NVK Vulkan Driver Adds Implicit Pipeline Caching To Boost DXVK Performance

[phoronix]

Phoronix: NVK Vulkan Driver Adds Implicit Pipeline Caching To Boost DXVK Performance

Adding to the impressive number of features to be found in this quarter's Mesa 24.1 release is now the open-source NVIDIA "NVK" Vulkan driver supporting implicit pipeline caching...

https://www.phoronix.com/news/NVK-Implicit-Pipeline-Caching

[Developer12]

Looking forward to the glorious Nova explicit-sync driver future, with NVK vulkan, Zink OpenGL, and rusticl OpenCL.

[oiaohm]

Looking forward to the glorious Nova explicit-sync driver future, with NVK vulkan, Zink OpenGL, and rusticl OpenCL.

Redhat developed Nova is going to support the kernel provided implicit sync interfaces

Its one thing to say explicit sync is the future. Its like here with DXVK there are particular areas of your driver you need implicit synced because the applications should not have access to particular information so you need some third party part taking care of that by implicit means.

card to kernel this can be all explicit sync. kernel to application some of this has to be implicit sync because OS security and stability forbids giving direct control of different things to applications. The kernel to application restrictions have been something Nvidia driver developers have tried to avoid having to obey.

We don't need another mistake like eglstreams where a party tries to use more explicit sync than what it sane to-do across the kernel to application bridge.

[Developer12]

Redhat developed Nova is going to support the kernel provided implicit sync interfaces

Its one thing to say explicit sync is the future. Its like here with DXVK there are particular areas of your driver you need implicit synced because the applications should not have access to particular information so you need some third party part taking care of that by implicit means.

card to kernel this can be all explicit sync. kernel to application some of this has to be implicit sync because OS security and stability forbids giving direct control of different things to applications. The kernel to application restrictions have been something Nvidia driver developers have tried to avoid having to obey.

We don't need another mistake like eglstreams where a party tries to use more explicit sync than what it sane to-do across the kernel to application bridge.

Everything you said is irrelevant. I've already done the legwork you haven't and asked directly how the Nova driver will be architected. It will be an explicit-sync-only driver in the same vein as AGX and the intel Xe driver.

[oiaohm]

Everything you said is irrelevant. I've already done the legwork you haven't and asked directly how the Nova driver will be architected. It will be an explicit-sync-only driver in the same vein as AGX and the intel Xe driver.

Redhat developed Nova is going to support the kernel provided implicit sync interfaces

Read that line more carefully. I did not say that the driver would be implicit sync. DRM/DRI layer in the kernel does have implicit sync on top of explicit sync to provide implicit sync into userspace where it needed.

Its like how zink is able to do implicit sync on top of explicit sync as well.

card to kernel this can be all explicit sync

I also wrote that in the hope it would also be clear. The question is not if the driver is explicit sync. The question is will the driver support the kernel wrapper to implicit sync for what it need to be between the kernel and the application.

Yes question is the driver explicit sync only gives you half the answer. You got the answer driver explicit sync that requires asking the next question. Is this driver going to support generic interface to provide implicit sync to legacy application and to perform implicit sync where need for stable application to/from kernel communication? Nova and AGX this is yes. Nvidia own open source module the next question after is this explicit sync is no I am not supporting the implicit sync emulation when it needed/requested.

I did a little more leg work than you I asked one more question.

[Developer12]

Redhat developed Nova is going to support the kernel provided implicit sync interfaces

Read that line more carefully. I did not say that the driver would be implicit sync. DRM/DRI layer in the kernel does have implicit sync on top of explicit sync to provide implicit sync into userspace where it needed.

Its like how zink is able to do implicit sync on top of explicit sync as well.

card to kernel this can be all explicit sync

I also wrote that in the hope it would also be clear. The question is not if the driver is explicit sync. The question is will the driver support the kernel wrapper to implicit sync for what it need to be between the kernel and the application.

Yes question is the driver explicit sync only gives you half the answer. You got the answer driver explicit sync that requires asking the next question. Is this driver going to support generic interface to provide implicit sync to legacy application and to perform implicit sync where need for stable application to/from kernel communication? Nova and AGX this is yes. Nvidia own open source module the next question after is this explicit sync is no I am not supporting the implicit sync emulation when it needed/requested.

I did a little more leg work than you I asked one more question.

Maybe you should ponder the implications (or lack thereof) of what I've said in this thread before you waste everyone's time with mountains of text.

[JoshuaAshton]

Redhat developed Nova is going to support the kernel provided implicit sync interfaces

Its one thing to say explicit sync is the future. Its like here with DXVK there are particular areas of your driver you need implicit synced because the applications should not have access to particular information so you need some third party part taking care of that by implicit means.

card to kernel this can be all explicit sync. kernel to application some of this has to be implicit sync because OS security and stability forbids giving direct control of different things to applications. The kernel to application restrictions have been something Nvidia driver developers have tried to avoid having to obey.

We don't need another mistake like eglstreams where a party tries to use more explicit sync than what it sane to-do across the kernel to application bridge.

I have no idea if the driver will support implicit sync or not but the rest of what you wrote is just complete gibberish and you have no idea what you are talking about.

[oiaohm]

I have no idea if the driver will support implicit sync or not but the rest of what you wrote is just complete gibberish and you have no idea what you are talking about.

https://lore.kernel.org/lkml/[email protected]/

Its not that I don't have idea what I am talking about.

Its that it not as straight forwards as one would think. The kernel need to know what it need to do if application has been force terminated. Percentage of operations have to be implicit sync with the kernel so the kernel has a list to clean up in case of application forced termination. There are other examples like this.

DMABUF section of the Linux kernel does implement some implicit sync on top of a explicit sync driver for mainline drivers.