FreeBSD CURRENT: testsuite failures depending on CPU model

[SB211]

Compiling POCL 0.10.0rc1 on FreeBSD 11.0-CURRENT with CLANG/LLVM 3.4.2 performs well until it comes to the testing.

On a Core2Duo system (4GB,

I see the following tests of the OpenCL book failing:
4: example1: dot product (SPIR32) skipped (testsuite.at:66)
71: struct kernel arguments - lp:987905 expected failure (testsuite-regression.at:108)
83: block expected failure (testsuite-regression.at:209)

and

TCE tests

100: A basic ttasim driver test skipped (testsuite-tce.at:6)
101: Half-precision floats on ttasim skipped (testsuite-tce.at:15)
102: A basic TCEMC test skipped (testsuite-tce.at:39)

OpenCL Programming Guide Samples

103: Building the sources against pocl FAILED (testsuite-samples.at:23)
104: Run Chapter 2: Hello World FAILED (testsuite-samples.at:32)
105: Run Chapter 3: OpenCLConvolution FAILED (testsuite-samples.at:42)
106: Run Chapter 6: HelloBinaryWorld FAILED (testsuite-samples.at:55)
107: Run Chapter 7: SimpleBufferSubBuffer FAILED (testsuite-samples.at:70)
108: Run Chapter 8: ImageFilter2D FAILED (testsuite-samples.at:82)
109: Run Chapter 12: VectorAdd (C++ bindings) FAILED (testsuite-samples.at:106)

On systems with CPU architectures (Intel amd64) > Core2Duo, like Sandy Bridge and Ivy Bridge (I tested on), additionally I receive those errors to the Core2Duo system's failure:

25: Kernel functions: shuffle charN FAILED (testsuite.at:306)
26: Kernel functions: shuffle shortN FAILED (testsuite.at:314)
27: Kernel functions: shuffle ushortN FAILED (testsuite.at:322)
28: Kernel functions: shuffle halfN skipped (testsuite.at:330)
29: Kernel functions: shuffle intN FAILED (testsuite.at:341)
30: Kernel functions: shuffle uintN FAILED (testsuite.at:349)
31: Kernel functions: shuffle floatN FAILED (testsuite.at:357)
32: Kernel functions: shuffle longN FAILED (testsuite.at:365)
33: Kernel functions: shuffle ulongN FAILED (testsuite.at:373)
34: Kernel functions: shuffle doubleN FAILED (testsuite.at:381)

See https://github.com/pocl/pocl/wiki/pocl-0.10-release-testing

Other tests fail all but this is probably due to a port's related configuration issue to be solved for testing.

[pjaaskel]

"skipped" is OK: it's a disabled test (e.g. if you have the test suite disabled). 71 I should remove as it tests an issue that depends on the target/ABI, thus breaks and works depending on the machine. I haven't tested the Programming Guide for a while as it works only without ICD. Try running 'make prepare-examples' first? However, the shuffle test is something to look at. Can you paste a snippet of the error log, e.g. tests/testsuite.dir/025/testsuite.log ?

[pjaaskel]

@kraiskil did you have a FreeBSD VM lying around somewhere?

[kraiskil]

No longer, alas. But I see the bug in shuffle. Now I just wonder how only BSD caught it...

Incidentally, does anyone want/need suffle for size-3 vectors. OCL specs do not seem to mention them.

[kraiskil]

No, that was a red herring.
From the output posted off-line, it looks like a codegen or alignment error, but it doesn't repeat for me.

[kraiskil]

Tried with IvyBridge & SandyBridge CPUs, with Ubuntu 12.04 & LLVM 3.4.2 - doesn't repeat.