llvm/test/MC/AMDGPU/gfx12_asm_vop3p.s

   1 // RUN: llvm-mc -triple=amdgcn -show-encoding -mcpu=gfx1200 %s | FileCheck --check-prefix=GFX12 %s
   2
   3 v_dot2_f32_bf16 v5, v1, v2, v3
   4 // GFX12: [0x05,0x40,0x1a,0xcc,0x01,0x05,0x0e,0x1c]
   5
   6 v_dot2_f32_bf16 v5, v1, v2, s3
   7 // GFX12: [0x05,0x40,0x1a,0xcc,0x01,0x05,0x0e,0x18]
   8
   9 v_dot2_f32_bf16 v5, v255, v255, s105
  10 // GFX12: [0x05,0x40,0x1a,0xcc,0xff,0xff,0xa7,0x19]
  11
  12 v_dot2_f32_bf16 v5, s1, s2, v255
  13 // GFX12: [0x05,0x40,0x1a,0xcc,0x01,0x04,0xfc,0x1f]
  14
  15 v_dot2_f32_bf16 v5, s105, s105, m0
  16 // GFX12: [0x05,0x40,0x1a,0xcc,0x69,0xd2,0xf4,0x19]
  17
  18 v_dot2_f32_bf16 v5, vcc_lo, ttmp15, vcc_lo
  19 // GFX12: [0x05,0x40,0x1a,0xcc,0x6a,0xf6,0xa8,0x19]
  20
  21 v_dot2_f32_bf16 v5, vcc_hi, 0xfe0b, vcc_hi
  22 // GFX12: [0x05,0x40,0x1a,0xcc,0x6b,0xfe,0xad,0x19,0x0b,0xfe,0x00,0x00]
  23
  24 v_dot2_f32_bf16 v5, ttmp15, src_scc, ttmp15
  25 // GFX12: [0x05,0x40,0x1a,0xcc,0x7b,0xfa,0xed,0x19]
  26
  27 v_dot2_f32_bf16 v5, m0, -1, exec_hi
  28 // GFX12: [0x05,0x40,0x1a,0xcc,0x7d,0x82,0xfd,0x19]
  29
  30 v_dot2_f32_bf16 v5, exec_lo, null, exec_lo
  31 // GFX12: [0x05,0x40,0x1a,0xcc,0x7e,0xf8,0xf8,0x19]
  32
  33 v_dot2_f32_bf16 v5, exec_hi, exec_lo, null
  34 // GFX12: [0x05,0x40,0x1a,0xcc,0x7f,0xfc,0xf0,0x19]
  35
  36 v_dot2_f32_bf16 v5, null, exec_hi, -1
  37 // GFX12: [0x05,0x40,0x1a,0xcc,0x7c,0xfe,0x04,0x1b]
  38
  39 v_dot2_f32_bf16 v5, -1, m0, 0xaf123456
  40 // GFX12: [0x05,0x40,0x1a,0xcc,0xc1,0xfa,0xfc,0x1b,0x56,0x34,0x12,0xaf]
  41
  42 v_dot2_f32_bf16 v5, src_scc, vcc_lo, src_scc neg_lo:[1,0,0] neg_hi:[1,0,0]
  43 // GFX12: [0x05,0x41,0x1a,0xcc,0xfd,0xd4,0xf4,0x3b]
  44
  45 v_dot2_f32_bf16 v255, 0xfe0b, vcc_hi, 0.5 neg_lo:[0,0,0] neg_hi:[0,0,0] clamp
  46 // GFX12: [0xff,0xc0,0x1a,0xcc,0xff,0xd6,0xc0,0x1b,0x0b,0xfe,0x00,0x00]
  47
  48 v_dot2_f32_f16 v5, v1, v2, s3
  49 // GFX12: [0x05,0x40,0x13,0xcc,0x01,0x05,0x0e,0x18]
  50
  51 v_dot2_f32_f16 v5, v255, s2, s105
  52 // GFX12: [0x05,0x40,0x13,0xcc,0xff,0x05,0xa4,0x19]
  53
  54 v_dot2_f32_f16 v5, s1, v255, exec_hi
  55 // GFX12: [0x05,0x40,0x13,0xcc,0x01,0xfe,0xff,0x19]
  56
  57 v_dot2_f32_f16 v5, s105, s105, exec_lo
  58 // GFX12: [0x05,0x40,0x13,0xcc,0x69,0xd2,0xf8,0x19]
  59
  60 v_dot2_f32_f16 v5, vcc_lo, ttmp15, v3
  61 // GFX12: [0x05,0x40,0x13,0xcc,0x6a,0xf6,0x0c,0x1c]
  62
  63 v_dot2_f32_f16 v5, vcc_hi, 0xfe0b, v255
  64 // GFX12: [0x05,0x40,0x13,0xcc,0x6b,0xfe,0xfd,0x1f,0x0b,0xfe,0x00,0x00]
  65
  66 v_dot2_f32_f16 v5, ttmp15, src_scc, ttmp15
  67 // GFX12: [0x05,0x40,0x13,0xcc,0x7b,0xfa,0xed,0x19]
  68
  69 v_dot2_f32_f16 v5, m0, 0.5, m0
  70 // GFX12: [0x05,0x40,0x13,0xcc,0x7d,0xe0,0xf5,0x19]
  71
  72 v_dot2_f32_f16 v5, exec_lo, -1, vcc_hi
  73 // GFX12: [0x05,0x40,0x13,0xcc,0x7e,0x82,0xad,0x19]
  74
  75 v_dot2_f32_f16 v5, exec_hi, null, vcc_lo
  76 // GFX12: [0x05,0x40,0x13,0xcc,0x7f,0xf8,0xa8,0x19]
  77
  78 v_dot2_f32_f16 v5, null, exec_lo, 0xaf123456
  79 // GFX12: [0x05,0x40,0x13,0xcc,0x7c,0xfc,0xfc,0x1b,0x56,0x34,0x12,0xaf]
  80
  81 v_dot2_f32_f16 v5, -1, exec_hi, src_scc
  82 // GFX12: [0x05,0x40,0x13,0xcc,0xc1,0xfe,0xf4,0x1b]
  83
  84 v_dot2_f32_f16 v5, 0.5, m0, 0.5 neg_lo:[1,0,0] neg_hi:[1,0,0]
  85 // GFX12: [0x05,0x41,0x13,0xcc,0xf0,0xfa,0xc0,0x3b]
  86
  87 v_dot2_f32_f16 v5, src_scc, vcc_lo, -1 neg_lo:[0,1,0] neg_hi:[0,1,0]
  88 // GFX12: [0x05,0x42,0x13,0xcc,0xfd,0xd4,0x04,0x5b]
  89
  90 v_dot2_f32_f16 v255, 0xfe0b, vcc_hi, null neg_lo:[0,0,0] neg_hi:[0,0,0] clamp
  91 // GFX12: [0xff,0xc0,0x13,0xcc,0xff,0xd6,0xf0,0x19,0x0b,0xfe,0x00,0x00]
  92
  93 v_dot4_i32_iu8 v5, v1, v2, s3
  94 // GFX12: [0x05,0x40,0x16,0xcc,0x01,0x05,0x0e,0x18]
  95
  96 v_dot4_i32_iu8 v5, v255, s2, s105
  97 // GFX12: [0x05,0x40,0x16,0xcc,0xff,0x05,0xa4,0x19]
  98
  99 v_dot4_i32_iu8 v5, s1, v255, exec_hi
 100 // GFX12: [0x05,0x40,0x16,0xcc,0x01,0xfe,0xff,0x19]
 101
 102 v_dot4_i32_iu8 v5, s105, s105, exec_lo
 103 // GFX12: [0x05,0x40,0x16,0xcc,0x69,0xd2,0xf8,0x19]
 104
 105 v_dot4_i32_iu8 v5, vcc_lo, ttmp15, v3
 106 // GFX12: [0x05,0x40,0x16,0xcc,0x6a,0xf6,0x0c,0x1c]
 107
 108 v_dot4_i32_iu8 v5, vcc_hi, 0xaf123456, v255
 109 // GFX12: [0x05,0x40,0x16,0xcc,0x6b,0xfe,0xfd,0x1f,0x56,0x34,0x12,0xaf]
 110
 111 v_dot4_i32_iu8 v5, ttmp15, src_scc, ttmp15
 112 // GFX12: [0x05,0x40,0x16,0xcc,0x7b,0xfa,0xed,0x19]
 113
 114 v_dot4_i32_iu8 v5, m0, 0.5, m0
 115 // GFX12: [0x05,0x40,0x16,0xcc,0x7d,0xe0,0xf5,0x19]
 116
 117 v_dot4_i32_iu8 v5, exec_lo, -1, vcc_hi
 118 // GFX12: [0x05,0x40,0x16,0xcc,0x7e,0x82,0xad,0x19]
 119
 120 v_dot4_i32_iu8 v5, exec_hi, null, vcc_lo
 121 // GFX12: [0x05,0x40,0x16,0xcc,0x7f,0xf8,0xa8,0x19]
 122
 123 v_dot4_i32_iu8 v5, null, exec_lo, 0xaf123456
 124 // GFX12: [0x05,0x40,0x16,0xcc,0x7c,0xfc,0xfc,0x1b,0x56,0x34,0x12,0xaf]
 125
 126 v_dot4_i32_iu8 v5, -1, exec_hi, src_scc
 127 // GFX12: [0x05,0x40,0x16,0xcc,0xc1,0xfe,0xf4,0x1b]
 128
 129 v_dot4_i32_iu8 v5, 0.5, m0, 0.5 neg_lo:[1,0,0]
 130 // GFX12: [0x05,0x40,0x16,0xcc,0xf0,0xfa,0xc0,0x3b]
 131
 132 v_dot4_i32_iu8 v5, src_scc, vcc_lo, -1 neg_lo:[0,1,0]
 133 // GFX12: [0x05,0x40,0x16,0xcc,0xfd,0xd4,0x04,0x5b]
 134
 135 v_dot4_i32_iu8 v255, 0xaf123456, vcc_hi, null neg_lo:[0,0,0]
 136 // GFX12: [0xff,0x40,0x16,0xcc,0xff,0xd6,0xf0,0x19,0x56,0x34,0x12,0xaf]
 137
 138 v_dot4_u32_u8 v5, v1, v2, s3
 139 // GFX12: [0x05,0x40,0x17,0xcc,0x01,0x05,0x0e,0x18]
 140
 141 v_dot4_u32_u8 v5, v255, s2, s105
 142 // GFX12: [0x05,0x40,0x17,0xcc,0xff,0x05,0xa4,0x19]
 143
 144 v_dot4_u32_u8 v5, s1, v255, exec_hi
 145 // GFX12: [0x05,0x40,0x17,0xcc,0x01,0xfe,0xff,0x19]
 146
 147 v_dot4_u32_u8 v5, s105, s105, exec_lo
 148 // GFX12: [0x05,0x40,0x17,0xcc,0x69,0xd2,0xf8,0x19]
 149
 150 v_dot4_u32_u8 v5, vcc_lo, ttmp15, v3
 151 // GFX12: [0x05,0x40,0x17,0xcc,0x6a,0xf6,0x0c,0x1c]
 152
 153 v_dot4_u32_u8 v5, vcc_hi, 0xaf123456, v255
 154 // GFX12: [0x05,0x40,0x17,0xcc,0x6b,0xfe,0xfd,0x1f,0x56,0x34,0x12,0xaf]
 155
 156 v_dot4_u32_u8 v5, ttmp15, src_scc, ttmp15
 157 // GFX12: [0x05,0x40,0x17,0xcc,0x7b,0xfa,0xed,0x19]
 158
 159 v_dot4_u32_u8 v5, m0, 0.5, m0
 160 // GFX12: [0x05,0x40,0x17,0xcc,0x7d,0xe0,0xf5,0x19]
 161
 162 v_dot4_u32_u8 v5, exec_lo, -1, vcc_hi
 163 // GFX12: [0x05,0x40,0x17,0xcc,0x7e,0x82,0xad,0x19]
 164
 165 v_dot4_u32_u8 v5, exec_hi, null, vcc_lo
 166 // GFX12: [0x05,0x40,0x17,0xcc,0x7f,0xf8,0xa8,0x19]
 167
 168 v_dot4_u32_u8 v5, null, exec_lo, 0xaf123456
 169 // GFX12: [0x05,0x40,0x17,0xcc,0x7c,0xfc,0xfc,0x1b,0x56,0x34,0x12,0xaf]
 170
 171 v_dot4_u32_u8 v5, -1, exec_hi, src_scc
 172 // GFX12: [0x05,0x40,0x17,0xcc,0xc1,0xfe,0xf4,0x1b]
 173
 174 v_dot4_u32_u8 v5, 0.5, m0, 0.5
 175 // GFX12: [0x05,0x40,0x17,0xcc,0xf0,0xfa,0xc0,0x1b]
 176
 177 v_dot4_u32_u8 v5, src_scc, vcc_lo, -1
 178 // GFX12: [0x05,0x40,0x17,0xcc,0xfd,0xd4,0x04,0x1b]
 179
 180 v_dot4_u32_u8 v255, 0xaf123456, vcc_hi, null
 181 // GFX12: [0xff,0x40,0x17,0xcc,0xff,0xd6,0xf0,0x19,0x56,0x34,0x12,0xaf]
 182
 183 v_dot8_i32_iu4 v5, v1, v2, s3
 184 // GFX12: [0x05,0x40,0x18,0xcc,0x01,0x05,0x0e,0x18]
 185
 186 v_dot8_i32_iu4 v5, v255, s2, s105
 187 // GFX12: [0x05,0x40,0x18,0xcc,0xff,0x05,0xa4,0x19]
 188
 189 v_dot8_i32_iu4 v5, s1, v255, exec_hi
 190 // GFX12: [0x05,0x40,0x18,0xcc,0x01,0xfe,0xff,0x19]
 191
