test/MC/AMDGPU/gfx10_asm_dpp8.s

   1 // RUN: not llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=+WavefrontSize32,-WavefrontSize64 -show-encoding %s | FileCheck --check-prefixes=GFX10,W32 %s
   2 // RUN: not llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=-WavefrontSize32,+WavefrontSize64 -show-encoding %s | FileCheck --check-prefixes=GFX10,W64 %s
   3 // RUN: not llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=+WavefrontSize32,-WavefrontSize64 -show-encoding %s 2>&1 | FileCheck --check-prefixes=GFX10-ERR,W32-ERR %s
   4 // RUN: not llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=-WavefrontSize32,+WavefrontSize64 -show-encoding %s 2>&1 | FileCheck --check-prefixes=GFX10-ERR,W64-ERR %s
   5
   6 v_mov_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
   7 // GFX10: encoding: [0xe9,0x02,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
   8
   9 v_cvt_f32_i32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  10 // GFX10: encoding: [0xe9,0x0a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  11
  12 v_cvt_f32_u32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  13 // GFX10: encoding: [0xe9,0x0c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  14
  15 v_cvt_u32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  16 // GFX10: encoding: [0xe9,0x0e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  17
  18 v_cvt_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  19 // GFX10: encoding: [0xe9,0x10,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  20
  21 v_mov_fed_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  22 // GFX10: encoding: [0xe9,0x12,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  23
  24 v_cvt_f16_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  25 // GFX10: encoding: [0xe9,0x14,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  26
  27 v_cvt_f32_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  28 // GFX10: encoding: [0xe9,0x16,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  29
  30 v_cvt_rpi_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  31 // GFX10: encoding: [0xe9,0x18,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  32
  33 v_cvt_flr_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  34 // GFX10: encoding: [0xe9,0x1a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  35
  36 v_cvt_off_f32_i4_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  37 // GFX10: encoding: [0xe9,0x1c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  38
  39 v_cvt_f32_ubyte0_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  40 // GFX10: encoding: [0xe9,0x22,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  41
  42 v_cvt_f32_ubyte1_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  43 // GFX10: encoding: [0xe9,0x24,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  44
  45 v_cvt_f32_ubyte2_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  46 // GFX10: encoding: [0xe9,0x26,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  47
  48 v_cvt_f32_ubyte3_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  49 // GFX10: encoding: [0xe9,0x28,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  50
  51 v_fract_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  52 // GFX10: encoding: [0xe9,0x40,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  53
  54 v_trunc_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  55 // GFX10: encoding: [0xe9,0x42,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  56
  57 v_ceil_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  58 // GFX10: encoding: [0xe9,0x44,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  59
  60 v_rndne_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  61 // GFX10: encoding: [0xe9,0x46,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  62
  63 v_floor_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  64 // GFX10: encoding: [0xe9,0x48,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  65
  66 v_exp_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  67 // GFX10: encoding: [0xe9,0x4a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  68
  69 v_log_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  70 // GFX10: encoding: [0xe9,0x4e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  71
  72 v_rcp_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  73 // GFX10: encoding: [0xe9,0x54,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  74
  75 v_rcp_iflag_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  76 // GFX10: encoding: [0xe9,0x56,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  77
  78 v_rsq_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  79 // GFX10: encoding: [0xe9,0x5c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  80
