[llvm] [Offload] Add device UID (PR #164391)
Robert Imschweiler via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 21 12:32:58 PDT 2025
https://github.com/ro-i updated https://github.com/llvm/llvm-project/pull/164391
>From bb3ff9c7a1efbdea0008c2d0b11b42e05162cdd9 Mon Sep 17 00:00:00 2001
From: Robert Imschweiler <robert.imschweiler at amd.com>
Date: Mon, 20 Oct 2025 03:06:58 -0500
Subject: [PATCH 1/3] [Offload] Add device UID
Introduced in OpenMP 6.0, the device UID shall be a unique identifier of
a device on a given system. (Not necessarily a UUID.)
Since it is not guaranteed that the (U)UIDs defined by the device vendor
libraries, such as HSA, do not overlap with those of other vendors, the
device UIDs in offload are always combined with the offload plugin name.
In case the vendor library does not specify any device UID for a given
device, we fall back to the offload-internal device ID.
Note that the "fake" multiple host devices do *not* get different UIDs
because they are effectively the same device.
The device UID can be retrieved using the `llvm-offload-device-info`
tool.
---
offload/liboffload/API/Device.td | 1 +
offload/liboffload/src/OffloadImpl.cpp | 7 +++++--
.../amdgpu/dynamic_hsa/hsa_ext_amd.h | 1 +
offload/plugins-nextgen/amdgpu/src/rtl.cpp | 14 +++++++++++++
.../common/include/PluginInterface.h | 15 +++++++++++++-
.../common/src/PluginInterface.cpp | 14 +++++++++++--
offload/plugins-nextgen/host/src/rtl.cpp | 4 +++-
.../deviceinfo/llvm-offload-device-info.cpp | 1 +
.../OffloadAPI/device/olGetDeviceInfo.cpp | 20 +++++++++++++++++++
.../OffloadAPI/device/olGetDeviceInfoSize.cpp | 1 +
10 files changed, 72 insertions(+), 6 deletions(-)
diff --git a/offload/liboffload/API/Device.td b/offload/liboffload/API/Device.td
index 5b54c79d83f9d..e9c154818c4a1 100644
--- a/offload/liboffload/API/Device.td
+++ b/offload/liboffload/API/Device.td
@@ -29,6 +29,7 @@ def ol_device_info_t : Enum {
TaggedEtor<"PLATFORM", "ol_platform_handle_t", "the platform associated with the device">,
TaggedEtor<"NAME", "char[]", "Device name">,
TaggedEtor<"PRODUCT_NAME", "char[]", "Device user-facing marketing name">,
+ TaggedEtor<"UID", "char[]", "Device UID">,
TaggedEtor<"VENDOR", "char[]", "Device vendor">,
TaggedEtor<"DRIVER_VERSION", "char[]", "Driver version">,
TaggedEtor<"MAX_WORK_GROUP_SIZE", "uint32_t", "Maximum total work group size in work items">,
diff --git a/offload/liboffload/src/OffloadImpl.cpp b/offload/liboffload/src/OffloadImpl.cpp
index 6d22faeb0e57e..ed306470a0d76 100644
--- a/offload/liboffload/src/OffloadImpl.cpp
+++ b/offload/liboffload/src/OffloadImpl.cpp
@@ -147,8 +147,8 @@ llvm::Error ol_platform_impl_t::init() {
if (llvm::Error Err = Plugin->initDevice(Id))
return Err;
- auto Device = &Plugin->getDevice(Id);
- auto Info = Device->obtainInfoImpl();
+ GenericDeviceTy *Device = &Plugin->getDevice(Id);
+ llvm::Expected<InfoTreeNode> Info = Device->obtainInfo();
if (llvm::Error Err = Info.takeError())
return Err;
Devices.emplace_back(std::make_unique<ol_device_impl_t>(Id, Device, *this,
@@ -467,6 +467,7 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
switch (PropName) {
case OL_DEVICE_INFO_NAME:
case OL_DEVICE_INFO_PRODUCT_NAME:
+ case OL_DEVICE_INFO_UID:
case OL_DEVICE_INFO_VENDOR:
case OL_DEVICE_INFO_DRIVER_VERSION: {
// String values
@@ -544,6 +545,8 @@ Error olGetDeviceInfoImplDetailHost(ol_device_handle_t Device,
return Info.writeString("Virtual Host Device");
case OL_DEVICE_INFO_PRODUCT_NAME:
return Info.writeString("Virtual Host Device");
+ case OL_DEVICE_INFO_UID:
+ return Info.writeString(GenericDeviceTy::getHostDeviceUid());
case OL_DEVICE_INFO_VENDOR:
return Info.writeString("Liboffload");
