[SYCL][E2E] Fix lit for arch selection and l0v2 adapter #20728

mateuszpn · 2025-11-21T15:33:01Z

Fix llvm-lit to enable the simultaneous use of _v1 and _v2 to select the l0 adapter and the use of arch to select the GPU, eg.:
llvm-lit --param "sycl_devices=level_zero_v2:arch-intel_gpu_mtl_u" sycl/test-e2e
Necessary to run tests on integrated GPU on system with another GPU
Includes fixed setting of ZE_AFFINITY_MASK in multi-gpu systems

aelovikov-intel · 2025-11-21T22:14:55Z

sycl/test-e2e/format.py

+                backend, device = parsed_dev_name.split(":", 1)
+                device_selector = parsed_dev_name
+                if backend == "level_zero" and device.isdigit():
+                    extra_env.append(f"ZE_AFFINITY_MASK={device}")
+                    device_selector = f"{backend}:0"


This needs a comment.

Some local lit configs, eg. sycl/test-e2e/AddressSanitizer/lit.local.cfg set affinity mask to 0. It needs to be overridden if we do not use level_zero:0 device. Comment added in the code

Why should we do that here and not fix AddressSanitizer? IMO, this file should just save {device} somewhere in the config and the AddressSanitizer/lit.local.cfg should use that for setting ZE_AFFINITY_MASK.

Oh sorry I missed that this was specifically for asan.

@aelovikov-intel Setting of ZE_AFFINITY_MASK moved to local configs.

aelovikov-intel · 2025-11-21T22:20:01Z

sycl/test-e2e/lit.cfg.py

    "opencl": ("cpu", "gpu", "fpga"),
    "cuda": "gpu",
-    "level_zero": "gpu",
+    "level_zero": ("gpu", "0", "1"),


Are you trying to allow llvm-lit --param sycl_devices=level_zero:0? We intentionally decided to disallow that in review of #18197

Although the arch- setting ends up pointing at particular device, eg. level_zero:1, we do not do it in the params. Removed.

sarnex · 2025-11-21T22:27:47Z

sycl/test-e2e/format.py

                expanded = "env"

                extra_env = get_extra_env([parsed_dev_name])
+                backend, device = parsed_dev_name.split(":", 1)


The example of the intended lit string after the change in the PR description is

llvm-lit --param "sycl_devices=level_zero_v2:arch-intel_gpu_mtl_u" sycl/test-e2e

but I was expecting something like

llvm-lit --param "sycl_devices='level_zero_v1:gpu;level_zero_v2:gpu" sycl/test-e2e

Can you clarify on the intended use case? Thanks

The goal is to force tests to run on specified device, eg. integrated gpu, even when other devices are present. The option was available before (see lit.cpg.py:901). Although it is rarely used, it is still perfect for testing integrated GPUs. Unfortunately, due to a bug, it did not work with the _v1 or _v2 postfixes in the backend name. The PR fixes this bug.

sarnex

I can't comment if ZE_AFFINITY_MASK is the right way to pin it to a specific device but other than that lgtm, thanks!

Signed-off-by: Mateusz P. Nowak <[email protected]>

aelovikov-intel · 2025-11-25T15:55:42Z

sycl/test-e2e/lit.cfg.py

    features.update(sg_size_features)
    features.update(architecture_feature)
    features.update(device_family)
+    features.update(aspects)


Why this line? Isn't 1120 enough?

aelovikov-intel · 2025-11-25T15:57:27Z

sycl/test-e2e/lit.cfg.py

-import lit.formats
-import lit.util
-


aelovikov-intel · 2025-11-25T15:58:42Z

sycl/test-e2e/AddressSanitizer/lit.local.cfg

    config.unsupported=True

-config.environment["ZE_AFFINITY_MASK"] = "0"
+if hasattr(config, 'ze_affinity_mask') and config.ze_affinity_mask is not None:


You have

# Determine ZE_AFFINITY_MASK for Level Zero devices. # Sanitizer tests need to set ZE_AFFINITY_MASK to a single device index config.ze_affinity_mask = None

below, is hasattr check needed?

aelovikov-intel · 2025-11-25T16:01:01Z

sycl/test-e2e/AddressSanitizer/lit.local.cfg

+else:
+    config.environment["ZE_AFFINITY_MASK"] = "0"


Is else really needed? IIUC, config.ze_affinity_mask would be set for L0 devices, and for CPU this env. variable is meaningless anyway.

