[L0] Disabling Driver In Order Lists by default

[nrspruit]

• Due to L0 Driver issues related to performance using driver in order lists, the feature is being disabled for the time being.
• Once issues with the L0 Drivers implementing this feature are resolved, then this feature will be re-enabled.

[EwanC]

In command-buffer code we use ZE_COMMAND_LIST_FLAG_IN_ORDER command-lists without a precondition on this useDriverInOrderLists() check (it's based on a flag passed from SYCL-RT)

unified-runtime/source/adapters/level_zero/command_buffer.cpp

Line 575 in 3fdf7e3

ZeCommandListDesc.flags = IsInOrder ? ZE_COMMAND_LIST_FLAG_IN_ORDER

Is that something that will suffer from the problem that's motivating this change?

[Bensuo]

In command-buffer code we use ZE_COMMAND_LIST_FLAG_IN_ORDER command-lists without a precondition on this useDriverInOrderLists() check (it's based on a flag passed from SYCL-RT)

unified-runtime/source/adapters/level_zero/command_buffer.cpp

Line 575 in 3fdf7e3

ZeCommandListDesc.flags = IsInOrder ? ZE_COMMAND_LIST_FLAG_IN_ORDER

Is that something that will suffer from the problem that's motivating this change?

We do the same check here that's removed in this PR to confirm whether we can use inorder command lists or not, so I would guess that should also be changed.

[nrspruit]

In command-buffer code we use ZE_COMMAND_LIST_FLAG_IN_ORDER command-lists without a precondition on this useDriverInOrderLists() check (it's based on a flag passed from SYCL-RT)

unified-runtime/source/adapters/level_zero/command_buffer.cpp

Line 575 in 3fdf7e3

ZeCommandListDesc.flags = IsInOrder ? ZE_COMMAND_LIST_FLAG_IN_ORDER

. Is that something that will suffer from the problem that's motivating this change?

If this is based on a flag passed from SYCL RT, then it should not be changed. This is about changing the default behavior to use standard lists and not in order by default.

[nrspruit]

Ah, I missed the "canBeInOrder" let me change that too.

[nrspruit]

@EwanC , please check the command buffer changes, now both are disabled for now.

[nrspruit]

@EwanC , please check the command buffer changes, now both are disabled for now.

@EwanC
had a problem in my patch, please re-check, unfortunately this will disable driver in order in command buffer for now until the driver issues can be resolved.

[nrspruit]

@EwanC / @Bensuo please re-review so we can get this change in as soon as possible to create the release branch.

This updated PR will disable in both common code and command buffers.

[Bensuo]

@EwanC / @Bensuo please re-review so we can get this change in as soon as possible to create the release branch.

This updated PR will disable in both common code and command buffers.

I'm not sure offhand why we don't have this already, but could you add the check for UR_L0_USE_DRIVER_INORDER_LISTS to canBeInOrder() so that we can still enable them if needed? Since that's still in the general implementation I think that should be fine for now.

[nrspruit]

@EwanC / @Bensuo please re-review so we can get this change in as soon as possible to create the release branch.
This updated PR will disable in both common code and command buffers.

I'm not sure offhand why we don't have this already, but could you add the check for UR_L0_USE_DRIVER_INORDER_LISTS to canBeInOrder() so that we can still enable them if needed? Since that's still in the general implementation I think that should be fine for now.

Ok, let me add that in such that in both cases it is "possible" to use the feature.

[Bensuo]

LGTM, thanks for making those changes!

[nrspruit]

@EwanC / @Bensuo please re-review so we can get this change in as soon as possible to create the release branch.
This updated PR will disable in both common code and command buffers.

I'm not sure offhand why we don't have this already, but could you add the check for UR_L0_USE_DRIVER_INORDER_LISTS to canBeInOrder() so that we can still enable them if needed? Since that's still in the general implementation I think that should be fine for now.

Ok, let me add that in such that in both cases it is "possible" to use the feature.

@EwanC , I have updated the check in command buffer such that it also checks the env for driver in order. This will enable both paths to work now.

[nrspruit]

This change is ready to merge, but we are fighting CI issues with a very long delay in execution.

[EwanC]

Sorry @nrspruit I logged off for the day after leaving that comment so didn't see your replies, but Ben has reviewed this on behalf of the command-buffer group and happy with the final patch.