case OL_DEVICE_INFO_DRIVER_VERSION:
diff --git a/offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa_ext_amd.h b/offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa_ext_amd.h
index 29cfe78082dbb..ddfa65c76cf2d 100644
--- a/offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa_ext_amd.h
+++ b/offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa_ext_amd.h
@@ -72,6 +72,7 @@ typedef enum hsa_amd_agent_info_s {
HSA_AMD_AGENT_INFO_MAX_WAVES_PER_CU = 0xA00A,
HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU = 0xA00B,
HSA_AMD_AGENT_INFO_COOPERATIVE_QUEUES = 0xA010,
+ HSA_AMD_AGENT_INFO_UUID = 0xA011,
HSA_AMD_AGENT_INFO_TIMESTAMP_FREQUENCY = 0xA016,
} hsa_amd_agent_info_t;
diff --git a/offload/plugins-nextgen/amdgpu/src/rtl.cpp b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
index a7723b8598815..bcde829923b0d 100644
--- a/offload/plugins-nextgen/amdgpu/src/rtl.cpp
+++ b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
@@ -2069,6 +2069,20 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
return Err;
ComputeUnitKind = GPUName;
+ // From the ROCm HSA documentation:
+ // Query the UUID of the agent. The value is an Ascii string with a maximum
+ // of 21 chars including NUL. The string value consists of two parts: header
+ // and body. The header identifies the device type (GPU, CPU, DSP) while the
+ // body encodes the UUID as a 16 digit hex string.
+ //
+ // Agents that do not support UUID will return the string "GPU-XX" or
+ // "CPU-XX" or "DSP-XX" depending on their device type.
+ char Uuid[24];
+ if (auto Err = getDeviceAttr(HSA_AMD_AGENT_INFO_UUID, Uuid))
+ return Err;
+ if (strcmp(Uuid + 3, "-XX") != 0)
+ DeviceUid = std::string(Plugin.getName()) + "-" + std::string(Uuid);
+
// Get the wavefront size.
uint32_t WavefrontSize = 0;
if (auto Err = getDeviceAttr(HSA_AGENT_INFO_WAVEFRONT_SIZE, WavefrontSize))
diff --git a/offload/plugins-nextgen/common/include/PluginInterface.h b/offload/plugins-nextgen/common/include/PluginInterface.h
index 8c530bba3882c..6fe90d6ace1a0 100644
--- a/offload/plugins-nextgen/common/include/PluginInterface.h
+++ b/offload/plugins-nextgen/common/include/PluginInterface.h
@@ -791,6 +791,10 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
/// this id is not unique between different plugins; they may overlap.
int32_t getDeviceId() const { return DeviceId; }
+ /// Get the unique identifier of the device.
+ const char *getDeviceUid() const { return DeviceUid.c_str(); }
+ static constexpr const char *getHostDeviceUid() { return "HOST"; }
+
/// Set the context of the device if needed, before calling device-specific
/// functions. Plugins may implement this function as a no-op if not needed.
virtual Error setContext() = 0;
@@ -989,9 +993,12 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
Error syncEvent(void *EventPtr);
virtual Error syncEventImpl(void *EventPtr) = 0;
+ /// Obtain information about the device.
+ Expected<InfoTreeNode> obtainInfo();
+ virtual Expected<InfoTreeNode> obtainInfoImpl() = 0;
+
/// Print information about the device.
Error printInfo();
- virtual Expected<InfoTreeNode> obtainInfoImpl() = 0;
/// Return true if the device has work that is either queued or currently
/// running
@@ -1196,6 +1203,12 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
/// global device id and is not the device id visible to the OpenMP user.
const int32_t DeviceId;
+ /// The unique identifier of the device.
+ /// Per default, the unique identifier of the device is set to the device id,
+ /// combined with the plugin name, since the offload device id may overlap
+ /// between different plugins.
+ std::string DeviceUid;
+
/// The default grid values used for this device.
