Converting 16-bit shorts to floating point samples with SSE

sse-bufconvert.c

#include <stdlib.h>
#include <malloc.h>
#include <xmmintrin.h>

#if 1
#define ITERATIONS 1
#define BUFSIZE 32
#else
#define ITERATIONS 1000000
#define BUFSIZE 512
#endif

int main() {
  // Visual studio: __declspec(align(16))
  short shortSamples[BUFSIZE] __attribute((aligned(16)));
  float floatSamples[BUFSIZE + 16] __attribute((aligned(16)));
  //__attribute((aligned(16))) short* shortSamples;
  //__attribute((aligned(16))) float *floatSamples;

  // Visual studio: _aligned_malloc
  //posix_memalign((void**)&shortSamples, 16, sizeof(short) * BUFSIZE);
  //posix_memalign((void**)&floatSamples, 16, sizeof(float) * BUFSIZE);

  // Short -> float
#if 0
  for(int i = 0; i < BUFSIZE; ++i) {
    shortSamples[i] = i;
  }

#if 0
  for(int i = 0; i < BUFSIZE; ++i) {
    shortSamples[i] = i;
    floatSamples[i] = 0.0f;
  }
  for(int j = 0; j < ITERATIONS; j++) {
    for(int i = 0; i < BUFSIZE; i++) {
      floatSamples[i] = (float)shortSamples[i] / 32767.0f;
    }
  }
#else
  for(int i = 0; i < BUFSIZE; ++i) {
    shortSamples[i] = i;
    floatSamples[i] = 0.0f;
  }

  __m128 factor =  { 1.0f/32767, 1.0f/32767, 1.0f/32767, 1.0f/32767 };
  __m64 zero1 =  { 0,0 };
  __m128i zero2 =  { 0,0 };
  __m128i intValue;
  __m128 floatValue;
  for(int j = 0; j < ITERATIONS; j++) {
    __m64 *ps = (__m64*)shortSamples;
    __m128 *pd = (__m128*)floatSamples;
    for (int i = 0; i < BUFSIZE; i+=4) {
      intValue = _mm_unpacklo_epi16(_mm_set_epi64(zero1, *ps), zero2);
      intValue = _mm_srai_epi32(_mm_slli_epi32(intValue, 16), 16);
      floatValue = _mm_cvtepi32_ps(intValue);
      *pd = _mm_mul_ps(floatValue, factor);
      pd++;
      ps++;
    }
  }
#endif
#endif

  // Float -> short
#if 1
  for(int i = 0; i < BUFSIZE; i++) {
    shortSamples[i] = 1.0f / (float)i;
  }

#if 0
  for(int j = 0; j < ITERATIONS; j++) {
    for(int i = 0; i < BUFSIZE; i++) {
      shortSamples[i] = (short)(floatSamples[i] * 32767.0f);
      printf("%f, ", floatSamples[i]);
    }
  }
#else
  __m128 factor =  { 32767.0f, 32767.0f, 32767.0f, 32767.0f };
  __m64 zero1 =  { 0, 0 };
  __m128i zero2 =  { 0, 0 };
  __m128i intValue;
  __m128 floatValue;
  for(int j = 0; j < ITERATIONS; j++) {
    __m64 *ps = (__m64*)shortSamples;
    __m128 *pd = (__m128*)floatSamples;
    for (int i = 0; i < BUFSIZE; i+=4) {
      floatValue = _mm_mul_ps(*floatSamples, value);
      intValue = _mm_unpacklo_epi16(_mm_set_epi64(zero1, *ps), zero2);
      intValue = _mm_srai_epi32(_mm_slli_epi32(intValue, 16), 16);
      floatValue = _mm_cvtepi32_ps(intValue);
    }
  }
#endif
#endif

  return 0;
}