llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop1_dpp16-fake16.s

   1 // RUN: llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize32,-real-true16 -show-encoding %s | FileCheck --check-prefix=GFX12 %s
   2
   3 v_bfrev_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
   4 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
   5
   6 v_bfrev_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
   7 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
   8
   9 v_bfrev_b32_e64_dpp v5, v1 row_mirror
  10 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
  11
  12 v_bfrev_b32_e64_dpp v5, v1 row_half_mirror
  13 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
  14
  15 v_bfrev_b32_e64_dpp v5, v1 row_shl:1
  16 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
  17
  18 v_bfrev_b32_e64_dpp v5, v1 row_shl:15
  19 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
  20
  21 v_bfrev_b32_e64_dpp v5, v1 row_shr:1
  22 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
  23
  24 v_bfrev_b32_e64_dpp v5, v1 row_shr:15
  25 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
  26
  27 v_bfrev_b32_e64_dpp v5, v1 row_ror:1
  28 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
  29
  30 v_bfrev_b32_e64_dpp v5, v1 row_ror:15
  31 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
  32
  33 v_bfrev_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
  34 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
  35
  36 v_bfrev_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
  37 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
  38
  39 v_bfrev_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
  40 // GFX12: [0x05,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
  41
  42 v_bfrev_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
  43 // GFX12: [0xff,0x00,0xb8,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
  44
  45 v_ceil_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
  46 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
  47
  48 v_ceil_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
  49 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
  50
  51 v_ceil_f16_e64_dpp v5, v1 row_mirror
  52 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
  53
  54 v_ceil_f16_e64_dpp v5, v1 row_half_mirror
  55 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
  56
  57 v_ceil_f16_e64_dpp v5, v1 row_shl:1
  58 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
  59
  60 v_ceil_f16_e64_dpp v5, v1 row_shl:15
  61 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
  62
  63 v_ceil_f16_e64_dpp v5, v1 row_shr:1
  64 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
  65
  66 v_ceil_f16_e64_dpp v5, v1 row_shr:15
  67 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
  68
  69 v_ceil_f16_e64_dpp v5, v1 row_ror:1
  70 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
  71
  72 v_ceil_f16_e64_dpp v5, v1 row_ror:15
  73 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
  74
  75 v_ceil_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
  76 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
  77
  78 v_ceil_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
  79 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
  80
  81 v_ceil_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
  82 // GFX12: [0x05,0x00,0xdc,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
  83
  84 v_ceil_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
  85 // GFX12: [0xff,0x81,0xdc,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
  86
  87 v_ceil_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
  88 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
  89
  90 v_ceil_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
  91 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
  92
  93 v_ceil_f32_e64_dpp v5, v1 row_mirror
  94 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
  95
  96 v_ceil_f32_e64_dpp v5, v1 row_half_mirror
  97 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
  98
  99 v_ceil_f32_e64_dpp v5, v1 row_shl:1
 100 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 101
 102 v_ceil_f32_e64_dpp v5, v1 row_shl:15
 103 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 104
 105 v_ceil_f32_e64_dpp v5, v1 row_shr:1
 106 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 107
 108 v_ceil_f32_e64_dpp v5, v1 row_shr:15
 109 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 110
 111 v_ceil_f32_e64_dpp v5, v1 row_ror:1
 112 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 113
 114 v_ceil_f32_e64_dpp v5, v1 row_ror:15
 115 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 116
 117 v_ceil_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 118 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 119
 120 v_ceil_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 121 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 122
 123 v_ceil_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 124 // GFX12: [0x05,0x00,0xa2,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 125
 126 v_ceil_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 127 // GFX12: [0xff,0x81,0xa2,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
 128
 129 v_cls_i32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 130 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 131
 132 v_cls_i32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 133 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 134
 135 v_cls_i32_e64_dpp v5, v1 row_mirror
 136 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 137
 138 v_cls_i32_e64_dpp v5, v1 row_half_mirror
 139 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 140
 141 v_cls_i32_e64_dpp v5, v1 row_shl:1
 142 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 143
 144 v_cls_i32_e64_dpp v5, v1 row_shl:15
 145 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 146
 147 v_cls_i32_e64_dpp v5, v1 row_shr:1
 148 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 149
 150 v_cls_i32_e64_dpp v5, v1 row_shr:15
 151 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 152
 153 v_cls_i32_e64_dpp v5, v1 row_ror:1
 154 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 155
 156 v_cls_i32_e64_dpp v5, v1 row_ror:15
 157 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 158
 159 v_cls_i32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 160 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 161
 162 v_cls_i32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 163 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 164
 165 v_cls_i32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 166 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 167
 168 v_cls_i32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 169 // GFX12: [0xff,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
 170
 171 v_clz_i32_u32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 172 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 173
 174 v_clz_i32_u32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 175 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 176
 177 v_clz_i32_u32_e64_dpp v5, v1 row_mirror
 178 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 179
 180 v_clz_i32_u32_e64_dpp v5, v1 row_half_mirror
 181 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 182
 183 v_clz_i32_u32_e64_dpp v5, v1 row_shl:1
 184 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 185
 186 v_clz_i32_u32_e64_dpp v5, v1 row_shl:15
 187 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 188
 189 v_clz_i32_u32_e64_dpp v5, v1 row_shr:1
 190 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 191
 192 v_clz_i32_u32_e64_dpp v5, v1 row_shr:15
 193 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 194
 195 v_clz_i32_u32_e64_dpp v5, v1 row_ror:1
 196 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 197
 198 v_clz_i32_u32_e64_dpp v5, v1 row_ror:15
 199 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 200
 201 v_clz_i32_u32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 202 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 203
 204 v_clz_i32_u32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 205 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 206
 207 v_clz_i32_u32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 208 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 209
 210 v_clz_i32_u32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 211 // GFX12: [0xff,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
 212
 213 v_cos_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 214 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 215
 216 v_cos_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 217 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 218
 219 v_cos_f16_e64_dpp v5, v1 row_mirror
 220 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 221
 222 v_cos_f16_e64_dpp v5, v1 row_half_mirror
 223 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 224
 225 v_cos_f16_e64_dpp v5, v1 row_shl:1
 226 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 227
 228 v_cos_f16_e64_dpp v5, v1 row_shl:15
 229 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 230
 231 v_cos_f16_e64_dpp v5, v1 row_shr:1
 232 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 233
 234 v_cos_f16_e64_dpp v5, v1 row_shr:15
 235 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 236
 237 v_cos_f16_e64_dpp v5, v1 row_ror:1
 238 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 239
 240 v_cos_f16_e64_dpp v5, v1 row_ror:15
 241 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 242
 243 v_cos_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 244 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 245
 246 v_cos_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 247 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 248
 249 v_cos_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 250 // GFX12: [0x05,0x00,0xe1,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 251
 252 v_cos_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 253 // GFX12: [0xff,0x81,0xe1,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
 254
 255 v_cos_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 256 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 257
 258 v_cos_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 259 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 260
 261 v_cos_f32_e64_dpp v5, v1 row_mirror
 262 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 263
 264 v_cos_f32_e64_dpp v5, v1 row_half_mirror
 265 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 266
 267 v_cos_f32_e64_dpp v5, v1 row_shl:1
 268 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 269
 270 v_cos_f32_e64_dpp v5, v1 row_shl:15
 271 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 272
 273 v_cos_f32_e64_dpp v5, v1 row_shr:1
 274 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 275
 276 v_cos_f32_e64_dpp v5, v1 row_shr:15
 277 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 278
 279 v_cos_f32_e64_dpp v5, v1 row_ror:1
 280 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 281
 282 v_cos_f32_e64_dpp v5, v1 row_ror:15
 283 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 284
 285 v_cos_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 286 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 287
 288 v_cos_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 289 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 290
 291 v_cos_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 292 // GFX12: [0x05,0x00,0xb6,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 293
 294 v_cos_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 295 // GFX12: [0xff,0x81,0xb6,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
 296
 297 v_ctz_i32_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 298 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 299
 300 v_ctz_i32_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 301 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 302
 303 v_ctz_i32_b32_e64_dpp v5, v1 row_mirror
 304 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 305
 306 v_ctz_i32_b32_e64_dpp v5, v1 row_half_mirror
 307 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 308
 309 v_ctz_i32_b32_e64_dpp v5, v1 row_shl:1
 310 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 311
 312 v_ctz_i32_b32_e64_dpp v5, v1 row_shl:15
 313 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 314
 315 v_ctz_i32_b32_e64_dpp v5, v1 row_shr:1
 316 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 317
 318 v_ctz_i32_b32_e64_dpp v5, v1 row_shr:15
 319 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 320
 321 v_ctz_i32_b32_e64_dpp v5, v1 row_ror:1
 322 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 323
 324 v_ctz_i32_b32_e64_dpp v5, v1 row_ror:15
 325 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 326
 327 v_ctz_i32_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 328 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 329
 330 v_ctz_i32_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 331 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 332
 333 v_ctz_i32_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 334 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 335
 336 v_ctz_i32_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 337 // GFX12: [0xff,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
 338
 339 V_CVT_F32_FP8_e64_dpp v5, v1 quad_perm:[3,1,2,0] row_mask:0x2 bank_mask:0xd
 340 // GFX12: encoding: [0x05,0x00,0xec,0xd5,0xfa,0x00,0x00,0x00,0x01,0x27,0x00,0x2d]
 341
 342 V_CVT_F32_FP8_e64_dpp v1, v3 quad_perm:[2,1,0,3] row_mask:0x5 bank_mask:0xe
 343 // GFX12: encoding: [0x01,0x00,0xec,0xd5,0xfa,0x00,0x00,0x00,0x03,0xc6,0x00,0x5e]
 344
 345 V_CVT_F32_BF8_e64_dpp v5, v1 quad_perm:[0,3,2,1] row_mask:0x2 bank_mask:0xd
 346 // GFX12: encoding: [0x05,0x00,0xed,0xd5,0xfa,0x00,0x00,0x00,0x01,0x6c,0x00,0x2d]
 347
 348 V_CVT_F32_BF8_e64_dpp v1, v3 quad_perm:[0,1,3,2] row_mask:0x5 bank_mask:0xe
