llvm/test/MC/AMDGPU/gfx12_asm_vop3_dpp16-fake16.s

   1 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize32,-real-true16 -show-encoding %s | FileCheck --check-prefixes=GFX12,W32 %s
   2 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize64,-real-true16 -show-encoding %s | FileCheck --check-prefixes=GFX12,W64 %s
   3 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize32,-real-true16 %s 2>&1 | FileCheck --check-prefixes=GFX12-ERR,W32-ERR --implicit-check-not=error: %s
   4 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize64,-real-true16 %s 2>&1 | FileCheck --check-prefixes=GFX12-ERR,W64-ERR --implicit-check-not=error: %s
   5
   6 v_add3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
   7 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
   8
   9 v_add3_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
  10 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
  11
  12 v_add3_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
  13 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
  14
  15 v_add3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
  16 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
  17
  18 v_add3_u32_e64_dpp v5, v1, v2, v3 row_mirror
  19 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
  20
  21 v_add3_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
  22 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
  23
  24 v_add3_u32_e64_dpp v5, v1, v2, s105 row_shl:1
  25 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
  26
  27 v_add3_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
  28 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
  29
  30 v_add3_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
  31 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
  32
  33 v_add3_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
  34 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
  35
  36 v_add3_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
  37 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
  38
  39 v_add3_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
  40 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
  41
  42 v_add3_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
  43 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
  44
  45 v_add3_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
  46 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
  47
  48 v_add3_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
  49 // GFX12: [0x05,0x00,0x55,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
  50
  51 v_add3_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
  52 // GFX12: [0xff,0x00,0x55,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
  53
  54 v_add_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[3,2,1,0]
  55 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
  56 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  57
  58 v_add_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[0,1,2,3]
  59 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
  60 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  61
  62 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_mirror
  63 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
  64 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  65
  66 v_add_co_u32_e64_dpp v5, s6, v1, s2 row_mirror
  67 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x00,0x00,0x01,0x40,0x01,0xff]
  68 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  69
  70 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_half_mirror
  71 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
  72 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  73
  74 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_shl:1
  75 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
  76 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  77
  78 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_shl:15
  79 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
  80 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  81
  82 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_shr:1
  83 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
  84 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  85
  86 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_shr:15
  87 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
  88 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  89
  90 v_add_co_u32_e64_dpp v5, s6, v1, v2 row_ror:1
  91 // W32: [0x05,0x06,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
  92 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  93
  94 v_add_co_u32_e64_dpp v5, s105, v1, v2 row_ror:15
  95 // W32: [0x05,0x69,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
  96 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
  97
  98 v_add_co_u32_e64_dpp v5, vcc_lo, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
  99 // W32: [0x05,0x6a,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 100 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 101
 102 v_add_co_u32_e64_dpp v5, vcc_hi, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 103 // W32: [0x05,0x6b,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 104 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 105
 106 v_add_co_u32_e64_dpp v5, ttmp15, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 107 // W32: [0x05,0x7b,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 108 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 109
 110 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[3,2,1,0]
 111 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 112 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 113
 114 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[0,1,2,3]
 115 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 116 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 117
 118 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_mirror
 119 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 120 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 121
 122 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_half_mirror
 123 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 124 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 125
 126 v_add_co_u32_e64_dpp v5, s[12:13], v1, s2 row_half_mirror
 127 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x00,0x00,0x01,0x41,0x01,0xff]
 128 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 129
 130 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:1
 131 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 132 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 133
 134 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:15
 135 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 136 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 137
 138 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:1
 139 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 140 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 141
 142 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:15
 143 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 144 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 145
 146 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:1
 147 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 148 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 149
 150 v_add_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:15
 151 // W64: [0x05,0x0c,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 152 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 153
 154 v_add_co_u32_e64_dpp v5, s[104:105], v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 155 // W64: [0x05,0x68,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 156 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 157
 158 v_add_co_u32_e64_dpp v5, vcc, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 159 // W64: [0x05,0x6a,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 160 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 161
 162 v_add_co_u32_e64_dpp v5, ttmp[14:15], v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 163 // W64: [0x05,0x7a,0x00,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 164 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 165
 166 v_add_co_u32_e64_dpp v255, null, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 167 // GFX12: [0xff,0xfc,0x00,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 168
 169 v_add_lshl_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 170 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 171
 172 v_add_lshl_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 173 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 174
 175 v_add_lshl_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 176 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 177
 178 v_add_lshl_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 179 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 180
 181 v_add_lshl_u32_e64_dpp v5, v1, v2, v3 row_mirror
 182 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 183
 184 v_add_lshl_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
 185 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 186
 187 v_add_lshl_u32_e64_dpp v5, v1, v2, s105 row_shl:1
 188 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 189
 190 v_add_lshl_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 191 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 192
 193 v_add_lshl_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 194 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 195
 196 v_add_lshl_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
 197 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
 198
 199 v_add_lshl_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
 200 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
 201
 202 v_add_lshl_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
 203 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
 204
 205 v_add_lshl_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
 206 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
 207
 208 v_add_lshl_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
 209 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
 210
 211 v_add_lshl_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 212 // GFX12: [0x05,0x00,0x47,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
 213
 214 v_add_lshl_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 215 // GFX12: [0xff,0x00,0x47,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 216
 217 v_add_nc_i16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 218 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 219
 220 v_add_nc_i16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 221 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 222
 223 v_add_nc_i16_e64_dpp v5, v1, v2 row_mirror
 224 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 225
 226 v_add_nc_i16_e64_dpp v5, v1, v2 row_half_mirror
 227 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 228
 229 v_add_nc_i16_e64_dpp v5, v1, v2 row_shl:1
 230 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 231
 232 v_add_nc_i16_e64_dpp v5, v1, v2 row_shl:15
 233 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 234
 235 v_add_nc_i16_e64_dpp v5, v1, v2 row_shr:1
 236 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 237
 238 v_add_nc_i16_e64_dpp v5, v1, v2 row_shr:15
 239 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 240
 241 v_add_nc_i16_e64_dpp v5, v1, v2 row_ror:1
 242 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 243
 244 v_add_nc_i16_e64_dpp v5, v1, v2 row_ror:15
 245 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 246
 247 v_add_nc_i16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 248 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 249
 250 v_add_nc_i16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 251 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 252
 253 v_add_nc_i16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 254 // GFX12: [0x05,0x00,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 255
 256 v_add_nc_i16_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 257 // GFX12: [0xff,0x80,0x0d,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 258
 259 v_add_nc_i32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 260 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 261
 262 v_add_nc_i32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 263 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 264
 265 v_add_nc_i32_e64_dpp v5, v1, v2 row_mirror
 266 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 267
 268 v_add_nc_i32_e64_dpp v5, v1, v2 row_half_mirror
 269 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 270
 271 v_add_nc_i32_e64_dpp v5, v1, v2 row_shl:1
 272 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 273
 274 v_add_nc_i32_e64_dpp v5, v1, v2 row_shl:15
 275 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 276
 277 v_add_nc_i32_e64_dpp v5, v1, v2 row_shr:1
 278 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 279
 280 v_add_nc_i32_e64_dpp v5, v1, v2 row_shr:15
 281 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 282
 283 v_add_nc_i32_e64_dpp v5, v1, v2 row_ror:1
 284 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 285
 286 v_add_nc_i32_e64_dpp v5, v1, v2 row_ror:15
 287 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 288
 289 v_add_nc_i32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 290 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 291
 292 v_add_nc_i32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 293 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 294
 295 v_add_nc_i32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 296 // GFX12: [0x05,0x00,0x26,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 297
 298 v_add_nc_i32_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 299 // GFX12: [0xff,0x80,0x26,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 300
 301 v_add_nc_u16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 302 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 303
 304 v_add_nc_u16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 305 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 306
 307 v_add_nc_u16_e64_dpp v5, v1, v2 row_mirror
 308 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 309
 310 v_add_nc_u16_e64_dpp v5, v1, v2 row_half_mirror
 311 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 312
 313 v_add_nc_u16_e64_dpp v5, v1, v2 row_shl:1
 314 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 315
 316 v_add_nc_u16_e64_dpp v5, v1, v2 row_shl:15
 317 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 318
 319 v_add_nc_u16_e64_dpp v5, v1, v2 row_shr:1
 320 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 321
 322 v_add_nc_u16_e64_dpp v5, v1, v2 row_shr:15
 323 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 324
 325 v_add_nc_u16_e64_dpp v5, v1, v2 row_ror:1
 326 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 327
 328 v_add_nc_u16_e64_dpp v5, v1, v2 row_ror:15
 329 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 330
 331 v_add_nc_u16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 332 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 333
 334 v_add_nc_u16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 335 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 336
 337 v_add_nc_u16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 338 // GFX12: [0x05,0x00,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 339
 340 v_add_nc_u16_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 341 // GFX12: [0xff,0x80,0x03,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 342
 343 v_alignbit_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 344 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 345
 346 v_alignbit_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 347 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 348
 349 v_alignbit_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 350 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 351
 352 v_alignbit_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 353 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 354
 355 v_alignbit_b32_e64_dpp v5, v1, v2, v3 row_mirror
 356 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 357
 358 v_alignbit_b32_e64_dpp v5, v1, v2, v3 row_half_mirror
 359 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
 360
 361 v_alignbit_b32_e64_dpp v5, v1, v2, v255 row_shl:1
 362 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
 363
 364 v_alignbit_b32_e64_dpp v5, v1, v2, s105 row_shl:15
 365 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
 366
 367 v_alignbit_b32_e64_dpp v5, v1, v2, vcc_hi row_shr:1
 368 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
 369
 370 v_alignbit_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:15
 371 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
 372
 373 v_alignbit_b32_e64_dpp v5, v1, v2, ttmp15 row_ror:1
 374 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
 375
 376 v_alignbit_b32_e64_dpp v5, v1, v2, exec_hi row_ror:15
 377 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
 378
 379 v_alignbit_b32_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
 380 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
 381
 382 v_alignbit_b32_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
 383 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
 384
 385 v_alignbit_b32_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 386 // GFX12: [0x05,0x00,0x16,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
 387
 388 v_alignbit_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 389 // GFX12: [0xff,0x00,0x16,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 390
 391 v_alignbyte_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 392 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 393
 394 v_alignbyte_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 395 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 396
 397 v_alignbyte_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 398 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 399
 400 v_alignbyte_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 401 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 402
 403 v_alignbyte_b32_e64_dpp v5, v1, v2, v3 row_mirror
 404 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 405
 406 v_alignbyte_b32_e64_dpp v5, v1, v2, v3 row_half_mirror
 407 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
 408
 409 v_alignbyte_b32_e64_dpp v5, v1, v2, v255 row_shl:1
 410 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
 411
 412 v_alignbyte_b32_e64_dpp v5, v1, v2, s105 row_shl:15
 413 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
 414
 415 v_alignbyte_b32_e64_dpp v5, v1, v2, vcc_hi row_shr:1
 416 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
 417
 418 v_alignbyte_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:15
 419 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
 420
 421 v_alignbyte_b32_e64_dpp v5, v1, v2, ttmp15 row_ror:1
 422 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
 423
 424 v_alignbyte_b32_e64_dpp v5, v1, v2, exec_hi row_ror:15
 425 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
 426
 427 v_alignbyte_b32_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
 428 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
 429
 430 v_alignbyte_b32_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
 431 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
 432
 433 v_alignbyte_b32_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 434 // GFX12: [0x05,0x00,0x17,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
 435
 436 v_alignbyte_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 437 // GFX12: [0xff,0x00,0x17,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 438
 439 v_and_b16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 440 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 441
 442 v_and_b16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 443 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 444
 445 v_and_b16_e64_dpp v5, v1, v2 row_mirror
 446 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 447
 448 v_and_b16_e64_dpp v5, v1, v2 row_half_mirror
 449 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 450
 451 v_and_b16_e64_dpp v5, v1, v2 row_shl:1
 452 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 453
 454 v_and_b16_e64_dpp v5, v1, v2 row_shl:15
 455 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 456
 457 v_and_b16_e64_dpp v5, v1, v2 row_shr:1
 458 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 459
 460 v_and_b16_e64_dpp v5, v1, v2 row_shr:15
 461 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 462
 463 v_and_b16_e64_dpp v5, v1, v2 row_ror:1
 464 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 465
 466 v_and_b16_e64_dpp v5, v1, v2 row_ror:15
 467 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 468
 469 v_and_b16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 470 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 471
 472 v_and_b16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 473 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 474
 475 v_and_b16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 476 // GFX12: [0x05,0x00,0x62,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 477
 478 v_and_b16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 479 // GFX12: [0xff,0x00,0x62,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 480
 481 v_and_or_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 482 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 483
 484 v_and_or_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 485 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 486
 487 v_and_or_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 488 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 489
 490 v_and_or_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 491 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 492
 493 v_and_or_b32_e64_dpp v5, v1, v2, v3 row_mirror
 494 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 495
 496 v_and_or_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
 497 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 498
 499 v_and_or_b32_e64_dpp v5, v1, v2, s105 row_shl:1
 500 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 501
 502 v_and_or_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 503 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 504
 505 v_and_or_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 506 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 507
 508 v_and_or_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
 509 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
 510
 511 v_and_or_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
 512 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
 513
 514 v_and_or_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
 515 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
 516
 517 v_and_or_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
 518 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
 519
 520 v_and_or_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
 521 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
 522
 523 v_and_or_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 524 // GFX12: [0x05,0x00,0x57,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
 525
 526 v_and_or_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 527 // GFX12: [0xff,0x00,0x57,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 528
 529 v_ashrrev_i16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 530 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 531
 532 v_ashrrev_i16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 533 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 534
 535 v_ashrrev_i16_e64_dpp v5, v1, v2 row_mirror
 536 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 537
 538 v_ashrrev_i16_e64_dpp v5, v1, v2 row_half_mirror
 539 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 540
 541 v_ashrrev_i16_e64_dpp v5, v1, v2 row_shl:1
 542 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 543
 544 v_ashrrev_i16_e64_dpp v5, v1, v2 row_shl:15
 545 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 546
 547 v_ashrrev_i16_e64_dpp v5, v1, v2 row_shr:1
 548 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 549
 550 v_ashrrev_i16_e64_dpp v5, v1, v2 row_shr:15
 551 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 552
 553 v_ashrrev_i16_e64_dpp v5, v1, v2 row_ror:1
 554 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 555
 556 v_ashrrev_i16_e64_dpp v5, v1, v2 row_ror:15
 557 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 558
 559 v_ashrrev_i16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 560 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 561
 562 v_ashrrev_i16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 563 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 564
 565 v_ashrrev_i16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 566 // GFX12: [0x05,0x00,0x3a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 567
 568 v_ashrrev_i16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 569 // GFX12: [0xff,0x00,0x3a,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 570
 571 v_bcnt_u32_b32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 572 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 573
 574 v_bcnt_u32_b32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 575 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 576
 577 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_mirror
 578 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 579
 580 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_half_mirror
 581 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 582
 583 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_shl:1
 584 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 585
 586 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_shl:15
 587 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 588
 589 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_shr:1
 590 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 591
 592 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_shr:15
 593 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 594
 595 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_ror:1
 596 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 597
 598 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_ror:15
 599 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 600
 601 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 602 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 603
 604 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 605 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 606
 607 v_bcnt_u32_b32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 608 // GFX12: [0x05,0x00,0x1e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 609
 610 v_bcnt_u32_b32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 611 // GFX12: [0xff,0x00,0x1e,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 612
 613 v_bfe_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 614 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 615
