[llvm] 4ed2b6c - [X86][AMX] Fix a bug on tile config.
via llvm-commits
llvm-commits at lists.llvm.org
Wed May 26 06:59:12 PDT 2021
Author: Luo, Yuanke
Date: 2021-05-26T21:57:49+08:00
New Revision: 4ed2b6cccdef912013c84244c3f5ee4c018de9b1
URL: https://github.com/llvm/llvm-project/commit/4ed2b6cccdef912013c84244c3f5ee4c018de9b1
DIFF: https://github.com/llvm/llvm-project/commit/4ed2b6cccdef912013c84244c3f5ee4c018de9b1.diff
LOG: [X86][AMX] Fix a bug on tile config.
The previous code detect if a MBB is bottom block to determine if it is
a backedge of a loop. We should check latch block instead of bottom
block and we should check the header and the bottom block are in the
same loop.
Differential Revision: https://reviews.llvm.org/D103145
Added:
llvm/test/CodeGen/X86/AMX/amx-gemm.ll
Modified:
llvm/lib/Target/X86/X86PreTileConfig.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86PreTileConfig.cpp b/llvm/lib/Target/X86/X86PreTileConfig.cpp
index 2b3a7ac019111..b85a0b61d6f64 100644
--- a/llvm/lib/Target/X86/X86PreTileConfig.cpp
+++ b/llvm/lib/Target/X86/X86PreTileConfig.cpp
@@ -125,8 +125,13 @@ class X86PreTileConfig : public MachineFunctionPass {
/// Check if it is an edge from loop bottom to loop head.
bool isLoopBackEdge(MachineBasicBlock *Header, MachineBasicBlock *Bottom) {
- return MLI->isLoopHeader(Header) &&
- MLI->getLoopFor(Header)->getBottomBlock() == Bottom;
+ if (!MLI->isLoopHeader(Header))
+ return false;
+ auto *ML = MLI->getLoopFor(Header);
+ if (ML->contains(Bottom) && ML->isLoopLatch(Bottom))
+ return true;
+
+ return false;
}
/// Collect the shape def information for later use.
diff --git a/llvm/test/CodeGen/X86/AMX/amx-gemm.ll b/llvm/test/CodeGen/X86/AMX/amx-gemm.ll
new file mode 100644
index 0000000000000..8f5d0c7383187
--- /dev/null
+++ b/llvm/test/CodeGen/X86/AMX/amx-gemm.ll
@@ -0,0 +1,210 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -verify-machineinstrs | FileCheck %s
+
+; #include <immintrin.h>
+;
+; #define TILE_SZ 16
+; void inner_product(int *A_mem, int *B_mem, int *C_mem, int M, int N, int K) {
+; const int m = M / TILE_SZ;
+; const int n = N / TILE_SZ;
+; const int k = K / TILE_SZ;
+;
+; for (int i = 0; i < m; i++)
+; for (int j = 0; j < n; j++) {
+; __tile1024i c = {TILE_SZ, TILE_SZ*sizeof(int)};
+; __tile_zero(&c);
+; for (int l = 0; l < k; l++) {
+; __tile1024i a = {TILE_SZ, TILE_SZ*sizeof(int)};
+; __tile1024i b = {TILE_SZ, TILE_SZ*sizeof(int)};
+; __tile_loadd(&a, A_mem+(i*TILE_SZ)*K+l*TILE_SZ, K*sizeof(int));
+; __tile_loadd(&b, B_mem+(l*TILE_SZ)*N+j*TILE_SZ, N*sizeof(int));
+; __tile_dpbssd(&c, a, b);
+; }
+; __tile_stored(C_mem+(i*TILE_SZ)*M+j*TILE_SZ, N*sizeof(int), c);
+; }
+; }
+
+; CHECK: ldtilecfg
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @inner_product(i32* %A_mem, i32* %B_mem, i32* %C_mem, i32 %M, i32 %N, i32 %K) local_unnamed_addr {
+entry:
+ %mul = shl i32 %K, 4
+ %conv = sext i32 %K to i64
+ %mul15 = shl nsw i64 %conv, 2
+ %conv23 = sext i32 %N to i64
+ %mul24 = shl nsw i64 %conv23, 2
+ %cmp8163 = icmp sgt i32 %K, 15
+ %mul25 = shl i32 %M, 4
+ %cmp4173 = icmp sgt i32 %N, 15
+ %cmp187 = icmp sgt i32 %M, 15
+ br i1 %cmp187, label %for.cond3.preheader.preheader, label %for.cond.cleanup
+
