460 lines · plain
1// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx908 | FileCheck %s --check-prefixes=CHECK,GFX9,GFX9082// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx90a | FileCheck %s --check-prefixes=CHECK,GFX9,GFX90A3// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx942 | FileCheck %s --check-prefixes=CHECK,GFX9,GFX9424// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx1030 | FileCheck %s --check-prefixes=CHECK,GFX10,RDNA5// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx1100 | FileCheck %s --check-prefixes=CHECK,GFX11,RDNA6// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx1201 | FileCheck %s --check-prefixes=CHECK,GFX12,RDNA7 8// Note: #gpu.address_space<global> is hardcoded to `1` here because the9// test pass doesn't set up the GPU address space conversions.10 11// CHECK: #[[$MMRA_TAG:.+]] = #llvm.mmra_tag<"amdgpu-synchronize-as":"local">12 13#gpu_global_addrspace = 114 15// CHECK-LABEL: func @fat_raw_buffer_cast16func.func @fat_raw_buffer_cast(%buf: memref<8xi32, #gpu_global_addrspace>) -> memref<8xi32, #amdgpu.address_space<fat_raw_buffer>> {17 // CHECK: %[[desc:.*]] = builtin.unrealized_conversion_cast %{{.*}} : memref<8xi32, 1> to !llvm.struct<(ptr<1>, ptr<1>, i64, array<1 x i64>, array<1 x i64>)>18 // CHECK-DAG: %[[base:.*]] = llvm.extractvalue %[[desc]][1]19 // CHECK-DAG: %[[offset:.*]] = llvm.extractvalue %[[desc]][2]20 // CHECK-DAG: %[[sizes:.*]] = llvm.extractvalue %[[desc]][3]21 // CHECK-DAG: %[[strides:.*]] = llvm.extractvalue %[[desc]][4]22 // CHECK-DAG: %[[numRecords:.*]] = llvm.mlir.constant(32 : i64) : i6423 // CHECK-DAG: %[[strideArg:.*]] = llvm.mlir.constant(0 : i16) : i1624 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)25 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)26 // CHECK: %[[fatBuf:.*]] = rocdl.make.buffer.rsrc %[[base]], %[[strideArg]], %[[numRecords]], %[[flags]] : <1> to <7>27 // CHECK: %[[ret0:.*]] = llvm.mlir.poison : !llvm.struct<(ptr<7>, ptr<7>, i64, array<1 x i64>, array<1 x i64>)>28 // CHECK: %[[ret1:.*]] = llvm.insertvalue %[[fatBuf]], %[[ret0]][0]29 // CHECK: %[[ret2:.*]] = llvm.insertvalue %[[fatBuf]], %[[ret1]][1]30 // CHECK: %[[ret3:.*]] = llvm.insertvalue %[[offset]], %[[ret2]][2]31 // CHECK: %[[ret4:.*]] = llvm.insertvalue %[[sizes]], %[[ret3]][3]32 // CHECK: %[[ret5:.*]] = llvm.insertvalue %[[strides]], %[[ret4]][4]33 // CHECK: builtin.unrealized_conversion_cast %[[ret5]]34 %ret = amdgpu.fat_raw_buffer_cast %buf : memref<8xi32, #gpu_global_addrspace> to memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>35 return %ret : memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>36}37 38// CHECK-LABEL: func @fat_raw_buffer_cast_0d39func.func @fat_raw_buffer_cast_0d(%buf: memref<i32, #gpu_global_addrspace>) -> memref<i32, #amdgpu.address_space<fat_raw_buffer>> {40 // CHECK: %[[desc:.*]] = builtin.unrealized_conversion_cast %{{.*}} : memref<i32, 1> to !llvm.struct<(ptr<1>, ptr<1>, i64)>41 // CHECK-DAG: %[[base:.*]] = llvm.extractvalue %[[desc]][1]42 // CHECK-DAG: %[[offset:.*]] = llvm.extractvalue %[[desc]][2]43 // CHECK-DAG: %[[numRecords:.*]] = llvm.mlir.constant(4 : i64) : i6444 // CHECK-DAG: %[[strideArg:.*]] = llvm.mlir.constant(0 : i16) : i1645 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)46 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)47 // CHECK: %[[fatBuf:.*]] = rocdl.make.buffer.rsrc %[[base]], %[[strideArg]], %[[numRecords]], %[[flags]]48 // CHECK: %[[ret0:.*]] = llvm.mlir.poison : !llvm.struct<(ptr<7>, ptr<7>, i64)>49 // CHECK: %[[ret1:.