Cycles AVX kernel detection leads to crash when CPU does support AVX but OS does NOT #38815

New Issue

Martijn Berger · 2014-02-24T20:39:28+01:00

Martijn Berger commented

2014-02-24 20:39:28 +01:00

System Information
Any OS without AVX support (windows 7 pre SP1 and Vista 64 are most common with this)

Blender Version
Broken: Any build after AVX kernel was introduced

Short description of error
http://insufficientlycomplicated.wordpress.com/2011/11/07/detecting-intel-advanced-vector-extensions-avx-in-visual-studio/

Exact steps for others to reproduce the error
Open Blender try to render

**System Information** Any OS without AVX support (windows 7 pre SP1 and Vista 64 are most common with this) **Blender Version** Broken: Any build after AVX kernel was introduced **Short description of error** http://insufficientlycomplicated.wordpress.com/2011/11/07/detecting-intel-advanced-vector-extensions-avx-in-visual-studio/ **Exact steps for others to reproduce the error** Open Blender try to render

Martijn Berger commented

2014-02-24 20:39:28 +01:00

Changed status to: 'Open'

Martijn Berger self-assigned this 2014-02-24 20:39:28 +01:00

Martijn Berger commented

2014-02-24 20:39:28 +01:00

Added subscriber: @MartijnBerger

Martijn Berger commented

2014-02-24 20:40:02 +01:00

Possible fix

Martijn Berger commented

2014-02-24 20:40:35 +01:00

Added subscribers: @ThomasDinges, @brecht

Martijn Berger commented

2014-02-24 22:54:40 +01:00

Tested on OS 10.9.1/macbook pro 2013.

With patch still detects AVX successful

Tested on OS 10.9.1/macbook pro 2013. With patch still detects AVX successful

Martijn Berger commented

2014-02-25 08:47:34 +01:00

The final check need to be (xcr_feature_mask & 0x6) == 0x6.

Tested also on linux x64 and windows x64 and for the true positive case the check still works on all platforms it seems

The final check need to be (xcr_feature_mask & 0x6) == 0x6. Tested also on linux x64 and windows x64 and for the true positive case the check still works on all platforms it seems

Martijn Berger commented

2014-02-25 09:12:13 +01:00

llvm does this:

  static bool OSHasAVXSupport() {

#if defined(__GNUC__)
  *Check xgetbv; this uses a .byte sequence instead of the instruction* directly because older assemblers do not include support for xgetbv and

    // there is no easy way to conditionally compile based on the assembler used.
    int rEAX, rEDX;
    __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (rEAX), "=d" (rEDX) : "c" (0));

#elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)

    unsigned long long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);

#else

    int rEAX = 0; // Ensures we return false

#endif

    return (rEAX & 6) == 6;
  }

If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV indicates that the AVX registers will be saved and restored on context

// switch, then we have full AVX support.
const unsigned AVXBits = (1 << 27) | (1 << 28);
bool HasAVX = ((ECX & AVXBits) == AVXBits) && OSHasAVXSupport();

llvm does this: ``` static bool OSHasAVXSupport() { ``` #if defined(__GNUC__) *Check xgetbv; this uses a .byte sequence instead of the instruction* directly because older assemblers do not include support for xgetbv and ``` // there is no easy way to conditionally compile based on the assembler used. int rEAX, rEDX; __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (rEAX), "=d" (rEDX) : "c" (0)); ``` #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK) ``` unsigned long long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); ``` #else ``` int rEAX = 0; // Ensures we return false ``` #endif ``` return (rEAX & 6) == 6; } ``` *If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV* indicates that the AVX registers will be saved and restored on context ``` // switch, then we have full AVX support. const unsigned AVXBits = (1 << 27) | (1 << 28); bool HasAVX = ((ECX & AVXBits) == AVXBits) && OSHasAVXSupport();

Martijn Berger commented

2014-02-25 15:32:47 +01:00

make asm more safe