// RUN: triton-opt %s -split-input-file -tritongpu-coalesce -canonicalize | FileCheck %s #blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}> #blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1]}> #blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 4], order = [0, 1]}> #slice1dim1 = #triton_gpu.slice<{dim = 1, parent = #blocked1}> #slice2dim0 = #triton_gpu.slice<{dim = 0, parent = #blocked2}> module attributes {"triton_gpu.num-warps" = 4 : i32} { // CHECK: [[row_layout:#.*]] = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 16], warpsPerCTA = [4, 1], order = [1, 0]}> // CHECK: [[col_layout:#.*]] = #triton_gpu.blocked<{sizePerThread = [4, 1], threadsPerWarp = [16, 2], warpsPerCTA = [1, 4], order = [0, 1]}> // CHECK: [[load_ptr:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64x!tt.ptr, [[row_layout]]> // CHECK: [[load_mask:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xi1, [[row_layout]]> // CHECK: [[load_other:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xf32, [[row_layout]]> // CHECK: [[load_val:%.*]] = tt.load [[load_ptr]], [[load_mask]], [[load_other]] {{.*}} : tensor<64x64xf32, [[row_layout]]> // CHECK: [[store_ptr:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64x!tt.ptr, [[col_layout]]> // CHECK: [[store_val:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xf32, [[col_layout]]> // CHECK: [[store_mask:%.*]] = triton_gpu.convert_layout {{.*}} -> tensor<64x64xi1, [[col_layout]]> // CHECK: tt.store [[store_ptr]], [[store_val]], [[store_mask]] func @transpose(%arg0: !tt.ptr {tt.divisibility = 16 : i32}, %arg1: i32 {tt.divisibility = 16 : i32}, %arg2: !tt.ptr {tt.divisibility = 16 : i32}, %arg3: i32 {tt.divisibility = 16 : i32}) { %cst = arith.constant dense : tensor<64x64xi1, #blocked1> %cst_0 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1> %00 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice1dim1> %01 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice2dim0> %1 = tt.expand_dims %00 {axis = 1 : i32} : (tensor<64xi32, #slice1dim1>) -> tensor<64x1xi32, #blocked1> %2 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, #blocked1> %3 = arith.muli %1, %2 : tensor<64x1xi32, #blocked1> %4 = tt.splat %arg0 : (!tt.ptr) -> tensor<64x1x!tt.ptr, #blocked1> %5 = tt.addptr %4, %3 : tensor<64x1x!tt.ptr, #blocked1>, tensor<64x1xi32, #blocked1> %6 = tt.expand_dims %01 {axis = 0 : i32} : (tensor<64xi32, #slice2dim0>) -> tensor<1x64xi32, #blocked2> %7 = tt.broadcast %5 : (tensor<64x1x!tt.ptr, #blocked1>) -> tensor<64x64x!tt.ptr, #blocked1> %8 = tt.broadcast %6 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2> %9 = triton_gpu.convert_layout %8 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1> %10 = tt.addptr %7, %9 : tensor<64x64x!tt.ptr, #blocked1>, tensor<64x64xi32, #blocked1> %11 = tt.splat %arg2 : (!tt.ptr) -> tensor<64x1x!tt.ptr, #blocked1> %12 = tt.addptr %11, %1 : tensor<64x1x!tt.ptr, #blocked1>, tensor<64x1xi32, #blocked1> %13 = tt.splat %arg3 : (i32) -> tensor<1x64xi32, #blocked2> %14 = arith.muli %6, %13 : tensor<1x64xi32, #blocked2> %15 = tt.broadcast %12 : (tensor<64x1x!tt.ptr, #blocked1>) -> tensor<64x64x!tt.ptr, #blocked1> %16 = tt.broadcast %14 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2> %17 = triton_gpu.convert_layout %16 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1> %18 = tt.addptr %15, %17 : tensor<64x64x!tt.ptr, #blocked1>, tensor<64x64xi32, #blocked1> %19 = tt.load %10, %cst, %cst_0 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<64x64xf32, #blocked1> tt.store %18, %19, %cst : tensor<64x64xf32, #blocked1> return } }