 349 // GFX12: encoding: [0x01,0x00,0xed,0xd5,0xfa,0x00,0x00,0x00,0x03,0xb4,0x00,0x5e]
 350
 351 v_cvt_f16_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 352 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 353
 354 v_cvt_f16_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 355 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 356
 357 v_cvt_f16_f32_e64_dpp v5, v1 row_mirror
 358 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 359
 360 v_cvt_f16_f32_e64_dpp v5, v1 row_half_mirror
 361 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 362
 363 v_cvt_f16_f32_e64_dpp v5, v1 row_shl:1
 364 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 365
 366 v_cvt_f16_f32_e64_dpp v5, v1 row_shl:15
 367 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 368
 369 v_cvt_f16_f32_e64_dpp v5, v1 row_shr:1
 370 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 371
 372 v_cvt_f16_f32_e64_dpp v5, v1 row_shr:15
 373 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 374
 375 v_cvt_f16_f32_e64_dpp v5, v1 row_ror:1
 376 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 377
 378 v_cvt_f16_f32_e64_dpp v5, v1 row_ror:15
 379 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 380
 381 v_cvt_f16_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 382 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 383
 384 v_cvt_f16_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 385 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 386
 387 v_cvt_f16_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 388 // GFX12: [0x05,0x00,0x8a,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 389
 390 v_cvt_f16_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 391 // GFX12: [0xff,0x81,0x8a,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
 392
 393 v_cvt_f16_i16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 394 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 395
 396 v_cvt_f16_i16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 397 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 398
 399 v_cvt_f16_i16_e64_dpp v5, v1 row_mirror
 400 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 401
 402 v_cvt_f16_i16_e64_dpp v5, v1 row_half_mirror
 403 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 404
 405 v_cvt_f16_i16_e64_dpp v5, v1 row_shl:1
 406 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 407
 408 v_cvt_f16_i16_e64_dpp v5, v1 row_shl:15
 409 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 410
 411 v_cvt_f16_i16_e64_dpp v5, v1 row_shr:1
 412 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 413
 414 v_cvt_f16_i16_e64_dpp v5, v1 row_shr:15
 415 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 416
 417 v_cvt_f16_i16_e64_dpp v5, v1 row_ror:1
 418 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 419
 420 v_cvt_f16_i16_e64_dpp v5, v1 row_ror:15
 421 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 422
 423 v_cvt_f16_i16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 424 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 425
 426 v_cvt_f16_i16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 427 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 428
 429 v_cvt_f16_i16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 430 // GFX12: [0x05,0x00,0xd1,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 431
 432 v_cvt_f16_i16_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 433 // GFX12: [0xff,0x80,0xd1,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 434
 435 v_cvt_f16_u16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 436 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 437
 438 v_cvt_f16_u16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 439 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 440
 441 v_cvt_f16_u16_e64_dpp v5, v1 row_mirror
 442 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 443
 444 v_cvt_f16_u16_e64_dpp v5, v1 row_half_mirror
 445 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 446
 447 v_cvt_f16_u16_e64_dpp v5, v1 row_shl:1
 448 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 449
 450 v_cvt_f16_u16_e64_dpp v5, v1 row_shl:15
 451 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 452
 453 v_cvt_f16_u16_e64_dpp v5, v1 row_shr:1
 454 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 455
 456 v_cvt_f16_u16_e64_dpp v5, v1 row_shr:15
 457 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 458
 459 v_cvt_f16_u16_e64_dpp v5, v1 row_ror:1
 460 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 461
 462 v_cvt_f16_u16_e64_dpp v5, v1 row_ror:15
 463 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 464
 465 v_cvt_f16_u16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 466 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 467
 468 v_cvt_f16_u16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 469 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 470
 471 v_cvt_f16_u16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 472 // GFX12: [0x05,0x00,0xd0,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 473
 474 v_cvt_f16_u16_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 475 // GFX12: [0xff,0x80,0xd0,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 476
 477 v_cvt_f32_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 478 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 479
 480 v_cvt_f32_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 481 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 482
 483 v_cvt_f32_f16_e64_dpp v5, v1 row_mirror
 484 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 485
 486 v_cvt_f32_f16_e64_dpp v5, v1 row_half_mirror
 487 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 488
 489 v_cvt_f32_f16_e64_dpp v5, v1 row_shl:1
 490 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 491
 492 v_cvt_f32_f16_e64_dpp v5, v1 row_shl:15
 493 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 494
 495 v_cvt_f32_f16_e64_dpp v5, v1 row_shr:1
 496 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 497
 498 v_cvt_f32_f16_e64_dpp v5, v1 row_shr:15
 499 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 500
 501 v_cvt_f32_f16_e64_dpp v5, v1 row_ror:1
 502 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 503
 504 v_cvt_f32_f16_e64_dpp v5, v1 row_ror:15
 505 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 506
 507 v_cvt_f32_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 508 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 509
 510 v_cvt_f32_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 511 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 512
 513 v_cvt_f32_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 514 // GFX12: [0x05,0x00,0x8b,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 515
 516 v_cvt_f32_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 517 // GFX12: [0xff,0x81,0x8b,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
 518
 519 v_cvt_f32_fp8 v1, v2 quad_perm:[0,1,2,3]
 520 // GFX12: v_cvt_f32_fp8_dpp v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0xfa,0xd8,0x02,0x7e,0x02,0xe4,0x00,0xff]
 521
 522 v_cvt_f32_fp8 v1, v2 byte_sel:0 quad_perm:[0,1,2,3]
 523 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x00,0xec,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 524
 525 v_cvt_f32_fp8 v1, v2 byte_sel:1 quad_perm:[0,1,2,3]
 526 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:1 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x10,0xec,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 527
 528 v_cvt_f32_fp8 v1, v2 byte_sel:2 quad_perm:[0,1,2,3]
 529 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x08,0xec,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 530
 531 v_cvt_f32_fp8 v1, v2 byte_sel:3 quad_perm:[0,1,2,3]
 532 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:3 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x18,0xec,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 533
 534 v_cvt_f32_bf8 v1, v2 quad_perm:[0,1,2,3]
 535 // GFX12: v_cvt_f32_bf8_dpp v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0xfa,0xda,0x02,0x7e,0x02,0xe4,0x00,0xff]
 536
 537 v_cvt_f32_bf8 v1, v2 byte_sel:0 quad_perm:[0,1,2,3]
 538 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x00,0xed,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 539
 540 v_cvt_f32_bf8 v1, v2 byte_sel:1 quad_perm:[0,1,2,3]
 541 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:1 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x10,0xed,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 542
 543 v_cvt_f32_bf8 v1, v2 byte_sel:2 quad_perm:[0,1,2,3]
 544 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x08,0xed,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 545
 546 v_cvt_f32_bf8 v1, v2 byte_sel:3 quad_perm:[0,1,2,3]
 547 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:3 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x18,0xed,0xd5,0xfa,0x00,0x00,0x00,0x02,0xe4,0x00,0xff]
 548
 549 v_cvt_f32_i32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 550 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 551
 552 v_cvt_f32_i32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 553 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 554
 555 v_cvt_f32_i32_e64_dpp v5, v1 row_mirror
 556 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 557
 558 v_cvt_f32_i32_e64_dpp v5, v1 row_half_mirror
 559 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 560
 561 v_cvt_f32_i32_e64_dpp v5, v1 row_shl:1
 562 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 563
 564 v_cvt_f32_i32_e64_dpp v5, v1 row_shl:15
 565 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 566
 567 v_cvt_f32_i32_e64_dpp v5, v1 row_shr:1
 568 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 569
 570 v_cvt_f32_i32_e64_dpp v5, v1 row_shr:15
 571 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 572
 573 v_cvt_f32_i32_e64_dpp v5, v1 row_ror:1
 574 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 575
 576 v_cvt_f32_i32_e64_dpp v5, v1 row_ror:15
 577 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 578
 579 v_cvt_f32_i32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 580 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 581
 582 v_cvt_f32_i32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 583 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 584
 585 v_cvt_f32_i32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 586 // GFX12: [0x05,0x00,0x85,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 587
 588 v_cvt_f32_i32_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 589 // GFX12: [0xff,0x80,0x85,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 590
 591 v_cvt_f32_u32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 592 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 593
 594 v_cvt_f32_u32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 595 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 596
 597 v_cvt_f32_u32_e64_dpp v5, v1 row_mirror
 598 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 599
 600 v_cvt_f32_u32_e64_dpp v5, v1 row_half_mirror
 601 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 602
 603 v_cvt_f32_u32_e64_dpp v5, v1 row_shl:1
 604 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 605
 606 v_cvt_f32_u32_e64_dpp v5, v1 row_shl:15
 607 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 608
 609 v_cvt_f32_u32_e64_dpp v5, v1 row_shr:1
 610 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 611
 612 v_cvt_f32_u32_e64_dpp v5, v1 row_shr:15
 613 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 614
 615 v_cvt_f32_u32_e64_dpp v5, v1 row_ror:1
 616 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 617
 618 v_cvt_f32_u32_e64_dpp v5, v1 row_ror:15
 619 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 620
 621 v_cvt_f32_u32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 622 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 623
 624 v_cvt_f32_u32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 625 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 626
 627 v_cvt_f32_u32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 628 // GFX12: [0x05,0x00,0x86,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 629
 630 v_cvt_f32_u32_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 631 // GFX12: [0xff,0x80,0x86,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 632
 633 v_cvt_f32_ubyte0_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 634 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 635
 636 v_cvt_f32_ubyte0_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 637 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 638
 639 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_mirror
 640 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 641
 642 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_half_mirror
 643 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 644
 645 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_shl:1
 646 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 647
 648 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_shl:15
 649 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 650
 651 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_shr:1
 652 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 653
 654 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_shr:15
 655 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 656
 657 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_ror:1
 658 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 659
 660 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_ror:15
 661 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 662
 663 v_cvt_f32_ubyte0_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 664 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 665
 666 v_cvt_f32_ubyte0_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 667 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 668
 669 v_cvt_f32_ubyte0_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 670 // GFX12: [0x05,0x00,0x91,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 671
