[Libclc-dev] [PATCH 4/4] rootn: Port from amd_builtins
Jeroen Ketema via Libclc-dev
libclc-dev at lists.llvm.org
Wed Jan 31 14:16:51 PST 2018
> On 31 Jan 2018, at 22:51, Jan Vesely <jan.vesely at rutgers.edu> wrote:
> On Thu, 2018-01-18 at 20:33 +0100, Jeroen Ketema wrote:
>>> On 18 Jan 2018, at 00:49, Jan Vesely <jan.vesely at rutgers.edu> wrote:
>>> On Thu, 2018-01-18 at 00:39 +0100, Jeroen Ketema wrote:
>>>> Hi Jan,
>>>> Triggered by your patch below, I started looking a bit more closely at __clc_fp32_subnormals_supported.
>>>> Am I correct that this function currently always returns false no matter the target? Or does some
>>>> override occur somewhere?
>>> Looks like it. Tom might be better qualified to answer this.
>>> The original idea was to link default or disabled to select proper
>>> operation at kernel link time, but it looks likes it only applies to
>>> That is weird since fp64 denormals are required by specs. I'd say we
>>> can drop the entire mechanism.
>> That’s not totally true. Check out the cl-denorms-are-zero option.
> sorry, this one slipped under my radar. You're right. The mechanism was
> aimed at AMD gpus that only care about fp64 denormals (fp32 denorms are
> preferred disabled for performance/compliance reasons).
> I don't mind someone reworking it to be more general (and idealy in CLC
> instead of llvm ir).
Thanks for the clarification. In that light the implementation makes sense to me.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Libclc-dev