const void *onnxModel, size_t onnxModelSize,
+                       uint32_t weightCount,
+                       const onnxTensorDescriptorV1 *weightDescriptors,
+                       void *deferedBlobReader) override;
 
   /// Async run HostManagerGraph with the given ExecutionContext \p ctx then
   /// signal \p outputEvent when done. \p phNameToOnnxTensorOutputs is a mapping

lib/Onnxifi/InlineOnnxifi.cpp

@@ -43,10 +43,9 @@ void computeModelHash(const void *onnxModel, size_t onnxModelSize,
 }
 } // namespace
 
-onnxStatus
-InlineGraph::initGraph(const void *onnxModel, size_t onnxModelSize,
-                       uint32_t weightCount,
-                       const onnxTensorDescriptorV1 *weightDescriptors) {
+onnxStatus InlineGraph::initGraph(
+    const void *onnxModel, size_t onnxModelSize, uint32_t weightCount,
+    const onnxTensorDescriptorV1 *weightDescriptors, void *deferedBlobReader) {
   function_ = executionEngine_.getModule().createFunction("function");
 
   std::unique_ptr<ONNXIFIModelLoader> loader =

lib/Onnxifi/InlineOnnxifi.h

@@ -31,10 +31,12 @@ class InlineGraph : public Graph {
       : Graph(backendPtr), quantizationMode_(quantizationMode) {}
 
   /// Init Glow graph based on the ONNX model \p onnxModel and
-  /// static trained weights \p weightDescriptors.
-  onnxStatus
-  initGraph(const void *onnxModel, size_t onnxModelSize, uint32_t weightCount,
-            const onnxTensorDescriptorV1 *weightDescriptors) override;
+  /// static trained weights \p weightDescriptors. Weights can be read in later
+  /// by a \p deferedBlobReader.
+  onnxStatus initGraph(const void *onnxModel, size_t onnxModelSize,
+                       uint32_t weightCount,
+                       const onnxTensorDescriptorV1 *weightDescriptors,
+                       void *deferedBlobReader) override;
 
   onnxStatus run(std::unique_ptr<ExecutionContext> ctx, EventPtr outputEvent,
                  onnxTraceEventList *traceEvents) override;

lib/Onnxifi/onnxifiGlow.cpp

@@ -401,7 +401,8 @@ EXTERNC ONNXIFI_PUBLIC ONNXIFI_CHECK_RESULT onnxStatus ONNXIFI_ABI
 GLOW_ONNXIFI_LIBRARY_FUNCTION_WRAPPER(onnxInitGraph)(
     onnxBackend backend, const uint64_t *auxPropertiesList,
     size_t onnxModelSize, const void *onnxModel, uint32_t weightsCount,
-    const onnxTensorDescriptorV1 *weightDescriptors, onnxGraph *graph) {
+    const onnxTensorDescriptorV1 *weightDescriptors, onnxGraph *graph,
+    void *deferredBlobReader) {
   if (!onnxModel || (!weightDescriptors && weightsCount) || !graph) {
     return ONNXIFI_STATUS_INVALID_POINTER;
   }
@@ -427,7 +428,7 @@ GLOW_ONNXIFI_LIBRARY_FUNCTION_WRAPPER(onnxInitGraph)(
 
   auto *glowGraph = manager.createGraph(glowBackend, quantizationMode);
   auto ret = glowGraph->initGraph(onnxModel, onnxModelSize, weightsCount,
-                                  weightDescriptors);
+                                  weightDescriptors, deferredBlobReader);
   if (ret != ONNXIFI_STATUS_SUCCESS) {
     manager.release(glowGraph);
     return ret;