 672 v_cvt_f32_ubyte0_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 673 // GFX12: [0xff,0x80,0x91,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 674
 675 v_cvt_f32_ubyte1_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 676 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 677
 678 v_cvt_f32_ubyte1_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 679 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 680
 681 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_mirror
 682 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 683
 684 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_half_mirror
 685 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 686
 687 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_shl:1
 688 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 689
 690 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_shl:15
 691 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 692
 693 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_shr:1
 694 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 695
 696 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_shr:15
 697 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 698
 699 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_ror:1
 700 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 701
 702 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_ror:15
 703 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 704
 705 v_cvt_f32_ubyte1_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 706 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 707
 708 v_cvt_f32_ubyte1_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 709 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 710
 711 v_cvt_f32_ubyte1_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 712 // GFX12: [0x05,0x00,0x92,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 713
 714 v_cvt_f32_ubyte1_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 715 // GFX12: [0xff,0x80,0x92,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 716
 717 v_cvt_f32_ubyte2_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 718 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 719
 720 v_cvt_f32_ubyte2_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 721 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 722
 723 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_mirror
 724 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 725
 726 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_half_mirror
 727 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 728
 729 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_shl:1
 730 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 731
 732 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_shl:15
 733 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 734
 735 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_shr:1
 736 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 737
 738 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_shr:15
 739 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 740
 741 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_ror:1
 742 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 743
 744 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_ror:15
 745 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 746
 747 v_cvt_f32_ubyte2_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 748 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 749
 750 v_cvt_f32_ubyte2_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 751 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 752
 753 v_cvt_f32_ubyte2_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 754 // GFX12: [0x05,0x00,0x93,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 755
 756 v_cvt_f32_ubyte2_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 757 // GFX12: [0xff,0x80,0x93,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 758
 759 v_cvt_f32_ubyte3_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 760 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 761
 762 v_cvt_f32_ubyte3_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 763 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 764
 765 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_mirror
 766 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 767
 768 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_half_mirror
 769 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 770
 771 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_shl:1
 772 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 773
 774 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_shl:15
 775 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 776
 777 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_shr:1
 778 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 779
 780 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_shr:15
 781 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 782
 783 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_ror:1
 784 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 785
 786 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_ror:15
 787 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 788
 789 v_cvt_f32_ubyte3_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 790 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 791
 792 v_cvt_f32_ubyte3_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 793 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
 794
 795 v_cvt_f32_ubyte3_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 796 // GFX12: [0x05,0x00,0x94,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
 797
 798 v_cvt_f32_ubyte3_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 799 // GFX12: [0xff,0x80,0x94,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
 800
 801 v_cvt_floor_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 802 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 803
 804 v_cvt_floor_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 805 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 806
 807 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_mirror
 808 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 809
 810 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_half_mirror
 811 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 812
 813 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_shl:1
 814 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 815
 816 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_shl:15
 817 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 818
 819 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_shr:1
 820 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 821
 822 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_shr:15
 823 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 824
 825 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_ror:1
 826 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 827
 828 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_ror:15
 829 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 830
 831 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 832 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 833
 834 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 835 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 836
 837 v_cvt_floor_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 838 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 839
 840 v_cvt_floor_i32_f32_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 841 // GFX12: [0xff,0x01,0x8d,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
 842
 843 v_cvt_flr_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 844 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 845
 846 v_cvt_flr_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 847 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 848
 849 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_mirror
 850 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 851
 852 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_half_mirror
 853 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 854
 855 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_shl:1
 856 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 857
 858 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_shl:15
 859 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 860
 861 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_shr:1
 862 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 863
 864 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_shr:15
 865 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 866
 867 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_ror:1
 868 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 869
 870 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_ror:15
 871 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 872
 873 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 874 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 875
 876 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 877 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 878
 879 v_cvt_flr_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 880 // GFX12: [0x05,0x00,0x8d,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 881
 882 v_cvt_flr_i32_f32_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 883 // GFX12: [0xff,0x01,0x8d,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
 884
 885 v_cvt_i16_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 886 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 887
 888 v_cvt_i16_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 889 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 890
 891 v_cvt_i16_f16_e64_dpp v5, v1 row_mirror
 892 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 893
 894 v_cvt_i16_f16_e64_dpp v5, v1 row_half_mirror
 895 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 896
 897 v_cvt_i16_f16_e64_dpp v5, v1 row_shl:1
 898 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 899
 900 v_cvt_i16_f16_e64_dpp v5, v1 row_shl:15
 901 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 902
 903 v_cvt_i16_f16_e64_dpp v5, v1 row_shr:1
 904 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 905
 906 v_cvt_i16_f16_e64_dpp v5, v1 row_shr:15
 907 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 908
 909 v_cvt_i16_f16_e64_dpp v5, v1 row_ror:1
 910 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 911
 912 v_cvt_i16_f16_e64_dpp v5, v1 row_ror:15
 913 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 914
 915 v_cvt_i16_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 916 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 917
 918 v_cvt_i16_f16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 919 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 920
 921 v_cvt_i16_f16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 922 // GFX12: [0x05,0x00,0xd3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 923
 924 v_cvt_i16_f16_e64_dpp v255, -|v255| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 925 // GFX12: [0xff,0x81,0xd3,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
 926
 927 v_cvt_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 928 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 929
 930 v_cvt_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 931 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 932
 933 v_cvt_i32_f32_e64_dpp v5, v1 row_mirror
 934 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 935
 936 v_cvt_i32_f32_e64_dpp v5, v1 row_half_mirror
 937 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 938
 939 v_cvt_i32_f32_e64_dpp v5, v1 row_shl:1
 940 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 941
 942 v_cvt_i32_f32_e64_dpp v5, v1 row_shl:15
 943 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 944
 945 v_cvt_i32_f32_e64_dpp v5, v1 row_shr:1
 946 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 947
 948 v_cvt_i32_f32_e64_dpp v5, v1 row_shr:15
 949 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 950
 951 v_cvt_i32_f32_e64_dpp v5, v1 row_ror:1
 952 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 953
 954 v_cvt_i32_f32_e64_dpp v5, v1 row_ror:15
 955 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 956
 957 v_cvt_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
 958 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
 959
 960 v_cvt_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
 961 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
 962
 963 v_cvt_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 964 // GFX12: [0x05,0x00,0x88,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
 965
 966 v_cvt_i32_f32_e64_dpp v255, -|v255| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 967 // GFX12: [0xff,0x81,0x88,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
 968
 969 v_cvt_i32_i16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
 970 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
 971
 972 v_cvt_i32_i16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
 973 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
 974
 975 v_cvt_i32_i16_e64_dpp v5, v1 row_mirror
 976 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
 977
 978 v_cvt_i32_i16_e64_dpp v5, v1 row_half_mirror
 979 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
 980
 981 v_cvt_i32_i16_e64_dpp v5, v1 row_shl:1
 982 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
 983
 984 v_cvt_i32_i16_e64_dpp v5, v1 row_shl:15
 985 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
 986
 987 v_cvt_i32_i16_e64_dpp v5, v1 row_shr:1
 988 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
 989
 990 v_cvt_i32_i16_e64_dpp v5, v1 row_shr:15
 991 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
 992
 993 v_cvt_i32_i16_e64_dpp v5, v1 row_ror:1
 994 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
 995
 996 v_cvt_i32_i16_e64_dpp v5, v1 row_ror:15
 997 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
 998
