| #blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 8], 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" = 8 : i32, "triton_gpu.threads-per-warp" = 32 : i32} { | |
| 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<1x1xf32, #blocked> | |
| %cst_0 = arith.constant dense<-1> : tensor<1x1xi64, #blocked> | |
| %cst_1 = arith.constant dense<0.000000e+00> : tensor<1x2048xf32, #blocked> | |
| %cst_2 = arith.constant dense<50257> : tensor<1x2048xi64, #blocked> | |
| %c0_i32 = arith.constant 0 : i32 | |
| %c2048_i32 = arith.constant 2048 : i32 | |
| %c50257_i32 = arith.constant 50257 : i32 | |
| %c50257_i64 = arith.constant 50257 : i64 | |
| %cst_3 = arith.constant dense<0.000000e+00> : tensor<1x2048xbf16, #blocked> | |
| %0 = tt.get_program_id x : i32 | |
| %1 = arith.extsi %0 : i32 to i64 | |
| %2 = tt.make_range {end = 2048 : i32, start = 0 : i32} : tensor<2048xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>> | |
| %3 = tt.expand_dims %2 {axis = 0 : i32} : (tensor<2048xi32, #triton_gpu.slice<{dim = 0, parent = #blocked}>>) -> tensor<1x2048xi32, #blocked> | |
| %4 = arith.extsi %3 : tensor<1x2048xi32, #blocked> to tensor<1x2048xi64, #blocked> | |
| %5 = tt.addptr %arg1, %1 : !tt.ptr<i64, 1>, i64 | |
| %6 = tt.splat %5 : (!tt.ptr<i64, 1>) -> tensor<1x1x!tt.ptr<i64, 1>, #blocked> | |
| %7 = tt.load %6 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x1xi64, #blocked> | |
| %8 = tt.addptr %arg2, %c0_i32 : !tt.ptr<f32, 1>, i32 | |
| %9 = tt.load %8 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : f32 | |
| %10 = tt.addptr %arg3, %c0_i32 : !tt.ptr<f32, 1>, i32 | |
| %11 = tt.load %10 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : f32 | |
| %12 = arith.muli %1, %c50257_i64 : i64 | |
| %13 = tt.splat %12 : (i64) -> tensor<1x2048xi64, #blocked> | |
| %14 = tt.splat %arg0 : (!tt.ptr<f32, 1>) -> tensor<1x2048x!tt.ptr<f32, 1>, #blocked> | |
| %15 = arith.cmpi ne, %7, %cst_0 : tensor<1x1xi64, #blocked> | |
| %16 = arith.divf %9, %11 : f32 | |
| %17 = tt.splat %16 : (f32) -> tensor<1x1xf32, #blocked> | |
| %18 = arith.select %15, %17, %cst : tensor<1x1xi1, #blocked>, tensor<1x1xf32, #blocked> | |
| %19 = tt.broadcast %18 : (tensor<1x1xf32, #blocked>) -> tensor<1x2048xf32, #blocked> | |
| %20 = scf.for %arg9 = %c0_i32 to %c50257_i32 step %c2048_i32 iter_args(%arg10 = %cst_1) -> (tensor<1x2048xf32, #blocked>) : i32 { | |
| %27 = arith.extsi %arg9 : i32 to i64 | |
| %28 = tt.splat %27 : (i64) -> tensor<1x2048xi64, #blocked> | |
| %29 = arith.addi %28, %4 : tensor<1x2048xi64, #blocked> | |
| %30 = arith.cmpi slt, %29, %cst_2 : tensor<1x2048xi64, #blocked> | |
| %31 = arith.addi %29, %13 : tensor<1x2048xi64, #blocked> | |
| %32 = tt.addptr %14, %31 : tensor<1x2048x!tt.ptr<f32, 1>, #blocked>, tensor<1x2048xi64, #blocked> | |
| %33 = tt.load %32, %30, %cst_1 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x2048xf32, #blocked> | |
