Re: [boost] [simd] Hardware support

Subject: Re: [boost] [simd] Hardware support
From: Joel FALCOU (joel.falcou_at_[hidden])
Date: 2017-04-10 12:58:07

On 10/04/2017 14:44, Bjorn Reese via Boost wrote:
> On 04/08/2017 01:06 PM, Mathias Gaunard via Boost wrote:
>> On 8 April 2017 at 11:14, Bjorn Reese via Boost <boost_at_[hidden]>
>> wrote:
>
>> What does Boost.SIMD currently do on unsupported platforms? Revert
>>> back to SISD?
>>>
>>
>> Yes, likewise if you're on x86 but requesting an operation that your
>> hardware does not support, it will still work using the best possible
>> implementation.
>
> Is there any run-time overhead (e.g. induced by pack, load, save, etc.)
> in the SISD case?

This is a case we don't benchmark much but the strategy is to statically 
unroll the scalar calls over the pack element which are store din an 
std::array. So I suspect the overhead is not worth than having called 
the scalar function Nth times.