 616 v_bfe_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 617 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 618
 619 v_bfe_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 620 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 621
 622 v_bfe_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 623 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 624
 625 v_bfe_i32_e64_dpp v5, v1, v2, v3 row_mirror
 626 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 627
 628 v_bfe_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
 629 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 630
 631 v_bfe_i32_e64_dpp v5, v1, v2, s105 row_shl:1
 632 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 633
 634 v_bfe_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 635 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 636
 637 v_bfe_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 638 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 639
 640 v_bfe_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
 641 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
 642
 643 v_bfe_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
 644 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
 645
 646 v_bfe_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
 647 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
 648
 649 v_bfe_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
 650 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
 651
 652 v_bfe_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
 653 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
 654
 655 v_bfe_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 656 // GFX12: [0x05,0x00,0x11,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
 657
 658 v_bfe_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 659 // GFX12: [0xff,0x00,0x11,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 660
 661 v_bfe_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 662 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 663
 664 v_bfe_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 665 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 666
 667 v_bfe_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 668 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 669
 670 v_bfe_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 671 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 672
 673 v_bfe_u32_e64_dpp v5, v1, v2, v3 row_mirror
 674 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 675
 676 v_bfe_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
 677 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 678
 679 v_bfe_u32_e64_dpp v5, v1, v2, s105 row_shl:1
 680 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 681
 682 v_bfe_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 683 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 684
 685 v_bfe_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 686 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 687
 688 v_bfe_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
 689 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
 690
 691 v_bfe_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
 692 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
 693
 694 v_bfe_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
 695 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
 696
 697 v_bfe_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
 698 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
 699
 700 v_bfe_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
 701 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
 702
 703 v_bfe_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 704 // GFX12: [0x05,0x00,0x10,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
 705
 706 v_bfe_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 707 // GFX12: [0xff,0x00,0x10,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 708
 709 v_bfi_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 710 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 711
 712 v_bfi_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 713 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 714
 715 v_bfi_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
 716 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
 717
 718 v_bfi_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 719 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 720
 721 v_bfi_b32_e64_dpp v5, v1, v2, v3 row_mirror
 722 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 723
 724 v_bfi_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
 725 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 726
 727 v_bfi_b32_e64_dpp v5, v1, v2, s105 row_shl:1
 728 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 729
 730 v_bfi_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 731 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 732
 733 v_bfi_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 734 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 735
 736 v_bfi_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
 737 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
 738
 739 v_bfi_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
 740 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
 741
 742 v_bfi_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
 743 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
 744
 745 v_bfi_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
 746 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
 747
 748 v_bfi_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
 749 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
 750
 751 v_bfi_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 752 // GFX12: [0x05,0x00,0x12,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
 753
 754 v_bfi_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 755 // GFX12: [0xff,0x00,0x12,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
 756
 757 v_bfm_b32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
 758 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
 759
 760 v_bfm_b32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
 761 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
 762
 763 v_bfm_b32_e64_dpp v5, v1, v2 row_mirror
 764 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
 765
 766 v_bfm_b32_e64_dpp v5, v1, v2 row_half_mirror
 767 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
 768
 769 v_bfm_b32_e64_dpp v5, v1, v2 row_shl:1
 770 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
 771
 772 v_bfm_b32_e64_dpp v5, v1, v2 row_shl:15
 773 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
 774
 775 v_bfm_b32_e64_dpp v5, v1, v2 row_shr:1
 776 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
 777
 778 v_bfm_b32_e64_dpp v5, v1, v2 row_shr:15
 779 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
 780
 781 v_bfm_b32_e64_dpp v5, v1, v2 row_ror:1
 782 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
 783
 784 v_bfm_b32_e64_dpp v5, v1, v2 row_ror:15
 785 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
 786
 787 v_bfm_b32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
 788 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
 789
 790 v_bfm_b32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
 791 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
 792
 793 v_bfm_b32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 794 // GFX12: [0x05,0x00,0x1d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
 795
 796 v_bfm_b32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 797 // GFX12: [0xff,0x00,0x1d,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
 798
 799 v_cndmask_b16_e64_dpp v5, v1, v2, s3 quad_perm:[3,2,1,0]
 800 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x1b,0x00,0xff]
 801 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 802
 803 v_cndmask_b16_e64_dpp v5, v1, v2, s3 quad_perm:[0,1,2,3]
 804 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0xe4,0x00,0xff]
 805 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 806
 807 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_mirror
 808 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x40,0x01,0xff]
 809 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 810
 811 v_cndmask_b16_e64_dpp v5, v1, s2, s3 row_mirror
 812 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0c,0x00,0x01,0x40,0x01,0xff]
 813 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 814
 815 v_cndmask_b16_e64_dpp v5, v1, 10, s3 row_mirror
 816 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x14,0x0d,0x00,0x01,0x40,0x01,0xff]
 817 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 818
 819 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_half_mirror
 820 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x41,0x01,0xff]
 821 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 822
 823 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_shl:1
 824 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x01,0x01,0xff]
 825 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 826
 827 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_shl:15
 828 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x0f,0x01,0xff]
 829 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 830
 831 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_shr:1
 832 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x11,0x01,0xff]
 833 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 834
 835 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_shr:15
 836 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x1f,0x01,0xff]
 837 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 838
 839 v_cndmask_b16_e64_dpp v5, v1, v2, s3 row_ror:1
 840 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x0e,0x00,0x01,0x21,0x01,0xff]
 841 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 842
 843 v_cndmask_b16_e64_dpp v5, v1, v2, s105 row_ror:15
 844 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x2f,0x01,0xff]
 845 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 846
 847 v_cndmask_b16_e64_dpp v5, v1, v2, vcc_hi row_share:0 row_mask:0xf bank_mask:0xf
 848 // W32: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x50,0x01,0xff]
 849 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 850
 851 v_cndmask_b16_e64_dpp v5, |v1|, -v2, vcc_lo row_share:15 row_mask:0x0 bank_mask:0x1
 852 // W32: [0x05,0x01,0x5d,0xd6,0xfa,0x04,0xaa,0x41,0x01,0x5f,0x01,0x01]
 853 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 854
 855 v_cndmask_b16_e64_dpp v5, -v1, |v2|, ttmp15 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 856 // W32: [0x05,0x02,0x5d,0xd6,0xfa,0x04,0xee,0x21,0x01,0x60,0x09,0x13]
 857 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 858
 859 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] quad_perm:[3,2,1,0]
 860 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x1b,0x00,0xff]
 861 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 862
 863 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] quad_perm:[0,1,2,3]
 864 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0xe4,0x00,0xff]
 865 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 866
 867 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_mirror
 868 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x40,0x01,0xff]
 869 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 870
 871 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_half_mirror
 872 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x41,0x01,0xff]
 873 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 874
 875 v_cndmask_b16_e64_dpp v5, v1, s2, s[6:7] row_half_mirror
 876 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x18,0x00,0x01,0x41,0x01,0xff]
 877 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 878
 879 v_cndmask_b16_e64_dpp v5, v1, 10, s[6:7] row_half_mirror
 880 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x14,0x19,0x00,0x01,0x41,0x01,0xff]
 881 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 882
 883 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_shl:1
 884 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x01,0x01,0xff]
 885 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 886
 887 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_shl:15
 888 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x0f,0x01,0xff]
 889 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 890
 891 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_shr:1
 892 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x11,0x01,0xff]
 893 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 894
 895 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_shr:15
 896 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x1f,0x01,0xff]
 897 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 898
 899 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_ror:1
 900 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x21,0x01,0xff]
 901 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 902
 903 v_cndmask_b16_e64_dpp v5, v1, v2, s[6:7] row_ror:15
 904 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0x1a,0x00,0x01,0x2f,0x01,0xff]
 905 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 906
 907 v_cndmask_b16_e64_dpp v5, v1, v2, s[104:105] row_share:0 row_mask:0xf bank_mask:0xf
 908 // W64: [0x05,0x00,0x5d,0xd6,0xfa,0x04,0xa2,0x01,0x01,0x50,0x01,0xff]
 909 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 910
 911 v_cndmask_b16_e64_dpp v5, |v1|, -v2, vcc row_share:15 row_mask:0x0 bank_mask:0x1
 912 // W64: [0x05,0x01,0x5d,0xd6,0xfa,0x04,0xaa,0x41,0x01,0x5f,0x01,0x01]
 913 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 914
 915 v_cndmask_b16_e64_dpp v5, -v1, |v2|, ttmp[14:15] row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 916 // W64: [0x05,0x02,0x5d,0xd6,0xfa,0x04,0xea,0x21,0x01,0x60,0x09,0x13]
 917 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
 918
 919 v_cndmask_b16_e64_dpp v255, -|v255|, -|v255|, null row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 920 // GFX12: [0xff,0x03,0x5d,0xd6,0xfa,0xfe,0xf3,0x61,0xff,0x6f,0x05,0x30]
 921
 922 v_cubeid_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 923 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 924
 925 v_cubeid_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 926 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 927
 928 v_cubeid_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
 929 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
 930
 931 v_cubeid_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 932 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 933
 934 v_cubeid_f32_e64_dpp v5, v1, v2, v3 row_mirror
 935 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 936
 937 v_cubeid_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
 938 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 939
 940 v_cubeid_f32_e64_dpp v5, v1, v2, s105 row_shl:1
 941 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 942
 943 v_cubeid_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 944 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 945
 946 v_cubeid_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 947 // GFX12: [0x05,0x00,0x0c,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 948
 949 v_cubeid_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
 950 // GFX12: [0x05,0x01,0x0c,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
 951
 952 v_cubeid_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
 953 // GFX12: [0x05,0x02,0x0c,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
 954
 955 v_cubeid_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
 956 // GFX12: [0x05,0x04,0x0c,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
 957
 958 v_cubeid_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
 959 // GFX12: [0x05,0x03,0x0c,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
 960
 961 v_cubeid_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
 962 // GFX12: [0x05,0x05,0x0c,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
 963
 964 v_cubeid_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
 965 // GFX12: [0x05,0x06,0x0c,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
 966
 967 v_cubeid_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
 968 // GFX12: [0xff,0x87,0x0c,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
 969
 970 v_cubema_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
 971 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
 972
 973 v_cubema_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
 974 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
 975
 976 v_cubema_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
 977 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
 978
 979 v_cubema_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
 980 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
 981
 982 v_cubema_f32_e64_dpp v5, v1, v2, v3 row_mirror
 983 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
 984
 985 v_cubema_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
 986 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
 987
 988 v_cubema_f32_e64_dpp v5, v1, v2, s105 row_shl:1
 989 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
 990
 991 v_cubema_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
 992 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
 993
 994 v_cubema_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
 995 // GFX12: [0x05,0x00,0x0f,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
 996
 997 v_cubema_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
 998 // GFX12: [0x05,0x01,0x0f,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
 999
1000 v_cubema_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1001 // GFX12: [0x05,0x02,0x0f,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1002
1003 v_cubema_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1004 // GFX12: [0x05,0x04,0x0f,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1005
1006 v_cubema_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1007 // GFX12: [0x05,0x03,0x0f,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1008
1009 v_cubema_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1010 // GFX12: [0x05,0x05,0x0f,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
1011
1012 v_cubema_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1013 // GFX12: [0x05,0x06,0x0f,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
1014
1015 v_cubema_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1016 // GFX12: [0xff,0x87,0x0f,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
1017
1018 v_cubesc_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1019 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1020
1021 v_cubesc_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1022 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1023
1024 v_cubesc_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1025 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1026
1027 v_cubesc_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1028 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1029
1030 v_cubesc_f32_e64_dpp v5, v1, v2, v3 row_mirror
1031 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1032
1033 v_cubesc_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
1034 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1035
1036 v_cubesc_f32_e64_dpp v5, v1, v2, s105 row_shl:1
1037 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1038
1039 v_cubesc_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1040 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1041
1042 v_cubesc_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1043 // GFX12: [0x05,0x00,0x0d,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1044
1045 v_cubesc_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
1046 // GFX12: [0x05,0x01,0x0d,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
1047
1048 v_cubesc_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1049 // GFX12: [0x05,0x02,0x0d,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1050
1051 v_cubesc_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1052 // GFX12: [0x05,0x04,0x0d,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1053
1054 v_cubesc_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1055 // GFX12: [0x05,0x03,0x0d,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1056
1057 v_cubesc_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1058 // GFX12: [0x05,0x05,0x0d,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
1059
1060 v_cubesc_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1061 // GFX12: [0x05,0x06,0x0d,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
1062
1063 v_cubesc_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1064 // GFX12: [0xff,0x87,0x0d,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
1065
1066 v_cubetc_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1067 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1068
1069 v_cubetc_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1070 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1071
1072 v_cubetc_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1073 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1074
1075 v_cubetc_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1076 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1077
1078 v_cubetc_f32_e64_dpp v5, v1, v2, v3 row_mirror
1079 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1080
1081 v_cubetc_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
1082 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1083
1084 v_cubetc_f32_e64_dpp v5, v1, v2, s105 row_shl:1
1085 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1086
1087 v_cubetc_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1088 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1089
1090 v_cubetc_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1091 // GFX12: [0x05,0x00,0x0e,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1092
1093 v_cubetc_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
1094 // GFX12: [0x05,0x01,0x0e,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
1095
1096 v_cubetc_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1097 // GFX12: [0x05,0x02,0x0e,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1098
1099 v_cubetc_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1100 // GFX12: [0x05,0x04,0x0e,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1101
1102 v_cubetc_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1103 // GFX12: [0x05,0x03,0x0e,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1104
1105 v_cubetc_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1106 // GFX12: [0x05,0x05,0x0e,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
1107
1108 v_cubetc_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1109 // GFX12: [0x05,0x06,0x0e,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
1110
1111 v_cubetc_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1112 // GFX12: [0xff,0x87,0x0e,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
1113
1114 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1115 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1116
1117 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[0,1,2,3]
1118 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0xe4,0x00,0xff]
1119
1120 v_cvt_pk_bf8_f32_e64_dpp v6, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1121 // GFX12: encoding: [0x06,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1122
1123 v_cvt_pk_bf8_f32_e64_dpp v1, -v6, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1124 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x06,0x1b,0x00,0xed]
1125
1126 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v255| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1127 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0xfe,0x03,0x20,0x02,0x1b,0x00,0xed]
1128
1129 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[0,2,1,3] row_mask:0xe bank_mask:0xd
1130 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0xd8,0x00,0xed]
1131
1132 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0x2 bank_mask:0xd
1133 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0x2d]
1134
1135 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0x5
1136 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xe5]
1137
1138 v_cvt_pk_bf8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd fi:1
1139 // GFX12: encoding: [0x01,0x02,0x6a,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x04,0xed]
1140
1141 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1142 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1143
1144 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[0,1,2,3]
1145 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0xe4,0x00,0xff]
1146
1147 v_cvt_pk_fp8_f32_e64_dpp v6, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1148 // GFX12: encoding: [0x06,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1149
1150 v_cvt_pk_fp8_f32_e64_dpp v1, -v6, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1151 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x06,0x1b,0x00,0xed]
1152
1153 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v255| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1154 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0xfe,0x03,0x20,0x02,0x1b,0x00,0xed]
1155
1156 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[0,2,1,3] row_mask:0xe bank_mask:0xd
1157 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0xd8,0x00,0xed]
1158
1159 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0x2 bank_mask:0xd
1160 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0x2d]
1161
1162 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0x5
1163 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xe5]
1164
1165 v_cvt_pk_fp8_f32_e64_dpp v1, -v2, |v3| quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd fi:1
1166 // GFX12: encoding: [0x01,0x02,0x69,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x04,0xed]
1167
1168 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1169 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1170
1171 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf
1172 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0xe4,0x00,0xff]
1173
1174 v_cvt_sr_bf8_f32_e64_dpp v6, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1175 // GFX12: encoding: [0x06,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1176
1177 v_cvt_sr_bf8_f32_e64_dpp v1, -v6, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1178 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x06,0x1b,0x00,0xed]
1179
1180 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v255 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1181 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0xfe,0x03,0x20,0x02,0x1b,0x00,0xed]
1182
1183 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[0,2,1,3] row_mask:0xe bank_mask:0xd
1184 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0xd8,0x00,0xed]
1185
1186 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0x2 bank_mask:0xd
1187 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0x2d]
1188
1189 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0x5
