This is an archive of the discontinued LLVM Phabricator instance.

Differential D39403

[LV][X86] update the cost of interleaving mem. access of floats
ClosedPublic

Authored by magabari on Oct 29 2017, 2:56 AM.

Details

Reviewers
delena
dorit
zvi
RKSimon
craig.topper
Commits
rG669175836406: [LV][X86] update the cost of interleaving mem. access of floats
rGacd69dbc7c57: [REVERT][LV][X86] update the cost of interleaving mem. access of floats
rGf74c767de692: [LV][X86] update the cost of interleaving mem. access of floats
rL317471: [LV][X86] update the cost of interleaving mem. access of floats
rL317433: [REVERT][LV][X86] update the cost of interleaving mem. access of floats
rL317432: [LV][X86] update the cost of interleaving mem. access of floats
Summary

This patch contains update of the costs of interleaved loads of v8floats of stride 3 and 8.

Diff Detail

Repository
rL LLVM

Event Timeline

magabari created this revision.Oct 29 2017, 2:56 AM
RKSimon added inline comments.Oct 29 2017, 4:43 AM
lib/Target/X86/X86TargetTransformInfo.cpp
2622 ↗(On Diff #120742)

9f32? or 24f32?

magabari updated this revision to Diff 120780.Oct 30 2017, 1:50 AM
magabari marked an inline comment as done.

updating diff with the comment fix

magabari updated this revision to Diff 120781.Oct 30 2017, 1:52 AM
magabari added a comment.Oct 31 2017, 3:08 AM

@RKSimon can you take a look on the changes please.

RKSimon edited edge metadata.Nov 1 2017, 6:11 AM
RKSimon added a subscriber: m_zuckerman.

Didn't @m_zuckerman put the previous tests into llvm/test/Analysis/CostModel/interleaved-*.ll ? Maybe match that?

RKSimon added a subscriber: llvm-commits.Nov 1 2017, 6:12 AM

Also, please don't forget to CC llvm-commits as a subscriber in your patches.

magabari added a comment.Nov 1 2017, 6:33 AM
In D39403#912670, @RKSimon wrote:

Didn't @m_zuckerman put the previous tests into llvm/test/Analysis/CostModel/interleaved-*.ll ? Maybe match that?

Do you mean renaming the test to interleaved-load-float.ll?
because as i see those tests is for i8 and i32 while i am checking new datatype.

dcaballe added a subscriber: dcaballe.Nov 1 2017, 12:02 PM
RKSimon added a comment.Nov 1 2017, 1:32 PM
In D39403#912689, @magabari wrote:
In D39403#912670, @RKSimon wrote:

Didn't @m_zuckerman put the previous tests into llvm/test/Analysis/CostModel/interleaved-*.ll ? Maybe match that?

Do you mean renaming the test to interleaved-load-float.ll?
because as i see those tests is for i8 and i32 while i am checking new datatype.

Yes please, move it to llvm/test/Analysis/CostModel/ and rename it interleaved-load-float.ll

magabari updated this revision to Diff 121239.Nov 2 2017, 12:45 AM

moving the test to cost model directory (and renaming it)

RKSimon accepted this revision.Nov 3 2017, 5:31 AM

LGTM

This revision is now accepted and ready to land.Nov 3 2017, 5:31 AM
Closed by commit rL317432: [LV][X86] update the cost of interleaving mem. access of floats (authored by magabari). · Explain WhyNov 5 2017, 1:06 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
llvm/
trunk/
lib/
Target/
X86/
5 lines
test/
Analysis/
CostModel/
141 lines