*]] = llvm.insertvalue %[[fatBuf]], %[[ret0]][0]50 // CHECK: %[[ret2:.*]] = llvm.insertvalue %[[fatBuf]], %[[ret1]][1]51 // CHECK: %[[ret3:.*]] = llvm.insertvalue %[[offset]], %[[ret2]][2]52 // CHECK: builtin.unrealized_conversion_cast %[[ret3]]53 %ret = amdgpu.fat_raw_buffer_cast %buf : memref<i32, #gpu_global_addrspace> to memref<i32, #amdgpu.address_space<fat_raw_buffer>>54 return %ret : memref<i32, #amdgpu.address_space<fat_raw_buffer>>55}56 57// CHECK-LABEL: func @fat_raw_buffer_cast_dyn_size_offset58func.func @fat_raw_buffer_cast_dyn_size_offset(%buf: memref<?xi32, strided<[1], offset: ?>, #gpu_global_addrspace>) -> memref<?xi32, strided<[1], offset: ?>, #amdgpu.address_space<fat_raw_buffer>> {59 // CHECK: %[[size0:.*]] = llvm.extractvalue %{{.*}}[3, 0]60 // CHECK: %[[stride0:.*]] = llvm.extractvalue %{{.*}}[4, 0]61 // CHECK: %[[maxVals:.*]] = llvm.mul %[[size0]], %[[stride0]]62 // CHECK: %[[byteSize:.*]] = llvm.mlir.constant(4 : i64) : i6463 // CHECK: %[[numRecords:.*]] = llvm.mul %[[maxVals]], %[[byteSize]]64 // CHECK: %[[offset:.*]] = llvm.extractvalue %{{.*}}[2]65 // CHECK: rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}66 // CHECK: llvm.insertvalue %[[offset]], %{{.*}}[2]67 %ret = amdgpu.fat_raw_buffer_cast %buf : memref<?xi32, strided<[1], offset: ?>, #gpu_global_addrspace> to memref<?xi32, strided<[1], offset: ?>, #amdgpu.address_space<fat_raw_buffer>>68 return %ret : memref<?xi32, strided<[1], offset: ?>, #amdgpu.address_space<fat_raw_buffer>>69}70 71// CHECK-LABEL: func @fat_raw_buffer_cast_reset_offset72func.func @fat_raw_buffer_cast_reset_offset(%buf: memref<?xi32, strided<[1], offset: ?>, #gpu_global_addrspace>) -> memref<?xi32, #amdgpu.address_space<fat_raw_buffer>> {73 // CHECK: %[[desc:.*]] = builtin.unrealized_conversion_cast %{{.*}} : memref<?xi32, strided<[1], offset: ?>, 1> to !llvm.struct<(ptr<1>, ptr<1>, i64, array<1 x i64>, array<1 x i64>)>74 // CHECK-DAG: %[[memRefPtr:.*]] = llvm.extractvalue %[[desc]][1]75 // CHECK-DAG: %[[memRefOff:.*]] = llvm.extractvalue %[[desc]][2]76 // CHECK-DAG: %[[basePtr:.*]] = llvm.getelementptr %[[memRefPtr]][%[[memRefOff]]]77 // CHECK-DAG: %[[zeroOff:.*]] = llvm.mlir.constant(0 : index) : i6478 // CHECK: %[[fatBuf:.*]] = rocdl.make.buffer.rsrc %[[basePtr]], %{{.*}}, %{{.*}}, %{{.*}}79 // CHECK: llvm.insertvalue %[[fatBuf]], %{{.*}}[1]80 // CHECK: llvm.insertvalue %[[zeroOff]], %{{.*}}[2]81 %ret = amdgpu.fat_raw_buffer_cast %buf resetOffset : memref<?xi32, strided<[1], offset: ?>, #gpu_global_addrspace> to memref<?xi32, #amdgpu.address_space<fat_raw_buffer>>82 return %ret : memref<?xi32, #amdgpu.address_space<fat_raw_buffer>>83}84 85// CHECK-LABEL: func @fat_raw_buffer_cast_valid_bytes86func.func @fat_raw_buffer_cast_valid_bytes(%buf: memref<8xi32, #gpu_global_addrspace>) -> memref<8xi32, #amdgpu.address_space<fat_raw_buffer>> {87 // CHECK: %[[numRecords:.*]] = arith.constant -1 : i6488 // CHECK: rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}89 %cu64_max = arith.constant -1 : i6490 %ret = amdgpu.fat_raw_buffer_cast %buf validBytes(%cu64_max) : memref<8xi32, #gpu_global_addrspace> to memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>91 return %ret : memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>92}93 94// CHECK-LABEL: func @fat_raw_buffer_cast_bounds_check95func.func @fat_raw_buffer_cast_bounds_check(%buf: memref<8xi32, #gpu_global_addrspace>) -> memref<8xi32, #amdgpu.address_space<fat_raw_buffer>> {96 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)97 // RDNA: %[[flags:.