  81 v_sqrt_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  82 // GFX10: encoding: [0xe9,0x66,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  83
  84 v_sin_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  85 // GFX10: encoding: [0xe9,0x6a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  86
  87 v_cos_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  88 // GFX10: encoding: [0xe9,0x6c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  89
  90 v_not_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  91 // GFX10: encoding: [0xe9,0x6e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  92
  93 v_bfrev_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  94 // GFX10: encoding: [0xe9,0x70,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  95
  96 v_ffbh_u32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
  97 // GFX10: encoding: [0xe9,0x72,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
  98
  99 v_ffbl_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 100 // GFX10: encoding: [0xe9,0x74,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 101
 102 v_ffbh_i32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 103 // GFX10: encoding: [0xe9,0x76,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 104
 105 v_frexp_exp_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 106 // GFX10: encoding: [0xe9,0x7e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 107
 108 v_frexp_mant_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 109 // GFX10: encoding: [0xe9,0x80,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 110
 111 v_cvt_f16_u16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 112 // GFX10: encoding: [0xe9,0xa0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 113
 114 v_cvt_f16_i16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 115 // GFX10: encoding: [0xe9,0xa2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 116
 117 v_cvt_u16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 118 // GFX10: encoding: [0xe9,0xa4,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 119
 120 v_cvt_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 121 // GFX10: encoding: [0xe9,0xa6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 122
 123 v_rcp_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 124 // GFX10: encoding: [0xe9,0xa8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 125
 126 v_sqrt_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 127 // GFX10: encoding: [0xe9,0xaa,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 128
 129 v_rsq_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 130 // GFX10: encoding: [0xe9,0xac,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 131
 132 v_log_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 133 // GFX10: encoding: [0xe9,0xae,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 134
 135 v_exp_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 136 // GFX10: encoding: [0xe9,0xb0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 137
 138 v_frexp_mant_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 139 // GFX10: encoding: [0xe9,0xb2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 140
 141 v_frexp_exp_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 142 // GFX10: encoding: [0xe9,0xb4,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 143
 144 v_floor_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 145 // GFX10: encoding: [0xe9,0xb6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 146
 147 v_ceil_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 148 // GFX10: encoding: [0xe9,0xb8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 149
 150 v_trunc_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 151 // GFX10: encoding: [0xe9,0xba,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 152
 153 v_rndne_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 154 // GFX10: encoding: [0xe9,0xbc,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 155
 156 v_fract_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 157 // GFX10: encoding: [0xe9,0xbe,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 158
 159 v_sin_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 160 // GFX10: encoding: [0xe9,0xc0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 161
 162 v_cos_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 163 // GFX10: encoding: [0xe9,0xc2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 164
 165 v_cvt_norm_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 166 // GFX10: encoding: [0xe9,0xc6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 167
 168 v_cvt_norm_u16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
 169 // GFX10: encoding: [0xe9,0xc8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 170
 171 v_add_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 172 // GFX10: encoding: [0xe9,0x04,0x0a,0x06,0x01,0x88,0xc6,0xfa]