llvm::omp::GV GridValues;
diff --git a/offload/plugins-nextgen/common/src/PluginInterface.cpp b/offload/plugins-nextgen/common/src/PluginInterface.cpp
index db43cbe49cc2b..cad91668c5e0b 100644
--- a/offload/plugins-nextgen/common/src/PluginInterface.cpp
+++ b/offload/plugins-nextgen/common/src/PluginInterface.cpp
@@ -715,6 +715,9 @@ GenericDeviceTy::GenericDeviceTy(GenericPluginTy &Plugin, int32_t DeviceId,
DeviceId(DeviceId), GridValues(OMPGridValues),
PeerAccesses(NumDevices, PeerAccessState::PENDING), PeerAccessesLock(),
PinnedAllocs(*this), RPCServer(nullptr) {
+ DeviceUid = std::string(Plugin.getName()) + "-" +
+ std::to_string(static_cast<uint64_t>(DeviceId));
+
#ifdef OMPT_SUPPORT
OmptInitialized.store(false);
// Bind the callbacks to this device's member functions
@@ -1524,11 +1527,18 @@ Error GenericDeviceTy::enqueueHostCall(void (*Callback)(void *), void *UserData,
return Err;
}
+Expected<InfoTreeNode> GenericDeviceTy::obtainInfo() {
+ Expected<InfoTreeNode> Info = obtainInfoImpl();
+ if (Info)
+ Info->add("UID", DeviceUid, "", DeviceInfo::UID);
+ return Info;
+}
+
Error GenericDeviceTy::printInfo() {
- auto Info = obtainInfoImpl();
+ Expected<InfoTreeNode> Info = obtainInfo();
// Get the vendor-specific info entries describing the device properties.
- if (auto Err = Info.takeError())
+ if (Error Err = Info.takeError())
return Err;
// Print all info entries.
diff --git a/offload/plugins-nextgen/host/src/rtl.cpp b/offload/plugins-nextgen/host/src/rtl.cpp
index eb4ecac9907a1..f75c9cd29b2ad 100644
--- a/offload/plugins-nextgen/host/src/rtl.cpp
+++ b/offload/plugins-nextgen/host/src/rtl.cpp
@@ -148,7 +148,9 @@ struct GenELF64DeviceTy : public GenericDeviceTy {
/// Create the device with a specific id.
GenELF64DeviceTy(GenericPluginTy &Plugin, int32_t DeviceId,
int32_t NumDevices)
- : GenericDeviceTy(Plugin, DeviceId, NumDevices, GenELF64GridValues) {}
+ : GenericDeviceTy(Plugin, DeviceId, NumDevices, GenELF64GridValues) {
+ DeviceUid = getHostDeviceUid();
+ }
~GenELF64DeviceTy() {}
diff --git a/offload/tools/deviceinfo/llvm-offload-device-info.cpp b/offload/tools/deviceinfo/llvm-offload-device-info.cpp
index 9b58d67f017ca..42ffb97d6d77c 100644
--- a/offload/tools/deviceinfo/llvm-offload-device-info.cpp
+++ b/offload/tools/deviceinfo/llvm-offload-device-info.cpp
@@ -176,6 +176,7 @@ ol_result_t printDevice(std::ostream &S, ol_device_handle_t D) {
printDeviceValue<const char *>(S, D, OL_DEVICE_INFO_NAME, "Name"));
OFFLOAD_ERR(printDeviceValue<const char *>(S, D, OL_DEVICE_INFO_PRODUCT_NAME,
"Product Name"));
+ OFFLOAD_ERR(printDeviceValue<const char *>(S, D, OL_DEVICE_INFO_UID, "UID"));
OFFLOAD_ERR(
printDeviceValue<ol_device_type_t>(S, D, OL_DEVICE_INFO_TYPE, "Type"));
OFFLOAD_ERR(printDeviceValue<const char *>(
diff --git a/offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp b/offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp
index 8cb0b8065c33e..30eafee026316 100644
--- a/offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp
+++ b/offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp
@@ -98,6 +98,16 @@ TEST_P(olGetDeviceInfoTest, SuccessProductName) {
ASSERT_EQ(std::strlen(Name.data()), Size - 1);
}
+TEST_P(olGetDeviceInfoTest, SuccessUID) {
+ size_t Size = 0;
+ ASSERT_SUCCESS(olGetDeviceInfoSize(Device, OL_DEVICE_INFO_UID, &Size));
+ ASSERT_GT(Size, 0ul);
+ std::vector<char> UID;
+ UID.resize(Size);
+ ASSERT_SUCCESS(olGetDeviceInfo(Device, OL_DEVICE_INFO_UID, Size, UID.data()));
+ ASSERT_EQ(std::strlen(UID.data()), Size - 1);
+}
+
TEST_P(olGetDeviceInfoTest, HostProductName) {
size_t Size = 0;
ASSERT_SUCCESS(olGetDeviceInfoSize(Host, OL_DEVICE_INFO_PRODUCT_NAME, &Size));
@@ -109,6 +119,16 @@ TEST_P(olGetDeviceInfoTest, HostProductName) {
ASSERT_EQ(std::strlen(Name.data()), Size - 1);
}
+TEST_P(olGetDeviceInfoTest, HostUID) {
+ size_t Size = 0;
