00:42 <HdkR> macc24: Sounds like you want mesh shaders then :P

09:38 <macc24> HdkR: it was easier for me to write a doom-style ray marching renderer than get anything opengl working

09:45 <HdkR> Sounds reasonable

10:00 <robmur01> "are you saying you would rather have t620 drivers than drivers for the mt8192?"

10:03 <macc24> robmur01: for the sake of "every midgard and bifrost mali supported" :P

10:16 <rasterman> valhall support will come ... it is the sound of inevitability...

10:16 <rasterman> https://www.youtube.com/watch?v=x5m1A7zoIcc

10:30 <HdkR> I'm looking forward to that :)

10:54 <macc24> i'm looking forward to seeing mainline linux run on mt8192 somewhat usably :D

13:48 <alyssa> bbrezillon: currently adding valhall support to the blitter

13:48 <alyssa> Valhall doesn't have RSDs, instead it has much smaller shader descriptors and separate depth/stencil and blend descriptors and a ton of state in the draw call descriptor

13:49 <alyssa> so looking at the functions to get RSDs

13:49 <alyssa> I see that, not only do we cache blit shaders, but also cache blit RSDs on top of that

13:49 <alyssa> I assume this was performance minded?

13:50 <alyssa> I'm wondering

13:50 <alyssa> 1) whether rebuilding the RSD is actually more expensive than rebuilding and hashing the key

13:50 <alyssa> 2) whether it matters at all

13:51 <alyssa> on one hand, RSDs are quite large

13:51 <alyssa> on the other hand, the issues with rebuilding RSDs come when issues tens of thousands of draws per second

13:52 <alyssa> for GLES, the blit infrastructure is only used for preframe shaders, i.e. constant in the number of render pass submissions (the overhead head is dominated by the kernel submission)

13:52 <alyssa> for VK, it's used in full so presumably panvk is much more sensitive to the overhead

13:55 <bbrezillon> alyssa: I don't recall if blitter RSD are put in GPU buffers directly, or if it's just a template that needs to be copied to the descriptor pool

13:58 <bbrezillon> if this is the latter, then I guess rebuilding the RSD is not a big deal, since I suspect most of the time is spent writing to this uncached memory buffer anyway

14:00 <bbrezillon> but it's things are prepared once in GPU mem, and re-used when we hit a cache entry, the cost is likely to be higher (for both memory and CPU usage)

14:03 <bbrezillon> alyssa: IIUC, you'd still benefit from re-using the shader descriptor if the blit shader already exists in the blitter cache, even for valhall, right?

14:04 <bbrezillon> or is the shader descriptor directly embedded in job?

14:24 <alyssa> bbrezillon: Fair enough

14:24 <alyssa> Reusing the shader descirptor would benefit on Valhall, yes

14:24 <alyssa> but the logical place to cache that is the blit shader cache, not the blit RSD cache

14:28 <bbrezillon> guess you can add an extra 'sd' (shader descriptor) field to the blit_shader_data struct so that you save a hash_table search when arch==valhall

14:31 <bbrezillon> actually, all you'll need is the shader descriptor in that case, so you can even re-use pan_blit_shader_data.address

15:08 <alyssa> ah true

22:32 <alyssa> making progress on the valhall port of pan_blitter

22:32 <alyssa> admittedly e getting pretty messy :|