[Piglit] [PATCH] gl-1.0-blend-func: skip some blend tests when using LLVM 3.8
Brian Paul
brianp at vmware.com
Thu Jan 18 20:04:14 UTC 2018
On 01/18/2018 12:38 PM, Roland Scheidegger wrote:
> Am 18.01.2018 um 20:07 schrieb Brian Paul:
>> To avoid an infinite loop. See code comments for details.
>> ---
>> tests/spec/gl-1.0/blend.c | 39 +++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 39 insertions(+)
>>
>> diff --git a/tests/spec/gl-1.0/blend.c b/tests/spec/gl-1.0/blend.c
>> index 35e940f..e69ed31 100644
>> --- a/tests/spec/gl-1.0/blend.c
>> +++ b/tests/spec/gl-1.0/blend.c
>> @@ -76,6 +76,9 @@ static int test_stride = 1;
>> #define img_width drawing_size
>> #define img_height drawing_size
>>
>> +static bool using_llvm_3_8 = false;
>> +
>> +
>> PIGLIT_GL_TEST_CONFIG_BEGIN
>>
>> config.supports_gl_compat_version = 10;
>> @@ -177,6 +180,10 @@ image_init(struct image *image)
>> GL_RGBA, GL_FLOAT, image->data);
>> glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
>> glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
>> +
>> + if (strstr("LLVM 3.8", (const char *) glGetString(GL_RENDERER)) == 0) {
>> + using_llvm_3_8 = true;
>> + }
>> } /* image_init */
> FWIW I think this might mistakenly catch other drivers which use llvm
> (radeonsi).
I'll fix that by also testing for "llvmpipe".
>
> I am not really sure it's worth the trouble. Personally, I couldn't care
> less about llvm 3.8 causing a hang here (distros which updated llvm 3.8
> with the fixes from the 3.8 stable branch won't have a problem).
As I mentioned in my other email, I recently updated to Mint 18.3
(latest version) and this version of LLVM is the default. Mint 18 is
based on Ubuntu 16.04 which has LLVM 3.8 also (but I haven't tested it
to see if has this bug or is fixed). Same story with Ubuntu 16.10. So
this buggy LLVM may be common.
> So all
> the trouble just for avoiding a bug with some old-ish llvm version seems
> a bit overkill to me. I am not sure anymore if it actually was an issue
> both on AVX(2) and non-AVX path neither.
> Plus an actual app using the affected blend factor combinations will
> just hang too, so hiding it here doesn't really seem helpful. (IIRC I
> thought about trying to work around it in llvmpipe code, but there was
> no trivial solution (for not just avoiding the hang but still correctly
> rendering) and I didn't care enough about that particular llvm version,
> even more so since it was fixed in the stable branch...)
>
> But in any case if you really think it's worthwile, I am not opposing it
> neither, so
> Reviewed-by: Roland Scheidegger <sroland at vmware.com>
It'll probably save me some grief so I'd like to push it.
-Brian
>
>
>>
>> @@ -532,6 +539,32 @@ apply_blend(GLenum src_factor_rgb, GLenum src_factor_a,
>> } /* apply_blend */
>>
>>
>> +/**
>> + * With unpactched LLVM 3.8, llvmpipe can hit an bug in LLVM which results
>> + * in an infinite loop. See https://bugs.llvm.org/show_bug.cgi?id=27689
>> + * This function tries to determine if the test case will hit that bug so
>> + * we can skip the test.
>> + */
>> +bool
>> +skip_on_llvmpipe(GLenum src_factor_rgb, GLenum src_factor_a,
>> + GLenum dst_factor_rgb, GLenum dst_factor_a)
>> +{
>> + if (!using_llvm_3_8)
>> + return false;
>> +
>> + /* This could probably be a bit tighter, but this seems to catch
>> + * the troublesome cases.
>> + */
>> + if (src_factor_rgb == GL_CONSTANT_COLOR ||
>> + dst_factor_rgb == GL_CONSTANT_COLOR ||
>> + dst_factor_a == GL_CONSTANT_COLOR ||
>> + dst_factor_a == GL_CONSTANT_ALPHA)
>> + return true;
>> +
>> + return false;
>> +}
>> +
>> +
>> /* Test for one set of factor levels */
>> bool
>> run_factor_set(GLenum src_factor_rgb, GLenum src_factor_a,
>> @@ -542,6 +575,12 @@ run_factor_set(GLenum src_factor_rgb, GLenum src_factor_a,
>> int i, j;
>> bool pass = true, p;
>>
>> + if (skip_on_llvmpipe(src_factor_rgb, src_factor_a,
>> + dst_factor_rgb, dst_factor_a)) {
>> + /*printf("Skipping Blend test to avoid LLVM bug\n");*/
>> + return true;
>> + }
>> +
>> glDisable(GL_DITHER);
>> glClear(GL_COLOR_BUFFER_BIT);
>>
>>
More information about the Piglit
mailing list