neon_fast_exp benchmark

bench.cpp

#include <arm_neon.h>
#include <benchmark/benchmark.h>

inline float16x8_t neon_fast_exp(float16x8_t x) {
  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f)));     // clamp under with -14
  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f)));      // clamp over with 14

  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
  float16x8_t fpart = vsubq_f16(x, ipart);

  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);

  // generate 2**ipart in the floating point representation using integer
  // bitshifting
  int16x8_t epart = vcvtq_s16_f16(ipart);
  epart = vaddq_s16(epart, vdupq_n_s16(15));
  epart = vshlq_n_s16(epart, 10);

  return vmulq_f16(vreinterpretq_f16_s16(epart), x);
}

inline float16x8_t neon_fast_exp_fixed(float16x8_t x) {
  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f)));     // clamp under with -14
  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f)));      // clamp over with 14

  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
  float16x8_t fpart = vsubq_f16(x, ipart);

  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);

  // generate 2**ipart in the floating point representation using integer
  // bitshifting
  int16x8_t epart = vcvtq_s16_f16(ipart);
  epart = vaddq_s16(epart, vdupq_n_s16(15));
  epart = vshlq_n_s16(epart, 10);

  return vmulq_f16(vreinterpretq_f16_s16(epart), x);
}

// Benchmark function
template <float16x8_t(Func)(float16x8_t)>
static void BM_NeonFastExp(benchmark::State &state) {
  float16x8_t input =
      vdupq_n_f16(1.0f); // Example input vector with all elements as 1.0f
  float16x8_t result;

  for (auto _ : state) {
    result = Func(input);
    benchmark::DoNotOptimize(
        result); // Prevent the compiler from optimizing away the call
  }
}

// Register the function as a benchmark
BENCHMARK(BM_NeonFastExp<neon_fast_exp>);
BENCHMARK(BM_NeonFastExp<neon_fast_exp_fixed>);

// Main function to run the benchmarks
BENCHMARK_MAIN();