Use Dpp::Zip instead of blend-based zip to save registers - Githubissues

ROCm / rocWMMA

rocWMMA

https://rocm.docs.amd.com/projects/rocWMMA/

MIT License

91 stars 26 forks source link

Use Dpp::Zip instead of blend-based zip to save registers #376

Closed cgmillette closed 7 months ago

cgmillette commented 7 months ago

Updates documentation of cross-lane operations
Adds Zip4/8/16/32 functionality to DPP
DPP doesn't require conditional mask, so we can save registers