1190 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xe5]
1191
1192 v_cvt_sr_bf8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd fi:1
1193 // GFX12: encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x04,0xed]
1194
1195 v_cvt_sr_bf8_f32 v1, v2, v3 byte_sel:0 quad_perm:[3,2,1,0]
1196 // GFX12: v_cvt_sr_bf8_f32_e64_dpp v1, v2, v3 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x00,0x6c,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1197
1198 v_cvt_sr_bf8_f32 v1, v2, v3 byte_sel:1 quad_perm:[3,2,1,0]
1199 // GFX12: v_cvt_sr_bf8_f32_e64_dpp v1, v2, v3 byte_sel:1 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x20,0x6c,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1200
1201 v_cvt_sr_bf8_f32 v1, v2, v3 byte_sel:2 quad_perm:[3,2,1,0]
1202 // GFX12: v_cvt_sr_bf8_f32_e64_dpp v1, v2, v3 byte_sel:2 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x40,0x6c,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1203
1204 v_cvt_sr_bf8_f32 v1, v2, v3 byte_sel:3 quad_perm:[3,2,1,0]
1205 // GFX12: v_cvt_sr_bf8_f32_e64_dpp v1, v2, v3 byte_sel:3 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x60,0x6c,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1206
1207 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1208 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1209
1210 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf
1211 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0xe4,0x00,0xff]
1212
1213 v_cvt_sr_fp8_f32_e64_dpp v6, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1214 // GFX12: encoding: [0x06,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xed]
1215
1216 v_cvt_sr_fp8_f32_e64_dpp v1, -v6, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1217 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x06,0x1b,0x00,0xed]
1218
1219 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v255 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd
1220 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0xfe,0x03,0x20,0x02,0x1b,0x00,0xed]
1221
1222 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[0,2,1,3] row_mask:0xe bank_mask:0xd
1223 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0xd8,0x00,0xed]
1224
1225 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0x2 bank_mask:0xd
1226 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0x2d]
1227
1228 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0x5
1229 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x00,0xe5]
1230
1231 v_cvt_sr_fp8_f32_e64_dpp v1, -v2, v3 quad_perm:[3,2,1,0] row_mask:0xe bank_mask:0xd fi:1
1232 // GFX12: encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x20,0x02,0x1b,0x04,0xed]
1233
1234 v_cvt_sr_fp8_f32 v1, v2, v3 byte_sel:0 quad_perm:[3,2,1,0]
1235 // GFX12: v_cvt_sr_fp8_f32_e64_dpp v1, v2, v3 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x00,0x6b,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1236
1237 v_cvt_sr_fp8_f32 v1, v2, v3 byte_sel:1 quad_perm:[3,2,1,0]
1238 // GFX12: v_cvt_sr_fp8_f32_e64_dpp v1, v2, v3 byte_sel:1 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x20,0x6b,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1239
1240 v_cvt_sr_fp8_f32 v1, v2, v3 byte_sel:2 quad_perm:[3,2,1,0]
1241 // GFX12: v_cvt_sr_fp8_f32_e64_dpp v1, v2, v3 byte_sel:2 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x40,0x6b,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1242
1243 v_cvt_sr_fp8_f32 v1, v2, v3 byte_sel:3 quad_perm:[3,2,1,0]
1244 // GFX12: v_cvt_sr_fp8_f32_e64_dpp v1, v2, v3 byte_sel:3 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf ; encoding: [0x01,0x60,0x6b,0xd7,0xfa,0x06,0x02,0x00,0x02,0x1b,0x00,0xff]
1245
1246 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1247 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1248
1249 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1250 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1251
1252 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_mirror
1253 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1254
1255 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_half_mirror
1256 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1257
1258 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_shl:1
1259 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1260
1261 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_shl:15
1262 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1263
1264 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_shr:1
1265 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1266
1267 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_shr:15
1268 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1269
1270 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_ror:1
1271 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1272
1273 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_ror:15
1274 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1275
1276 v_cvt_pk_i16_f32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1277 // GFX12: [0x05,0x00,0x06,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1278
1279 v_cvt_pk_i16_f32_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1280 // GFX12: [0x05,0x01,0x06,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1281
1282 v_cvt_pk_i16_f32_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1283 // GFX12: [0x05,0x02,0x06,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1284
1285 v_cvt_pk_i16_f32_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1286 // GFX12: [0xff,0x03,0x06,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1287
1288 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1289 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1290
1291 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1292 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1293
1294 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_mirror
1295 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1296
1297 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_half_mirror
1298 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1299
1300 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_shl:1
1301 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1302
1303 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_shl:15
1304 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1305
1306 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_shr:1
1307 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1308
1309 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_shr:15
1310 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1311
1312 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_ror:1
1313 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1314
1315 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_ror:15
1316 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1317
1318 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1319 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1320
1321 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
1322 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
1323
1324 v_cvt_pk_i16_i32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1325 // GFX12: [0x05,0x00,0x24,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
1326
1327 v_cvt_pk_i16_i32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1328 // GFX12: [0xff,0x00,0x24,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
1329
1330 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1331 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1332
1333 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1334 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1335
1336 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_mirror
1337 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1338
1339 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_half_mirror
1340 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1341
1342 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shl:1
1343 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1344
1345 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shl:15
1346 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1347
1348 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shr:1
1349 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1350
1351 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shr:15
1352 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1353
1354 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_ror:1
1355 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1356
1357 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_ror:15
1358 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1359
1360 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1361 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1362
1363 v_cvt_pk_norm_i16_f16_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1364 // GFX12: [0x05,0x01,0x12,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1365
1366 v_cvt_pk_norm_i16_f16_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1367 // GFX12: [0x05,0x02,0x12,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1368
1369 v_cvt_pk_norm_i16_f16_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1370 // GFX12: [0xff,0x03,0x12,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1371
1372 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1373 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1374
1375 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1376 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1377
1378 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_mirror
1379 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1380
1381 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_half_mirror
1382 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1383
1384 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shl:1
1385 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1386
1387 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shl:15
1388 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1389
1390 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shr:1
1391 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1392
1393 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shr:15
1394 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1395
1396 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_ror:1
1397 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1398
1399 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_ror:15
1400 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1401
1402 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1403 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1404
1405 v_cvt_pk_norm_u16_f16_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1406 // GFX12: [0x05,0x01,0x13,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1407
1408 v_cvt_pk_norm_u16_f16_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1409 // GFX12: [0x05,0x02,0x13,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1410
1411 v_cvt_pk_norm_u16_f16_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1412 // GFX12: [0xff,0x03,0x13,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1413
1414 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1415 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1416
1417 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1418 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1419
1420 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_mirror
1421 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1422
1423 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_half_mirror
1424 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1425
1426 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_shl:1
1427 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1428
1429 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_shl:15
1430 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1431
1432 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_shr:1
1433 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1434
1435 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_shr:15
1436 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1437
1438 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_ror:1
1439 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1440
1441 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_ror:15
1442 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1443
1444 v_cvt_pk_u16_f32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1445 // GFX12: [0x05,0x00,0x07,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1446
1447 v_cvt_pk_u16_f32_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1448 // GFX12: [0x05,0x01,0x07,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1449
1450 v_cvt_pk_u16_f32_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1451 // GFX12: [0x05,0x02,0x07,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1452
1453 v_cvt_pk_u16_f32_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1454 // GFX12: [0xff,0x03,0x07,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1455
1456 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1457 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1458
1459 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1460 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1461
1462 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_mirror
1463 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1464
1465 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_half_mirror
1466 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1467
1468 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_shl:1
1469 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1470
1471 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_shl:15
1472 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1473
1474 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_shr:1
1475 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1476
1477 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_shr:15
1478 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1479
1480 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_ror:1
1481 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1482
1483 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_ror:15
1484 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1485
1486 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1487 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1488
1489 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
1490 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
1491
1492 v_cvt_pk_u16_u32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1493 // GFX12: [0x05,0x00,0x23,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
1494
1495 v_cvt_pk_u16_u32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1496 // GFX12: [0xff,0x00,0x23,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
1497
1498 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1499 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1500
1501 v_cvt_pk_u8_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1502 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1503
1504 v_cvt_pk_u8_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1505 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1506
1507 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1508 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1509
1510 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, v3 row_mirror
1511 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1512
1513 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
1514 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1515
1516 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, s105 row_shl:1
1517 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1518
1519 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1520 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1521
1522 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1523 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1524
1525 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
1526 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
1527
1528 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, exec_hi row_ror:1
1529 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
1530
1531 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, exec_lo row_ror:15
1532 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
1533
1534 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
1535 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
1536
1537 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
1538 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
1539
1540 v_cvt_pk_u8_f32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1541 // GFX12: [0x05,0x00,0x26,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
1542
1543 v_cvt_pk_u8_f32_e64_dpp v255, -|v255|, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1544 // GFX12: [0xff,0x01,0x26,0xd6,0xfa,0xfe,0xf7,0x23,0xff,0x6f,0x05,0x30]
1545
1546 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1547 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1548
1549 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1550 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1551
1552 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_mirror
1553 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1554
1555 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_half_mirror
1556 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1557
1558 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shl:1
1559 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1560
1561 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shl:15
1562 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1563
1564 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shr:1
1565 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1566
1567 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_shr:15
1568 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1569
1570 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_ror:1
1571 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1572
1573 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_ror:15
1574 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1575
1576 v_cvt_pk_norm_i16_f16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1577 // GFX12: [0x05,0x00,0x12,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1578
1579 v_cvt_pk_norm_i16_f16_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1580 // GFX12: [0x05,0x01,0x12,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1581
1582 v_cvt_pk_norm_i16_f16_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1583 // GFX12: [0x05,0x02,0x12,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1584
1585 v_cvt_pk_norm_i16_f16_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1586 // GFX12: [0xff,0x03,0x12,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1587
1588 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1589 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1590
1591 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1592 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1593
1594 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_mirror
1595 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1596
1597 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_half_mirror
1598 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1599
1600 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_shl:1
1601 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1602
1603 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_shl:15
1604 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1605
1606 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_shr:1
1607 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1608
1609 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_shr:15
1610 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1611
1612 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_ror:1
1613 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1614
1615 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_ror:15
1616 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1617
1618 v_cvt_pk_norm_i16_f32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1619 // GFX12: [0x05,0x00,0x21,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1620
1621 v_cvt_pk_norm_i16_f32_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1622 // GFX12: [0x05,0x01,0x21,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1623
1624 v_cvt_pk_norm_i16_f32_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1625 // GFX12: [0x05,0x02,0x21,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1626
1627 v_cvt_pk_norm_i16_f32_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1628 // GFX12: [0xff,0x03,0x21,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1629
1630 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1631 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1632
1633 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1634 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1635
1636 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_mirror
1637 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1638
1639 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_half_mirror
1640 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1641
1642 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shl:1
1643 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1644
1645 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shl:15
1646 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1647
1648 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shr:1
1649 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1650
1651 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_shr:15
1652 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1653
1654 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_ror:1
1655 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1656
1657 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_ror:15
1658 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1659
1660 v_cvt_pk_norm_u16_f16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1661 // GFX12: [0x05,0x00,0x13,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1662
1663 v_cvt_pk_norm_u16_f16_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1664 // GFX12: [0x05,0x01,0x13,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1665
1666 v_cvt_pk_norm_u16_f16_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1667 // GFX12: [0x05,0x02,0x13,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1668
1669 v_cvt_pk_norm_u16_f16_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1670 // GFX12: [0xff,0x03,0x13,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1671
1672 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1673 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1674
1675 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1676 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1677
1678 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_mirror
1679 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1680
1681 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_half_mirror
1682 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1683
1684 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_shl:1
1685 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1686
1687 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_shl:15
1688 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1689
1690 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_shr:1
1691 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1692
1693 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_shr:15
1694 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1695
1696 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_ror:1
1697 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1698
1699 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_ror:15
1700 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1701
1702 v_cvt_pk_norm_u16_f32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1703 // GFX12: [0x05,0x00,0x22,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1704
1705 v_cvt_pk_norm_u16_f32_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
1706 // GFX12: [0x05,0x01,0x22,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
1707
1708 v_cvt_pk_norm_u16_f32_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1709 // GFX12: [0x05,0x02,0x22,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
1710
1711 v_cvt_pk_norm_u16_f32_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1712 // GFX12: [0xff,0x03,0x22,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
1713
1714 v_div_fixup_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1715 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1716
1717 v_div_fixup_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1718 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1719
1720 v_div_fixup_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1721 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1722
1723 v_div_fixup_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1724 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1725
1726 v_div_fixup_f16_e64_dpp v5, v1, v2, v3 row_mirror
1727 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1728
1729 v_div_fixup_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
1730 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1731
1732 v_div_fixup_f16_e64_dpp v5, v1, v2, s105 row_shl:1
1733 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1734
1735 v_div_fixup_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1736 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1737
1738 v_div_fixup_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1739 // GFX12: [0x05,0x00,0x54,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1740
1741 v_div_fixup_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
1742 // GFX12: [0x05,0x01,0x54,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
1743
1744 v_div_fixup_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1745 // GFX12: [0x05,0x02,0x54,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1746
1747 v_div_fixup_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1748 // GFX12: [0x05,0x04,0x54,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1749
1750 v_div_fixup_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1751 // GFX12: [0x05,0x03,0x54,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1752
1753 v_div_fixup_f16_e64_dpp v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
1754 // GFX12: [0x05,0x05,0x54,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
1755
1756 v_div_fixup_f16_e64_dpp v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1757 // GFX12: [0x05,0x06,0x54,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
1758
1759 v_div_fixup_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1760 // GFX12: [0xff,0x87,0x54,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
1761
1762 v_fma_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1763 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1764
1765 v_fma_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1766 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1767
1768 v_fma_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1769 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1770