| %34 = arith.mulf %33, %19 : tensor<1x2048xf32, #blocked> | |
| %35 = arith.addf %arg10, %34 : tensor<1x2048xf32, #blocked> | |
| %36 = arith.select %30, %35, %arg10 : tensor<1x2048xi1, #blocked>, tensor<1x2048xf32, #blocked> | |
| scf.yield %36 : tensor<1x2048xf32, #blocked> | |
| } | |
| %21 = "tt.reduce"(%20) <{axis = 1 : i32}> ({ | |
| ^bb0(%arg9: f32, %arg10: f32): | |
| %27 = arith.addf %arg9, %arg10 : f32 | |
| tt.reduce.return %27 : f32 | |
| }) : (tensor<1x2048xf32, #blocked>) -> tensor<1xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>> | |
| %22 = tt.expand_dims %21 {axis = 1 : i32} : (tensor<1xf32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<1x1xf32, #blocked> | |
| %23 = tt.splat %arg4 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>, #blocked> | |
| %24 = tt.splat %arg5 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>, #blocked> | |
| %25 = tt.broadcast %22 : (tensor<1x1xf32, #blocked>) -> tensor<1x2048xf32, #blocked> | |
| %26 = tt.splat %arg6 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>, #blocked> | |
| scf.for %arg9 = %c0_i32 to %c50257_i32 step %c2048_i32 : i32 { | |
| %27 = arith.extsi %arg9 : i32 to i64 | |
| %28 = tt.splat %27 : (i64) -> tensor<1x2048xi64, #blocked> | |
| %29 = arith.addi %28, %4 : tensor<1x2048xi64, #blocked> | |
| %30 = arith.cmpi slt, %29, %cst_2 : tensor<1x2048xi64, #blocked> | |
| %31 = arith.addi %29, %13 : tensor<1x2048xi64, #blocked> | |
| %32 = tt.addptr %23, %31 : tensor<1x2048x!tt.ptr<bf16, 1>, #blocked>, tensor<1x2048xi64, #blocked> | |
| %33 = tt.load %32, %30, %cst_3 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xbf16, #blocked> | |
| %34 = arith.extf %33 : tensor<1x2048xbf16, #blocked> to tensor<1x2048xf32, #blocked> | |
| %35 = tt.addptr %14, %31 : tensor<1x2048x!tt.ptr<f32, 1>, #blocked>, tensor<1x2048xi64, #blocked> | |
| %36 = tt.load %35, %30, %cst_1 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xf32, #blocked> | |
| %37 = tt.addptr %24, %31 : tensor<1x2048x!tt.ptr<bf16, 1>, #blocked>, tensor<1x2048xi64, #blocked> | |
| %38 = tt.load %37, %30, %cst_3 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xbf16, #blocked> | |
| %39 = arith.extf %38 : tensor<1x2048xbf16, #blocked> to tensor<1x2048xf32, #blocked> | |
| %40 = arith.mulf %36, %19 : tensor<1x2048xf32, #blocked> | |
| %41 = math.exp %39 : tensor<1x2048xf32, #blocked> | |
| %42 = arith.mulf %41, %25 : tensor<1x2048xf32, #blocked> | |
| %43 = arith.subf %40, %42 : tensor<1x2048xf32, #blocked> | |
| %44 = arith.addf %34, %43 : tensor<1x2048xf32, #blocked> | |
| %45 = tt.addptr %26, %31 : tensor<1x2048x!tt.ptr<bf16, 1>, #blocked>, tensor<1x2048xi64, #blocked> | |
| %46 = arith.truncf %44 : tensor<1x2048xf32, #blocked> to tensor<1x2048xbf16, #blocked> | |
| tt.store %45, %46, %30 {cache = 1 : i32, evict = 1 : i32} : tensor<1x2048xbf16, #blocked> | |
| } | |
| tt.return | |
| } | |
| } | |