*]] = llvm.mlir.constant(553807872 : i32)98 // CHECK: %[[rsrc:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %{{.*}}, %[[flags]]99 %ret = amdgpu.fat_raw_buffer_cast %buf boundsCheck(false) : memref<8xi32, #gpu_global_addrspace> to memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>100 return %ret : memref<8xi32, #amdgpu.address_space<fat_raw_buffer>>101}102 103// CHECK-LABEL: func @fat_raw_buffer_cast_cache_swizzle104// CHECK-SAME: (%{{.*}}: memref<64x64xi32, 1>, %[[stride:.*]]: i14)105func.func @fat_raw_buffer_cast_cache_swizzle(%buf: memref<64x64xi32, #gpu_global_addrspace>, %stride: i14) -> memref<64x64xi32, #amdgpu.address_space<fat_raw_buffer>> {106 // GFX908: %[[stride:.*]] = llvm.mlir.constant(0 : i16) : i16107 // GFX90A: %[[stride:.*]] = llvm.mlir.constant(0 : i16) : i16108 // RDNA: %[[stride:.*]] = llvm.mlir.constant(0 : i16) : i16109 // GFX942: %[[asI16:.*]] = llvm.zext %[[stride]] : i14 to i16110 // GFX942: %[[cacheSwizzleOn:.*]] = llvm.mlir.constant(16384 : i16) : i16111 // GFX942: %[[stride:.*]] = llvm.or disjoint %[[asI16]], %[[cacheSwizzleOn]]112 // CHECK: rocdl.make.buffer.rsrc %{{.*}}, %[[stride]], %{{.*}}, %{{.*}}113 %ret = amdgpu.fat_raw_buffer_cast %buf cacheSwizzleStride(%stride) : memref<64x64xi32, #gpu_global_addrspace> to memref<64x64xi32, #amdgpu.address_space<fat_raw_buffer>>114 return %ret : memref<64x64xi32, #amdgpu.address_space<fat_raw_buffer>>115}116 117// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_scalar_i32118func.func @gpu_gcn_raw_buffer_load_scalar_i32(%buf: memref<i32>) -> i32 {119 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(4 : i64)120 // CHECK: %[[stride:.*]] = llvm.mlir.constant(0 : i16)121 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)122 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)123 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %[[stride]], %[[numRecords]], %[[flags]] : !llvm.ptr to <8>124 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32125 // CHECK: return %[[ret]]126 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[] : memref<i32> -> i32127 func.return %0 : i32128}129 130// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i32131func.func @gpu_gcn_raw_buffer_load_i32(%buf: memref<64xi32>, %idx: i32) -> i32 {132 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)133 // CHECK: %[[stride:.*]] = llvm.mlir.constant(0 : i16)134 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)135 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)136 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %[[stride]], %[[numRecords]], %[[flags]] : !llvm.ptr to <8>137 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32138 // CHECK: return %[[ret]]139 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi32>, i32 -> i32140 func.return %0 : i32141}142 143// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i32_strided144func.func @gpu_gcn_raw_buffer_load_i32_strided(%buf: memref<16x16xi32, strided<[?, ?], offset: ?>>, %i: i32, %j: i32) -> i32 {145 // CHECK: %[[descriptor:.*]] = builtin.unrealized_conversion_cast %{{.*}} : memref<16x16xi32, strided<[?, ?], offset: ?>> to !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>146 // CHECK: %[[elem_size:.*]] = llvm.mlir.constant(4 : i32) : i32147 // CHECK: %[[algn_ptr:.*]] = llvm.extractvalue %[[descriptor]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>148 // CHECK: %[[offset:.*]] = llvm.extractvalue %[[descriptor]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>149 // CHECK: %[[ptr:.