This is an archive of the discontinued LLVM Phabricator instance.

Differential D93173

[X86] Add test case for commit e52bc1d2bba794b.
ClosedPublic

Authored by LuoYuanke on Dec 12 2020, 7:15 PM.

Details

Reviewers
lebedev.ri
craig.topper
pengfei
Commits
rG18925dd8722e: [X86] Add test case for commit e52bc1d2bba794b.
Summary

Add test case for commit e52bc1d2bba794b.

Diff Detail

Event Timeline

LuoYuanke created this revision.Dec 12 2020, 7:15 PM
Herald added a subscriber: pengfei. · View Herald TranscriptDec 12 2020, 7:15 PM
LuoYuanke requested review of this revision.Dec 12 2020, 7:15 PM
Herald added a project: Restricted Project. · View Herald TranscriptDec 12 2020, 7:15 PM
Herald added a subscriber: llvm-commits. · View Herald Transcript
LuoYuanke edited the summary of this revision. (Show Details)Dec 12 2020, 7:17 PM
LuoYuanke added reviewers: lebedev.ri, craig.topper, pengfei.
LuoYuanke added a subscriber: annita.zhang.
Harbormaster completed remote builds in B82172: Diff 311424.Dec 12 2020, 7:48 PM
pengfei added inline comments.Dec 14 2020, 3:50 AM
llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll
5

The triple is different from RUN command, you can just remove it. Or remove the mtriple from command line.

13

Does the adding chain that generates more __tile1024i initialization here?
But why do we need to generate more initialization for the same row and col?

64

I think you can add avx512f in "target-features" and remove -mattr=+amx-int8 -mattr=+avx512f in run line.

LuoYuanke updated this revision to Diff 311547.Dec 14 2020, 4:07 AM

Address Pengfei's comments.

LuoYuanke marked 2 inline comments as done.Dec 14 2020, 4:08 AM
LuoYuanke added inline comments.Dec 14 2020, 4:11 AM
llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll
13

Adding chain doesn't generate more __tile1024i initialization. There 5 movb for row, 5 movw for column and movb $1, ... for palette. There are 5 tmm register allocated, so the initialization is correct.

LuoYuanke added inline comments.Dec 14 2020, 4:12 AM
llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll
13

One of the tilestore is deleted during instruction selection if we don't add chain.

LuoYuanke updated this revision to Diff 311549.Dec 14 2020, 4:13 AM

Rename function name.

Harbormaster completed remote builds in B82249: Diff 311547.Dec 14 2020, 4:39 AM
Harbormaster completed remote builds in B82251: Diff 311549.Dec 14 2020, 5:20 AM
LuoYuanke updated this revision to Diff 311742.Dec 14 2020, 4:29 PM

Fix function name check.

Harbormaster completed remote builds in B82358: Diff 311742.Dec 14 2020, 4:30 PM
LuoYuanke updated this revision to Diff 311743.Dec 14 2020, 4:33 PM

Rebase.

Harbormaster completed remote builds in B82359: Diff 311743.Dec 14 2020, 4:34 PM
pengfei accepted this revision.Dec 14 2020, 6:23 PM

LGTM.

llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll
13

I see. Thanks.

This revision is now accepted and ready to land.Dec 14 2020, 6:23 PM
This revision was landed with ongoing or failed builds.Dec 14 2020, 7:14 PM
Closed by commit rG18925dd8722e: [X86] Add test case for commit e52bc1d2bba794b. (authored by LuoYuanke). · Explain Why
This revision was automatically updated to reflect the committed changes.
LuoYuanke added a commit: rG18925dd8722e: [X86] Add test case for commit e52bc1d2bba794b..

