gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[OPTIMIZER] Made layout simplification pass efficient for fused attention kernels (#790)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-21 16:52:15 -07:00
committed by GitHub
parent c4726333bf
commit bb0f9235d1
26 changed files with 683 additions and 229 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -662,7 +662,7 @@ module attributes {"triton_gpu.num-warps" = 1 : i32} {
// CHECK-SAME: mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32
// CHECK: llvm.inline_asm
// CHECK-SAME: mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32
%D = tt.dot %AA, %BB, %cst0 {allowTF32 = true} : tensor<16x16xf16, #shared0> * tensor<16x16xf16, #shared0> -> tensor<16x16xf32, #mma0>
%D = tt.dot %AA, %BB, %cst0 {allowTF32 = true, transA = false, transB = false} : tensor<16x16xf16, #shared0> * tensor<16x16xf16, #shared0> -> tensor<16x16xf32, #mma0>
return
}