*]] = llvm.getelementptr %[[algn_ptr]][%[[offset]]] : (!llvm.ptr, i64) -> !llvm.ptr, i32150 // CHECK: %[[sz_i:.*]] = llvm.extractvalue %[[descriptor]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>151 // CHECK: %[[stride_i:.*]] = llvm.extractvalue %[[descriptor]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>152 // CHECK: %[[ext_i:.*]] = llvm.mul %[[sz_i]], %[[stride_i]] : i64153 // CHECK: %[[sz_j:.*]] = llvm.extractvalue %[[descriptor]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>154 // CHECK: %[[stride_j:.*]] = llvm.extractvalue %[[descriptor]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>155 // CHECK: %[[ext_j:.*]] = llvm.mul %[[sz_j]], %[[stride_j]] : i64156 // CHECK: %[[num_records:.*]] = llvm.intr.umax(%[[ext_i]], %[[ext_j]]) : (i64, i64) -> i64157 // CHECK: %[[elem_size_2:.*]] = llvm.mlir.constant(4 : i64) : i64158 // CHECK: %[[num_rec_bytes:.*]] = llvm.mul %[[num_records]], %[[elem_size_2]] : i64159 // CHECK: %[[stride:.*]] = llvm.mlir.constant(0 : i16) : i16160 // CHECK: %[[rsrc:.*]] = rocdl.make.buffer.rsrc %[[ptr]], %[[stride]], %[[num_rec_bytes]], %{{.*}} : !llvm.ptr to <8>161 // CHECK: %[[stride_i_1:.*]] = llvm.extractvalue %[[descriptor]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>162 // CHECK: %[[stride_i_i32:.*]] = llvm.trunc %[[stride_i_1]] : i64 to i32163 // CHECK: %[[t_0:.*]] = llvm.mul %{{.*}}, %[[stride_i_i32]] : i32164 // CHECK: %[[stride_j_1:.*]] = llvm.extractvalue %[[descriptor]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>165 // CHECK: %[[stride_j_i32:.*]] = llvm.trunc %[[stride_j_1]] : i64 to i32166 // CHECK: %[[t_1:.*]] = llvm.mul %{{.*}}, %[[stride_j_i32]] : i32167 // CHECK: %[[index:.*]] = llvm.add %[[t_0]], %[[t_1]] : i32168 // CHECK: %[[vgpr_off:.*]] = llvm.mul %[[index]], %[[elem_size]] : i32169 // CHECK: %[[zero_0:.*]] = llvm.mlir.constant(0 : i32) : i32170 // CHECK: %[[sgpr_off:.*]] = llvm.mul %[[zero_0]], %[[elem_size]] : i32171 // CHECK: %[[zero_1:.*]] = llvm.mlir.constant(0 : i32) : i32172 // CHECK: %[[v:.*]] = rocdl.raw.ptr.buffer.load %[[rsrc]], %[[vgpr_off]], %[[sgpr_off]], %[[zero_1]] : i32173 // CHECK: return %[[v]] : i32174 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%i, %j] : memref<16x16xi32, strided<[?, ?], offset: ?>>, i32, i32 -> i32175 func.return %0 : i32176}177 178// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i32_oob_off179func.func @gpu_gcn_raw_buffer_load_i32_oob_off(%buf: memref<64xi32>, %idx: i32) -> i32 {180 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)181 // RDNA: %[[flags:.*]] = llvm.mlir.constant(553807872 : i32)182 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %{{.*}}, %[[flags]]183 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32184 // CHECK: return %[[ret]]185 %0 = amdgpu.raw_buffer_load {boundsCheck = false} %buf[%idx] : memref<64xi32>, i32 -> i32186 func.return %0 : i32187}188 189// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_1xi32190func.func @gpu_gcn_raw_buffer_load_1xi32(%buf: memref<64xi32>, %idx: i32) -> vector<1xi32> {191 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : i32192 // CHECK: %[[cast:.