File size: 11,002 Bytes
d742687 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 |
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [1, 32], warpsPerCTA = [2, 2], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 2], threadsPerWarp = [1, 32], warpsPerCTA = [1, 4], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
module attributes {"triton_gpu.compute-capability" = 89 : i32, "triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32, "triton_gpu.threads-per-warp" = 32 : i32} {
tt.func public @triton__0d1d2d3d4d5de6de(%arg0: !tt.ptr<i64, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<f32, 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: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg6: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}) attributes {noinline = false} {
%cst = arith.constant dense<512> : tensor<2x1xi32, #blocked>
%cst_0 = arith.constant dense<256> : tensor<1x256xi32, #blocked>
%cst_1 = arith.constant dense<256> : tensor<1x256xi32, #blocked1>
%cst_2 = arith.constant dense<256> : tensor<2x1xi32, #blocked>
%cst_3 = arith.constant dense<1.000000e+00> : tensor<1x256xf32, #blocked>
%cst_4 = arith.constant dense<0.000000e+00> : tensor<1x256xf32, #blocked>
%cst_5 = arith.constant dense<256> : tensor<2x1xi64, #blocked>
%cst_6 = arith.constant dense<50257> : tensor<2x1xi64, #blocked>
%cst_7 = arith.constant dense<0> : tensor<2x1xi64, #blocked>
%cst_8 = arith.constant dense<0> : tensor<2x1xi64, #blocked2>
%cst_9 = arith.constant dense<50257> : tensor<2x1xi64, #blocked2>
%cst_10 = arith.constant 0.000000e+00 : f32
%cst_11 = arith.constant dense<9.99999974E-6> : tensor<2x1xf32, #blocked>
%cst_12 = arith.constant dense<2.560000e+02> : tensor<2x1xf32, #blocked>
%cst_13 = arith.constant dense<0.000000e+00> : tensor<2x256xf32, #blocked>
%cst_14 = arith.constant dense<0.000000e+00> : tensor<1x256xf32, #blocked1>
%c2_i32 = arith.constant 2 : i32
%0 = tt.get_program_id x : i32
%1 = arith.muli %0, %c2_i32 : i32
%2 = tt.make_range {end = 2 : i32, start = 0 : i32} : tensor<2xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>
%3 = tt.make_range {end = 2 : i32, start = 0 : i32} : tensor<2xi32, #triton_gpu.slice<{dim = 1, parent = #blocked2}>>
%4 = tt.expand_dims %2 {axis = 1 : i32} : (tensor<2xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<2x1xi32, #blocked>
%5 = tt.expand_dims %3 {axis = 1 : i32} : (tensor<2xi32, #triton_gpu.slice<{dim = 1, parent = #blocked2}>>) -> tensor<2x1xi32, #blocked2>
%6 = tt.splat %1 : (i32) -> tensor<2x1xi32, #blocked>
%7 = tt.splat %1 : (i32) -> tensor<2x1xi32, #blocked2>
%8 = arith.addi %6, %4 : tensor<2x1xi32, #blocked>
%9 = arith.addi %7, %5 : tensor<2x1xi32, #blocked2>
%10 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>>
%11 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu.slice<{dim = 0, parent = #blocked1}>>
%12 = tt.expand_dims %10 {axis = 0 : i32} : (tensor<256xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>>) -> tensor<1x256xi32, #blocked>
%13 = tt.expand_dims %11 {axis = 0 : i32} : (tensor<256xi32, #triton_gpu.slice<{dim = 0, parent = #blocked1}>>) -> tensor<1x256xi32, #blocked1>
%14 = tt.splat %arg0 : (!tt.ptr<i64, 1>) -> tensor<2x1x!tt.ptr<i64, 1>, #blocked>
%15 = tt.splat %arg0 : (!tt.ptr<i64, 1>) -> tensor<2x1x!tt.ptr<i64, 1>, #blocked2>
