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

[FRONTEND] Add scalar type support for some ops (#661)

Browse Source 
This PR adds basic support for scalar-type inputs to some ops (cast and pointer arithmetics) for Triton-MLIR. Also renames getelementptr -> addptr
This commit is contained in:
Shintaro Iwasaki
2022-09-15 16:12:52 -07:00
committed by GitHub
parent 2e08450c80
commit 43be75ad42
27 changed files with 203 additions and 129 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
test/TritonGPU/loop-pipeline.mlir
View File

@@ -55,8 +55,8 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_a_ptr = tt.getelementptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.getelementptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
%next_a_ptr = tt.addptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
scf.yield %next_a_ptr, %next_b_ptr, %c : tensor<128x32x!tt.ptr<f16>, #AL>, tensor<32x128x!tt.ptr<f16>, #BL>, tensor<128x128xf32, #C>
}
return
@@ -112,8 +112,8 @@ func @matmul_loop_nested(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_a_ptr = tt.getelementptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.getelementptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
%next_a_ptr = tt.addptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
scf.yield %next_a_ptr, %next_b_ptr, %c : tensor<128x32x!tt.ptr<f16>, #AL>, tensor<32x128x!tt.ptr<f16>, #BL>, tensor<128x128xf32, #C>
}
}
@@ -161,7 +161,7 @@ func @matmul_loop_single_pipeline(%lb : index, %ub : index, %step : index, %A :
%b_ = tt.load %b_ptr, %b_mask, %b_other {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<32x128xf16, #BL>
%b = triton_gpu.convert_layout %b_ : (tensor<32x128xf16, #BL>) -> tensor<32x128xf16, #B>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_b_ptr = tt.getelementptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
scf.yield %next_b_ptr, %c : tensor<32x128x!tt.ptr<f16>, #BL>, tensor<128x128xf32, #C>
}
return