Any way to know if memcpy is submitted into compute engine or copy engine?

[GHGmc2]

Is there any way to know if memcpy is submitted into compute engine or copy engine?
Or more generally, any env variables or logs on which command queue are command lists submitted to?

[jandres742]

For OpenCL, copy operations use the copy engine if available.
For Level Zero, application needs to use queue groups to select the copy engine. See here https://spec.oneapi.com/level-zero/latest/core/PROG.html#command-queue-groups

For OpenCL, you can use the following debug keys to steer behavior:

EnableBlitterOperationsSupport
EnableBlitterForEnqueueOperations
EnableBlitterForReadWriteImage

compute-runtime/shared/source/debug_settings/debug_variables_base.inl

Line 250 in 233aab8

DECLARE_DEBUG_VARIABLE(int32_t, EnableBlitterOperationsSupport, -1, "-1: default, 0: disable, 1: enable")

[GHGmc2]

Thanks!
On application(framework) side, we do not call Level Zero APIs directly, but through SYCL APIs and DPC++ compiler implementations, any way(with env variables or other ways) to know if we are using copy engine on application side?

[MichalMrozek]

You may also use this variable -> PrintDeviceAndEngineIdOnSubmission it prints following details:

• to what device driver submits
• what is the engine that is getting submission ( compute / copy , etc.)

[GHGmc2]

Thanks, I'll have a try.

[jandres742]

For SYCL APIs when using L0, you need to use SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=1

https://github.com/intel/llvm/blob/06e4ebc73bc9093d2a2b327598833a787b51d2a0/sycl/plugins/level_zero/pi_level_zero.cpp#L548

For SYCL over OpenCL, same rules mentioned before applied. Also, you may use debug key shared by @MichalMrozek