Revision Contents

PathSize
llvm/
test/
CodeGen/
X86/
AMX/
59 lines

Diff 311770

llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll

  • This file was added.
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -verify-machineinstrs | FileCheck %s
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
define dso_local void @test_chain(i8* %A_mem, i8* %B_mem, i8* %C_mem) local_unnamed_addr {
pengfeiUnsubmitted
Done
ReplyInline Actions

The triple is different from RUN command, you can just remove it. Or remove the mtriple from command line.

pengfei: The triple is different from RUN command, you can just remove it. Or remove the mtriple from…
; CHECK-LABEL: test_chain:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: vpxord %zmm0, %zmm0, %zmm0
; CHECK-NEXT: vmovdqu64 %zmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $1, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $16, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movw $64, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $16, -{{[0-9]+}}(%rsp)
pengfeiUnsubmitted
Not Done
ReplyInline Actions

Does the adding chain that generates more __tile1024i initialization here?
But why do we need to generate more initialization for the same row and col?

pengfei: Does the adding chain that generates more `__tile1024i` initialization here? But why do we need…
LuoYuankeAuthorUnsubmitted
Done
ReplyInline Actions

Adding chain doesn't generate more __tile1024i initialization. There 5 movb for row, 5 movw for column and movb $1, ... for palette. There are 5 tmm register allocated, so the initialization is correct.

LuoYuanke: Adding chain doesn't generate more `__tile1024i` initialization. There 5 movb for row, 5 movw…
LuoYuankeAuthorUnsubmitted
Done
ReplyInline Actions

One of the tilestore is deleted during instruction selection if we don't add chain.

LuoYuanke: One of the tilestore is deleted during instruction selection if we don't add chain.
pengfeiUnsubmitted
Not Done
ReplyInline Actions

I see. Thanks.

pengfei: I see. Thanks.
; CHECK-NEXT: movw $64, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $16, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movw $64, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $16, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movw $64, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movb $16, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movw $64, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: ldtilecfg -{{[0-9]+}}(%rsp)
; CHECK-NEXT: movl $64, %r8d
; CHECK-NEXT: movw $64, %cx
; CHECK-NEXT: movw $16, %ax
; CHECK-NEXT: tileloadd (%rdi,%r8), %tmm0
; CHECK-NEXT: addq $1024, %rdi # imm = 0x400
; CHECK-NEXT: tileloadd (%rdi,%r8), %tmm1
; CHECK-NEXT: tileloadd (%rdx,%r8), %tmm3
; CHECK-NEXT: leaq 1024(%rdx), %rdi
; CHECK-NEXT: tileloadd (%rdi,%r8), %tmm2
; CHECK-NEXT: tileloadd (%rsi,%r8), %tmm4
; CHECK-NEXT: tdpbssd %tmm4, %tmm0, %tmm3
; CHECK-NEXT: tilestored %tmm3, (%rdx,%r8)
; CHECK-NEXT: tdpbssd %tmm4, %tmm1, %tmm2
; CHECK-NEXT: tilestored %tmm2, (%rdi,%r8)
; CHECK-NEXT: tilerelease
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
entry:
%a1 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %A_mem, i64 64)
%addr = getelementptr inbounds i8, i8* %A_mem, i64 1024
%a2 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %addr, i64 64)
%c1 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64)
%caddr = getelementptr inbounds i8, i8* %C_mem, i64 1024
%c2 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64)
br label %dotpd
dotpd:
%b = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %B_mem, i64 64)
%dp1 = call <256 x i32> @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, <256 x i32> %c1, <256 x i32> %a1, <256 x i32> %b)
call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64, <256 x i32> %dp1)
%dp2 = call <256 x i32> @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, <256 x i32> %c2, <256 x i32> %a2, <256 x i32> %b)
call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64, <256 x i32> %dp2)
ret void
}
declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>)
declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>)
pengfeiUnsubmitted
Done
ReplyInline Actions

I think you can add avx512f in "target-features" and remove -mattr=+amx-int8 -mattr=+avx512f in run line.

pengfei: I think you can add `avx512f` in `"target-features"` and remove `-mattr=+amx-int8…