*]] = llvm.bitcast %[[ret]] : i32 to vector<1xi32>193 // CHECK: return %[[cast]]194 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi32>, i32 -> vector<1xi32>195 func.return %0 : vector<1xi32>196}197 198// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_2xi32199func.func @gpu_gcn_raw_buffer_load_2xi32(%buf: memref<64xi32>, %idx: i32) -> vector<2xi32> {200 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : vector<2xi32>201 // CHECK: return %[[ret]]202 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi32>, i32 -> vector<2xi32>203 func.return %0 : vector<2xi32>204}205 206// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i8207func.func @gpu_gcn_raw_buffer_load_i8(%buf: memref<64xi8>, %idx: i32) -> i8 {208 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(64 : i64)209 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}210 // CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i8211 // CHECK: return %[[ret]]212 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi8>, i32 -> i8213 func.return %0 : i8214}215 216// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_2xi8217func.func @gpu_gcn_raw_buffer_load_2xi8(%buf: memref<64xi8>, %idx: i32) -> vector<2xi8> {218 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(64 : i64)219 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}220 // CHECK: %[[loaded:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i16221 // CHECK: %[[ret:.*]] = llvm.bitcast %[[loaded]] : i16 to vector<2xi8>222 // CHECK: return %[[ret]]223 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi8>, i32 -> vector<2xi8>224 func.return %0 : vector<2xi8>225}226 227// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_16xi8228func.func @gpu_gcn_raw_buffer_load_16xi8(%buf: memref<64xi8>, %idx: i32) -> vector<16xi8> {229 // CHECK: %[[loaded:.*]] = rocdl.raw.ptr.buffer.load %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : vector<4xi32>230 // CHECK: %[[ret:.*]] = llvm.bitcast %[[loaded]] : vector<4xi32> to vector<16xi8>231 // CHECK: return %[[ret]]232 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi8>, i32 -> vector<16xi8>233 func.return %0 : vector<16xi8>234}235 236// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_f8E5M2FNUZ237func.func @gpu_gcn_raw_buffer_load_f8E5M2FNUZ(%buf: memref<64xf8E5M2FNUZ>, %idx: i32) -> f8E5M2FNUZ {238 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(64 : i64)239 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}240 // CHECK: %[[loaded:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i8241 // CHECK: %[[ret:.*]] = builtin.unrealized_conversion_cast %[[loaded]] : i8 to f8E5M2FNUZ242 // CHECK: return %[[ret]]243 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xf8E5M2FNUZ>, i32 -> f8E5M2FNUZ244 func.return %0 : f8E5M2FNUZ245}246 247// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_4xf8E4M3FNUZ248func.func @gpu_gcn_raw_buffer_load_4xf8E4M3FNUZ(%buf: memref<64xf8E4M3FNUZ>, %idx: i32) -> vector<4xf8E4M3FNUZ> {249 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(64 : i64)250 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %{{.*}}251 // CHECK: %[[loaded:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32252 // CHECK: %[[cast:.*]] = llvm.bitcast %[[loaded]] : i32 to vector<4xi8>253 // CHECK: %[[ret:.*]] = builtin.unrealized_conversion_cast %[[cast]] : vector<4xi8> to vector<4xf8E4M3FNUZ>254 // CHECK: return %[[ret]]255 %0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xf8E4M3FNUZ>, i32 -> vector<4xf8E4M3FNUZ>256 func.return %0 : vector<4xf8E4M3FNUZ>257}258 259// Since the lowering logic is shared with loads, only bitcasts need to be rechecked260// CHECK-LABEL: func @gpu_gcn_raw_buffer_store_scalar_i32261func.func @gpu_gcn_raw_buffer_store_scalar_i32(%value: i32, %buf: memref<i32>) {262 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)263 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)264 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %{{.