Prototype new command-list submission entrypoint #1975
Conversation
github-actions
bot
added
level-zero
| urCommandBufferCreateExp(ur_context_handle_t Context, ur_device_handle_t Device, | ||
| const ur_exp_command_buffer_desc_t *CommandBufferDesc, | ||
| ur_exp_command_buffer_handle_t *CommandBuffer) { | ||
| if (Device->isPVC() && Device->ImmCommandListUsed) { |
There was a problem hiding this comment.
Should probably check for the existance of the extension here in the driver instead, something like
if (!CommandBuffer->Context->getPlatform()
->ZeDriverAppendImmediateCommandListExtensionFound)
There was a problem hiding this comment.
I'm checking the driver version in the latest commit. But this is definitely better. Will update the code to do it in this way.
There was a problem hiding this comment.
I implemented this but it seems that the level-zero driver is not yet reporting the extension correctly. I left a TODO to switch to this later.
fabiomestre
force-pushed
the
fabio/immediate_append_exp
branch
2 times, most recently
from
12fcfa5
to
6c0b66c
Compare
Adds a new path that uses the zeCommandListImmediateAppendCommandListsExp to submit command-buffers on PVC hardware.
fabiomestre
force-pushed
the
fabio/immediate_append_exp
branch
from
6c0b66c
to
a27ba98
Compare
| bool IsInOrder = canBeInOrder(Context, CommandBufferDesc); | ||
| bool EnableProfiling = | ||
| CommandBufferDesc && CommandBufferDesc->enableProfiling; | ||
| CommandBufferDesc && CommandBufferDesc->enableProfiling && !IsInOrder; |
There was a problem hiding this comment.
Does this change mean we can't profile in-order command-lists?
There was a problem hiding this comment.
In-order command-lists don't have sync-point events. So there is nothing to profile. This check already existed on the original implementation, but it was only checked during enqueue with:
const bool DoProfiling =
(UrQueue->Properties & UR_QUEUE_FLAG_PROFILING_ENABLE) &&
(!CommandBuffer->IsInOrderCmdList) &&
(CommandBuffer->IsProfilingEnabled) && Event;
Work in progress. Waiting for the new level-zero entrypoint to work before continuing.