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 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 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#DCSR = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : compressed, d1 : compressed) }> 35 36// An example of a quantized sparse matmul. With the zero offset for the 37// sparse input, the sparsifier generates very efficient code for the 38// x(i,j) += (ext(a(i,k)) - 2) * ext(b(k,j)) 39// operation. 40module { 41 42 func.func @quantized_matmul(%input1: tensor<5x3xi8>, 43 %input2: tensor<3x6xi8, #DCSR>, 44 %output: tensor<5x6xi32>) -> tensor<5x6xi32> { 45 %c0 = arith.constant 0 : i32 46 %c2 = arith.constant 2 : i32 47 %0 = linalg.quantized_matmul 48 ins(%input1, %input2, %c2, %c0 : tensor<5x3xi8>, tensor<3x6xi8, #DCSR>, i32, i32) 49 outs(%output : tensor<5x6xi32>) -> tensor<5x6xi32> 50 return %0: tensor<5x6xi32> 51 } 52 53 func.func @main() { 54 %c0 = arith.constant 0 : index 55 %i0 = arith.constant 0 : i32 56 57 %input1 = arith.constant dense<[ 58 [ -128, 3, 127 ], 59 [ 0, 0, 0 ], 60 [ 11, 1, 0 ], 61 [ 0, 5, -1 ], 62 [ 13, 0, 3 ] 63 ]> : tensor<5x3xi8> 64 65 %input2 = arith.constant dense<[ 66 [ 127, 0, -128, 0, 0, 3 ], 67 [ 0, 0, 0, 0, 0, 0 ], 68 [ 0, 0, 0, 100, 10, 0 ] 69 ]> : tensor<3x6xi8> 70 71 %sparse_input2 = sparse_tensor.convert %input2 : tensor<3x6xi8> to tensor<3x6xi8, #DCSR> 72 73 // Call the kernel. 74 %output = arith.constant dense<0> : tensor<5x6xi32> 75 %0 = call @quantized_matmul(%input1, %sparse_input2, %output) 76 : (tensor<5x3xi8>, 77 tensor<3x6xi8, #DCSR>, 78 tensor<5x6xi32>) -> tensor<5x6xi32> 79 80 // 81 // Verify the output. 82 // 83 // CHECK: ( ( -16510, 0, 16640, 12500, 1250, -390 ), 84 // CHECK-SAME: ( -254, 0, 256, -200, -20, -6 ), 85 // CHECK-SAME: ( 1143, 0, -1152, -200, -20, 27 ), 86 // CHECK-SAME: ( -254, 0, 256, -300, -30, -6 ), 87 // CHECK-SAME: ( 1397, 0, -1408, 100, 10, 33 ) ) 88 // 89 %v = vector.transfer_read %0[%c0, %c0], %i0 90 : tensor<5x6xi32>, vector<5x6xi32> 91 vector.print %v : vector<5x6xi32> 92 93 // Release the resources. 94 bufferization.dealloc_tensor %sparse_input2 : tensor<3x6xi8, #DCSR> 95 bufferization.dealloc_tensor %0 : tensor<5x6xi32> 96 97 return 98 } 99} 100