[TritonIR] Make mask operand optional (#74)

This commit is contained in:
Shintaro Iwasaki
2022-08-22 22:00:17 -07:00
committed by GitHub
parent de2dd04c8a
commit 0ebef11c77
14 changed files with 113 additions and 102 deletions

View File

@@ -3,7 +3,7 @@
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1]}>
#blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 4], order = [0, 1]}>
// CHECK: [[row_layout:#.*]] = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 16], warpsPerCTA = [1, 4], order = [1, 0]}>
// CHECK: [[col_layout:#.*]] = #triton_gpu.blocked<{sizePerThread = [4, 1], threadsPerWarp = [16, 2], warpsPerCTA = [4, 1], order = [0, 1]}>
@@ -15,9 +15,9 @@
// CHECK: [[store_val:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xf32, [[col_layout]]>
// CHECK: [[store_mask:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xi1, [[col_layout]]>
// CHECK: tt.store [[store_ptr]], [[store_val]], [[store_mask]]
func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%arg1: i32 {tt.divisibility = 16 : i32},
%arg2: !tt.ptr<f32> {tt.divisibility = 16 : i32},
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}) {
%cst = arith.constant dense<true> : tensor<64x64xi1, #blocked1>
%cst_0 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
@@ -41,6 +41,6 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%17 = triton_gpu.convert_layout %16 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1>
%18 = tt.getelementptr %15, %17 : tensor<64x64x!tt.ptr<f32>, #blocked1>
%19 = tt.load %10, %cst, %cst_0 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<64x64xf32, #blocked1>
tt.store %18, %19, %cst, : tensor<64x64xf32, #blocked1>
tt.store %18, %19, %cst : tensor<64x64xf32, #blocked1>
return
}
}

View File

@@ -86,7 +86,7 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {tt
// 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.getelementptr %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: tt.store %22, %23, %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>
@@ -117,7 +117,7 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {tt
%24 = triton_gpu.convert_layout %18 : (tensor<64x64x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked4>
%25 = triton_gpu.convert_layout %23 : (tensor<64x64xf32, #blocked1>) -> tensor<64x64xf32, #blocked4>
%26 = triton_gpu.convert_layout %cst_0 : (tensor<64x64xi1, #blocked1>) -> tensor<64x64xi1, #blocked4>
tt.store %24, %25, %26, : tensor<64x64xf32, #blocked4>
tt.store %24, %25, %26 : tensor<64x64xf32, #blocked4>
return
}
@@ -170,6 +170,6 @@ func @loop(%arg0: !tt.ptr<f32>, %arg1: i32, %arg2: !tt.ptr<f32>, %arg3: i32, %ar
%20 = triton_gpu.convert_layout %19 : (tensor<64x64x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked1>
%21 = triton_gpu.convert_layout %11#0 : (tensor<64x64xf32, #blocked1>) -> tensor<64x64xf32, #blocked1>
%22 = triton_gpu.convert_layout %cst : (tensor<64x64xi1, #blocked1>) -> tensor<64x64xi1, #blocked1>
tt.store %20, %21, %22, : tensor<64x64xf32, #blocked1>
tt.store %20, %21, %22 : tensor<64x64xf32, #blocked1>
return
}
}