 173
 174 v_sub_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 175 // GFX10: encoding: [0xe9,0x04,0x0a,0x08,0x01,0x88,0xc6,0xfa]
 176
 177 v_subrev_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 178 // GFX10: encoding: [0xe9,0x04,0x0a,0x0a,0x01,0x88,0xc6,0xfa]
 179
 180 v_mul_legacy_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 181 // GFX10: encoding: [0xe9,0x04,0x0a,0x0e,0x01,0x88,0xc6,0xfa]
 182
 183 v_mul_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 184 // GFX10: encoding: [0xe9,0x04,0x0a,0x10,0x01,0x88,0xc6,0xfa]
 185
 186 v_mul_i32_i24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 187 // GFX10: encoding: [0xe9,0x04,0x0a,0x12,0x01,0x88,0xc6,0xfa]
 188
 189 v_mul_hi_i32_i24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 190 // GFX10: encoding: [0xe9,0x04,0x0a,0x14,0x01,0x88,0xc6,0xfa]
 191
 192 v_mul_u32_u24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 193 // GFX10: encoding: [0xe9,0x04,0x0a,0x16,0x01,0x88,0xc6,0xfa]
 194
 195 v_mul_hi_u32_u24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 196 // GFX10: encoding: [0xe9,0x04,0x0a,0x18,0x01,0x88,0xc6,0xfa]
 197
 198 v_min_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 199 // GFX10: encoding: [0xe9,0x04,0x0a,0x1e,0x01,0x88,0xc6,0xfa]
 200
 201 v_max_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 202 // GFX10: encoding: [0xe9,0x04,0x0a,0x20,0x01,0x88,0xc6,0xfa]
 203
 204 v_min_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 205 // GFX10: encoding: [0xe9,0x04,0x0a,0x22,0x01,0x88,0xc6,0xfa]
 206
 207 v_max_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 208 // GFX10: encoding: [0xe9,0x04,0x0a,0x24,0x01,0x88,0xc6,0xfa]
 209
 210 v_min_u32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 211 // GFX10: encoding: [0xe9,0x04,0x0a,0x26,0x01,0x88,0xc6,0xfa]
 212
 213 v_max_u32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 214 // GFX10: encoding: [0xe9,0x04,0x0a,0x28,0x01,0x88,0xc6,0xfa]
 215
 216 v_lshrrev_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 217 // GFX10: encoding: [0xe9,0x04,0x0a,0x2c,0x01,0x88,0xc6,0xfa]
 218
 219 v_ashrrev_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 220 // GFX10: encoding: [0xe9,0x04,0x0a,0x30,0x01,0x88,0xc6,0xfa]
 221
 222 v_lshlrev_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 223 // GFX10: encoding: [0xe9,0x04,0x0a,0x34,0x01,0x88,0xc6,0xfa]
 224
 225 v_and_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 226 // GFX10: encoding: [0xe9,0x04,0x0a,0x36,0x01,0x88,0xc6,0xfa]
 227
 228 v_or_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 229 // GFX10: encoding: [0xe9,0x04,0x0a,0x38,0x01,0x88,0xc6,0xfa]
 230
 231 v_xor_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 232 // GFX10: encoding: [0xe9,0x04,0x0a,0x3a,0x01,0x88,0xc6,0xfa]
 233
 234 v_xnor_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 235 // GFX10: encoding: [0xe9,0x04,0x0a,0x3c,0x01,0x88,0xc6,0xfa]
 236
 237 v_add_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 238 // GFX10: encoding: [0xe9,0x04,0x0a,0x64,0x01,0x88,0xc6,0xfa]
 239
 240 v_sub_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 241 // GFX10: encoding: [0xe9,0x04,0x0a,0x66,0x01,0x88,0xc6,0xfa]
 242
 243 v_subrev_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 244 // GFX10: encoding: [0xe9,0x04,0x0a,0x68,0x01,0x88,0xc6,0xfa]
 245
 246 v_mul_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 247 // GFX10: encoding: [0xe9,0x04,0x0a,0x6a,0x01,0x88,0xc6,0xfa]
 248
 249 v_max_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 250 // GFX10: encoding: [0xe9,0x04,0x0a,0x72,0x01,0x88,0xc6,0xfa]
 251
 252 v_min_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 253 // GFX10: encoding: [0xe9,0x04,0x0a,0x74,0x01,0x88,0xc6,0xfa]
 254
 255 v_ldexp_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7]
 256 // GFX10: encoding: [0xe9,0x04,0x0a,0x76,0x01,0x88,0xc6,0xfa]
 257
 258 v_mov_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:0
 259 // GFX10: encoding: [0xe9,0x02,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 260
 261 v_mov_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 262 // GFX10: encoding: [0xea,0x02,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 263
 264 v_cvt_f32_i32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 265 // GFX10: encoding: [0xea,0x0a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 266
 267 v_cvt_f32_u32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 268 // GFX10: encoding: [0xea,0x0c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 269
 270 v_cvt_u32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 271 // GFX10: encoding: [0xea,0x0e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 272
 273 v_cvt_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 274 // GFX10: encoding: [0xea,0x10,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 275
