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

[OPTIMIZER] Updated TritonGPU-combine pass (#784)

Browse Source 
WIP but should work int t…he cases we need so far
This commit is contained in:
Philippe Tillet
2022-10-16 21:19:42 -07:00
committed by GitHub
parent e948a618b3
commit 38a80664b5
2 changed files with 165 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
test/TritonGPU/combine.mlir
View File

@@ -45,8 +45,8 @@ func @remat(%arg0: i32) -> tensor<1024xi32, #layout1> {
// CHECK: %1 = tt.make_range {end = 1024 : i32, start = 0 : i32} : tensor<1024xi32, [[target_layout]]>
// CHECK: %2 = tt.make_range {end = 1024 : i32, start = 0 : i32} : tensor<1024xi32, [[target_layout]]>
// CHECK: %3 = tt.make_range {end = 1024 : i32, start = 0 : i32} : tensor<1024xi32, [[target_layout]]>
// CHECK: %4 = arith.muli %2, %3 : tensor<1024xi32, [[target_layout]]>
// CHECK: %5 = arith.muli %0, %1 : tensor<1024xi32, [[target_layout]]>
// CHECK: %4 = arith.muli %0, %2 : tensor<1024xi32, [[target_layout]]>
// CHECK: %5 = arith.muli %1, %3 : tensor<1024xi32, [[target_layout]]>
// CHECK: %6 = arith.addi %4, %5 : tensor<1024xi32, [[target_layout]]>
// CHECK: return %6 : tensor<1024xi32, [[target_layout]]>
}
@@ -61,34 +61,10 @@ func @remat(%arg0: i32) -> tensor<1024xi32, #layout1> {
// CHECK-LABEL: transpose
func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg3: i32 {tt.divisibility = 16 : i32}) {
// CHECK: %cst = arith.constant dense<true> : tensor<64x64xi1, [[row_layout]]>
// CHECK: %cst_0 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, [[row_layout]]>
// CHECK: %cst_1 = arith.constant dense<true> : tensor<64x64xi1, [[col_layout]]>
// CHECK: %0 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = [[col_layout]]}>>
// CHECK: %1 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = [[row_layout]]}>>
// CHECK: %2 = tt.expand_dims %1 {axis = 1 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = [[row_layout]]}>>) -> tensor<64x1xi32, [[row_layout]]>
// CHECK: %3 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, [[row_layout]]>
// CHECK: %4 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, [[row_layout]]>
// CHECK: %5 = arith.muli %2, %3 : tensor<64x1xi32, [[row_layout]]>
// CHECK: %6 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 0, parent = [[col_layout]]}>>
// CHECK: %7 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 0, parent = [[row_layout]]}>>
// CHECK: %8 = tt.addptr %4, %5 : tensor<64x1x!tt.ptr<f32>, [[row_layout]]>
// CHECK: %9 = tt.expand_dims %7 {axis = 0 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 0, parent = [[row_layout]]}>>) -> tensor<1x64xi32, [[row_layout]]>
// CHECK: %10 = tt.broadcast %8 : (tensor<64x1x!tt.ptr<f32>, [[row_layout]]>) -> tensor<64x64x!tt.ptr<f32>, [[row_layout]]>
// CHECK: %11 = tt.broadcast %9 : (tensor<1x64xi32, [[row_layout]]>) -> tensor<64x64xi32, [[row_layout]]>
// CHECK: %12 = tt.splat %arg2 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, [[col_layout]]>
// CHECK: %13 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = [[col_layout]]}>>) -> tensor<64x1xi32, [[col_layout]]>
// CHECK: %14 = tt.expand_dims %6 {axis = 0 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 0, parent = [[col_layout]]}>>) -> tensor<1x64xi32, [[col_layout]]>
// CHECK: %15 = tt.splat %arg3 : (i32) -> tensor<1x64xi32, [[col_layout]]>
// CHECK: %16 = tt.addptr %12, %13 : tensor<64x1x!tt.ptr<f32>, [[col_layout]]>
// CHECK: %17 = arith.muli %14, %15 : tensor<1x64xi32, [[col_layout]]>
// CHECK: %18 = tt.broadcast %16 : (tensor<64x1x!tt.ptr<f32>, [[col_layout]]>) -> tensor<64x64x!tt.ptr<f32>, [[col_layout]]>
// CHECK: %19 = tt.broadcast %17 : (tensor<1x64xi32, [[col_layout]]>) -> tensor<64x64xi32, [[col_layout]]>
// CHECK: %20 = tt.addptr %10, %11 : tensor<64x64x!tt.ptr<f32>, [[row_layout]]>
// CHECK: %21 = tt.load %20, %cst, %cst_0 {cache = 1 : i32, evict = 1 : i32, isOtherUnspecified = false, isVolatile = false} : tensor<64x64xf32, [[row_layout]]>
// CHECK: %22 = tt.addptr %18, %19 : tensor<64x64x!tt.ptr<f32>, [[col_layout]]>
// CHECK: %23 = triton_gpu.convert_layout %21 : (tensor<64x64xf32, [[row_layout]]>) -> tensor<64x64xf32, [[col_layout]]>
// CHECK: tt.store %22, %23, %cst_1 : tensor<64x64xf32, [[col_layout]]>
// CHECK-NOT: triton_gpu.convert_layout
// CHECK: [[loaded_val:%.*]] = tt.load {{.*}}, %cst, %cst_0 {cache = 1 : i32, evict = 1 : i32, isOtherUnspecified = false, isVolatile = false} : tensor<64x64xf32, [[row_layout]]>
// CHECK: [[cvt_val:%.*]] = triton_gpu.convert_layout [[loaded_val]] : (tensor<64x64xf32, [[row_layout]]>) -> tensor<64x64xf32, [[col_layout]]>
// CHECK: tt.store {{.*}}, [[cvt_val]], %cst_1 : tensor<64x64xf32, [[col_layout]]>
// CHECK: return
%cst = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
%cst_0 = arith.constant dense<true> : tensor<64x64xi1, #blocked1>