 999 v_cvt_i32_i16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1000 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1001
1002 v_cvt_i32_i16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1003 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1004
1005 v_cvt_i32_i16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1006 // GFX12: [0x05,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1007
1008 v_cvt_i32_i16_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1009 // GFX12: [0xff,0x00,0xea,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
1010
1011 v_cvt_nearest_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1012 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1013
1014 v_cvt_nearest_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1015 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1016
1017 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_mirror
1018 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1019
1020 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_half_mirror
1021 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1022
1023 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_shl:1
1024 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1025
1026 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_shl:15
1027 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1028
1029 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_shr:1
1030 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1031
1032 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_shr:15
1033 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1034
1035 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_ror:1
1036 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1037
1038 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_ror:15
1039 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1040
1041 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1042 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1043
1044 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1045 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1046
1047 v_cvt_nearest_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1048 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1049
1050 v_cvt_nearest_i32_f32_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1051 // GFX12: [0xff,0x01,0x8c,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1052
1053 v_cvt_norm_i16_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1054 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1055
1056 v_cvt_norm_i16_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1057 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1058
1059 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_mirror
1060 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1061
1062 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_half_mirror
1063 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1064
1065 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_shl:1
1066 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1067
1068 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_shl:15
1069 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1070
1071 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_shr:1
1072 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1073
1074 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_shr:15
1075 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1076
1077 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_ror:1
1078 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1079
1080 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_ror:15
1081 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1082
1083 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1084 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1085
1086 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1087 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1088
1089 v_cvt_norm_i16_f16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1090 // GFX12: [0x05,0x00,0xe3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1091
1092 v_cvt_norm_i16_f16_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1093 // GFX12: [0xff,0x01,0xe3,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1094
1095 v_cvt_norm_u16_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1096 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1097
1098 v_cvt_norm_u16_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1099 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1100
1101 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_mirror
1102 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1103
1104 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_half_mirror
1105 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1106
1107 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_shl:1
1108 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1109
1110 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_shl:15
1111 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1112
1113 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_shr:1
1114 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1115
1116 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_shr:15
1117 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1118
1119 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_ror:1
1120 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1121
1122 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_ror:15
1123 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1124
1125 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1126 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1127
1128 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1129 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1130
1131 v_cvt_norm_u16_f16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1132 // GFX12: [0x05,0x00,0xe4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1133
1134 v_cvt_norm_u16_f16_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1135 // GFX12: [0xff,0x01,0xe4,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1136
1137 v_cvt_off_f32_i4_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1138 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1139
1140 v_cvt_off_f32_i4_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1141 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1142
1143 v_cvt_off_f32_i4_e64_dpp v5, v1 row_mirror
1144 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1145
1146 v_cvt_off_f32_i4_e64_dpp v5, v1 row_half_mirror
1147 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1148
1149 v_cvt_off_f32_i4_e64_dpp v5, v1 row_shl:1
1150 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1151
1152 v_cvt_off_f32_i4_e64_dpp v5, v1 row_shl:15
1153 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1154
1155 v_cvt_off_f32_i4_e64_dpp v5, v1 row_shr:1
1156 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1157
1158 v_cvt_off_f32_i4_e64_dpp v5, v1 row_shr:15
1159 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1160
1161 v_cvt_off_f32_i4_e64_dpp v5, v1 row_ror:1
1162 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1163
1164 v_cvt_off_f32_i4_e64_dpp v5, v1 row_ror:15
1165 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1166
1167 v_cvt_off_f32_i4_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1168 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1169
1170 v_cvt_off_f32_i4_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1171 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1172
1173 v_cvt_off_f32_i4_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1174 // GFX12: [0x05,0x00,0x8e,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1175
1176 v_cvt_off_f32_i4_e64_dpp v255, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1177 // GFX12: [0xff,0x80,0x8e,0xd5,0xfa,0x00,0x00,0x18,0xff,0x6f,0x05,0x30]
1178
1179 v_cvt_rpi_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1180 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1181
1182 v_cvt_rpi_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1183 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1184
1185 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_mirror
1186 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1187
1188 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_half_mirror
1189 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1190
1191 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_shl:1
1192 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1193
1194 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_shl:15
1195 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1196
1197 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_shr:1
1198 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1199
1200 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_shr:15
1201 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1202
1203 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_ror:1
1204 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1205
1206 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_ror:15
1207 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1208
1209 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1210 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1211
1212 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1213 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1214
1215 v_cvt_rpi_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1216 // GFX12: [0x05,0x00,0x8c,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1217
1218 v_cvt_rpi_i32_f32_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1219 // GFX12: [0xff,0x01,0x8c,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1220
1221 v_cvt_u16_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1222 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1223
1224 v_cvt_u16_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1225 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1226
1227 v_cvt_u16_f16_e64_dpp v5, v1 row_mirror
1228 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1229
1230 v_cvt_u16_f16_e64_dpp v5, v1 row_half_mirror
1231 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1232
1233 v_cvt_u16_f16_e64_dpp v5, v1 row_shl:1
1234 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1235
1236 v_cvt_u16_f16_e64_dpp v5, v1 row_shl:15
1237 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1238
1239 v_cvt_u16_f16_e64_dpp v5, v1 row_shr:1
1240 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1241
1242 v_cvt_u16_f16_e64_dpp v5, v1 row_shr:15
1243 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1244
1245 v_cvt_u16_f16_e64_dpp v5, v1 row_ror:1
1246 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1247
1248 v_cvt_u16_f16_e64_dpp v5, v1 row_ror:15
1249 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1250
1251 v_cvt_u16_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1252 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1253
1254 v_cvt_u16_f16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1255 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1256
1257 v_cvt_u16_f16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1258 // GFX12: [0x05,0x00,0xd2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1259
1260 v_cvt_u16_f16_e64_dpp v255, -|v255| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1261 // GFX12: [0xff,0x81,0xd2,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1262
1263 v_cvt_u32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1264 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1265
1266 v_cvt_u32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1267 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1268
1269 v_cvt_u32_f32_e64_dpp v5, v1 row_mirror
1270 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1271
1272 v_cvt_u32_f32_e64_dpp v5, v1 row_half_mirror
1273 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1274
1275 v_cvt_u32_f32_e64_dpp v5, v1 row_shl:1
1276 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1277
1278 v_cvt_u32_f32_e64_dpp v5, v1 row_shl:15
1279 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1280
1281 v_cvt_u32_f32_e64_dpp v5, v1 row_shr:1
1282 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1283
1284 v_cvt_u32_f32_e64_dpp v5, v1 row_shr:15
1285 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1286
1287 v_cvt_u32_f32_e64_dpp v5, v1 row_ror:1
1288 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1289
1290 v_cvt_u32_f32_e64_dpp v5, v1 row_ror:15
1291 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1292
1293 v_cvt_u32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1294 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1295
1296 v_cvt_u32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1297 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1298
1299 v_cvt_u32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1300 // GFX12: [0x05,0x00,0x87,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1301
1302 v_cvt_u32_f32_e64_dpp v255, -|v255| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1303 // GFX12: [0xff,0x81,0x87,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1304
1305 v_cvt_u32_u16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1306 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1307
1308 v_cvt_u32_u16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1309 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1310
1311 v_cvt_u32_u16_e64_dpp v5, v1 row_mirror
1312 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1313
1314 v_cvt_u32_u16_e64_dpp v5, v1 row_half_mirror
1315 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1316
1317 v_cvt_u32_u16_e64_dpp v5, v1 row_shl:1
1318 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1319
1320 v_cvt_u32_u16_e64_dpp v5, v1 row_shl:15
1321 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1322
1323 v_cvt_u32_u16_e64_dpp v5, v1 row_shr:1
1324 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1325
1326 v_cvt_u32_u16_e64_dpp v5, v1 row_shr:15
1327 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1328
