.triton/dump/284d053ae6736ef59b97361c588791fb/triton_.ttgir

#blocked = #triton_gpu.blocked<{sizePerThread = [4, 1], threadsPerWarp = [16, 2], warpsPerCTA = [1, 4], order = [0, 1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [2, 2], order = [0, 1], 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__0d1d2d3de4de(%arg0: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg3: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg4: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}) attributes {noinline = false} {
    %cst = arith.constant dense<256> : tensor<64x1xi32, #blocked>
    %cst_0 = arith.constant dense<32768> : tensor<64x1xi32, #blocked>
    %cst_1 = arith.constant dense<256> : tensor<1x8xi32, #blocked>
    %cst_2 = arith.constant dense<128> : tensor<1x8xi32, #blocked>
    %c0_i32 = arith.constant 0 : i32
    %c128_i32 = arith.constant 128 : i32
    %c8_i32 = arith.constant 8 : i32
    %cst_3 = arith.constant dense<0.000000e+00> : tensor<64x8xf32, #blocked>
    %cst_4 = arith.constant dense<0.000000e+00> : tensor<64x8xbf16, #blocked>
    %c64_i32 = arith.constant 64 : i32
    %0 = tt.get_program_id x : i32
    %1 = arith.muli %0, %c64_i32 : i32
    %2 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>
    %3 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = #blocked1}>>
    %4 = tt.expand_dims %2 {axis = 1 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<64x1xi32, #blocked>
    %5 = tt.expand_dims %3 {axis = 1 : i32} : (tensor<64xi32, #triton_gpu.slice<{dim = 1, parent = #blocked1}>>) -> tensor<64x1xi32, #blocked1>
    %6 = tt.splat %1 : (i32) -> tensor<64x1xi32, #blocked>
    %7 = tt.splat %1 : (i32) -> tensor<64x1xi32, #blocked1>
    %8 = arith.addi %6, %4 : tensor<64x1xi32, #blocked>
    %9 = arith.addi %7, %5 : tensor<64x1xi32, #blocked1>
    %10 = tt.make_range {end = 8 : i32, start = 0 : i32} : tensor<8xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>>
    %11 = tt.expand_dims %10 {axis = 0 : i32} : (tensor<8xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>>) -> tensor<1x8xi32, #blocked>
    %12 = arith.remsi %8, %cst : tensor<64x1xi32, #blocked>
    %13 = arith.divsi %8, %cst : tensor<64x1xi32, #blocked>
    %14 = tt.broadcast %12 : (tensor<64x1xi32, #blocked>) -> tensor<64x8xi32, #blocked>
    %15 = arith.muli %13, %cst_0 : tensor<64x1xi32, #blocked>
    %16 = tt.broadcast %15 : (tensor<64x1xi32, #blocked>) -> tensor<64x8xi32, #blocked>
    %17 = tt.splat %arg0 : (!tt.ptr<bf16, 1>) -> tensor<64x8x!tt.ptr<bf16, 1>, #blocked>
    %18 = tt.splat %arg1 : (!tt.ptr<f32, 1>) -> tensor<64x8x!tt.ptr<f32, 1>, #blocked>
    %19 = scf.for %arg5 = %c0_i32 to %c128_i32 step %c8_i32 iter_args(%arg6 = %cst_3) -> (tensor<64x8xf32, #blocked>)  : i32 {
      %25 = tt.splat %arg5 : (i32) -> tensor<1x8xi32, #blocked>
      %26 = arith.addi %25, %11 : tensor<1x8xi32, #blocked>
      %27 = arith.cmpi slt, %26, %cst_2 : tensor<1x8xi32, #blocked>
      %28 = arith.muli %26, %cst_1 : tensor<1x8xi32, #blocked>
      %29 = tt.broadcast %28 : (tensor<1x8xi32, #blocked>) -> tensor<64x8xi32, #blocked>
      %30 = arith.addi %14, %29 : tensor<64x8xi32, #blocked>
      %31 = arith.addi %30, %16 : tensor<64x8xi32, #blocked>
      %32 = tt.addptr %17, %31 : tensor<64x8x!tt.ptr<bf16, 1>, #blocked>, tensor<64x8xi32, #blocked>
      %33 = tt.broadcast %27 : (tensor<1x8xi1, #blocked>) -> tensor<64x8xi1, #blocked>
      %34 = tt.load %32, %33, %cst_4 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x8xbf16, #blocked>
      %35 = arith.extf %34 : tensor<64x8xbf16, #blocked> to tensor<64x8xf32, #blocked>
      %36 = tt.addptr %18, %31 : tensor<64x8x!tt.ptr<f32, 1>, #blocked>, tensor<64x8xi32, #blocked>
      %37 = tt.load %36, %33, %cst_3 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x8xf32, #blocked>
      %38 = arith.mulf %35, %37 : tensor<64x8xf32, #blocked>
      %39 = arith.addf %arg6, %38 : tensor<64x8xf32, #blocked>
      %40 = arith.select %33, %39, %arg6 : tensor<64x8xi1, #blocked>, tensor<64x8xf32, #blocked>
      scf.yield %40 : tensor<64x8xf32, #blocked>
    }
    %20 = "tt.reduce"(%19) <{axis = 1 : i32}> ({
    ^bb0(%arg5: f32, %arg6: f32):
      %25 = arith.addf %arg5, %arg6 : f32
      tt.reduce.return %25 : f32
    }) : (tensor<64x8xf32, #blocked>) -> tensor<64xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>
    %21 = triton_gpu.convert_layout %20 : (tensor<64xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<64xf32, #triton_gpu.slice<{dim = 1, parent = #blocked1}>>
    %22 = tt.expand_dims %21 {axis = 1 : i32} : (tensor<64xf32, #triton_gpu.slice<{dim = 1, parent = #blocked1}>>) -> tensor<64x1xf32, #blocked1>
    %23 = tt.splat %arg2 : (!tt.ptr<f32, 1>) -> tensor<64x1x!tt.ptr<f32, 1>, #blocked1>
    %24 = tt.addptr %23, %9 : tensor<64x1x!tt.ptr<f32, 1>, #blocked1>, tensor<64x1xi32, #blocked1>
    tt.store %24, %22 {cache = 1 : i32, evict = 1 : i32} : tensor<64x1xf32, #blocked1>
    tt.return
  }
}