*}}, %[[flags]]265 // CHECK: rocdl.raw.ptr.buffer.store %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32266 amdgpu.raw_buffer_store {boundsCheck = true} %value -> %buf[] : i32 -> memref<i32>267 func.return268}269 270// CHECK-LABEL: func @gpu_gcn_raw_buffer_store_i32271func.func @gpu_gcn_raw_buffer_store_i32(%value: i32, %buf: memref<64xi32>, %idx: i32) {272 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)273 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)274 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)275 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]276 // CHECK: rocdl.raw.ptr.buffer.store %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32277 amdgpu.raw_buffer_store {boundsCheck = true} %value -> %buf[%idx] : i32 -> memref<64xi32>, i32278 func.return279}280 281// CHECK-LABEL: func @gpu_gcn_raw_buffer_store_1xf32282func.func @gpu_gcn_raw_buffer_store_1xf32(%value: vector<1xf32>, %buf: memref<64xf32>, %idx: i32) {283 // CHECK: %[[cast:.*]] = llvm.bitcast %{{.*}} : vector<1xf32> to f32284 // CHECK: rocdl.raw.ptr.buffer.store %[[cast]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : f32285 amdgpu.raw_buffer_store {boundsCheck = true} %value -> %buf[%idx] : vector<1xf32> -> memref<64xf32>, i32286 func.return287}288 289// CHECK-LABEL: func @gpu_gcn_raw_buffer_store_2xi8290func.func @gpu_gcn_raw_buffer_store_2xi8(%value: vector<2xi8>, %buf: memref<64xi8>, %idx: i32) {291 // CHECK: %[[cast:.*]] = llvm.bitcast %{{.*}} : vector<2xi8> to i16292 // CHECK: rocdl.raw.ptr.buffer.store %[[cast]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : i16293 amdgpu.raw_buffer_store {boundsCheck = true} %value -> %buf[%idx] : vector<2xi8> -> memref<64xi8>, i32294 func.return295}296 297// CHECK-LABEL: func @gpu_gcn_raw_buffer_store_16xi8298func.func @gpu_gcn_raw_buffer_store_16xi8(%value: vector<16xi8>, %buf: memref<64xi8>, %idx: i32) {299 // CHECK: %[[cast:.*]] = llvm.bitcast %{{.*}} : vector<16xi8> to vector<4xi32>300 // CHECK: rocdl.raw.ptr.buffer.store %[[cast]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : vector<4xi32>301 amdgpu.raw_buffer_store {boundsCheck = true} %value -> %buf[%idx] : vector<16xi8> -> memref<64xi8>, i32302 func.return303}304 305// And more so for atomic add306// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_fadd_f32307func.func @gpu_gcn_raw_buffer_atomic_fadd_f32(%value: f32, %buf: memref<64xf32>, %idx: i32) {308 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)309 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)310 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)311 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]312 // CHECK: rocdl.raw.ptr.buffer.atomic.fadd %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : f32313 amdgpu.raw_buffer_atomic_fadd {boundsCheck = true} %value -> %buf[%idx] : f32 -> memref<64xf32>, i32314 func.return315}316 317// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_fadd_v2f16318func.func @gpu_gcn_raw_buffer_atomic_fadd_v2f16(%value: vector<2xf16>, %buf: memref<64xf16>, %idx: i32) {319 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(128 : i64)320 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)321 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)322 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]323 // CHECK: rocdl.raw.ptr.buffer.atomic.fadd %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : vector<2xf16>324 amdgpu.raw_buffer_atomic_fadd {boundsCheck = true} %value -> %buf[%idx] : vector<2xf16> -> memref<64xf16>, i32325 func.return326}327 328// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_fadd_v2bf16329func.func @gpu_gcn_raw_buffer_atomic_fadd_v2bf16(%value: vector<2xbf16>, %buf: memref<64xbf16>, %idx: i32) {330 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(128 : i64)331 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)332 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)333 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]334 // CHECK: rocdl.raw.ptr.buffer.atomic.fadd %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : vector<2xbf16>335 amdgpu.raw_buffer_atomic_fadd {boundsCheck = true} %value -> %buf[%idx] : vector<2xbf16> -> memref<64xbf16>, i32336 func.return337}338 339// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_fmax_f32340func.func @gpu_gcn_raw_buffer_atomic_fmax_f32(%value: f32, %buf: memref<64xf32>, %idx: i32) {341 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)342 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)343 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)344 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]345 // CHECK: rocdl.raw.ptr.buffer.atomic.fmax %{{.*}}, %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : f32346 amdgpu.raw_buffer_atomic_fmax {boundsCheck = true} %value -> %buf[%idx] : f32 -> memref<64xf32>, i32347 func.return348}349 350// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_smax_i32351func.func @gpu_gcn_raw_buffer_atomic_smax_i32(%value: i32, %buf: memref<64xi32>, %idx: i32) {352 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)353 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)354 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)355 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]356 // CHECK: rocdl.raw.ptr.buffer.atomic.smax %{{.*}} %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32357 amdgpu.raw_buffer_atomic_smax {boundsCheck = true} %value -> %buf[%idx] : i32 -> memref<64xi32>, i32358 func.return359}360 361// CHECK-LABEL: func @gpu_gcn_raw_buffer_atomic_umin_i32362func.func @gpu_gcn_raw_buffer_atomic_umin_i32(%value: i32, %buf: memref<64xi32>, %idx: i32) {363 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)364 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)365 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)366 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]367 // CHECK: rocdl.raw.ptr.buffer.atomic.umin %{{.*}} %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32368 amdgpu.raw_buffer_atomic_umin {boundsCheck = true} %value -> %buf[%idx] : i32 -> memref<64xi32>, i32369 func.return370}371 372// CHECK-LABEL: func @amdgpu_raw_buffer_atomic_cmpswap_f32373// CHECK-SAME: (%[[src:.*]]: f32, %[[cmp:.*]]: f32, {{.*}})374func.func @amdgpu_raw_buffer_atomic_cmpswap_f32(%src : f32, %cmp : f32, %buf : memref<64xf32>, %idx: i32) -> f32 {375 // CHECK: %[[srcCast:.*]] = llvm.bitcast %[[src]] : f32 to i32376 // CHECK: %[[cmpCast:.*]] = llvm.bitcast %[[cmp]] : f32 to i32377 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(256 : i64)378 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)379 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)380 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]381 // CHECK: %[[dst:.*]] = rocdl.raw.ptr.buffer.atomic.cmpswap %[[srcCast]], %[[cmpCast]], %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32382 // CHECK: %[[dstCast:.*]] = llvm.bitcast %[[dst]] : i32 to f32383 // CHECK: return %[[dstCast]]384 %dst = amdgpu.raw_buffer_atomic_cmpswap {boundsCheck = true} %src, %cmp -> %buf[%idx] : f32 -> memref<64xf32>, i32385 func.return %dst : f32386}387 388// CHECK-LABEL: func @amdgpu_raw_buffer_atomic_cmpswap_i64389// CHECK-SAME: (%[[src:.