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

[FRONTEND] Added ExpandDimsOp primitive (#36)

Browse Source
This commit is contained in:
Philippe Tillet
2022-08-04 18:41:06 -07:00
committed by GitHub
parent a7b49b3227
commit 78ebbe24c7
8 changed files with 98 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -10,7 +10,7 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [128] ; Divisibility: [65536] ; Constancy: [1]
%1 = tt.make_range {end = 128 : i32, start = 0 : i32} : tensor<128xi32>
// CHECK-NEXT: Contiguity: [128, 1] ; Divisibility: [65536, 1] ; Constancy: [1, 1]
%2 = tt.view %0 : (tensor<128xi32>) -> tensor<128x1xi32>
%2 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<128xi32>) -> tensor<128x1xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [128, 1]
%3 = tt.splat %arg1 : (i32) -> tensor<128x1xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [1048576, 16] ; Constancy: [1, 1]
@@ -20,7 +20,7 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 1]
%6 = tt.getelementptr %5, %4 : tensor<128x1x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [1, 1]
%7 = tt.view %1 : (tensor<128xi32>) -> tensor<1x128xi32>
%7 = tt.expand_dims %1 {axis = 0 : i32}: (tensor<128xi32>) -> tensor<1x128xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 128]
%8 = tt.broadcast %6 : (tensor<128x1x!tt.ptr<f32>>) -> tensor<128x128x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [128, 1]
@@ -28,13 +28,13 @@ func @permute_2d(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {t
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 16] ; Constancy: [1, 1]
%10 = tt.getelementptr %8, %9 : tensor<128x128x!tt.ptr<f32>>
// CHECK-NEXT: Contiguity: [128, 1] ; Divisibility: [65536, 1] ; Constancy: [1, 1]
%11 = tt.view %0 : (tensor<128xi32>) -> tensor<128x1xi32>
%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>>
// CHECK-NEXT: Contiguity: [1, 128] ; Divisibility: [1, 65536] ; Constancy: [1, 1]
%14 = tt.view %1 : (tensor<128xi32>) -> tensor<1x128xi32>
%14 = tt.expand_dims %1 {axis = 0 : i32} : (tensor<128xi32>) -> tensor<1x128xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 16] ; Constancy: [1, 128]
%15 = tt.splat %arg3 : (i32) -> tensor<1x128xi32>
// CHECK-NEXT: Contiguity: [1, 1] ; Divisibility: [16, 1048576] ; Constancy: [1, 1]

4
test/TritonGPU/coalesce.mlir
View File

@@ -22,12 +22,12 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%cst = arith.constant dense<true> : tensor<64x64xi1, #blocked1>
%cst_0 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
%0 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #blocked0>
%1 = tt.view %0 : (tensor<64xi32, #blocked0>) -> tensor<64x1xi32, #blocked1>
%1 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<64xi32, #blocked0>) -> tensor<64x1xi32, #blocked1>
%2 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, #blocked1>
%3 = arith.muli %1, %2 : tensor<64x1xi32, #blocked1>
%4 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, #blocked1>
%5 = tt.getelementptr %4, %3 : tensor<64x1x!tt.ptr<f32>, #blocked1>
%6 = tt.view %0 : (tensor<64xi32, #blocked0>) -> tensor<1x64xi32, #blocked2>
%6 = tt.expand_dims %0 {axis = 0 : i32} : (tensor<64xi32, #blocked0>) -> tensor<1x64xi32, #blocked2>
%7 = tt.broadcast %5 : (tensor<64x1x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked1>
%8 = tt.broadcast %6 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2>
%9 = triton_gpu.convert_layout %8 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1>