1771 v_fma_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1772 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1773
1774 v_fma_f16_e64_dpp v5, v1, v2, v3 row_mirror
1775 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1776
1777 v_fma_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
1778 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1779
1780 v_fma_f16_e64_dpp v5, v1, v2, s105 row_shl:1
1781 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1782
1783 v_fma_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1784 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1785
1786 v_fma_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1787 // GFX12: [0x05,0x00,0x48,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1788
1789 v_fma_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
1790 // GFX12: [0x05,0x01,0x48,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
1791
1792 v_fma_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1793 // GFX12: [0x05,0x02,0x48,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1794
1795 v_fma_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1796 // GFX12: [0x05,0x04,0x48,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1797
1798 v_fma_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1799 // GFX12: [0x05,0x03,0x48,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1800
1801 v_fma_f16_e64_dpp v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
1802 // GFX12: [0x05,0x05,0x48,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
1803
1804 v_fma_f16_e64_dpp v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1805 // GFX12: [0x05,0x06,0x48,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
1806
1807 v_fma_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1808 // GFX12: [0xff,0x87,0x48,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
1809
1810 v_fma_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1811 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1812
1813 v_fma_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1814 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1815
1816 v_fma_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
1817 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
1818
1819 v_fma_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1820 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1821
1822 v_fma_f32_e64_dpp v5, v1, v2, v3 row_mirror
1823 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1824
1825 v_fma_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
1826 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1827
1828 v_fma_f32_e64_dpp v5, v1, v2, s105 row_shl:1
1829 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1830
1831 v_fma_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1832 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1833
1834 v_fma_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1835 // GFX12: [0x05,0x00,0x13,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1836
1837 v_fma_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
1838 // GFX12: [0x05,0x01,0x13,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
1839
1840 v_fma_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
1841 // GFX12: [0x05,0x02,0x13,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
1842
1843 v_fma_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
1844 // GFX12: [0x05,0x04,0x13,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
1845
1846 v_fma_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
1847 // GFX12: [0x05,0x03,0x13,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
1848
1849 v_fma_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1850 // GFX12: [0x05,0x05,0x13,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
1851
1852 v_fma_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1853 // GFX12: [0x05,0x06,0x13,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
1854
1855 v_fma_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1856 // GFX12: [0xff,0x87,0x13,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
1857
1858 v_ldexp_f32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
1859 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
1860
1861 v_ldexp_f32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
1862 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
1863
1864 v_ldexp_f32_e64_dpp v5, v1, v2 row_mirror
1865 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
1866
1867 v_ldexp_f32_e64_dpp v5, v1, v2 row_half_mirror
1868 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
1869
1870 v_ldexp_f32_e64_dpp v5, v1, v2 row_shl:1
1871 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
1872
1873 v_ldexp_f32_e64_dpp v5, v1, v2 row_shl:15
1874 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
1875
1876 v_ldexp_f32_e64_dpp v5, v1, v2 row_shr:1
1877 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
1878
1879 v_ldexp_f32_e64_dpp v5, v1, v2 row_shr:15
1880 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
1881
1882 v_ldexp_f32_e64_dpp v5, v1, v2 row_ror:1
1883 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
1884
1885 v_ldexp_f32_e64_dpp v5, v1, v2 row_ror:15
1886 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
1887
1888 v_ldexp_f32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
1889 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
1890
1891 v_ldexp_f32_e64_dpp v5, v1, v2 mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
1892 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x08,0x01,0x5f,0x01,0x01]
1893
1894 v_ldexp_f32_e64_dpp v5, v1, v2 mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1895 // GFX12: [0x05,0x00,0x1c,0xd7,0xfa,0x04,0x02,0x10,0x01,0x60,0x09,0x13]
1896
1897 v_ldexp_f32_e64_dpp v255, -|v255|, v255 clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1898 // GFX12: [0xff,0x81,0x1c,0xd7,0xfa,0xfe,0x03,0x38,0xff,0x6f,0x05,0x30]
1899
1900 v_lerp_u8_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1901 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1902
1903 v_lerp_u8_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1904 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1905
1906 v_lerp_u8_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1907 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1908
1909 v_lerp_u8_e64_dpp v5, v1, v2, v3 row_mirror
1910 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1911
1912 v_lerp_u8_e64_dpp v5, v1, v2, v255 row_half_mirror
1913 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1914
1915 v_lerp_u8_e64_dpp v5, v1, v2, s105 row_shl:1
1916 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1917
1918 v_lerp_u8_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1919 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1920
1921 v_lerp_u8_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1922 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1923
1924 v_lerp_u8_e64_dpp v5, v1, v2, ttmp15 row_shr:15
1925 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
1926
1927 v_lerp_u8_e64_dpp v5, v1, v2, exec_hi row_ror:1
1928 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
1929
1930 v_lerp_u8_e64_dpp v5, v1, v2, exec_lo row_ror:15
1931 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
1932
1933 v_lerp_u8_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
1934 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
1935
1936 v_lerp_u8_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
1937 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
1938
1939 v_lerp_u8_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1940 // GFX12: [0x05,0x00,0x15,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
1941
1942 v_lerp_u8_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1943 // GFX12: [0xff,0x00,0x15,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
1944
1945 v_lshl_add_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1946 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1947
1948 v_lshl_add_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1949 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1950
1951 v_lshl_add_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
1952 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
1953
1954 v_lshl_add_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
1955 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
1956
1957 v_lshl_add_u32_e64_dpp v5, v1, v2, v3 row_mirror
1958 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
1959
1960 v_lshl_add_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
1961 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
1962
1963 v_lshl_add_u32_e64_dpp v5, v1, v2, s105 row_shl:1
1964 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
1965
1966 v_lshl_add_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
1967 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
1968
1969 v_lshl_add_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
1970 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
1971
1972 v_lshl_add_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
1973 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
1974
1975 v_lshl_add_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
1976 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
1977
1978 v_lshl_add_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
1979 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
1980
1981 v_lshl_add_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
1982 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
1983
1984 v_lshl_add_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
1985 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
1986
1987 v_lshl_add_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
1988 // GFX12: [0x05,0x00,0x46,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
1989
1990 v_lshl_add_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
1991 // GFX12: [0xff,0x00,0x46,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
1992
1993 v_lshl_or_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
1994 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
1995
1996 v_lshl_or_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
1997 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
1998
1999 v_lshl_or_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2000 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2001
2002 v_lshl_or_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2003 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2004
2005 v_lshl_or_b32_e64_dpp v5, v1, v2, v3 row_mirror
2006 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2007
2008 v_lshl_or_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
2009 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2010
2011 v_lshl_or_b32_e64_dpp v5, v1, v2, s105 row_shl:1
2012 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2013
2014 v_lshl_or_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2015 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2016
2017 v_lshl_or_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2018 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2019
2020 v_lshl_or_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2021 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2022
2023 v_lshl_or_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
2024 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2025
2026 v_lshl_or_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
2027 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2028
2029 v_lshl_or_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2030 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2031
2032 v_lshl_or_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2033 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2034
2035 v_lshl_or_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2036 // GFX12: [0x05,0x00,0x56,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2037
2038 v_lshl_or_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2039 // GFX12: [0xff,0x00,0x56,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2040
2041 v_lshlrev_b16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
2042 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
2043
2044 v_lshlrev_b16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
2045 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
2046
2047 v_lshlrev_b16_e64_dpp v5, v1, v2 row_mirror
2048 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
2049
2050 v_lshlrev_b16_e64_dpp v5, v1, v2 row_half_mirror
2051 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
2052
2053 v_lshlrev_b16_e64_dpp v5, v1, v2 row_shl:1
2054 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
2055
2056 v_lshlrev_b16_e64_dpp v5, v1, v2 row_shl:15
2057 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
2058
2059 v_lshlrev_b16_e64_dpp v5, v1, v2 row_shr:1
2060 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
2061
2062 v_lshlrev_b16_e64_dpp v5, v1, v2 row_shr:15
2063 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
2064
2065 v_lshlrev_b16_e64_dpp v5, v1, v2 row_ror:1
2066 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
2067
2068 v_lshlrev_b16_e64_dpp v5, v1, v2 row_ror:15
2069 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
2070
2071 v_lshlrev_b16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
2072 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
2073
2074 v_lshlrev_b16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
2075 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
2076
2077 v_lshlrev_b16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2078 // GFX12: [0x05,0x00,0x38,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
2079
2080 v_lshlrev_b16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2081 // GFX12: [0xff,0x00,0x38,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
2082
2083 v_lshrrev_b16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
2084 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
2085
2086 v_lshrrev_b16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
2087 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
2088
2089 v_lshrrev_b16_e64_dpp v5, v1, v2 row_mirror
2090 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
2091
2092 v_lshrrev_b16_e64_dpp v5, v1, v2 row_half_mirror
2093 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
2094
2095 v_lshrrev_b16_e64_dpp v5, v1, v2 row_shl:1
2096 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
2097
2098 v_lshrrev_b16_e64_dpp v5, v1, v2 row_shl:15
2099 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
2100
2101 v_lshrrev_b16_e64_dpp v5, v1, v2 row_shr:1
2102 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
2103
2104 v_lshrrev_b16_e64_dpp v5, v1, v2 row_shr:15
2105 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
2106
2107 v_lshrrev_b16_e64_dpp v5, v1, v2 row_ror:1
2108 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
2109
2110 v_lshrrev_b16_e64_dpp v5, v1, v2 row_ror:15
2111 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
2112
2113 v_lshrrev_b16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
2114 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
2115
2116 v_lshrrev_b16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
2117 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
2118
2119 v_lshrrev_b16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2120 // GFX12: [0x05,0x00,0x39,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
2121
2122 v_lshrrev_b16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2123 // GFX12: [0xff,0x00,0x39,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
2124
2125 v_mad_i16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2126 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2127
2128 v_mad_i16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2129 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2130
2131 v_mad_i16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2132 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2133
2134 v_mad_i16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2135 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2136
2137 v_mad_i16_e64_dpp v5, v1, v2, v3 row_mirror
2138 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2139
2140 v_mad_i16_e64_dpp v5, v1, v2, v3 row_half_mirror
2141 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
2142
2143 v_mad_i16_e64_dpp v5, v1, v2, v255 row_shl:1
2144 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
2145
2146 v_mad_i16_e64_dpp v5, v1, v2, s105 row_shl:15
2147 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
2148
2149 v_mad_i16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
2150 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
2151
2152 v_mad_i16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
2153 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
2154
2155 v_mad_i16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
2156 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
2157
2158 v_mad_i16_e64_dpp v5, v1, v2, exec_hi row_ror:15
2159 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
2160
2161 v_mad_i16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
2162 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
2163
2164 v_mad_i16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
2165 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
2166
2167 v_mad_i16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2168 // GFX12: [0x05,0x00,0x53,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
2169
2170 v_mad_i16_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2171 // GFX12: [0xff,0x80,0x53,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2172
2173 v_mad_i32_i16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2174 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2175
2176 v_mad_i32_i16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2177 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2178
2179 v_mad_i32_i16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2180 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2181
2182 v_mad_i32_i16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2183 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2184
2185 v_mad_i32_i16_e64_dpp v5, v1, v2, v3 row_mirror
2186 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2187
2188 v_mad_i32_i16_e64_dpp v5, v1, v2, v255 row_half_mirror
2189 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2190
2191 v_mad_i32_i16_e64_dpp v5, v1, v2, s105 row_shl:1
2192 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2193
2194 v_mad_i32_i16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2195 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2196
2197 v_mad_i32_i16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2198 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2199
2200 v_mad_i32_i16_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2201 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2202
2203 v_mad_i32_i16_e64_dpp v5, v1, v2, exec_hi row_ror:1
2204 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2205
2206 v_mad_i32_i16_e64_dpp v5, v1, v2, exec_lo row_ror:15
2207 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2208
2209 v_mad_i32_i16_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2210 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2211
2212 v_mad_i32_i16_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2213 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2214
2215 v_mad_i32_i16_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2216 // GFX12: [0x05,0x00,0x5a,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2217
2218 v_mad_i32_i16_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2219 // GFX12: [0xff,0x80,0x5a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2220
2221 v_mad_i32_i24_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2222 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2223
2224 v_mad_i32_i24_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2225 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2226
2227 v_mad_i32_i24_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2228 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2229
2230 v_mad_i32_i24_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2231 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2232
2233 v_mad_i32_i24_e64_dpp v5, v1, v2, v3 row_mirror
2234 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2235
2236 v_mad_i32_i24_e64_dpp v5, v1, v2, v255 row_half_mirror
2237 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2238
2239 v_mad_i32_i24_e64_dpp v5, v1, v2, s105 row_shl:1
2240 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2241
2242 v_mad_i32_i24_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2243 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2244
2245 v_mad_i32_i24_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2246 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2247
2248 v_mad_i32_i24_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2249 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2250
2251 v_mad_i32_i24_e64_dpp v5, v1, v2, exec_hi row_ror:1
2252 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2253
2254 v_mad_i32_i24_e64_dpp v5, v1, v2, exec_lo row_ror:15
2255 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2256
2257 v_mad_i32_i24_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2258 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2259
2260 v_mad_i32_i24_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2261 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2262
2263 v_mad_i32_i24_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2264 // GFX12: [0x05,0x00,0x0a,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2265
2266 v_mad_i32_i24_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2267 // GFX12: [0xff,0x80,0x0a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2268
2269 v_mad_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2270 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2271
2272 v_mad_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2273 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2274
2275 v_mad_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2276 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2277
2278 v_mad_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2279 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2280
2281 v_mad_u16_e64_dpp v5, v1, v2, v3 row_mirror
2282 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2283
2284 v_mad_u16_e64_dpp v5, v1, v2, v3 row_half_mirror
2285 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
2286
2287 v_mad_u16_e64_dpp v5, v1, v2, v255 row_shl:1
2288 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
2289
2290 v_mad_u16_e64_dpp v5, v1, v2, s105 row_shl:15
2291 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
2292
2293 v_mad_u16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
2294 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
2295
2296 v_mad_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
2297 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
2298
2299 v_mad_u16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
2300 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
2301
2302 v_mad_u16_e64_dpp v5, v1, v2, exec_hi row_ror:15
2303 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
2304
2305 v_mad_u16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
2306 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
2307
2308 v_mad_u16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
2309 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
2310
2311 v_mad_u16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2312 // GFX12: [0x05,0x00,0x41,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
2313
2314 v_mad_u16_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2315 // GFX12: [0xff,0x80,0x41,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2316
2317 v_mad_u32_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2318 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2319
2320 v_mad_u32_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2321 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2322
2323 v_mad_u32_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2324 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2325
2326 v_mad_u32_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2327 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2328
2329 v_mad_u32_u16_e64_dpp v5, v1, v2, v3 row_mirror
2330 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2331
2332 v_mad_u32_u16_e64_dpp v5, v1, v2, v255 row_half_mirror
2333 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2334
2335 v_mad_u32_u16_e64_dpp v5, v1, v2, s105 row_shl:1
2336 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2337
2338 v_mad_u32_u16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2339 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2340
2341 v_mad_u32_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2342 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2343
2344 v_mad_u32_u16_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2345 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2346
2347 v_mad_u32_u16_e64_dpp v5, v1, v2, exec_hi row_ror:1
2348 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2349
2350 v_mad_u32_u16_e64_dpp v5, v1, v2, exec_lo row_ror:15
2351 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2352
2353 v_mad_u32_u16_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2354 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2355
2356 v_mad_u32_u16_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2357 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2358
2359 v_mad_u32_u16_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2360 // GFX12: [0x05,0x00,0x59,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2361
2362 v_mad_u32_u16_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2363 // GFX12: [0xff,0x80,0x59,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2364
2365 v_mad_u32_u24_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2366 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2367
2368 v_mad_u32_u24_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2369 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2370
2371 v_mad_u32_u24_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2372 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2373
2374 v_mad_u32_u24_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2375 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2376
2377 v_mad_u32_u24_e64_dpp v5, v1, v2, v3 row_mirror
2378 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2379
2380 v_mad_u32_u24_e64_dpp v5, v1, v2, v255 row_half_mirror
2381 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2382
2383 v_mad_u32_u24_e64_dpp v5, v1, v2, s105 row_shl:1
2384 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2385
2386 v_mad_u32_u24_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2387 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2388
2389 v_mad_u32_u24_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2390 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2391
2392 v_mad_u32_u24_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2393 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2394
2395 v_mad_u32_u24_e64_dpp v5, v1, v2, exec_hi row_ror:1
2396 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2397
