gcc/testsuite/gcc.target/i386/avx512bw-vpcmpuw-2.c

   1 /* { dg-do run } */
   2 /* { dg-options "-O2 -mavx512bw -DAVX512BW" } */
   3 /* { dg-require-effective-target avx512bw } */
   4
   5 #include "avx512f-helper.h"
   6
   7 #include <math.h>
   8 #define SIZE (AVX512F_LEN / 16)
   9 #include "avx512f-mask-type.h"
  10
  11 #if AVX512F_LEN == 512
  12 #undef CMP
  13 #define CMP(imm, rel)                                   \
  14     dst_ref = 0;                                        \
  15     for (i = 0; i < 32; i++)                            \
  16     {                                                   \
  17       dst_ref = ((rel) << i) | dst_ref;                 \
  18     }                                                   \
  19     source1.x = _mm512_loadu_si512 (s1);                \
  20     source2.x = _mm512_loadu_si512 (s2);                \
  21     dst1 = _mm512_cmp_epu16_mask (source1.x, source2.x, imm);\
  22     dst2 = _mm512_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  23     if (dst_ref != dst1) abort();                       \
  24     if ((mask & dst_ref) != dst2) abort();
  25 #endif
  26
  27 #if AVX512F_LEN == 256
  28 #undef CMP
  29 #define CMP(imm, rel)                                   \
  30     dst_ref = 0;                                        \
  31     for (i = 0; i < 16; i++)                            \
  32     {                                                   \
  33       dst_ref = ((rel) << i) | dst_ref;                 \
  34     }                                                   \
  35     source1.x = _mm256_loadu_si256 ((__m256i*)s1);      \
  36     source2.x = _mm256_loadu_si256 ((__m256i*)s2);      \
  37     dst1 = _mm256_cmp_epu16_mask (source1.x, source2.x, imm);\
  38     dst2 = _mm256_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  39     if (dst_ref != dst1) abort();                       \
  40     if ((mask & dst_ref) != dst2) abort();
  41 #endif
  42
  43 #if AVX512F_LEN == 128
  44 #undef CMP
  45 #define CMP(imm, rel)                                   \
  46     dst_ref = 0;                                        \
  47     for (i = 0; i < 8; i++)                             \
  48     {                                                   \
  49       dst_ref = ((rel) << i) | dst_ref;                 \
  50     }                                                   \
  51     source1.x = _mm_loadu_si128 ((__m128i*)s1);         \
  52     source2.x = _mm_loadu_si128 ((__m128i*)s2);         \
  53     dst1 = _mm_cmp_epu16_mask (source1.x, source2.x, imm);\
  54     dst2 = _mm_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  55     if (dst_ref != dst1) abort();                       \
  56     if ((mask & dst_ref) != dst2) abort();
  57 #endif
  58
  59 void
  60 TEST ()
  61 {
  62     unsigned short s1[32] = {2134,  6678,  453, 54646,
  63                             231, 5674,  111, 23241,
  64                             12314,  145,  671, 7575,
  65                             23455,  166, 5321, 5673,
  66                             2134,  6678,  453, 54646,
  67                             231, 5674,  111, 23241,
  68                             12314,  145,  671, 7575,
  69                             23455,  166, 5321, 5673};
  70     unsigned short s2[32] = {41124, 6678, 8653,   856,
  71                             231, 4646,  111,   124,
  72                             2745, 4567, 3676,   123,
  73                             714, 3589, 5683,  5673,
  74                             41124, 6678, 8653,   856,
  75                             231, 4646,  111,   124,
  76                             2745, 4567, 3676,   123,
  77                             714, 3589, 5683,  5673};
  78     UNION_TYPE (AVX512F_LEN, i_w) source1, source2;
  79     MASK_TYPE dst1, dst2, dst_ref;
  80     MASK_TYPE mask = MASK_VALUE;
  81     int i;
  82
  83     CMP(0x00, s1[i] == s2[i]);
  84     CMP(0x01, s1[i] < s2[i]);
  85     CMP(0x02, s1[i] <= s2[i]);
  86     CMP(0x03, 0);
  87     CMP(0x04, s1[i] != s2[i]);
  88     CMP(0x05, s1[i] >= s2[i]);
  89     CMP(0x06, s1[i] > s2[i]);
  90     CMP(0x07, 1);
  91 }