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

4
test/Analysis/test-alias.mlir
View File

@@ -26,8 +26,8 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
%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_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

8
test/Analysis/test-alignment.mlir
View File

@@ -18,7 +18,7 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [128, 1]
%5 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<128x1x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 1]
%6 = tt.getelementptr %5, %4 : tensor<128x1x!tt.ptr<f32>>
%6 = tt.addptr %5, %4 : tensor<128x1x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [1, 1]
%7 = tt.expand_dims %1 {axis = 0 : i32}: (tensor<128xi32>) -> tensor<1x128xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 128]
@@ -26,13 +26,13 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [128, 1]
%9 = tt.broadcast %7 : (tensor<1x128xi32>) -> tensor<128x128xi32>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 16] ; Constancy: [1, 1]
%10 = tt.getelementptr %8, %9 : tensor<128x128x!tt.ptr<f32>>
%10 = tt.addptr %8, %9 : tensor<128x128x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [128, 1] ; Divisibility: [65536, 1] ; Constancy: [1, 1]
%11 = tt.expand_dims %0 {axis = 1 : i32}: (tensor<128xi32>) -> tensor<128x1xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [128, 1]
%12 = tt.splat %arg2 : (!tt.ptr<f32>) -> tensor<128x1x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [128, 1] ; Divisibility: [16, 1] ; Constancy: [1, 1]
%13 = tt.getelementptr %12, %11 : tensor<128x1x!tt.ptr<f32>>
%13 = tt.addptr %12, %11 : tensor<128x1x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [1, 1]
%14 = tt.expand_dims %1 {axis = 0 : i32} : (tensor<128xi32>) -> tensor<1x128xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 128]
@@ -44,7 +44,7 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 1048576] ; Constancy: [128, 1]
%18 = tt.broadcast %16 : (tensor<1x128xi32>) -> tensor<128x128xi32>
// CHECK-NEXT: Contiguity: [128, 1] ; Divisibility: [16, 1] ; Constancy: [1, 1]
%19 = tt.getelementptr %17, %18 : tensor<128x128x!tt.ptr<f32>>
%19 = tt.addptr %17, %18 : tensor<128x128x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [1, 1] ; Constancy: [1, 1]
%20 = tt.load %10, %cst, %cst_0 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x128xf32>
tt.store %19, %20, %cst : tensor<128x128xf32>

4
test/Analysis/test-allocation.mlir
View File

@@ -30,8 +30,8 @@ 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

4
test/Analysis/test-membar.mlir
View File

@@ -28,8 +28,8 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
// CHECK: Membar 13
%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