2398 v_mad_u32_u24_e64_dpp v5, v1, v2, exec_lo row_ror:15
2399 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2400
2401 v_mad_u32_u24_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2402 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2403
2404 v_mad_u32_u24_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2405 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2406
2407 v_mad_u32_u24_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2408 // GFX12: [0x05,0x00,0x0b,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2409
2410 v_mad_u32_u24_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2411 // GFX12: [0xff,0x80,0x0b,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2412
2413 v_max3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2414 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2415
2416 v_max3_num_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2417 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2418
2419 v_max3_num_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
2420 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
2421
2422 v_max3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2423 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2424
2425 v_max3_num_f16_e64_dpp v5, v1, v2, v3 row_mirror
2426 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2427
2428 v_max3_num_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
2429 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2430
2431 v_max3_num_f16_e64_dpp v5, v1, v2, s105 row_shl:1
2432 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2433
2434 v_max3_num_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2435 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2436
2437 v_max3_num_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2438 // GFX12: [0x05,0x00,0x2c,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2439
2440 v_max3_num_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
2441 // GFX12: [0x05,0x01,0x2c,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
2442
2443 v_max3_num_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
2444 // GFX12: [0x05,0x02,0x2c,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
2445
2446 v_max3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
2447 // GFX12: [0x05,0x04,0x2c,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
2448
2449 v_max3_num_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
2450 // GFX12: [0x05,0x03,0x2c,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
2451
2452 v_max3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
2453 // GFX12: [0x05,0x05,0x2c,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
2454
2455 v_max3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2456 // GFX12: [0x05,0x06,0x2c,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
2457
2458 v_max3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2459 // GFX12: [0xff,0x87,0x2c,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
2460
2461 v_max3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2462 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2463
2464 v_max3_num_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2465 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2466
2467 v_max3_num_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
2468 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
2469
2470 v_max3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2471 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2472
2473 v_max3_num_f32_e64_dpp v5, v1, v2, v3 row_mirror
2474 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2475
2476 v_max3_num_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
2477 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2478
2479 v_max3_num_f32_e64_dpp v5, v1, v2, s105 row_shl:1
2480 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2481
2482 v_max3_num_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2483 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2484
2485 v_max3_num_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2486 // GFX12: [0x05,0x00,0x2a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2487
2488 v_max3_num_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
2489 // GFX12: [0x05,0x01,0x2a,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
2490
2491 v_max3_num_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
2492 // GFX12: [0x05,0x02,0x2a,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
2493
2494 v_max3_num_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
2495 // GFX12: [0x05,0x04,0x2a,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
2496
2497 v_max3_num_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
2498 // GFX12: [0x05,0x03,0x2a,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
2499
2500 v_max3_num_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2501 // GFX12: [0x05,0x05,0x2a,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
2502
2503 v_max3_num_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2504 // GFX12: [0x05,0x06,0x2a,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
2505
2506 v_max3_num_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2507 // GFX12: [0xff,0x87,0x2a,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
2508
2509 v_max3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2510 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2511
2512 v_max3_i16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2513 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2514
2515 v_max3_i16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2516 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2517
2518 v_max3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2519 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2520
2521 v_max3_i16_e64_dpp v5, v1, v2, v3 row_mirror
2522 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2523
2524 v_max3_i16_e64_dpp v5, v1, v2, v3 row_half_mirror
2525 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
2526
2527 v_max3_i16_e64_dpp v5, v1, v2, v255 row_shl:1
2528 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
2529
2530 v_max3_i16_e64_dpp v5, v1, v2, s105 row_shl:15
2531 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
2532
2533 v_max3_i16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
2534 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
2535
2536 v_max3_i16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
2537 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
2538
2539 v_max3_i16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
2540 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
2541
2542 v_max3_i16_e64_dpp v5, v1, v2, exec_hi row_ror:15
2543 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
2544
2545 v_max3_i16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
2546 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
2547
2548 v_max3_i16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
2549 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
2550
2551 v_max3_i16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2552 // GFX12: [0x05,0x00,0x4d,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
2553
2554 v_max3_i16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2555 // GFX12: [0xff,0x00,0x4d,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2556
2557 v_max3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2558 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2559
2560 v_max3_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2561 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2562
2563 v_max3_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2564 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2565
2566 v_max3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2567 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2568
2569 v_max3_i32_e64_dpp v5, v1, v2, v3 row_mirror
2570 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2571
2572 v_max3_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
2573 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2574
2575 v_max3_i32_e64_dpp v5, v1, v2, s105 row_shl:1
2576 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2577
2578 v_max3_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2579 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2580
2581 v_max3_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2582 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2583
2584 v_max3_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2585 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2586
2587 v_max3_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
2588 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2589
2590 v_max3_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
2591 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2592
2593 v_max3_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2594 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2595
2596 v_max3_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2597 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2598
2599 v_max3_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2600 // GFX12: [0x05,0x00,0x1d,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2601
2602 v_max3_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2603 // GFX12: [0xff,0x00,0x1d,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2604
2605 v_max3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2606 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2607
2608 v_max3_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2609 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2610
2611 v_max3_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2612 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2613
2614 v_max3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2615 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2616
2617 v_max3_u16_e64_dpp v5, v1, v2, v3 row_mirror
2618 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2619
2620 v_max3_u16_e64_dpp v5, v1, v2, v3 row_half_mirror
2621 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
2622
2623 v_max3_u16_e64_dpp v5, v1, v2, v255 row_shl:1
2624 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
2625
2626 v_max3_u16_e64_dpp v5, v1, v2, s105 row_shl:15
2627 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
2628
2629 v_max3_u16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
2630 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
2631
2632 v_max3_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
2633 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
2634
2635 v_max3_u16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
2636 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
2637
2638 v_max3_u16_e64_dpp v5, v1, v2, exec_hi row_ror:15
2639 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
2640
2641 v_max3_u16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
2642 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
2643
2644 v_max3_u16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
2645 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
2646
2647 v_max3_u16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2648 // GFX12: [0x05,0x00,0x4e,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
2649
2650 v_max3_u16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2651 // GFX12: [0xff,0x00,0x4e,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2652
2653 v_max3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2654 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2655
2656 v_max3_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2657 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2658
2659 v_max3_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2660 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2661
2662 v_max3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2663 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2664
2665 v_max3_u32_e64_dpp v5, v1, v2, v3 row_mirror
2666 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2667
2668 v_max3_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
2669 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2670
2671 v_max3_u32_e64_dpp v5, v1, v2, s105 row_shl:1
2672 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2673
2674 v_max3_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2675 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2676
2677 v_max3_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2678 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2679
2680 v_max3_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2681 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2682
2683 v_max3_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
2684 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2685
2686 v_max3_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
2687 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2688
2689 v_max3_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2690 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2691
2692 v_max3_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2693 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2694
2695 v_max3_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2696 // GFX12: [0x05,0x00,0x1e,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2697
2698 v_max3_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2699 // GFX12: [0xff,0x00,0x1e,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2700
2701 v_max_i16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
2702 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
2703
2704 v_max_i16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
2705 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
2706
2707 v_max_i16_e64_dpp v5, v1, v2 row_mirror
2708 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
2709
2710 v_max_i16_e64_dpp v5, v1, v2 row_half_mirror
2711 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
2712
2713 v_max_i16_e64_dpp v5, v1, v2 row_shl:1
2714 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
2715
2716 v_max_i16_e64_dpp v5, v1, v2 row_shl:15
2717 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
2718
2719 v_max_i16_e64_dpp v5, v1, v2 row_shr:1
2720 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
2721
2722 v_max_i16_e64_dpp v5, v1, v2 row_shr:15
2723 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
2724
2725 v_max_i16_e64_dpp v5, v1, v2 row_ror:1
2726 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
2727
2728 v_max_i16_e64_dpp v5, v1, v2 row_ror:15
2729 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
2730
2731 v_max_i16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
2732 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
2733
2734 v_max_i16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
2735 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
2736
2737 v_max_i16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2738 // GFX12: [0x05,0x00,0x0a,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
2739
2740 v_max_i16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2741 // GFX12: [0xff,0x00,0x0a,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
2742
2743 v_max_u16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
2744 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
2745
2746 v_max_u16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
2747 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
2748
2749 v_max_u16_e64_dpp v5, v1, v2 row_mirror
2750 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
2751
2752 v_max_u16_e64_dpp v5, v1, v2 row_half_mirror
2753 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
2754
2755 v_max_u16_e64_dpp v5, v1, v2 row_shl:1
2756 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
2757
2758 v_max_u16_e64_dpp v5, v1, v2 row_shl:15
2759 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
2760
2761 v_max_u16_e64_dpp v5, v1, v2 row_shr:1
2762 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
2763
2764 v_max_u16_e64_dpp v5, v1, v2 row_shr:15
2765 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
2766
2767 v_max_u16_e64_dpp v5, v1, v2 row_ror:1
2768 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
2769
2770 v_max_u16_e64_dpp v5, v1, v2 row_ror:15
2771 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
2772
2773 v_max_u16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
2774 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
2775
2776 v_max_u16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
2777 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
2778
2779 v_max_u16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2780 // GFX12: [0x05,0x00,0x09,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
2781
2782 v_max_u16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2783 // GFX12: [0xff,0x00,0x09,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
2784
2785 v_maxmin_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2786 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2787
2788 v_maxmin_num_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2789 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2790
2791 v_maxmin_num_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
2792 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
2793
2794 v_maxmin_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2795 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2796
2797 v_maxmin_num_f16_e64_dpp v5, v1, v2, v3 row_mirror
2798 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2799
2800 v_maxmin_num_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
2801 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2802
2803 v_maxmin_num_f16_e64_dpp v5, v1, v2, s105 row_shl:1
2804 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2805
2806 v_maxmin_num_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2807 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2808
2809 v_maxmin_num_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2810 // GFX12: [0x05,0x00,0x6b,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2811
2812 v_maxmin_num_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
2813 // GFX12: [0x05,0x01,0x6b,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
2814
2815 v_maxmin_num_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
2816 // GFX12: [0x05,0x02,0x6b,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
2817
2818 v_maxmin_num_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
2819 // GFX12: [0x05,0x04,0x6b,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
2820
2821 v_maxmin_num_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
2822 // GFX12: [0x05,0x03,0x6b,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
2823
2824 v_maxmin_num_f16_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2825 // GFX12: [0x05,0x05,0x6b,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
2826
2827 v_maxmin_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2828 // GFX12: [0x05,0x06,0x6b,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
2829
2830 v_maxmin_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2831 // GFX12: [0xff,0x87,0x6b,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
2832
2833 v_maxmin_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2834 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2835
2836 v_maxmin_num_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2837 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2838
2839 v_maxmin_num_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
2840 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
2841
2842 v_maxmin_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2843 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2844
2845 v_maxmin_num_f32_e64_dpp v5, v1, v2, v3 row_mirror
2846 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2847
2848 v_maxmin_num_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
2849 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2850
2851 v_maxmin_num_f32_e64_dpp v5, v1, v2, s105 row_shl:1
2852 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2853
2854 v_maxmin_num_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2855 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2856
2857 v_maxmin_num_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2858 // GFX12: [0x05,0x00,0x69,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2859
2860 v_maxmin_num_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
2861 // GFX12: [0x05,0x01,0x69,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
2862
2863 v_maxmin_num_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
2864 // GFX12: [0x05,0x02,0x69,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
2865
2866 v_maxmin_num_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
2867 // GFX12: [0x05,0x04,0x69,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
2868
2869 v_maxmin_num_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
2870 // GFX12: [0x05,0x03,0x69,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
2871
2872 v_maxmin_num_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
2873 // GFX12: [0x05,0x05,0x69,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
2874
2875 v_maxmin_num_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2876 // GFX12: [0x05,0x06,0x69,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
2877
2878 v_maxmin_num_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2879 // GFX12: [0xff,0x87,0x69,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
2880
2881 v_maxmin_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2882 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2883
2884 v_maxmin_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2885 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2886
2887 v_maxmin_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2888 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2889
2890 v_maxmin_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2891 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2892
2893 v_maxmin_i32_e64_dpp v5, v1, v2, v3 row_mirror
2894 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2895
2896 v_maxmin_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
2897 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2898
2899 v_maxmin_i32_e64_dpp v5, v1, v2, s105 row_shl:1
2900 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2901
2902 v_maxmin_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2903 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2904
2905 v_maxmin_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2906 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2907
2908 v_maxmin_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2909 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2910
2911 v_maxmin_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
2912 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2913
2914 v_maxmin_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
2915 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2916
2917 v_maxmin_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2918 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2919
2920 v_maxmin_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2921 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2922
2923 v_maxmin_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2924 // GFX12: [0x05,0x00,0x64,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2925
2926 v_maxmin_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2927 // GFX12: [0xff,0x00,0x64,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2928
2929 v_maxmin_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
2930 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
2931
2932 v_maxmin_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
2933 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
2934
2935 v_maxmin_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
2936 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
2937
2938 v_maxmin_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
2939 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
2940
2941 v_maxmin_u32_e64_dpp v5, v1, v2, v3 row_mirror
2942 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
2943
2944 v_maxmin_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
2945 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
2946
2947 v_maxmin_u32_e64_dpp v5, v1, v2, s105 row_shl:1
2948 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
2949
2950 v_maxmin_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
2951 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
2952
2953 v_maxmin_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
2954 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
2955
2956 v_maxmin_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
2957 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
2958
2959 v_maxmin_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
2960 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
2961
2962 v_maxmin_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
2963 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
2964
2965 v_maxmin_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
2966 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
2967
2968 v_maxmin_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
2969 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
2970
2971 v_maxmin_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
2972 // GFX12: [0x05,0x00,0x62,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
2973
2974 v_maxmin_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
2975 // GFX12: [0xff,0x00,0x62,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
2976
2977 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
2978 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
2979
2980 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
2981 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
2982
2983 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_mirror
2984 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
2985
2986 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_half_mirror
2987 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
2988
2989 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_shl:1
2990 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
2991
2992 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_shl:15
2993 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
2994
2995 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_shr:1
2996 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
2997
2998 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_shr:15
2999 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
3000
3001 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_ror:1
3002 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
3003
3004 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_ror:15
3005 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
3006
3007 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
3008 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
3009
3010 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
3011 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
3012
3013 v_mbcnt_hi_u32_b32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3014 // GFX12: [0x05,0x00,0x20,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
3015