+for.cond3.preheader.preheader: ; preds = %entry
+ %div2 = sdiv i32 %K, 16
+ %div1 = sdiv i32 %N, 16
+ %div209 = lshr i32 %M, 4
+ %wide.trip.count207 = zext i32 %div209 to i64
+ %wide.trip.count203 = zext i32 %div1 to i64
+ %wide.trip.count = zext i32 %div2 to i64
+ %0 = add nsw i64 %wide.trip.count, -1
+ %xtraiter = and i64 %wide.trip.count, 7
+ %1 = icmp ult i64 %0, 7
+ %unroll_iter = and i64 %wide.trip.count, 4294967288
+ %lcmp.mod.not = icmp eq i64 %xtraiter, 0
+ br label %for.cond3.preheader
+
+for.cond3.preheader: ; preds = %for.cond3.preheader.preheader, %for.cond.cleanup5
+ %indvars.iv205 = phi i64 [ 0, %for.cond3.preheader.preheader ], [ %indvars.iv.next206, %for.cond.cleanup5 ]
+ %2 = trunc i64 %indvars.iv205 to i32
+ %mul11 = mul i32 %mul, %2
+ %idx.ext = sext i32 %mul11 to i64
+ %add.ptr = getelementptr inbounds i32, i32* %A_mem, i64 %idx.ext
+ %mul26 = mul i32 %mul25, %2
+ %idx.ext27 = sext i32 %mul26 to i64
+ %add.ptr28 = getelementptr inbounds i32, i32* %C_mem, i64 %idx.ext27
+ br i1 %cmp4173, label %for.body6, label %for.cond.cleanup5
+
+for.cond.cleanup: ; preds = %for.cond.cleanup5, %entry
+ ret void
+
+for.cond.cleanup5: ; preds = %for.cond.cleanup9, %for.cond3.preheader
+ %indvars.iv.next206 = add nuw nsw i64 %indvars.iv205, 1
+ %exitcond208.not = icmp eq i64 %indvars.iv.next206, %wide.trip.count207
+ br i1 %exitcond208.not, label %for.cond.cleanup, label %for.cond3.preheader
+
+for.body6: ; preds = %for.cond3.preheader, %for.cond.cleanup9
+ %indvars.iv199 = phi i64 [ %indvars.iv.next200, %for.cond.cleanup9 ], [ 0, %for.cond3.preheader ]
+ %3 = tail call x86_amx @llvm.x86.tilezero.internal(i16 16, i16 64)
+ %4 = shl nsw i64 %indvars.iv199, 4
+ br i1 %cmp8163, label %for.body10.preheader, label %for.cond.cleanup9
+
+for.body10.preheader: ; preds = %for.body6
+ %add.ptr19 = getelementptr inbounds i32, i32* %B_mem, i64 %4
+ br i1 %1, label %for.cond.cleanup9.loopexit.unr-lcssa, label %for.body10
+
+for.cond.cleanup9.loopexit.unr-lcssa: ; preds = %for.body10, %for.body10.preheader
+ %.lcssa.ph = phi x86_amx [ undef, %for.body10.preheader ], [ %68, %for.body10 ]
+ %indvars.iv.unr = phi i64 [ 0, %for.body10.preheader ], [ %indvars.iv.next.7, %for.body10 ]
+ %c.sroa.8127.2.in164.unr = phi x86_amx [ %3, %for.body10.preheader ], [ %68, %for.body10 ]
+ br i1 %lcmp.mod.not, label %for.cond.cleanup9, label %for.body10.epil
+
+for.body10.epil: ; preds = %for.cond.cleanup9.loopexit.unr-lcssa, %for.body10.epil
+ %indvars.iv.epil = phi i64 [ %indvars.iv.next.epil, %for.body10.epil ], [ %indvars.iv.unr, %for.cond.cleanup9.loopexit.unr-lcssa ]
+ %c.sroa.8127.2.in164.epil = phi x86_amx [ %11, %for.body10.epil ], [ %c.sroa.8127.2.in164.unr, %for.cond.cleanup9.loopexit.unr-lcssa ]
+ %epil.iter = phi i64 [ %epil.iter.sub, %for.body10.epil ], [ %xtraiter, %for.cond.cleanup9.loopexit.unr-lcssa ]
+ %5 = shl nsw i64 %indvars.iv.epil, 4
+ %add.ptr14.epil = getelementptr inbounds i32, i32* %add.ptr, i64 %5
+ %6 = bitcast i32* %add.ptr14.epil to i8*
+ %7 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* %6, i64 %mul15)
+ %8 = mul nsw i64 %5, %conv23
+ %add.ptr22.epil = getelementptr inbounds i32, i32* %add.ptr19, i64 %8
+ %9 = bitcast i32* %add.ptr22.epil to i8*
