Bug#881054: libarrayfire-opencl3: "INTERNAL KERNEL BUILD ERROR" from af::matmulTN

[Ralf Stubner]

On 08.11.2017 15:38, Ralf Stubner wrote:
> However, I have just found a very similar issue
> where clBLAS is named as the source of the error:
> https://github.com/Theano/libgpuarray/issues/216#issuecomment-232225071

I have found another similar bug report:
https://github.com/clMathLibraries/clBLAS/issues/20

Following the suggestions there I have now build a custom version of
libclblas2 with the patch:

--- clblas-2.12.orig/src/library/blas/generic/common.c
+++ clblas-2.12/src/library/blas/generic/common.c
@@ -539,8 +539,8 @@ setupBuildOpts(

 #if !defined NDEBUG
     // Nvidia runtime does not appear to support the -g flag, at least
in their OpenCL v1.1 runtime
-    if( target.ident.vendor != VENDOR_NVIDIA )
-        addBuildOpt( opts, BUILD_OPTS_MAXLEN, "-g" );
+    //if( target.ident.vendor != VENDOR_NVIDIA )
+    //    addBuildOpt( opts, BUILD_OPTS_MAXLEN, "-g" );
 #endif  /* NDEBUG */

     if (target.ident.vendor == VENDOR_NVIDIA &&

[Of course, this patch does not make sense for general consumption.]

With this patched libclblas2 installed, my fastLm program works also
with libarrayfire-opencl3.

> The binary distribution from upstream is not linked with the system
> clBLAS. It might actually use clBlast instead (cf
> https://arrayfire.com/arrayfire-v3-5-official-release/).

It is interesting to note that the clBLAS used by upstream contains the
same code:

https://github.com/arrayfire/clBLAS/blob/arrayfire-release/src/library/blas/generic/common.c#L540-L544

However, it seems that they define NDEBUG during compilation, so that
the relevant code parts are never triggered. Maybe defining NDEBUG would
be the right patch for libclblas?

> In the end
> latest upstream version might solve this issue, but I am not sure.

I meanwhile think that it probably would not help. However, upstream's
binary version is quite a bit faster on my hardware, so an updated
version would still be appreciated.

cheerio
ralf

-- 
Ralf Stubner
Senior Software Engineer / Trainer

R Institute GmbH
Dortustraße 48
14467 Potsdam

T: +49 331 23 70 81 66
F: +49 331 23 70 81 67
M: +49 162 20 91 196

Mail: ralf.stubner at r-institute.com

Sitz: Potsdam
Register: AG Potsdam HRB 27966 P
Ust.-IdNr.: DE300072622
Geschäftsführer: Prof. Dr. Dr. Karl-Kuno Kunze

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4008 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.alioth.debian.org/pipermail/debian-science-maintainers/attachments/20171108/0512f06d/attachment.bin>