You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2020/08/18 21:52:14 UTC

[GitHub] [incubator-mxnet] b-shi commented on issue #17980: When compiled with MKL, fully_connected calls DNNL while dot and batch_dot call MKL

b-shi commented on issue #17980:
URL: https://github.com/apache/incubator-mxnet/issues/17980#issuecomment-675738532


   Hi, we benchmarked gemm performance for those sizes internally using oneDNN vs oneMKL directly and here is what we got:
   
   | Max (20K trials) GFlops | oneMKL (1Th) | oneDNN (1Th) | oneMKL (4Th) | oneDNN (4Th) | oneMKL (8Th) | oneDNN (8Th) | oneMKL (28Th) | oneDNN (28Th) |
   |-------------------------|--------------|--------------|--------------|--------------|--------------|--------------|---------------|---------------|
   | 5, 512, 512               |        34.49 |        40.96 |       174.76 |       262.14 |       291.27 |       436.91 |        374.49 |        524.29 |
   | 5, 512, 1536              |        26.57 |        27.40 |       178.73 |       231.30 |       357.47 |       491.52 |        561.74 |        873.81 |
   | 5, 512, 2048              |        25.83 |        26.55 |       143.64 |       174.76 |       338.25 |       436.91 |        616.81 |       1048.58 |
   | 5, 2048, 512              |        25.51 |        26.82 |       139.81 |       183.96 |       388.36 |       524.29 |        873.81 |       1310.72 |
   | 4, 512, 512               |        31.78 |        42.80 |       161.32 |       233.02 |       262.14 |       419.43 |        299.59 |        419.43 |
   
   | Avg (20K trials) GFlops | oneMKL (1Th) | oneDNN (1Th) | oneMKL (4Th) | oneDNN (4Th) | oneMKL (8Th) | oneDNN (8Th) | oneMKL (28Th) | oneDNN (28Th) |
   |-------------------------|--------------|--------------|--------------|--------------|--------------|--------------|---------------|---------------|
   | 5, 512, 512               |        33.96 |        39.45 |       169.05 |       233.88 |       258.32 |       364.93 |        328.68 |        399.95 |
   | 5, 512, 1536              |        26.30 |        27.13 |       174.04 |       226.75 |       342.51 |       463.81 |        492.14 |        765.09 |
   | 5, 512, 2048              |        25.58 |        26.35 |       140.57 |       171.42 |       328.70 |       429.25 |        589.81 |        873.36 |
   | 5, 2048, 512              |        25.27 |        26.57 |       137.35 |       179.91 |       368.89 |       508.38 |        792.29 |       1101.67 |
   | 4, 512, 512               |        31.02 |        41.91 |       146.59 |       216.20 |       226.68 |       335.54 |        271.98 |        334.98 |
   
   


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org