Cherry-pick release:1.23.0 PRs to rel-1.23.0 (microsoft#25985)

This PR cherry-picks the following PRs to the rel-1.23.0 branch:

* microsoft#25938
* microsoft#25957
* microsoft#25960
* microsoft#25968
* microsoft#25971

---------

Co-authored-by: Dmitri Smirnov <[email protected]>
Co-authored-by: Adrian Lizarraga <[email protected]>
Co-authored-by: Hariharan Seshadri <[email protected]>

Author: 4 people

.github/workflows/android.yml

@@ -78,8 +78,8 @@ jobs:
       run: |
         set -e -x
         BINARY_SIZE_THRESHOLD_ARGS=""
-        echo "Binary size threshold in bytes: 1306224"
-        BINARY_SIZE_THRESHOLD_ARGS="--threshold_size_in_bytes 1306224"
+        echo "Binary size threshold in bytes: 1436672"
+        BINARY_SIZE_THRESHOLD_ARGS="--threshold_size_in_bytes 1436672"
 
         # Ensure ANDROID_NDK_HOME is available and get its real path
         if [ -z "$ANDROID_NDK_HOME" ]; then

include/onnxruntime/core/framework/ortmemoryinfo.h

@@ -13,18 +13,14 @@ struct OrtMemoryInfo {
   OrtMemoryInfo() = default;  // to allow default construction of Tensor
 
   // use string for name, so we could have customized allocator in execution provider.
-  const char* name = nullptr;
+  std::string name;
   OrtMemType mem_type = OrtMemTypeDefault;
   OrtAllocatorType alloc_type = OrtInvalidAllocator;
   OrtDevice device;
 
-  constexpr OrtMemoryInfo(const char* name_, OrtAllocatorType type_, OrtDevice device_ = OrtDevice(),
-                          OrtMemType mem_type_ = OrtMemTypeDefault)
-#if ((defined(__GNUC__) && __GNUC__ > 4) || defined(__clang__))
-      // this causes a spurious error in CentOS gcc 4.8 build so disable if GCC version < 5
-      __attribute__((nonnull))
-#endif
-      : name(name_),
+  OrtMemoryInfo(std::string name_, OrtAllocatorType type_, OrtDevice device_ = OrtDevice(),
+                OrtMemType mem_type_ = OrtMemTypeDefault)
+      : name(std::move(name_)),
         mem_type(mem_type_),
         alloc_type(type_),
         device(device_) {
@@ -39,7 +35,7 @@ struct OrtMemoryInfo {
     if (device != other.device)
       return device < other.device;
 
-    return strcmp(name, other.name) < 0;
+    return name < other.name;
   }
 
   // This is to make OrtMemoryInfo a valid key in hash tables
@@ -68,7 +64,7 @@ inline bool operator==(const OrtMemoryInfo& left, const OrtMemoryInfo& other) {
   return left.mem_type == other.mem_type &&
          left.alloc_type == other.alloc_type &&
          left.device == other.device &&
-         strcmp(left.name, other.name) == 0;
+         left.name == other.name;
 }
 
 inline bool operator!=(const OrtMemoryInfo& lhs, const OrtMemoryInfo& rhs) { return !(lhs == rhs); }

