Bug ID: JDK-8214751 X86: Support for VNNI Instructions

JDK-8214751 : X86: Support for VNNI Instructions

Type: Enhancement
Component: hotspot
Sub-Component: compiler
Affected Version: 12

Priority: P4
Status: Resolved
Resolution: Fixed
CPU: x86

Submitted: 2018-12-04
Updated: 2022-05-16
Resolved: 2018-12-12

Versions (Unresolved/Resolved/Fixed)

The Version table provides details related to the release that this issue/RFE will be addressed.

Unresolved : Release in which this issue/RFE will be addressed.
Resolved: Release in which this issue/RFE has been resolved.
Fixed : Release in which this issue/RFE has been fixed. The release containing this fix may be available for download as an Early Access Release or a General Availability Release.

To download the current JDK release, click here.

JDK 12	JDK 13
12 b24Fixed	13Fixed

Related Reports

Relates :	JDK-8216050 - Superword optimization fails with assert(0 <= i && i < _len) failed: illegal index
Relates :	JDK-8229694 - JVM crash in SWPointer during C2 OSR compilation
Relates :	JDK-8230185 - assert(is_Loop()) failed: invalid node class
Relates :	JDK-8230078 - compiler/loopopts/superword/Vec_MulAddS2I.java is unexpectedly slow in windows
Relates :	JDK-8239549 - AArch64: Backend support for MulAddVS2VI node
Relates :	JDK-8219151 - Illegal instruction exception on JDK 12 due to incorrect CPU feature bits
Relates :	JDK-8236701 - [TESTBUG] compiler/loopopts/superword/Vec_MulAddS2I.java uses wrong flag -XX:-SuperWord
Relates :	JDK-8215353 - x86_32 build failures after JDK-8214751 (X86: Support for VNNI Instructions)
Relates :	JDK-8215891 - X86: Support for VNNI byte Instruction VPDPBUSD
Relates :	JDK-8216580 - Fix generation of VNNI vector code by allowing adjacent LoadS nodes to be isomorphic

Description

This is 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]));

This patch is useful for AI ML/DL applications such as convolution based Neural Nets.

More information on VNNI can be found here:
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf
Code contributed by: razvan.a.lupusoru@intel.com and vdeshpande(vivek.r.deshpande@intel.com)

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:
http://cr.openjdk.java.net/~vdeshpande/8214751/VNNI/webrev.00/

Comments

Git URL: https://github.com/openjdk/jdk/commit/05e175bf1beeaecc24c120846595347cf08dd2c0
16-05-2022
URL: http://hg.openjdk.java.net/jdk/jdk/rev/4bb6e0871bf7 User: kvn Date: 2018-12-12 22:47:53 +0000
12-12-2018
Testing webrev.03 passed with known unrelated failures.
12-12-2018
Thanks Vladimir. I tested the change on my development machine.
12-12-2018
I added is_valid_counted_loop() in convert_add_to_muladd() and refactored code a little: http://cr.openjdk.java.net/~kvn/8214751/webrev.03/
12-12-2018
The updated webrev is here: http://cr.openjdk.java.net/~vdeshpande/8214751/VNNI/webrev.02/
10-12-2018