More AMD "GFX940" Enablement Work Landing In LLVM

Earlier this month AMD began publishing code for their "GFX940" graphics block as a new CDNA GPU, presumably what will be the AMD Instinct MI300 series as their next-gen datacenter GPU. More GFX940 open-source driver enablement work is getting underway.

Since that original GFX940 target being added to LLVM's AMDGPU shader compiler back-end near the start of the month, it's been rather quiet until yesterday when another batch of multiple patches hit the LLVM Git code-base. The latest item to note includes formalizing the GFX940 memory model and additional commits working on supporting new opcodes and instructions among other open-source GFX940 patches working their way out. A few days ago were also some new floating point atomics support for GFX940.

When it comes to the GFX940 memory model, it appears to be largely the same as GFX90A for prior CDNA accelerators. One change noted though is, "the gfx940 can be configured as a number of smaller agents with each having a single L2 shared by all CUs on the same agent, or as fewer (possibly one) larger agents with groups of CUs on each agent each sharing separate L2 caches." The other changes in the GFX940 memory model documentation appear to be mostly subtle from my initial look.

In any event look for more AMD GFX940 open-source Linux driver patches to begin flowing soon. On the AMDGPU Linux kernel driver side the new Instinct support like Radeon GPUs is being brought up in their IP-based enumeration approach / block by block.