Bug#877316: clblas: Crashes on single-precision-only hardware, due to double-precision literals
Ghislain Vaillant
ghisvail at gmail.com
Wed Oct 4 03:06:18 UTC 2017
Hi Rebecca,
On Sat, 30 Sep 2017 14:47:03 +0100 "Rebecca N. Palmer"
<rebecca_palmer at zoho.com> wrote:
> Package: libclblas2
> Version: 2.12-1
> Control: tags -1 upstream
> Control: affects -1 beignet-opencl-icd
>
> Some clblas operations use '0.0' (a double-precision literal) not '0.0f'
> (a single-precision literal) even when processing single-precision arrays.
>
> This causes it to crash on GPUs that don't support double precision:
>
> ASSERTION FAILED: sel.hasDoubleType()
> at file
> /build/beignet-1.3.1/backend/src/backend/gen_insn_selection.cpp,
> function void
> gbe::ConvertInstructionPattern::convertBetweenFloatDouble(gbe::Selection::Opaque&,
> const gbe::ir::ConvertInstruction&, bool&) const, line 6148
>
> This particular 0.0 appears to have come from
> http://sources.debian.net/src/clblas/2.12-1/src/library/blas/AutoGemm/KernelOpenCL.py/#L368,
> but there may well be more.
>
> This issue also exists in upstream git.
Are you aware of an existing bug filed upstream for this? If so, could
you link it to this bug report with an appropriate forward comand ?
Otherwise, would you please consider filing the bug upstream, since you
already triaged it as non Debian specific.
Thanks for investigating this.
Ghis
More information about the debian-science-maintainers
mailing list