[UR] Static OpenCL adapter by mateuszpn · Pull Request #21983 · intel/llvm

mateuszpn · 2026-05-11T15:32:12Z

Builds the OpenCL adapter as a static library embedded in the UR loader instead of a separate runtime-loaded .so/.dll. Controlled by UR_STATIC_ADAPTER_OPENCL (default ON for SYCL builds). The OpenCL ICD itself is still resolved dynamically at runtime, so there's no OpenCL link-time dependency.

Key changes

Dynamic ICD loading — new ocl_dynamic_lib.{cpp,hpp} loads the ICD at init: LoadLibraryExA("OpenCL.dll", LOAD_LIBRARY_SEARCH_SYSTEM32) on Windows, dlopen("libOpenCL.so.1") (→ libOpenCL.so) on Linux, resolving entry points through function pointers (core_functions.def, new ocl_functions.def).
Namespace isolation — all adapter entry points moved into namespace ur::opencl to avoid symbol collisions; the loader wires DDI tables directly via ur::opencl::urGet*ProcAddrTable.
Loader integration — under UR_STATIC_ADAPTER_OPENCL the adapter is probed/registered in-process (no adapter dlopen), and discarded cleanly when no platforms are present (ur_loader.cpp, ur_adapter_registry.hpp, loader/CMakeLists.txt).
Build/linking control — opencl/CMakeLists.txt builds the static lib, drops the OpenCL link, and zeroes UR_APIEXPORT/UR_APICALL/UR_DLLEXPORT so definitions carry no __declspec(dllimport/dllexport). BuildUnifiedRuntime.cmake forces the option on, propagates it into the Windows ExternalProject, and suppresses the now-nonexistent ur_adapter_opencl.dll in the copy/install logic.
SYCL runtime — os_util.hpp: with the adapter static, __SYCL_OCL_CALL probes libOpenCL.so.1/libOpenCL.so for clRetain*/clRelease* directly (the libur_adapter_opencl.so it used to look them up in no longer exists);

pbalcer · 2026-05-13T08:00:09Z

  }
 #endif
+#ifdef UR_STATIC_ADAPTER_OPENCL
+  if (!adapter_registry.adaptersForceLoaded()) {


don't duplicate this if.

pbalcer · 2026-05-13T08:03:32Z

@@ -0,0 +1,120 @@
+// OpenCL function definitions for dynamic loading
+// Format: OCL_FUNC(name, required)
+// where 'required' is 1 for OpenCL 1.0 functions that must be present, 0 for optional


how did you determine which functions are required, and which are not?
some of these don't look right to me.

The split was based on "OpenCL 1.0 spec membership = required, anything later = optional". However, not all the functions are used by adapter, and those which are used - are required. Now the file is reduced to functions actually used by adapter.

pbalcer · 2026-05-26T08:35:04Z

+    parser.add_argument(
+        "--static-opencl-adapter",
+        action="store_true",
+        help="Build OpenCL adapter as static library with dynamic OpenCL loading (no OpenCL dependency at link time).",
+    )


I don't think we need to add this option. We didn't add any other backend. I think this would be a very rarely used function. And we want to enable static adapters by default anyway.

pbalcer · 2026-05-26T08:44:13Z

Suggested change

*pNumAdapters = liveAdapter != nullptr ? 1 : 0;

?
what is the purpose of checking ocl::loadOCLLibrary() result here?

Applied the suggestion

pbalcer · 2026-05-26T08:45:05Z

does the namespace have to be ifdefed?

In general, I suggest limiting the number of ifdefs to the minimum required, preferring to share implementations wherever possible.

pbalcer · 2026-05-26T08:47:08Z

please follow the same pattern used in level-zero adapter, with these symbols defined in ur_interface_loader.hpp. You should be able to reuse the exact same templates, just with different namespace name.

pbalcer · 2026-05-26T08:51:02Z

wouldn't it be simpler to rename the struct fields?

pbalcer · 2026-05-26T08:54:11Z

namespace it in both modes, and remove the ifdef.

pbalcer · 2026-05-26T08:56:21Z

did you modify this file manually?

please use the existing ur_interface_loader template. It shouldn't require modifications.

Yes, that hand-edit is gone. generate_code.py now drives the existing ur_interface_loader.{cpp,hpp}.mako templates for opencl too. Only template change was parameterizing two L0-hardcoded strings (Level Zero →
${adapter}, UR_STATIC_ADAPTER_LEVEL_ZERO → UR_STATIC_ADAPTER_${Adapter});

pbalcer · 2026-05-26T08:59:43Z

why? is there a symbol conflict?