onnxruntime/core/framework/allocator.cc

@@ -6,6 +6,7 @@
 #include "core/common/safeint.h"
 #include "core/common/status.h"
 #include "core/framework/allocator.h"
+#include "core/framework/error_code_helper.h"
 #include "core/mlas/inc/mlas.h"
 #include "core/framework/utils.h"
 #include "core/session/ort_apis.h"
@@ -185,22 +186,32 @@ std::ostream& operator<<(std::ostream& out, const OrtMemoryInfo& info) { return
 #endif
 ORT_API_STATUS_IMPL(OrtApis::CreateMemoryInfo, _In_ const char* name1, enum OrtAllocatorType type, int id1,
                     enum OrtMemType mem_type1, _Outptr_ OrtMemoryInfo** out) {
+  API_IMPL_BEGIN
+
+  if (name1 == nullptr) {
+    return OrtApis::CreateStatus(ORT_INVALID_ARGUMENT, "MemoryInfo name cannot be null.");
+  }
+
+  if (out == nullptr) {
+    return OrtApis::CreateStatus(ORT_INVALID_ARGUMENT, "Output memory info cannot be null.");
+  }
+
   auto device_id = static_cast<OrtDevice::DeviceId>(id1);
   if (strcmp(name1, onnxruntime::CPU) == 0) {
     *out = new OrtMemoryInfo(onnxruntime::CPU, type, OrtDevice(), mem_type1);
   } else if (strcmp(name1, onnxruntime::CUDA) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::CUDA, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::NVIDIA, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::OpenVINO_GPU) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::OpenVINO_GPU, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::INTEL, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::HIP) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::HIP, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::AMD, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::WEBGPU_BUFFER) == 0 ||
@@ -212,45 +223,56 @@ ORT_API_STATUS_IMPL(OrtApis::CreateMemoryInfo, _In_ const char* name1, enum OrtA
 
   } else if (strcmp(name1, onnxruntime::DML) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::DML, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::MICROSOFT, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::OpenVINO_RT_NPU) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::OpenVINO_RT_NPU, type,
         OrtDevice(OrtDevice::NPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::INTEL, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::CUDA_PINNED) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::CUDA_PINNED, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::NVIDIA, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::HIP_PINNED) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::HIP_PINNED, type,
         OrtDevice(OrtDevice::GPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::AMD, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::QNN_HTP_SHARED) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::QNN_HTP_SHARED, type,
         OrtDevice(OrtDevice::CPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::QUALCOMM, device_id),
         mem_type1);
   } else if (strcmp(name1, onnxruntime::CPU_ALIGNED_4K) == 0) {
     *out = new OrtMemoryInfo(
-        name1, type,
+        onnxruntime::CPU_ALIGNED_4K, type,
         OrtDevice(OrtDevice::CPU, OrtDevice::MemType::DEFAULT, OrtDevice::VendorIds::NONE, device_id,
                   onnxruntime::kAlloc4KAlignment),
         mem_type1);
   } else {
     return OrtApis::CreateStatus(ORT_INVALID_ARGUMENT, "Specified device is not supported. Try CreateMemoryInfo_V2.");
   }
+  API_IMPL_END
   return nullptr;
 }
 
 ORT_API_STATUS_IMPL(OrtApis::CreateMemoryInfo_V2, _In_ const char* name, _In_ enum OrtMemoryInfoDeviceType device_type,
                     _In_ uint32_t vendor_id, _In_ int32_t device_id, _In_ enum OrtDeviceMemoryType mem_type,
                     _In_ size_t alignment, enum OrtAllocatorType type,
                     _Outptr_ OrtMemoryInfo** out) {
+  API_IMPL_BEGIN
+
+  if (name == nullptr) {
+    return OrtApis::CreateStatus(ORT_INVALID_ARGUMENT, "MemoryInfo name cannot be null.");
+  }
+
+  if (out == nullptr) {
+    return OrtApis::CreateStatus(ORT_INVALID_ARGUMENT, "Output memory info cannot be null.");
+  }
+
   // map the public enum values to internal OrtDevice values
   OrtDevice::MemoryType mt = mem_type == OrtDeviceMemoryType_DEFAULT ? OrtDevice::MemType::DEFAULT
                                                                      : OrtDevice::MemType::HOST_ACCESSIBLE;
@@ -275,6 +297,7 @@ ORT_API_STATUS_IMPL(OrtApis::CreateMemoryInfo_V2, _In_ const char* name, _In_ en
 
   *out = new OrtMemoryInfo(name, type, OrtDevice{dt, mt, vendor_id, narrow<int16_t>(device_id), alignment},
                            mem_type == OrtDeviceMemoryType_DEFAULT ? OrtMemTypeDefault : OrtMemTypeCPU);
+  API_IMPL_END
   return nullptr;
 }
 
