radv and the vulkan deferred demo - no fps left behind!

A little while back I took to wondering why one particular demo from the Sascha Willems vulkan demos was a lot slower on radv compared to amdgpu-pro. Like half the speed slow.

I internally titled this my "no fps left behind" project.

The deferred demo, does an offscreen rendering to 3 2048x2048 color attachments and one 2048x2048 D32S8 depth attachment. It then does a rendering using those down to as 1280x720 screen image.

Bas identifed the first cause was probably the fact we were doing clear color eliminations on the offscreen surfaces when we didn't need to. AMD GPU have a delta-color compression feature, and with certain clear values you don't need to do the clear color eliminations step. This brought me back from about 1/2 the FPS to about 3/4, however it took me quite a while to figure out where the rest of the FPS were hiding.

I took a few diversions in my testing, I pulled in some experimental patches to allow the depth buffer to be texture cache compatible, so could bypass the depth decompression pass, however this didn't seem to budge the number too much.

I found a bunch of registers we were setting different values from -pro, nothing too much came of these.

I found some places we were using a compute shader to fill some DCC or htile surfaces to a value, then doing a clear and overwriting the values, not much help.

I noticed the vertex descriptions and buffer attachments on amdgpu-pro were done quite different to how radv does it. With vulkan you have vertex descriptors and bindings, with radv we generate a set of hw descriptors from the combination of both descriptors and bindings. The pro driver uses typed buffer loads in the shader to embed the descriptor contents in the shader, then it only updates the hw descriptors for the buffer bindings. This seems like it might be more efficient, guess what, no help. (LLVM just grew support for typed buffer loads, so we could probably move to this scheme if we wished now).

I dug out some patches that inline all the push constants and some descriptors so our shaders had less overhead, (really helps our meta shaders have less impact), no helps.

I noticed they export the shader results in a different order from the fragment shader, and always at the end. (no help). The vertex shader emits pos first, (no help). The vertex shader uses off exports for unused channels, (no help).

I went on holidays for a week and came back to stare at the traces again, when I my brain finally noticed something I'd missed. When binding the 3 color buffers, the addresses given as the base address were unusual. A surface has a 40-bit address, normally for alignment and tiling the bottom 16-bits are 0, and we shift 8 of those off completely before writing them. This leaves the bottom 8 bits of the base address has should be 0, and the CIK docs from AMD say that. However the pro traces didn't have these at 0. It appears from earlier evergreen/cayman documents these register control some tiling offset bits. After writing a hacky patch to set the values, I managed to get back the rest of the FPS I was missing in the deferred demo. I discussed with AMD developers, and we worked out the addrlib library has an API for working out these values, and it seems that it allows better memory bandwidth utilisation. I've written a patch to try and use these values correctly and sent it out along with the DCC avoidance patch.

Now I'm not sure this will help any real apps, we may not be hitting limitations in that area, and I'm never happy with the benchmarks I run myself. I thought I saw some FPS difference with some madmax scenes, but I might be lying to myself. Once the patches land in mesa I'm sure others will run benchmarks and we can see if there is any use case where they have an effect. The AMD radeonsi OpenGL driver can also do the same tweaks so hopefully there as well there will be some benefit.

Otherwise I can just write this off as making deferred run at equality and removing at least one of the deltas that radv has compared to the pro driver. Some of the other differences I discovered along the way might also have some promise in other scenarios, so I'll keep an eye on them.

Thanks to Bas, Marek and Christian for looking into what the magic meant!