+ %10 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* %9, i64 %mul24)
+ %11 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %c.sroa.8127.2.in164.epil, x86_amx %7, x86_amx %10)
+ %indvars.iv.next.epil = add nuw nsw i64 %indvars.iv.epil, 1
+ %epil.iter.sub = add i64 %epil.iter, -1
+ %epil.iter.cmp.not = icmp eq i64 %epil.iter.sub, 0
+ br i1 %epil.iter.cmp.not, label %for.cond.cleanup9, label %for.body10.epil
+
+for.cond.cleanup9: ; preds = %for.cond.cleanup9.loopexit.unr-lcssa, %for.body10.epil, %for.body6
+ %c.sroa.8127.2.in.lcssa = phi x86_amx [ %3, %for.body6 ], [ %.lcssa.ph, %for.cond.cleanup9.loopexit.unr-lcssa ], [ %11, %for.body10.epil ]
+ %add.ptr31 = getelementptr inbounds i32, i32* %add.ptr28, i64 %4
+ %12 = bitcast i32* %add.ptr31 to i8*
+ tail call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* %12, i64 %mul24, x86_amx %c.sroa.8127.2.in.lcssa)
+ %indvars.iv.next200 = add nuw nsw i64 %indvars.iv199, 1
+ %exitcond204.not = icmp eq i64 %indvars.iv.next200, %wide.trip.count203
+ br i1 %exitcond204.not, label %for.cond.cleanup5, label %for.body6
+
+for.body10: ; preds = %for.body10.preheader, %for.body10
+ %indvars.iv = phi i64 [ %indvars.iv.next.7, %for.body10 ], [ 0, %for.body10.preheader ]
+ %c.sroa.8127.2.in164 = phi x86_amx [ %68, %for.body10 ], [ %3, %for.body10.preheader ]
+ %niter = phi i64 [ %niter.nsub.7, %for.body10 ], [ %unroll_iter, %for.body10.preheader ]
+ %13 = shl nsw i64 %indvars.iv, 4
+ %add.ptr14 = getelementptr inbounds i32, i32* %add.ptr, i64 %13
+ %14 = bitcast i32* %add.ptr14 to i8*
+ %15 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* %14, i64 %mul15)
+ %16 = mul nsw i64 %13, %conv23
+ %add.ptr22 = getelementptr inbounds i32, i32* %add.ptr19, i64 %16
+ %17 = bitcast i32* %add.ptr22 to i8*
+ %18 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* %17, i64 %mul24)
+ %19 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %c.sroa.8127.2.in164, x86_amx %15, x86_amx %18)
+ %indvars.iv.next = shl i64 %indvars.iv, 4
+ %20 = or i64 %indvars.iv.next, 16
+ %add.ptr14.1 = getelementptr inbounds i32, i32* %add.ptr, i64 %20
+ %21 = bitcast i32* %add.ptr14.1 to i8*
+ %22 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %21, i64 %mul15)
+ %23 = mul nsw i64 %20, %conv23
+ %add.ptr22.1 = getelementptr inbounds i32, i32* %add.ptr19, i64 %23
+ %24 = bitcast i32* %add.ptr22.1 to i8*
+ %25 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %24, i64 %mul24)
+ %26 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %19, x86_amx %22, x86_amx %25)
+ %indvars.iv.next.1 = shl i64 %indvars.iv, 4
+ %27 = or i64 %indvars.iv.next.1, 32
+ %add.ptr14.2 = getelementptr inbounds i32, i32* %add.ptr, i64 %27
+ %28 = bitcast i32* %add.ptr14.2 to i8*
+ %29 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %28, i64 %mul15)
+ %30 = mul nsw i64 %27, %conv23
+ %add.ptr22.2 = getelementptr inbounds i32, i32* %add.ptr19, i64 %30
+ %31 = bitcast i32* %add.ptr22.2 to i8*
+ %32 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %31, i64 %mul24)
+ %33 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %26, x86_amx %29, x86_amx %32)
+ %indvars.iv.next.2 = shl i64 %indvars.iv, 4
+ %34 = or i64 %indvars.iv.next.2, 48