@@ -283,7 +306,7 @@ ORT_API(void, OrtApis::ReleaseMemoryInfo, _Frees_ptr_opt_ OrtMemoryInfo* p) { de
 #pragma warning(pop)
 #endif
 ORT_API_STATUS_IMPL(OrtApis::MemoryInfoGetName, _In_ const OrtMemoryInfo* ptr, _Out_ const char** out) {
-  *out = ptr->name;
+  *out = ptr->name.c_str();
   return nullptr;
 }
 

onnxruntime/core/framework/bfc_arena.cc

@@ -13,7 +13,7 @@ BFCArena::BFCArena(std::unique_ptr<IAllocator> resource_allocator,
                    int max_dead_bytes_per_chunk,
                    int initial_growth_chunk_size_bytes,
                    int64_t max_power_of_two_extend_bytes)
-    : IAllocator(OrtMemoryInfo(resource_allocator->Info().name,
+    : IAllocator(OrtMemoryInfo(resource_allocator->Info().name.c_str(),
                                OrtAllocatorType::OrtArenaAllocator,
                                resource_allocator->Info().device,
                                resource_allocator->Info().mem_type)),

onnxruntime/core/mlas/lib/qnbitgemm.h

@@ -53,16 +53,25 @@ struct PackedQuantBDataStruct {
     {
         const size_t PackedQuantBDataSize = N * BlockCountK * MlasQNBitBlkDataSizeInBytes(BlkBitWidth, BlkLen);
         size_t BlkSumSize = MlasDivRoundup(N, 16) * BlockCountK * 16 * sizeof(T);
-        if constexpr (BlkBitWidth == 8) {
-            PackedQuantBData = (std::byte*)MlasAlignAddress(PackedQuantBWorkspace, 32);
-        } else {
 #if defined(MLAS_TARGET_AMD64_IX86)
         // avx512 requires alignment on a 64-byte boundary
         PackedQuantBData = (std::byte*)MlasAlignAddress(PackedQuantBWorkspace, 64);
+#elif defined (MLAS_TARGET_ARM64)
+        // Only for 8-bit Gemms is the `PackedQuantBData` is to be 32-byte aligned and
+        // there is enough memory allocated to support this alignment.
+        // See QNBitGemmPackQuantBDataSize().
+        // When bit width is 4, there is no alignment guarantee.
+        // TODO(hasesh): Can we unify the alignment for 4-bit and 8-bit ARM64 Gemms so as to
+        // simpify this logic and make code here cleaner ?
+        if constexpr (BlkBitWidth == 8) {
+            PackedQuantBData = (std::byte*)MlasAlignAddress(PackedQuantBWorkspace, 32);        
+        }
+        else {
+            PackedQuantBData = (std::byte*)PackedQuantBWorkspace;
+        }
 #else
         PackedQuantBData = (std::byte*)PackedQuantBWorkspace;
 #endif
-        }
 
         QuantBBlkSum = (T*)(PackedQuantBData + PackedQuantBDataSize);
         QuantBBlkSum = (T*)MlasAlignAddress(QuantBBlkSum, MlasQNBitQuantBBlkSumAlignment());

onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.cpp

@@ -927,7 +927,7 @@ namespace Dml
 
     bool IsGpuTensor(const onnxruntime::Tensor& tensor)
     {
-        return strcmp(tensor.Location().name, onnxruntime::CPU) &&
+        return strcmp(tensor.Location().name.c_str(), onnxruntime::CPU) &&
             !(tensor.Location().mem_type == ::OrtMemType::OrtMemTypeCPUOutput || tensor.Location().mem_type == ::OrtMemType::OrtMemTypeCPUInput);
     }
 

onnxruntime/core/providers/dml/DmlExecutionProvider/src/MLOperatorAuthorImpl.cpp

@@ -98,7 +98,7 @@ namespace Windows::AI::MachineLearning::Adapter
 
     bool IsAllocationInterface(const ::OrtMemoryInfo& info)
     {
-        return strcmp(info.name, onnxruntime::CPU) && !(info.mem_type == ::OrtMemType::OrtMemTypeCPUOutput || info.mem_type == ::OrtMemType::OrtMemTypeCPUInput);
+        return strcmp(info.name.c_str(), onnxruntime::CPU) && !(info.mem_type == ::OrtMemType::OrtMemTypeCPUOutput || info.mem_type == ::OrtMemType::OrtMemTypeCPUInput);
     }
 
