onnxruntime
05fc0c60
- [MLAS][AArch64] SQNBitGemm CompInt8 - Use 4x2 tiles (#21380)
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
1 year ago
[MLAS][AArch64] SQNBitGemm CompInt8 - Use 4x2 tiles (#21380) Update SQNBitGemm ARM NEON kernel to compute 4x2 tile of output. Note: Also tried 2x4 and 4x4 tiles but observed the best microbenchmark results with 4x2 tiles.
References
#21380 - [MLAS][AArch64] SQNBitGemm CompInt8 - Use 4x2 tiles
Author
edgchen1
Parents
92f66de7
Loading