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

   1 /* { dg-do run } */
   2 /* { dg-options "-O2 -mavx512bw" } */
   3 /* { dg-require-effective-target avx512bw } */
   4
   5 #define AVX512BW
   6 #include "avx512f-helper.h"
   7
   8 #include <math.h>
   9 #define SIZE (AVX512F_LEN / 16)
  10 #include "avx512f-mask-type.h"
  11
  12 #if AVX512F_LEN == 512
  13 #undef CMP
  14 #define CMP(imm, rel)                                   \
  15     dst_ref = 0;                                        \
  16     for (i = 0; i < 32; i++)                            \
  17     {                                                   \
  18       dst_ref = ((rel) << i) | dst_ref;                 \
  19     }                                                   \
  20     source1.x = _mm512_loadu_si512 (s1);                \
  21     source2.x = _mm512_loadu_si512 (s2);                \
  22     dst1 = _mm512_cmp_epu16_mask (source1.x, source2.x, imm);\
  23     dst2 = _mm512_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  24     if (dst_ref != dst1) abort();                       \
  25     if ((mask & dst_ref) != dst2) abort();
  26 #endif
  27
  28 #if AVX512F_LEN == 256
  29 #undef CMP
  30 #define CMP(imm, rel)                                   \
  31     dst_ref = 0;                                        \
  32     for (i = 0; i < 16; i++)                            \
  33     {                                                   \
  34       dst_ref = ((rel) << i) | dst_ref;                 \
  35     }                                                   \
  36     source1.x = _mm256_loadu_si256 ((__m256i*)s1);      \
  37     source2.x = _mm256_loadu_si256 ((__m256i*)s2);      \
  38     dst1 = _mm256_cmp_epu16_mask (source1.x, source2.x, imm);\
  39     dst2 = _mm256_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  40     if (dst_ref != dst1) abort();                       \
  41     if ((mask & dst_ref) != dst2) abort();
  42 #endif
  43
  44 #if AVX512F_LEN == 128
  45 #undef CMP
  46 #define CMP(imm, rel)                                   \
  47     dst_ref = 0;                                        \
  48     for (i = 0; i < 8; i++)                             \
  49     {                                                   \
  50       dst_ref = ((rel) << i) | dst_ref;                 \
  51     }                                                   \
  52     source1.x = _mm_loadu_si128 ((__m128i*)s1);         \
  53     source2.x = _mm_loadu_si128 ((__m128i*)s2);         \
  54     dst1 = _mm_cmp_epu16_mask (source1.x, source2.x, imm);\
  55     dst2 = _mm_mask_cmp_epu16_mask (mask, source1.x, source2.x, imm);\
  56     if (dst_ref != dst1) abort();                       \
  57     if ((mask & dst_ref) != dst2) abort();
  58 #endif
  59
  60 void
  61 TEST ()
  62 {
  63     unsigned short s1[32] = {2134,  6678,  453, 54646,
  64                             231, 5674,  111, 23241,
  65                             12314,  145,  671, 7575,
  66                             23455,  166, 5321, 5673,
  67                             2134,  6678,  453, 54646,
  68                             231, 5674,  111, 23241,
  69                             12314,  145,  671, 7575,
  70                             23455,  166, 5321, 5673};
  71     unsigned short s2[32] = {41124, 6678, 8653,   856,
  72                             231, 4646,  111,   124,
  73                             2745, 4567, 3676,   123,
  74                             714, 3589, 5683,  5673,
  75                             41124, 6678, 8653,   856,
  76                             231, 4646,  111,   124,
  77                             2745, 4567, 3676,   123,
  78                             714, 3589, 5683,  5673};
  79     UNION_TYPE (AVX512F_LEN, i_w) source1, source2;
  80     MASK_TYPE dst1, dst2, dst_ref;
  81     MASK_TYPE mask = MASK_VALUE;
  82     int i;
  83
  84     CMP(0x00, s1[i] == s2[i]);
  85     CMP(0x01, s1[i] < s2[i]);
  86     CMP(0x02, s1[i] <= s2[i]);
  87     CMP(0x03, 0);
  88     CMP(0x04, s1[i] != s2[i]);
  89     CMP(0x05, s1[i] >= s2[i]);
  90     CMP(0x06, s1[i] > s2[i]);
  91     CMP(0x07, 1);
  92 }