101 lines · plain
1// RUN: mlir-opt %s -test-lower-to-llvm | \2// RUN: mlir-runner -e entry -entry-point-result=void \3// RUN: -shared-libs=%mlir_c_runner_utils | \4// RUN: FileCheck %s5 6!vector_type_A = vector<8xf32>7!vector_type_B = vector<8xf32>8!vector_type_C = vector<8x8xf32>9 10!vector_type_X = vector<2xf32>11!vector_type_Y = vector<3xf32>12!vector_type_Z = vector<2x3xf32>13 14!vector_type_R = vector<7xf32>15 16func.func @vector_outerproduct_splat_8x8(%fa: f32, %fb: f32, %fc: f32) -> !vector_type_C {17 %a = vector.broadcast %fa: f32 to !vector_type_A18 %b = vector.broadcast %fb: f32 to !vector_type_B19 %c = vector.broadcast %fc: f32 to !vector_type_C20 %d = vector.outerproduct %a, %b, %c : !vector_type_A, !vector_type_B21 return %d: !vector_type_C22}23 24func.func @vector_outerproduct_vec_2x3(%x : !vector_type_X,25 %y : !vector_type_Y) -> !vector_type_Z {26 %o = vector.outerproduct %x, %y : !vector_type_X, !vector_type_Y27 return %o: !vector_type_Z28}29 30func.func @vector_outerproduct_vec_2x3_acc(%x : !vector_type_X,31 %y : !vector_type_Y,32 %z : !vector_type_Z) -> !vector_type_Z {33 %o = vector.outerproduct %x, %y, %z : !vector_type_X, !vector_type_Y34 return %o: !vector_type_Z35}36 37func.func @entry() {38 %f0 = arith.constant 0.0: f3239 %f1 = arith.constant 1.0: f3240 %f2 = arith.constant 2.0: f3241 %f3 = arith.constant 3.0: f3242 %f4 = arith.constant 4.0: f3243 %f5 = arith.constant 5.0: f3244 %f10 = arith.constant 10.0: f3245 46 // Simple case, splat scalars into vectors, then take outer product.47 %v = call @vector_outerproduct_splat_8x8(%f1, %f2, %f10)48 : (f32, f32, f32) -> (!vector_type_C)49 vector.print %v : !vector_type_C50 //51 // outer product 8x8:52 //53 // CHECK-COUNT-8: ( 12, 12, 12, 12, 12, 12, 12, 12 )54 55 // Direct outerproduct on vectors with different size.56 %0 = vector.broadcast %f1 : f32 to !vector_type_X57 %x = vector.insert %f2, %0[1] : f32 into !vector_type_X58 %1 = vector.broadcast %f3 : f32 to !vector_type_Y59 %2 = vector.insert %f4, %1[1] : f32 into !vector_type_Y60 %y = vector.insert %f5, %2[2] : f32 into !vector_type_Y61 62 %p = call @vector_outerproduct_vec_2x3(%x, %y)63 : (!vector_type_X, !vector_type_Y) -> (!vector_type_Z)64 vector.print %p : !vector_type_Z65 //66 // outer product 2x3:67 //68 // CHECK: ( ( 3, 4, 5 ), ( 6, 8, 10 ) )69 70 %q = call @vector_outerproduct_vec_2x3_acc(%x, %y, %p)71 : (!vector_type_X, !vector_type_Y, !vector_type_Z) -> (!vector_type_Z)72 vector.print %q : !vector_type_Z73 //74 // outer product 2x3:75 //76 // CHECK: ( ( 6, 8, 10 ), ( 12, 16, 20 ) )77 78 %3 = vector.broadcast %f0 : f32 to !vector_type_R79 %4 = vector.insert %f1, %3[1] : f32 into !vector_type_R80 %5 = vector.insert %f2, %4[2] : f32 into !vector_type_R81 %6 = vector.insert %f3, %5[3] : f32 into !vector_type_R82 %7 = vector.insert %f4, %6[4] : f32 into !vector_type_R83 %8 = vector.insert %f5, %7[5] : f32 into !vector_type_R84 %9 = vector.insert %f10, %8[6] : f32 into !vector_type_R85 86 %o = vector.broadcast %f1 : f32 to !vector_type_R87 88 %axpy1 = vector.outerproduct %9, %f2 : !vector_type_R, f3289 %axpy2 = vector.outerproduct %9, %f2, %o : !vector_type_R, f3290 91 vector.print %axpy1 : !vector_type_R92 vector.print %axpy2 : !vector_type_R93 //94 // axpy operations:95 //96 // CHECK: ( 0, 2, 4, 6, 8, 10, 20 )97 // CHECK: ( 1, 3, 5, 7, 9, 11, 21 )98 99 return100}101