[311Volt]

Liboffload implementations of QueueMemcpy, QueueInOrderMemcpy, UsmFillImmediate, UsmMemoryAllocation and UsmBatchMemoryAllocation have been merged to Compute Benchmarks. Run the newly-merged liboffload implementations of the memcpy, USM allocation and USM fill Compute Benchmarks in the "SYCL Run Benchmarks" nightly job, next to their existing baselines if applicable.

For the fill and memcpy benchmarks, add Offload instances against existing SYCL/UR baselines where possible. Additionally, add L0 instances for UsmMemoryAllocation and QueueInOrderMemcpy, leveraging previously existing Compute Benchmarks implementations, to better gauge API overhead of the other higher-level APIs. Rewrite the memcpy classes to no longer hardcode SYCL and be parametrized on the API instead.

Add benchmark instances for UsmFillImmediate - L0 and OL with a 4-byte pattern, and OL with a 3-byte pattern, on a 4 KiB and a 512 MiB device buffer. The intent is to exercise liboffload's non-pow2 pattern support, which may use a slow path underneath (on L0 plugin it appends O(log n) memcpy commands). Small fills report time in µs, large fills report bandwidth in GB/s.

Bump Compute Benchmarks to de729a5340ac (Oct 30, 2026) to pick up the new liboffload benchmarks. The new version requires --strategy for UsmMemoryAllocation. Pass --strategy=sync, to preserve pre-bump behavior.

Full testing matrix for affected benchmarks after the patch is provided as a PR comment.

[311Volt]

Full testing params matrix

• "T+C" = TIMER + CPU_COUNTER profilers (label suffix CPU count).
• Results are µs (CPU count: instr), lower is better, except the large fills: GB/s, higher is better.

A. Memcpy: memory_benchmark_{sycl,l0,ol}

QueueMemcpy: --iterations=10000 --sourcePlacement --destinationPlacement --size --profilerType. No L0 implementation.

Runtime	src -> dst	size	profilers	status
SYCL	Device -> Device	1024	T+C	existing
OL	Device -> Device	1024	T+C	new

QueueInOrderMemcpy. --iterations=10000 --IsCopyOnly=0 --count=100 --withCopyOffload=0 --profilerType, plus --doNotPrintBandwidth on L0.

Runtime	src -> dst	size	profilers	status
SYCL	Device -> Device	1024	T+C	existing
SYCL	Host -> Device	1024	T+C	existing
L0	Device -> Device	1024	TIMER	new
L0	Host -> Device	1024	TIMER	new
OL	Device -> Device	1024	T+C	new
OL	Host -> Device	1024	T+C	new

Total: 14 instances, 8 new.

B. Alloc: api_overhead_benchmark_{ur,l0,ol}

UsmMemoryAllocation: --iterations=10000 --measureMode=Both --strategy=sync.

Runtime	type	size	status
UR	Device	256, 262144	existing
L0	Device	256, 262144	new
OL	Device	256, 262144	new

UsmBatchMemoryAllocation: --iterations=1000 --measureMode=Both. No L0 implementation.

Runtime	type	allocationCount	size	status
UR	Device	128	256, 16384, 131072	existing
OL	Device	128	256, 16384, 131072	new

Total: 12 instances, 7 new. Groups: UsmMemoryAllocation, UsmBatchMemoryAllocation.

C. Fill: memory_benchmark_{l0,ol} UsmFillImmediate

All rows: --memory=Device --contents=Zeros --patternContents=Random --forceBlitter=0 --useEvents=0. Small rows also pass --doNotPrintBandwidth.

Variant	Runtime	patternSize	small (nominal 4096)	large (nominal 536870912)	liboffload path
L0	L0	4	4096	536870912	n/a (native L0)
OL	OL	4	4096	536870912	native fill
OL non-pow2	OL	3	4095 (truncate 1)	536870913 (pad 1)	replicate

size class	--iterations	metric	unit	better	group
small	10000	time per fill (host-timed, incl. sync)	µs	lower	UsmFillImmediate Device, small fill
large	100	fill bandwidth (size / host time)	GB/s	higher	UsmFillImmediate Device, large fill

Total: 6 instances, all new.

[lukaszstolarczuk]

LGTM

[lukaszstolarczuk]

LGTM