[SYCL][UR] Improve diagnostics for missing platforms in sycl-ls by inteltdelisle · Pull Request #23542 · intel/llvm

inteltdelisle · 2026-10-19T14:33:18Z

When a backend's platforms or devices fail to show up in sycl-ls, users currently get little indication of where in the stack things went wrong. This change adds some logging and guidance to help narrow it down:

UR loader: log each adapter that is loaded (statically or dynamically) and the total number of adapters loaded.
UR adapters (CUDA, HIP, Level Zero, OpenCL, Native CPU, Offload): log the number of platforms each adapter initializes at INFO level. The Level Zero message is raised from DEBUG to INFO for consistency.
OpenCL adapter: log a message when clGetPlatformIDs fails or returns no platforms, and when a platform reports CL_DEVICE_NOT_FOUND.
sycl-ls --verbose: print backend-specific hints at the end of the output (clinfo, nvidia-smi, rocminfo/amd-smi, ZE_ENABLE_LOADER_DEBUG_TRACE, strace).

The new messages can be seen with UR_LOG_LOADER / UR_LOG_ set to "level:info;output:stdout" (or similar).

This doesn't fully resolve the issue, since sycl-ls still can't tell users why a backend is missing, but it should make the cause much easier to find.

Refs: #22313

When a backend's platforms or devices fail to show up in sycl-ls, users currently get little indication of where in the stack things went wrong. This change adds some logging and guidance to help narrow it down: * UR loader: log each adapter that is loaded (statically or dynamically) and the total number of adapters loaded. * UR adapters (CUDA, HIP, Level Zero, OpenCL, Native CPU, Offload): log the number of platforms each adapter initializes at INFO level. The Level Zero message is raised from DEBUG to INFO for consistency. * OpenCL adapter: log an error when clGetPlatformIDs fails or returns no platforms, and when a platform reports CL_DEVICE_NOT_FOUND. * sycl-ls --verbose: print backend-specific hints at the end of the output (clinfo, nvidia-smi, rocminfo/amd-smi, ZE_ENABLE_LOADER_DEBUG_TRACE, /dev/dri/renderD* permissions, strace), and say which UR logger messages point to a configuration problem. The new messages can be seen with UR_LOG_LOADER / UR_LOG_<ADAPTER> set to "level:info;output:stdout" (or similar). This doesn't fully resolve the issue, since sycl-ls still can't tell users *why* a backend is missing, but it should make the cause much easier to find. Refs: intel#22313 Signed-off-by: Delisle, Thierry <thierry.delisle@intel.com>

bratpiorka

CUDA/HIP/NativeCPU/OpenCL LGTM

ldorau

The description says more than the code does. It says:

OpenCL adapter logs an error when clGetPlatformIDs fails and for CL_DEVICE_NOT_FOUND,
sycl-ls prints a hint about permissions on /dev/dri/renderD*,
sycl-ls explains which UR logger messages point to a configuration problem.

I cannot find any of these in the code. The OpenCL adapter adds only one INFO message ("CL No platforms found").
Please update the description (and the commit message) to match the code, or add the missing parts.

ldorau · 2026-10-20T08:38:49Z

  cl_uint NumPlatforms = 0;
  cl_int Result = ocl::clGetPlatformIDs_ptr(0, nullptr, &NumPlatforms);
  if (Result != CL_SUCCESS || NumPlatforms == 0) {
+    UR_LOG(INFO, "CL No platforms found (Result = {})", Result);


Nit: the new messages use different styles:

CUDA Initialized {} platforms

HIP Initialized {} platforms

LZ Initialized {} platforms

CL Initialized {} platforms

CL No platforms found (Result = {})

NATIVE_CPU: 1 platform available

OFFLOAD: {} platforms available

Could you use one style for all of them? For example:
<Adapter name>: found {} platforms.
Also "OpenCL" is easier to understand than "CL".

ldorau · 2026-10-20T08:40:18Z

  UR_ASSERT(pNumPlatforms || phPlatforms, UR_RESULT_ERROR_INVALID_VALUE);

  if (pNumPlatforms) {
+    UR_LOG(INFO, "NATIVE_CPU: 1 platform available");


Nit: this is inside urPlatformGet(), which is called many times.
So this message is printed again and again. The other adapters print it only once,
during initialization. Could you move it there, or keep it but know it will repeat?

Unless I'm missing something, I don't believe this particular adapter has initialization code. So I moved the log into a static initilization.

ldorau · 2026-10-20T08:40:49Z

              ur_platform_handle_t *phPlatforms, uint32_t *pNumPlatforms) {

  if (pNumPlatforms) {
+    UR_LOG(INFO, "OFFLOAD: {} platforms available", Adapter->Platforms.size());


Nit: this is inside urPlatformGet(), which is called many times.
So this message is printed again and again. The other adapters print it only once,
during initialization. Could you move it there, or keep it but know it will repeat?

inteltdelisle · 2026-10-20T13:07:49Z

The description says more than the code does. It says:

OpenCL adapter logs an error when clGetPlatformIDs fails and for CL_DEVICE_NOT_FOUND,

sycl-ls prints a hint about permissions on /dev/dri/renderD*,

sycl-ls explains which UR logger messages point to a configuration problem.

I cannot find any of these in the code. The OpenCL adapter adds only one INFO message ("CL No platforms found"). Please update the description (and the commit message) to match the code, or add the missing parts.

Updated the description:

OpenCL logs a message, I forgot I downgraded this to info.
I believe the code already checked for /dev/dri/renderD* in checkRenderGroupPermission.
The hints at the end used to point to existing logger messages, but I felt they could go out of sync too easily.

Signed-off-by: Delisle, Thierry <thierry.delisle@intel.com>

inteltdelisle requested a review from 0x12CC October 19, 2026 14:33

inteltdelisle requested review from a team as code owners October 19, 2026 14:33

inteltdelisle requested review from bratpiorka and uditagarwal97 October 19, 2026 14:33

bratpiorka reviewed Oct 20, 2026

View reviewed changes

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

bratpiorka approved these changes Oct 20, 2026

View reviewed changes

ldorau requested changes Oct 20, 2026

View reviewed changes

Address review comments.

44ed92c

Signed-off-by: Delisle, Thierry <thierry.delisle@intel.com>

Navigation Menu

Repository navigation

[SYCL][UR] Improve diagnostics for missing platforms in sycl-ls - #23542

[SYCL][UR] Improve diagnostics for missing platforms in sycl-ls#23542
inteltdelisle wants to merge 2 commits into
intel:syclfrom
inteltdelisle:unified-runtime/verbose_sycl_ls

inteltdelisle commented Oct 19, 2026 •

edited

Loading

Uh oh!

Uh oh!

bratpiorka left a comment

Uh oh!

ldorau left a comment

Uh oh!

Uh oh!

ldorau Oct 20, 2026

Uh oh!

inteltdelisle Oct 20, 2026

Uh oh!

ldorau Oct 20, 2026

Uh oh!

inteltdelisle Oct 20, 2026

Uh oh!

ldorau Oct 20, 2026

Uh oh!

Uh oh!

inteltdelisle commented Oct 20, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

inteltdelisle commented Oct 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

bratpiorka left a comment

Choose a reason for hiding this comment

Uh oh!

ldorau left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ldorau Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

inteltdelisle Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

ldorau Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

inteltdelisle Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

ldorau Oct 20, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

inteltdelisle commented Oct 20, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

inteltdelisle commented Oct 19, 2026 •

edited

Loading