[llvm] [AMDGPU] Enable WMMA256bInsts + Wave32 for gfx1200/gfx1201 + SISchedule fix + TargetParser gfx1200 propagation (PR #202093)

[via llvm-commits]

clearnature wrote:

One area prone to issues is the power stability and quality of data center GPUs, which differ significantly from their consumer-grade counterparts; factors such as regional voltage and frequency variations can impact the hardware. As manufacturing processes shrink to the nanometer scale, the electron tunneling effect becomes increasingly pronounced. It is necessary to acknowledge the technological differences between AMD and NVIDIA. Reverse-engineering microarchitectures—such as cracking the "black box" of CUDA's NVPTX—is a topic frequently discussed online, yet the process is undeniably difficult; for instance, AMD's approach to sparse matrix operations via outer-product engines differs fundamentally from traditional methods. Obtaining the necessary data for analysis is a challenging task. GPU matrix computation provides data that cannot be faked, making it an ideal benchmark for testing and ensuring deterministic performance.

https://github.com/llvm/llvm-project/pull/202093