+ ASSERT_SUCCESS(olGetDeviceInfoSize(Host, OL_DEVICE_INFO_UID, &Size));
+ ASSERT_GT(Size, 0ul);
+ std::vector<char> UID;
+ UID.resize(Size);
+ ASSERT_SUCCESS(olGetDeviceInfo(Host, OL_DEVICE_INFO_UID, Size, UID.data()));
+ ASSERT_EQ(std::strlen(UID.data()), Size - 1);
+}
+
TEST_P(olGetDeviceInfoTest, SuccessVendor) {
size_t Size = 0;
ASSERT_SUCCESS(olGetDeviceInfoSize(Device, OL_DEVICE_INFO_VENDOR, &Size));
diff --git a/offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp b/offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp
index c4a3c2d5e3c75..79a18c1d133dc 100644
--- a/offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp
+++ b/offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp
@@ -32,6 +32,7 @@ OL_DEVICE_INFO_SIZE_TEST_EQ(Platform, ol_platform_handle_t,
OL_DEVICE_INFO_PLATFORM);
OL_DEVICE_INFO_SIZE_TEST_NONZERO(Name, OL_DEVICE_INFO_NAME);
OL_DEVICE_INFO_SIZE_TEST_NONZERO(ProductName, OL_DEVICE_INFO_PRODUCT_NAME);
+OL_DEVICE_INFO_SIZE_TEST_NONZERO(UID, OL_DEVICE_INFO_UID);
OL_DEVICE_INFO_SIZE_TEST_NONZERO(Vendor, OL_DEVICE_INFO_VENDOR);
OL_DEVICE_INFO_SIZE_TEST_NONZERO(DriverVersion, OL_DEVICE_INFO_DRIVER_VERSION);
OL_DEVICE_INFO_SIZE_TEST_EQ(MaxWorkGroupSize, uint32_t,
>From 3b353fadf98fce49aff9eaff2baba5371ba6c12c Mon Sep 17 00:00:00 2001
From: Robert Imschweiler <robert.imschweiler at amd.com>
Date: Tue, 21 Oct 2025 13:58:21 -0500
Subject: [PATCH 2/3] align use of auto to the file
---
.../plugins-nextgen/common/src/PluginInterface.cpp | 14 +++++++-------
1 file changed, 7 insertions(+), 7 deletions(-)
diff --git a/offload/plugins-nextgen/common/src/PluginInterface.cpp b/offload/plugins-nextgen/common/src/PluginInterface.cpp
index cad91668c5e0b..78327e1480796 100644
--- a/offload/plugins-nextgen/common/src/PluginInterface.cpp
+++ b/offload/plugins-nextgen/common/src/PluginInterface.cpp
@@ -1528,21 +1528,21 @@ Error GenericDeviceTy::enqueueHostCall(void (*Callback)(void *), void *UserData,
}
Expected<InfoTreeNode> GenericDeviceTy::obtainInfo() {
- Expected<InfoTreeNode> Info = obtainInfoImpl();
- if (Info)
- Info->add("UID", DeviceUid, "", DeviceInfo::UID);
- return Info;
+ auto InfoOrErr = obtainInfoImpl();
+ if (InfoOrErr)
+ InfoOrErr->add("UID", getDeviceUid(), "", DeviceInfo::UID);
+ return InfoOrErr;
}
Error GenericDeviceTy::printInfo() {
- Expected<InfoTreeNode> Info = obtainInfo();
+ auto InfoOrErr = obtainInfo();
// Get the vendor-specific info entries describing the device properties.
- if (Error Err = Info.takeError())
+ if (auto Err = InfoOrErr.takeError())
return Err;
// Print all info entries.
- Info->print();
+ InfoOrErr->print();
return Plugin::success();
}
>From 7dbac00ff833d338d3a8e5ddce3fcc6e941010de Mon Sep 17 00:00:00 2001
From: Robert Imschweiler <robert.imschweiler at amd.com>
Date: Tue, 21 Oct 2025 14:31:15 -0500
Subject: [PATCH 3/3] add cuda UID
---
offload/plugins-nextgen/amdgpu/src/rtl.cpp | 2 +-
offload/plugins-nextgen/common/include/PluginInterface.h | 2 ++
offload/plugins-nextgen/common/src/PluginInterface.cpp | 4 ++++
offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp | 1 +
offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h | 2 ++
offload/plugins-nextgen/cuda/src/rtl.cpp | 6 ++++++
6 files changed, 16 insertions(+), 1 deletion(-)
diff --git a/offload/plugins-nextgen/amdgpu/src/rtl.cpp b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
index bcde829923b0d..4e1db2d5688cd 100644
--- a/offload/plugins-nextgen/amdgpu/src/rtl.cpp
+++ b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
@@ -2081,7 +2081,7 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
if (auto Err = getDeviceAttr(HSA_AMD_AGENT_INFO_UUID, Uuid))
return Err;
if (strcmp(Uuid + 3, "-XX") != 0)
- DeviceUid = std::string(Plugin.getName()) + "-" + std::string(Uuid);
+ setDeviceUidFromVendorUid(Uuid);
// Get the wavefront size.
