On Mon, Jan 06, 2025 at 02:42:16PM +0100, Uwe Kleine-König wrote: > Control: merge -1 1015871 > > Hello, > > This request came up already in the past, see https://bugs.debian.org/1015871 >
Thank you very much for merging this bug! Actually my request contains two feature, and I'm sorry for mixing them in one bug report: 1. Enabling CONFIG_PCI_P2PDMA, which precisely duplicates bug 1015871. > Back then this looked more like some niche thing that mostly helps people > who need a custom kernel anyhow. Does this new bug report mean that > things have changed? In https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1015871#37 > Anyone using this option to do GPU P2PDMA transfers are certainly using > out of tree (and likely proprietary) modules as the upstream kernel does > not yet appear to support anything like that at this time. My use case is P2P of multiple AMD GPUs, the kernel is not tainted. The amdgpu kernel module in tree make use of P2PDMA to transfer data between GPU. This is important when performing compute on multiple cards. This was also mentioned in https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1015871#52 I'd say multi-GPU setup is still only common in workstation and servers. I guess there is a trend people put more GPU devices into server for accelerated computing. 2. Enabling CONFIG_HSA_AMD_SVM CONFIG_HSA_AMD_SVM is useful in general compute even on single GPU system: > On Mon, Dec 16, 2024 at 12:05:23PM +0800, Yiyang Wu wrote: > > The heterogeneous memory management (HMM), is a mechanism to treat CPU > > memory > > and GPU memory in a joint address space. It is an essential part of Unified > > Memory in heterogeneous computing, which frees the developer from manually > > handling data migration across different types of memories. Data will be > > automatically transferred and cached into the memory of the compute node. > > This > > programming model is gaining popularity, and for AMD GPU and ROCm, cards > > that > > support xnack replay can utilize this model if CONFIG_HSA_AMD_SVM=y This is also enabled in Fedora Linux 41's kernel-x86_64-fedora.config > > Best regards > Uwe Best regards, Yiyang