 276 v_mov_fed_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 277 // GFX10: encoding: [0xea,0x12,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 278
 279 v_cvt_f16_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 280 // GFX10: encoding: [0xea,0x14,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 281
 282 v_cvt_f32_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 283 // GFX10: encoding: [0xea,0x16,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 284
 285 v_cvt_rpi_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 286 // GFX10: encoding: [0xea,0x18,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 287
 288 v_cvt_flr_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 289 // GFX10: encoding: [0xea,0x1a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 290
 291 v_cvt_off_f32_i4_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 292 // GFX10: encoding: [0xea,0x1c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 293
 294 v_cvt_f32_ubyte0_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 295 // GFX10: encoding: [0xea,0x22,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 296
 297 v_cvt_f32_ubyte1_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 298 // GFX10: encoding: [0xea,0x24,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 299
 300 v_cvt_f32_ubyte2_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 301 // GFX10: encoding: [0xea,0x26,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 302
 303 v_cvt_f32_ubyte3_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 304 // GFX10: encoding: [0xea,0x28,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 305
 306 v_fract_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 307 // GFX10: encoding: [0xea,0x40,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 308
 309 v_trunc_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 310 // GFX10: encoding: [0xea,0x42,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 311
 312 v_ceil_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 313 // GFX10: encoding: [0xea,0x44,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 314
 315 v_rndne_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 316 // GFX10: encoding: [0xea,0x46,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 317
 318 v_floor_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 319 // GFX10: encoding: [0xea,0x48,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 320
 321 v_exp_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 322 // GFX10: encoding: [0xea,0x4a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 323
 324 v_log_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 325 // GFX10: encoding: [0xea,0x4e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 326
 327 v_rcp_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 328 // GFX10: encoding: [0xea,0x54,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 329
 330 v_rcp_iflag_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 331 // GFX10: encoding: [0xea,0x56,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 332
 333 v_rsq_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 334 // GFX10: encoding: [0xea,0x5c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 335
 336 v_sqrt_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 337 // GFX10: encoding: [0xea,0x66,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 338
 339 v_sin_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 340 // GFX10: encoding: [0xea,0x6a,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 341
 342 v_cos_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 343 // GFX10: encoding: [0xea,0x6c,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 344
 345 v_not_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 346 // GFX10: encoding: [0xea,0x6e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 347
 348 v_bfrev_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 349 // GFX10: encoding: [0xea,0x70,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 350
 351 v_ffbh_u32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 352 // GFX10: encoding: [0xea,0x72,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 353
 354 v_ffbl_b32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 355 // GFX10: encoding: [0xea,0x74,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 356
 357 v_ffbh_i32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 358 // GFX10: encoding: [0xea,0x76,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 359
 360 v_frexp_exp_i32_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 361 // GFX10: encoding: [0xea,0x7e,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 362