1329 v_cvt_u32_u16_e64_dpp v5, v1 row_ror:1
1330 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1331
1332 v_cvt_u32_u16_e64_dpp v5, v1 row_ror:15
1333 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1334
1335 v_cvt_u32_u16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1336 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1337
1338 v_cvt_u32_u16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1339 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1340
1341 v_cvt_u32_u16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1342 // GFX12: [0x05,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1343
1344 v_cvt_u32_u16_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1345 // GFX12: [0xff,0x00,0xeb,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
1346
1347 v_exp_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1348 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1349
1350 v_exp_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1351 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1352
1353 v_exp_f16_e64_dpp v5, v1 row_mirror
1354 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1355
1356 v_exp_f16_e64_dpp v5, v1 row_half_mirror
1357 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1358
1359 v_exp_f16_e64_dpp v5, v1 row_shl:1
1360 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1361
1362 v_exp_f16_e64_dpp v5, v1 row_shl:15
1363 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1364
1365 v_exp_f16_e64_dpp v5, v1 row_shr:1
1366 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1367
1368 v_exp_f16_e64_dpp v5, v1 row_shr:15
1369 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1370
1371 v_exp_f16_e64_dpp v5, v1 row_ror:1
1372 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1373
1374 v_exp_f16_e64_dpp v5, v1 row_ror:15
1375 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1376
1377 v_exp_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1378 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1379
1380 v_exp_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1381 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1382
1383 v_exp_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1384 // GFX12: [0x05,0x00,0xd8,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1385
1386 v_exp_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1387 // GFX12: [0xff,0x81,0xd8,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1388
1389 v_exp_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1390 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1391
1392 v_exp_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1393 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1394
1395 v_exp_f32_e64_dpp v5, v1 row_mirror
1396 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1397
1398 v_exp_f32_e64_dpp v5, v1 row_half_mirror
1399 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1400
1401 v_exp_f32_e64_dpp v5, v1 row_shl:1
1402 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1403
1404 v_exp_f32_e64_dpp v5, v1 row_shl:15
1405 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1406
1407 v_exp_f32_e64_dpp v5, v1 row_shr:1
1408 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1409
1410 v_exp_f32_e64_dpp v5, v1 row_shr:15
1411 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1412
1413 v_exp_f32_e64_dpp v5, v1 row_ror:1
1414 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1415
1416 v_exp_f32_e64_dpp v5, v1 row_ror:15
1417 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1418
1419 v_exp_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1420 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1421
1422 v_exp_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1423 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1424
1425 v_exp_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1426 // GFX12: [0x05,0x00,0xa5,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1427
1428 v_exp_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1429 // GFX12: [0xff,0x81,0xa5,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1430
1431 v_ffbh_i32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1432 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1433
1434 v_ffbh_i32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1435 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1436
1437 v_ffbh_i32_e64_dpp v5, v1 row_mirror
1438 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1439
1440 v_ffbh_i32_e64_dpp v5, v1 row_half_mirror
1441 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1442
1443 v_ffbh_i32_e64_dpp v5, v1 row_shl:1
1444 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1445
1446 v_ffbh_i32_e64_dpp v5, v1 row_shl:15
1447 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1448
1449 v_ffbh_i32_e64_dpp v5, v1 row_shr:1
1450 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1451
1452 v_ffbh_i32_e64_dpp v5, v1 row_shr:15
1453 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1454
1455 v_ffbh_i32_e64_dpp v5, v1 row_ror:1
1456 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1457
1458 v_ffbh_i32_e64_dpp v5, v1 row_ror:15
1459 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1460
1461 v_ffbh_i32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1462 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1463
1464 v_ffbh_i32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1465 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1466
1467 v_ffbh_i32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1468 // GFX12: [0x05,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1469
1470 v_ffbh_i32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1471 // GFX12: [0xff,0x00,0xbb,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
1472
1473 v_ffbh_u32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1474 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1475
1476 v_ffbh_u32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1477 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1478
1479 v_ffbh_u32_e64_dpp v5, v1 row_mirror
1480 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1481
1482 v_ffbh_u32_e64_dpp v5, v1 row_half_mirror
1483 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1484
1485 v_ffbh_u32_e64_dpp v5, v1 row_shl:1
1486 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1487
1488 v_ffbh_u32_e64_dpp v5, v1 row_shl:15
1489 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1490
1491 v_ffbh_u32_e64_dpp v5, v1 row_shr:1
1492 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1493
1494 v_ffbh_u32_e64_dpp v5, v1 row_shr:15
1495 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1496
1497 v_ffbh_u32_e64_dpp v5, v1 row_ror:1
1498 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1499
1500 v_ffbh_u32_e64_dpp v5, v1 row_ror:15
1501 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1502
1503 v_ffbh_u32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1504 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1505
1506 v_ffbh_u32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1507 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1508
1509 v_ffbh_u32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1510 // GFX12: [0x05,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1511
1512 v_ffbh_u32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1513 // GFX12: [0xff,0x00,0xb9,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
1514
1515 v_ffbl_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1516 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1517
1518 v_ffbl_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1519 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1520
1521 v_ffbl_b32_e64_dpp v5, v1 row_mirror
1522 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1523
1524 v_ffbl_b32_e64_dpp v5, v1 row_half_mirror
1525 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1526
1527 v_ffbl_b32_e64_dpp v5, v1 row_shl:1
1528 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1529
1530 v_ffbl_b32_e64_dpp v5, v1 row_shl:15
1531 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1532
1533 v_ffbl_b32_e64_dpp v5, v1 row_shr:1
1534 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1535
1536 v_ffbl_b32_e64_dpp v5, v1 row_shr:15
1537 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1538
1539 v_ffbl_b32_e64_dpp v5, v1 row_ror:1
1540 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1541
1542 v_ffbl_b32_e64_dpp v5, v1 row_ror:15
1543 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1544
1545 v_ffbl_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1546 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1547
1548 v_ffbl_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1549 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1550
1551 v_ffbl_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1552 // GFX12: [0x05,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1553
1554 v_ffbl_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1555 // GFX12: [0xff,0x00,0xba,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
1556
1557 v_floor_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1558 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1559
1560 v_floor_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1561 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1562
1563 v_floor_f16_e64_dpp v5, v1 row_mirror
1564 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1565
1566 v_floor_f16_e64_dpp v5, v1 row_half_mirror
1567 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1568
1569 v_floor_f16_e64_dpp v5, v1 row_shl:1
1570 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1571
1572 v_floor_f16_e64_dpp v5, v1 row_shl:15
1573 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1574
1575 v_floor_f16_e64_dpp v5, v1 row_shr:1
1576 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1577
1578 v_floor_f16_e64_dpp v5, v1 row_shr:15
1579 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1580
1581 v_floor_f16_e64_dpp v5, v1 row_ror:1
1582 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1583
1584 v_floor_f16_e64_dpp v5, v1 row_ror:15
1585 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1586
1587 v_floor_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1588 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1589
1590 v_floor_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1591 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1592
1593 v_floor_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1594 // GFX12: [0x05,0x00,0xdb,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1595
1596 v_floor_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1597 // GFX12: [0xff,0x81,0xdb,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1598
1599 v_floor_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1600 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1601
1602 v_floor_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1603 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1604
1605 v_floor_f32_e64_dpp v5, v1 row_mirror
1606 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1607
1608 v_floor_f32_e64_dpp v5, v1 row_half_mirror
1609 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1610
1611 v_floor_f32_e64_dpp v5, v1 row_shl:1
1612 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1613
1614 v_floor_f32_e64_dpp v5, v1 row_shl:15
1615 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1616
1617 v_floor_f32_e64_dpp v5, v1 row_shr:1
1618 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1619
1620 v_floor_f32_e64_dpp v5, v1 row_shr:15
1621 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1622
1623 v_floor_f32_e64_dpp v5, v1 row_ror:1
1624 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1625
1626 v_floor_f32_e64_dpp v5, v1 row_ror:15
1627 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1628
1629 v_floor_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1630 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1631
1632 v_floor_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1633 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1634
1635 v_floor_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1636 // GFX12: [0x05,0x00,0xa4,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1637
1638 v_floor_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1639 // GFX12: [0xff,0x81,0xa4,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1640
1641 v_fract_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1642 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1643
1644 v_fract_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1645 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1646
1647 v_fract_f16_e64_dpp v5, v1 row_mirror
1648 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1649
1650 v_fract_f16_e64_dpp v5, v1 row_half_mirror
1651 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1652
1653 v_fract_f16_e64_dpp v5, v1 row_shl:1
1654 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1655
1656 v_fract_f16_e64_dpp v5, v1 row_shl:15
1657 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1658
1659 v_fract_f16_e64_dpp v5, v1 row_shr:1
1660 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1661
1662 v_fract_f16_e64_dpp v5, v1 row_shr:15
1663 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1664
1665 v_fract_f16_e64_dpp v5, v1 row_ror:1
1666 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1667
1668 v_fract_f16_e64_dpp v5, v1 row_ror:15
1669 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1670
1671 v_fract_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1672 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1673
1674 v_fract_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1675 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1676
1677 v_fract_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1678 // GFX12: [0x05,0x00,0xdf,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1679
1680 v_fract_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1681 // GFX12: [0xff,0x81,0xdf,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1682
1683 v_fract_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1684 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1685
