Package: clblas Version: 2.12-5 Severity: important Tags: patch
Executing any of the clblas.gemm functions in libclBLAS.so.2.12.0 (libclblas2 2.12-1build1 on bionic) aborts the calling program with this error message: OpenCL error -11 [...] clblas-2.12/src/library/blas/xgemm.cc:244: void makeGemmKernel(_cl_kernel**, cl_command_queue, const char*, const char*, const unsigned char**, size_t*, const char*): Assertion `false' failed. The error occurs when the backend is an OpenCL 1.1 device. This includes all GPUs supported by libclc through mesa-opencl-icd (Clover), that is effectively all AMD GPUs. The error can be easily reproduced by running the example from the upstream README.md[1] on a machine with an AMD GPU, with environment entry OCL_ICD_VENDORS=mesa.icd. This bug also affects all BLAS functionality in libarrayfire-opencl3, as can be seen by running e.g. examples/benchmarks/blas or any of the lin_algebra examples from the libarrayfire-doc package. The good news is, there is a fix :-) I have submitted this PR[2] upstream, and am attaching a dquilt-ready patch as upstream releases (and even pulls to master) are very infrequent. [1] https://github.com/clMathLibraries/clBLAS#example [2] https://github.com/clMathLibraries/clBLAS/pull/342 from: https://bugs.launchpad.net/ubuntu/+source/clblas/+bug/1816887 diff -Nru clblas-2.12/debian/patches/342.patch clblas-2.12/debian/patches/342.patch --- clblas-2.12/debian/patches/342.patch 1970-01-01 01:00:00.000000000 +0100 +++ clblas-2.12/debian/patches/342.patch 2024-08-02 09:32:06.000000000 +0200 @@ -0,0 +1,34 @@ +From 8dd6df3f1bf543db3765383b4f9646df50d23d08 Mon Sep 17 00:00:00 2001 +From: Marco van Zwetselaar <[email protected]> +Date: Wed, 20 Feb 2019 15:48:19 +0300 +Subject: [PATCH] Do not pass the -cl-std option to kernel builds + +- Resolves the abort in xgemm.cc:244 for OpenCL 1.1 devices +--- + src/library/blas/AutoGemm/Includes.py | 6 +++--- + 1 file changed, 3 insertions(+), 3 deletions(-) + +diff --git a/src/library/blas/AutoGemm/Includes.py b/src/library/blas/AutoGemm/Includes.py +index 0f616956..e32694d8 100644 +--- a/src/library/blas/AutoGemm/Includes.py ++++ b/src/library/blas/AutoGemm/Includes.py +@@ -272,8 +272,8 @@ def addKernel(self, kernel): + kernelName = kernel.getName() + self.incStr += "extern const char * const %s_srcBuildOptions;\n" \ + % kernelName +- self.cppStr += "const char * const %s_srcBuildOptions = \"-cl-std=CL%s\";\n" \ +- % (kernelName, Common.getClCompilerVersion() ) ++ self.cppStr += "const char * const %s_srcBuildOptions = \"\";\n" \ ++ % kernelName + + self.incFile.write( self.incStr ) + self.incStr = "" +@@ -314,7 +314,7 @@ def __init__(self): + def addKernel(self, kernel): + kernelName = kernel.getName() + self.incStr += "extern const char * const %s_binBuildOptions;\n" % kernelName +- self.cppStr += "const char * const %s_binBuildOptions = \"-cl-std=CL%s\";\n" % (kernelName, Common.getClCompilerVersion() ) ++ self.cppStr += "const char * const %s_binBuildOptions = \"\";\n" % kernelName + + self.incFile.write( self.incStr ) + self.incStr = "" thanks for considering it. Gianfranco
OpenPGP_signature.asc
Description: OpenPGP digital signature
