Merge branch 'main' of https://github.com/microsoft/onnxruntime

microsoft · Feb 26, 2024 · ca03fdb · ca03fdb
2 parents 4b7df02 + a956893
commit ca03fdb
Show file tree

Hide file tree

Showing 61 changed files with 6,308 additions and 702 deletions.
diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
@@ -1729,14 +1729,12 @@ if(onnxruntime_BUILD_KERNEL_EXPLORER)
 endif()
 
 # When GDK_PLATFORM is set then WINAPI_FAMILY is defined in gdk_toolchain.cmake (along with other relevant flags/definitions).
-if (WIN32 AND NOT GDK_PLATFORM)
+if (WIN32 AND NOT GDK_PLATFORM AND NOT CMAKE_CROSSCOMPILING)
   if (NOT CMAKE_CXX_STANDARD_LIBRARIES MATCHES kernel32.lib)
     # On onecore, link to the onecore build of the MSVC runtime
     get_filename_component(msvc_path "${CMAKE_C_COMPILER}/../../../.." ABSOLUTE)
     link_directories(BEFORE "${msvc_path}/lib/onecore/${onnxruntime_target_platform}")
-    # The .lib files in the MSVC runtime have a DEFAULITLIB entry for onecore.lib, which in turn links to reverse forwarders.
-    # We ignore that entry and use onecore_apiset.lib instead, since system components must not rely on reverse forwarders.
-    add_link_options("/NODEFAULTLIB:onecore.lib")
+    # The .lib files in the MSVC runtime have a DEFAULITLIB entry for onecore.lib, but it shold not cause any conflict with onecoreuap.lib
   endif()
 endif()
 

diff --git a/cmake/wcos_rules_override.cmake b/cmake/wcos_rules_override.cmake
@@ -1,2 +1,2 @@
-set(CMAKE_C_STANDARD_LIBRARIES_INIT onecoreuap_apiset.lib)
-set(CMAKE_CXX_STANDARD_LIBRARIES_INIT onecoreuap_apiset.lib)
+set(CMAKE_C_STANDARD_LIBRARIES_INIT onecoreuap.lib)
+set(CMAKE_CXX_STANDARD_LIBRARIES_INIT onecoreuap.lib)
diff --git a/js/web/lib/wasm/jsep/webgpu/ops/concat.ts b/js/web/lib/wasm/jsep/webgpu/ops/concat.ts
@@ -154,7 +154,9 @@ const createConcatProgramInfo = (inputs: readonly TensorView[], axis: number): P
 
 export const concat = (context: ComputeContext, attributes: ConcatAttributes): void => {
   validateInputs(context.inputs);
-  context.compute(createConcatProgramInfo(context.inputs, attributes.axis));
+  // 0 length tensors are valid for concat, remove them
+  const nonEmptyInputs = context.inputs.filter(input => ShapeUtil.size(input.dims) > 0);
+  context.compute(createConcatProgramInfo(nonEmptyInputs, attributes.axis), {inputs: nonEmptyInputs});
 };
 
 export const parseConcatAttributes = (attributes: Record<string, unknown>): ConcatAttributes =>

diff --git a/js/web/lib/wasm/jsep/webgpu/ops/gather.ts b/js/web/lib/wasm/jsep/webgpu/ops/gather.ts
@@ -55,7 +55,7 @@ const createGatherProgramInfo = (inputs: readonly TensorView[], attributes: Gath
           if (idx${x} < 0) {
             idx${x} = idx${x} + uniforms.axisDimLimit;
           }
-          var dataIndices${x} = ${data.type.indices}(0);
+          var dataIndices${x} : ${data.type.indices};
         `;
       for (let i = 0, j = 0; i < inputRank; i++) {
         if (i === axis) {

diff --git a/js/web/test/data/ops/add.jsonc b/js/web/test/data/ops/add.jsonc
@@ -157,28 +157,6 @@
             "type": "float32"
           }
         ]
-      },
-      {
-        "name": "T[2,0] T[2,1]",
-        "inputs": [
-          {
-            "data": [],
-            "dims": [2, 0],
-            "type": "float32"
-          },
-          {
-            "data": [1, 2],
-            "dims": [2, 1],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [],
-            "dims": [2, 0],
-            "type": "float32"
-          }
-        ]
       }
     ]
   }