1686 v_fract_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1687 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1688
1689 v_fract_f32_e64_dpp v5, v1 row_mirror
1690 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1691
1692 v_fract_f32_e64_dpp v5, v1 row_half_mirror
1693 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1694
1695 v_fract_f32_e64_dpp v5, v1 row_shl:1
1696 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1697
1698 v_fract_f32_e64_dpp v5, v1 row_shl:15
1699 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1700
1701 v_fract_f32_e64_dpp v5, v1 row_shr:1
1702 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1703
1704 v_fract_f32_e64_dpp v5, v1 row_shr:15
1705 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1706
1707 v_fract_f32_e64_dpp v5, v1 row_ror:1
1708 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1709
1710 v_fract_f32_e64_dpp v5, v1 row_ror:15
1711 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1712
1713 v_fract_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1714 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1715
1716 v_fract_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1717 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1718
1719 v_fract_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1720 // GFX12: [0x05,0x00,0xa0,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1721
1722 v_fract_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1723 // GFX12: [0xff,0x81,0xa0,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1724
1725 v_frexp_exp_i16_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1726 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1727
1728 v_frexp_exp_i16_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1729 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1730
1731 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_mirror
1732 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1733
1734 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_half_mirror
1735 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1736
1737 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_shl:1
1738 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1739
1740 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_shl:15
1741 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1742
1743 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_shr:1
1744 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1745
1746 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_shr:15
1747 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1748
1749 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_ror:1
1750 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1751
1752 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_ror:15
1753 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1754
1755 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1756 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1757
1758 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1759 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1760
1761 v_frexp_exp_i16_f16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1762 // GFX12: [0x05,0x00,0xda,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1763
1764 v_frexp_exp_i16_f16_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1765 // GFX12: [0xff,0x01,0xda,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1766
1767 v_frexp_exp_i32_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1768 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1769
1770 v_frexp_exp_i32_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1771 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1772
1773 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_mirror
1774 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1775
1776 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_half_mirror
1777 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1778
1779 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_shl:1
1780 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1781
1782 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_shl:15
1783 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1784
1785 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_shr:1
1786 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1787
1788 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_shr:15
1789 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1790
1791 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_ror:1
1792 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1793
1794 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_ror:15
1795 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1796
1797 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1798 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1799
1800 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
1801 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
1802
1803 v_frexp_exp_i32_f32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1804 // GFX12: [0x05,0x00,0xbf,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
1805
1806 v_frexp_exp_i32_f32_e64_dpp v255, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1807 // GFX12: [0xff,0x01,0xbf,0xd5,0xfa,0x00,0x00,0x20,0xff,0x6f,0x05,0x30]
1808
1809 v_frexp_mant_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1810 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1811
1812 v_frexp_mant_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1813 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1814
1815 v_frexp_mant_f16_e64_dpp v5, v1 row_mirror
1816 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1817
1818 v_frexp_mant_f16_e64_dpp v5, v1 row_half_mirror
1819 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1820
1821 v_frexp_mant_f16_e64_dpp v5, v1 row_shl:1
1822 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1823
1824 v_frexp_mant_f16_e64_dpp v5, v1 row_shl:15
1825 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1826
1827 v_frexp_mant_f16_e64_dpp v5, v1 row_shr:1
1828 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1829
1830 v_frexp_mant_f16_e64_dpp v5, v1 row_shr:15
1831 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1832
1833 v_frexp_mant_f16_e64_dpp v5, v1 row_ror:1
1834 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1835
1836 v_frexp_mant_f16_e64_dpp v5, v1 row_ror:15
1837 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1838
1839 v_frexp_mant_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1840 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1841
1842 v_frexp_mant_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1843 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1844
1845 v_frexp_mant_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1846 // GFX12: [0x05,0x00,0xd9,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1847
1848 v_frexp_mant_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1849 // GFX12: [0xff,0x81,0xd9,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1850
1851 v_frexp_mant_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1852 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1853
1854 v_frexp_mant_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1855 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1856
1857 v_frexp_mant_f32_e64_dpp v5, v1 row_mirror
1858 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1859
1860 v_frexp_mant_f32_e64_dpp v5, v1 row_half_mirror
1861 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1862
1863 v_frexp_mant_f32_e64_dpp v5, v1 row_shl:1
1864 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1865
1866 v_frexp_mant_f32_e64_dpp v5, v1 row_shl:15
1867 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1868
1869 v_frexp_mant_f32_e64_dpp v5, v1 row_shr:1
1870 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1871
1872 v_frexp_mant_f32_e64_dpp v5, v1 row_shr:15
1873 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1874
1875 v_frexp_mant_f32_e64_dpp v5, v1 row_ror:1
1876 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1877
1878 v_frexp_mant_f32_e64_dpp v5, v1 row_ror:15
1879 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1880
1881 v_frexp_mant_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1882 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1883
1884 v_frexp_mant_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1885 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1886
1887 v_frexp_mant_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1888 // GFX12: [0x05,0x00,0xc0,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1889
1890 v_frexp_mant_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1891 // GFX12: [0xff,0x81,0xc0,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1892
1893 v_log_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1894 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1895
1896 v_log_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1897 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1898
1899 v_log_f16_e64_dpp v5, v1 row_mirror
1900 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1901
1902 v_log_f16_e64_dpp v5, v1 row_half_mirror
1903 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1904
1905 v_log_f16_e64_dpp v5, v1 row_shl:1
1906 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1907
1908 v_log_f16_e64_dpp v5, v1 row_shl:15
1909 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1910
1911 v_log_f16_e64_dpp v5, v1 row_shr:1
1912 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1913
1914 v_log_f16_e64_dpp v5, v1 row_shr:15
1915 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1916
1917 v_log_f16_e64_dpp v5, v1 row_ror:1
1918 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1919
1920 v_log_f16_e64_dpp v5, v1 row_ror:15
1921 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1922
1923 v_log_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1924 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1925
1926 v_log_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1927 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1928
1929 v_log_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1930 // GFX12: [0x05,0x00,0xd7,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1931
1932 v_log_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1933 // GFX12: [0xff,0x81,0xd7,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1934
1935 v_log_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1936 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1937
1938 v_log_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1939 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1940
1941 v_log_f32_e64_dpp v5, v1 row_mirror
1942 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1943
1944 v_log_f32_e64_dpp v5, v1 row_half_mirror
1945 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1946
1947 v_log_f32_e64_dpp v5, v1 row_shl:1
1948 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1949
1950 v_log_f32_e64_dpp v5, v1 row_shl:15
1951 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1952
1953 v_log_f32_e64_dpp v5, v1 row_shr:1
1954 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1955
1956 v_log_f32_e64_dpp v5, v1 row_shr:15
1957 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
1958
1959 v_log_f32_e64_dpp v5, v1 row_ror:1
1960 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
1961
1962 v_log_f32_e64_dpp v5, v1 row_ror:15
1963 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
1964
1965 v_log_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
1966 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
1967
1968 v_log_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1969 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
1970
1971 v_log_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1972 // GFX12: [0x05,0x00,0xa7,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
1973
1974 v_log_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1975 // GFX12: [0xff,0x81,0xa7,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
1976
1977 v_mov_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
1978 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
1979
1980 v_mov_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
1981 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
1982
1983 v_mov_b32_e64_dpp v5, v1 row_mirror
1984 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
1985
1986 v_mov_b32_e64_dpp v5, v1 row_half_mirror
1987 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
1988
1989 v_mov_b32_e64_dpp v5, v1 row_shl:1
1990 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
1991
1992 v_mov_b32_e64_dpp v5, v1 row_shl:15
1993 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
1994
1995 v_mov_b32_e64_dpp v5, v1 row_shr:1
1996 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
1997
1998 v_mov_b32_e64_dpp v5, v1 row_shr:15
1999 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2000
2001 v_mov_b32_e64_dpp v5, v1 row_ror:1
2002 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2003
2004 v_mov_b32_e64_dpp v5, v1 row_ror:15
2005 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2006
2007 v_mov_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2008 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2009
2010 v_mov_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
2011 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
2012
2013 v_mov_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2014 // GFX12: [0x05,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
2015
2016 v_mov_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2017 // GFX12: [0xff,0x00,0x81,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
2018
2019 v_movrels_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2020 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2021
2022 v_movrels_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2023 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2024
2025 v_movrels_b32_e64_dpp v5, v1 row_mirror
2026 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2027
