[clang] [llvm] [TargetParser] Add a list of Intel GPUs, and use it in offload-arch (PR #222072)

[Srividya Sundaram via cfe-commits]

================
@@ -173,11 +189,42 @@ int printGPUsByLevelZero() {
     CALL_ZE_AND_CHECK(zeDeviceGet, Driver, &DeviceCount, Devices.data());
 
     for (auto Device : Devices) {
+      ze_device_ip_version_ext_t IPVersion = {};
+      IPVersion.stype = ZE_STRUCTURE_TYPE_DEVICE_IP_VERSION_EXT;
+      IPVersion.pNext = nullptr;
+
       ze_device_properties_t DeviceProperties = {};
       DeviceProperties.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
-      DeviceProperties.pNext = nullptr;
+      DeviceProperties.pNext = &IPVersion;
       CALL_ZE_AND_CHECK(zeDeviceGetProperties, Device, &DeviceProperties);
-      llvm::outs() << DeviceProperties.name << '\n';
+
+      // A driver that does not support the extension leaves the chained
+      // structure untouched, in which case there is no architecture to name.
+      if (IPVersion.ipVersion == 0) {
+        if (Verbose)
+          llvm::errs() << "Unable to query the IP version of device '"
+                       << DeviceProperties.name << "'\n";
+        continue;
+      }
+
+      if (Verbose)
+        llvm::errs() << "Found device '" << DeviceProperties.name << "'\n";
+
+      // Naming an unknown device after its GMDID would print something that
+      // --offload-arch cannot accept, because this build knows no IGCA level to
+      // compile for.  Report it instead, spelling out the GMDID so that the
+      // device can be identified.
----------------
srividya-sundaram wrote:

Detection should name the device, and the driver decides what is targetable, because
only the driver can name the bad value the user typed for --offload-arch and list the supported ones.

I think with the current code if the user has a device xe_35.11.0, and a clang whose .def predates it:

We will get:
```
$ clang++ -fsycl --offload-arch=native app.cpp
Unknown Intel GPU 'Intel(R) Graphics', which reports the architecture xe_35.11.0
clang++: error: cannot determine offload architecture: No GPU detected in the
system; consider passing it via '--offload-arch'  ---> wrong, the user did have a GPU installed.
```

Instead, if print `getNumericArchName()` and continue, instead of erroring, that seems helpful:

```
$ clang++ -fsycl --offload-arch=native app.cpp
clang++: error: unsupported Intel GPU architecture 'xe_35.11.0'
note: valid target IDs are: xe-pvc, xe-bmg-g21, xe-lnl-m, ...
```

So, offload-arch can just do detection and driver the validation.
Other backends also do detection, leaving validation and diagnostics to the driver.

Detection prints, never validates

https://github.com/llvm/llvm-project/blob/main/clang/tools/offload-arch/NVPTXArch.cpp#L115
https://github.com/llvm/llvm-project/blob/main/clang/tools/offload-arch/AMDGPUArchByKFD.cpp#L81-L82
https://github.com/llvm/llvm-project/blob/main/clang/tools/offload-arch/AMDGPUArchByHIP.cpp#L378
https://github.com/llvm/llvm-project/blob/main/clang/tools/offload-arch/AMDGPUArchByHIP.cpp#L440

Driver validates detection output through the user-input path

https://github.com/llvm/llvm-project/blob/main/clang/lib/Driver/Driver.cpp#L4011-L4025
https://github.com/llvm/llvm-project/blob/main/clang/lib/Driver/Driver.cpp#L3936-L3946
https://github.com/llvm/llvm-project/blob/main/clang/include/clang/Basic/DiagnosticDriverKinds.td#L60

-cc1 validates again and lists valid values

https://github.com/llvm/llvm-project/blob/main/clang/lib/Basic/Targets.cpp#L855-L862


https://github.com/llvm/llvm-project/pull/222072