3016 v_mbcnt_hi_u32_b32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3017 // GFX12: [0xff,0x00,0x20,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
3018
3019 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
3020 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
3021
3022 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
3023 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
3024
3025 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_mirror
3026 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
3027
3028 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_half_mirror
3029 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
3030
3031 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_shl:1
3032 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
3033
3034 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_shl:15
3035 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
3036
3037 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_shr:1
3038 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
3039
3040 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_shr:15
3041 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
3042
3043 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_ror:1
3044 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
3045
3046 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_ror:15
3047 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
3048
3049 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
3050 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
3051
3052 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
3053 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
3054
3055 v_mbcnt_lo_u32_b32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3056 // GFX12: [0x05,0x00,0x1f,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
3057
3058 v_mbcnt_lo_u32_b32_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3059 // GFX12: [0xff,0x00,0x1f,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
3060
3061 v_med3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3062 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3063
3064 v_med3_num_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3065 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3066
3067 v_med3_num_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3068 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3069
3070 v_med3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3071 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3072
3073 v_med3_num_f16_e64_dpp v5, v1, v2, v3 row_mirror
3074 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3075
3076 v_med3_num_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
3077 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3078
3079 v_med3_num_f16_e64_dpp v5, v1, v2, s105 row_shl:1
3080 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3081
3082 v_med3_num_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3083 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3084
3085 v_med3_num_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3086 // GFX12: [0x05,0x00,0x32,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3087
3088 v_med3_num_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3089 // GFX12: [0x05,0x01,0x32,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3090
3091 v_med3_num_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3092 // GFX12: [0x05,0x02,0x32,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3093
3094 v_med3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3095 // GFX12: [0x05,0x04,0x32,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3096
3097 v_med3_num_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3098 // GFX12: [0x05,0x03,0x32,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3099
3100 v_med3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
3101 // GFX12: [0x05,0x05,0x32,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
3102
3103 v_med3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3104 // GFX12: [0x05,0x06,0x32,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
3105
3106 v_med3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3107 // GFX12: [0xff,0x87,0x32,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
3108
3109 v_med3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3110 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3111
3112 v_med3_num_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3113 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3114
3115 v_med3_num_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3116 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3117
3118 v_med3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3119 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3120
3121 v_med3_num_f32_e64_dpp v5, v1, v2, v3 row_mirror
3122 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3123
3124 v_med3_num_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
3125 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3126
3127 v_med3_num_f32_e64_dpp v5, v1, v2, s105 row_shl:1
3128 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3129
3130 v_med3_num_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3131 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3132
3133 v_med3_num_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3134 // GFX12: [0x05,0x00,0x31,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3135
3136 v_med3_num_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3137 // GFX12: [0x05,0x01,0x31,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3138
3139 v_med3_num_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3140 // GFX12: [0x05,0x02,0x31,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3141
3142 v_med3_num_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3143 // GFX12: [0x05,0x04,0x31,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3144
3145 v_med3_num_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3146 // GFX12: [0x05,0x03,0x31,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3147
3148 v_med3_num_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
3149 // GFX12: [0x05,0x05,0x31,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
3150
3151 v_med3_num_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3152 // GFX12: [0x05,0x06,0x31,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
3153
3154 v_med3_num_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3155 // GFX12: [0xff,0x87,0x31,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
3156
3157 v_med3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3158 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3159
3160 v_med3_i16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3161 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3162
3163 v_med3_i16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3164 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3165
3166 v_med3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3167 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3168
3169 v_med3_i16_e64_dpp v5, v1, v2, v3 row_mirror
3170 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3171
3172 v_med3_i16_e64_dpp v5, v1, v2, v3 row_half_mirror
3173 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
3174
3175 v_med3_i16_e64_dpp v5, v1, v2, v255 row_shl:1
3176 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
3177
3178 v_med3_i16_e64_dpp v5, v1, v2, s105 row_shl:15
3179 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
3180
3181 v_med3_i16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
3182 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
3183
3184 v_med3_i16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
3185 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
3186
3187 v_med3_i16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
3188 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
3189
3190 v_med3_i16_e64_dpp v5, v1, v2, exec_hi row_ror:15
3191 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
3192
3193 v_med3_i16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
3194 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
3195
3196 v_med3_i16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
3197 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
3198
3199 v_med3_i16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3200 // GFX12: [0x05,0x00,0x50,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
3201
3202 v_med3_i16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3203 // GFX12: [0xff,0x00,0x50,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3204
3205 v_med3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3206 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3207
3208 v_med3_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3209 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3210
3211 v_med3_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3212 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3213
3214 v_med3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3215 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3216
3217 v_med3_i32_e64_dpp v5, v1, v2, v3 row_mirror
3218 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3219
3220 v_med3_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
3221 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3222
3223 v_med3_i32_e64_dpp v5, v1, v2, s105 row_shl:1
3224 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3225
3226 v_med3_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3227 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3228
3229 v_med3_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3230 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3231
3232 v_med3_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3233 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3234
3235 v_med3_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3236 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3237
3238 v_med3_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3239 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3240
3241 v_med3_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3242 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3243
3244 v_med3_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3245 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3246
3247 v_med3_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3248 // GFX12: [0x05,0x00,0x20,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3249
3250 v_med3_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3251 // GFX12: [0xff,0x00,0x20,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3252
3253 v_med3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3254 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3255
3256 v_med3_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3257 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3258
3259 v_med3_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3260 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3261
3262 v_med3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3263 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3264
3265 v_med3_u16_e64_dpp v5, v1, v2, v3 row_mirror
3266 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3267
3268 v_med3_u16_e64_dpp v5, v1, v2, v3 row_half_mirror
3269 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
3270
3271 v_med3_u16_e64_dpp v5, v1, v2, v255 row_shl:1
3272 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
3273
3274 v_med3_u16_e64_dpp v5, v1, v2, s105 row_shl:15
3275 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
3276
3277 v_med3_u16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
3278 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
3279
3280 v_med3_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
3281 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
3282
3283 v_med3_u16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
3284 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
3285
3286 v_med3_u16_e64_dpp v5, v1, v2, exec_hi row_ror:15
3287 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
3288
3289 v_med3_u16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
3290 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
3291
3292 v_med3_u16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
3293 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
3294
3295 v_med3_u16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3296 // GFX12: [0x05,0x00,0x51,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
3297
3298 v_med3_u16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3299 // GFX12: [0xff,0x00,0x51,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3300
3301 v_med3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3302 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3303
3304 v_med3_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3305 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3306
3307 v_med3_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3308 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3309
3310 v_med3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3311 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3312
3313 v_med3_u32_e64_dpp v5, v1, v2, v3 row_mirror
3314 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3315
3316 v_med3_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
3317 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3318
3319 v_med3_u32_e64_dpp v5, v1, v2, s105 row_shl:1
3320 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3321
3322 v_med3_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3323 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3324
3325 v_med3_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3326 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3327
3328 v_med3_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3329 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3330
3331 v_med3_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3332 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3333
3334 v_med3_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3335 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3336
3337 v_med3_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3338 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3339
3340 v_med3_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3341 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3342
3343 v_med3_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3344 // GFX12: [0x05,0x00,0x21,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3345
3346 v_med3_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3347 // GFX12: [0xff,0x00,0x21,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3348
3349 v_min3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3350 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3351
3352 v_min3_num_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3353 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3354
3355 v_min3_num_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3356 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3357
3358 v_min3_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3359 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3360
3361 v_min3_num_f16_e64_dpp v5, v1, v2, v3 row_mirror
3362 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3363
3364 v_min3_num_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
3365 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3366
3367 v_min3_num_f16_e64_dpp v5, v1, v2, s105 row_shl:1
3368 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3369
3370 v_min3_num_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3371 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3372
3373 v_min3_num_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3374 // GFX12: [0x05,0x00,0x2b,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3375
3376 v_min3_num_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3377 // GFX12: [0x05,0x01,0x2b,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3378
3379 v_min3_num_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3380 // GFX12: [0x05,0x02,0x2b,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3381
3382 v_min3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3383 // GFX12: [0x05,0x04,0x2b,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3384
3385 v_min3_num_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3386 // GFX12: [0x05,0x03,0x2b,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3387
3388 v_min3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
3389 // GFX12: [0x05,0x05,0x2b,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
3390
3391 v_min3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3392 // GFX12: [0x05,0x06,0x2b,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
3393
3394 v_min3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3395 // GFX12: [0xff,0x87,0x2b,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
3396
3397 v_min3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3398 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3399
3400 v_min3_num_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3401 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3402
3403 v_min3_num_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3404 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3405
3406 v_min3_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3407 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3408
3409 v_min3_num_f32_e64_dpp v5, v1, v2, v3 row_mirror
3410 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3411
3412 v_min3_num_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
3413 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3414
3415 v_min3_num_f32_e64_dpp v5, v1, v2, s105 row_shl:1
3416 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3417
3418 v_min3_num_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3419 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3420
3421 v_min3_num_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3422 // GFX12: [0x05,0x00,0x29,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3423
3424 v_min3_num_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3425 // GFX12: [0x05,0x01,0x29,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3426
3427 v_min3_num_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3428 // GFX12: [0x05,0x02,0x29,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3429
3430 v_min3_num_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3431 // GFX12: [0x05,0x04,0x29,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3432
3433 v_min3_num_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3434 // GFX12: [0x05,0x03,0x29,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3435
3436 v_min3_num_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
3437 // GFX12: [0x05,0x05,0x29,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
3438
3439 v_min3_num_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3440 // GFX12: [0x05,0x06,0x29,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
3441
3442 v_min3_num_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3443 // GFX12: [0xff,0x87,0x29,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
3444
3445 v_min3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3446 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3447
3448 v_min3_i16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3449 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3450
3451 v_min3_i16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3452 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3453
3454 v_min3_i16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3455 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3456
3457 v_min3_i16_e64_dpp v5, v1, v2, v3 row_mirror
3458 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3459
3460 v_min3_i16_e64_dpp v5, v1, v2, v3 row_half_mirror
3461 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
3462
3463 v_min3_i16_e64_dpp v5, v1, v2, v255 row_shl:1
3464 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
3465
3466 v_min3_i16_e64_dpp v5, v1, v2, s105 row_shl:15
3467 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
3468
3469 v_min3_i16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
3470 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
3471
3472 v_min3_i16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
3473 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
3474
3475 v_min3_i16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
3476 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
3477
3478 v_min3_i16_e64_dpp v5, v1, v2, exec_hi row_ror:15
3479 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
3480
3481 v_min3_i16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
3482 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
3483
3484 v_min3_i16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
3485 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
3486
3487 v_min3_i16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3488 // GFX12: [0x05,0x00,0x4a,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
3489
3490 v_min3_i16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3491 // GFX12: [0xff,0x00,0x4a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3492
3493 v_min3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3494 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3495
3496 v_min3_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3497 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3498
3499 v_min3_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3500 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3501
3502 v_min3_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3503 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3504
3505 v_min3_i32_e64_dpp v5, v1, v2, v3 row_mirror
3506 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3507
3508 v_min3_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
3509 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3510
3511 v_min3_i32_e64_dpp v5, v1, v2, s105 row_shl:1
3512 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3513
3514 v_min3_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3515 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3516
3517 v_min3_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3518 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3519
3520 v_min3_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3521 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3522
3523 v_min3_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3524 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3525
3526 v_min3_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3527 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3528
3529 v_min3_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3530 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3531
3532 v_min3_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3533 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3534
3535 v_min3_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3536 // GFX12: [0x05,0x00,0x1a,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3537
3538 v_min3_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3539 // GFX12: [0xff,0x00,0x1a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3540
3541 v_min3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3542 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3543
3544 v_min3_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3545 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3546
3547 v_min3_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3548 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3549
3550 v_min3_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3551 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3552
3553 v_min3_u16_e64_dpp v5, v1, v2, v3 row_mirror
3554 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3555
3556 v_min3_u16_e64_dpp v5, v1, v2, v3 row_half_mirror
3557 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x41,0x01,0xff]
3558
3559 v_min3_u16_e64_dpp v5, v1, v2, v255 row_shl:1
3560 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x01,0x01,0xff]
3561
3562 v_min3_u16_e64_dpp v5, v1, v2, s105 row_shl:15
3563 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x0f,0x01,0xff]
3564
3565 v_min3_u16_e64_dpp v5, v1, v2, vcc_hi row_shr:1
3566 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xae,0x01,0x01,0x11,0x01,0xff]
3567
3568 v_min3_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:15
3569 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x1f,0x01,0xff]
3570
3571 v_min3_u16_e64_dpp v5, v1, v2, ttmp15 row_ror:1
3572 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xee,0x01,0x01,0x21,0x01,0xff]
3573
3574 v_min3_u16_e64_dpp v5, v1, v2, exec_hi row_ror:15
3575 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
3576
3577 v_min3_u16_e64_dpp v5, v1, v2, exec_lo row_share:0 row_mask:0xf bank_mask:0xf
3578 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
3579
3580 v_min3_u16_e64_dpp v5, v1, v2, null row_share:15 row_mask:0x0 bank_mask:0x1
3581 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
3582
3583 v_min3_u16_e64_dpp v5, v1, v2, -1 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3584 // GFX12: [0x05,0x00,0x4b,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x09,0x13]
3585
3586 v_min3_u16_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3587 // GFX12: [0xff,0x00,0x4b,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3588
3589 v_min3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3590 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3591
3592 v_min3_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3593 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3594
3595 v_min3_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3596 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3597
3598 v_min3_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3599 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3600
3601 v_min3_u32_e64_dpp v5, v1, v2, v3 row_mirror
3602 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3603
3604 v_min3_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
3605 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3606
3607 v_min3_u32_e64_dpp v5, v1, v2, s105 row_shl:1
3608 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3609
3610 v_min3_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3611 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3612
3613 v_min3_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3614 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3615
3616 v_min3_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3617 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3618
3619 v_min3_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3620 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3621
3622 v_min3_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3623 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3624
3625 v_min3_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3626 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3627
3628 v_min3_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3629 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3630
3631 v_min3_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3632 // GFX12: [0x05,0x00,0x1b,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3633