 192 v_dot8_i32_iu4 v5, s105, s105, exec_lo
 193 // GFX12: [0x05,0x40,0x18,0xcc,0x69,0xd2,0xf8,0x19]
 194
 195 v_dot8_i32_iu4 v5, vcc_lo, ttmp15, v3
 196 // GFX12: [0x05,0x40,0x18,0xcc,0x6a,0xf6,0x0c,0x1c]
 197
 198 v_dot8_i32_iu4 v5, vcc_hi, 0xaf123456, v255
 199 // GFX12: [0x05,0x40,0x18,0xcc,0x6b,0xfe,0xfd,0x1f,0x56,0x34,0x12,0xaf]
 200
 201 v_dot8_i32_iu4 v5, ttmp15, src_scc, ttmp15
 202 // GFX12: [0x05,0x40,0x18,0xcc,0x7b,0xfa,0xed,0x19]
 203
 204 v_dot8_i32_iu4 v5, m0, 0.5, m0
 205 // GFX12: [0x05,0x40,0x18,0xcc,0x7d,0xe0,0xf5,0x19]
 206
 207 v_dot8_i32_iu4 v5, exec_lo, -1, vcc_hi
 208 // GFX12: [0x05,0x40,0x18,0xcc,0x7e,0x82,0xad,0x19]
 209
 210 v_dot8_i32_iu4 v5, exec_hi, null, vcc_lo
 211 // GFX12: [0x05,0x40,0x18,0xcc,0x7f,0xf8,0xa8,0x19]
 212
 213 v_dot8_i32_iu4 v5, null, exec_lo, 0xaf123456
 214 // GFX12: [0x05,0x40,0x18,0xcc,0x7c,0xfc,0xfc,0x1b,0x56,0x34,0x12,0xaf]
 215
 216 v_dot8_i32_iu4 v5, -1, exec_hi, src_scc
 217 // GFX12: [0x05,0x40,0x18,0xcc,0xc1,0xfe,0xf4,0x1b]
 218
 219 v_dot8_i32_iu4 v5, 0.5, m0, 0.5 neg_lo:[1,0,0]
 220 // GFX12: [0x05,0x40,0x18,0xcc,0xf0,0xfa,0xc0,0x3b]
 221
 222 v_dot8_i32_iu4 v5, src_scc, vcc_lo, -1 neg_lo:[0,1,0]
 223 // GFX12: [0x05,0x40,0x18,0xcc,0xfd,0xd4,0x04,0x5b]
 224
 225 v_dot8_i32_iu4 v255, 0xaf123456, vcc_hi, null neg_lo:[0,0,0] clamp
 226 // GFX12: [0xff,0xc0,0x18,0xcc,0xff,0xd6,0xf0,0x19,0x56,0x34,0x12,0xaf]
 227
 228 v_dot8_u32_u4 v5, v1, v2, s3
 229 // GFX12: [0x05,0x40,0x19,0xcc,0x01,0x05,0x0e,0x18]
 230
 231 v_dot8_u32_u4 v5, v255, s2, s105
 232 // GFX12: [0x05,0x40,0x19,0xcc,0xff,0x05,0xa4,0x19]
 233
 234 v_dot8_u32_u4 v5, s1, v255, exec_hi
 235 // GFX12: [0x05,0x40,0x19,0xcc,0x01,0xfe,0xff,0x19]
 236
 237 v_dot8_u32_u4 v5, s105, s105, exec_lo
 238 // GFX12: [0x05,0x40,0x19,0xcc,0x69,0xd2,0xf8,0x19]
 239
 240 v_dot8_u32_u4 v5, vcc_lo, ttmp15, v3
 241 // GFX12: [0x05,0x40,0x19,0xcc,0x6a,0xf6,0x0c,0x1c]
 242
 243 v_dot8_u32_u4 v5, vcc_hi, 0xaf123456, v255
 244 // GFX12: [0x05,0x40,0x19,0xcc,0x6b,0xfe,0xfd,0x1f,0x56,0x34,0x12,0xaf]
 245
 246 v_dot8_u32_u4 v5, ttmp15, src_scc, ttmp15
 247 // GFX12: [0x05,0x40,0x19,0xcc,0x7b,0xfa,0xed,0x19]
 248
 249 v_dot8_u32_u4 v5, m0, 0.5, m0
 250 // GFX12: [0x05,0x40,0x19,0xcc,0x7d,0xe0,0xf5,0x19]
 251
 252 v_dot8_u32_u4 v5, exec_lo, -1, vcc_hi
 253 // GFX12: [0x05,0x40,0x19,0xcc,0x7e,0x82,0xad,0x19]
 254
 255 v_dot8_u32_u4 v5, exec_hi, null, vcc_lo
 256 // GFX12: [0x05,0x40,0x19,0xcc,0x7f,0xf8,0xa8,0x19]
 257
 258 v_dot8_u32_u4 v5, null, exec_lo, 0xaf123456
 259 // GFX12: [0x05,0x40,0x19,0xcc,0x7c,0xfc,0xfc,0x1b,0x56,0x34,0x12,0xaf]
 260
 261 v_dot8_u32_u4 v5, -1, exec_hi, src_scc
 262 // GFX12: [0x05,0x40,0x19,0xcc,0xc1,0xfe,0xf4,0x1b]
 263
 264 v_dot8_u32_u4 v5, 0.5, m0, 0.5
 265 // GFX12: [0x05,0x40,0x19,0xcc,0xf0,0xfa,0xc0,0x1b]
 266
 267 v_dot8_u32_u4 v5, src_scc, vcc_lo, -1
 268 // GFX12: [0x05,0x40,0x19,0xcc,0xfd,0xd4,0x04,0x1b]
 269
 270 v_dot8_u32_u4 v255, 0xaf123456, vcc_hi, null clamp
 271 // GFX12: [0xff,0xc0,0x19,0xcc,0xff,0xd6,0xf0,0x19,0x56,0x34,0x12,0xaf]
 272
 273 v_fma_mix_f32 v5, v1, v2, s3
 274 // GFX12: [0x05,0x00,0x20,0xcc,0x01,0x05,0x0e,0x00]
 275
 276 v_fma_mix_f32 v5, v255, v255, s105
 277 // GFX12: [0x05,0x00,0x20,0xcc,0xff,0xff,0xa7,0x01]
 278
 279 v_fma_mix_f32 v5, s1, s2, v3
 280 // GFX12: [0x05,0x00,0x20,0xcc,0x01,0x04,0x0c,0x04]
 281
 282 v_fma_mix_f32 v5, s105, s105, m0
 283 // GFX12: [0x05,0x00,0x20,0xcc,0x69,0xd2,0xf4,0x01]
 284
 285 v_fma_mix_f32 v5, vcc_lo, ttmp15, ttmp15
 286 // GFX12: [0x05,0x00,0x20,0xcc,0x6a,0xf6,0xec,0x01]
 287
 288 v_fma_mix_f32 v5, vcc_hi, src_scc, v255
 289 // GFX12: [0x05,0x00,0x20,0xcc,0x6b,0xfa,0xfd,0x07]
 290
 291 v_fma_mix_f32 v5, |ttmp15|, 0.5, -vcc_hi
 292 // GFX12: [0x05,0x01,0x20,0xcc,0x7b,0xe0,0xad,0x81]
 293
 294 v_fma_mix_f32 v5, -m0, -1, |vcc_lo|
 295 // GFX12: [0x05,0x04,0x20,0xcc,0x7d,0x82,0xa9,0x21]
 296
 297 v_fma_mix_f32 v5, -|exec_lo|, null, -|src_scc|
 298 // GFX12: [0x05,0x05,0x20,0xcc,0x7e,0xf8,0xf4,0xa3]
 299
 300 v_fma_mix_f32 v5, -|exec_hi|, -|exec_lo|, -|exec_lo| op_sel:[1,1,1] op_sel_hi:[1,1,1]
 301 // GFX12: [0x05,0x7f,0x20,0xcc,0x7f,0xfc,0xf8,0xf9]
 302
 303 v_fma_mix_f32 v5, null, exec_hi, 0.5 op_sel:[0,0,0] op_sel_hi:[0,0,1]
 304 // GFX12: [0x05,0x40,0x20,0xcc,0x7c,0xfe,0xc0,0x03]
 305
 306 v_fma_mix_f32 v5, -1, -|m0|, -1 op_sel:[1,0,0] op_sel_hi:[0,1,0]
 307 // GFX12: [0x05,0x0a,0x20,0xcc,0xc1,0xfa,0x04,0x53]
 308
 309 v_fma_mix_f32 v5, 0.5, -|vcc_lo|, -|exec_hi| op_sel:[0,1,0] op_sel_hi:[1,0,0]
 310 // GFX12: [0x05,0x16,0x20,0xcc,0xf0,0xd4,0xfc,0xc9]
 311
 312 v_fma_mix_f32 v255, -|src_scc|, -|vcc_hi|, null op_sel:[0,0,1] op_sel_hi:[0,0,0] clamp
 313 // GFX12: [0xff,0xa3,0x20,0xcc,0xfd,0xd6,0xf0,0x61]
 314
 315 v_fma_mixhi_f16 v5, v1, v2, s3
 316 // GFX12: [0x05,0x00,0x22,0xcc,0x01,0x05,0x0e,0x00]
 317
 318 v_fma_mixhi_f16 v5, v255, v255, s105
 319 // GFX12: [0x05,0x00,0x22,0xcc,0xff,0xff,0xa7,0x01]
 320
 321 v_fma_mixhi_f16 v5, s1, s2, v3
 322 // GFX12: [0x05,0x00,0x22,0xcc,0x01,0x04,0x0c,0x04]
 323
 324 v_fma_mixhi_f16 v5, s105, s105, m0
 325 // GFX12: [0x05,0x00,0x22,0xcc,0x69,0xd2,0xf4,0x01]
 326
 327 v_fma_mixhi_f16 v5, vcc_lo, ttmp15, ttmp15
 328 // GFX12: [0x05,0x00,0x22,0xcc,0x6a,0xf6,0xec,0x01]
 329
 330 v_fma_mixhi_f16 v5, vcc_hi, src_scc, v255
 331 // GFX12: [0x05,0x00,0x22,0xcc,0x6b,0xfa,0xfd,0x07]
 332
 333 v_fma_mixhi_f16 v5, |ttmp15|, 0.5, -vcc_hi
 334 // GFX12: [0x05,0x01,0x22,0xcc,0x7b,0xe0,0xad,0x81]
 335
 336 v_fma_mixhi_f16 v5, -m0, -1, |vcc_lo|
 337 // GFX12: [0x05,0x04,0x22,0xcc,0x7d,0x82,0xa9,0x21]
 338
 339 v_fma_mixhi_f16 v5, -|exec_lo|, null, -|src_scc|
 340 // GFX12: [0x05,0x05,0x22,0xcc,0x7e,0xf8,0xf4,0xa3]
 341
 342 v_fma_mixhi_f16 v5, -|exec_hi|, -|exec_lo|, -|exec_lo| op_sel:[1,1,1] op_sel_hi:[1,1,1]
 343 // GFX12: [0x05,0x7f,0x22,0xcc,0x7f,0xfc,0xf8,0xf9]
 344
 345 v_fma_mixhi_f16 v5, null, exec_hi, 0.5 op_sel:[0,0,0] op_sel_hi:[0,0,1]
 346 // GFX12: [0x05,0x40,0x22,0xcc,0x7c,0xfe,0xc0,0x03]
 347
 348 v_fma_mixhi_f16 v5, -1, -|m0|, -1 op_sel:[1,0,0] op_sel_hi:[0,1,0]
 349 // GFX12: [0x05,0x0a,0x22,0xcc,0xc1,0xfa,0x04,0x53]
 350
 351 v_fma_mixhi_f16 v5, 0.5, -|vcc_lo|, -|exec_hi| op_sel:[0,1,0] op_sel_hi:[1,0,0]
 352 // GFX12: [0x05,0x16,0x22,0xcc,0xf0,0xd4,0xfc,0xc9]
 353
 354 v_fma_mixhi_f16 v255, -|src_scc|, -|vcc_hi|, null op_sel:[0,0,1] op_sel_hi:[0,0,0] clamp
 355 // GFX12: [0xff,0xa3,0x22,0xcc,0xfd,0xd6,0xf0,0x61]
 356
 357 v_fma_mixlo_f16 v5, v1, v2, s3
 358 // GFX12: [0x05,0x00,0x21,0xcc,0x01,0x05,0x0e,0x00]
 359
 360 v_fma_mixlo_f16 v5, v255, v255, s105
 361 // GFX12: [0x05,0x00,0x21,0xcc,0xff,0xff,0xa7,0x01]
 362
 363 v_fma_mixlo_f16 v5, s1, s2, v3
 364 // GFX12: [0x05,0x00,0x21,0xcc,0x01,0x04,0x0c,0x04]
 365
 366 v_fma_mixlo_f16 v5, s105, s105, m0
 367 // GFX12: [0x05,0x00,0x21,0xcc,0x69,0xd2,0xf4,0x01]
 368
 369 v_fma_mixlo_f16 v5, vcc_lo, ttmp15, ttmp15
 370 // GFX12: [0x05,0x00,0x21,0xcc,0x6a,0xf6,0xec,0x01]
 371
 372 v_fma_mixlo_f16 v5, vcc_hi, src_scc, v255
 373 // GFX12: [0x05,0x00,0x21,0xcc,0x6b,0xfa,0xfd,0x07]
 374
 375 v_fma_mixlo_f16 v5, |ttmp15|, 0.5, -vcc_hi
 376 // GFX12: [0x05,0x01,0x21,0xcc,0x7b,0xe0,0xad,0x81]
 377
 378 v_fma_mixlo_f16 v5, -m0, -1, |vcc_lo|
 379 // GFX12: [0x05,0x04,0x21,0xcc,0x7d,0x82,0xa9,0x21]
 380
