mlir/test/mlir-cpu-runner/X86Vector/math-polynomial-approx-avx2.mlir

   1 // RUN:   mlir-opt %s -test-math-polynomial-approximation="enable-avx2"        \
   2 // RUN:               -convert-arith-to-llvm                                   \
   3 // RUN:               -convert-vector-to-llvm="enable-x86vector"               \
   4 // RUN:               -convert-math-to-llvm                                    \
   5 // RUN:               -convert-std-to-llvm                                     \
   6 // RUN:               -reconcile-unrealized-casts                              \
   7 // RUN: | mlir-cpu-runner                                                      \
   8 // RUN:     -e main -entry-point-result=void -O0                               \
   9 // RUN:     -shared-libs=%linalg_test_lib_dir/libmlir_c_runner_utils%shlibext  \
  10 // RUN:     -shared-libs=%linalg_test_lib_dir/libmlir_runner_utils%shlibext    \
  11 // RUN: | FileCheck %s
  12
  13 // -------------------------------------------------------------------------- //
  14 // rsqrt.
  15 // -------------------------------------------------------------------------- //
  16
  17 func @rsqrt() {
  18   // Sanity-check that the scalar rsqrt still works OK.
  19   // CHECK: inf
  20   %0 = arith.constant 0.0 : f32
  21   %rsqrt_0 = math.rsqrt %0 : f32
  22   vector.print %rsqrt_0 : f32
  23   // CHECK: 0.707107
  24   %two = arith.constant 2.0: f32
  25   %rsqrt_two = math.rsqrt %two : f32
  26   vector.print %rsqrt_two : f32
  27
  28   // Check that the vectorized approximation is reasonably accurate.
  29   // CHECK: 0.707107, 0.707107, 0.707107, 0.707107, 0.707107, 0.707107, 0.707107, 0.707107
  30   %vec8 = arith.constant dense<2.0> : vector<8xf32>
  31   %rsqrt_vec8 = math.rsqrt %vec8 : vector<8xf32>
  32   vector.print %rsqrt_vec8 : vector<8xf32>
  33
  34   return
  35 }
  36
  37 func @main() {
  38   call @rsqrt(): () -> ()
  39   return
  40 }