%16 = tt.addptr %14, %8 : tensor<2x1x!tt.ptr<i64, 1>, #blocked>, tensor<2x1xi32, #blocked>
%17 = tt.addptr %15, %9 : tensor<2x1x!tt.ptr<i64, 1>, #blocked2>, tensor<2x1xi32, #blocked2>
%18 = tt.load %16 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<2x1xi64, #blocked>
%19 = tt.load %17 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<2x1xi64, #blocked2>
%20 = arith.remsi %8, %cst : tensor<2x1xi32, #blocked>
%21 = arith.cmpi slt, %12, %cst_0 : tensor<1x256xi32, #blocked>
%22 = arith.cmpi slt, %13, %cst_1 : tensor<1x256xi32, #blocked1>
%23 = arith.muli %20, %cst_2 : tensor<2x1xi32, #blocked>
%24 = tt.broadcast %12 : (tensor<1x256xi32, #blocked>) -> tensor<2x256xi32, #blocked>
%25 = tt.broadcast %23 : (tensor<2x1xi32, #blocked>) -> tensor<2x256xi32, #blocked>
%26 = arith.addi %24, %25 : tensor<2x256xi32, #blocked>
%27 = tt.splat %arg2 : (!tt.ptr<f32, 1>) -> tensor<2x256x!tt.ptr<f32, 1>, #blocked>
%28 = tt.addptr %27, %26 : tensor<2x256x!tt.ptr<f32, 1>, #blocked>, tensor<2x256xi32, #blocked>
%29 = tt.broadcast %21 : (tensor<1x256xi1, #blocked>) -> tensor<2x256xi1, #blocked>
%30 = tt.load %28, %29, %cst_13 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<2x256xf32, #blocked>
%31 = arith.addi %18, %cst_6 : tensor<2x1xi64, #blocked>
%32 = arith.addi %19, %cst_9 : tensor<2x1xi64, #blocked2>
%33 = arith.cmpi slt, %18, %cst_7 : tensor<2x1xi64, #blocked>
%34 = arith.cmpi slt, %19, %cst_8 : tensor<2x1xi64, #blocked2>
%35 = arith.select %33, %31, %18 : tensor<2x1xi1, #blocked>, tensor<2x1xi64, #blocked>
%36 = arith.select %34, %32, %19 : tensor<2x1xi1, #blocked2>, tensor<2x1xi64, #blocked2>
%37 = arith.cmpi sge, %36, %cst_8 : tensor<2x1xi64, #blocked2>
%38 = arith.cmpi slt, %36, %cst_9 : tensor<2x1xi64, #blocked2>
%39 = arith.andi %37, %38 : tensor<2x1xi1, #blocked2>
tt.assert %39, "index out of bounds: 0 <= tmp3 < 50257", "<frozen importlib._bootstrap_external>", "_call_with_frames_removed", 883 : tensor<2x1xi1, #blocked2>
%40 = arith.muli %35, %cst_5 : tensor<2x1xi64, #blocked>
%41 = tt.broadcast %40 : (tensor<2x1xi64, #blocked>) -> tensor<2x256xi64, #blocked>
%42 = arith.extsi %12 : tensor<1x256xi32, #blocked> to tensor<1x256xi64, #blocked>
%43 = tt.broadcast %42 : (tensor<1x256xi64, #blocked>) -> tensor<2x256xi64, #blocked>
%44 = arith.addi %43, %41 : tensor<2x256xi64, #blocked>
%45 = tt.splat %arg1 : (!tt.ptr<f32, 1>) -> tensor<2x256x!tt.ptr<f32, 1>, #blocked>
%46 = tt.addptr %45, %44 : tensor<2x256x!tt.ptr<f32, 1>, #blocked>, tensor<2x256xi64, #blocked>
%47 = tt.load %46, %29, %cst_13 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<2x256xf32, #blocked>
%48 = arith.addf %47, %30 : tensor<2x256xf32, #blocked>
%49 = arith.addf %48, %cst_13 : tensor<2x256xf32, #blocked>
%50 = arith.subf %48, %49 : tensor<2x256xf32, #blocked>
%51 = arith.mulf %48, %50 : tensor<2x256xf32, #blocked>
%52 = arith.addf %51, %cst_13 : tensor<2x256xf32, #blocked>
%53 = arith.select %29, %49, %cst_13 : tensor<2x256xi1, #blocked>, tensor<2x256xf32, #blocked>
%54 = arith.select %29, %52, %cst_13 : tensor<2x256xi1, #blocked>, tensor<2x256xf32, #blocked>
%55 = arith.select %21, %cst_3, %cst_4 : tensor<1x256xi1, #blocked>, tensor<1x256xf32, #blocked>
%56 = tt.broadcast %55 : (tensor<1x256xf32, #blocked>) -> tensor<2x256xf32, #blocked>
%57:3 = "tt.reduce"(%53, %54, %56) <{axis = 1 : i32}> ({
