158 lines · plain
1// RUN: mlir-opt %s -transform-interpreter -split-input-file -verify-diagnostics2 3func.func @set_desc_layout(%arg0: memref<4096x4096xf16>) {4 %c32 = arith.constant 32 : index // expected-note {{target op}}5 return6}7 8module attributes {transform.with_named_sequence} {9 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {10 %0 = transform.structured.match ops{["arith.constant"]} in %arg1 : (!transform.any_op) -> !transform.any_op11 // expected-error@below {{Expected a xegpu.create_nd_desc op, but got: arith.constant}}12 %1 = transform.xegpu.set_desc_layout %0 sg_layout = [8, 4] sg_data = [32, 32] : (!transform.any_op) -> !transform.any_op13 transform.yield14 }15}16 17// -----18 19// CHECK-LABEL: @set_op_layout_attr_bad_result_index20func.func @set_op_layout_attr_bad_result_index(%arg0: memref<4096x4096xf16>) {21 %0 = xegpu.create_nd_tdesc %arg0 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x32xf16>22 %1 = xegpu.load_nd %0[0, 0] : !xegpu.tensor_desc<256x32xf16> -> vector<256x32xf16>23 %2 = arith.extf %1 : vector<256x32xf16> to vector<256x32xf32>24 return25}26 27module attributes {transform.with_named_sequence} {28 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {29 %0 = transform.structured.match ops{["arith.extf"]} in %arg1 : (!transform.any_op) -> !transform.any_op30 // expected-error@below {{Index exceeds the number of op results}}31 transform.xegpu.set_op_layout_attr %0 result index = 1 sg_layout = [8, 4] sg_data = [32, 64] : !transform.any_op32 transform.yield33 }34}35 36// -----37 38// CHECK-LABEL: @set_op_layout_attr_bad_operand_index39func.func @set_op_layout_attr_bad_operand_index(%arg0: memref<4096x4096xf16>) {40 %0 = xegpu.create_nd_tdesc %arg0 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x32xf16>41 %1 = xegpu.load_nd %0[0, 0] : !xegpu.tensor_desc<256x32xf16> -> vector<256x32xf16>42 %2 = arith.extf %1 : vector<256x32xf16> to vector<256x32xf32>43 return44}45 46module attributes {transform.with_named_sequence} {47 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {48 %0 = transform.structured.match ops{["arith.extf"]} in %arg1 : (!transform.any_op) -> !transform.any_op49 // expected-error@below {{Index exceeds the number of op operands}}50 transform.xegpu.set_op_layout_attr %0 index = 1 sg_layout = [8, 4] sg_data = [32, 64] : !transform.any_op51 transform.yield52 }53}54 55// -----56 57// CHECK-LABEL: @set_op_layout_attr_multiple58func.func @set_op_layout_attr_multiple(%arg0: memref<4096x4096xf16>) {59 %0 = xegpu.create_nd_tdesc %arg0 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x32xf16>60 %1 = xegpu.load_nd %0[0, 0] : !xegpu.tensor_desc<256x32xf16> -> vector<256x32xf16>61 %2 = arith.extf %1 : vector<256x32xf16> to vector<256x32xf32>62 %3 = arith.extf %2 : vector<256x32xf32> to vector<256x32xf64>63 return64}65 66module attributes {transform.with_named_sequence} {67 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {68 %0 = transform.structured.match ops{["arith.extf"]} in %arg1 : (!transform.any_op) -> !transform.any_op69 // expected-error@below {{Requires exactly one targetOp handle (got 2)}}70 transform.xegpu.set_op_layout_attr %0 sg_layout = [8, 4] sg_data = [32, 64] : !transform.any_op71 transform.yield72 }73}74 75// -----76 77func.func @set_gpu_launch_threads_bad_handle(%arg0: memref<4096x4096xf16>) {78 %c32 = arith.constant 32 : index // expected-note {{target op}}79 return80}81 82module attributes {transform.with_named_sequence} {83 