[LLVMdev] supporting SAD in loop vectorizer

Tue Nov 11 06:54:01 PST 2014

----- Original Message -----
> From: "James Molloy" <james at jamesmolloy.co.uk>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "Dibyendu Das" <Dibyendu.Das at amd.com>, llvmdev at cs.uiuc.edu
> Sent: Tuesday, November 11, 2014 8:21:37 AM
> Subject: Re: [LLVMdev] supporting SAD in loop vectorizer
> 
> 
> If you'd like to contribute support for this, look at
> isHorizontalBinOp and go from there. Feel free to ask questions if
> you get stuck.
> 
> 
> 
> FWIW, I've looked at isHorizontalBinOp for inspiration for matching
> AArch64 ADDV-and-friends (horizontal reduction operations), and
> thought it was rather temperamental and noticed it being prone to
> breaking depending on the exact format of the IR. Given that we
> don't have a canonical form for reductions, I think it wrong that we
> expect targets to undo quite complex patterns.
> 
> 
> The reduction pattern is a log2(n) sequence of shuffles and binops,
> that are really rather complex. These sort of things should, IMHO,
> be intrinsics. I chatted with Arnold about this at the devmtg and
> was going to send a patch to do exactly that in a week or so.

Sounds good. We should try hard to canonicalize into the intrinsic in InstCombine from the shuffles (in addition to emitting it directly from the vectorizer), but it is likely easier to do there than in the backend.

