[UR][L0v2 adapter] Fix buffers for integrated gpu #20495

mateuszpn · 2025-10-28T15:43:18Z

This PR implements zero-copy buffers for integrated GPUs.

Signed-off-by: Mateusz P. Nowak <mateusz.p.nowak@intel.com>

pbalcer · 2025-11-05T09:06:49Z

sycl/test-e2e/lit.cfg.py

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


Changed for local tests, but may be useful in CI, too, depending on hw config. Allows to point out which gpu should be used for tests (eg. integrated/discrete). Example:
llvm-lit --param "sycl_devices=level_zero_v2:1" - the tests are run on device level_zero:1 with v2 adapter

After closer look, as some tests appeared as unsupported, additional one-line fix was necessary.
So, if any of reviewers thinks the change does not belong to this PR, I am also OK with that.

I don't think we should be changing the default config. CI and other dev systems where this is used may have a different configuration where :0 and :1 mean e.g., different dGPUs. Or there may not be a second GPU at all.

pbalcer · 2025-11-05T09:10:05Z

unified-runtime/source/adapters/level_zero/v2/memory.cpp

-    void *rawPtr;
-    UR_CALL_THROWS(hContext->getDefaultUSMPool()->allocate(
-        hContext, nullptr, nullptr, UR_USM_TYPE_HOST, size, &rawPtr));
+  if (hostPtr) {


when can hostPtr be null?

Just checking the argument. Can we be sure it will never be null?

pbalcer · 2025-11-05T09:16:49Z

unified-runtime/source/adapters/level_zero/v2/memory.cpp

+    auto ret =
+        getMemoryAttrs(hContext->getZeHandle(), hostPtr, nullptr, &memProps);
+
+    if (ret == UR_RESULT_SUCCESS && memProps.type != ZE_MEMORY_TYPE_UNKNOWN) {


Why check against ZE_MEMORY_TYPE_UNKNOWN and not something specific?

If I understand specification right, mem type can be known to be USM (host, device or shared) or unknown. The condition checks if mem type is USM.

pbalcer · 2025-11-05T09:17:00Z

unified-runtime/source/adapters/level_zero/v2/memory.cpp

    wait_list_view & /*waitListView*/) {
-  // TODO: if writeBackPtr is set, we should copy the data back
-  /* nop */
+  // No-op: integrated buffers use zero-copy, no synchronization needed


what if we go in this path when creating the buffer:
"// Import failed - allocate backing buffer and set up copy-back"
?

If "import failed" is a possible scenario, we should have a test. If it's not, that code path should be removed.

pbalcer · 2025-11-05T09:17:13Z

unified-runtime/source/adapters/level_zero/v2/memory.cpp

-  // TODO: if writeBackPtr is set, we should map to that pointer
-  // because that's what SYCL expects, SYCL will attempt to call free
-  // on the resulting pointer leading to double free with the current
-  // implementation. Investigate the SYCL implementation.


was this comment false?

You are right, the comment was not false as long as map/unmap functions missed full handling of copy-back path. Last commit addresses the issue.

Signed-off-by: Mateusz P. Nowak <mateusz.p.nowak@intel.com>

mateuszpn temporarily deployed to WindowsCILock October 28, 2025 15:43 — with GitHub Actions Inactive

mateuszpn had a problem deploying to WindowsCILock October 28, 2025 16:11 — with GitHub Actions Failure

mateuszpn temporarily deployed to WindowsCILock October 28, 2025 16:11 — with GitHub Actions Inactive

Enable teporary disabled tests

203620d

mateuszpn temporarily deployed to WindowsCILock October 29, 2025 09:25 — with GitHub Actions Inactive

mateuszpn had a problem deploying to WindowsCILock November 4, 2025 11:39 — with GitHub Actions Error

detect and handle host mem in integrated gpu

c3ce32b

Signed-off-by: Mateusz P. Nowak <mateusz.p.nowak@intel.com>

mateuszpn had a problem deploying to WindowsCILock November 4, 2025 11:41 — with GitHub Actions Error

mateuszpn force-pushed the int-buffers branch from 5516bc4 to c3ce32b Compare November 4, 2025 11:43

mateuszpn had a problem deploying to WindowsCILock November 4, 2025 11:43 — with GitHub Actions Error

