yongqli

  • Joined on
@yongqli
"Acc (Scalar a)" vs "Exp a"
@yongqli
Blacklist is active, even when unrezzed
@yongqli

Could ! somehow get attached to a different expression? Meaning if I have not (x && y) end up with not x && y somehow? I just tried a simple exampl…

@yongqli

Also, how can I look at the AST and C++ code that is generated?

@yongqli

Can you make sure that a similar problem isn't happening with other operators, such as <* and >*? I have a case where I strongly suspect accelerate…

@yongqli
"error: expression must be a modifiable lvalue"
@yongqli

No, I haven't tried it yet. I've settled on just class Lower c e where type Lowest c e lower :: c e -> Lowest c e so that I can define lower for …

@yongqli

She does, but not when you install something, which always consumes her ability (since it's on first install). Thus sometimes I forget that I alrea…

@yongqli

I'm seeing this start to pop up again.

@yongqli

Yes, I've also encountered this phenomenon with multi-access.

@yongqli
Mark Hayley somehow
@yongqli

Ok, I've got an implementation of this working. Basically, the problem is that we might already have a function f :: Num a => (V3 (V3 a), a) -> a.…

@yongqli
Interpreter backend bit rot?
@yongqli

This turned out to be a bug in my code. It looks like the CUDA backend is more tolerant of accessing out of bounds memory.

@yongqli

No, I'm not using that. I have a bunch of simple operations on linear vectors like V2, V3, etc.

@yongqli

Data.Array.Accelerate.Array.Representation.toIndex.\, called from Data.Array.Accelerate.Array.Representation.toIndex, called from Data.Array.Accele…

@yongqli
Interpreter backend bit rot?
@yongqli

It looks like most accelerate programs generate the cuCtxPopCurrent_v2 returned (0xc9) warning.

@yongqli

Now it segfaults even without profiling turned on. But it does not segfault when run under cuda-gdb.

@yongqli

Running under cuda-gdb I see some warnings: warning: Cuda API error detected: cuCtxPopCurrent_v2 returned (0xc9) warning: Cuda API error detected: …

@yongqli

Actually, after moving some stuff around, it doesn't seem to happen when the program is exiting after all. The core dump is still the same: (cuda-g…

@yongqli

Actually, that's just because I ran it under cuda-gdb instead. The core dump is the same. It looks like the segfault is something else.

@yongqli

The "context is destroyed" errors are gone, but the segfaults still happen. cuda-memcheck doesn't have much to say: ========= Error: process didn't…

@yongqli

Yes.

@yongqli

I can only reproduce this when the code is run from cabal test. The exact same code doesn't cause any problems when put in its own program.

@yongqli

cuda-memcheck says ========= Program hit CUDA_ERROR_CONTEXT_IS_DESTROYED (error 709) due to "context is destroyed" on CUDA API call to cuMemFree_v2…

@yongqli
Segfaults with profiling turned on
yongqli opened pull request AccelerateHS/accelerate#259
@yongqli
support ghc 7.10
1 commit with 7 additions and 5 deletions
@yongqli
CUDA Exception: an illegal memory access was encountered
@yongqli

That didn't fix the problem, but the issue is with our code and not accelerate. Thanks for fixing the bug.