uint32_t WavefrontSize = 0;
diff --git a/offload/plugins-nextgen/common/include/PluginInterface.h b/offload/plugins-nextgen/common/include/PluginInterface.h
index 6fe90d6ace1a0..70ba66582b94e 100644
--- a/offload/plugins-nextgen/common/include/PluginInterface.h
+++ b/offload/plugins-nextgen/common/include/PluginInterface.h
@@ -1208,6 +1208,8 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
/// combined with the plugin name, since the offload device id may overlap
/// between different plugins.
std::string DeviceUid;
+ /// Construct the device UID from the vendor (U)UID.
+ void setDeviceUidFromVendorUid(const char *VendorUid);
/// The default grid values used for this device.
llvm::omp::GV GridValues;
diff --git a/offload/plugins-nextgen/common/src/PluginInterface.cpp b/offload/plugins-nextgen/common/src/PluginInterface.cpp
index 78327e1480796..99d7367548592 100644
--- a/offload/plugins-nextgen/common/src/PluginInterface.cpp
+++ b/offload/plugins-nextgen/common/src/PluginInterface.cpp
@@ -1609,6 +1609,10 @@ Error GenericDeviceTy::syncEvent(void *EventPtr) {
bool GenericDeviceTy::useAutoZeroCopy() { return useAutoZeroCopyImpl(); }
+void GenericDeviceTy::setDeviceUidFromVendorUid(const char *VendorUid) {
+ DeviceUid = std::string(Plugin.getName()) + "-" + std::string(VendorUid);
+}
+
Error GenericPluginTy::init() {
if (Initialized)
return Plugin::success();
diff --git a/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp b/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp
index f5b2d074a47e7..e7a1ca38b3c13 100644
--- a/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp
+++ b/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.cpp
@@ -35,6 +35,7 @@ DLWRAP(cuFuncSetAttribute, 3)
// Device info
DLWRAP(cuDeviceGetName, 3)
+DLWRAP(cuDeviceGetUuid, 2)
DLWRAP(cuDeviceTotalMem, 2)
DLWRAP(cuDriverGetVersion, 1)
diff --git a/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h b/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h
index dec4e33508c62..cc1ba4b577645 100644
--- a/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h
+++ b/offload/plugins-nextgen/cuda/dynamic_cuda/cuda.h
@@ -33,6 +33,7 @@ typedef struct CUfunc_st *CUfunction;
typedef void (*CUhostFn)(void *userData);
typedef struct CUstream_st *CUstream;
typedef struct CUevent_st *CUevent;
+typedef struct CUuuid_st *CUuuid;
#define CU_DEVICE_INVALID ((CUdevice)(-2))
@@ -301,6 +302,7 @@ CUresult cuFuncSetAttribute(CUfunction, CUfunction_attribute, int);
// Device info
CUresult cuDeviceGetName(char *, int, CUdevice);
+CUresult cuDeviceGetUuid(CUuuid *, CUdevice);
CUresult cuDeviceTotalMem(size_t *, CUdevice);
CUresult cuDriverGetVersion(int *);
diff --git a/offload/plugins-nextgen/cuda/src/rtl.cpp b/offload/plugins-nextgen/cuda/src/rtl.cpp
index db94f7f2dd995..468581d522f51 100644
--- a/offload/plugins-nextgen/cuda/src/rtl.cpp
+++ b/offload/plugins-nextgen/cuda/src/rtl.cpp
@@ -293,6 +293,12 @@ struct CUDADeviceTy : public GenericDeviceTy {
if (auto Err = Plugin::check(Res, "error in cuDeviceGet: %s"))
return Err;
+ char Uuid[16];
+ Res = cuDeviceGetUuid(Uuid, Device);
+ if (auto Err = Plugin::check(Res, "error in cuDeviceGetUuid: %s"))
+ return Err;
+ setDeviceUidFromVendorUid(Uuid);
+
// Query the current flags of the primary context and set its flags if
// it is inactive.
unsigned int FormerPrimaryCtxFlags = 0;
More information about the llvm-commits
mailing list