This should get resolved by creating a common library for the adapter. #21845 already does that here:
https://github.com/intel/llvm/pull/21845/changes#diff-e83af30e7502eb85108ac495d4a4d50775e6b0eba4e27e8b682c797f8b8a168fR26

Reverted for now, should be fixed after #21845 merge

pbalcer · 2026-06-26T08:33:37Z

+  // The device above is a raw reference into the platform's device list, which
+  // is owned by the platform. Hold a shared_ptr to the owning platform so that
+  // `MDevice` (and the per-device queue list it is registered in) is guaranteed
+  // to outlive this queue. Without this, a queue_impl kept alive by a deferred
+  // scheduler command can outlive its device_impl and crash in ~queue_impl when
+  // unregistering from the (already freed) device's queue list.
+  const std::shared_ptr<platform_impl> MPlatform;


how is this related?

The fix in queue_impl is not directly related to the PR's core purpose, still the issue (pre-existing, but static linking probably changed timing and the problem surfaced) prevented Windows tests to pass - "Queue not found" assert was fired on destructor of queue_impl.
You are probably right to move the fix to separate PR. I did #22481 for this.

pbalcer · 2026-06-26T08:33:56Z

+ur_result_t mock_urAdapterGetInfo(void *pParams) {
+  const auto &params = *static_cast<ur_adapter_get_info_params_t *>(pParams);
+  if (*params.ppropName == UR_ADAPTER_INFO_BACKEND) {
+    if (*params.ppPropValue) {
+      *reinterpret_cast<ur_backend_t *>(*params.ppPropValue) =
+          UR_BACKEND_OPENCL;
+    }
+    if (*params.ppPropSizeRet) {
+      **params.ppPropSizeRet = sizeof(ur_backend_t);
+    }
+  }
+  return UR_RESULT_SUCCESS;
+}
+


pbalcer · 2026-06-26T08:57:01Z

+    {
+      // Probe the adapter before registering: urAdapterGet triggers
+      // loadOCLLibrary(). If the OpenCL library is absent the global
+      // ocl::*_ptr function pointers stay null, and any later call through
+      // the DDI table (e.g. urPlatformGetInfo) would segfault.
+      ur_dditable_t ocl_ddi = {};
+      ur::opencl::urAdapterGetDdiTables(&ocl_ddi);
+      ur_adapter_handle_t hAdapter = nullptr;
+      if (ocl_ddi.Adapter.pfnGet &&
+          ocl_ddi.Adapter.pfnGet(1, &hAdapter, nullptr) == UR_RESULT_SUCCESS) {
+        ocl_ddi.Adapter.pfnRelease(hAdapter);
+        auto &opencl = platforms.emplace_back(nullptr);
+        opencl.dditable = ocl_ddi;
+        staticOCLRegistered = true;
+      }
+    }


How was this handled in the normal dynamic case? I don't think statically linking changes anything here. The OpenCL adapter should return 0 platforms, and users wouldn't call any functions on it.

pbalcer

lgtm, just one nit.

pbalcer · 2026-06-30T12:57:10Z

+
  void enableMock() {
    adaptersLoadPaths.clear();
+    forceLoaded = true;


Please add a comment why this is added here.

kswiecicki · 2026-07-01T09:25:10Z

@intel/dpcpp-devops-reviewers, @intel/bindless-images-reviewers, @intel/sycl-graphs-reviewers could you take a look please.

slawekptak

SYCL changes LGTM

kswiecicki · 2026-07-01T12:47:54Z

Why was this changed?

mmichel11

Command buffer changes LGTM

sarnex

cmake lgtm minus a nit

sarnex · 2026-07-01T14:20:08Z

+    ur_adapter_is_static(${adapter} adapter_is_static)
+    if(adapter_is_static)
+      # No DLL for statically linked adapter
+    else()


we can just do
if(NOT adapter_is_static) instead of having an empty branch

dyniols

bindless images changes LGTM

mateuszpn · 2026-07-02T16:25:08Z

@intel/llvm-gatekeepers Please consider merge. MacOS workflow failure appears unrelated to the changes in this PR and looks like a workflow configuration issue.

