1// RUN: mlir-opt %s \ 2// RUN: --sparsification --sparse-tensor-conversion \ 3// RUN: --convert-vector-to-scf --convert-scf-to-std \ 4// RUN: --func-bufferize --tensor-constant-bufferize --tensor-bufferize \ 5// RUN: --std-bufferize --finalizing-bufferize --lower-affine \ 6// RUN: --convert-vector-to-llvm --convert-memref-to-llvm --convert-std-to-llvm --reconcile-unrealized-casts | \ 7// RUN: TENSOR0="%mlir_integration_test_dir/data/test.mtx" \ 8// RUN: mlir-cpu-runner \ 9// RUN: -e entry -entry-point-result=void \ 10// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \ 11// RUN: FileCheck %s 12// 13// Do the same run, but now with SIMDization as well. This should not change the outcome. 14// 15// RUN: mlir-opt %s \ 16// RUN: --sparsification="vectorization-strategy=2 vl=4" --sparse-tensor-conversion \ 17// RUN: --convert-vector-to-scf --convert-scf-to-std \ 18// RUN: --func-bufferize --tensor-constant-bufferize --tensor-bufferize \ 19// RUN: --std-bufferize --finalizing-bufferize --lower-affine \ 20// RUN: --convert-vector-to-llvm --convert-memref-to-llvm --convert-std-to-llvm --reconcile-unrealized-casts | \ 21// RUN: TENSOR0="%mlir_integration_test_dir/data/test.mtx" \ 22// RUN: mlir-cpu-runner \ 23// RUN: -e entry -entry-point-result=void \ 24// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \ 25// RUN: FileCheck %s 26 27!Filename = type !llvm.ptr<i8> 28 29#DCSR = #sparse_tensor.encoding<{ 30 dimLevelType = [ "compressed", "compressed" ], 31 dimOrdering = affine_map<(i,j) -> (i,j)> 32}> 33 34#eltwise_mult = { 35 indexing_maps = [ 36 affine_map<(i,j) -> (i,j)> // X (out) 37 ], 38 iterator_types = ["parallel", "parallel"], 39 doc = "X(i,j) *= X(i,j)" 40} 41 42// 43// Integration test that lowers a kernel annotated as sparse to 44// actual sparse code, initializes a matching sparse storage scheme 45// from file, and runs the resulting code with the JIT compiler. 46// 47module { 48 // 49 // A kernel that multiplies a sparse matrix A with itself 50 // in an element-wise fashion. In this operation, we have 51 // a sparse tensor as output, but although the values of the 52 // sparse tensor change, its nonzero structure remains the same. 53 // 54 func @kernel_eltwise_mult(%argx: tensor<?x?xf64, #DCSR> {linalg.inplaceable = true}) 55 -> tensor<?x?xf64, #DCSR> { 56 %0 = linalg.generic #eltwise_mult 57 outs(%argx: tensor<?x?xf64, #DCSR>) { 58 ^bb(%x: f64): 59 %0 = arith.mulf %x, %x : f64 60 linalg.yield %0 : f64 61 } -> tensor<?x?xf64, #DCSR> 62 return %0 : tensor<?x?xf64, #DCSR> 63 } 64 65 func private @getTensorFilename(index) -> (!Filename) 66 67 // 68 // Main driver that reads matrix from file and calls the sparse kernel. 69 // 70 func @entry() { 71 %d0 = arith.constant 0.0 : f64 72 %c0 = arith.constant 0 : index 73 74 // Read the sparse matrix from file, construct sparse storage. 75 %fileName = call @getTensorFilename(%c0) : (index) -> (!Filename) 76 %x = sparse_tensor.new %fileName : !Filename to tensor<?x?xf64, #DCSR> 77 78 // Call kernel. 79 %0 = call @kernel_eltwise_mult(%x) : (tensor<?x?xf64, #DCSR>) -> tensor<?x?xf64, #DCSR> 80 81 // Print the result for verification. 82 // 83 // CHECK: ( 1, 1.96, 4, 6.25, 9, 16.81, 16, 27.04, 25 ) 84 // 85 %m = sparse_tensor.values %0 : tensor<?x?xf64, #DCSR> to memref<?xf64> 86 %v = vector.transfer_read %m[%c0], %d0: memref<?xf64>, vector<9xf64> 87 vector.print %v : vector<9xf64> 88 89 // Release the resources. 90 sparse_tensor.release %x : tensor<?x?xf64, #DCSR> 91 92 return 93 } 94} 95