 381 v_fma_mixlo_f16 v5, -|exec_lo|, null, -|src_scc|
 382 // GFX12: [0x05,0x05,0x21,0xcc,0x7e,0xf8,0xf4,0xa3]
 383
 384 v_fma_mixlo_f16 v5, -|exec_hi|, -|exec_lo|, -|exec_lo| op_sel:[1,1,1] op_sel_hi:[1,1,1]
 385 // GFX12: [0x05,0x7f,0x21,0xcc,0x7f,0xfc,0xf8,0xf9]
 386
 387 v_fma_mixlo_f16 v5, null, exec_hi, 0.5 op_sel:[0,0,0] op_sel_hi:[0,0,1]
 388 // GFX12: [0x05,0x40,0x21,0xcc,0x7c,0xfe,0xc0,0x03]
 389
 390 v_fma_mixlo_f16 v5, -1, -|m0|, -1 op_sel:[1,0,0] op_sel_hi:[0,1,0]
 391 // GFX12: [0x05,0x0a,0x21,0xcc,0xc1,0xfa,0x04,0x53]
 392
 393 v_fma_mixlo_f16 v5, 0.5, -|vcc_lo|, -|exec_hi| op_sel:[0,1,0] op_sel_hi:[1,0,0]
 394 // GFX12: [0x05,0x16,0x21,0xcc,0xf0,0xd4,0xfc,0xc9]
 395
 396 v_fma_mixlo_f16 v255, -|src_scc|, -|vcc_hi|, null op_sel:[0,0,1] op_sel_hi:[0,0,0] clamp
 397 // GFX12: [0xff,0xa3,0x21,0xcc,0xfd,0xd6,0xf0,0x61]
 398
 399 v_pk_add_f16 v5, v1, v2
 400 // GFX12: [0x05,0x40,0x0f,0xcc,0x01,0x05,0x02,0x18]
 401
 402 v_pk_add_f16 v5, v255, v255
 403 // GFX12: [0x05,0x40,0x0f,0xcc,0xff,0xff,0x03,0x18]
 404
 405 v_pk_add_f16 v5, s1, s2
 406 // GFX12: [0x05,0x40,0x0f,0xcc,0x01,0x04,0x00,0x18]
 407
 408 v_pk_add_f16 v5, s105, s105
 409 // GFX12: [0x05,0x40,0x0f,0xcc,0x69,0xd2,0x00,0x18]
 410
 411 v_pk_add_f16 v5, vcc_lo, ttmp15
 412 // GFX12: [0x05,0x40,0x0f,0xcc,0x6a,0xf6,0x00,0x18]
 413
 414 v_pk_add_f16 v5, vcc_hi, 0xfe0b
 415 // GFX12: [0x05,0x40,0x0f,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 416
 417 v_pk_add_f16 v5, ttmp15, src_scc
 418 // GFX12: [0x05,0x40,0x0f,0xcc,0x7b,0xfa,0x01,0x18]
 419
 420 v_pk_add_f16 v5, m0, 0.5
 421 // GFX12: [0x05,0x40,0x0f,0xcc,0x7d,0xe0,0x01,0x18]
 422
 423 v_pk_add_f16 v5, exec_lo, -1
 424 // GFX12: [0x05,0x40,0x0f,0xcc,0x7e,0x82,0x01,0x18]
 425
 426 v_pk_add_f16 v5, exec_hi, null
 427 // GFX12: [0x05,0x40,0x0f,0xcc,0x7f,0xf8,0x00,0x18]
 428
 429 v_pk_add_f16 v5, null, exec_lo
 430 // GFX12: [0x05,0x40,0x0f,0xcc,0x7c,0xfc,0x00,0x18]
 431
 432 v_pk_add_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
 433 // GFX12: [0x05,0x59,0x0f,0xcc,0xc1,0xfe,0x00,0x20]
 434
 435 v_pk_add_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
 436 // GFX12: [0x05,0x42,0x0f,0xcc,0xf0,0xfa,0x00,0x58]
 437
 438 v_pk_add_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
 439 // GFX12: [0x05,0x48,0x0f,0xcc,0xfd,0xd4,0x00,0x10]
 440
 441 v_pk_add_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
 442 // GFX12: [0xff,0xd3,0x0f,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
 443
 444 v_pk_add_i16 v5, v1, v2
 445 // GFX12: [0x05,0x40,0x02,0xcc,0x01,0x05,0x02,0x18]
 446
 447 v_pk_add_i16 v5, v255, v255
 448 // GFX12: [0x05,0x40,0x02,0xcc,0xff,0xff,0x03,0x18]
 449
 450 v_pk_add_i16 v5, s1, s2
 451 // GFX12: [0x05,0x40,0x02,0xcc,0x01,0x04,0x00,0x18]
 452
 453 v_pk_add_i16 v5, s105, s105
 454 // GFX12: [0x05,0x40,0x02,0xcc,0x69,0xd2,0x00,0x18]
 455
 456 v_pk_add_i16 v5, vcc_lo, ttmp15
 457 // GFX12: [0x05,0x40,0x02,0xcc,0x6a,0xf6,0x00,0x18]
 458
 459 v_pk_add_i16 v5, vcc_hi, 0xfe0b
 460 // GFX12: [0x05,0x40,0x02,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 461
 462 v_pk_add_i16 v5, ttmp15, src_scc
 463 // GFX12: [0x05,0x40,0x02,0xcc,0x7b,0xfa,0x01,0x18]
 464
 465 v_pk_add_i16 v5, m0, 0.5
 466 // GFX12: [0x05,0x40,0x02,0xcc,0x7d,0xe0,0x01,0x18]
 467
 468 v_pk_add_i16 v5, exec_lo, -1
 469 // GFX12: [0x05,0x40,0x02,0xcc,0x7e,0x82,0x01,0x18]
 470
 471 v_pk_add_i16 v5, exec_hi, null
 472 // GFX12: [0x05,0x40,0x02,0xcc,0x7f,0xf8,0x00,0x18]
 473
 474 v_pk_add_i16 v5, null, exec_lo
 475 // GFX12: [0x05,0x40,0x02,0xcc,0x7c,0xfc,0x00,0x18]
 476
 477 v_pk_add_i16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 478 // GFX12: [0x05,0x58,0x02,0xcc,0xc1,0xfe,0x00,0x00]
 479
 480 v_pk_add_i16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 481 // GFX12: [0x05,0x40,0x02,0xcc,0xf0,0xfa,0x00,0x18]
 482
 483 v_pk_add_i16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 484 // GFX12: [0x05,0x48,0x02,0xcc,0xfd,0xd4,0x00,0x10]
 485
 486 v_pk_add_i16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] clamp
 487 // GFX12: [0xff,0xd0,0x02,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 488
 489 v_pk_add_u16 v5, v1, v2
 490 // GFX12: [0x05,0x40,0x0a,0xcc,0x01,0x05,0x02,0x18]
 491
 492 v_pk_add_u16 v5, v255, v255
 493 // GFX12: [0x05,0x40,0x0a,0xcc,0xff,0xff,0x03,0x18]
 494
 495 v_pk_add_u16 v5, s1, s2
 496 // GFX12: [0x05,0x40,0x0a,0xcc,0x01,0x04,0x00,0x18]
 497
 498 v_pk_add_u16 v5, s105, s105
 499 // GFX12: [0x05,0x40,0x0a,0xcc,0x69,0xd2,0x00,0x18]
 500
 501 v_pk_add_u16 v5, vcc_lo, ttmp15
 502 // GFX12: [0x05,0x40,0x0a,0xcc,0x6a,0xf6,0x00,0x18]
 503
 504 v_pk_add_u16 v5, vcc_hi, 0xfe0b
 505 // GFX12: [0x05,0x40,0x0a,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 506
 507 v_pk_add_u16 v5, ttmp15, src_scc
 508 // GFX12: [0x05,0x40,0x0a,0xcc,0x7b,0xfa,0x01,0x18]
 509
 510 v_pk_add_u16 v5, m0, 0.5
 511 // GFX12: [0x05,0x40,0x0a,0xcc,0x7d,0xe0,0x01,0x18]
 512
 513 v_pk_add_u16 v5, exec_lo, -1
 514 // GFX12: [0x05,0x40,0x0a,0xcc,0x7e,0x82,0x01,0x18]
 515
 516 v_pk_add_u16 v5, exec_hi, null
 517 // GFX12: [0x05,0x40,0x0a,0xcc,0x7f,0xf8,0x00,0x18]
 518
 519 v_pk_add_u16 v5, null, exec_lo
 520 // GFX12: [0x05,0x40,0x0a,0xcc,0x7c,0xfc,0x00,0x18]
 521
 522 v_pk_add_u16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 523 // GFX12: [0x05,0x58,0x0a,0xcc,0xc1,0xfe,0x00,0x00]
 524
 525 v_pk_add_u16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 526 // GFX12: [0x05,0x40,0x0a,0xcc,0xf0,0xfa,0x00,0x18]
 527
 528 v_pk_add_u16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 529 // GFX12: [0x05,0x48,0x0a,0xcc,0xfd,0xd4,0x00,0x10]
 530
 531 v_pk_add_u16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] clamp
 532 // GFX12: [0xff,0xd0,0x0a,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 533
 534 v_pk_ashrrev_i16 v5, v1, v2
 535 // GFX12: [0x05,0x40,0x06,0xcc,0x01,0x05,0x02,0x18]
 536
 537 v_pk_ashrrev_i16 v5, v255, v255
 538 // GFX12: [0x05,0x40,0x06,0xcc,0xff,0xff,0x03,0x18]
 539
 540 v_pk_ashrrev_i16 v5, s1, s2
 541 // GFX12: [0x05,0x40,0x06,0xcc,0x01,0x04,0x00,0x18]
 542
 543 v_pk_ashrrev_i16 v5, s105, s105
 544 // GFX12: [0x05,0x40,0x06,0xcc,0x69,0xd2,0x00,0x18]
 545
 546 v_pk_ashrrev_i16 v5, vcc_lo, ttmp15
 547 // GFX12: [0x05,0x40,0x06,0xcc,0x6a,0xf6,0x00,0x18]
 548
 549 v_pk_ashrrev_i16 v5, vcc_hi, 0xfe0b
 550 // GFX12: [0x05,0x40,0x06,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 551
 552 v_pk_ashrrev_i16 v5, ttmp15, src_scc
 553 // GFX12: [0x05,0x40,0x06,0xcc,0x7b,0xfa,0x01,0x18]
 554
 555 v_pk_ashrrev_i16 v5, m0, 0.5
 556 // GFX12: [0x05,0x40,0x06,0xcc,0x7d,0xe0,0x01,0x18]
 557
 558 v_pk_ashrrev_i16 v5, exec_lo, -1
 559 // GFX12: [0x05,0x40,0x06,0xcc,0x7e,0x82,0x01,0x18]
 560
 561 v_pk_ashrrev_i16 v5, exec_hi, null
 562 // GFX12: [0x05,0x40,0x06,0xcc,0x7f,0xf8,0x00,0x18]
 563
 564 v_pk_ashrrev_i16 v5, null, exec_lo
 565 // GFX12: [0x05,0x40,0x06,0xcc,0x7c,0xfc,0x00,0x18]
 566
 567 v_pk_ashrrev_i16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 568 // GFX12: [0x05,0x58,0x06,0xcc,0xc1,0xfe,0x00,0x00]
 569
 570 v_pk_ashrrev_i16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 571 // GFX12: [0x05,0x40,0x06,0xcc,0xf0,0xfa,0x00,0x18]
 572
 573 v_pk_ashrrev_i16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 574 // GFX12: [0x05,0x48,0x06,0xcc,0xfd,0xd4,0x00,0x10]