2028 v_movrels_b32_e64_dpp v5, v1 row_half_mirror
2029 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2030
2031 v_movrels_b32_e64_dpp v5, v1 row_shl:1
2032 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2033
2034 v_movrels_b32_e64_dpp v5, v1 row_shl:15
2035 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2036
2037 v_movrels_b32_e64_dpp v5, v1 row_shr:1
2038 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2039
2040 v_movrels_b32_e64_dpp v5, v1 row_shr:15
2041 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2042
2043 v_movrels_b32_e64_dpp v5, v1 row_ror:1
2044 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2045
2046 v_movrels_b32_e64_dpp v5, v1 row_ror:15
2047 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2048
2049 v_movrels_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2050 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2051
2052 v_movrels_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
2053 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
2054
2055 v_movrels_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2056 // GFX12: [0x05,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
2057
2058 v_movrels_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2059 // GFX12: [0xff,0x00,0xc3,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
2060
2061 v_not_b16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2062 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2063
2064 v_not_b16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2065 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2066
2067 v_not_b16_e64_dpp v5, v1 row_mirror
2068 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2069
2070 v_not_b16_e64_dpp v5, v1 row_half_mirror
2071 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2072
2073 v_not_b16_e64_dpp v5, v1 row_shl:1
2074 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2075
2076 v_not_b16_e64_dpp v5, v1 row_shl:15
2077 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2078
2079 v_not_b16_e64_dpp v5, v1 row_shr:1
2080 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2081
2082 v_not_b16_e64_dpp v5, v1 row_shr:15
2083 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2084
2085 v_not_b16_e64_dpp v5, v1 row_ror:1
2086 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2087
2088 v_not_b16_e64_dpp v5, v1 row_ror:15
2089 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2090
2091 v_not_b16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2092 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2093
2094 v_not_b16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
2095 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
2096
2097 v_not_b16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2098 // GFX12: [0x05,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
2099
2100 v_not_b16_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2101 // GFX12: [0xff,0x00,0xe9,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
2102
2103 v_not_b32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2104 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2105
2106 v_not_b32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2107 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2108
2109 v_not_b32_e64_dpp v5, v1 row_mirror
2110 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2111
2112 v_not_b32_e64_dpp v5, v1 row_half_mirror
2113 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2114
2115 v_not_b32_e64_dpp v5, v1 row_shl:1
2116 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2117
2118 v_not_b32_e64_dpp v5, v1 row_shl:15
2119 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2120
2121 v_not_b32_e64_dpp v5, v1 row_shr:1
2122 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2123
2124 v_not_b32_e64_dpp v5, v1 row_shr:15
2125 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2126
2127 v_not_b32_e64_dpp v5, v1 row_ror:1
2128 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2129
2130 v_not_b32_e64_dpp v5, v1 row_ror:15
2131 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2132
2133 v_not_b32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2134 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2135
2136 v_not_b32_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
2137 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
2138
2139 v_not_b32_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2140 // GFX12: [0x05,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
2141
2142 v_not_b32_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2143 // GFX12: [0xff,0x00,0xb7,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
2144
2145 v_rcp_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2146 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2147
2148 v_rcp_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2149 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2150
2151 v_rcp_f16_e64_dpp v5, v1 row_mirror
2152 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2153
2154 v_rcp_f16_e64_dpp v5, v1 row_half_mirror
2155 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2156
2157 v_rcp_f16_e64_dpp v5, v1 row_shl:1
2158 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2159
2160 v_rcp_f16_e64_dpp v5, v1 row_shl:15
2161 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2162
2163 v_rcp_f16_e64_dpp v5, v1 row_shr:1
2164 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2165
2166 v_rcp_f16_e64_dpp v5, v1 row_shr:15
2167 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2168
2169 v_rcp_f16_e64_dpp v5, v1 row_ror:1
2170 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2171
2172 v_rcp_f16_e64_dpp v5, v1 row_ror:15
2173 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2174
2175 v_rcp_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2176 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2177
2178 v_rcp_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2179 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2180
2181 v_rcp_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2182 // GFX12: [0x05,0x00,0xd4,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2183
2184 v_rcp_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2185 // GFX12: [0xff,0x81,0xd4,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2186
2187 v_rcp_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2188 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2189
2190 v_rcp_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2191 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2192
2193 v_rcp_f32_e64_dpp v5, v1 row_mirror
2194 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2195
2196 v_rcp_f32_e64_dpp v5, v1 row_half_mirror
2197 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2198
2199 v_rcp_f32_e64_dpp v5, v1 row_shl:1
2200 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2201
2202 v_rcp_f32_e64_dpp v5, v1 row_shl:15
2203 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2204
2205 v_rcp_f32_e64_dpp v5, v1 row_shr:1
2206 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2207
2208 v_rcp_f32_e64_dpp v5, v1 row_shr:15
2209 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2210
2211 v_rcp_f32_e64_dpp v5, v1 row_ror:1
2212 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2213
2214 v_rcp_f32_e64_dpp v5, v1 row_ror:15
2215 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2216
2217 v_rcp_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2218 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2219
2220 v_rcp_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2221 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2222
2223 v_rcp_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2224 // GFX12: [0x05,0x00,0xaa,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2225
2226 v_rcp_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2227 // GFX12: [0xff,0x81,0xaa,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2228
2229 v_rcp_iflag_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2230 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2231
2232 v_rcp_iflag_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2233 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2234
2235 v_rcp_iflag_f32_e64_dpp v5, v1 row_mirror
2236 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2237
2238 v_rcp_iflag_f32_e64_dpp v5, v1 row_half_mirror
2239 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2240
2241 v_rcp_iflag_f32_e64_dpp v5, v1 row_shl:1
2242 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2243
2244 v_rcp_iflag_f32_e64_dpp v5, v1 row_shl:15
2245 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2246
2247 v_rcp_iflag_f32_e64_dpp v5, v1 row_shr:1
2248 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2249
2250 v_rcp_iflag_f32_e64_dpp v5, v1 row_shr:15
2251 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2252
2253 v_rcp_iflag_f32_e64_dpp v5, v1 row_ror:1
2254 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2255
2256 v_rcp_iflag_f32_e64_dpp v5, v1 row_ror:15
2257 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2258
2259 v_rcp_iflag_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2260 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2261
2262 v_rcp_iflag_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2263 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2264
2265 v_rcp_iflag_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2266 // GFX12: [0x05,0x00,0xab,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2267
2268 v_rcp_iflag_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2269 // GFX12: [0xff,0x81,0xab,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2270
2271 v_rndne_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2272 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2273
2274 v_rndne_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2275 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2276
2277 v_rndne_f16_e64_dpp v5, v1 row_mirror
2278 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2279
2280 v_rndne_f16_e64_dpp v5, v1 row_half_mirror
2281 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2282
2283 v_rndne_f16_e64_dpp v5, v1 row_shl:1
2284 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2285
2286 v_rndne_f16_e64_dpp v5, v1 row_shl:15
2287 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2288
2289 v_rndne_f16_e64_dpp v5, v1 row_shr:1
2290 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2291
2292 v_rndne_f16_e64_dpp v5, v1 row_shr:15
2293 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2294
2295 v_rndne_f16_e64_dpp v5, v1 row_ror:1
2296 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2297
2298 v_rndne_f16_e64_dpp v5, v1 row_ror:15
2299 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2300
2301 v_rndne_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2302 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2303
2304 v_rndne_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2305 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2306
2307 v_rndne_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2308 // GFX12: [0x05,0x00,0xde,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2309
2310 v_rndne_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2311 // GFX12: [0xff,0x81,0xde,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2312
2313 v_rndne_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2314 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2315
2316 v_rndne_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2317 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2318
2319 v_rndne_f32_e64_dpp v5, v1 row_mirror
2320 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2321
2322 v_rndne_f32_e64_dpp v5, v1 row_half_mirror
2323 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2324
2325 v_rndne_f32_e64_dpp v5, v1 row_shl:1
2326 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2327
2328 v_rndne_f32_e64_dpp v5, v1 row_shl:15
2329 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2330
2331 v_rndne_f32_e64_dpp v5, v1 row_shr:1
2332 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2333
2334 v_rndne_f32_e64_dpp v5, v1 row_shr:15
2335 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2336
2337 v_rndne_f32_e64_dpp v5, v1 row_ror:1
2338 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2339
2340 v_rndne_f32_e64_dpp v5, v1 row_ror:15
2341 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2342
2343 v_rndne_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2344 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2345
2346 v_rndne_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2347 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2348
2349 v_rndne_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2350 // GFX12: [0x05,0x00,0xa3,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2351
2352 v_rndne_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2353 // GFX12: [0xff,0x81,0xa3,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2354
2355 v_rsq_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2356 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2357
2358 v_rsq_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2359 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2360
2361 v_rsq_f16_e64_dpp v5, v1 row_mirror
2362 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2363
2364 v_rsq_f16_e64_dpp v5, v1 row_half_mirror
2365 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2366
2367 v_rsq_f16_e64_dpp v5, v1 row_shl:1
2368 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2369
2370 v_rsq_f16_e64_dpp v5, v1 row_shl:15
2371 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2372
2373 v_rsq_f16_e64_dpp v5, v1 row_shr:1
2374 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2375
2376 v_rsq_f16_e64_dpp v5, v1 row_shr:15
2377 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2378
2379 v_rsq_f16_e64_dpp v5, v1 row_ror:1
2380 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2381
2382 v_rsq_f16_e64_dpp v5, v1 row_ror:15
2383 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2384