+ %add.ptr14.3 = getelementptr inbounds i32, i32* %add.ptr, i64 %34
+ %35 = bitcast i32* %add.ptr14.3 to i8*
+ %36 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %35, i64 %mul15)
+ %37 = mul nsw i64 %34, %conv23
+ %add.ptr22.3 = getelementptr inbounds i32, i32* %add.ptr19, i64 %37
+ %38 = bitcast i32* %add.ptr22.3 to i8*
+ %39 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %38, i64 %mul24)
+ %40 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %33, x86_amx %36, x86_amx %39)
+ %indvars.iv.next.3 = shl i64 %indvars.iv, 4
+ %41 = or i64 %indvars.iv.next.3, 64
+ %add.ptr14.4 = getelementptr inbounds i32, i32* %add.ptr, i64 %41
+ %42 = bitcast i32* %add.ptr14.4 to i8*
+ %43 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %42, i64 %mul15)
+ %44 = mul nsw i64 %41, %conv23
+ %add.ptr22.4 = getelementptr inbounds i32, i32* %add.ptr19, i64 %44
+ %45 = bitcast i32* %add.ptr22.4 to i8*
+ %46 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %45, i64 %mul24)
+ %47 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %40, x86_amx %43, x86_amx %46)
+ %indvars.iv.next.4 = shl i64 %indvars.iv, 4
+ %48 = or i64 %indvars.iv.next.4, 80
+ %add.ptr14.5 = getelementptr inbounds i32, i32* %add.ptr, i64 %48
+ %49 = bitcast i32* %add.ptr14.5 to i8*
+ %50 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %49, i64 %mul15)
+ %51 = mul nsw i64 %48, %conv23
+ %add.ptr22.5 = getelementptr inbounds i32, i32* %add.ptr19, i64 %51
+ %52 = bitcast i32* %add.ptr22.5 to i8*
+ %53 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %52, i64 %mul24)
+ %54 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %47, x86_amx %50, x86_amx %53)
+ %indvars.iv.next.5 = shl i64 %indvars.iv, 4
+ %55 = or i64 %indvars.iv.next.5, 96
+ %add.ptr14.6 = getelementptr inbounds i32, i32* %add.ptr, i64 %55
+ %56 = bitcast i32* %add.ptr14.6 to i8*
+ %57 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %56, i64 %mul15)
+ %58 = mul nsw i64 %55, %conv23
+ %add.ptr22.6 = getelementptr inbounds i32, i32* %add.ptr19, i64 %58
+ %59 = bitcast i32* %add.ptr22.6 to i8*
+ %60 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %59, i64 %mul24)
+ %61 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %54, x86_amx %57, x86_amx %60)
+ %indvars.iv.next.6 = shl i64 %indvars.iv, 4
+ %62 = or i64 %indvars.iv.next.6, 112
+ %add.ptr14.7 = getelementptr inbounds i32, i32* %add.ptr, i64 %62
+ %63 = bitcast i32* %add.ptr14.7 to i8*
+ %64 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %63, i64 %mul15)
+ %65 = mul nsw i64 %62, %conv23
+ %add.ptr22.7 = getelementptr inbounds i32, i32* %add.ptr19, i64 %65
+ %66 = bitcast i32* %add.ptr22.7 to i8*
+ %67 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %66, i64 %mul24)
+ %68 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %61, x86_amx %64, x86_amx %67)
+ %indvars.iv.next.7 = add nuw nsw i64 %indvars.iv, 8
+ %niter.nsub.7 = add i64 %niter, -8
+ %niter.ncmp.7 = icmp eq i64 %niter.nsub.7, 0
+ br i1 %niter.ncmp.7, label %for.cond.cleanup9.loopexit.unr-lcssa, label %for.body10
+}
+
+declare x86_amx @llvm.x86.tilezero.internal(i16, i16)
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
More information about the llvm-commits
mailing list