3634 v_min3_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3635 // GFX12: [0xff,0x00,0x1b,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3636
3637 v_min_i16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
3638 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
3639
3640 v_min_i16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
3641 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
3642
3643 v_min_i16_e64_dpp v5, v1, v2 row_mirror
3644 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
3645
3646 v_min_i16_e64_dpp v5, v1, v2 row_half_mirror
3647 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
3648
3649 v_min_i16_e64_dpp v5, v1, v2 row_shl:1
3650 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
3651
3652 v_min_i16_e64_dpp v5, v1, v2 row_shl:15
3653 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
3654
3655 v_min_i16_e64_dpp v5, v1, v2 row_shr:1
3656 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
3657
3658 v_min_i16_e64_dpp v5, v1, v2 row_shr:15
3659 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
3660
3661 v_min_i16_e64_dpp v5, v1, v2 row_ror:1
3662 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
3663
3664 v_min_i16_e64_dpp v5, v1, v2 row_ror:15
3665 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
3666
3667 v_min_i16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
3668 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
3669
3670 v_min_i16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
3671 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
3672
3673 v_min_i16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3674 // GFX12: [0x05,0x00,0x0c,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
3675
3676 v_min_i16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3677 // GFX12: [0xff,0x00,0x0c,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
3678
3679 v_min_u16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
3680 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
3681
3682 v_min_u16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
3683 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
3684
3685 v_min_u16_e64_dpp v5, v1, v2 row_mirror
3686 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
3687
3688 v_min_u16_e64_dpp v5, v1, v2 row_half_mirror
3689 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
3690
3691 v_min_u16_e64_dpp v5, v1, v2 row_shl:1
3692 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
3693
3694 v_min_u16_e64_dpp v5, v1, v2 row_shl:15
3695 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
3696
3697 v_min_u16_e64_dpp v5, v1, v2 row_shr:1
3698 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
3699
3700 v_min_u16_e64_dpp v5, v1, v2 row_shr:15
3701 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
3702
3703 v_min_u16_e64_dpp v5, v1, v2 row_ror:1
3704 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
3705
3706 v_min_u16_e64_dpp v5, v1, v2 row_ror:15
3707 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
3708
3709 v_min_u16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
3710 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
3711
3712 v_min_u16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
3713 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
3714
3715 v_min_u16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3716 // GFX12: [0x05,0x00,0x0b,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
3717
3718 v_min_u16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3719 // GFX12: [0xff,0x00,0x0b,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
3720
3721 v_minmax_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3722 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3723
3724 v_minmax_num_f16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3725 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3726
3727 v_minmax_num_f16_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3728 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3729
3730 v_minmax_num_f16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3731 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3732
3733 v_minmax_num_f16_e64_dpp v5, v1, v2, v3 row_mirror
3734 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3735
3736 v_minmax_num_f16_e64_dpp v5, v1, v2, v255 row_half_mirror
3737 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3738
3739 v_minmax_num_f16_e64_dpp v5, v1, v2, s105 row_shl:1
3740 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3741
3742 v_minmax_num_f16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3743 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3744
3745 v_minmax_num_f16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3746 // GFX12: [0x05,0x00,0x6a,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3747
3748 v_minmax_num_f16_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3749 // GFX12: [0x05,0x01,0x6a,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3750
3751 v_minmax_num_f16_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3752 // GFX12: [0x05,0x02,0x6a,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3753
3754 v_minmax_num_f16_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3755 // GFX12: [0x05,0x04,0x6a,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3756
3757 v_minmax_num_f16_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3758 // GFX12: [0x05,0x03,0x6a,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3759
3760 v_minmax_num_f16_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
3761 // GFX12: [0x05,0x05,0x6a,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
3762
3763 v_minmax_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3764 // GFX12: [0x05,0x06,0x6a,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
3765
3766 v_minmax_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3767 // GFX12: [0xff,0x87,0x6a,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
3768
3769 v_minmax_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3770 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3771
3772 v_minmax_num_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3773 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3774
3775 v_minmax_num_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
3776 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
3777
3778 v_minmax_num_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3779 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3780
3781 v_minmax_num_f32_e64_dpp v5, v1, v2, v3 row_mirror
3782 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3783
3784 v_minmax_num_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
3785 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3786
3787 v_minmax_num_f32_e64_dpp v5, v1, v2, s105 row_shl:1
3788 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3789
3790 v_minmax_num_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3791 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3792
3793 v_minmax_num_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3794 // GFX12: [0x05,0x00,0x68,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3795
3796 v_minmax_num_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
3797 // GFX12: [0x05,0x01,0x68,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
3798
3799 v_minmax_num_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
3800 // GFX12: [0x05,0x02,0x68,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
3801
3802 v_minmax_num_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
3803 // GFX12: [0x05,0x04,0x68,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
3804
3805 v_minmax_num_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
3806 // GFX12: [0x05,0x03,0x68,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
3807
3808 v_minmax_num_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
3809 // GFX12: [0x05,0x05,0x68,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
3810
3811 v_minmax_num_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3812 // GFX12: [0x05,0x06,0x68,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
3813
3814 v_minmax_num_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3815 // GFX12: [0xff,0x87,0x68,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
3816
3817 v_minmax_i32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3818 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3819
3820 v_minmax_i32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3821 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3822
3823 v_minmax_i32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3824 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3825
3826 v_minmax_i32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3827 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3828
3829 v_minmax_i32_e64_dpp v5, v1, v2, v3 row_mirror
3830 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3831
3832 v_minmax_i32_e64_dpp v5, v1, v2, v255 row_half_mirror
3833 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3834
3835 v_minmax_i32_e64_dpp v5, v1, v2, s105 row_shl:1
3836 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3837
3838 v_minmax_i32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3839 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3840
3841 v_minmax_i32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3842 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3843
3844 v_minmax_i32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3845 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3846
3847 v_minmax_i32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3848 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3849
3850 v_minmax_i32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3851 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3852
3853 v_minmax_i32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3854 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3855
3856 v_minmax_i32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3857 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3858
3859 v_minmax_i32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3860 // GFX12: [0x05,0x00,0x65,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3861
3862 v_minmax_i32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3863 // GFX12: [0xff,0x00,0x65,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3864
3865 v_minmax_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3866 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3867
3868 v_minmax_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3869 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3870
3871 v_minmax_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
3872 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
3873
3874 v_minmax_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3875 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3876
3877 v_minmax_u32_e64_dpp v5, v1, v2, v3 row_mirror
3878 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3879
3880 v_minmax_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
3881 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3882
3883 v_minmax_u32_e64_dpp v5, v1, v2, s105 row_shl:1
3884 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3885
3886 v_minmax_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3887 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3888
3889 v_minmax_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3890 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3891
3892 v_minmax_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3893 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3894
3895 v_minmax_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
3896 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3897
3898 v_minmax_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
3899 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3900
3901 v_minmax_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3902 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3903
3904 v_minmax_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3905 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3906
3907 v_minmax_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3908 // GFX12: [0x05,0x00,0x63,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3909
3910 v_minmax_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3911 // GFX12: [0xff,0x00,0x63,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3912
3913 v_msad_u8_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
3914 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
3915
3916 v_msad_u8_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
3917 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
3918
3919 v_msad_u8_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
3920 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
3921
3922 v_msad_u8_e64_dpp v5, v1, v2, v3 row_mirror
3923 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
3924
3925 v_msad_u8_e64_dpp v5, v1, v2, v255 row_half_mirror
3926 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
3927
3928 v_msad_u8_e64_dpp v5, v1, v2, s105 row_shl:1
3929 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
3930
3931 v_msad_u8_e64_dpp v5, v1, v2, vcc_hi row_shl:15
3932 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
3933
3934 v_msad_u8_e64_dpp v5, v1, v2, vcc_lo row_shr:1
3935 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
3936
3937 v_msad_u8_e64_dpp v5, v1, v2, ttmp15 row_shr:15
3938 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
3939
3940 v_msad_u8_e64_dpp v5, v1, v2, exec_hi row_ror:1
3941 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
3942
3943 v_msad_u8_e64_dpp v5, v1, v2, exec_lo row_ror:15
3944 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
3945
3946 v_msad_u8_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
3947 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
3948
3949 v_msad_u8_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
3950 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
3951
3952 v_msad_u8_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3953 // GFX12: [0x05,0x00,0x39,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
3954
3955 v_msad_u8_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3956 // GFX12: [0xff,0x80,0x39,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
3957
3958 v_mul_lo_u16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
3959 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
3960
3961 v_mul_lo_u16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
3962 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
3963
3964 v_mul_lo_u16_e64_dpp v5, v1, v2 row_mirror
3965 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
3966
3967 v_mul_lo_u16_e64_dpp v5, v1, v2 row_half_mirror
3968 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
3969
3970 v_mul_lo_u16_e64_dpp v5, v1, v2 row_shl:1
3971 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
3972
3973 v_mul_lo_u16_e64_dpp v5, v1, v2 row_shl:15
3974 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
3975
3976 v_mul_lo_u16_e64_dpp v5, v1, v2 row_shr:1
3977 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
3978
3979 v_mul_lo_u16_e64_dpp v5, v1, v2 row_shr:15
3980 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
3981
3982 v_mul_lo_u16_e64_dpp v5, v1, v2 row_ror:1
3983 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
3984
3985 v_mul_lo_u16_e64_dpp v5, v1, v2 row_ror:15
3986 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
3987
3988 v_mul_lo_u16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
3989 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
3990
3991 v_mul_lo_u16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
3992 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
3993
3994 v_mul_lo_u16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
3995 // GFX12: [0x05,0x00,0x05,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
3996
3997 v_mul_lo_u16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
3998 // GFX12: [0xff,0x00,0x05,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
3999
4000 v_mullit_f32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4001 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4002
4003 v_mullit_f32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4004 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4005
4006 v_mullit_f32_e64_dpp v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
4007 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
4008
4009 v_mullit_f32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4010 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4011
4012 v_mullit_f32_e64_dpp v5, v1, v2, v3 row_mirror
4013 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4014
4015 v_mullit_f32_e64_dpp v5, v1, v2, v255 row_half_mirror
4016 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4017
4018 v_mullit_f32_e64_dpp v5, v1, v2, s105 row_shl:1
4019 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4020
4021 v_mullit_f32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4022 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4023
4024 v_mullit_f32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4025 // GFX12: [0x05,0x00,0x18,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4026
4027 v_mullit_f32_e64_dpp v5, |v1|, v2, -ttmp15 row_shr:15
4028 // GFX12: [0x05,0x01,0x18,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
4029
4030 v_mullit_f32_e64_dpp v5, v1, -|v2|, exec_hi row_ror:1
4031 // GFX12: [0x05,0x02,0x18,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
4032
4033 v_mullit_f32_e64_dpp v5, -v1, v2, |exec_lo| row_ror:15
4034 // GFX12: [0x05,0x04,0x18,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
4035
4036 v_mullit_f32_e64_dpp v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
4037 // GFX12: [0x05,0x03,0x18,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
4038
4039 v_mullit_f32_e64_dpp v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
4040 // GFX12: [0x05,0x05,0x18,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
4041
4042 v_mullit_f32_e64_dpp v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4043 // GFX12: [0x05,0x06,0x18,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
4044
4045 v_mullit_f32_e64_dpp v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4046 // GFX12: [0xff,0x87,0x18,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
4047
4048 v_or3_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4049 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4050
4051 v_or3_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4052 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4053
4054 v_or3_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4055 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4056
4057 v_or3_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4058 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4059
4060 v_or3_b32_e64_dpp v5, v1, v2, v3 row_mirror
4061 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4062
4063 v_or3_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
4064 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4065
4066 v_or3_b32_e64_dpp v5, v1, v2, s105 row_shl:1
4067 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4068
4069 v_or3_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4070 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4071
4072 v_or3_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4073 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4074
4075 v_or3_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4076 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4077
4078 v_or3_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
4079 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4080
4081 v_or3_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
4082 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4083
4084 v_or3_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4085 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4086
4087 v_or3_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4088 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4089
4090 v_or3_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4091 // GFX12: [0x05,0x00,0x58,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4092
4093 v_or3_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4094 // GFX12: [0xff,0x00,0x58,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4095
4096 v_or_b16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4097 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4098
4099 v_or_b16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4100 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4101
4102 v_or_b16_e64_dpp v5, v1, v2 row_mirror
4103 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4104
4105 v_or_b16_e64_dpp v5, v1, v2 row_half_mirror
4106 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4107
4108 v_or_b16_e64_dpp v5, v1, v2 row_shl:1
4109 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4110
4111 v_or_b16_e64_dpp v5, v1, v2 row_shl:15
4112 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4113
4114 v_or_b16_e64_dpp v5, v1, v2 row_shr:1
4115 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4116
4117 v_or_b16_e64_dpp v5, v1, v2 row_shr:15
4118 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4119
4120 v_or_b16_e64_dpp v5, v1, v2 row_ror:1
4121 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4122
4123 v_or_b16_e64_dpp v5, v1, v2 row_ror:15
4124 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4125
4126 v_or_b16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4127 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4128
4129 v_or_b16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4130 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4131
4132 v_or_b16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4133 // GFX12: [0x05,0x00,0x63,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4134
4135 v_or_b16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4136 // GFX12: [0xff,0x00,0x63,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4137
4138 v_pack_b32_f16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4139 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4140
4141 v_pack_b32_f16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4142 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4143
4144 v_pack_b32_f16_e64_dpp v5, v1, v2 row_mirror
4145 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4146
4147 v_pack_b32_f16_e64_dpp v5, v1, v2 row_half_mirror
4148 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4149
4150 v_pack_b32_f16_e64_dpp v5, v1, v2 row_shl:1
4151 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4152
4153 v_pack_b32_f16_e64_dpp v5, v1, v2 row_shl:15
4154 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4155
4156 v_pack_b32_f16_e64_dpp v5, v1, v2 row_shr:1
4157 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4158
4159 v_pack_b32_f16_e64_dpp v5, v1, v2 row_shr:15
4160 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4161
4162 v_pack_b32_f16_e64_dpp v5, v1, v2 row_ror:1
4163 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4164
4165 v_pack_b32_f16_e64_dpp v5, v1, v2 row_ror:15
4166 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4167
4168 v_pack_b32_f16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4169 // GFX12: [0x05,0x00,0x11,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4170
4171 v_pack_b32_f16_e64_dpp v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
4172 // GFX12: [0x05,0x01,0x11,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
4173
4174 v_pack_b32_f16_e64_dpp v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4175 // GFX12: [0x05,0x02,0x11,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
4176
4177 v_pack_b32_f16_e64_dpp v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4178 // GFX12: [0xff,0x03,0x11,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
4179
4180 v_perm_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4181 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4182
4183 v_perm_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4184 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4185
4186 v_perm_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4187 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4188
4189 v_perm_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4190 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4191
4192 v_perm_b32_e64_dpp v5, v1, v2, v3 row_mirror
4193 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4194
4195 v_perm_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
4196 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4197
4198 v_perm_b32_e64_dpp v5, v1, v2, s105 row_shl:1
4199 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4200
4201 v_perm_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4202 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4203
4204 v_perm_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4205 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4206
4207 v_perm_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4208 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4209
4210 v_perm_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
4211 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4212
4213 v_perm_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
4214 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4215
4216 v_perm_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4217 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4218
4219 v_perm_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4220 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4221
4222 v_perm_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4223 // GFX12: [0x05,0x00,0x44,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4224
4225 v_perm_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4226 // GFX12: [0xff,0x00,0x44,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4227
4228 v_sad_hi_u8_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4229 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4230
4231 v_sad_hi_u8_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4232 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4233
4234 v_sad_hi_u8_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4235 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4236
4237 v_sad_hi_u8_e64_dpp v5, v1, v2, v3 row_mirror
4238 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4239
4240 v_sad_hi_u8_e64_dpp v5, v1, v2, v255 row_half_mirror
4241 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4242
4243 v_sad_hi_u8_e64_dpp v5, v1, v2, s105 row_shl:1
4244 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4245
4246 v_sad_hi_u8_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4247 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4248
4249 v_sad_hi_u8_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4250 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4251
4252 v_sad_hi_u8_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4253 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4254
4255 v_sad_hi_u8_e64_dpp v5, v1, v2, exec_hi row_ror:1
4256 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4257
4258 v_sad_hi_u8_e64_dpp v5, v1, v2, exec_lo row_ror:15
4259 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4260