mateuszpn force-pushed the ocl-static-link branch 5 times, most recently from 61e281d to 2857569 Compare May 12, 2026 12:08

mateuszpn requested a review from kswiecicki May 12, 2026 13:12

pbalcer reviewed May 13, 2026

View reviewed changes

mateuszpn force-pushed the ocl-static-link branch from 42ffe80 to f002c34 Compare May 25, 2026 12:39

mateuszpn marked this pull request as ready for review May 25, 2026 13:44

mateuszpn requested review from a team as code owners May 25, 2026 13:44

pbalcer reviewed May 26, 2026

View reviewed changes

kswiecicki reviewed May 27, 2026

View reviewed changes

Comment thread unified-runtime/source/adapters/opencl/ocl_dynamic_lib.cpp Outdated

mateuszpn force-pushed the ocl-static-link branch from 4148cd8 to 7e87593 Compare May 27, 2026 13:47

mateuszpn requested review from a team as code owners May 27, 2026 13:47

mateuszpn requested a review from mmichel11 May 27, 2026 13:47

mateuszpn force-pushed the ocl-static-link branch from 7e87593 to 5918f7d Compare May 27, 2026 14:10

mateuszpn requested review from kswiecicki and pbalcer May 27, 2026 14:38

mateuszpn force-pushed the ocl-static-link branch 7 times, most recently from a3e5bbe to cb907b8 Compare June 1, 2026 14:11

mateuszpn requested a review from a team as a code owner June 1, 2026 14:11

kswiecicki approved these changes Jun 23, 2026

View reviewed changes

mateuszpn force-pushed the ocl-static-link branch 2 times, most recently from 7170f4b to 1e647d8 Compare June 23, 2026 14:30

mateuszpn added 3 commits June 23, 2026 14:31

Merge branch 'intel:sycl' into ocl-static-link

a6bde63

add missing implementations after rebase

1e647d8

Merge branch 'intel:sycl' into ocl-static-link

d7ccdbd

mateuszpn force-pushed the ocl-static-link branch from c755601 to 3e27cc9 Compare June 24, 2026 14:02

mateuszpn added 2 commits June 24, 2026 15:03

fix the Queue-not-found bug

3e27cc9

apply comment

c5035fe

mateuszpn force-pushed the ocl-static-link branch from 554972b to 50ebd35 Compare June 25, 2026 11:52

Merge branch 'intel:sycl' into ocl-static-link

50ebd35

mateuszpn requested a review from pbalcer June 26, 2026 08:23

kswiecicki reviewed Jun 26, 2026

View reviewed changes

Comment thread unified-runtime/source/adapters/opencl/adapter.cpp Outdated

pbalcer reviewed Jun 26, 2026

View reviewed changes

mateuszpn added 2 commits June 29, 2026 17:07

apply comments

db8849c

Merge branch 'intel:sycl' into ocl-static-link

6ea84e3

mateuszpn requested a review from pbalcer June 30, 2026 12:06

pbalcer approved these changes Jun 30, 2026

View reviewed changes

mateuszpn added 2 commits June 30, 2026 14:23

apply comment

3153ec0

Merge remote-tracking branch 'origin/sycl' into ocl-static-link

0c477cc

slawekptak approved these changes Jul 1, 2026

View reviewed changes

kswiecicki reviewed Jul 1, 2026

View reviewed changes

Comment thread sycl/source/detail/queue_impl.hpp

kswiecicki Jul 1, 2026

Copy link
Copy Markdown

Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why was this changed?

kswiecicki approved these changes Jul 1, 2026

View reviewed changes

mmichel11 approved these changes Jul 1, 2026

View reviewed changes

sarnex approved these changes Jul 1, 2026

View reviewed changes

dyniols approved these changes Jul 1, 2026

View reviewed changes

mateuszpn added 2 commits July 2, 2026 10:43

fix BuildUnifiedRuntime.cmake

b26a0b8

Merge branch 'intel:sycl' into ocl-static-link

f1a497a

Uh oh!

Conversation

mateuszpn commented May 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbalcer left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kswiecicki commented Jul 1, 2026

Uh oh!

slawekptak left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mmichel11 left a comment

Choose a reason for hiding this comment

Uh oh!

sarnex left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dyniols left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mateuszpn commented Jul 2, 2026

Uh oh!

Reviewers

mateuszpn commented May 11, 2026 •

edited

Loading

dyniols left a comment •

edited

Loading