src/etc/platform-intrinsics/x86/avx2.json

   1 {
   2     "llvm_prefix": "llvm.x86.avx2.",
   3     "intrinsics": [
   4         {
   5             "intrinsic": "256_abs_{0.data_type}",
   6             "width": [256],
   7             "llvm": "pabs.{0.data_type_short}",
   8             "ret": "s(8-32)",
   9             "args": ["0"]
  10         },
  11         {
  12             "intrinsic": "256_adds_{0.data_type}",
  13             "width": [256],
  14             "llvm": "padd{0.kind_short}s.{0.data_type_short}",
  15             "ret": "i(8-16)",
  16             "args": ["0", "0"]
  17         },
  18         {
  19             "intrinsic": "256_avg_{0.data_type}",
  20             "width": [256],
  21             "llvm": "pavg.{0.data_type_short}",
  22             "ret": "u(8-16)",
  23             "args": ["0", "0"]
  24         },
  25         {
  26             "intrinsic": "256_hadd_{0.data_type}",
  27             "width": [256],
  28             "llvm": "phadd.{0.data_type_short}",
  29             "ret": "s(16-32)",
  30             "args": ["0", "0"]
  31         },
  32         {
  33             "intrinsic": "256_hadds_epi16",
  34             "width": [256],
  35             "llvm": "phadd.sw",
  36             "ret": "s16",
  37             "args": ["0", "0"]
  38         },
  39         {
  40             "intrinsic": "256_hsub_{0.data_type}",
  41             "width": [256],
  42             "llvm": "phsub.{0.data_type_short}",
  43             "ret": "s(16-32)",
  44             "args": ["0", "0"]
  45         },
  46         {
  47             "intrinsic": "256_hsubs_epi16",
  48             "width": [256],
  49             "llvm": "phsub.sw",
  50             "ret": "s16",
  51             "args": ["0", "0"]
  52         },
  53         {
  54             "intrinsic": "256_madd_epi16",
  55             "width": [256],
  56             "llvm": "pmadd.wd",
  57             "ret": "s32",
  58             "args": ["s16", "s16"]
  59         },
  60         {
  61             "intrinsic": "256_maddubs_epi16",
  62             "width": [256],
  63             "llvm": "pmadd.ub.sw",
  64             "ret": "s16",
  65             "args": ["s8", "s8"]
  66         },
  67         {
  68             "intrinsic": "{0.width_mm}_mask_i32gather_{0.data_type}",
  69             "width": [128, 256],
  70             "llvm": "gather.d.{0.data_type_short}{0.width_suffix}",
  71             "ret": ["s32", "f32"],
  72             "args": ["0", "0SPc/S8", "s32", "0s->0", "S32/8"]
  73         },
  74         {
  75             "intrinsic": "{0.width_mm}_mask_i32gather_{0.data_type}",
  76             "width": [128, 256],
  77             "llvm": "gather.d.{0.data_type_short}{0.width_suffix}",
  78             "ret": ["s64", "f64"],
  79             "args": ["0", "0SPc/S8", "s32x128", "0s->0", "S32/8"]
  80         },
  81         {
  82             "intrinsic": "{3.width_mm}_mask_i64gather_{0.data_type}",
  83             "width": [128, 256],
  84             "llvm": "gather.q.{0.data_type_short}{0.width_suffix}",
  85             "ret": ["s32x128", "f32x128"],
  86             "args": ["0", "0SPc/S8", "s64", "0s->0", "S32/8"]
  87         },
  88         {
  89             "intrinsic": "{0.width_mm}_mask_i64gather_{0.data_type}",
  90             "width": [128, 256],
  91             "llvm": "gather.q.{0.data_type_short}{0.width_suffix}",
  92             "ret": ["s64", "f64"],
  93             "args": ["0", "0SPc/S8", "s64", "0s->0", "S32/8"]
  94         },
  95         {
  96             "intrinsic": "{0.width_mm}_maskload_{0.data_type}",
