[LLVMdev] AVX2 Cost Table in X86TargetTransformInfo

[suyog sarda]

+LLVMdev (sorry for not broadcasting earlier)
On 5 May 2015 12:40, "suyog sarda" <sardask01 at gmail.com> wrote:

> Hi Nadav,
>
> I stumbled upon one more question (sorry for not specifying earlier).
> Below query is when -mavx2 is specified as target feature.
>
> As i understand correctly, AVX1 is subset of AVX2. In SLP, we get scalar
> reduction cost in getReduction() function, which queries the TTI
> (TargetTransformInfo) via getArithmeticInstrCost().
>
> Now for integer ADD, since AVX2 added support for integer arithmetic, the
> entry for ADD (SUB/MUL) are missing in AVX2CostTable (which is what you
> also specified earlier).
> It fails to find the entry and goes for subsequent checks further. When it
> comes to AVX1 check, it specifically checks if AVX2 is not specified.
>
> (ST->hasAVX() && !ST->hasAVX2())
>
> since, we have specified -mavx2 this check also fails falls back to
> BaseTTI.
>
> Shouldn't it just check for hasAVX(), since AVX1 is subset of AVX2 ?
>
> (ST->hasAVX())
>
> I have a situation where i have integer ADD as reduction op. When i
> specify AVX2, the scalar cost is much less than AVX1. And hence, it doesn't
> vectorize the code at all.
>  If AVX2 vector instructions are costly, shouldn't it fall back to AVX1
> and generate AVX1 vector instructions?
>
> Correct me if i am wrong somewhere. Awaiting for your comments :)
>
> Thanks.
>
> Regards,
> Suyog
>
>
>
>
>
>
> On Mon, May 4, 2015 at 11:20 PM, Nadav Rotem <nrotem at apple.com> wrote:
>
>>
>> On May 4, 2015, at 10:23 AM, suyog sarda <sardask01 at gmail.com> wrote:
>>
>> Thanks Nadav for the info. It clears my query :)
>>
>> Yes its an integer ADD, and since AVX2 supports 256 bits integer
>> arithmetic, so its cost is less than AVX1.
>>
>> One query though - shouldn't then the cost of integer ADD/SUB/MUL (which
>> would be 1) be explicitly specified in AVX2 cost table? Because right now
>> this entry is missing and cost of these operations are taken from BaseTTI
>> (which is generic). IMO, it will make things more clear.
>>
>> Your thoughts on this??
>>
>>
>> I prefer that we continue to rely on TargetLowering in order to avoid
>> duplicating the cost information.
>>
>> Regards,
>> Suyog Sarda
>> On 4 May 2015 21:57, "Nadav Rotem" <nrotem at apple.com> wrote:
>>
>>>
>>> > On May 4, 2015, at 2:36 AM, suyog sarda <sardask01 at gmail.com> wrote:
>>> >
>>> > Hi all,
>>> >
>>> > I have a query regarding Cost Table for AVX2 in TargetTransformInfo.
>>> >
>>> > The table consist of entries for shift and div operations only. There
>>> are no entries for ADD, SUB and MUL for AVX2 cost table. Those entries are
>>> present in Cost Table for AVX.
>>>
>>> Most of the cost information is inferred from the TargetLowering tables
>>> (where operations are marked as Legal, Custom, etc.)  Only exceptional
>>> instructions need to be recorded in the TargetTransformInfo cost tables.
>>>
>>> >
>>> > The reason for query is - when my sub target feature is AVX2, in SLP
>>> Vectorization,  while calculating scalar cost of ADD, it doesn't see the
>>> entry in cost table and falls back to default implementation returning cost
>>> 1. While for AVX, it finds the ADD in cost table and returns 4 as scalar
>>> cost.
>>>
>>> >
>>> > I am suspecting this is something specific to architecture difference
>>> between AVX and AVX2. I am naive to architecture specifics in this case.
>>>
>>> I assume that this is integer ADD, because AVX1 only supported floating
>>> point arithmetic on 256bit vectors, while AVX2 added support for 256bit
>>> integer arithmetic. So, it makes sense that the cost that AVX1 gives this
>>> operation is much higher.
>>>
>>>
>>> >
>>> > I would be glad if someone clarifies on this.
>>> >
>>> > Thanks.
>>> >
>>> > Regards,
>>> > Suyog Sarda
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150505/270fa5d5/attachment.html>