[matmul] Quantify various techniques we are looking

nod-ai / SHARK-ModelDev

Unified compiler/runtime for interfacing with PyTorch Dynamo.

Apache License 2.0

95 stars 48 forks source link

[matmul] Quantify various techniques we are looking #703

Closed antiagainst closed 3 months ago

antiagainst commented 6 months ago

Advanced instruction scheduling in llvm amdgpu backend #515
Advanced software pipelining #686
etc.

antiagainst commented 6 months ago

Update 5/23: Expected impact on scheduling. Turning off gives large difference on a specific shape. Need to further play with it.

antiagainst commented 6 months ago

Update 5/28: will try to use CK to turn on/off scheduling intrinsic features to get a sense of perf impacts. Update 5/30: turning on/off CK llvm scheduling barrier. ~20%. @Groverkss to quantify advanced software pipelining expectations.

antiagainst commented 3 months ago

Closing this--will have more clear separate tracking tasks