  97             "width": [128, 256],
  98             "llvm": "maskload.{0.data_type_short}{0.width_suffix}",
  99             "ret": ["s(32-64)"],
 100             "args": ["0Pc/S8", "0"]
 101         },
 102         {
 103             "intrinsic": "{2.width_mm}_maskstore_{2.data_type}",
 104             "width": [128, 256],
 105             "llvm": "maskstore.{2.data_type_short}{2.width_suffix}",
 106             "ret": "V",
 107             "args": ["S(32-64)Pm/S8", "1Dv", "2"]
 108         },
 109         {
 110             "intrinsic": "256_max_{0.data_type}",
 111             "width": [256],
 112             "llvm": "pmax{0.kind}.{0.data_type_short}",
 113             "ret": "i(8-32)",
 114             "args": ["0", "0"]
 115         },
 116         {
 117             "intrinsic": "256_min_{0.data_type}",
 118             "width": [256],
 119             "llvm": "pmin{0.kind}.{0.data_type_short}",
 120             "ret": "i(8-32)",
 121             "args": ["0", "0"]
 122         },
 123         {
 124             "intrinsic": "256_movemask_epi8",
 125             "width": [256],
 126             "llvm": "pmovmskb",
 127             "ret": "S32",
 128             "args": ["s8"]
 129         },
 130         {
 131             "intrinsic": "256_mpsadbw_epu8",
 132             "width": [256],
 133             "llvm": "mpsadbw",
 134             "ret": "u16",
 135             "args": ["u8", "u8", "S32/8"]
 136         },
 137         {
 138             "intrinsic": "256_mul_{0.data_type}",
 139             "width": [256],
 140             "llvm": "pmul{0.data_type_short}.dq",
 141             "ret": "i64",
 142             "args": ["0dn", "0dn"]
 143         },
 144         {
 145             "intrinsic": "256_mulhi_{0.data_type}",
 146             "width": [256],
 147             "llvm": "pmulh{0.data_type_short}.w",
 148             "ret": "i16",
 149             "args": ["0", "0"]
 150         },
 151         {
 152             "intrinsic": "256_mulhrs_epi16",
 153             "width": [256],
 154             "llvm": "pmul.hr.sw",
 155             "ret": "s16",
 156             "args": ["0", "0"]
 157         },
 158         {
 159             "intrinsic": "256_pack{0.kind_short}s_{1.data_type}",
 160             "width": [256],
 161             "llvm": "pack{0.kind}s{1.data_type_short}{0.data_type_short}",
 162             "ret": "i(8-16)",
 163             "args": ["0hws", "0hws"]
 164         },
 165         {
 166             "intrinsic": "256_permutevar8x32_{0.data_type}",
 167             "width": [256],
 168             "llvm": "perm{0.data_type_short}",
 169             "ret": ["s32", "f32"],
 170             "args": ["0", "0s"]
 171         },
 172         {
 173             "intrinsic": "256_sad_epu8",
 174             "width": [256],
 175             "llvm": "psad.bw",
 176             "ret": "u8",
 177             "args": ["0", "0"]
 178         },
 179         {
 180             "intrinsic": "256_shuffle_epi8",
 181             "width": [256],
 182             "llvm": "pshuf.b",
 183             "ret": "s8",
 184             "args": ["0", "0"]
 185         },
 186         {
 187             "intrinsic": "256_sign_{0.data_type}",
 188             "width": [256],
 189             "llvm": "psign.{0.data_type_short}",
 190             "ret": "s(8-32)",
 191             "args": ["0", "0"]
 192         },
 193         {
 194             "intrinsic": "256_subs_{0.data_type}",
 195             "width": [256],
 196             "llvm": "psub{0.kind_short}s.{0.data_type_short}",
 197             "ret": "i(8-16)",
 198             "args": ["0", "0"]
 199         }
 200     ]
 201 }