search

clMathLibraries / clBLAS

a software library containing BLAS functions written in OpenCL
Apache License 2.0
844 stars 237 forks source link

numerical problems with clMagma and clBLAS #8

Closed wznxn closed 10 years ago

wznxn commented 11 years ago

I tested clmagma1.0.0 and clBLAS with AMD FirePro W8000, and got some weird testing results below.

Both MKL 11.0.2.146 and Acml5.3.0 are used to obtain the same problems.

We are wondering whether it is a driver issue. Have you seen this kind of errors before? Any suggestion? Thanks.

./testing_sgemm

M N K clAmdBlas GFLop/s (sec) CPU GFlop/s (sec) error

1024 1024 1024 335.91 ( 0.01) 50.63 ( 0.04) 1.831055e-04 1280 1280 1280 614.83 ( 0.01) 103.87 ( 0.04) 2.746582e-04 1600 1600 1600 605.87 ( 0.01) 138.04 ( 0.06) 3.433228e-04 2000 2000 2000 1198.67 ( 0.01) 136.84 ( 0.12) 4.577637e-04 2500 2500 2500 1173.58 ( 0.03) 37.49 ( 0.83) 8.087158e-04 3125 3125 3125 1250.90 ( 0.05) 151.47 ( 0.40) 1.068115e-03 3906 3906 3906 1226.68 ( 0.10) 166.15 ( 0.72) 1.312256e-03 4882 4882 4882 1250.68 ( 0.19) 194.26 ( 1.20) nan 6102 6102 6102 1260.94 ( 0.36) 217.01 ( 2.09) nan

./testing_dgemm

M N K clAmdBlas GFLop/s (sec) CPU GFlop/s (sec) error

1024 1024 1024 210.07 ( 0.01) 39.60 ( 0.05) 3.126388e-13 1280 1280 1280 334.66 ( 0.01) 55.77 ( 0.08) 5.115908e-13 1600 1600 1600 501.54 ( 0.02) 36.27 ( 0.23) 6.536993e-13 2000 2000 2000 505.43 ( 0.03) 71.73 ( 0.22) 8.242296e-13 2500 2500 2500 482.68 ( 0.06) 83.53 ( 0.37) nan 3125 3125 3125 431.77 ( 0.14) 94.39 ( 0.65) nan 3906 3906 3906 424.89 ( 0.28) 100.95 ( 1.18) nan 4882 4882 4882 419.76 ( 0.55) 112.94 ( 2.06) nan 6102 6102 6102 422.29 ( 1.08) 128.65 ( 3.53) 5.627498e-12

./testing_sgesv_gpu

N NRHS GPU GFlop/s (sec) ||B - AX|| / ||A||*||X||

1024 100 31.37 ( 0.03) 7.35e-08 2048 100 77.00 ( 0.09) 1.35e-07 3072 100 130.27 ( 0.16) 2.02e-07 4032 100 214.66 ( 0.22) 2.27e-07 5184 100 270.23 ( 0.36) nan magma_sposv had error 257. 6384 100 308.80 ( 0.59) nan magma_sposv had error 195. 8385 100 434.03 ( 0.94) nan

According to clinfo, Max memory allocation is 1073741824. Global memory size is 3221225472.

From dmesg: [fglrx] module loaded - fglrx 12.10.5 [May 30 2013]

pavanky commented 11 years ago

@wznxn Can you post the test code as a gist ?

kknox commented 11 years ago

@pavanky The testing_sgemm program is part of the clMagma download suite from the ICL website; we are currently trying to reproduce this issue here in our labs. I will assign the issue to us to represent that we are working on it.

bragadeesh commented 11 years ago

We believe the bug has been fixed in the latest beta public driver: 13.11 Beta. Users running into this issue should try with the beta driver to confirm resolution of this failure.

bragadeesh commented 10 years ago

AMD has released the GA version of Catalyst 13.12. With this driver release, this issue is fixed.