^bb0(%arg7: f32, %arg8: f32, %arg9: f32, %arg10: f32, %arg11: f32, %arg12: f32):
%82 = arith.subf %arg10, %arg7 : f32
%83 = arith.addf %arg9, %arg12 : f32
%84 = arith.cmpf oeq, %83, %cst_10 : f32
%85 = arith.divf %arg12, %83 : f32
%86 = arith.select %84, %cst_10, %85 : f32
%87 = arith.mulf %82, %86 : f32
%88 = arith.addf %arg7, %87 : f32
%89 = arith.addf %arg8, %arg11 : f32
%90 = arith.mulf %82, %82 : f32
%91 = arith.mulf %90, %arg9 : f32
%92 = arith.mulf %91, %86 : f32
%93 = arith.addf %89, %92 : f32
tt.reduce.return %88, %93, %83 : f32, f32, f32
}) : (tensor<2x256xf32, #blocked>, tensor<2x256xf32, #blocked>, tensor<2x256xf32, #blocked>) -> (tensor<2xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>, tensor<2xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>, tensor<2xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>)
%58 = tt.expand_dims %57#0 {axis = 1 : i32} : (tensor<2xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<2x1xf32, #blocked>
%59 = tt.expand_dims %57#1 {axis = 1 : i32} : (tensor<2xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<2x1xf32, #blocked>
%60 = tt.load %28, %29, %cst_13 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<2x256xf32, #blocked>
%61 = tt.splat %arg3 : (!tt.ptr<f32, 1>) -> tensor<1x256x!tt.ptr<f32, 1>, #blocked1>
%62 = tt.addptr %61, %13 : tensor<1x256x!tt.ptr<f32, 1>, #blocked1>, tensor<1x256xi32, #blocked1>
%63 = tt.load %62, %22, %cst_14 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x256xf32, #blocked1>
tt.assert %39, "index out of bounds: 0 <= tmp13 < 50257", "<frozen importlib._bootstrap_external>", "_call_with_frames_removed", 883 : tensor<2x1xi1, #blocked2>
%64 = tt.load %46, %29, %cst_13 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<2x256xf32, #blocked>
%65 = arith.addf %64, %60 : tensor<2x256xf32, #blocked>
%66 = tt.broadcast %58 : (tensor<2x1xf32, #blocked>) -> tensor<2x256xf32, #blocked>
%67 = arith.subf %65, %66 : tensor<2x256xf32, #blocked>
%68 = arith.divf %59, %cst_12 : tensor<2x1xf32, #blocked>
%69 = arith.addf %68, %cst_11 : tensor<2x1xf32, #blocked>
%70 = tt.extern_elementwise %69 {libname = "libdevice", libpath = "/usr/local/lib/python3.10/dist-packages/triton/language/../third_party/cuda/lib/libdevice.10.bc", pure = true, symbol = "__nv_rsqrtf"} : (tensor<2x1xf32, #blocked>) -> tensor<2x1xf32, #blocked>
%71 = tt.broadcast %70 : (tensor<2x1xf32, #blocked>) -> tensor<2x256xf32, #blocked>
%72 = arith.mulf %67, %71 : tensor<2x256xf32, #blocked>
%73 = triton_gpu.convert_layout %63 : (tensor<1x256xf32, #blocked1>) -> tensor<1x256xf32, #blocked>
%74 = tt.broadcast %73 : (tensor<1x256xf32, #blocked>) -> tensor<2x256xf32, #blocked>
%75 = arith.mulf %72, %74 : tensor<2x256xf32, #blocked>
%76 = arith.muli %8, %cst_2 : tensor<2x1xi32, #blocked>
%77 = tt.broadcast %76 : (tensor<2x1xi32, #blocked>) -> tensor<2x256xi32, #blocked>
%78 = arith.addi %24, %77 : tensor<2x256xi32, #blocked>
%79 = tt.splat %arg4 : (!tt.ptr<bf16, 1>) -> tensor<2x256x!tt.ptr<bf16, 1>, #blocked>
%80 = tt.addptr %79, %78 : tensor<2x256x!tt.ptr<bf16, 1>, #blocked>, tensor<2x256xi32, #blocked>
%81 = arith.truncf %75 : tensor<2x256xf32, #blocked> to tensor<2x256xbf16, #blocked>
tt.store %80, %81, %29 {cache = 1 : i32, evict = 1 : i32} : tensor<2x256xbf16, #blocked>
tt.return
}
}
|