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

…baijumeswani/fix-pipelines
microsoft · Mar 25, 2024 · 4d15d36 · 4d15d36
2 parents f3bd524 + 7d976cf
commit 4d15d36
Show file tree

Hide file tree

Showing 37 changed files with 3,938 additions and 599 deletions.
diff --git a/cmake/patches/abseil/absl_windows.patch b/cmake/patches/abseil/absl_windows.patch
@@ -25,17 +25,91 @@ index a6efc98e..8c4de8e7 100644
      "/wd4800",
  ]
 diff --git a/absl/copts/copts.py b/absl/copts/copts.py
-index 0d6c1ec3..75fd935f 100644
+index e6e11949..0aa7d868 100644
 --- a/absl/copts/copts.py
 +++ b/absl/copts/copts.py
-@@ -132,10 +132,6 @@ COPT_VARS = {
-             "/wd4068",  # unknown pragma
-             # qualifier applied to function type has no meaning; ignored
-             "/wd4180",
--            # conversion from 'type1' to 'type2', possible loss of data
--            "/wd4244",
--            # conversion from 'size_t' to 'type', possible loss of data
--            "/wd4267",
-             # The decorated name was longer than the compiler limit
-             "/wd4503",
-             # forcing value to bool 'true' or 'false' (performance warning)
+@@ -115,10 +115,6 @@ MSVC_WARNING_FLAGS = [
+     "/wd4068",  # unknown pragma
+     # qualifier applied to function type has no meaning; ignored
+     "/wd4180",
+-    # conversion from 'type1' to 'type2', possible loss of data
+-    "/wd4244",
+-    # conversion from 'size_t' to 'type', possible loss of data
+-    "/wd4267",
+     # The decorated name was longer than the compiler limit
+     "/wd4503",
+     # forcing value to bool 'true' or 'false' (performance warning)
+diff --git a/absl/debugging/symbolize_win32.inc b/absl/debugging/symbolize_win32.inc
+index 53a099a1..34d210d6 100644
+--- a/absl/debugging/symbolize_win32.inc
++++ b/absl/debugging/symbolize_win32.inc
+@@ -35,15 +35,15 @@ ABSL_NAMESPACE_BEGIN
+
+ static HANDLE process = NULL;
+
+-void InitializeSymbolizer(const char*) {
+-  if (process != nullptr) {
+-    return;
+-  }
++namespace {
++void InitializeSymbolizerImpl() {
++
+   process = GetCurrentProcess();
+
+   // Symbols are not loaded until a reference is made requiring the
+   // symbols be loaded. This is the fastest, most efficient way to use
+   // the symbol handler.
++
+   SymSetOptions(SYMOPT_DEFERRED_LOADS | SYMOPT_UNDNAME);
+   if (!SymInitialize(process, nullptr, true)) {
+     // GetLastError() returns a Win32 DWORD, but we assign to
+@@ -54,6 +54,36 @@ void InitializeSymbolizer(const char*) {
+   }
+ }
+
++bool LookupAndInitialize(const void* pc, SYMBOL_INFO* symbol) {
++  auto hProcess = (process != NULL) ? process : GetCurrentProcess();
++  if (SymFromAddr(hProcess, reinterpret_cast<DWORD64>(pc), nullptr, symbol) != TRUE) {
++    if (GetLastError() == ERROR_INVALID_HANDLE && process == NULL) {
++      InitializeSymbolizerImpl();
++      if (SymFromAddr(process, reinterpret_cast<DWORD64>(pc), nullptr, symbol) != TRUE) {
++        return false;
++      }
++    } else {
++      return false;
++    }
++    return false;
++  }
++  return true;
++}
++}
++
++void InitializeSymbolizer(const char*) {
++  if (process != nullptr) {
++    return;
++  }
++
++  alignas(SYMBOL_INFO) char buf[sizeof(SYMBOL_INFO) + MAX_SYM_NAME];
++  SYMBOL_INFO* symbol = reinterpret_cast<SYMBOL_INFO*>(buf);
++  symbol->SizeOfStruct = sizeof(SYMBOL_INFO);
++  symbol->MaxNameLen = MAX_SYM_NAME;
++
++  static_cast<void>(LookupAndInitialize(reinterpret_cast<const void*>(&InitializeSymbolizer), symbol));
++}
++
+ bool Symbolize(const void* pc, char* out, int out_size) {
+   if (out_size <= 0) {
+     return false;
+@@ -62,9 +92,11 @@ bool Symbolize(const void* pc, char* out, int out_size) {
+   SYMBOL_INFO* symbol = reinterpret_cast<SYMBOL_INFO*>(buf);
+   symbol->SizeOfStruct = sizeof(SYMBOL_INFO);
+   symbol->MaxNameLen = MAX_SYM_NAME;
+-  if (!SymFromAddr(process, reinterpret_cast<DWORD64>(pc), nullptr, symbol)) {
++
++  if(!LookupAndInitialize(pc, symbol)) {
+     return false;
+   }
++
+   const size_t out_size_t = static_cast<size_t>(out_size);
+   strncpy(out, symbol->Name, out_size_t);
+   if (out[out_size_t - 1] != '\0') {
diff --git a/js/web/lib/wasm/jsep/webgpu/ops/3rd-party/conv_backprop_mm_webgpu.ts b/js/web/lib/wasm/jsep/webgpu/ops/3rd-party/conv_backprop_mm_webgpu.ts
@@ -164,17 +164,14 @@ export const createConv2DTransposeMatMulProgramInfo =
       const outWidth = isChannelsLast ? outputShape[2] : outputShape[3];
       const outHeight = isChannelsLast ? outputShape[1] : outputShape[2];
       const outChannels = isChannelsLast ? outputShape[3] : outputShape[1];
-      const isVec4 =
-          isChannelsLast ? inChannels % 4 === 0 && outChannels % 4 === 0 : outWidth % 4 === 0 && outChannels % 4 === 0;
+      // TODO: enable vec4 for NCHW
+      const isVec4 = isChannelsLast && (inChannels % 4 === 0 && inChannels % 3) && outChannels % 4 === 0;
 
       // TODO: fine tune size
       const dispatchX = isChannelsLast ? outChannels : outWidth * outHeight;
       const dispatchY = isChannelsLast ? outWidth * outHeight : outChannels;
-      const workGroupSize: [number, number, number] = isVec4 ?
-          [8, 8, 1] :
-          [(dispatchX <= 4 || dispatchY <= 4) ? 4 : 16, dispatchX > 4 && dispatchY <= 4 ? 4 : 16, 1];
-      const elementsPerThread =
-          isVec4 ? [4, 4, 1] : [dispatchX <= 4 ? 1 : 4, dispatchX > 4 && dispatchY <= 4 ? 1 : 4, 1];
+      const workGroupSize: [number, number, number] = [8, 8, 1];
+      const elementsPerThread = dimAOuter <= 8 ? [4, 1, 1] : [4, 4, 1];
       const dispatch = [
         Math.ceil(dispatchX / workGroupSize[0] / elementsPerThread[0]),
         Math.ceil(dispatchY / workGroupSize[1] / elementsPerThread[1]),