gonum / blas / 9

44%

Pull #22

travis-ci

Simplify dgemm to use fewer goroutines and fewer messages.

The old code was wasteful. It sent 2 messages every block computation, and it launched a new goroutine at each block computation. This was done to synchronize the computations along the k dimension, but this is silly, because the k dimension is being computed in serial anyway. This refactors the code such that the k dimension is computed in serial by a worker before a new {i,j} pair is received. This reduced the messages sent, reduces the number of goroutines, and should help cache coherency. This seems to improve performance somewhat, but even if the performance gains were negligible, the simplification makes it worth it. When the inner loop is sped up (bounds checking, vectorization), this code will almost certainly be faster.

BenchmarkDgemmSmSmSm              2372           2313           -2.49%
BenchmarkDgemmSmSmSm-2            2280           2267           -0.57%
BenchmarkDgemmSmSmSm-4            2290           2295           +0.22%
BenchmarkDgemmMedMedMed           1302412        1293271        -0.70%
BenchmarkDgemmMedMedMed-2         723841         717892         -0.82%
BenchmarkDgemmMedMedMed-4         563282         575358         +2.14%
BenchmarkDgemmMedLgMed            12601547       12633168       +0.25%
BenchmarkDgemmMedLgMed-2          6668578        6727320        +0.88%
BenchmarkDgemmMedLgMed-4          4005838        4029760        +0.60%
BenchmarkDgemmLgLgLg              1284515732     1317360458     +2.56%
BenchmarkDgemmLgLgLg-2            663440326      653124639      -1.55%
BenchmarkDgemmLgLgLg-4            442994398      443350897      +0.08%
BenchmarkDgemmLgSmLg              15146665       15495804       +2.31%
BenchmarkDgemmLgSmLg-2            7894837        7818921        -0.96%
BenchmarkDgemmLgSmLg-4            4770521        5022796        +5.29%
BenchmarkDgemmLgLgSm              15419348       14768983       -4.22%
BenchmarkDgemmLgLgSm-2            7854456        ... (continued)

57 of 57 new or added lines in 1 file covered. (100.0%)

1339 of 4997 relevant lines covered (26.8%)

0.54 hits per line