brintos

brintos / llvm-project-archived public Read only

0
0
Text · 2.9 KiB · 0bb2650 Raw
80 lines · plain
1// RUN: mlir-opt -pass-pipeline="builtin.module(func.func(convert-linalg-to-loops,lower-affine,convert-scf-to-cf,convert-arith-to-llvm),convert-vector-to-llvm,finalize-memref-to-llvm,convert-func-to-llvm,convert-cf-to-llvm,reconcile-unrealized-casts)" %s | mlir-runner -O3 -e main -entry-point-result=void -shared-libs=%mlir_c_runner_utils | FileCheck %s2 3func.func @main() {4  %A = memref.alloc() : memref<16x16xf32>5  %B = memref.alloc() : memref<16x16xf32>6  %C = memref.alloc() : memref<16x16xf32>7 8  %cf1 = arith.constant 1.00000e+00 : f329 10  linalg.fill ins(%cf1 : f32) outs(%A : memref<16x16xf32>)11  linalg.fill ins(%cf1 : f32) outs(%B : memref<16x16xf32>)12 13  %num_reps = arith.constant 5 : index14 15  %t_start = call @rtclock() : () -> f6416  affine.for %arg0 = 0 to %num_reps {17    linalg.fill ins(%cf1 : f32) outs(%C : memref<16x16xf32>)18    func.call @sgemm_naive(%A, %B, %C) : (memref<16x16xf32>, memref<16x16xf32>, memref<16x16xf32>) -> ()19  }20  %t_end = call @rtclock() : () -> f6421  %t = arith.subf %t_end, %t_start : f6422 23  %res = affine.load %C[0, 0]: memref<16x16xf32>24  vector.print %res: f3225 26  %c0 = arith.constant 0 : index27  %c1 = arith.constant 1 : index28  %c2 = arith.constant 2 : index29 30  %M = memref.dim %C, %c0 : memref<16x16xf32>31  %N = memref.dim %C, %c1 : memref<16x16xf32>32  %K = memref.dim %A, %c1 : memref<16x16xf32>33 34  // num_flops_per_iter = 2*M*N*K35  %f1 = arith.muli %M, %N : index36  %f2 = arith.muli %f1, %K : index37  %num_flops_per_iter = arith.muli %c2, %f2 : index38 39  // num_flops_total = num_flops_per_iter * num_reps40  %num_flops_total = arith.muli %num_flops_per_iter, %num_reps: index41 42  // Print the number of flops per second43  %num_flops_total_i = arith.index_cast %num_flops_total : index to i1644  %num_flops_total_f = arith.uitofp %num_flops_total_i : i16 to f6445  %flops_per_s = arith.divf %num_flops_total_f, %t : f6446  call @printFlops(%flops_per_s) : (f64) -> ()47 48  memref.dealloc %A : memref<16x16xf32>49  memref.dealloc %B : memref<16x16xf32>50  memref.dealloc %C : memref<16x16xf32>51  return52}53// CHECK: 1754 55func.func @sgemm_naive(%arg0: memref<16x16xf32>, %arg1: memref<16x16xf32>, %arg2: memref<16x16xf32>) {56  %c0 = arith.constant 0 : index57  affine.for %arg3 = 0 to 16 {58    affine.for %arg4 = 0 to 16 {59      %m = memref.alloc() : memref<1xf32>60      %v = affine.load %arg2[%arg3, %arg4] : memref<16x16xf32>61      affine.store %v, %m[%c0] : memref<1xf32>62      affine.for %arg5 = 0 to 16 {63        %3 = affine.load %arg0[%arg3, %arg5] : memref<16x16xf32>64        %4 = affine.load %arg1[%arg5, %arg4] : memref<16x16xf32>65        %5 = affine.load %m[0] : memref<1xf32>66        %6 = arith.mulf %3, %4 : f3267        %7 = arith.addf %6, %5 : f3268        affine.store %7, %m[0] : memref<1xf32>69      }70      %s = affine.load %m[%c0] : memref<1xf32>71      affine.store %s, %arg2[%arg3, %arg4] : memref<16x16xf32>72      memref.dealloc %m : memref<1xf32>73    }74  }75  return76}77 78func.func private @printFlops(f64)79func.func private @rtclock() -> f6480