 363 v_frexp_mant_f32_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 364 // GFX10: encoding: [0xea,0x80,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 365
 366 v_cvt_f16_u16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 367 // GFX10: encoding: [0xea,0xa0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 368
 369 v_cvt_f16_i16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 370 // GFX10: encoding: [0xea,0xa2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 371
 372 v_cvt_u16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 373 // GFX10: encoding: [0xea,0xa4,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 374
 375 v_cvt_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 376 // GFX10: encoding: [0xea,0xa6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 377
 378 v_rcp_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 379 // GFX10: encoding: [0xea,0xa8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 380
 381 v_sqrt_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 382 // GFX10: encoding: [0xea,0xaa,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 383
 384 v_rsq_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 385 // GFX10: encoding: [0xea,0xac,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 386
 387 v_log_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 388 // GFX10: encoding: [0xea,0xae,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 389
 390 v_exp_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 391 // GFX10: encoding: [0xea,0xb0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 392
 393 v_frexp_mant_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 394 // GFX10: encoding: [0xea,0xb2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 395
 396 v_frexp_exp_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 397 // GFX10: encoding: [0xea,0xb4,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 398
 399 v_floor_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 400 // GFX10: encoding: [0xea,0xb6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 401
 402 v_ceil_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 403 // GFX10: encoding: [0xea,0xb8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 404
 405 v_trunc_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 406 // GFX10: encoding: [0xea,0xba,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 407
 408 v_rndne_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 409 // GFX10: encoding: [0xea,0xbc,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 410
 411 v_fract_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 412 // GFX10: encoding: [0xea,0xbe,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 413
 414 v_sin_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 415 // GFX10: encoding: [0xea,0xc0,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 416
 417 v_cos_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 418 // GFX10: encoding: [0xea,0xc2,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 419
 420 v_cvt_norm_i16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 421 // GFX10: encoding: [0xea,0xc6,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 422
 423 v_cvt_norm_u16_f16_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7] fi:1
 424 // GFX10: encoding: [0xea,0xc8,0x0a,0x7e,0x01,0x88,0xc6,0xfa]
 425
 426 v_add_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 427 // GFX10: encoding: [0xea,0x04,0x0a,0x06,0x01,0x88,0xc6,0xfa]
 428
 429 v_sub_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 430 // GFX10: encoding: [0xea,0x04,0x0a,0x08,0x01,0x88,0xc6,0xfa]
 431
 432 v_subrev_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 433 // GFX10: encoding: [0xea,0x04,0x0a,0x0a,0x01,0x88,0xc6,0xfa]
 434
 435 v_mul_legacy_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 436 // GFX10: encoding: [0xea,0x04,0x0a,0x0e,0x01,0x88,0xc6,0xfa]
 437
 438 v_mul_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 439 // GFX10: encoding: [0xea,0x04,0x0a,0x10,0x01,0x88,0xc6,0xfa]
 440
 441 v_mul_i32_i24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 442 // GFX10: encoding: [0xea,0x04,0x0a,0x12,0x01,0x88,0xc6,0xfa]
 443
 444 v_mul_hi_i32_i24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 445 // GFX10: encoding: [0xea,0x04,0x0a,0x14,0x01,0x88,0xc6,0xfa]
 446
 447 v_mul_u32_u24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 448 // GFX10: encoding: [0xea,0x04,0x0a,0x16,0x01,0x88,0xc6,0xfa]
 449
 450 v_mul_hi_u32_u24_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 451 // GFX10: encoding: [0xea,0x04,0x0a,0x18,0x01,0x88,0xc6,0xfa]
 452
 453 v_min_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 454 // GFX10: encoding: [0xea,0x04,0x0a,0x1e,0x01,0x88,0xc6,0xfa]
 455
 456 v_max_f32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 457 // GFX10: encoding: [0xea,0x04,0x0a,0x20,0x01,0x88,0xc6,0xfa]
 458