 575
 576 v_pk_ashrrev_i16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
 577 // GFX12: [0xff,0x50,0x06,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 578
 579 v_pk_fma_f16 v5, v1, v2, s3
 580 // GFX12: [0x05,0x40,0x0e,0xcc,0x01,0x05,0x0e,0x18]
 581
 582 v_pk_fma_f16 v5, v255, s2, s105
 583 // GFX12: [0x05,0x40,0x0e,0xcc,0xff,0x05,0xa4,0x19]
 584
 585 v_pk_fma_f16 v5, s1, v255, exec_hi
 586 // GFX12: [0x05,0x40,0x0e,0xcc,0x01,0xfe,0xff,0x19]
 587
 588 v_pk_fma_f16 v5, s105, s105, exec_lo
 589 // GFX12: [0x05,0x40,0x0e,0xcc,0x69,0xd2,0xf8,0x19]
 590
 591 v_pk_fma_f16 v5, vcc_lo, ttmp15, v3
 592 // GFX12: [0x05,0x40,0x0e,0xcc,0x6a,0xf6,0x0c,0x1c]
 593
 594 v_pk_fma_f16 v5, vcc_hi, 0xfe0b, v255
 595 // GFX12: [0x05,0x40,0x0e,0xcc,0x6b,0xfe,0xfd,0x1f,0x0b,0xfe,0x00,0x00]
 596
 597 v_pk_fma_f16 v5, ttmp15, src_scc, ttmp15
 598 // GFX12: [0x05,0x40,0x0e,0xcc,0x7b,0xfa,0xed,0x19]
 599
 600 v_pk_fma_f16 v5, m0, 0.5, m0 op_sel_hi:[0,0,0]
 601 // GFX12: [0x05,0x00,0x0e,0xcc,0x7d,0xe0,0xf5,0x01]
 602
 603 v_pk_fma_f16 v5, exec_lo, -1, vcc_hi op_sel_hi:[0,0,1]
 604 // GFX12: [0x05,0x40,0x0e,0xcc,0x7e,0x82,0xad,0x01]
 605
 606 v_pk_fma_f16 v5, exec_hi, null, vcc_lo op_sel_hi:[0,1,0]
 607 // GFX12: [0x05,0x00,0x0e,0xcc,0x7f,0xf8,0xa8,0x11]
 608
 609 v_pk_fma_f16 v5, null, exec_lo, 0xfe0b op_sel:[1,1,1] op_sel_hi:[1,0,0] neg_lo:[1,0,0] neg_hi:[1,0,0]
 610 // GFX12: [0x05,0x39,0x0e,0xcc,0x7c,0xfc,0xfc,0x2b,0x0b,0xfe,0x00,0x00]
 611
 612 v_pk_fma_f16 v5, -1, exec_hi, src_scc op_sel:[0,0,0] op_sel_hi:[1,1,1] neg_lo:[0,1,0] neg_hi:[0,1,0]
 613 // GFX12: [0x05,0x42,0x0e,0xcc,0xc1,0xfe,0xf4,0x5b]
 614
 615 v_pk_fma_f16 v5, 0.5, m0, 0.5 op_sel:[1,0,0] op_sel_hi:[0,1,1] neg_lo:[0,0,1] neg_hi:[0,0,1]
 616 // GFX12: [0x05,0x4c,0x0e,0xcc,0xf0,0xfa,0xc0,0x93]
 617
 618 v_pk_fma_f16 v5, src_scc, vcc_lo, -1 op_sel:[0,1,0] op_sel_hi:[1,0,1] neg_lo:[0,0,0] neg_hi:[0,0,0]
 619 // GFX12: [0x05,0x50,0x0e,0xcc,0xfd,0xd4,0x04,0x0b]
 620
 621 v_pk_fma_f16 v255, 0xfe0b, vcc_hi, null op_sel:[0,0,1] op_sel_hi:[1,1,0] neg_lo:[1,1,1] neg_hi:[1,1,1] clamp
 622 // GFX12: [0xff,0xa7,0x0e,0xcc,0xff,0xd6,0xf0,0xf9,0x0b,0xfe,0x00,0x00]
 623
 624 v_pk_lshlrev_b16 v5, v1, v2
 625 // GFX12: [0x05,0x40,0x04,0xcc,0x01,0x05,0x02,0x18]
 626
 627 v_pk_lshlrev_b16 v5, v255, v255
 628 // GFX12: [0x05,0x40,0x04,0xcc,0xff,0xff,0x03,0x18]
 629
 630 v_pk_lshlrev_b16 v5, s1, s2
 631 // GFX12: [0x05,0x40,0x04,0xcc,0x01,0x04,0x00,0x18]
 632
 633 v_pk_lshlrev_b16 v5, s105, s105
 634 // GFX12: [0x05,0x40,0x04,0xcc,0x69,0xd2,0x00,0x18]
 635
 636 v_pk_lshlrev_b16 v5, vcc_lo, ttmp15
 637 // GFX12: [0x05,0x40,0x04,0xcc,0x6a,0xf6,0x00,0x18]
 638
 639 v_pk_lshlrev_b16 v5, vcc_hi, 0xfe0b
 640 // GFX12: [0x05,0x40,0x04,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 641
 642 v_pk_lshlrev_b16 v5, ttmp15, src_scc
 643 // GFX12: [0x05,0x40,0x04,0xcc,0x7b,0xfa,0x01,0x18]
 644
 645 v_pk_lshlrev_b16 v5, m0, 0.5
 646 // GFX12: [0x05,0x40,0x04,0xcc,0x7d,0xe0,0x01,0x18]
 647
 648 v_pk_lshlrev_b16 v5, exec_lo, -1
 649 // GFX12: [0x05,0x40,0x04,0xcc,0x7e,0x82,0x01,0x18]
 650
 651 v_pk_lshlrev_b16 v5, exec_hi, null
 652 // GFX12: [0x05,0x40,0x04,0xcc,0x7f,0xf8,0x00,0x18]
 653
 654 v_pk_lshlrev_b16 v5, null, exec_lo
 655 // GFX12: [0x05,0x40,0x04,0xcc,0x7c,0xfc,0x00,0x18]
 656
 657 v_pk_lshlrev_b16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 658 // GFX12: [0x05,0x58,0x04,0xcc,0xc1,0xfe,0x00,0x00]
 659
 660 v_pk_lshlrev_b16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 661 // GFX12: [0x05,0x40,0x04,0xcc,0xf0,0xfa,0x00,0x18]
 662
 663 v_pk_lshlrev_b16 v5, 0x3800, m0 op_sel:[0,0] op_sel_hi:[1,1]
 664 // GFX12: [0x05,0x40,0x04,0xcc,0xff,0xfa,0x00,0x18,0x00,0x38,0x00,0x00]
 665
 666 v_pk_lshlrev_b16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 667 // GFX12: [0x05,0x48,0x04,0xcc,0xfd,0xd4,0x00,0x10]
 668
 669 v_pk_lshlrev_b16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
 670 // GFX12: [0xff,0x50,0x04,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 671
 672 v_pk_lshrrev_b16 v5, v1, v2
 673 // GFX12: [0x05,0x40,0x05,0xcc,0x01,0x05,0x02,0x18]
 674
 675 v_pk_lshrrev_b16 v5, v255, v255
 676 // GFX12: [0x05,0x40,0x05,0xcc,0xff,0xff,0x03,0x18]
 677
 678 v_pk_lshrrev_b16 v5, s1, s2
 679 // GFX12: [0x05,0x40,0x05,0xcc,0x01,0x04,0x00,0x18]
 680
 681 v_pk_lshrrev_b16 v5, s105, s105
 682 // GFX12: [0x05,0x40,0x05,0xcc,0x69,0xd2,0x00,0x18]
 683
 684 v_pk_lshrrev_b16 v5, vcc_lo, ttmp15
 685 // GFX12: [0x05,0x40,0x05,0xcc,0x6a,0xf6,0x00,0x18]
 686
 687 v_pk_lshrrev_b16 v5, vcc_hi, 0xfe0b
 688 // GFX12: [0x05,0x40,0x05,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 689
 690 v_pk_lshrrev_b16 v5, ttmp15, src_scc
 691 // GFX12: [0x05,0x40,0x05,0xcc,0x7b,0xfa,0x01,0x18]
 692
 693 v_pk_lshrrev_b16 v5, m0, 0.5
 694 // GFX12: [0x05,0x40,0x05,0xcc,0x7d,0xe0,0x01,0x18]
 695
 696 v_pk_lshrrev_b16 v5, exec_lo, -1
 697 // GFX12: [0x05,0x40,0x05,0xcc,0x7e,0x82,0x01,0x18]
 698
 699 v_pk_lshrrev_b16 v5, exec_hi, null
 700 // GFX12: [0x05,0x40,0x05,0xcc,0x7f,0xf8,0x00,0x18]
 701
 702 v_pk_lshrrev_b16 v5, null, exec_lo
 703 // GFX12: [0x05,0x40,0x05,0xcc,0x7c,0xfc,0x00,0x18]
 704
 705 v_pk_lshrrev_b16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 706 // GFX12: [0x05,0x58,0x05,0xcc,0xc1,0xfe,0x00,0x00]
 707
 708 v_pk_lshrrev_b16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 709 // GFX12: [0x05,0x40,0x05,0xcc,0xf0,0xfa,0x00,0x18]
 710
 711 v_pk_lshrrev_b16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 712 // GFX12: [0x05,0x48,0x05,0xcc,0xfd,0xd4,0x00,0x10]
 713
 714 v_pk_lshrrev_b16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
 715 // GFX12: [0xff,0x50,0x05,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 716
 717 v_pk_mad_i16 v5, v1, v2, s3
 718 // GFX12: [0x05,0x40,0x00,0xcc,0x01,0x05,0x0e,0x18]
 719
 720 v_pk_mad_i16 v5, v255, s2, s105
 721 // GFX12: [0x05,0x40,0x00,0xcc,0xff,0x05,0xa4,0x19]
 722
 723 v_pk_mad_i16 v5, s1, v255, exec_hi
 724 // GFX12: [0x05,0x40,0x00,0xcc,0x01,0xfe,0xff,0x19]
 725
 726 v_pk_mad_i16 v5, s105, s105, exec_lo
 727 // GFX12: [0x05,0x40,0x00,0xcc,0x69,0xd2,0xf8,0x19]
 728
 729 v_pk_mad_i16 v5, vcc_lo, ttmp15, v3
 730 // GFX12: [0x05,0x40,0x00,0xcc,0x6a,0xf6,0x0c,0x1c]
 731
 732 v_pk_mad_i16 v5, vcc_hi, 0xfe0b, v255
 733 // GFX12: [0x05,0x40,0x00,0xcc,0x6b,0xfe,0xfd,0x1f,0x0b,0xfe,0x00,0x00]
 734
 735 v_pk_mad_i16 v5, ttmp15, src_scc, ttmp15
 736 // GFX12: [0x05,0x40,0x00,0xcc,0x7b,0xfa,0xed,0x19]
 737
 738 v_pk_mad_i16 v5, m0, 0.5, m0 op_sel_hi:[0,0,0]
 739 // GFX12: [0x05,0x00,0x00,0xcc,0x7d,0xe0,0xf5,0x01]
 740
 741 v_pk_mad_i16 v5, exec_lo, -1, vcc_hi op_sel_hi:[0,0,1]
 742 // GFX12: [0x05,0x40,0x00,0xcc,0x7e,0x82,0xad,0x01]
 743
 744 v_pk_mad_i16 v5, exec_hi, null, vcc_lo op_sel_hi:[0,1,0]
 745 // GFX12: [0x05,0x00,0x00,0xcc,0x7f,0xf8,0xa8,0x11]
 746
 747 v_pk_mad_i16 v5, null, exec_lo, 0xfe0b op_sel:[1,1,1] op_sel_hi:[1,0,0]
 748 // GFX12: [0x05,0x38,0x00,0xcc,0x7c,0xfc,0xfc,0x0b,0x0b,0xfe,0x00,0x00]
 749
 750 v_pk_mad_i16 v5, -1, exec_hi, src_scc op_sel:[0,0,0] op_sel_hi:[1,1,1]
 751 // GFX12: [0x05,0x40,0x00,0xcc,0xc1,0xfe,0xf4,0x1b]
 752
 753 v_pk_mad_i16 v5, 0.5, m0, 0.5 op_sel:[1,0,0] op_sel_hi:[0,1,1]
