157 lines · plain
1// RUN: mlir-opt %s --gpu-to-llvm | FileCheck %s2 3module attributes {gpu.container_module} {4 5 // CHECK-LABEL: func @matvec6 // CHECK: llvm.call @mgpuStreamCreate7 // CHECK: llvm.call @mgpuMemAlloc8 // CHECK: llvm.call @mgpuMemAlloc9 // CHECK: llvm.call @mgpuCreateCoo10 // CHECK: llvm.call @mgpuCreateDnVec11 // CHECK: llvm.call @mgpuSpMVBufferSize12 // CHECK: llvm.call @mgpuSpMV13 // CHECK: llvm.call @mgpuDestroySpMat14 // CHECK: llvm.call @mgpuDestroyDnVec15 // CHECK: llvm.call @mgpuStreamSynchronize16 // CHECK: llvm.call @mgpuStreamDestroy17 func.func @matvec(%arg0: index) {18 %token0 = gpu.wait async19 %mem1, %token1 = gpu.alloc async [%token0] (%arg0) : memref<?xindex>20 %mem2, %token2 = gpu.alloc async [%token1] (%arg0) : memref<?xf64>21 %spmat, %token4 = gpu.create_coo async [%token2] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf64>22 %dnvec, %token5 = gpu.create_dn_tensor async [%token4] %mem2, %arg0 : index into memref<?xf64>23 %bufferSz, %token6 = gpu.spmv_buffer_size async [%token5] %spmat, %dnvec, %dnvec into f6424 %token7 = gpu.spmv async [%token6] %spmat, %dnvec, %dnvec, %mem2 : memref<?xf64> into f6425 %token8 = gpu.destroy_sp_mat async [%token7] %spmat26 %token9 = gpu.destroy_dn_tensor async [%token8] %dnvec27 gpu.wait [%token9]28 return29 }30 31 // CHECK-LABEL: func @matmul32 // CHECK: llvm.call @mgpuStreamCreate33 // CHECK: llvm.call @mgpuMemAlloc34 // CHECK: llvm.call @mgpuMemAlloc35 // CHECK: llvm.call @mgpuCreateCsr36 // CHECK: llvm.call @mgpuCreateDnMat37 // CHECK: llvm.call @mgpuSpMMBufferSize38 // CHECK: llvm.call @mgpuSpMM39 // CHECK: llvm.call @mgpuDestroySpMat40 // CHECK: llvm.call @mgpuDestroyDnMat41 // CHECK: llvm.call @mgpuStreamSynchronize42 // CHECK: llvm.call @mgpuStreamDestroy43 func.func @matmul(%arg0: index) {44 %token0 = gpu.wait async45 %mem1, %token1 = gpu.alloc async [%token0] (%arg0) : memref<?xindex>46 %mem2, %token2 = gpu.alloc async [%token1] (%arg0) : memref<?xf64>47 %spmat, %token4 = gpu.create_csr async [%token2] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf64>48 %dnmat, %token5 = gpu.create_dn_tensor async [%token4] %mem2, %arg0, %arg0 : index, index into memref<?xf64>49 %bufferSz, %token6 = gpu.spmm_buffer_size async [%token5] %spmat, %dnmat, %dnmat : index into f6450 %token7 = gpu.spmm async [%token6] %spmat, %dnmat, %dnmat, %mem2 : memref<?xf64> into f6451 %token8 = gpu.destroy_sp_mat async [%token7] %spmat52 %token9 = gpu.destroy_dn_tensor async [%token8] %dnmat53 gpu.wait [%token9]54 return55 }56 57 // CHECK-LABEL: func @spgemm58 // CHECK: llvm.call @mgpuStreamCreate59 // CHECK: llvm.call @mgpuMemAlloc60 // CHECK: llvm.call @mgpuMemAlloc61 // CHECK: llvm.call @mgpuCreateCsr62 // CHECK: llvm.call @mgpuCreateCsr63 // CHECK: llvm.call @mgpuCreateCsr64 // CHECK: llvm.call @mgpuSpGEMMCreateDescr65 // CHECK: llvm.call @mgpuSpGEMMWorkEstimation66 // CHECK: llvm.call @mgpuSpGEMMCompute67 // CHECK: llvm.call @mgpuSpMatGetSize68 // CHECK: llvm.call @mgpuSetCsrPointers69 // CHECK: llvm.call @mgpuSpGEMMCopy70 // CHECK: llvm.call @mgpuSpGEMMDestroyDescr71 // CHECK: llvm.call @mgpuDestroySpMat72 // CHECK: llvm.call @mgpuDestroySpMat73 // CHECK: llvm.call @mgpuDestroySpMat74 // CHECK: llvm.call @mgpuStreamSynchronize75 // CHECK: llvm.call @mgpuStreamDestroy76 func.func @spgemm(%arg0: index) {77 %token0 = gpu.wait async78 %mem1, %token1 = gpu.alloc async [%token0] (%arg0) : memref<?xindex>79 %mem2, %token2 = gpu.alloc async [%token1] (%arg0) : memref<?xf32>80 %spmatA, %token3 = gpu.create_csr async [%token2] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf32>81 %spmatB, %token4 = gpu.create_csr async [%token3] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf32>82 %spmatC, %token5 = gpu.create_csr async [%token4] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf32>83 %spgemmDesc, %token6 = gpu.spgemm_create_descr async [%token5]84 %alloc = memref.alloc() : memref<0xi8> // nullptr85 %c0 = arith.constant 0 : index86 %bufferSz1, %token7 = gpu.spgemm_work_estimation_or_compute async87 [%token6]{WORK_ESTIMATION}88 %spmatA, %spmatB, %spmatC,89 %spgemmDesc, %c0, %alloc: f32 into memref<0xi8>90 %bufferSz2, %token8 = gpu.spgemm_work_estimation_or_compute async91 [%token7]{COMPUTE}92 %spmatA, %spmatB, %spmatC,93 %spgemmDesc, %c0, %alloc: f32 into memref<0xi8>94 %rows, %cols, %nnz, %token9 = gpu.spmat_get_size async [%token8] %spmatC95 %token10 = gpu.set_csr_pointers async [%token8] %spmatC, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf32>96 %token11 = gpu.spgemm_copy async [%token10] %spmatA, %spmatB, %spmatC, %spgemmDesc: f3297 %token12 = gpu.spgemm_destroy_descr async [%token11] %spgemmDesc98 %token13 = gpu.destroy_sp_mat async [%token12] %spmatA99 %token14 = gpu.destroy_sp_mat async [%token13] %spmatB100 %token15 = gpu.destroy_sp_mat async [%token14] %spmatC101 gpu.wait [%token15]102 return103 }104 105 // CHECK-LABEL: func @sddmm106 // CHECK: llvm.call @mgpuStreamCreate107 // CHECK: llvm.call @mgpuMemAlloc108 // CHECK: llvm.call @mgpuMemAlloc109 // CHECK: llvm.call @mgpuCreateCsr110 // CHECK: llvm.call @mgpuCreateDnMat111 // CHECK: llvm.call @mgpuSDDMMBufferSize112 // CHECK: llvm.call @mgpuSDDMM113 // CHECK: llvm.call @mgpuDestroySpMat114 // CHECK: llvm.call @mgpuDestroyDnMat115 // CHECK: llvm.call @mgpuStreamSynchronize116 // CHECK: llvm.call @mgpuStreamDestroy117 func.func @sddmm(%arg0: index) {118 %token0 = gpu.wait async119 %mem1, %token1 = gpu.alloc async [%token0] (%arg0) : memref<?xindex>120 %mem2, %token2 = gpu.alloc async [%token1] (%arg0) : memref<?xf64>121 %spmat, %token4 = gpu.create_csr async [%token2] %arg0, %arg0, %arg0, %mem1, %mem1, %mem2 : memref<?xindex>, memref<?xindex>, memref<?xf64>122 %dnmat, %token5 = gpu.create_dn_tensor async [%token4] %mem2, %arg0, %arg0 : index, index into memref<?xf64>123 %bufferSz, %token6 = gpu.sddmm_buffer_size async [%token5] %dnmat, %dnmat, %spmat into f64124 %token7 = gpu.sddmm async [%token6] %dnmat, %dnmat, %spmat, %mem2 : memref<?xf64> into f64125 %token8 = gpu.destroy_sp_mat async [%token7] %spmat126 %token9 = gpu.destroy_dn_tensor async [%token8] %dnmat127 gpu.wait [%token9]128 return129 }130 131 // CHECK-LABEL: func @csc_and_bsr132 // CHECK: llvm.call @mgpuStreamCreate133 // CHECK: llvm.call @mgpuMemAlloc134 // CHECK: llvm.call @mgpuMemAlloc135 // CHECK: llvm.call @mgpuCreateCsc136 // CHECK: llvm.call @mgpuCreateBsr137 // CHECK: llvm.call @mgpuDestroySpMat138 // CHECK: llvm.call @mgpuDestroySpMat139 // CHECK: llvm.call @mgpuStreamSynchronize140 // CHECK: llvm.call @mgpuStreamDestroy141 func.func @csc_and_bsr(%arg0: index) {142 %token0 = gpu.wait async143 %mem1, %token1 = gpu.alloc async [%token0] (%arg0) : memref<?xindex>144 %mem2, %token2 = gpu.alloc async [%token1] (%arg0) : memref<?xf64>145 %csc, %token3 = gpu.create_csc async [%token2]146 %arg0, %arg0, %arg0, %mem1, %mem1, %mem2147 : memref<?xindex>, memref<?xindex>, memref<?xf64>148 %bsr, %token4 = gpu.create_bsr async [%token3]149 %arg0, %arg0, %arg0, %arg0, %arg0, %mem1, %mem1, %mem2150 : memref<?xindex>, memref<?xindex>, memref<?xf64>151 %token5 = gpu.destroy_sp_mat async [%token4] %csc152 %token6 = gpu.destroy_sp_mat async [%token5] %bsr153 gpu.wait [%token6]154 return155 }156}157