gcc/testsuite/gcc.target/i386/ssse3-pmaddubsw.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target ssse3 } */
   3 /* { dg-options "-O2 -mssse3" } */
   4
   5 #include "ssse3-check.h"
   6 #include "ssse3-vals.h"
   7
   8 #include <tmmintrin.h>
   9
  10 /* Test the 64-bit form */
  11 static void
  12 ssse3_test_pmaddubsw (int *i1, int *i2, int *r)
  13 {
  14   __m64 t1 = *(__m64 *) i1;
  15   __m64 t2 = *(__m64 *) i2;
  16   *(__m64 *) r = _mm_maddubs_pi16 (t1, t2);
  17   _mm_empty ();
  18 }
  19
  20 /* Test the 128-bit form */
  21 static void
  22 ssse3_test_pmaddubsw128 (int *i1, int *i2, int *r)
  23 {
  24   /* Assumes incoming pointers are 16-byte aligned */
  25   __m128i t1 = *(__m128i *) i1;
  26   __m128i t2 = *(__m128i *) i2;
  27   *(__m128i *) r = _mm_maddubs_epi16 (t1, t2);
  28 }
  29
  30 static short
  31 signed_saturate_to_word(int x)
  32 {
  33   if (x > (int) 0x7fff)
  34     return 0x7fff;
  35
  36   if (x < (int) 0xffff8000)
  37     return 0x8000;
  38
  39   return (short) x;
  40 }
  41
  42 /* Routine to manually compute the results */
  43 static void
  44 compute_correct_result (int *i1, int *i2, int *r)
  45 {
  46   unsigned char *ub1 = (unsigned char *) i1;
  47   char *sb2 = (char *) i2;
  48   short *sout = (short *) r;
  49   int t0;
  50   int i;
  51
  52   for (i = 0; i < 8; i++)
  53     {
  54       t0 = ((int) ub1[2 * i] * (int) sb2[2 * i] +
  55             (int) ub1[2 * i + 1] * (int) sb2[2 * i + 1]);
  56       sout[i] = signed_saturate_to_word (t0);
  57     }
  58 }
  59
  60 static void
  61 ssse3_test (void)
  62 {
  63   int i;
  64   int r [4] __attribute__ ((aligned(16)));
  65   int ck [4];
  66   int fail = 0;
  67
  68   for (i = 0; i < 256; i += 8)
  69     {
  70       /* Manually compute the result */
  71       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  72
  73       /* Run the 64-bit tests */
  74       ssse3_test_pmaddubsw (&vals[i + 0], &vals[i + 4], &r[0]);
  75       ssse3_test_pmaddubsw (&vals[i + 2], &vals[i + 6], &r[2]);
  76       fail += chk_128 (ck, r);
  77
  78       /* Run the 128-bit tests */
  79       ssse3_test_pmaddubsw128 (&vals[i + 0], &vals[i + 4], r);
  80       fail += chk_128 (ck, r);
  81     }
  82
  83   if (fail != 0)
  84     abort ();
  85 }