 754 // GFX12: [0x05,0x48,0x00,0xcc,0xf0,0xfa,0xc0,0x13]
 755
 756 v_pk_mad_i16 v5, src_scc, vcc_lo, -1 op_sel:[0,1,0] op_sel_hi:[1,0,1]
 757 // GFX12: [0x05,0x50,0x00,0xcc,0xfd,0xd4,0x04,0x0b]
 758
 759 v_pk_mad_i16 v255, 0xfe0b, vcc_hi, null op_sel:[0,0,1] op_sel_hi:[1,1,0] clamp
 760 // GFX12: [0xff,0xa0,0x00,0xcc,0xff,0xd6,0xf0,0x19,0x0b,0xfe,0x00,0x00]
 761
 762 v_pk_mad_u16 v5, v1, v2, s3
 763 // GFX12: [0x05,0x40,0x09,0xcc,0x01,0x05,0x0e,0x18]
 764
 765 v_pk_mad_u16 v5, v255, s2, s105
 766 // GFX12: [0x05,0x40,0x09,0xcc,0xff,0x05,0xa4,0x19]
 767
 768 v_pk_mad_u16 v5, s1, v255, exec_hi
 769 // GFX12: [0x05,0x40,0x09,0xcc,0x01,0xfe,0xff,0x19]
 770
 771 v_pk_mad_u16 v5, s105, s105, exec_lo
 772 // GFX12: [0x05,0x40,0x09,0xcc,0x69,0xd2,0xf8,0x19]
 773
 774 v_pk_mad_u16 v5, vcc_lo, ttmp15, v3
 775 // GFX12: [0x05,0x40,0x09,0xcc,0x6a,0xf6,0x0c,0x1c]
 776
 777 v_pk_mad_u16 v5, vcc_hi, 0xfe0b, v255
 778 // GFX12: [0x05,0x40,0x09,0xcc,0x6b,0xfe,0xfd,0x1f,0x0b,0xfe,0x00,0x00]
 779
 780 v_pk_mad_u16 v5, ttmp15, src_scc, ttmp15
 781 // GFX12: [0x05,0x40,0x09,0xcc,0x7b,0xfa,0xed,0x19]
 782
 783 v_pk_mad_u16 v5, m0, 0.5, m0 op_sel_hi:[0,0,0]
 784 // GFX12: [0x05,0x00,0x09,0xcc,0x7d,0xe0,0xf5,0x01]
 785
 786 v_pk_mad_u16 v5, exec_lo, -1, vcc_hi op_sel_hi:[0,0,1]
 787 // GFX12: [0x05,0x40,0x09,0xcc,0x7e,0x82,0xad,0x01]
 788
 789 v_pk_mad_u16 v5, exec_hi, null, vcc_lo op_sel_hi:[0,1,0]
 790 // GFX12: [0x05,0x00,0x09,0xcc,0x7f,0xf8,0xa8,0x11]
 791
 792 v_pk_mad_u16 v5, null, exec_lo, 0xfe0b op_sel:[1,1,1] op_sel_hi:[1,0,0]
 793 // GFX12: [0x05,0x38,0x09,0xcc,0x7c,0xfc,0xfc,0x0b,0x0b,0xfe,0x00,0x00]
 794
 795 v_pk_mad_u16 v5, -1, exec_hi, src_scc op_sel:[0,0,0] op_sel_hi:[1,1,1]
 796 // GFX12: [0x05,0x40,0x09,0xcc,0xc1,0xfe,0xf4,0x1b]
 797
 798 v_pk_mad_u16 v5, 0.5, m0, 0.5 op_sel:[1,0,0] op_sel_hi:[0,1,1]
 799 // GFX12: [0x05,0x48,0x09,0xcc,0xf0,0xfa,0xc0,0x13]
 800
 801 v_pk_mad_u16 v5, src_scc, vcc_lo, -1 op_sel:[0,1,0] op_sel_hi:[1,0,1]
 802 // GFX12: [0x05,0x50,0x09,0xcc,0xfd,0xd4,0x04,0x0b]
 803
 804 v_pk_mad_u16 v255, 0xfe0b, vcc_hi, null op_sel:[0,0,1] op_sel_hi:[1,1,0] clamp
 805 // GFX12: [0xff,0xa0,0x09,0xcc,0xff,0xd6,0xf0,0x19,0x0b,0xfe,0x00,0x00]
 806
 807 v_pk_max_num_f16 v5, v1, v2
 808 // GFX12: [0x05,0x40,0x1c,0xcc,0x01,0x05,0x02,0x18]
 809
 810 v_pk_max_num_f16 v5, v255, v255
 811 // GFX12: [0x05,0x40,0x1c,0xcc,0xff,0xff,0x03,0x18]
 812
 813 v_pk_max_num_f16 v5, s1, s2
 814 // GFX12: [0x05,0x40,0x1c,0xcc,0x01,0x04,0x00,0x18]
 815
 816 v_pk_max_num_f16 v5, s105, s105
 817 // GFX12: [0x05,0x40,0x1c,0xcc,0x69,0xd2,0x00,0x18]
 818
 819 v_pk_max_num_f16 v5, vcc_lo, ttmp15
 820 // GFX12: [0x05,0x40,0x1c,0xcc,0x6a,0xf6,0x00,0x18]
 821
 822 v_pk_max_num_f16 v5, vcc_hi, 0xfe0b
 823 // GFX12: [0x05,0x40,0x1c,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 824
 825 v_pk_max_num_f16 v5, ttmp15, src_scc
 826 // GFX12: [0x05,0x40,0x1c,0xcc,0x7b,0xfa,0x01,0x18]
 827
 828 v_pk_max_num_f16 v5, m0, 0.5
 829 // GFX12: [0x05,0x40,0x1c,0xcc,0x7d,0xe0,0x01,0x18]
 830
 831 v_pk_max_num_f16 v5, exec_lo, -1
 832 // GFX12: [0x05,0x40,0x1c,0xcc,0x7e,0x82,0x01,0x18]
 833
 834 v_pk_max_num_f16 v5, exec_hi, null
 835 // GFX12: [0x05,0x40,0x1c,0xcc,0x7f,0xf8,0x00,0x18]
 836
 837 v_pk_max_num_f16 v5, null, exec_lo
 838 // GFX12: [0x05,0x40,0x1c,0xcc,0x7c,0xfc,0x00,0x18]
 839
 840 v_pk_max_num_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
 841 // GFX12: [0x05,0x59,0x1c,0xcc,0xc1,0xfe,0x00,0x20]
 842
 843 v_pk_max_num_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
 844 // GFX12: [0x05,0x42,0x1c,0xcc,0xf0,0xfa,0x00,0x58]
 845
 846 v_pk_max_num_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
 847 // GFX12: [0x05,0x48,0x1c,0xcc,0xfd,0xd4,0x00,0x10]
 848
 849 v_pk_max_num_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
 850 // GFX12: [0xff,0xd3,0x1c,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
 851
 852 v_pk_max_i16 v5, v1, v2
 853 // GFX12: [0x05,0x40,0x07,0xcc,0x01,0x05,0x02,0x18]
 854
 855 v_pk_max_i16 v5, v255, v255
 856 // GFX12: [0x05,0x40,0x07,0xcc,0xff,0xff,0x03,0x18]
 857
 858 v_pk_max_i16 v5, s1, s2
 859 // GFX12: [0x05,0x40,0x07,0xcc,0x01,0x04,0x00,0x18]
 860
 861 v_pk_max_i16 v5, s105, s105
 862 // GFX12: [0x05,0x40,0x07,0xcc,0x69,0xd2,0x00,0x18]
 863
 864 v_pk_max_i16 v5, vcc_lo, ttmp15
 865 // GFX12: [0x05,0x40,0x07,0xcc,0x6a,0xf6,0x00,0x18]
 866
 867 v_pk_max_i16 v5, vcc_hi, 0xfe0b
 868 // GFX12: [0x05,0x40,0x07,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 869
 870 v_pk_max_i16 v5, ttmp15, src_scc
 871 // GFX12: [0x05,0x40,0x07,0xcc,0x7b,0xfa,0x01,0x18]
 872
 873 v_pk_max_i16 v5, m0, 0.5
 874 // GFX12: [0x05,0x40,0x07,0xcc,0x7d,0xe0,0x01,0x18]
 875
 876 v_pk_max_i16 v5, exec_lo, -1
 877 // GFX12: [0x05,0x40,0x07,0xcc,0x7e,0x82,0x01,0x18]
 878
 879 v_pk_max_i16 v5, exec_hi, null
 880 // GFX12: [0x05,0x40,0x07,0xcc,0x7f,0xf8,0x00,0x18]
 881
 882 v_pk_max_i16 v5, null, exec_lo
 883 // GFX12: [0x05,0x40,0x07,0xcc,0x7c,0xfc,0x00,0x18]
 884
 885 v_pk_max_i16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 886 // GFX12: [0x05,0x58,0x07,0xcc,0xc1,0xfe,0x00,0x00]
 887
 888 v_pk_max_i16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 889 // GFX12: [0x05,0x40,0x07,0xcc,0xf0,0xfa,0x00,0x18]
 890
 891 v_pk_max_i16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 892 // GFX12: [0x05,0x48,0x07,0xcc,0xfd,0xd4,0x00,0x10]
 893
 894 v_pk_max_i16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
 895 // GFX12: [0xff,0x50,0x07,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 896
 897 v_pk_max_u16 v5, v1, v2
 898 // GFX12: [0x05,0x40,0x0c,0xcc,0x01,0x05,0x02,0x18]
 899
 900 v_pk_max_u16 v5, v255, v255
 901 // GFX12: [0x05,0x40,0x0c,0xcc,0xff,0xff,0x03,0x18]
 902
 903 v_pk_max_u16 v5, s1, s2
 904 // GFX12: [0x05,0x40,0x0c,0xcc,0x01,0x04,0x00,0x18]
 905
 906 v_pk_max_u16 v5, s105, s105
 907 // GFX12: [0x05,0x40,0x0c,0xcc,0x69,0xd2,0x00,0x18]
 908
 909 v_pk_max_u16 v5, vcc_lo, ttmp15
 910 // GFX12: [0x05,0x40,0x0c,0xcc,0x6a,0xf6,0x00,0x18]
 911
 912 v_pk_max_u16 v5, vcc_hi, 0xfe0b
 913 // GFX12: [0x05,0x40,0x0c,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 914
 915 v_pk_max_u16 v5, ttmp15, src_scc
 916 // GFX12: [0x05,0x40,0x0c,0xcc,0x7b,0xfa,0x01,0x18]
 917
 918 v_pk_max_u16 v5, m0, 0.5
 919 // GFX12: [0x05,0x40,0x0c,0xcc,0x7d,0xe0,0x01,0x18]
 920
 921 v_pk_max_u16 v5, exec_lo, -1
 922 // GFX12: [0x05,0x40,0x0c,0xcc,0x7e,0x82,0x01,0x18]
 923
 924 v_pk_max_u16 v5, exec_hi, null
 925 // GFX12: [0x05,0x40,0x0c,0xcc,0x7f,0xf8,0x00,0x18]
 926
 927 v_pk_max_u16 v5, null, exec_lo
 928 // GFX12: [0x05,0x40,0x0c,0xcc,0x7c,0xfc,0x00,0x18]
 929
 930 v_pk_max_u16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
 931 // GFX12: [0x05,0x58,0x0c,0xcc,0xc1,0xfe,0x00,0x00]
 932
 933 v_pk_max_u16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
 934 // GFX12: [0x05,0x40,0x0c,0xcc,0xf0,0xfa,0x00,0x18]
 935
 936 v_pk_max_u16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
 937 // GFX12: [0x05,0x48,0x0c,0xcc,0xfd,0xd4,0x00,0x10]
 938
 939 v_pk_max_u16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
 940 // GFX12: [0xff,0x50,0x0c,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
 941
 942 v_pk_min_num_f16 v5, v1, v2