Merge branch 'intel:sycl' into int-buffers

33b3f69

mateuszpn had a problem deploying to WindowsCILock November 4, 2025 11:44 — with GitHub Actions Error

some cleanups

f1b62a9

Signed-off-by: Mateusz P. Nowak <mateusz.p.nowak@intel.com>

mateuszpn temporarily deployed to WindowsCILock November 4, 2025 12:01 — with GitHub Actions Inactive

mateuszpn changed the title ~~[UR][L0v2 adapter] Fix SYCL buffer map implementation~~ [UR][L0v2 adapter] Fix buffers for integrated gpu Nov 4, 2025

mateuszpn temporarily deployed to WindowsCILock November 4, 2025 12:27 — with GitHub Actions Inactive

mateuszpn had a problem deploying to WindowsCILock November 4, 2025 15:25 — with GitHub Actions Error

mateuszpn force-pushed the int-buffers branch from 49057b5 to d002d08 Compare November 4, 2025 15:33

mateuszpn marked this pull request as ready for review November 4, 2025 15:34

mateuszpn requested review from a team as code owners November 4, 2025 15:34

mateuszpn requested a review from vinser52 November 4, 2025 15:34

mateuszpn temporarily deployed to WindowsCILock November 4, 2025 15:34 — with GitHub Actions Inactive

mateuszpn temporarily deployed to WindowsCILock November 4, 2025 15:57 — with GitHub Actions Inactive

pbalcer reviewed Nov 5, 2025

View reviewed changes

mateuszpn had a problem deploying to WindowsCILock November 6, 2025 11:18 — with GitHub Actions Error

mateuszpn force-pushed the int-buffers branch from cd3bd6e to 5c52b4c Compare November 6, 2025 11:32

mateuszpn had a problem deploying to WindowsCILock November 6, 2025 11:33 — with GitHub Actions Error

copy-back for non-USM memory

2a0d2df

Signed-off-by: Mateusz P. Nowak <mateusz.p.nowak@intel.com>

mateuszpn force-pushed the int-buffers branch from 5c52b4c to 2a0d2df Compare November 6, 2025 11:45

mateuszpn temporarily deployed to WindowsCILock November 6, 2025 11:46 — with GitHub Actions Inactive

mateuszpn temporarily deployed to WindowsCILock November 6, 2025 12:08 — with GitHub Actions Inactive

Merge branch 'intel:sycl' into int-buffers

197201e

mateuszpn temporarily deployed to WindowsCILock November 6, 2025 12:27 — with GitHub Actions Inactive

mateuszpn temporarily deployed to WindowsCILock November 6, 2025 12:51 — with GitHub Actions Inactive

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[UR][L0v2 adapter] Fix buffers for integrated gpu #20495

[UR][L0v2 adapter] Fix buffers for integrated gpu #20495

mateuszpn commented Oct 28, 2025 •

edited

Loading

Uh oh!

pbalcer Nov 5, 2025

Uh oh!

mateuszpn Nov 5, 2025

Uh oh!

mateuszpn Nov 6, 2025

Uh oh!

pbalcer Nov 6, 2025

Uh oh!

pbalcer Nov 5, 2025

Uh oh!

mateuszpn Nov 6, 2025

Uh oh!

pbalcer Nov 5, 2025

Uh oh!

mateuszpn Nov 6, 2025

Uh oh!

pbalcer Nov 5, 2025 •

edited

Loading

Uh oh!

pbalcer Nov 5, 2025

Uh oh!

mateuszpn Nov 6, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

[UR][L0v2 adapter] Fix buffers for integrated gpu #20495

Are you sure you want to change the base?

[UR][L0v2 adapter] Fix buffers for integrated gpu #20495

Conversation

mateuszpn commented Oct 28, 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!

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!

pbalcer Nov 5, 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mateuszpn commented Oct 28, 2025 •

edited

Loading

pbalcer Nov 5, 2025 •

edited

Loading