     // Translate the data object stored in a tensor to the type which will be returned through
@@ -1774,7 +1774,9 @@ namespace Windows::AI::MachineLearning::Adapter
         }
 
         // tells caller whether this tensor is in CPU memory
-        return !strcmp(m_impl->Location().name, onnxruntime::CPU) || m_impl->Location().mem_type == ::OrtMemType::OrtMemTypeCPUOutput || m_impl->Location().mem_type == ::OrtMemType::OrtMemTypeCPUInput;
+        return !strcmp(m_impl->Location().name.c_str(), onnxruntime::CPU) 
+            || m_impl->Location().mem_type == ::OrtMemType::OrtMemTypeCPUOutput
+            || m_impl->Location().mem_type == ::OrtMemType::OrtMemTypeCPUInput;
     }
 
     bool STDMETHODCALLTYPE TensorWrapper::IsDataInterface() const noexcept

onnxruntime/core/providers/qnn/qnn_provider_factory.cc

@@ -219,9 +219,11 @@ struct QnnEpFactory : OrtEpFactory {
         OrtKeyValuePairs* ep_options = nullptr;
         factory->ort_api.CreateKeyValuePairs(&ep_options);
         factory->ort_api.AddKeyValuePair(ep_options, "backend_path", factory->qnn_backend_path.c_str());
-        ORT_API_RETURN_IF_ERROR(
-            factory->ort_api.GetEpApi()->CreateEpDevice(factory, &device, nullptr, ep_options,
-                                                        &ep_devices[num_ep_devices++]));
+        OrtStatus* status = factory->ort_api.GetEpApi()->CreateEpDevice(factory, &device, nullptr, ep_options,
+                                                                        &ep_devices[num_ep_devices++]);
+
+        factory->ort_api.ReleaseKeyValuePairs(ep_options);
+        ORT_API_RETURN_IF_ERROR(status);
       }
     }
 

onnxruntime/core/providers/webgpu/webgpu_context.cc

@@ -210,7 +210,7 @@ Status WebGpuContext::Run(ComputeContext& context, const ProgramBase& program) {
                   return tensor != nullptr &&
                          tensor->Location().mem_type == OrtMemType::OrtMemTypeDefault &&
                          tensor->Location().device.Type() == OrtDevice::GPU &&
-                         !strcmp(tensor->Location().name, WEBGPU_BUFFER);
+                         !strcmp(tensor->Location().name.c_str(), WEBGPU_BUFFER);
                 }),
                 "All inputs must be tensors on WebGPU buffers.");
 
@@ -219,7 +219,7 @@ Status WebGpuContext::Run(ComputeContext& context, const ProgramBase& program) {
                   return tensor != nullptr &&
                          tensor->Location().mem_type == OrtMemType::OrtMemTypeDefault &&
                          tensor->Location().device.Type() == OrtDevice::GPU &&
-                         !strcmp(tensor->Location().name, WEBGPU_BUFFER);
+                         !strcmp(tensor->Location().name.c_str(), WEBGPU_BUFFER);
                 }),
                 "All outputs must be tensors on WebGPU buffers.");
   }

onnxruntime/core/session/environment.cc

@@ -79,7 +79,7 @@ static bool AreOrtMemoryInfosEquivalent(
     bool ignore_alignment = false) {
   return left.mem_type == right.mem_type &&
          (ignore_alignment ? left.device.EqualIgnoringAlignment(right.device) : left.device == right.device) &&
-         (!match_name || strcmp(left.name, right.name) == 0);
+         (!match_name || left.name == right.name);
 }
 
 std::vector<AllocatorPtr>::const_iterator FindExistingAllocator(const std::vector<AllocatorPtr>& allocators,