 943 // GFX12: [0x05,0x40,0x1b,0xcc,0x01,0x05,0x02,0x18]
 944
 945 v_pk_min_num_f16 v5, v255, v255
 946 // GFX12: [0x05,0x40,0x1b,0xcc,0xff,0xff,0x03,0x18]
 947
 948 v_pk_min_num_f16 v5, s1, s2
 949 // GFX12: [0x05,0x40,0x1b,0xcc,0x01,0x04,0x00,0x18]
 950
 951 v_pk_min_num_f16 v5, s105, s105
 952 // GFX12: [0x05,0x40,0x1b,0xcc,0x69,0xd2,0x00,0x18]
 953
 954 v_pk_min_num_f16 v5, vcc_lo, ttmp15
 955 // GFX12: [0x05,0x40,0x1b,0xcc,0x6a,0xf6,0x00,0x18]
 956
 957 v_pk_min_num_f16 v5, vcc_hi, 0xfe0b
 958 // GFX12: [0x05,0x40,0x1b,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
 959
 960 v_pk_min_num_f16 v5, ttmp15, src_scc
 961 // GFX12: [0x05,0x40,0x1b,0xcc,0x7b,0xfa,0x01,0x18]
 962
 963 v_pk_min_num_f16 v5, m0, 0.5
 964 // GFX12: [0x05,0x40,0x1b,0xcc,0x7d,0xe0,0x01,0x18]
 965
 966 v_pk_min_num_f16 v5, exec_lo, -1
 967 // GFX12: [0x05,0x40,0x1b,0xcc,0x7e,0x82,0x01,0x18]
 968
 969 v_pk_min_num_f16 v5, exec_hi, null
 970 // GFX12: [0x05,0x40,0x1b,0xcc,0x7f,0xf8,0x00,0x18]
 971
 972 v_pk_min_num_f16 v5, null, exec_lo
 973 // GFX12: [0x05,0x40,0x1b,0xcc,0x7c,0xfc,0x00,0x18]
 974
 975 v_pk_min_num_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
 976 // GFX12: [0x05,0x59,0x1b,0xcc,0xc1,0xfe,0x00,0x20]
 977
 978 v_pk_min_num_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
 979 // GFX12: [0x05,0x42,0x1b,0xcc,0xf0,0xfa,0x00,0x58]
 980
 981 v_pk_min_num_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
 982 // GFX12: [0x05,0x48,0x1b,0xcc,0xfd,0xd4,0x00,0x10]
 983
 984 v_pk_min_num_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
 985 // GFX12: [0xff,0xd3,0x1b,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
 986
 987 v_pk_min_i16 v5, v1, v2
 988 // GFX12: [0x05,0x40,0x08,0xcc,0x01,0x05,0x02,0x18]
 989
 990 v_pk_min_i16 v5, v255, v255
 991 // GFX12: [0x05,0x40,0x08,0xcc,0xff,0xff,0x03,0x18]
 992
 993 v_pk_min_i16 v5, s1, s2
 994 // GFX12: [0x05,0x40,0x08,0xcc,0x01,0x04,0x00,0x18]
 995
 996 v_pk_min_i16 v5, s105, s105
 997 // GFX12: [0x05,0x40,0x08,0xcc,0x69,0xd2,0x00,0x18]
 998
 999 v_pk_min_i16 v5, vcc_lo, ttmp15
1000 // GFX12: [0x05,0x40,0x08,0xcc,0x6a,0xf6,0x00,0x18]
1001
1002 v_pk_min_i16 v5, vcc_hi, 0xfe0b
1003 // GFX12: [0x05,0x40,0x08,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1004
1005 v_pk_min_i16 v5, ttmp15, src_scc
1006 // GFX12: [0x05,0x40,0x08,0xcc,0x7b,0xfa,0x01,0x18]
1007
1008 v_pk_min_i16 v5, m0, 0.5
1009 // GFX12: [0x05,0x40,0x08,0xcc,0x7d,0xe0,0x01,0x18]
1010
1011 v_pk_min_i16 v5, exec_lo, -1
1012 // GFX12: [0x05,0x40,0x08,0xcc,0x7e,0x82,0x01,0x18]
1013
1014 v_pk_min_i16 v5, exec_hi, null
1015 // GFX12: [0x05,0x40,0x08,0xcc,0x7f,0xf8,0x00,0x18]
1016
1017 v_pk_min_i16 v5, null, exec_lo
1018 // GFX12: [0x05,0x40,0x08,0xcc,0x7c,0xfc,0x00,0x18]
1019
1020 v_pk_min_i16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
1021 // GFX12: [0x05,0x58,0x08,0xcc,0xc1,0xfe,0x00,0x00]
1022
1023 v_pk_min_i16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
1024 // GFX12: [0x05,0x40,0x08,0xcc,0xf0,0xfa,0x00,0x18]
1025
1026 v_pk_min_i16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
1027 // GFX12: [0x05,0x48,0x08,0xcc,0xfd,0xd4,0x00,0x10]
1028
1029 v_pk_min_i16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
1030 // GFX12: [0xff,0x50,0x08,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
1031
1032 v_pk_min_u16 v5, v1, v2
1033 // GFX12: [0x05,0x40,0x0d,0xcc,0x01,0x05,0x02,0x18]
1034
1035 v_pk_min_u16 v5, v255, v255
1036 // GFX12: [0x05,0x40,0x0d,0xcc,0xff,0xff,0x03,0x18]
1037
1038 v_pk_min_u16 v5, s1, s2
1039 // GFX12: [0x05,0x40,0x0d,0xcc,0x01,0x04,0x00,0x18]
1040
1041 v_pk_min_u16 v5, s105, s105
1042 // GFX12: [0x05,0x40,0x0d,0xcc,0x69,0xd2,0x00,0x18]
1043
1044 v_pk_min_u16 v5, vcc_lo, ttmp15
1045 // GFX12: [0x05,0x40,0x0d,0xcc,0x6a,0xf6,0x00,0x18]
1046
1047 v_pk_min_u16 v5, vcc_hi, 0xfe0b
1048 // GFX12: [0x05,0x40,0x0d,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1049
1050 v_pk_min_u16 v5, ttmp15, src_scc
1051 // GFX12: [0x05,0x40,0x0d,0xcc,0x7b,0xfa,0x01,0x18]
1052
1053 v_pk_min_u16 v5, m0, 0.5
1054 // GFX12: [0x05,0x40,0x0d,0xcc,0x7d,0xe0,0x01,0x18]
1055
1056 v_pk_min_u16 v5, exec_lo, -1
1057 // GFX12: [0x05,0x40,0x0d,0xcc,0x7e,0x82,0x01,0x18]
1058
1059 v_pk_min_u16 v5, exec_hi, null
1060 // GFX12: [0x05,0x40,0x0d,0xcc,0x7f,0xf8,0x00,0x18]
1061
1062 v_pk_min_u16 v5, null, exec_lo
1063 // GFX12: [0x05,0x40,0x0d,0xcc,0x7c,0xfc,0x00,0x18]
1064
1065 v_pk_min_u16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
1066 // GFX12: [0x05,0x58,0x0d,0xcc,0xc1,0xfe,0x00,0x00]
1067
1068 v_pk_min_u16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
1069 // GFX12: [0x05,0x40,0x0d,0xcc,0xf0,0xfa,0x00,0x18]
1070
1071 v_pk_min_u16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
1072 // GFX12: [0x05,0x48,0x0d,0xcc,0xfd,0xd4,0x00,0x10]
1073
1074 v_pk_min_u16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
1075 // GFX12: [0xff,0x50,0x0d,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
1076
1077 v_pk_mul_f16 v5, v1, v2
1078 // GFX12: [0x05,0x40,0x10,0xcc,0x01,0x05,0x02,0x18]
1079
1080 v_pk_mul_f16 v5, v255, v255
1081 // GFX12: [0x05,0x40,0x10,0xcc,0xff,0xff,0x03,0x18]
1082
1083 v_pk_mul_f16 v5, s1, s2
1084 // GFX12: [0x05,0x40,0x10,0xcc,0x01,0x04,0x00,0x18]
1085
1086 v_pk_mul_f16 v5, s105, s105
1087 // GFX12: [0x05,0x40,0x10,0xcc,0x69,0xd2,0x00,0x18]
1088
1089 v_pk_mul_f16 v5, vcc_lo, ttmp15
1090 // GFX12: [0x05,0x40,0x10,0xcc,0x6a,0xf6,0x00,0x18]
1091
1092 v_pk_mul_f16 v5, vcc_hi, 0xfe0b
1093 // GFX12: [0x05,0x40,0x10,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1094
1095 v_pk_mul_f16 v5, ttmp15, src_scc
1096 // GFX12: [0x05,0x40,0x10,0xcc,0x7b,0xfa,0x01,0x18]
1097
1098 v_pk_mul_f16 v5, m0, 0.5
1099 // GFX12: [0x05,0x40,0x10,0xcc,0x7d,0xe0,0x01,0x18]
1100
1101 v_pk_mul_f16 v5, exec_lo, -1
1102 // GFX12: [0x05,0x40,0x10,0xcc,0x7e,0x82,0x01,0x18]
1103
1104 v_pk_mul_f16 v5, exec_hi, null
1105 // GFX12: [0x05,0x40,0x10,0xcc,0x7f,0xf8,0x00,0x18]
1106
1107 v_pk_mul_f16 v5, null, exec_lo
1108 // GFX12: [0x05,0x40,0x10,0xcc,0x7c,0xfc,0x00,0x18]
1109
1110 v_pk_mul_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
1111 // GFX12: [0x05,0x59,0x10,0xcc,0xc1,0xfe,0x00,0x20]
1112
1113 v_pk_mul_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
1114 // GFX12: [0x05,0x42,0x10,0xcc,0xf0,0xfa,0x00,0x58]
1115
1116 v_pk_mul_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
1117 // GFX12: [0x05,0x48,0x10,0xcc,0xfd,0xd4,0x00,0x10]
1118
1119 v_pk_mul_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
1120 // GFX12: [0xff,0xd3,0x10,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
1121
1122 v_pk_mul_lo_u16 v5, v1, v2
1123 // GFX12: [0x05,0x40,0x01,0xcc,0x01,0x05,0x02,0x18]
1124
1125 v_pk_mul_lo_u16 v5, v255, v255
1126 // GFX12: [0x05,0x40,0x01,0xcc,0xff,0xff,0x03,0x18]
1127
1128 v_pk_mul_lo_u16 v5, s1, s2
1129 // GFX12: [0x05,0x40,0x01,0xcc,0x01,0x04,0x00,0x18]
1130
1131 v_pk_mul_lo_u16 v5, s105, s105
1132 // GFX12: [0x05,0x40,0x01,0xcc,0x69,0xd2,0x00,0x18]
1133
1134 v_pk_mul_lo_u16 v5, vcc_lo, ttmp15
1135 // GFX12: [0x05,0x40,0x01,0xcc,0x6a,0xf6,0x00,0x18]
1136
1137 v_pk_mul_lo_u16 v5, vcc_hi, 0xfe0b
1138 // GFX12: [0x05,0x40,0x01,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1139
1140 v_pk_mul_lo_u16 v5, ttmp15, src_scc
1141 // GFX12: [0x05,0x40,0x01,0xcc,0x7b,0xfa,0x01,0x18]
1142
1143 v_pk_mul_lo_u16 v5, m0, 0.5
1144 // GFX12: [0x05,0x40,0x01,0xcc,0x7d,0xe0,0x01,0x18]
1145
1146 v_pk_mul_lo_u16 v5, exec_lo, -1
1147 // GFX12: [0x05,0x40,0x01,0xcc,0x7e,0x82,0x01,0x18]
1148
1149 v_pk_mul_lo_u16 v5, exec_hi, null
1150 // GFX12: [0x05,0x40,0x01,0xcc,0x7f,0xf8,0x00,0x18]
1151
1152 v_pk_mul_lo_u16 v5, null, exec_lo
1153 // GFX12: [0x05,0x40,0x01,0xcc,0x7c,0xfc,0x00,0x18]
1154
1155 v_pk_mul_lo_u16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
1156 // GFX12: [0x05,0x58,0x01,0xcc,0xc1,0xfe,0x00,0x00]
1157
1158 v_pk_mul_lo_u16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
1159 // GFX12: [0x05,0x40,0x01,0xcc,0xf0,0xfa,0x00,0x18]
1160
1161 v_pk_mul_lo_u16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
1162 // GFX12: [0x05,0x48,0x01,0xcc,0xfd,0xd4,0x00,0x10]
1163
1164 v_pk_mul_lo_u16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0]
1165 // GFX12: [0xff,0x50,0x01,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
1166
1167 v_pk_sub_i16 v5, v1, v2
1168 // GFX12: [0x05,0x40,0x03,0xcc,0x01,0x05,0x02,0x18]
1169
1170 v_pk_sub_i16 v5, v255, v255
1171 // GFX12: [0x05,0x40,0x03,0xcc,0xff,0xff,0x03,0x18]
1172
1173 v_pk_sub_i16 v5, s1, s2
1174 // GFX12: [0x05,0x40,0x03,0xcc,0x01,0x04,0x00,0x18]
1175
1176 v_pk_sub_i16 v5, s105, s105
1177 // GFX12: [0x05,0x40,0x03,0xcc,0x69,0xd2,0x00,0x18]
1178
1179 v_pk_sub_i16 v5, vcc_lo, ttmp15
1180 // GFX12: [0x05,0x40,0x03,0xcc,0x6a,0xf6,0x00,0x18]
1181
1182 v_pk_sub_i16 v5, vcc_hi, 0xfe0b
1183 // GFX12: [0x05,0x40,0x03,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1184
1185 v_pk_sub_i16 v5, ttmp15, src_scc
1186 // GFX12: [0x05,0x40,0x03,0xcc,0x7b,0xfa,0x01,0x18]
1187
1188 v_pk_sub_i16 v5, m0, 0.5
1189 // GFX12: [0x05,0x40,0x03,0xcc,0x7d,0xe0,0x01,0x18]
1190
1191 v_pk_sub_i16 v5, exec_lo, -1
1192 // GFX12: [0x05,0x40,0x03,0xcc,0x7e,0x82,0x01,0x18]
1193
1194 v_pk_sub_i16 v5, exec_hi, null
1195 // GFX12: [0x05,0x40,0x03,0xcc,0x7f,0xf8,0x00,0x18]
1196
1197 v_pk_sub_i16 v5, null, exec_lo
1198 // GFX12: [0x05,0x40,0x03,0xcc,0x7c,0xfc,0x00,0x18]
1199
1200 v_pk_sub_i16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
1201 // GFX12: [0x05,0x58,0x03,0xcc,0xc1,0xfe,0x00,0x00]
1202
1203 v_pk_sub_i16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
1204 // GFX12: [0x05,0x40,0x03,0xcc,0xf0,0xfa,0x00,0x18]
1205
1206 v_pk_sub_i16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
1207 // GFX12: [0x05,0x48,0x03,0xcc,0xfd,0xd4,0x00,0x10]
1208
1209 v_pk_sub_i16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] clamp
1210 // GFX12: [0xff,0xd0,0x03,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
1211
1212 v_pk_sub_u16 v5, v1, v2
1213 // GFX12: [0x05,0x40,0x0b,0xcc,0x01,0x05,0x02,0x18]
1214
1215 v_pk_sub_u16 v5, v255, v255
1216 // GFX12: [0x05,0x40,0x0b,0xcc,0xff,0xff,0x03,0x18]
1217
1218 v_pk_sub_u16 v5, s1, s2
1219 // GFX12: [0x05,0x40,0x0b,0xcc,0x01,0x04,0x00,0x18]
1220
1221 v_pk_sub_u16 v5, s105, s105
1222 // GFX12: [0x05,0x40,0x0b,0xcc,0x69,0xd2,0x00,0x18]
1223
1224 v_pk_sub_u16 v5, vcc_lo, ttmp15
1225 // GFX12: [0x05,0x40,0x0b,0xcc,0x6a,0xf6,0x00,0x18]
1226
1227 v_pk_sub_u16 v5, vcc_hi, 0xfe0b
1228 // GFX12: [0x05,0x40,0x0b,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1229
1230 v_pk_sub_u16 v5, ttmp15, src_scc
1231 // GFX12: [0x05,0x40,0x0b,0xcc,0x7b,0xfa,0x01,0x18]
1232
1233 v_pk_sub_u16 v5, m0, 0.5
1234 // GFX12: [0x05,0x40,0x0b,0xcc,0x7d,0xe0,0x01,0x18]
1235
1236 v_pk_sub_u16 v5, exec_lo, -1
1237 // GFX12: [0x05,0x40,0x0b,0xcc,0x7e,0x82,0x01,0x18]
1238
1239 v_pk_sub_u16 v5, exec_hi, null
1240 // GFX12: [0x05,0x40,0x0b,0xcc,0x7f,0xf8,0x00,0x18]
1241
1242 v_pk_sub_u16 v5, null, exec_lo
1243 // GFX12: [0x05,0x40,0x0b,0xcc,0x7c,0xfc,0x00,0x18]
1244
1245 v_pk_sub_u16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0]
1246 // GFX12: [0x05,0x58,0x0b,0xcc,0xc1,0xfe,0x00,0x00]
1247
1248 v_pk_sub_u16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1]
1249 // GFX12: [0x05,0x40,0x0b,0xcc,0xf0,0xfa,0x00,0x18]
1250
1251 v_pk_sub_u16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1]
1252 // GFX12: [0x05,0x48,0x0b,0xcc,0xfd,0xd4,0x00,0x10]
1253
1254 v_pk_sub_u16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] clamp
1255 // GFX12: [0xff,0xd0,0x0b,0xcc,0xff,0xd6,0x00,0x08,0x0b,0xfe,0x00,0x00]
1256
1257 v_pk_minimum_f16 v5, v1, v2
1258 // GFX12: [0x05,0x40,0x1d,0xcc,0x01,0x05,0x02,0x18]
1259
1260 v_pk_minimum_f16 v5, v255, v255
1261 // GFX12: [0x05,0x40,0x1d,0xcc,0xff,0xff,0x03,0x18]
1262
1263 v_pk_minimum_f16 v5, s1, s2
1264 // GFX12: [0x05,0x40,0x1d,0xcc,0x01,0x04,0x00,0x18]
1265
1266 v_pk_minimum_f16 v5, s105, s105
1267 // GFX12: [0x05,0x40,0x1d,0xcc,0x69,0xd2,0x00,0x18]
1268
1269 v_pk_minimum_f16 v5, vcc_lo, ttmp15
1270 // GFX12: [0x05,0x40,0x1d,0xcc,0x6a,0xf6,0x00,0x18]
1271
1272 v_pk_minimum_f16 v5, vcc_hi, 0xfe0b
1273 // GFX12: [0x05,0x40,0x1d,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1274
1275 v_pk_minimum_f16 v5, ttmp15, src_scc
1276 // GFX12: [0x05,0x40,0x1d,0xcc,0x7b,0xfa,0x01,0x18]
1277
1278 v_pk_minimum_f16 v5, m0, 0.5
1279 // GFX12: [0x05,0x40,0x1d,0xcc,0x7d,0xe0,0x01,0x18]
1280
1281 v_pk_minimum_f16 v5, exec_lo, -1
1282 // GFX12: [0x05,0x40,0x1d,0xcc,0x7e,0x82,0x01,0x18]
1283
1284 v_pk_minimum_f16 v5, exec_hi, null
1285 // GFX12: [0x05,0x40,0x1d,0xcc,0x7f,0xf8,0x00,0x18]
1286
1287 v_pk_minimum_f16 v5, null, exec_lo
1288 // GFX12: [0x05,0x40,0x1d,0xcc,0x7c,0xfc,0x00,0x18]
1289
1290 v_pk_minimum_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
1291 // GFX12: [0x05,0x59,0x1d,0xcc,0xc1,0xfe,0x00,0x20]
1292
1293 v_pk_minimum_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
1294 // GFX12: [0x05,0x42,0x1d,0xcc,0xf0,0xfa,0x00,0x58]
1295
1296 v_pk_minimum_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
1297 // GFX12: [0x05,0x48,0x1d,0xcc,0xfd,0xd4,0x00,0x10]
1298
1299 v_pk_minimum_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
1300 // GFX12: [0xff,0xd3,0x1d,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
1301
1302 v_pk_maximum_f16 v5, v1, v2
1303 // GFX12: [0x05,0x40,0x1e,0xcc,0x01,0x05,0x02,0x18]
1304
1305 v_pk_maximum_f16 v5, v255, v255
1306 // GFX12: [0x05,0x40,0x1e,0xcc,0xff,0xff,0x03,0x18]
1307
1308 v_pk_maximum_f16 v5, s1, s2
1309 // GFX12: [0x05,0x40,0x1e,0xcc,0x01,0x04,0x00,0x18]
1310
1311 v_pk_maximum_f16 v5, s105, s105
1312 // GFX12: [0x05,0x40,0x1e,0xcc,0x69,0xd2,0x00,0x18]
1313
1314 v_pk_maximum_f16 v5, vcc_lo, ttmp15
1315 // GFX12: [0x05,0x40,0x1e,0xcc,0x6a,0xf6,0x00,0x18]
1316
1317 v_pk_maximum_f16 v5, vcc_hi, 0xfe0b
1318 // GFX12: [0x05,0x40,0x1e,0xcc,0x6b,0xfe,0x01,0x18,0x0b,0xfe,0x00,0x00]
1319
1320 v_pk_maximum_f16 v5, ttmp15, src_scc
1321 // GFX12: [0x05,0x40,0x1e,0xcc,0x7b,0xfa,0x01,0x18]
1322
1323 v_pk_maximum_f16 v5, m0, 0.5
1324 // GFX12: [0x05,0x40,0x1e,0xcc,0x7d,0xe0,0x01,0x18]
1325
1326 v_pk_maximum_f16 v5, exec_lo, -1
1327 // GFX12: [0x05,0x40,0x1e,0xcc,0x7e,0x82,0x01,0x18]
1328
1329 v_pk_maximum_f16 v5, exec_hi, null
1330 // GFX12: [0x05,0x40,0x1e,0xcc,0x7f,0xf8,0x00,0x18]
1331
1332 v_pk_maximum_f16 v5, null, exec_lo
1333 // GFX12: [0x05,0x40,0x1e,0xcc,0x7c,0xfc,0x00,0x18]
1334
1335 v_pk_maximum_f16 v5, -1, exec_hi op_sel:[1,1] op_sel_hi:[0,0] neg_lo:[1,0] neg_hi:[1,0]