Diff 121621

llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp

Show First 20 Lines • Show All 2,638 Lines • ▼ Show 20 Linesint X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy,
// The cost of the loads/stores is accounted for separately. // The cost of the loads/stores is accounted for separately.
// //
static const CostTblEntry AVX2InterleavedLoadTbl[] = { static const CostTblEntry AVX2InterleavedLoadTbl[] = {
{ 3, MVT::v2i8, 10 }, //(load 6i8 and) deinterleave into 3 x 2i8 { 3, MVT::v2i8, 10 }, //(load 6i8 and) deinterleave into 3 x 2i8
{ 3, MVT::v4i8, 4 }, //(load 12i8 and) deinterleave into 3 x 4i8 { 3, MVT::v4i8, 4 }, //(load 12i8 and) deinterleave into 3 x 4i8
{ 3, MVT::v8i8, 9 }, //(load 24i8 and) deinterleave into 3 x 8i8 { 3, MVT::v8i8, 9 }, //(load 24i8 and) deinterleave into 3 x 8i8
{ 3, MVT::v16i8, 11}, //(load 48i8 and) deinterleave into 3 x 16i8 { 3, MVT::v16i8, 11}, //(load 48i8 and) deinterleave into 3 x 16i8
{ 3, MVT::v32i8, 13}, //(load 96i8 and) deinterleave into 3 x 32i8 { 3, MVT::v32i8, 13}, //(load 96i8 and) deinterleave into 3 x 32i8
{ 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32
{ 4, MVT::v2i8, 12 }, //(load 8i8 and) deinterleave into 4 x 2i8 { 4, MVT::v2i8, 12 }, //(load 8i8 and) deinterleave into 4 x 2i8
{ 4, MVT::v4i8, 4 }, //(load 16i8 and) deinterleave into 4 x 4i8 { 4, MVT::v4i8, 4 }, //(load 16i8 and) deinterleave into 4 x 4i8
{ 4, MVT::v8i8, 20 }, //(load 32i8 and) deinterleave into 4 x 8i8 { 4, MVT::v8i8, 20 }, //(load 32i8 and) deinterleave into 4 x 8i8
{ 4, MVT::v16i8, 39 }, //(load 64i8 and) deinterleave into 4 x 16i8 { 4, MVT::v16i8, 39 }, //(load 64i8 and) deinterleave into 4 x 16i8
{ 4, MVT::v32i8, 80 } //(load 128i8 and) deinterleave into 4 x 32i8 { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8
{ 8, MVT::v8f32, 40 } //(load 64f32 and)deinterleave into 8 x 8f32
}; };
static const CostTblEntry AVX2InterleavedStoreTbl[] = { static const CostTblEntry AVX2InterleavedStoreTbl[] = {
{ 3, MVT::v2i8, 7 }, //interleave 3 x 2i8 into 6i8 (and store) { 3, MVT::v2i8, 7 }, //interleave 3 x 2i8 into 6i8 (and store)
{ 3, MVT::v4i8, 8 }, //interleave 3 x 4i8 into 12i8 (and store) { 3, MVT::v4i8, 8 }, //interleave 3 x 4i8 into 12i8 (and store)
{ 3, MVT::v8i8, 11 }, //interleave 3 x 8i8 into 24i8 (and store) { 3, MVT::v8i8, 11 }, //interleave 3 x 8i8 into 24i8 (and store)
{ 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store) { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store)
{ 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store) { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store)
▲ Show 20 Lines • Show All 175 Lines • Show Last 20 Lines