 -Hal

> 
> 
> Cheers,
> 
> 
> James
> 
> 
> On 11 November 2014 13:35, Hal Finkel < hfinkel at anl.gov > wrote:
> 
> 
> ----- Original Message -----
> > From: "Dibyendu Das" < Dibyendu.Das at amd.com >
> > To: "Hal Finkel" < hfinkel at anl.gov >, "Renato Golin" <
> > renato.golin at linaro.org >
> > Cc: llvmdev at cs.uiuc.edu
> > Sent: Tuesday, November 4, 2014 12:15:12 PM
> > Subject: RE: [LLVMdev] supporting SAD in loop vectorizer
> > 
> > Here's the simple SAD code:
> > ---------------------------------------------------
> > 1 #include <stdlib.h>
> > 2
> > 3 extern int ly,lx;
> > 4 int sad_c( unsigned char *pix1, unsigned char *pix2)
> > 5 {
> > 6 int i_sum = 0;
> > 7 for( int x = 0; x < lx; x++ )
> > 8 i_sum += abs( pix1[x] - pix2[x] );
> > 9 return i_sum;
> > 10 }
> > 11
> > -----------------------------------------------------
> > 
> > The loop vectorizer does vectorize the loop and then unrolls it
> > twice. The main body of the loop at the end looks like below where
> > we see the icmp, neg select pattern appearing twice.
> > Are we saying we pattern match this to PSADBW in target ?
> 
> Yes.
> 
> > That seems
> > to have some challenges
> 
> It does, but we already have code in the backend that matches other
> horizontal operations (see isHorizontalBinOp and its callers in
> lib/Target/X86/X86ISelLowering.cpp), and I suspect this won't be
> significantly more complicated.
> 
> > including the fact that we would need a
> > 4-way unroll to use all of 128b PSADBWs. Or am I
> > missing something ?
> 
> No, each unrolling will get its own, so you'll get a PSADBW from each
> time the loop is unrolled. Each unrolling is vectorized in terms of
> <4 x i32>, and that is the 128 bits you need.
> 
> If you'd like to contribute support for this, look at
> isHorizontalBinOp and go from there. Feel free to ask questions if
> you get stuck.
> 
> -Hal
> 
> 
> 
> > 
> > 2783 vector.body: ; preds =
> > %vector.body.preheader, %vector.body
> > 2784 %index = phi i64 [ %index.next, %vector.body ], [ 0,
> > %vector.body.preheader ]
> > 2785 %vec.phi = phi <4 x i32> [ %24, %vector.body ], [
> > zeroinitializer, %vector.body.preheader ]
> > 2786 %vec.phi9 = phi <4 x i32> [ %25, %vector.body ], [
> > zeroinitializer, %vector.body.preheader ]
> > 2787 %4 = getelementptr inbounds i8* %pix1, i64 %index
> > 2788 %5 = bitcast i8* %4 to <4 x i8>*
> > 2789 %wide.load = load <4 x i8>* %5, align 1
> > 2790 %.sum19 = or i64 %index, 4
> > 2791 %6 = getelementptr i8* %pix1, i64 %.sum19
> > 2792 %7 = bitcast i8* %6 to <4 x i8>*
> > 2793 %wide.load10 = load <4 x i8>* %7, align 1
> > 2794 %8 = zext <4 x i8> %wide.load to <4 x i32>
> > 2795 %9 = zext <4 x i8> %wide.load10 to <4 x i32>
> > 2796 %10 = getelementptr inbounds i8* %pix2, i64 %index
> > 2797 %11 = bitcast i8* %10 to <4 x i8>*
> > 2798 %wide.load11 = load <4 x i8>* %11, align 1
> > 2799 %.sum20 = or i64 %index, 4
> > 2800 %12 = getelementptr i8* %pix2, i64 %.sum20
> > 2801 %13 = bitcast i8* %12 to <4 x i8>*
> > 2802 %wide.load12 = load <4 x i8>* %13, align 1
> > 2803 %14 = zext <4 x i8> %wide.load11 to <4 x i32>
> > 2804 %15 = zext <4 x i8> %wide.load12 to <4 x i32>
> > 2805 %16 = sub nsw <4 x i32> %8, %14
> > 2806 %17 = sub nsw <4 x i32> %9, %15
> > 2807 %18 = icmp sgt <4 x i32> %16, <i32 -1, i32 -1, i32 -1, i32 -1>
> > 2808 %19 = icmp sgt <4 x i32> %17, <i32 -1, i32 -1, i32 -1, i32 -1>
> > 2809 %20 = sub <4 x i32> zeroinitializer, %16
> > 2810 %21 = sub <4 x i32> zeroinitializer, %17
> > 2811 %22 = select <4 x i1> %18, <4 x i32> %16, <4 x i32> %20
> > 2812 %23 = select <4 x i1> %19, <4 x i32> %17, <4 x i32> %21
> > 2813 %24 = add nsw <4 x i32> %22, %vec.phi
> > 2814 %25 = add nsw <4 x i32> %23, %vec.phi9
> > 2815 %index.next = add i64 %index, 8
> > 2816 %26 = icmp eq i64 %index.next, %n.vec
> > 2817 br i1 %26, label %middle.block.loopexit, label %vector.body,
> > !llvm.loop !1
> > -----------------------------------------------------
> > 
> > -----Original Message-----
> > From: Hal Finkel [mailto: hfinkel at anl.gov ]
> > Sent: Tuesday, November 04, 2014 9:54 PM
> > To: Renato Golin
> > Cc: llvmdev at cs.uiuc.edu ; Das, Dibyendu
> > Subject: Re: [LLVMdev] supporting SAD in loop vectorizer
> > 
> > ----- Original Message -----
> > > From: "Renato Golin" < renato.golin at linaro.org >
> > > To: "Dibyendu Das" < Dibyendu.Das at amd.com >
> > > Cc: llvmdev at cs.uiuc.edu
> > > Sent: Tuesday, November 4, 2014 5:23:30 AM
> > > Subject: Re: [LLVMdev] supporting SAD in loop vectorizer
> > > 
> > > On 4 November 2014 11:06, Das, Dibyendu < Dibyendu.Das at amd.com >
> > > wrote:
> > > > Is there any plan to support special idioms in the loop
> > > > vectorizer
> > > > like sum of absolute difference (SAD) ? We see some useful
> > > > cases
> > > > where llvm is losing performance at -O3 due to SADs not being
> > > > vectorized (hence PSADBWs not being generated).
> > > 
> > > It's been a while, but this could either be that the legalisation
> > > phase is not recognising the reduction or that the cost is not
> > > taking
> > > into account the lowered abs().
> > > 
> > > What does -debug-only=loop-vectorize say about it?
> > 
> > FWIW, I agree, this sounds like a cost-model problem. The
> > loop-vectorizer should be able to vectorize the 'icmp; neg; select'
> > pattern, and then the backend can pattern-patch that with the
> > reduction (which is a series of shuffles and extract_element) into
> > the single instruction PSADBW -- we're quite likely missing the
> > target code to do that.
> > 
> > -Hal
> > 
> > > 
> > > cheers,
> > > --renato
> > > _______________________________________________
> > > LLVM Developers mailing list
> > > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> > > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> > > 
> > 
> > --
> > Hal Finkel
> > Assistant Computational Scientist
> > Leadership Computing Facility
> > Argonne National Laboratory
> > 
> 
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> 
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory