RFR(M): 8181616: FMA Vectorization on x86
Vladimir Kozlov
vladimir.kozlov at oracle.com
Wed Jun 7 19:12:18 UTC 2017
Good. I will sponsor it.
Thanks,
Vladimir
On 6/7/17 11:04 AM, Deshpande, Vivek R wrote:
> Hi Vladimir
>
> I have updated the webrev with changes in supports_fma().
> The webrev is here: http://cr.openjdk.java.net/~vdeshpande/8181616/webrev.01/
> Thank you.
>
> Regards,
> Vivek
> -----Original Message-----
> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
> Sent: Tuesday, June 6, 2017 12:28 PM
> To: Deshpande, Vivek R; hotspot-compiler-dev at openjdk.java.net
> Cc: Viswanathan, Sandhya
> Subject: Re: RFR(M): 8181616: FMA Vectorization on x86
>
> Yes, Michael explained that FMA code should be generated only when AVX is available.
> In such case may be supports_fma() should check for AVX?
>
> I ran testing last night and results are good.
>
> Vladimir
>
> On 6/6/17 11:28 AM, Deshpande, Vivek R wrote:
>> Hi Vladimir
>>
>> Thanks for taking a look at the patch.
>> The FMA is introduced with Haswell microarchitecture along with AVX2. It is recommended to also check for AVX for using FMA instructions.
>>
>> Regards,
>> Vivek
>> -----Original Message-----
>> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
>> Sent: Monday, June 5, 2017 4:53 PM
>> To: Deshpande, Vivek R; hotspot-compiler-dev at openjdk.java.net
>> Cc: Viswanathan, Sandhya
>> Subject: Re: RFR(M): 8181616: FMA Vectorization on x86
>>
>> Thank you, Vivek
>>
>> Why you restrict FMA to AVX?:
>>
>> - if (supports_fma() && UseSSE >= 2) {
>> + if (supports_fma() && UseAVX > 0) {
>>
>> Otherwise changes look good.
>>
>> Thanks,
>> Vladimir
>>
>> On 6/5/17 2:42 PM, Deshpande, Vivek R wrote:
>>> Hi All
>>>
>>>
>>>
>>> The enhancement generates Vector FMA instructions using SuperWord Analysis for X86.
>>>
>>> We observed upto 6x gain on Skylake and KNL using 512 bit Vector FMA instructions using FMA API operating on independent arrays.
>>>
>>> Webrev:
>>>
>>> http://cr.openjdk.java.net/~vdeshpande/8181616/webrev.00/
>>>
>>> I have also updated the JBS entry.
>>>
>>> https://bugs.openjdk.java.net/browse/JDK-8181616
>>>
>>> Would you please review and sponsor it.
>>>
>>>
>>>
>>> Regards,
>>>
>>> Vivek
>>>
>>>
>>>
More information about the hotspot-compiler-dev
mailing list