152 lines · plain
1// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup" \2// RUN: | mlir-runner \3// RUN: --shared-libs=%mlir_levelzero_runtime \4// RUN: --shared-libs=%mlir_runner_utils \5// RUN: --entry-point-result=void \6// RUN: | FileCheck %s7 8#a = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], inst_data = [8, 16]>9#b = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 64], inst_data = [16, 16]>10#c = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 64], inst_data = [8, 16]>11#a_prefetch = #xegpu.layout<sg_layout = [32, 1], sg_data = [8, 32], inst_data = [8, 16]>12#b_prefetch = #xegpu.layout<sg_layout = [4, 8], sg_data = [8, 32], inst_data = [8, 16]>13module @gemm attributes {gpu.container_module} {14 func.func @test(%A: memref<256x256xf16>, %B: memref<256x256xf16>, %C: memref<256x256xf32>) -> memref<256x256xf32> attributes {llvm.emit_c_interface} {15 %c1 = arith.constant 1 : index16 %c4 = arith.constant 4 : index17 %c8 = arith.constant 8 : index18 %c16 = arith.constant 16 : index19 %c32 = arith.constant 32 : index20 %c64 = arith.constant 64 : index21 %c128 = arith.constant 128 : index22 %c512 = arith.constant 512 : index23 %A_gpu = gpu.alloc () : memref<256x256xf16>24 gpu.memcpy %A_gpu, %A : memref<256x256xf16>, memref<256x256xf16>25 %B_gpu = gpu.alloc () : memref<256x256xf16>26 gpu.memcpy %B_gpu, %B : memref<256x256xf16>, memref<256x256xf16>27 %C_gpu = gpu.alloc () : memref<256x256xf32>28 gpu.memcpy %C_gpu, %C : memref<256x256xf32>, memref<256x256xf32>29 // NOTE: Here we can't use [8, 64] wi threads following30 // the SG thread layout of [8, 4]. Because runtime will linearize31 // the x dimension first (we need y dimension to be linearized first).32 // So just use linearized thread layout of [512, 1] wi threads.33 gpu.launch_func @test_kernel::@test_kernel blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%A_gpu : memref<256x256xf16>, %B_gpu : memref<256x256xf16>, %C_gpu : memref<256x256xf32>)34 gpu.wait // Wait for the kernel to finish.35 gpu.memcpy %C, %C_gpu : memref<256x256xf32>, memref<256x256xf32>36 gpu.dealloc %A_gpu : memref<256x256xf16>37 gpu.dealloc %B_gpu : memref<256x256xf16>38 gpu.dealloc %C_gpu : memref<256x256xf32>39 return %C : memref<256x256xf32>40 }41 42 gpu.module @test_kernel {43 gpu.func @test_kernel(%A: memref<256x256xf16>, %B: memref<256x256xf16>, %C: memref<256x256xf32>) kernel {44 %c0 = arith.constant 0 : index45 %c1 = arith.constant 1 : index46 %c32 = arith.constant 32 : index47 %c64 = arith.constant 64 : index48 %c96 = arith.constant 96 : index49 %c256 = arith.constant 256 : index50 %c4096 = arith.constant 4096 : index51 %block_id_x = gpu.block_id x52 %block_id_y = gpu.block_id y53 %m = arith.muli %block_id_x, %c256 : index54 %n = arith.muli %block_id_y, %c256 : index55 %c_tdesc = xegpu.create_nd_tdesc %C : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #c>56 %c_init_value = xegpu.load_nd %c_tdesc[%m, %n] : !xegpu.tensor_desc<256x256xf32, #c> -> vector<256x256xf32>57 %a_tdesc = xegpu.create_nd_tdesc %A : memref<256x256xf16> -> !xegpu.tensor_desc<256x32xf16, #a>58 %b_tdesc = xegpu.create_nd_tdesc %B : memref<256x256xf16> -> !xegpu.tensor_desc<32x256xf16, #b>59 // Prefetch A 3 times.60 %a_prefetch_tdesc = xegpu.create_nd_tdesc %A : memref<256x256xf16> -> !xegpu.tensor_desc<256x32xf16, #a_prefetch>61 xegpu.prefetch_nd %a_prefetch_tdesc[%m, %c0] : !xegpu.tensor_desc<256x32xf16, #a_prefetch>62 xegpu.prefetch_nd %a_prefetch_tdesc[%m, %c32] : !xegpu.tensor_desc<256x32xf16, #a_prefetch>63 xegpu.prefetch_nd %a_prefetch_tdesc[%m, %c64] : !xegpu.tensor_desc<256x32xf16, #a_prefetch>64 // Prefetch B 3 times.65 %b_prefetch_tdesc = xegpu.create_nd_tdesc %B : memref<256x256xf16> -> !xegpu.tensor_desc<32x256xf16, #b_prefetch>66 xegpu.prefetch_nd %b_prefetch_tdesc[%c0, %n] : !xegpu.tensor_desc<32x256xf16, #b_prefetch>67 xegpu.prefetch_nd %b_prefetch_tdesc[%c32, %n] : !xegpu.tensor_desc<32x256xf16, #b_prefetch>68 xegpu.prefetch_nd %b_prefetch_tdesc[%c64, %n] : !xegpu.tensor_desc<32x256xf16, #b_prefetch>69 70 %out = scf.for %k = %c0 to %c256 step %c3271 iter_args(%c_value = %c_init_value)72 -> (vector<256x256xf32>) {73 %a_value = xegpu.load_nd %a_tdesc[%m, %k] : !xegpu.tensor_desc<256x32xf16, #a> -> vector<256x32xf16>74 %b_value = xegpu.load_nd %b_tdesc[%k, %n] : !xegpu.tensor_desc<32x256xf16, #b> -> vector<32x256xf16>75 // Prefetch next tiles.76 %prefetch_offset = arith.addi %k, %c96 : index77 xegpu.prefetch_nd %a_prefetch_tdesc[%m, %prefetch_offset] : !xegpu.tensor_desc<256x32xf16, #a_prefetch>78 xegpu.prefetch_nd %b_prefetch_tdesc[%prefetch_offset, %n] : !xegpu.tensor_desc<32x256xf16, #b_prefetch>79 %c_new_value = xegpu.dpas %a_value, %b_value, %c_value {layout_result_0 = #c}80 : vector<256x32xf16>, vector<32x256xf16>, vector<256x256xf32> -> vector<256x256xf32>81 scf.yield %c_new_value : vector<256x256xf32>82 }83 xegpu.store_nd %out, %c_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #c>84 gpu.return85 }86 }87 88 func.func @main() attributes {llvm.emit_c_interface} {89 %c0 = arith.constant 0 : index90 %c1 = arith.constant 1 : index91 %c1_f16 = arith.constant 1.0 : f1692 %c2_f16 = arith.constant 2.0 : f1693 %c256 = arith.constant 256 : index94 %cf_0 = arith.constant 0.0 : f1695 %cf_1 = arith.constant 1.0 : f1696 %A = memref.alloc() : memref<256x256xf16>97 %B = memref.alloc() : memref<256x256xf16>98 %C = memref.alloc() : memref<256x256xf32>99 %C_ref = memref.alloc() : memref<256x256xf32>100 %c_gen_int = arith.constant 0 : i1101 %cf_lower = arith.constant -0.5 : f32102 %cf_upper = arith.constant 0.5 : f32103 // Intialize matrix A ; A[i, j] = j104 scf.for %i = %c0 to %c256 step %c1 {105 scf.for %j = %c0 to %c256 step %c1 {106 %t = index.castu %j : index to i16107 %val = arith.uitofp %t : i16 to f16108 memref.store %val, %A[%i, %j] : memref<256x256xf16>109 }110 }111 112 // Initialize the B matrix113 // Make matrix B an identity matrix114 scf.for %i = %c0 to %c256 step %c1 {115 scf.for %j = %c0 to %c256 step %c1 {116 %i_i32 = index.castu %i : index to i32117 %j_i32 = index.castu %j : index to i32118 %i_j_same = arith.cmpi eq, %i_i32, %j_i32 : i32119 120 scf.if %i_j_same {121 memref.store %cf_1, %B[%i, %j] : memref<256x256xf16>122 } else {123 memref.store %cf_0, %B[%i, %j] : memref<256x256xf16>124 }125 }126 }127 128 // Initialize matrix C and C_ref ; C[i, j] = 0129 %c0_f32 = arith.constant 0.0 : f32130 scf.for %i = %c0 to %c256 step %c1 {131 scf.for %j = %c0 to %c256 step %c1 {132 memref.store %c0_f32, %C[%i, %j] : memref<256x256xf32>133 memref.store %c0_f32, %C_ref[%i, %j] : memref<256x256xf32>134 }135 }136 137 // Run GPU version.138 %2 = call @test(%A, %B, %C) : (memref<256x256xf16>, memref<256x256xf16>, memref<256x256xf32>) -> memref<256x256xf32>139 %gpu_result_cast = memref.cast %2 : memref<256x256xf32> to memref<*xf32>140 // CHECK: Unranked Memref base@ = 0x{{[0-9a-f]+}}141 // CHECK-COUNT-256: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]142 call @printMemrefF32(%gpu_result_cast) : (memref<*xf32>) -> ()143 144 memref.dealloc %A : memref<256x256xf16>145 memref.dealloc %B : memref<256x256xf16>146 memref.dealloc %C : memref<256x256xf32>147 memref.dealloc %C_ref : memref<256x256xf32>148 return149 }150 func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface}151}152