aelovikov-intel · 2025-11-25T16:02:01Z

sycl/test-e2e/format.py

+                # If ZE_AFFINITY_MASK is set, it filters devices so we should use :0
+                device_selector = parsed_dev_name
+                if "ZE_AFFINITY_MASK" in test.config.environment:
+                    backend, _ = parsed_dev_name.split(":", 1)
+                    device_selector = f"{backend}:0"


@sarnex , should we just unset it and report a warning instead? Somewhere around https://github.com/mateuszpn/llvm/blob/c5fa7266ca7ecfc0a265ba06da9af99a40cfd557/sycl/test-e2e/lit.cfg.py#L77-L96

Yeah I recommend this way, having code trying to play nice with existing envvars seems bad

fix lit for arch selection and l0v2 adapter

b262dc6

mateuszpn requested a review from a team as a code owner November 21, 2025 15:33

mateuszpn requested a review from maarquitos14 November 21, 2025 15:33

mateuszpn temporarily deployed to WindowsCILock November 21, 2025 15:33 — with GitHub Actions Inactive

mateuszpn temporarily deployed to WindowsCILock November 21, 2025 15:57 — with GitHub Actions Inactive

mateuszpn had a problem deploying to WindowsCILock November 21, 2025 15:57 — with GitHub Actions Failure

aelovikov-intel changed the title ~~[SYCL][e2e-tests] fix lit for arch selection and l0v2 adapter~~ [SYCL][E2E] Fix lit for arch selection and l0v2 adapter Nov 21, 2025

aelovikov-intel reviewed Nov 21, 2025

View reviewed changes

aelovikov-intel requested a review from sarnex November 21, 2025 22:22

sarnex reviewed Nov 21, 2025

View reviewed changes

mateuszpn had a problem deploying to WindowsCILock November 24, 2025 12:14 — with GitHub Actions Failure

apply comments

08db888

mateuszpn temporarily deployed to WindowsCILock November 24, 2025 13:23 — with GitHub Actions Inactive

mateuszpn requested review from aelovikov-intel and sarnex November 24, 2025 13:27

mateuszpn temporarily deployed to WindowsCILock November 24, 2025 13:48 — with GitHub Actions Inactive

sarnex approved these changes Nov 24, 2025

View reviewed changes

mateuszpn requested a review from a team as a code owner November 25, 2025 12:35

mateuszpn had a problem deploying to WindowsCILock November 25, 2025 12:35 — with GitHub Actions Error

mateuszpn had a problem deploying to WindowsCILock November 25, 2025 12:41 — with GitHub Actions Error

mateuszpn force-pushed the lit-fix-v2 branch from 0bd0caa to 3bbaef5 Compare November 25, 2025 12:41

mateuszpn had a problem deploying to WindowsCILock November 25, 2025 12:42 — with GitHub Actions Error

mateuszpn force-pushed the lit-fix-v2 branch from 3bbaef5 to 1afbf80 Compare November 25, 2025 12:42

mateuszpn had a problem deploying to WindowsCILock November 25, 2025 12:43 — with GitHub Actions Error

mateuszpn force-pushed the lit-fix-v2 branch from 1afbf80 to c8868c8 Compare November 25, 2025 12:43

mateuszpn had a problem deploying to WindowsCILock November 25, 2025 12:43 — with GitHub Actions Error

fix affinity mask setting

c5fa726

Signed-off-by: Mateusz P. Nowak <[email protected]>

mateuszpn force-pushed the lit-fix-v2 branch from c8868c8 to c5fa726 Compare November 25, 2025 12:48

mateuszpn temporarily deployed to WindowsCILock November 25, 2025 12:49 — with GitHub Actions Inactive

mateuszpn temporarily deployed to WindowsCILock November 25, 2025 13:14 — with GitHub Actions Inactive

aelovikov-intel reviewed Nov 25, 2025

View reviewed changes

		import lit.formats
		import lit.util

[SYCL][E2E] Fix lit for arch selection and l0v2 adapter #20728

Are you sure you want to change the base?

[SYCL][E2E] Fix lit for arch selection and l0v2 adapter #20728

Conversation

mateuszpn commented Nov 21, 2025 • 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!

aelovikov-intel Nov 24, 2025 • 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!

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!

sarnex 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!

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

mateuszpn commented Nov 21, 2025 •

edited

Loading

aelovikov-intel Nov 24, 2025 •

edited

Loading

sarnex left a comment •

edited

Loading