diff --git a/js/web/test/data/ops/add_zero-sized.jsonc b/js/web/test/data/ops/add_zero-sized.jsonc
@@ -0,0 +1,31 @@
+[
+  {
+    "name": "Add with no attributes",
+    "operator": "Add",
+    "attributes": [],
+    "cases": [
+      {
+        "name": "T[2,0] T[2,1]",
+        "inputs": [
+          {
+            "data": [],
+            "dims": [2, 0],
+            "type": "float32"
+          },
+          {
+            "data": [1, 2],
+            "dims": [2, 1],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [],
+            "dims": [2, 0],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
+  }
+]
diff --git a/js/web/test/data/ops/concat_zero-sized.jsonc b/js/web/test/data/ops/concat_zero-sized.jsonc
diff --git a/js/web/test/suite-test-list.jsonc b/js/web/test/suite-test-list.jsonc
@@ -1334,6 +1334,7 @@
       "acos.jsonc",
       "add.jsonc",
       "add_int32.jsonc",
+      "add_zero-sized.jsonc",
       //"and.jsonc",
       "asin.jsonc",
       "attention.jsonc",
@@ -1343,6 +1344,7 @@
       "ceil.jsonc",
       "concat.jsonc",
       "concat_int32.jsonc",
+      "concat_zero-sized.jsonc",
       "cast.jsonc",
       "conv.jsonc",
       "cos.jsonc",

diff --git a/onnxruntime/core/optimizer/gemm_activation_fusion.cc b/onnxruntime/core/optimizer/gemm_activation_fusion.cc
@@ -56,6 +56,13 @@ Status GemmActivationFusion::ApplyImpl(Graph& graph, bool& modified, int graph_l
       continue;
     }
 
+    NodeArg* node_output = node.MutableOutputDefs()[0];
+    auto data_type = node_output->TypeAsProto()->tensor_type().elem_type();
+    if (data_type != ONNX_NAMESPACE::TensorProto_DataType_FLOAT) {
+      // FusedGemm is only registered for float data type in fused_gemm.cc!
+      continue;
+    }
+
     const Node& next_node = *(node.OutputNodesBegin());
     if (!IsFusableActivation(next_node) || next_node.GetExecutionProviderType() != node.GetExecutionProviderType()) {
       continue;

diff --git a/onnxruntime/core/platform/windows/env.cc b/onnxruntime/core/platform/windows/env.cc
@@ -459,8 +459,8 @@ Status WindowsEnv::MapFileIntoMemory(_In_z_ const ORTCHAR_T* file_path,
 
   void* const mapped_base = MapViewOfFile(file_mapping_handle.get(),
                                           FILE_MAP_READ,
-                                          0,
-                                          static_cast<DWORD>(mapped_offset),
+                                          static_cast<DWORD>((mapped_offset >> 32) & 0xFFFFFFFF),
+                                          static_cast<DWORD>(mapped_offset & 0xFFFFFFFF),
                                           mapped_length);
   GSL_SUPPRESS(r.11)
   mapped_memory =

diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/inc/IWinmlExecutionProvider.h b/onnxruntime/core/providers/dml/DmlExecutionProvider/inc/IWinmlExecutionProvider.h
@@ -80,15 +80,10 @@ namespace Windows::AI::MachineLearning::Adapter
     };
 
     // This is the counterpart to the MLOperatorGraphDesc ABI struct which owns its memory and uses containers.
-    // Either nodesAsOperatorDesc or nodesAsIDMLOperator can have non-zero size.
     struct DmlGraphNodeCreateInfo
     {
         uint32_t nodeCount = 0;
-        std::vector<std::unique_ptr<AbstractOperatorDesc>> nodesAsOperatorDesc;
-
-        // TODO (jeffbloo): Remove this
-        std::vector<Microsoft::WRL::ComPtr<IDMLOperator>> nodesAsIDMLOperator;
-
+        std::vector<std::unique_ptr<AbstractOperatorDesc>> nodes;
         std::vector<DML_INPUT_GRAPH_EDGE_DESC> inputEdges;
         std::vector<DML_OUTPUT_GRAPH_EDGE_DESC> outputEdges;
         std::vector<DML_INTERMEDIATE_GRAPH_EDGE_DESC> intermediateEdges;