RFR(M):8214751: X86: Support for VNNI instruction
Deshpande, Vivek R
vivek.r.deshpande at intel.com
Mon Dec 10 23:46:54 UTC 2018
I have the patch which adds the check that all the nodes belong to same loop.
Also I have added a jtreg test.
Could you please take a look at the patch.
From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
Sent: Friday, December 7, 2018 11:22 AM
To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; hotspot-compiler-dev at openjdk.java.net compiler <hotspot-compiler-dev at openjdk.java.net>
Cc: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru <guru.raj at intel.com>
Subject: Re: RFR(M):8214751: X86: Support for VNNI instruction
On 12/7/18 10:33 AM, Deshpande, Vivek R wrote:
> Hi Vladimir
> This patch is useful for AI ML/DL applications such convolution based Neural Nets.
Add to RFE's Description this comment. May be have some JMH benchmark results to show improvement.
> I have updated the patch with your suggestion.
> I am creating the MulAddS2I patch late and before vectoriztion.
Add check that all combined nodes belongs to the same loop - you have this information since you inside loopopts.
> The updated webrev is here:
> I am also working on the test.
> -----Original Message-----
> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
> Sent: Thursday, December 6, 2018 11:59 AM
> To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>;
> hotspot-compiler-dev at openjdk.java.net compiler
> <hotspot-compiler-dev at openjdk.java.net>
> Subject: Re: RFR(M):8214751: X86: Support for VNNI instruction
> Hi Vivek,
> What applications benefit this optimizations?
> This optimization may prevent some constant folding and others IGVN optimizations and RA since MulAddS2INode is generated too early I think. The only benefit we will have only if vectors are generated. Can you generate vectors without MulAddS2INode? Or create MulAddS2INode just before vectorization and expand it if vectorization failed? I would prefer first solution to have a struct in SuperWord code which find such pattern and try to vectorize it.
> You need to add test to verify correctness of results.
> Add UseAVX == 0 check to predicates which use SSE2 code. Otherwise they may be selected even if UseAVX > 0.
> On 12/3/18 8:58 PM, Deshpande, Vivek R wrote:
>> Hi All
>> Could you please review the VNNI VPDPWSSD instruction support with autovectorization.
>> It can vectorize this operation in the loop:
>> out[i] += ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])); More
>> information on VNNI can be found here:
>> t ure-instruction-set-extensions-programming-reference.pdf
>> The initial performance gains with micro on skylake with AVX3 is 10.8x.
>> and it generates
>> vmovdqu xmm3, xmmword ptr [rbp+r8*2+0x10] vmovdqu xmm6, xmmword ptr
>> [rdx+r8*2+0x10] vpmaddwd xmm3, xmm6, xmm3 vpaddd xmm3, xmm3, xmmword
>> ptr [r9+rdi*4+0x10] vmovdqu xmmword ptr [r9+rdi*4+0x10], xmm3
>> It can generate vpdpwssd instruction on cascadelake.
>> The webrev is here:
>> The jbs entry for the same is here:
More information about the hotspot-compiler-dev