Dave Airlie Linux Graphics blog

 Sriram invited me to the oneAPI meetup, and I felt I hadn't summed up the state of compute and community development in a while. Enjoy 45 minutes of opinions! https://www.youtube.com/watch?v=HzzLY5TdnZo

The initial NVK (nouveau vulkan) experimental driver has been merged into mesa master[1], and although there's lots of work to be done before it's application ready, the main reason it was merged was because the initial kernel work needed was merged into drm-misc-next[2] and will then go to drm-next for the 6.6 merge window. (This work is separate from the GSP firmware enablement required for reclocking, that is a parallel development, needed to make nvk useable). Faith at Collabora will have a blog post about the Mesa side, this is more about the kernel journey. What was needed in the kernel? The nouveau kernel API was written 10 years or more ago, and was designed around OpenGL at the time. There were two major restrictions in the current uAPI that made it unsuitable for Vulkan. buffer objects (physical memory allocations) were allocated 1:1 with virtual memory allocations for a file descriptor. This meant the kernel managed the virtual address space. For proper Vulkan suppor

I recently came across tinygrad as a small powerful nn framework that had an OpenCL backend target and could run LLaMA model. I've been looking out for rusticl workloads, and this seemed like a good one, and I could jump on the AI train, and run an LLM in my house! I started it going on my Radeon 6700XT with the latest rusticl using radeonsi with the LLVM backend, and I could slowly interrogate a model with a question, and it would respond. I've no idea how performant it is vs ROCm yet which seems to be where tinygrad is more directed, but I may get to that next week. While I was there though I decided to give the Mesa ACO compiler backend a go, it's been tied into radeonsi recently, and I done some hacks before to get compute kernels to run. I reproduced said hacks on the modern code and gave it a run. tinygrad comes with a benchmark script called benchmark_train_efficientnet so I started playing with it to see what low hanging fruit I could find in an LLVM vs ACO shootout

 Thanks for all the suggestions, on here, and on twitter and on mastodon, anyway who noted I could use a single fd and avoid all the pain was correct! I hacked up an ever growing ftruncate/madvise memfd and it seemed to work fine. In order to use it for sparse I have to use it for all device memory allocations in lavapipe which means if I push forward I probably have to prove it works and scales a bit better to myself. I suspect layering some of the pb bufmgr code on top of an ever growing fd might work, or maybe just having multiple 2GB buffers might be enough. Not sure how best to do shaderResourceResidency, userfaultfd might be somewhat useful, mapping with PROT_NONE and then using write(2) to get a -EFAULT is also promising, but I'm not sure how best to avoid segfaults for read/writes to PROT_NONE regions. Once I got that going, though I ran headfirst into something that should have been obvious to me, but I hadn't thought through. llvmpipe allocates all it's textures l

Mike nerdsniped me into wondering how hard sparse memory support would be in lavapipe. The answer is unfortunately extremely. Sparse binding essentially allows creating a vulkan buffer/image of a certain size, then plugging in chunks of memory to back it in page-size multiple chunks. This works great with GPU APIs where we've designed this, but it's actually hard to pull off on the CPU. Currently lavapipe allocates memory with an aligned malloc. It allocates objects with no backing and non-sparse bindings connect objects to the malloced memory. However with sparse objects, the object creation should allocate a chunk of virtual memory space, then sparse binding should bind allocated device memory into the virtual memory space. Except Linux has no interfaces for doing this without using a file descriptor. You can't mmap a chunk of anonymous memory that you allocated with malloc to another location. So if I malloc backing memory A at 0x1234000, but the virtual memory I've

F38 just released and seeing a bunch of people complain that TF2 dies on AMD or other platforms when lavapipe is installed. Who's at fault? I've no real idea. How to fix it? I've no real idea. What's happening? AMD OpenGL drivers use LLVM as the backend compiler. Fedora 38 updated to LLVM 16. LLVM 16 is built with c++17 by default. C++17 introduces new "operator new/delete" interfaces[1]. TF2 ships with it's own libtcmalloc_minimal.so implementation, tcmalloc expects to replace all the new/delete interfaces, but the version in TF2 must not support or had incorrect support for the new align interfaces. What happens is when TF2 probes OpenGL and LLVM is loaded, when DenseMap initializes, one "new" path fails to go into tcmalloc, but the "delete" path does, and this causes tcmalloc to explode with "src/tcmalloc.cc:278] Attempt to free invalid pointer" Fixing it? I'll talk to Valve and see if we can work out something, LLVM 16