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

[FRONTEND][BACKEND] Fixes for cat / reshape / addptr (#959)

Browse Source 
Most notably, this PR:
- changes the traits (and assembly format) of addptr so it can handle offsets that have arbitrary integer width.
- adds support for `cat`
This commit is contained in:
Philippe Tillet
2022-12-06 23:29:50 -08:00
committed by GitHub
parent 981aee7f1e
commit b2b793dfb5
24 changed files with 199 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/tests/test_backend.py
View File

@@ -64,12 +64,12 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
%7 = tt.broadcast %6 : (tensor<1x128xi32, #src>) -> tensor<128x128xi32, #src>
%8 = tt.broadcast %5 : (tensor<128x1xi32, #src>) -> tensor<128x128xi32, #src>
%9 = arith.addi %8, %7 : tensor<128x128xi32, #src>
%10 = tt.addptr %2, %9 : tensor<128x128x!tt.ptr<f16>, #src>
%10 = tt.addptr %2, %9 : tensor<128x128x!tt.ptr<f16>, #src>, tensor<128x128xi32, #src>
%11 = tt.load %10 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x128xf16, #src>
%3 = tt.splat %arg1 : (!tt.ptr<f16>) -> tensor<128x128x!tt.ptr<f16>, #dst>
%12 = triton_gpu.convert_layout %9 : (tensor<128x128xi32, #src>) -> tensor<128x128xi32, #dst>
%13 = triton_gpu.convert_layout %11 : (tensor<128x128xf16, #src>) -> tensor<128x128xf16, #dst>
%14 = tt.addptr %3, %12 : tensor<128x128x!tt.ptr<f16>, #dst>
%14 = tt.addptr %3, %12 : tensor<128x128x!tt.ptr<f16>, #dst>, tensor<128x128xi32, #dst>
tt.store %14, %13 : tensor<128x128xf16, #dst>
return
}