File size: 6,413 Bytes
6f0bac9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 |
module {
tt.func public @triton__0d1d2d3d4d5d6d7de8(%arg0: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<i64, 1> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg3: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg4: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg5: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg6: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg7: i64 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg8: i64) attributes {noinline = false} {
%cst = arith.constant dense<0.000000e+00> : tensor<64x4xbf16>
%cst_0 = arith.constant dense<0.000000e+00> : tensor<64x1xf32>
%c50257_i32 = arith.constant 50257 : i32
%c4_i32 = arith.constant 4 : i32
%c0_i32 = arith.constant 0 : i32
%cst_1 = arith.constant dense<50257> : tensor<64x1xi64>
%cst_2 = arith.constant dense<50257> : tensor<1x4xi64>
%c64_i64 = arith.constant 64 : i64
%cst_3 = arith.constant dense<-1> : tensor<64x1xi64>
%cst_4 = arith.constant dense<0.000000e+00> : tensor<64x4xf32>
%0 = tt.get_program_id x : i32
%1 = arith.extsi %0 : i32 to i64
%2 = arith.muli %1, %c64_i64 : i64
%3 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32>
%4 = tt.expand_dims %3 {axis = 1 : i32} : (tensor<64xi32>) -> tensor<64x1xi32>
%5 = arith.extsi %4 : tensor<64x1xi32> to tensor<64x1xi64>
%6 = tt.splat %2 : (i64) -> tensor<64x1xi64>
%7 = arith.addi %6, %5 : tensor<64x1xi64>
%8 = tt.make_range {end = 4 : i32, start = 0 : i32} : tensor<4xi32>
%9 = tt.expand_dims %8 {axis = 0 : i32} : (tensor<4xi32>) -> tensor<1x4xi32>
%10 = arith.extsi %9 : tensor<1x4xi32> to tensor<1x4xi64>
%11 = tt.splat %arg1 : (!tt.ptr<i64, 1>) -> tensor<64x1x!tt.ptr<i64, 1>>
%12 = tt.addptr %11, %7 : tensor<64x1x!tt.ptr<i64, 1>>, tensor<64x1xi64>
%13 = tt.load %12 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<64x1xi64>
%14 = tt.addptr %arg2, %c0_i32 : !tt.ptr<f32, 1>, i32
%15 = tt.load %14 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : f32
%16 = tt.addptr %arg3, %c0_i32 : !tt.ptr<f32, 1>, i32
%17 = tt.load %16 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : f32
%18 = arith.muli %7, %cst_1 : tensor<64x1xi64>
%19 = tt.broadcast %18 : (tensor<64x1xi64>) -> tensor<64x4xi64>
%20 = tt.splat %arg0 : (!tt.ptr<f32, 1>) -> tensor<64x4x!tt.ptr<f32, 1>>
%21 = arith.cmpi ne, %13, %cst_3 : tensor<64x1xi64>
%22 = arith.divf %15, %17 : f32
%23 = tt.splat %22 : (f32) -> tensor<64x1xf32>
%24 = arith.select %21, %23, %cst_0 : tensor<64x1xi1>, tensor<64x1xf32>
%25 = tt.broadcast %24 : (tensor<64x1xf32>) -> tensor<64x4xf32>
%26 = scf.for %arg9 = %c0_i32 to %c50257_i32 step %c4_i32 iter_args(%arg10 = %cst_4) -> (tensor<64x4xf32>) : i32 {
%41 = arith.extsi %arg9 : i32 to i64
%42 = tt.splat %41 : (i64) -> tensor<1x4xi64>
%43 = arith.addi %42, %10 : tensor<1x4xi64>
%44 = arith.cmpi slt, %43, %cst_2 : tensor<1x4xi64>
%45 = tt.broadcast %43 : (tensor<1x4xi64>) -> tensor<64x4xi64>