 459 v_min_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 460 // GFX10: encoding: [0xea,0x04,0x0a,0x22,0x01,0x88,0xc6,0xfa]
 461
 462 v_max_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 463 // GFX10: encoding: [0xea,0x04,0x0a,0x24,0x01,0x88,0xc6,0xfa]
 464
 465 v_min_u32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 466 // GFX10: encoding: [0xea,0x04,0x0a,0x26,0x01,0x88,0xc6,0xfa]
 467
 468 v_max_u32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 469 // GFX10: encoding: [0xea,0x04,0x0a,0x28,0x01,0x88,0xc6,0xfa]
 470
 471 v_lshrrev_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 472 // GFX10: encoding: [0xea,0x04,0x0a,0x2c,0x01,0x88,0xc6,0xfa]
 473
 474 v_ashrrev_i32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 475 // GFX10: encoding: [0xea,0x04,0x0a,0x30,0x01,0x88,0xc6,0xfa]
 476
 477 v_lshlrev_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 478 // GFX10: encoding: [0xea,0x04,0x0a,0x34,0x01,0x88,0xc6,0xfa]
 479
 480 v_and_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 481 // GFX10: encoding: [0xea,0x04,0x0a,0x36,0x01,0x88,0xc6,0xfa]
 482
 483 v_or_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 484 // GFX10: encoding: [0xea,0x04,0x0a,0x38,0x01,0x88,0xc6,0xfa]
 485
 486 v_xor_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 487 // GFX10: encoding: [0xea,0x04,0x0a,0x3a,0x01,0x88,0xc6,0xfa]
 488
 489 v_xnor_b32_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 490 // GFX10: encoding: [0xea,0x04,0x0a,0x3c,0x01,0x88,0xc6,0xfa]
 491
 492 v_add_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 493 // GFX10: encoding: [0xea,0x04,0x0a,0x64,0x01,0x88,0xc6,0xfa]
 494
 495 v_sub_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 496 // GFX10: encoding: [0xea,0x04,0x0a,0x66,0x01,0x88,0xc6,0xfa]
 497
 498 v_subrev_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 499 // GFX10: encoding: [0xea,0x04,0x0a,0x68,0x01,0x88,0xc6,0xfa]
 500
 501 v_mul_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 502 // GFX10: encoding: [0xea,0x04,0x0a,0x6a,0x01,0x88,0xc6,0xfa]
 503
 504 v_max_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 505 // GFX10: encoding: [0xea,0x04,0x0a,0x72,0x01,0x88,0xc6,0xfa]
 506
 507 v_min_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 508 // GFX10: encoding: [0xea,0x04,0x0a,0x74,0x01,0x88,0xc6,0xfa]
 509
 510 v_ldexp_f16_dpp v5, v1, v2 dpp8:[0,1,2,3,4,5,6,7] fi:1
 511 // GFX10: encoding: [0xea,0x04,0x0a,0x76,0x01,0x88,0xc6,0xfa]
 512
 513 v_add_co_ci_u32_dpp v0, vcc_lo, v0, v0, vcc_lo dpp8:[7,6,5,4,3,2,1,0]
 514 // W32: [0xe9,0x00,0x00,0x50,0x00,0x77,0x39,0x05]
 515 // W64-ERR: error: instruction not supported on this GPU
 516
 517 v_sub_co_ci_u32_dpp v0, vcc_lo, v0, v0, vcc_lo dpp8:[7,6,5,4,3,2,1,0] fi:0
 518 // W32: [0xe9,0x00,0x00,0x52,0x00,0x77,0x39,0x05]
 519 // W64-ERR: error: instruction not supported on this GPU
 520
 521 v_subrev_co_ci_u32_dpp v0, vcc_lo, v0, v0, vcc_lo dpp8:[7,6,5,4,3,2,1,0] fi:1
 522 // W32: [0xea,0x00,0x00,0x54,0x00,0x77,0x39,0x05]
 523 // W64-ERR: error: instruction not supported on this GPU
 524
 525 v_add_co_ci_u32_dpp v0, vcc, v0, v0, vcc dpp8:[7,6,5,4,3,2,1,0] fi:1
 526 // W64: [0xea,0x00,0x00,0x50,0x00,0x77,0x39,0x05]
 527 // W32-ERR: error: instruction not supported on this GPU
 528
 529 v_sub_co_ci_u32_dpp v0, vcc, v0, v0, vcc dpp8:[7,6,5,4,3,2,1,0] fi:1
 530 // W64: [0xea,0x00,0x00,0x52,0x00,0x77,0x39,0x05]
 531 // W32-ERR: error: instruction not supported on this GPU
 532
 533 v_subrev_co_ci_u32_dpp v0, vcc, v0, v0, vcc dpp8:[7,6,5,4,3,2,1,0]
 534 // W64: [0xe9,0x00,0x00,0x54,0x00,0x77,0x39,0x05]
 535 // W32-ERR: error: instruction not supported on this GPU
 536
 537 v_add_nc_u32_dpp v5, v1, v255 dpp8:[7,6,5,4,3,2,1,0]
 538 // GFX10: [0xe9,0xfe,0x0b,0x4a,0x01,0x77,0x39,0x05]
 539
 540 v_add_nc_u32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1
 541 // GFX10: [0xea,0x04,0x0a,0x4a,0x01,0x77,0x39,0x05]
 542
 543 v_sub_nc_u32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0]
 544 // GFX10: [0xe9,0x04,0x0a,0x4c,0x01,0x77,0x39,0x05]
 545
 546 v_sub_nc_u32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1
 547 // GFX10: [0xea,0x04,0x0a,0x4c,0x01,0x77,0x39,0x05]
 548
 549 v_subrev_nc_u32_dpp v5, v1, v255 dpp8:[7,6,5,4,3,2,1,0]
 550 // GFX10: [0xe9,0xfe,0x0b,0x4e,0x01,0x77,0x39,0x05]
 551
 552 v_subrev_nc_u32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1
 553 // GFX10: [0xea,0x04,0x0a,0x4e,0x01,0x77,0x39,0x05]
 554
 555 v_mac_f32 v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0]
 556 // GFX10: v_mac_f32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0xe9,0x04,0x0a,0x3e,0x01,0x77,0x39,0x05]
 557
 558 v_mac_f32 v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1
 559 // GFX10: v_mac_f32_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 ; encoding: [0xea,0x04,0x0a,0x3e,0x01,0x77,0x39,0x05]