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {84 %0 = transform.structured.match ops{["arith.constant"]} in %arg1 : (!transform.any_op) -> !transform.any_op85 // expected-error@below {{Expected a gpu.launch op, but got: arith.constant}}86 transform.xegpu.set_gpu_launch_threads %0 threads = [8, 4, 1] : !transform.any_op87 transform.yield88 }89}90 91// -----92 93func.func @set_gpu_launch_threads_many_handles(%arg0: memref<4096x4096xf16>) {94 %c32 = arith.constant 32 : index95 %c64 = arith.constant 64 : index96 return97}98 99module attributes {transform.with_named_sequence} {100 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {101 %0 = transform.structured.match ops{["arith.constant"]} in %arg1 : (!transform.any_op) -> !transform.any_op102 // expected-error@below {{Requires exactly one targetOp handle (got 2)}}103 transform.xegpu.set_gpu_launch_threads %0 threads = [8, 4, 1] : !transform.any_op104 transform.yield105 }106}107 108// -----109 110func.func @set_gpu_launch_threads_bad_threads(%arg0: memref<4096x4096xf16>) {111 %c1 = arith.constant 1 : index112 %c16 = arith.constant 16 : index113 gpu.launch blocks(%arg3, %arg4, %arg5) in (%arg9 = %c16, %arg10 = %c16, %arg11 = %c1) threads(%arg6, %arg7, %arg8) in (%arg12 = %c1, %arg13 = %c1, %arg14 = %c1) {114 gpu.terminator115 }116 return117}118 119module attributes {transform.with_named_sequence} {120 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {121 %0 = transform.structured.match ops{["gpu.launch"]} in %arg1 : (!transform.any_op) -> !transform.any_op122 // expected-error@below {{Expected threads argument to consist of three values (got 2)}}123 transform.xegpu.set_gpu_launch_threads %0 threads = [8, 4] : !transform.any_op124 transform.yield125 }126}127 128// -----129 130// CHECK-LABEL: @insert_prefetch_dpas_c131func.func @insert_prefetch_dpas_c(%arg0: memref<4096x4096xf16>, %arg1: memref<4096x4096xf16>, %arg2: memref<4096x4096xf16>) {132 %c32 = arith.constant 32 : index133 %c4096 = arith.constant 4096 : index134 %c0 = arith.constant 0 : index135 %0 = xegpu.create_nd_tdesc %arg2 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x256xf16>136 // expected-note@below {{load op}}137 %1 = xegpu.load_nd %0[%c0, %c0] : !xegpu.tensor_desc<256x256xf16> -> vector<256x256xf16>138 %3 = xegpu.create_nd_tdesc %arg0 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x32xf16>139 %4 = xegpu.create_nd_tdesc %arg1 : memref<4096x4096xf16> -> !xegpu.tensor_desc<32x256xf16>140 %2 = scf.for %arg3 = %c0 to %c4096 step %c32 iter_args(%arg4 = %1) -> (vector<256x256xf16>) {141 %5 = xegpu.load_nd %3[%c0, %arg3] : !xegpu.tensor_desc<256x32xf16> -> vector<256x32xf16>142 %6 = xegpu.load_nd %4[%arg3, %c0] : !xegpu.tensor_desc<32x256xf16> -> vector<32x256xf16>143 %7 = xegpu.dpas %5, %6, %arg4 : vector<256x32xf16>, vector<32x256xf16>, vector<256x256xf16> -> vector<256x256xf16>144 scf.yield %7 : vector<256x256xf16>145 }146 return147}148 149module attributes {transform.with_named_sequence} {150 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {151 %0 = transform.structured.match ops{["xegpu.dpas"]} in %arg0 : (!transform.any_op) -> !transform.any_op152 %1 = transform.get_operand %0[2] : (!transform.any_op) -> !transform.any_value153 // expected-error@below {{Load op is not contained in a scf.for loop.}}154 %2 = transform.xegpu.insert_prefetch %1 nb_prefetch = 1 : (!transform.any_value) -> !transform.any_op155 transform.yield156 }157}158