2385 v_rsq_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2386 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2387
2388 v_rsq_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2389 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2390
2391 v_rsq_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2392 // GFX12: [0x05,0x00,0xd6,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2393
2394 v_rsq_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2395 // GFX12: [0xff,0x81,0xd6,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2396
2397 v_rsq_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2398 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2399
2400 v_rsq_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2401 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2402
2403 v_rsq_f32_e64_dpp v5, v1 row_mirror
2404 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2405
2406 v_rsq_f32_e64_dpp v5, v1 row_half_mirror
2407 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2408
2409 v_rsq_f32_e64_dpp v5, v1 row_shl:1
2410 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2411
2412 v_rsq_f32_e64_dpp v5, v1 row_shl:15
2413 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2414
2415 v_rsq_f32_e64_dpp v5, v1 row_shr:1
2416 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2417
2418 v_rsq_f32_e64_dpp v5, v1 row_shr:15
2419 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2420
2421 v_rsq_f32_e64_dpp v5, v1 row_ror:1
2422 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2423
2424 v_rsq_f32_e64_dpp v5, v1 row_ror:15
2425 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2426
2427 v_rsq_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2428 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2429
2430 v_rsq_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2431 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2432
2433 v_rsq_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2434 // GFX12: [0x05,0x00,0xae,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2435
2436 v_rsq_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2437 // GFX12: [0xff,0x81,0xae,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2438
2439 v_sat_pk_u8_i16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2440 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2441
2442 v_sat_pk_u8_i16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2443 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2444
2445 v_sat_pk_u8_i16_e64_dpp v5, v1 row_mirror
2446 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2447
2448 v_sat_pk_u8_i16_e64_dpp v5, v1 row_half_mirror
2449 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2450
2451 v_sat_pk_u8_i16_e64_dpp v5, v1 row_shl:1
2452 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2453
2454 v_sat_pk_u8_i16_e64_dpp v5, v1 row_shl:15
2455 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2456
2457 v_sat_pk_u8_i16_e64_dpp v5, v1 row_shr:1
2458 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2459
2460 v_sat_pk_u8_i16_e64_dpp v5, v1 row_shr:15
2461 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2462
2463 v_sat_pk_u8_i16_e64_dpp v5, v1 row_ror:1
2464 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2465
2466 v_sat_pk_u8_i16_e64_dpp v5, v1 row_ror:15
2467 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2468
2469 v_sat_pk_u8_i16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2470 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2471
2472 v_sat_pk_u8_i16_e64_dpp v5, v1 row_share:15 row_mask:0x0 bank_mask:0x1
2473 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x5f,0x01,0x01]
2474
2475 v_sat_pk_u8_i16_e64_dpp v5, v1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2476 // GFX12: [0x05,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0x01,0x60,0x09,0x13]
2477
2478 v_sat_pk_u8_i16_e64_dpp v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2479 // GFX12: [0xff,0x00,0xe2,0xd5,0xfa,0x00,0x00,0x00,0xff,0x6f,0x05,0x30]
2480
2481 v_sin_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2482 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2483
2484 v_sin_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2485 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2486
2487 v_sin_f16_e64_dpp v5, v1 row_mirror
2488 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2489
2490 v_sin_f16_e64_dpp v5, v1 row_half_mirror
2491 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2492
2493 v_sin_f16_e64_dpp v5, v1 row_shl:1
2494 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2495
2496 v_sin_f16_e64_dpp v5, v1 row_shl:15
2497 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2498
2499 v_sin_f16_e64_dpp v5, v1 row_shr:1
2500 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2501
2502 v_sin_f16_e64_dpp v5, v1 row_shr:15
2503 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2504
2505 v_sin_f16_e64_dpp v5, v1 row_ror:1
2506 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2507
2508 v_sin_f16_e64_dpp v5, v1 row_ror:15
2509 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2510
2511 v_sin_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2512 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2513
2514 v_sin_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2515 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2516
2517 v_sin_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2518 // GFX12: [0x05,0x00,0xe0,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2519
2520 v_sin_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2521 // GFX12: [0xff,0x81,0xe0,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2522
2523 v_sin_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2524 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2525
2526 v_sin_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2527 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2528
2529 v_sin_f32_e64_dpp v5, v1 row_mirror
2530 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2531
2532 v_sin_f32_e64_dpp v5, v1 row_half_mirror
2533 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2534
2535 v_sin_f32_e64_dpp v5, v1 row_shl:1
2536 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2537
2538 v_sin_f32_e64_dpp v5, v1 row_shl:15
2539 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2540
2541 v_sin_f32_e64_dpp v5, v1 row_shr:1
2542 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2543
2544 v_sin_f32_e64_dpp v5, v1 row_shr:15
2545 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2546
2547 v_sin_f32_e64_dpp v5, v1 row_ror:1
2548 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2549
2550 v_sin_f32_e64_dpp v5, v1 row_ror:15
2551 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2552
2553 v_sin_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2554 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2555
2556 v_sin_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2557 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2558
2559 v_sin_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2560 // GFX12: [0x05,0x00,0xb5,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2561
2562 v_sin_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2563 // GFX12: [0xff,0x81,0xb5,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2564
2565 v_sqrt_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2566 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2567
2568 v_sqrt_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2569 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2570
2571 v_sqrt_f16_e64_dpp v5, v1 row_mirror
2572 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2573
2574 v_sqrt_f16_e64_dpp v5, v1 row_half_mirror
2575 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2576
2577 v_sqrt_f16_e64_dpp v5, v1 row_shl:1
2578 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2579
2580 v_sqrt_f16_e64_dpp v5, v1 row_shl:15
2581 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2582
2583 v_sqrt_f16_e64_dpp v5, v1 row_shr:1
2584 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2585
2586 v_sqrt_f16_e64_dpp v5, v1 row_shr:15
2587 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2588
2589 v_sqrt_f16_e64_dpp v5, v1 row_ror:1
2590 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2591
2592 v_sqrt_f16_e64_dpp v5, v1 row_ror:15
2593 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2594
2595 v_sqrt_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2596 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2597
2598 v_sqrt_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2599 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2600
2601 v_sqrt_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2602 // GFX12: [0x05,0x00,0xd5,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2603
2604 v_sqrt_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2605 // GFX12: [0xff,0x81,0xd5,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2606
2607 v_sqrt_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2608 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2609
2610 v_sqrt_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2611 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2612
2613 v_sqrt_f32_e64_dpp v5, v1 row_mirror
2614 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2615
2616 v_sqrt_f32_e64_dpp v5, v1 row_half_mirror
2617 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2618
2619 v_sqrt_f32_e64_dpp v5, v1 row_shl:1
2620 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2621
2622 v_sqrt_f32_e64_dpp v5, v1 row_shl:15
2623 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2624
2625 v_sqrt_f32_e64_dpp v5, v1 row_shr:1
2626 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2627
2628 v_sqrt_f32_e64_dpp v5, v1 row_shr:15
2629 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2630
2631 v_sqrt_f32_e64_dpp v5, v1 row_ror:1
2632 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2633
2634 v_sqrt_f32_e64_dpp v5, v1 row_ror:15
2635 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2636
2637 v_sqrt_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2638 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2639
2640 v_sqrt_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2641 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2642
2643 v_sqrt_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2644 // GFX12: [0x05,0x00,0xb3,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2645
2646 v_sqrt_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2647 // GFX12: [0xff,0x81,0xb3,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2648
2649 v_trunc_f16_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2650 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2651
2652 v_trunc_f16_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2653 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2654
2655 v_trunc_f16_e64_dpp v5, v1 row_mirror
2656 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2657
2658 v_trunc_f16_e64_dpp v5, v1 row_half_mirror
2659 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2660
2661 v_trunc_f16_e64_dpp v5, v1 row_shl:1
2662 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2663
2664 v_trunc_f16_e64_dpp v5, v1 row_shl:15
2665 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2666
2667 v_trunc_f16_e64_dpp v5, v1 row_shr:1
2668 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2669
2670 v_trunc_f16_e64_dpp v5, v1 row_shr:15
2671 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2672
2673 v_trunc_f16_e64_dpp v5, v1 row_ror:1
2674 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2675
2676 v_trunc_f16_e64_dpp v5, v1 row_ror:15
2677 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2678
2679 v_trunc_f16_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2680 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2681
2682 v_trunc_f16_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2683 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2684
2685 v_trunc_f16_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2686 // GFX12: [0x05,0x00,0xdd,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2687
2688 v_trunc_f16_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2689 // GFX12: [0xff,0x81,0xdd,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]
2690
2691 v_trunc_f32_e64_dpp v5, v1 quad_perm:[3,2,1,0]
2692 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1b,0x00,0xff]
2693
2694 v_trunc_f32_e64_dpp v5, v1 quad_perm:[0,1,2,3]
2695 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0xe4,0x00,0xff]
2696
2697 v_trunc_f32_e64_dpp v5, v1 row_mirror
2698 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x40,0x01,0xff]
2699
2700 v_trunc_f32_e64_dpp v5, v1 row_half_mirror
2701 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x41,0x01,0xff]
2702
2703 v_trunc_f32_e64_dpp v5, v1 row_shl:1
2704 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x01,0x01,0xff]
2705
2706 v_trunc_f32_e64_dpp v5, v1 row_shl:15
2707 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x0f,0x01,0xff]
2708
2709 v_trunc_f32_e64_dpp v5, v1 row_shr:1
2710 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x11,0x01,0xff]
2711
2712 v_trunc_f32_e64_dpp v5, v1 row_shr:15
2713 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x1f,0x01,0xff]
2714
2715 v_trunc_f32_e64_dpp v5, v1 row_ror:1
2716 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x21,0x01,0xff]
2717
2718 v_trunc_f32_e64_dpp v5, v1 row_ror:15
2719 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x2f,0x01,0xff]
2720
2721 v_trunc_f32_e64_dpp v5, v1 row_share:0 row_mask:0xf bank_mask:0xf
2722 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x00,0x01,0x50,0x01,0xff]
2723
2724 v_trunc_f32_e64_dpp v5, v1 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2725 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x08,0x01,0x5f,0x01,0x01]
2726
2727 v_trunc_f32_e64_dpp v5, v1 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2728 // GFX12: [0x05,0x00,0xa1,0xd5,0xfa,0x00,0x00,0x10,0x01,0x60,0x09,0x13]
2729
2730 v_trunc_f32_e64_dpp v255, -|v255| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2731 // GFX12: [0xff,0x81,0xa1,0xd5,0xfa,0x00,0x00,0x38,0xff,0x6f,0x05,0x30]