frasercrmck wrote:
> I still have a few things to investigate to make sure of what is happening,
> but in the meantime would you guys approve to land this patch to make the
> implementation more stable whether the device supports denorm or not?
>
> ```
> - __CLC_INTN is_inf_nan_or_zero =
> -
rjodinchr wrote:
I'm clearly seeing those devices returning true when comparing subnormals to
zero. So either they all have a bug, or the SPIR-V code produced by `clspv` is
buggy.
If it ends up being a bug from the device, we might want to have something more
architecture friendly, or at least
rjodinchr wrote:
I've deployed that patch into ChromeOS, and now Intel, Qualcomm and Imagination
are also failing. I need to have a closer look to see if anything can be
changed in `clspv` to fix it, but I might end-up needing the fix mentioned a
few weeks ago.
https://github.com/llvm/llvm-pr
rjodinchr wrote:
I tend to agree. let's keep it that way.
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
frasercrmck wrote:
It strikes me as a swiftshader bug. If `x` is a denormal/subnormal number and
it is being untouched by this function and really reaching this comparison as
`0x1.008p-140` then I don't believe it should ever compare true against `0.0`.
My reading of flush-to-zero
([here](htt
frasercrmck wrote:
That's interesting. As I said alive2 was happy with the before/after IR diff of
the AMDGPU version of this builtin. Do you know what might be going on? Is that
the fp64 or fp32 test that's failing?
https://github.com/llvm/llvm-project/pull/129871
arsenm wrote:
Is this failure using -cl-denorms-are-zero (IIRC -z when using
math_brute_force)
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cf
rjodinchr wrote:
The following patch is fixing the issue on my side:
```
- __CLC_INTN is_inf_nan_or_zero =
- x == __CLC_FP_LIT(0.0) || __clc_isinf(x) || __clc_isnan(x);
+ __CLC_INTN is_inf_nan_or_zero =
+ ai == (__CLC_INTN)0 || e == (__CLC_INTN)129 || __clc_isinf(x) ||
__clc_isnan(
rjodinchr wrote:
But maybe it is just masking a swiftshader bug?
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
rjodinchr wrote:
This is failing with `fp32`. The failing value is very close to zero (it's a
denormal), I think something is going wrong with `x == __CLC_FP_LIT(0.0)` when
computing `is_inf_nan_or_zero`.
Maybe a bug inside Swiftshader with denormal comparison to zero?
https://github.com/llvm/
rjodinchr wrote:
yes wrong results:
```
ERROR: frexp: {-8404992.00, 139} ulp error at 0x1.008p-140: *{0x1.008p-1,
-139} vs. {0x1.008p-140, 0}
```
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.
frasercrmck wrote:
> This commit breaks running with clvk & Swiftshader. Not a big deal on my
> side. Just wanted to let you know.
Oh, what exactly breaks? As in wrong results?
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing
rjodinchr wrote:
This commit breaks running with clvk & Swiftshader. Not a big deal on my side.
Just wanted to let you know.
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.or
https://github.com/frasercrmck closed
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
frasercrmck wrote:
It's worth noting that alive2 considers this a correct transformation for the
before/after IR, for the scalar float/double versions of __clc_frexp for amdgcn.
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing
https://github.com/arsenm approved this pull request.
I haven't spent much time looking at this one, because we have instructions for
the two halves
https://github.com/llvm/llvm-project/pull/129871
___
cfe-commits mailing list
cfe-commits@lists.llvm.
https://github.com/frasercrmck created
https://github.com/llvm/llvm-project/pull/129871
Also replace some magic constants with named ones.
Checking against FP zero and using isnan and isinf functions allows the
optimizer to create one unified @llvm.is.fpclass intrinsic. This results in
fewer
17 matches
Mail list logo
