.triton/dump/3791d630ba27aad5e647360045d2f1b5/triton_.ttgir

#blocked = #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [2], order = [0], CTAsPerCGA = [1], CTASplitNum = [1], CTAOrder = [0]}>
module attributes {"triton_gpu.compute-capability" = 89 : i32, "triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 2 : i32, "triton_gpu.threads-per-warp" = 32 : i32} {
  tt.func public @triton__0d1d2d3d4d5d6d7de8de(%arg0: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<bf16, 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<f32, 1> {tt.divisibility = 16 : i32}, %arg5: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg6: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg7: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg8: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}) attributes {noinline = false} {
    %cst = arith.constant dense<256> : tensor<256xi32, #blocked>
    %cst_0 = arith.constant dense<2.560000e+02> : tensor<1xf32, #blocked>
    %cst_1 = arith.constant 0.000000e+00 : f32
    %c256_i32 = arith.constant 256 : i32
    %cst_2 = arith.constant dense<0.000000e+00> : tensor<256xf32, #blocked>
    %cst_3 = arith.constant dense<2.560000e+02> : tensor<256xf32, #blocked>
    %cst_4 = arith.constant dense<0.000000e+00> : tensor<256xbf16, #blocked>
    %0 = tt.get_program_id x : i32
    %1 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #blocked>
    %2 = arith.cmpi slt, %1, %cst : tensor<256xi32, #blocked>
    %3 = arith.muli %0, %c256_i32 : i32
    %4 = tt.splat %3 : (i32) -> tensor<256xi32, #blocked>
    %5 = arith.addi %1, %4 : tensor<256xi32, #blocked>
    %6 = tt.splat %arg1 : (!tt.ptr<bf16, 1>) -> tensor<256x!tt.ptr<bf16, 1>, #blocked>
    %7 = tt.addptr %6, %5 : tensor<256x!tt.ptr<bf16, 1>, #blocked>, tensor<256xi32, #blocked>
    %8 = tt.load %7, %2, %cst_4 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xbf16, #blocked>
    %9 = arith.extf %8 : tensor<256xbf16, #blocked> to tensor<256xf32, #blocked>
    %10 = tt.splat %arg2 : (!tt.ptr<f32, 1>) -> tensor<256x!tt.ptr<f32, 1>, #blocked>
    %11 = tt.addptr %10, %1 : tensor<256x!tt.ptr<f32, 1>, #blocked>, tensor<256xi32, #blocked>
    %12 = tt.load %11, %2, %cst_2 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<256xf32, #blocked>
    %13 = tt.splat %arg3 : (!tt.ptr<f32, 1>) -> tensor<256x!tt.ptr<f32, 1>, #blocked>
    %14 = tt.addptr %13, %5 : tensor<256x!tt.ptr<f32, 1>, #blocked>, tensor<256xi32, #blocked>
    %15 = tt.load %14, %2, %cst_2 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32, #blocked>
    %16 = tt.addptr %arg4, %0 : !tt.ptr<f32, 1>, i32
    %17 = tt.splat %16 : (!tt.ptr<f32, 1>) -> tensor<1x!tt.ptr<f32, 1>, #blocked>
    %18 = tt.load %17 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1xf32, #blocked>
    %19 = tt.addptr %arg5, %0 : !tt.ptr<f32, 1>, i32
    %20 = tt.splat %19 : (!tt.ptr<f32, 1>) -> tensor<1x!tt.ptr<f32, 1>, #blocked>
    %21 = tt.load %20 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1xf32, #blocked>
    %22 = tt.splat %arg0 : (!tt.ptr<f32, 1>) -> tensor<256x!tt.ptr<f32, 1>, #blocked>
    %23 = tt.addptr %22, %5 : tensor<256x!tt.ptr<f32, 1>, #blocked>, tensor<256xi32, #blocked>
    %24 = tt.load %23, %2, %cst_2 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32, #blocked>
    %25 = arith.mulf %9, %12 : tensor<256xf32, #blocked>
    %26 = arith.select %2, %25, %cst_2 : tensor<256xi1, #blocked>, tensor<256xf32, #blocked>
    %27 = "tt.reduce"(%26) <{axis = 0 : i32}> ({
    ^bb0(%arg9: f32, %arg10: f32):
      %50 = arith.addf %arg9, %arg10 : f32
      tt.reduce.return %50 : f32
    }) : (tensor<256xf32, #blocked>) -> f32
    %28 = arith.addf %27, %cst_1 : f32
    %29 = tt.broadcast %18 : (tensor<1xf32, #blocked>) -> tensor<256xf32, #blocked>
    %30 = arith.subf %15, %29 : tensor<256xf32, #blocked>
    %31 = tt.broadcast %21 : (tensor<1xf32, #blocked>) -> tensor<256xf32, #blocked>
    %32 = arith.mulf %30, %31 : tensor<256xf32, #blocked>
    %33 = arith.mulf %25, %32 : tensor<256xf32, #blocked>
    %34 = arith.select %2, %33, %cst_2 : tensor<256xi1, #blocked>, tensor<256xf32, #blocked>
    %35 = "tt.reduce"(%34) <{axis = 0 : i32}> ({
    ^bb0(%arg9: f32, %arg10: f32):
      %50 = arith.addf %arg9, %arg10 : f32
      tt.reduce.return %50 : f32
    }) : (tensor<256xf32, #blocked>) -> f32
    %36 = arith.addf %35, %cst_1 : f32
    %37 = arith.divf %21, %cst_0 : tensor<1xf32, #blocked>
    %38 = arith.mulf %25, %cst_3 : tensor<256xf32, #blocked>
    %39 = tt.splat %28 : (f32) -> tensor<256xf32, #blocked>
    %40 = arith.subf %38, %39 : tensor<256xf32, #blocked>
    %41 = tt.splat %36 : (f32) -> tensor<256xf32, #blocked>
    %42 = arith.mulf %32, %41 : tensor<256xf32, #blocked>
    %43 = arith.subf %40, %42 : tensor<256xf32, #blocked>
    %44 = tt.broadcast %37 : (tensor<1xf32, #blocked>) -> tensor<256xf32, #blocked>
    %45 = arith.mulf %44, %43 : tensor<256xf32, #blocked>
    %46 = arith.addf %24, %45 : tensor<256xf32, #blocked>
    tt.store %23, %46, %2 {cache = 1 : i32, evict = 1 : i32} : tensor<256xf32, #blocked>
    %47 = tt.splat %arg6 : (!tt.ptr<bf16, 1>) -> tensor<256x!tt.ptr<bf16, 1>, #blocked>
    %48 = tt.addptr %47, %5 : tensor<256x!tt.ptr<bf16, 1>, #blocked>, tensor<256xi32, #blocked>
    %49 = arith.truncf %46 : tensor<256xf32, #blocked> to tensor<256xbf16, #blocked>
    tt.store %48, %49, %2 {cache = 1 : i32, evict = 1 : i32} : tensor<256xbf16, #blocked>
    tt.return
  }
}