search

bluss / matrixmultiply

General matrix multiplication of f32 and f64 matrices in Rust. Supports matrices with general strides.
https://docs.rs/matrixmultiply/
Apache License 2.0
213 stars 25 forks source link

Use a 4-by-8 microkernel for sgemm #4

Closed bluss closed 8 years ago

bluss commented 8 years ago

Use a 4-by-8 microkernel for sgemm

Processing more data per kernel is a win. Some small sizes that need more zero padding regress by a few percent.

name                           sgemm4x4.log ns/iter  sgemm4x8.log ns/iter    diff ns/iter   diff %
mat_mul_f32::m004              153                   172                               19   12.42%
mat_mul_f32::m007              280                   271                               -9   -3.21%
mat_mul_f32::m008              305                   286                              -19   -6.23%
mat_mul_f32::m012              598                   650                               52    8.70%
mat_mul_f32::m016              1,044                 945                              -99   -9.48%
mat_mul_f32::m032              5,037                 4,638                           -399   -7.92%
mat_mul_f32::m064              30,305                27,748                        -2,557   -8.44%
mat_mul_f32::m127              208,380               188,977                      -19,403   -9.31%
mat_mul_f32::mix128x10000x128  16,291,680            14,293,288                -1,998,392  -12.27%
nonative_mat_mul_f32::m127     291,936               272,487                      -19,449   -6.66%