*]]: i64, %[[cmp:.*]]: i64, {{.*}})390func.func @amdgpu_raw_buffer_atomic_cmpswap_i64(%src : i64, %cmp : i64, %buf : memref<64xi64>, %idx: i32) -> i64 {391 // CHECK: %[[numRecords:.*]] = llvm.mlir.constant(512 : i64)392 // GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)393 // RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)394 // CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %{{.*}}, %[[numRecords]], %[[flags]]395 // CHECK: %[[dst:.*]] = rocdl.raw.ptr.buffer.atomic.cmpswap %[[src]], %[[cmp]], %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i64396 // CHECK: return %[[dst]]397 %dst = amdgpu.raw_buffer_atomic_cmpswap {boundsCheck = true} %src, %cmp -> %buf[%idx] : i64 -> memref<64xi64>, i32398 func.return %dst : i64399}400 401// CHECK-LABEL: func @amdgpu_raw_buffer_atomic_cmpswap_v2f16402// CHECK-SAME: (%[[src:.*]]: vector<2xf16>, %[[cmp:.*]]: vector<2xf16>, {{.*}})403func.func @amdgpu_raw_buffer_atomic_cmpswap_v2f16(%src : vector<2xf16>, %cmp : vector<2xf16>, %buf : memref<64xf16>, %idx: i32) -> vector<2xf16> {404 // CHECK-DAG: %[[srcBits:.+]] = llvm.bitcast %[[src]] : vector<2xf16> to i32405 // CHECK-DAG: %[[cmpBits:.+]] = llvm.bitcast %[[cmp]] : vector<2xf16> to i32406 // CHECK: %[[dstBits:.+]] = rocdl.raw.ptr.buffer.atomic.cmpswap %[[srcBits]], %[[cmpBits]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : i32407 // CHECK: %[[dst:.+]] = llvm.bitcast %[[dstBits]] : i32 to vector<2xf16>408 // CHECK: return %[[dst]]409 %dst = amdgpu.raw_buffer_atomic_cmpswap {boundsCheck = true} %src, %cmp -> %buf[%idx] : vector<2xf16> -> memref<64xf16>, i32410 func.return %dst : vector<2xf16>411}412 413// CHECK-LABEL: func @lds_barrier414func.func @lds_barrier() {415 // CHECK: llvm.fence syncscope("workgroup") release {llvm.mmra = #[[$MMRA_TAG]]}416 // GFX908: llvm.inline_asm has_side_effects asm_dialect = att417 // GFX908-SAME: ";;;WARNING: BREAKS DEBUG WATCHES\0As_barrier"418 // GFX90A-NEXT: rocdl.s.barrier419 // GFX942-NEXT: rocdl.s.barrier420 // GFX10-NEXT: rocdl.s.barrier421 // GFX11-NEXT: rocdl.s.barrier422 // GFX12-NEXT: rocdl.s.barrier.signal -1423 // GFX12-NEXT: rocdl.s.barrier.wait -1424 // CHECK-NEXT: llvm.fence syncscope("workgroup") acquire {llvm.mmra = #[[$MMRA_TAG]]}425 amdgpu.lds_barrier426 func.return427}428 429// CHECK-LABEL: func @sched_barrier430func.func @sched_barrier() {431 // CHECK: rocdl.sched.barrier 0432 amdgpu.sched_barrier allow = <none>433 // CHECK: rocdl.sched.barrier 1434 amdgpu.sched_barrier allow = <non_mem_non_sideffect>435 // CHECK: rocdl.sched.barrier 2436 amdgpu.sched_barrier allow = <valu>437 // CHECK: rocdl.sched.barrier 4438 amdgpu.sched_barrier allow = <salu>439 // CHECK: rocdl.sched.barrier 8440 amdgpu.sched_barrier allow = <mfma_wmma>441 // CHECK: rocdl.sched.barrier 16442 amdgpu.sched_barrier allow = <all_vmem>443 // CHECK: rocdl.sched.barrier 32444 amdgpu.sched_barrier allow = <vmem_read>445 // CHECK: rocdl.sched.barrier 64446 amdgpu.sched_barrier allow = <vmem_write>447 // CHECK: rocdl.sched.barrier 128448 amdgpu.sched_barrier allow = <all_ds>449 // CHECK: rocdl.sched.barrier 256450 amdgpu.sched_barrier allow = <ds_read>451 // CHECK: rocdl.sched.barrier 512452 amdgpu.sched_barrier allow = <ds_write>453 // CHECK: rocdl.sched.barrier 1024454 amdgpu.sched_barrier allow = <transcendental>455 // CHECK: rocdl.sched.barrier 18456 amdgpu.sched_barrier allow = <valu|all_vmem>457 func.return458}459 460