%46 = arith.addi %45, %19 : tensor<64x4xi64>
%47 = tt.addptr %20, %46 : tensor<64x4x!tt.ptr<f32, 1>>, tensor<64x4xi64>
%48 = tt.broadcast %44 : (tensor<1x4xi1>) -> tensor<64x4xi1>
%49 = tt.load %47, %48, %cst_4 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<64x4xf32>
%50 = arith.mulf %49, %25 : tensor<64x4xf32>
%51 = arith.addf %arg10, %50 : tensor<64x4xf32>
%52 = arith.select %48, %51, %arg10 : tensor<64x4xi1>, tensor<64x4xf32>
scf.yield %52 : tensor<64x4xf32>
}
%27 = "tt.reduce"(%26) <{axis = 1 : i32}> ({
^bb0(%arg9: f32, %arg10: f32):
%41 = arith.addf %arg9, %arg10 : f32
tt.reduce.return %41 : f32
}) : (tensor<64x4xf32>) -> tensor<64xf32>
%28 = tt.expand_dims %27 {axis = 1 : i32} : (tensor<64xf32>) -> tensor<64x1xf32>
%29 = arith.muli %7, %cst_1 : tensor<64x1xi64>
%30 = tt.broadcast %29 : (tensor<64x1xi64>) -> tensor<64x4xi64>
%31 = tt.splat %arg4 : (!tt.ptr<bf16, 1>) -> tensor<64x4x!tt.ptr<bf16, 1>>
%32 = tt.splat %arg0 : (!tt.ptr<f32, 1>) -> tensor<64x4x!tt.ptr<f32, 1>>
%33 = tt.splat %arg5 : (!tt.ptr<bf16, 1>) -> tensor<64x4x!tt.ptr<bf16, 1>>
%34 = arith.cmpi ne, %13, %cst_3 : tensor<64x1xi64>
%35 = arith.divf %15, %17 : f32
%36 = tt.splat %35 : (f32) -> tensor<64x1xf32>
%37 = arith.select %34, %36, %cst_0 : tensor<64x1xi1>, tensor<64x1xf32>
%38 = tt.broadcast %37 : (tensor<64x1xf32>) -> tensor<64x4xf32>
%39 = tt.broadcast %28 : (tensor<64x1xf32>) -> tensor<64x4xf32>
%40 = tt.splat %arg6 : (!tt.ptr<bf16, 1>) -> tensor<64x4x!tt.ptr<bf16, 1>>
scf.for %arg9 = %c0_i32 to %c50257_i32 step %c4_i32 : i32 {
%41 = arith.extsi %arg9 : i32 to i64
%42 = tt.splat %41 : (i64) -> tensor<1x4xi64>
%43 = arith.addi %42, %10 : tensor<1x4xi64>
%44 = arith.cmpi slt, %43, %cst_2 : tensor<1x4xi64>
%45 = tt.broadcast %43 : (tensor<1x4xi64>) -> tensor<64x4xi64>
%46 = arith.addi %45, %30 : tensor<64x4xi64>
%47 = tt.addptr %31, %46 : tensor<64x4x!tt.ptr<bf16, 1>>, tensor<64x4xi64>
%48 = tt.broadcast %44 : (tensor<1x4xi1>) -> tensor<64x4xi1>
%49 = tt.load %47, %48, %cst {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x4xbf16>
%50 = arith.extf %49 : tensor<64x4xbf16> to tensor<64x4xf32>
%51 = tt.addptr %32, %46 : tensor<64x4x!tt.ptr<f32, 1>>, tensor<64x4xi64>
%52 = tt.load %51, %48, %cst_4 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x4xf32>
%53 = tt.addptr %33, %46 : tensor<64x4x!tt.ptr<bf16, 1>>, tensor<64x4xi64>
%54 = tt.load %53, %48, %cst {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x4xbf16>
%55 = arith.extf %54 : tensor<64x4xbf16> to tensor<64x4xf32>
%56 = arith.mulf %52, %38 : tensor<64x4xf32>
%57 = math.exp %55 : tensor<64x4xf32>
%58 = arith.mulf %57, %39 : tensor<64x4xf32>
%59 = arith.subf %56, %58 : tensor<64x4xf32>
%60 = arith.addf %50, %59 : tensor<64x4xf32>
%61 = tt.addptr %40, %46 : tensor<64x4x!tt.ptr<bf16, 1>>, tensor<64x4xi64>
%62 = arith.truncf %60 : tensor<64x4xf32> to tensor<64x4xbf16>
tt.store %61, %62, %48 {cache = 1 : i32, evict = 1 : i32} : tensor<64x4xbf16>
}
tt.return
}
}
|