1336 // GFX12: [0x05,0x59,0x1e,0xcc,0xc1,0xfe,0x00,0x20]
1337
1338 v_pk_maximum_f16 v5, 0.5, m0 op_sel:[0,0] op_sel_hi:[1,1] neg_lo:[0,1] neg_hi:[0,1]
1339 // GFX12: [0x05,0x42,0x1e,0xcc,0xf0,0xfa,0x00,0x58]
1340
1341 v_pk_maximum_f16 v5, src_scc, vcc_lo op_sel:[1,0] op_sel_hi:[0,1] neg_lo:[0,0] neg_hi:[0,0]
1342 // GFX12: [0x05,0x48,0x1e,0xcc,0xfd,0xd4,0x00,0x10]
1343
1344 v_pk_maximum_f16 v255, 0xfe0b, vcc_hi op_sel:[0,1] op_sel_hi:[1,0] neg_lo:[1,1] neg_hi:[1,1] clamp
1345 // GFX12: [0xff,0xd3,0x1e,0xcc,0xff,0xd6,0x00,0x68,0x0b,0xfe,0x00,0x00]
1346
1347 v_dot4_f32_fp8_bf8 v0, v1, v2, v3
1348 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x05,0x0e,0x1c]
1349
1350 v_dot4_f32_fp8_bf8 v0, v1, v2, v3 neg_lo:[0,0,1]
1351 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, v3 neg_lo:[0,0,1] ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x05,0x0e,0x9c]
1352
1353 v_dot4_f32_fp8_bf8 v0, v1, v2, v3 neg_hi:[0,0,1]
1354 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, v3 neg_hi:[0,0,1] ; encoding: [0x00,0x44,0x24,0xcc,0x01,0x05,0x0e,0x1c]
1355
1356 v_dot4_f32_fp8_bf8 v0, s0, v2, v3
1357 // GFX12: v_dot4_f32_fp8_bf8 v0, s0, v2, v3 ; encoding: [0x00,0x40,0x24,0xcc,0x00,0x04,0x0e,0x1c]
1358
1359 v_dot4_f32_fp8_bf8 v0, v1, s0, v3
1360 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, s0, v3 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x01,0x0c,0x1c]
1361
1362 v_dot4_f32_fp8_bf8 v0, v1, v2, s0
1363 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, s0 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x05,0x02,0x18]
1364
1365 v_dot4_f32_fp8_bf8 v0, 1.0, v2, v3
1366 // GFX12: v_dot4_f32_fp8_bf8 v0, 1.0, v2, v3 ; encoding: [0x00,0x40,0x24,0xcc,0xf2,0x04,0x0e,0x1c]
1367
1368 v_dot4_f32_fp8_bf8 v0, v1, 1.0, v3
1369 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, 1.0, v3 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0xe5,0x0d,0x1c]
1370
1371 v_dot4_f32_fp8_bf8 v0, v1, v2, 1.0
1372 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, 1.0 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x05,0xca,0x1b]
1373
1374 v_dot4_f32_fp8_bf8 v0, v1, v2, 1
1375 // GFX12: v_dot4_f32_fp8_bf8 v0, v1, v2, 1 ; encoding: [0x00,0x40,0x24,0xcc,0x01,0x05,0x06,0x1a]
1376
1377 v_dot4_f32_bf8_fp8 v0, v1, v2, v3
1378 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x05,0x0e,0x1c]
1379
1380 v_dot4_f32_bf8_fp8 v0, v1, v2, v3 neg_lo:[0,0,1]
1381 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, v3 neg_lo:[0,0,1] ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x05,0x0e,0x9c]
1382
1383 v_dot4_f32_bf8_fp8 v0, v1, v2, v3 neg_hi:[0,0,1]
1384 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, v3 neg_hi:[0,0,1] ; encoding: [0x00,0x44,0x25,0xcc,0x01,0x05,0x0e,0x1c]
1385
1386 v_dot4_f32_bf8_fp8 v0, s0, v2, v3
1387 // GFX12: v_dot4_f32_bf8_fp8 v0, s0, v2, v3 ; encoding: [0x00,0x40,0x25,0xcc,0x00,0x04,0x0e,0x1c]
1388
1389 v_dot4_f32_bf8_fp8 v0, v1, s0, v3
1390 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, s0, v3 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x01,0x0c,0x1c]
1391
1392 v_dot4_f32_bf8_fp8 v0, v1, v2, s0
1393 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, s0 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x05,0x02,0x18]
1394
1395 v_dot4_f32_bf8_fp8 v0, 1.0, v2, v3
1396 // GFX12: v_dot4_f32_bf8_fp8 v0, 1.0, v2, v3 ; encoding: [0x00,0x40,0x25,0xcc,0xf2,0x04,0x0e,0x1c]
1397
1398 v_dot4_f32_bf8_fp8 v0, v1, 1.0, v3
1399 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, 1.0, v3 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0xe5,0x0d,0x1c]
1400
1401 v_dot4_f32_bf8_fp8 v0, v1, v2, 1.0
1402 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, 1.0 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x05,0xca,0x1b]
1403
1404 v_dot4_f32_bf8_fp8 v0, v1, v2, 1
1405 // GFX12: v_dot4_f32_bf8_fp8 v0, v1, v2, 1 ; encoding: [0x00,0x40,0x25,0xcc,0x01,0x05,0x06,0x1a]
1406
1407 v_dot4_f32_fp8_fp8 v0, v1, v2, v3
1408 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x05,0x0e,0x1c]
1409
1410 v_dot4_f32_fp8_fp8 v0, v1, v2, v3 neg_lo:[0,0,1]
1411 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, v3 neg_lo:[0,0,1] ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x05,0x0e,0x9c]
1412
1413 v_dot4_f32_fp8_fp8 v0, v1, v2, v3 neg_hi:[0,0,1]
1414 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, v3 neg_hi:[0,0,1] ; encoding: [0x00,0x44,0x26,0xcc,0x01,0x05,0x0e,0x1c]
1415
1416 v_dot4_f32_fp8_fp8 v0, s0, v2, v3
1417 // GFX12: v_dot4_f32_fp8_fp8 v0, s0, v2, v3 ; encoding: [0x00,0x40,0x26,0xcc,0x00,0x04,0x0e,0x1c]
1418
1419 v_dot4_f32_fp8_fp8 v0, v1, s0, v3
1420 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, s0, v3 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x01,0x0c,0x1c]
1421
1422 v_dot4_f32_fp8_fp8 v0, v1, v2, s0
1423 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, s0 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x05,0x02,0x18]
1424
1425 v_dot4_f32_fp8_fp8 v0, 1.0, v2, v3
1426 // GFX12: v_dot4_f32_fp8_fp8 v0, 1.0, v2, v3 ; encoding: [0x00,0x40,0x26,0xcc,0xf2,0x04,0x0e,0x1c]
1427
1428 v_dot4_f32_fp8_fp8 v0, v1, 1.0, v3
1429 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, 1.0, v3 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0xe5,0x0d,0x1c]
1430
1431 v_dot4_f32_fp8_fp8 v0, v1, v2, 1.0
1432 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, 1.0 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x05,0xca,0x1b]
1433
1434 v_dot4_f32_fp8_fp8 v0, v1, v2, 1
1435 // GFX12: v_dot4_f32_fp8_fp8 v0, v1, v2, 1 ; encoding: [0x00,0x40,0x26,0xcc,0x01,0x05,0x06,0x1a]
1436
1437 v_dot4_f32_bf8_bf8 v0, v1, v2, v3
1438 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x05,0x0e,0x1c]
1439
1440 v_dot4_f32_bf8_bf8 v0, v1, v2, v3 neg_lo:[0,0,1]
1441 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, v2, v3 neg_lo:[0,0,1] ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x05,0x0e,0x9c]
1442
1443 v_dot4_f32_bf8_bf8 v0, v1, v2, v3 neg_hi:[0,0,1]
1444 // GFX12: _dot4_f32_bf8_bf8 v0, v1, v2, v3 neg_hi:[0,0,1] ; encoding: [0x00,0x44,0x27,0xcc,0x01,0x05,0x0e,0x1c]
1445
1446 v_dot4_f32_bf8_bf8 v0, s0, v2, v3
1447 // GFX12: v_dot4_f32_bf8_bf8 v0, s0, v2, v3 ; encoding: [0x00,0x40,0x27,0xcc,0x00,0x04,0x0e,0x1c]
1448
1449 v_dot4_f32_bf8_bf8 v0, v1, s0, v3
1450 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, s0, v3 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x01,0x0c,0x1c]
1451
1452 v_dot4_f32_bf8_bf8 v0, v1, v2, s0
1453 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, v2, s0 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x05,0x02,0x18]
1454
1455 v_dot4_f32_bf8_bf8 v0, 1.0, v2, v3
1456 // GFX12: v_dot4_f32_bf8_bf8 v0, 1.0, v2, v3 ; encoding: [0x00,0x40,0x27,0xcc,0xf2,0x04,0x0e,0x1c]
1457
1458 v_dot4_f32_bf8_bf8 v0, v1, 1.0, v3
1459 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, 1.0, v3 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0xe5,0x0d,0x1c]
1460
1461 v_dot4_f32_bf8_bf8 v0, v1, v2, 1.0
1462 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, v2, 1.0 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x05,0xca,0x1b]
1463
1464 v_dot4_f32_bf8_bf8 v0, v1, v2, 1
1465 // GFX12: v_dot4_f32_bf8_bf8 v0, v1, v2, 1 ; encoding: [0x00,0x40,0x27,0xcc,0x01,0x05,0x06,0x1a]