[Question] Blocking calls for data transfers and kernel launch?

[jjfumero]

In the spec, it seems data transfers using this function

zeCommandListAppendMemoryCopy

are not blocking. Is there any variant for blocking calls? Or is there any equivalent to the OpenCL call clEnqueue{Read/Write}Buffer with CL_TRUE to indicate a blocking call?

Similar to the kernel launch, is there any way to specify a blocking call?

What I have found is to close a command lists, and then launch all pending command within the list after each data transfer or kernel launch. For example, by running the following sequence:

zeCommandListAppendMemoryCopy( .. )
zeCommandListClose( .. )
zeCommandQueueExecuteCommandLists ( .. )
zeCommandQueueSynchronize ( ..) 
zeCommandListReset ( .. )

but is there any other way to get blocking calls?

[bmyates]

We don't have an exact equivalent of clEnqueue{Read/Write}Buffer , but there are a few different ways to get the behavior you are looking for. Have you tried immediate command lists in synchronous mode? Something like...

ze_command_queue_desc_t desc = {};
desc.mode = ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS 
zeCommandListCreateImmediate(..., &desc, &hCommandList);
zeCommandListAppendMemoryCopy(... hCommandList)

[jjfumero]

That might work for me. Looking at the spec, immediate command lists are used for low latency, so I think they are even a better fit for what I am looking for. So, with the immediate command lists, using your example:

ze_command_queue_desc_t desc = {};
desc.mode = ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS 
zeCommandListCreateImmediate(..., &desc, &hCommandList);
zeCommandListAppendMemoryCopy(... hCommandList)

// >>>>>>>  At this point of execution, is there any guarantee that the copy is finished?  

Just to give you the context I am working on, I am writing a wrapper for Java, and I need to do blocking calls, otherwise, the Java GC can/might move the objects before the actual copy (data transfer).

[bmyates]

Yeah, in synchronous mode, zeCommandListAppendMemoryCopy(... hCommandList) should block until execution completes.

Per the spec:

ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS = 1 Device execution always completes immediately on execute; Host thread is blocked using wait on implicit synchronization object