1//-------------------------------------------------------------------------------------------------- 2// WHEN CREATING A NEW TEST, PLEASE JUST COPY & PASTE WITHOUT EDITS. 3// 4// Set-up that's shared across all tests in this directory. In principle, this 5// config could be moved to lit.local.cfg. However, there are downstream users that 6// do not use these LIT config files. Hence why this is kept inline. 7// 8// DEFINE: %{sparsifier_opts} = enable-runtime-library=true 9// DEFINE: %{sparsifier_opts_sve} = enable-arm-sve=true %{sparsifier_opts} 10// DEFINE: %{compile} = mlir-opt %s --sparsifier="%{sparsifier_opts}" 11// DEFINE: %{compile_sve} = mlir-opt %s --sparsifier="%{sparsifier_opts_sve}" 12// DEFINE: %{run_libs} = -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils 13// DEFINE: %{run_libs_sve} = -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils 14// DEFINE: %{run_opts} = -e main -entry-point-result=void 15// DEFINE: %{run} = mlir-runner %{run_opts} %{run_libs} 16// DEFINE: %{run_sve} = %mcr_aarch64_cmd --march=aarch64 --mattr="+sve" %{run_opts} %{run_libs_sve} 17// 18// DEFINE: %{env} = 19//-------------------------------------------------------------------------------------------------- 20 21// RUN: %{compile} | %{run} | FileCheck %s 22// 23// Do the same run, but now with direct IR generation. 24// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true 25// RUN: %{compile} | %{run} | FileCheck %s 26// 27// Do the same run, but now with direct IR generation and vectorization. 28// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true vl=2 reassociate-fp-reductions=true enable-index-optimizations=true 29// RUN: %{compile} | %{run} | FileCheck %s 30// 31// Do the same run, but now with direct IR generation and VLA vectorization. 32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %} 33 34// Test that test-bufferization-analysis-only works. This option is useful 35// for understanding why buffer copies were inserted. 36// RUN: mlir-opt %s --sparsifier="test-bufferization-analysis-only" -o /dev/null 37 38#Sparse1 = #sparse_tensor.encoding<{ 39 map = (i, j, k) -> ( 40 j : compressed, 41 k : compressed, 42 i : dense 43 ) 44}> 45 46#Sparse2 = #sparse_tensor.encoding<{ 47 map = (i, j, k) -> ( 48 i floordiv 2 : compressed, 49 j floordiv 2 : compressed, 50 k floordiv 2 : compressed, 51 i mod 2 : dense, 52 j mod 2 : dense, 53 k mod 2 : dense) 54}> 55 56module { 57 58 // 59 // Main driver that tests sparse tensor storage. 60 // 61 func.func @main() { 62 %c0 = arith.constant 0 : index 63 %i0 = arith.constant 0 : i32 64 65 // Setup input dense tensor and convert to two sparse tensors. 66 %d = arith.constant dense <[ 67 [ // i=0 68 [ 1, 0, 0, 0 ], 69 [ 0, 0, 0, 0 ], 70 [ 0, 0, 0, 0 ], 71 [ 0, 0, 5, 0 ] ], 72 [ // i=1 73 [ 2, 0, 0, 0 ], 74 [ 0, 0, 0, 0 ], 75 [ 0, 0, 0, 0 ], 76 [ 0, 0, 6, 0 ] ], 77 [ //i=2 78 [ 3, 0, 0, 0 ], 79 [ 0, 0, 0, 0 ], 80 [ 0, 0, 0, 0 ], 81 [ 0, 0, 7, 0 ] ], 82 //i=3 83 [ [ 4, 0, 0, 0 ], 84 [ 0, 0, 0, 0 ], 85 [ 0, 0, 0, 0 ], 86 [ 0, 0, 8, 0 ] ] 87 ]> : tensor<4x4x4xi32> 88 89 %a = sparse_tensor.convert %d : tensor<4x4x4xi32> to tensor<4x4x4xi32, #Sparse1> 90 %b = sparse_tensor.convert %d : tensor<4x4x4xi32> to tensor<4x4x4xi32, #Sparse2> 91 92 // 93 // If we store the two "fibers" [1,2,3,4] starting at index (0,0,0) and 94 // ending at index (3,0,0) and [5,6,7,8] starting at index (0,3,2) and 95 // ending at index (3,3,2)) with a “DCSR-flavored” along (j,k) with 96 // dense “fibers” in the i-dim, we end up with 8 stored entries. 97 // 98 // CHECK: ---- Sparse Tensor ---- 99 // CHECK-NEXT: nse = 8 100 // CHECK-NEXT: dim = ( 4, 4, 4 ) 101 // CHECK-NEXT: lvl = ( 4, 4, 4 ) 102 // CHECK-NEXT: pos[0] : ( 0, 2 ) 103 // CHECK-NEXT: crd[0] : ( 0, 3 ) 104 // CHECK-NEXT: pos[1] : ( 0, 1, 2 ) 105 // CHECK-NEXT: crd[1] : ( 0, 2 ) 106 // CHECK-NEXT: values : ( 1, 2, 3, 4, 5, 6, 7, 8 ) 107 // CHECK-NEXT: ---- 108 // 109 sparse_tensor.print %a : tensor<4x4x4xi32, #Sparse1> 110 111 // 112 // If we store full 2x2x2 3-D blocks in the original index order 113 // in a compressed fashion, we end up with 4 blocks to incorporate 114 // all the nonzeros, and thus 32 stored entries. 115 // 116 // CHECK: ---- Sparse Tensor ---- 117 // CHECK-NEXT: nse = 32 118 // CHECK-NEXT: dim = ( 4, 4, 4 ) 119 // CHECK-NEXT: lvl = ( 2, 2, 2, 2, 2, 2 ) 120 // CHECK-NEXT: pos[0] : ( 0, 2 ) 121 // CHECK-NEXT: crd[0] : ( 0, 1 ) 122 // CHECK-NEXT: pos[1] : ( 0, 2, 4 ) 123 // CHECK-NEXT: crd[1] : ( 0, 1, 0, 1 ) 124 // CHECK-NEXT: pos[2] : ( 0, 1, 2, 3, 4 ) 125 // CHECK-NEXT: crd[2] : ( 0, 1, 0, 1 ) 126 // CHECK-NEXT: values : ( 1, 0, 0, 0, 2, 0, 0, 0, 0, 0, 5, 0, 0, 0, 6, 0, 3, 0, 0, 0, 4, 0, 0, 0, 0, 0, 7, 0, 0, 0, 8, 0 ) 127 // CHECK-NEXT: ---- 128 // 129 sparse_tensor.print %b : tensor<4x4x4xi32, #Sparse2> 130 131 // Release the resources. 132 bufferization.dealloc_tensor %a : tensor<4x4x4xi32, #Sparse1> 133 bufferization.dealloc_tensor %b : tensor<4x4x4xi32, #Sparse2> 134 135 return 136 } 137} 138