llvm/trunk/test/Analysis/CostModel/interleaved-load-float.ll

; REQUIRES: asserts
; RUN: opt -S -loop-vectorize -debug-only=loop-vectorize -mcpu=skylake %s 2>&1 | FileCheck %s
target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"
target triple = "i386-unknown-linux-gnu"
@src = common local_unnamed_addr global [120 x float] zeroinitializer, align 4
@dst = common local_unnamed_addr global [120 x float] zeroinitializer, align 4
; Function Attrs: norecurse nounwind
define void @stride8(float %k, i32 %width_) {
entry:
; CHECK: Found an estimated cost of 48 for VF 8 For instruction: %0 = load float
%cmp72 = icmp sgt i32 %width_, 0
br i1 %cmp72, label %for.body.lr.ph, label %for.cond.cleanup
for.body.lr.ph: ; preds = %entry
br label %for.body
for.cond.cleanup.loopexit: ; preds = %for.body
br label %for.cond.cleanup
for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry
ret void
for.body: ; preds = %for.body.lr.ph, %for.body
%i.073 = phi i32 [ 0, %for.body.lr.ph ], [ %add46, %for.body ]
%arrayidx = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %i.073
%0 = load float, float* %arrayidx, align 4
%mul = fmul fast float %0, %k
%arrayidx2 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %i.073
%1 = load float, float* %arrayidx2, align 4
%add3 = fadd fast float %1, %mul
store float %add3, float* %arrayidx2, align 4
%add4 = or i32 %i.073, 1
%arrayidx5 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add4
%2 = load float, float* %arrayidx5, align 4
%mul6 = fmul fast float %2, %k
%arrayidx8 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add4
%3 = load float, float* %arrayidx8, align 4
%add9 = fadd fast float %3, %mul6
store float %add9, float* %arrayidx8, align 4
%add10 = or i32 %i.073, 2
%arrayidx11 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add10
%4 = load float, float* %arrayidx11, align 4
%mul12 = fmul fast float %4, %k
%arrayidx14 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add10
%5 = load float, float* %arrayidx14, align 4
%add15 = fadd fast float %5, %mul12
store float %add15, float* %arrayidx14, align 4
%add16 = or i32 %i.073, 3
%arrayidx17 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add16
%6 = load float, float* %arrayidx17, align 4
%mul18 = fmul fast float %6, %k
%arrayidx20 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add16
%7 = load float, float* %arrayidx20, align 4
%add21 = fadd fast float %7, %mul18
store float %add21, float* %arrayidx20, align 4
%add22 = or i32 %i.073, 4
%arrayidx23 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add22
%8 = load float, float* %arrayidx23, align 4
%mul24 = fmul fast float %8, %k
%arrayidx26 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add22
%9 = load float, float* %arrayidx26, align 4
%add27 = fadd fast float %9, %mul24
store float %add27, float* %arrayidx26, align 4
%add28 = or i32 %i.073, 5
%arrayidx29 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add28
%10 = load float, float* %arrayidx29, align 4
%mul30 = fmul fast float %10, %k
%arrayidx32 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add28
%11 = load float, float* %arrayidx32, align 4
%add33 = fadd fast float %11, %mul30
store float %add33, float* %arrayidx32, align 4
%add34 = or i32 %i.073, 6
%arrayidx35 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add34
%12 = load float, float* %arrayidx35, align 4
%mul36 = fmul fast float %12, %k
%arrayidx38 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add34
%13 = load float, float* %arrayidx38, align 4
%add39 = fadd fast float %13, %mul36
store float %add39, float* %arrayidx38, align 4
%add40 = or i32 %i.073, 7
%arrayidx41 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add40
%14 = load float, float* %arrayidx41, align 4
%mul42 = fmul fast float %14, %k
%arrayidx44 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add40
%15 = load float, float* %arrayidx44, align 4
%add45 = fadd fast float %15, %mul42
store float %add45, float* %arrayidx44, align 4
%add46 = add nuw nsw i32 %i.073, 8
%cmp = icmp slt i32 %add46, %width_
br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit
}
; Function Attrs: norecurse nounwind
define void @stride3(float %k, i32 %width_) {
entry:
; CHECK: Found an estimated cost of 20 for VF 8 For instruction: %0 = load float
%cmp27 = icmp sgt i32 %width_, 0
br i1 %cmp27, label %for.body.lr.ph, label %for.cond.cleanup
for.body.lr.ph: ; preds = %entry
br label %for.body
for.cond.cleanup: ; preds = %for.body, %entry
ret void
for.body: ; preds = %for.body.lr.ph, %for.body
%i.028 = phi i32 [ 0, %for.body.lr.ph ], [ %add16, %for.body ]
%arrayidx = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %i.028
%0 = load float, float* %arrayidx, align 4
%mul = fmul fast float %0, %k
%arrayidx2 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %i.028
%1 = load float, float* %arrayidx2, align 4
%add3 = fadd fast float %1, %mul
store float %add3, float* %arrayidx2, align 4
%add4 = add nuw nsw i32 %i.028, 1
%arrayidx5 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add4
%2 = load float, float* %arrayidx5, align 4
%mul6 = fmul fast float %2, %k
%arrayidx8 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add4
%3 = load float, float* %arrayidx8, align 4
%add9 = fadd fast float %3, %mul6
store float %add9, float* %arrayidx8, align 4
%add10 = add nuw nsw i32 %i.028, 2
%arrayidx11 = getelementptr inbounds [120 x float], [120 x float]* @src, i32 0, i32 %add10
%4 = load float, float* %arrayidx11, align 4
%mul12 = fmul fast float %4, %k
%arrayidx14 = getelementptr inbounds [120 x float], [120 x float]* @dst, i32 0, i32 %add10
%5 = load float, float* %arrayidx14, align 4
%add15 = fadd fast float %5, %mul12
store float %add15, float* %arrayidx14, align 4
%add16 = add nuw nsw i32 %i.028, 3
%cmp = icmp slt i32 %add16, %width_
br i1 %cmp, label %for.body, label %for.cond.cleanup
}