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

via llvm-commits llvm-commits at lists.llvm.org
Thu Jun 11 07:49:10 PDT 2026


clearnature wrote:

@Malexandra-de The 4326 TOPs number is not from the scheduling model alone — it's                                                                                                                                                                  
 from the K9 kernel design that uses staggered wave scheduling across                                                                                                                                                                
 the entire wavefront pool. The latency table enables the scheduler                                                                                                                                                                  
 to make correct decisions about wave interleaving, but the throughput                                                                                                                                                               
 comes from the kernel structure, not from the scheduling model.                                                                                                                                                                     
                                                                                                                                                                                                                                     
 Without the latency table, the scheduler defaults to a 5-cycle 
 estimate for matrix ops, which causes it to serialize independent 
 wavefronts. The table corrects this to the real 8/16-cycle latency, 
 which is all that's needed to allow the kernel's staggered wave 
 design to work. 
  
 Regarding physical ALU count: the 4326 figure is application-level 
 throughput measured end-to-end (entire sparse matmul kernel including 
 matrix loads, transforms, and stores), not raw ALU utilization. I'm 
 happy to share the benchmark methodology and measurement code. 

          he kernel design is already submitted upstream: rocWMMA SWMMAC                                                                                                                                                                     
 backend (#8089) and FP4 type definitions (#8146) in ROCm/rocm-libraries.                                                                                                                                                            
 The 4326 TOPs measurement is from this public kernel code running on                                                                                                                                                                
 production RX 9060 XT hardware (gfx1200).                                                                                                                                                                                           
                                      
 The latency table in this PR enables the scheduler to correctly                                                                                                                                                                     
 execute the wave-staggered kernel design. Without it, the scheduler                                                                                                                                                                 
 defaults to wrong latency assumptions and serializes wavefronts.                                                                                                                                                                    
                                                                                                                                                                                                                                     
 The code and benchmarks are public — anyone with a gfx1200 GPU can 
 reproduce the numbers. 


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


More information about the llvm-commits mailing list