4261 v_sad_hi_u8_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4262 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4263
4264 v_sad_hi_u8_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4265 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4266
4267 v_sad_hi_u8_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4268 // GFX12: [0x05,0x00,0x23,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4269
4270 v_sad_hi_u8_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4271 // GFX12: [0xff,0x80,0x23,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4272
4273 v_sad_u16_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4274 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4275
4276 v_sad_u16_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4277 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4278
4279 v_sad_u16_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4280 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4281
4282 v_sad_u16_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4283 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4284
4285 v_sad_u16_e64_dpp v5, v1, v2, v3 row_mirror
4286 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4287
4288 v_sad_u16_e64_dpp v5, v1, v2, v255 row_half_mirror
4289 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4290
4291 v_sad_u16_e64_dpp v5, v1, v2, s105 row_shl:1
4292 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4293
4294 v_sad_u16_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4295 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4296
4297 v_sad_u16_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4298 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4299
4300 v_sad_u16_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4301 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4302
4303 v_sad_u16_e64_dpp v5, v1, v2, exec_hi row_ror:1
4304 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4305
4306 v_sad_u16_e64_dpp v5, v1, v2, exec_lo row_ror:15
4307 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4308
4309 v_sad_u16_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4310 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4311
4312 v_sad_u16_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4313 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4314
4315 v_sad_u16_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4316 // GFX12: [0x05,0x00,0x24,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4317
4318 v_sad_u16_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4319 // GFX12: [0xff,0x80,0x24,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4320
4321 v_sad_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4322 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4323
4324 v_sad_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4325 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4326
4327 v_sad_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4328 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4329
4330 v_sad_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4331 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4332
4333 v_sad_u32_e64_dpp v5, v1, v2, v3 row_mirror
4334 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4335
4336 v_sad_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
4337 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4338
4339 v_sad_u32_e64_dpp v5, v1, v2, s105 row_shl:1
4340 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4341
4342 v_sad_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4343 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4344
4345 v_sad_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4346 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4347
4348 v_sad_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4349 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4350
4351 v_sad_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
4352 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4353
4354 v_sad_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
4355 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4356
4357 v_sad_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4358 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4359
4360 v_sad_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4361 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4362
4363 v_sad_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4364 // GFX12: [0x05,0x00,0x25,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4365
4366 v_sad_u32_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4367 // GFX12: [0xff,0x80,0x25,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4368
4369 v_sad_u8_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4370 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4371
4372 v_sad_u8_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4373 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4374
4375 v_sad_u8_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4376 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4377
4378 v_sad_u8_e64_dpp v5, v1, v2, v3 row_mirror
4379 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4380
4381 v_sad_u8_e64_dpp v5, v1, v2, v255 row_half_mirror
4382 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4383
4384 v_sad_u8_e64_dpp v5, v1, v2, s105 row_shl:1
4385 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4386
4387 v_sad_u8_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4388 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4389
4390 v_sad_u8_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4391 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4392
4393 v_sad_u8_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4394 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4395
4396 v_sad_u8_e64_dpp v5, v1, v2, exec_hi row_ror:1
4397 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4398
4399 v_sad_u8_e64_dpp v5, v1, v2, exec_lo row_ror:15
4400 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4401
4402 v_sad_u8_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4403 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4404
4405 v_sad_u8_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4406 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4407
4408 v_sad_u8_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4409 // GFX12: [0x05,0x00,0x22,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4410
4411 v_sad_u8_e64_dpp v255, v255, v255, src_scc clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4412 // GFX12: [0xff,0x80,0x22,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4413
4414 v_sub_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[3,2,1,0]
4415 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4416 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4417
4418 v_sub_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[0,1,2,3]
4419 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4420 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4421
4422 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_mirror
4423 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4424 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4425
4426 v_sub_co_u32_e64_dpp v5, s6, v1, s2 row_mirror
4427 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x00,0x00,0x01,0x40,0x01,0xff]
4428 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4429
4430 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_half_mirror
4431 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4432 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4433
4434 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_shl:1
4435 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4436 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4437
4438 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_shl:15
4439 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4440 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4441
4442 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_shr:1
4443 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4444 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4445
4446 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_shr:15
4447 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4448 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4449
4450 v_sub_co_u32_e64_dpp v5, s6, v1, v2 row_ror:1
4451 // W32: [0x05,0x06,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4452 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4453
4454 v_sub_co_u32_e64_dpp v5, s105, v1, v2 row_ror:15
4455 // W32: [0x05,0x69,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4456 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4457
4458 v_sub_co_u32_e64_dpp v5, vcc_lo, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4459 // W32: [0x05,0x6a,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4460 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4461
4462 v_sub_co_u32_e64_dpp v5, vcc_hi, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4463 // W32: [0x05,0x6b,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4464 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4465
4466 v_sub_co_u32_e64_dpp v5, ttmp15, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4467 // W32: [0x05,0x7b,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4468 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4469
4470 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[3,2,1,0]
4471 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4472 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4473
4474 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[0,1,2,3]
4475 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4476 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4477
4478 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_mirror
4479 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4480 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4481
4482 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_half_mirror
4483 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4484 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4485
4486 v_sub_co_u32_e64_dpp v5, s[12:13], v1, s2 row_half_mirror
4487 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x00,0x00,0x01,0x41,0x01,0xff]
4488 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4489
4490 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:1
4491 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4492 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4493
4494 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:15
4495 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4496 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4497
4498 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:1
4499 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4500 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4501
4502 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:15
4503 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4504 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4505
4506 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:1
4507 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4508 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4509
4510 v_sub_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:15
4511 // W64: [0x05,0x0c,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4512 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4513
4514 v_sub_co_u32_e64_dpp v5, s[104:105], v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4515 // W64: [0x05,0x68,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4516 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4517
4518 v_sub_co_u32_e64_dpp v5, vcc, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4519 // W64: [0x05,0x6a,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4520 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4521
4522 v_sub_co_u32_e64_dpp v5, ttmp[14:15], v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4523 // W64: [0x05,0x7a,0x01,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4524 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4525
4526 v_sub_co_u32_e64_dpp v255, null, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4527 // GFX12: [0xff,0xfc,0x01,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4528
4529 v_sub_nc_i16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4530 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4531
4532 v_sub_nc_i16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4533 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4534
4535 v_sub_nc_i16_e64_dpp v5, v1, v2 row_mirror
4536 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4537
4538 v_sub_nc_i16_e64_dpp v5, v1, v2 row_half_mirror
4539 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4540
4541 v_sub_nc_i16_e64_dpp v5, v1, v2 row_shl:1
4542 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4543
4544 v_sub_nc_i16_e64_dpp v5, v1, v2 row_shl:15
4545 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4546
4547 v_sub_nc_i16_e64_dpp v5, v1, v2 row_shr:1
4548 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4549
4550 v_sub_nc_i16_e64_dpp v5, v1, v2 row_shr:15
4551 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4552
4553 v_sub_nc_i16_e64_dpp v5, v1, v2 row_ror:1
4554 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4555
4556 v_sub_nc_i16_e64_dpp v5, v1, v2 row_ror:15
4557 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4558
4559 v_sub_nc_i16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4560 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4561
4562 v_sub_nc_i16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4563 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4564
4565 v_sub_nc_i16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4566 // GFX12: [0x05,0x00,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4567
4568 v_sub_nc_i16_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4569 // GFX12: [0xff,0x80,0x0e,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4570
4571 v_sub_nc_i32_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4572 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4573
4574 v_sub_nc_i32_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4575 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4576
4577 v_sub_nc_i32_e64_dpp v5, v1, v2 row_mirror
4578 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4579
4580 v_sub_nc_i32_e64_dpp v5, v1, v2 row_half_mirror
4581 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4582
4583 v_sub_nc_i32_e64_dpp v5, v1, v2 row_shl:1
4584 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4585
4586 v_sub_nc_i32_e64_dpp v5, v1, v2 row_shl:15
4587 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4588
4589 v_sub_nc_i32_e64_dpp v5, v1, v2 row_shr:1
4590 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4591
4592 v_sub_nc_i32_e64_dpp v5, v1, v2 row_shr:15
4593 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4594
4595 v_sub_nc_i32_e64_dpp v5, v1, v2 row_ror:1
4596 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4597
4598 v_sub_nc_i32_e64_dpp v5, v1, v2 row_ror:15
4599 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4600
4601 v_sub_nc_i32_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4602 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4603
4604 v_sub_nc_i32_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4605 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4606
4607 v_sub_nc_i32_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4608 // GFX12: [0x05,0x00,0x25,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4609
4610 v_sub_nc_i32_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4611 // GFX12: [0xff,0x80,0x25,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4612
4613 v_sub_nc_u16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4614 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4615
4616 v_sub_nc_u16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4617 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4618
4619 v_sub_nc_u16_e64_dpp v5, v1, v2 row_mirror
4620 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4621
4622 v_sub_nc_u16_e64_dpp v5, v1, v2 row_half_mirror
4623 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4624
4625 v_sub_nc_u16_e64_dpp v5, v1, v2 row_shl:1
4626 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4627
4628 v_sub_nc_u16_e64_dpp v5, v1, v2 row_shl:15
4629 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4630
4631 v_sub_nc_u16_e64_dpp v5, v1, v2 row_shr:1
4632 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4633
4634 v_sub_nc_u16_e64_dpp v5, v1, v2 row_shr:15
4635 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4636
4637 v_sub_nc_u16_e64_dpp v5, v1, v2 row_ror:1
4638 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4639
4640 v_sub_nc_u16_e64_dpp v5, v1, v2 row_ror:15
4641 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4642
4643 v_sub_nc_u16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4644 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4645
4646 v_sub_nc_u16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4647 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4648
4649 v_sub_nc_u16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4650 // GFX12: [0x05,0x00,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4651
4652 v_sub_nc_u16_e64_dpp v255, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4653 // GFX12: [0xff,0x80,0x04,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4654
4655 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[3,2,1,0]
4656 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4657 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4658
4659 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 quad_perm:[0,1,2,3]
4660 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4661 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4662
4663 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_mirror
4664 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4665 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4666
4667 v_subrev_co_u32_e64_dpp v5, s6, v1, s2 row_mirror
4668 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x00,0x00,0x01,0x40,0x01,0xff]
4669 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4670
4671 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_half_mirror
4672 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4673 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4674
4675 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_shl:1
4676 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4677 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4678
4679 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_shl:15
4680 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4681 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4682
4683 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_shr:1
4684 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4685 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4686
4687 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_shr:15
4688 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4689 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4690
4691 v_subrev_co_u32_e64_dpp v5, s6, v1, v2 row_ror:1
4692 // W32: [0x05,0x06,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4693 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4694
4695 v_subrev_co_u32_e64_dpp v5, s105, v1, v2 row_ror:15
4696 // W32: [0x05,0x69,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4697 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4698
4699 v_subrev_co_u32_e64_dpp v5, vcc_lo, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4700 // W32: [0x05,0x6a,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4701 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4702
4703 v_subrev_co_u32_e64_dpp v5, vcc_hi, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4704 // W32: [0x05,0x6b,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4705 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4706
4707 v_subrev_co_u32_e64_dpp v5, ttmp15, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4708 // W32: [0x05,0x7b,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4709 // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4710
4711 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[3,2,1,0]
4712 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4713 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4714
4715 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 quad_perm:[0,1,2,3]
4716 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4717 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4718
4719 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_mirror
4720 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4721 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4722
4723 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_half_mirror
4724 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4725 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4726
4727 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, s2 row_half_mirror
4728 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x00,0x00,0x01,0x41,0x01,0xff]
4729 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4730
4731 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:1
4732 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4733 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4734
4735 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shl:15
4736 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4737 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4738
4739 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:1
4740 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4741 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4742
4743 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_shr:15
4744 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4745 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4746
4747 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:1
4748 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4749 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4750
4751 v_subrev_co_u32_e64_dpp v5, s[12:13], v1, v2 row_ror:15
4752 // W64: [0x05,0x0c,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4753 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4754
4755 v_subrev_co_u32_e64_dpp v5, s[104:105], v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4756 // W64: [0x05,0x68,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4757 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4758
4759 v_subrev_co_u32_e64_dpp v5, vcc, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4760 // W64: [0x05,0x6a,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4761 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4762
4763 v_subrev_co_u32_e64_dpp v5, ttmp[14:15], v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4764 // W64: [0x05,0x7a,0x02,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4765 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction
4766
4767 v_subrev_co_u32_e64_dpp v255, null, v255, v255 clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4768 // GFX12: [0xff,0xfc,0x02,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4769
4770 v_xad_u32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4771 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4772
4773 v_xad_u32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4774 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4775
4776 v_xad_u32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4777 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4778
4779 v_xad_u32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4780 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4781
4782 v_xad_u32_e64_dpp v5, v1, v2, v3 row_mirror
4783 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4784
4785 v_xad_u32_e64_dpp v5, v1, v2, v255 row_half_mirror
4786 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4787
4788 v_xad_u32_e64_dpp v5, v1, v2, s105 row_shl:1
4789 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4790
4791 v_xad_u32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4792 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4793
4794 v_xad_u32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4795 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4796
4797 v_xad_u32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4798 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4799
4800 v_xad_u32_e64_dpp v5, v1, v2, exec_hi row_ror:1
4801 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4802
4803 v_xad_u32_e64_dpp v5, v1, v2, exec_lo row_ror:15
4804 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4805
4806 v_xad_u32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4807 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4808
4809 v_xad_u32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4810 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4811
4812 v_xad_u32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4813 // GFX12: [0x05,0x00,0x45,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4814
4815 v_xad_u32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4816 // GFX12: [0xff,0x00,0x45,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4817
4818 v_xor3_b32_e64_dpp v5, v1, v2, v3 quad_perm:[3,2,1,0]
4819 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
4820
4821 v_xor3_b32_e64_dpp v5, v1, s2, v3 quad_perm:[3,2,1,0]
4822 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
4823
4824 v_xor3_b32_e64_dpp v5, v1, 10, v3 quad_perm:[3,2,1,0]
4825 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x14,0x0d,0x04,0x01,0x1b,0x00,0xff]
4826
4827 v_xor3_b32_e64_dpp v5, v1, v2, v3 quad_perm:[0,1,2,3]
4828 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
4829
4830 v_xor3_b32_e64_dpp v5, v1, v2, v3 row_mirror
4831 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
4832
4833 v_xor3_b32_e64_dpp v5, v1, v2, v255 row_half_mirror
4834 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
4835
4836 v_xor3_b32_e64_dpp v5, v1, v2, s105 row_shl:1
4837 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
4838
4839 v_xor3_b32_e64_dpp v5, v1, v2, vcc_hi row_shl:15
4840 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
4841
4842 v_xor3_b32_e64_dpp v5, v1, v2, vcc_lo row_shr:1
4843 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
4844
4845 v_xor3_b32_e64_dpp v5, v1, v2, ttmp15 row_shr:15
4846 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xee,0x01,0x01,0x1f,0x01,0xff]
4847
4848 v_xor3_b32_e64_dpp v5, v1, v2, exec_hi row_ror:1
4849 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x21,0x01,0xff]
4850
4851 v_xor3_b32_e64_dpp v5, v1, v2, exec_lo row_ror:15
4852 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x2f,0x01,0xff]
4853
4854 v_xor3_b32_e64_dpp v5, v1, v2, null row_share:0 row_mask:0xf bank_mask:0xf
4855 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x50,0x01,0xff]
4856
4857 v_xor3_b32_e64_dpp v5, v1, v2, -1 row_share:15 row_mask:0x0 bank_mask:0x1
4858 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0x06,0x03,0x01,0x5f,0x01,0x01]
4859
4860 v_xor3_b32_e64_dpp v5, v1, v2, 0.5 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4861 // GFX12: [0x05,0x00,0x40,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x09,0x13]
4862
4863 v_xor3_b32_e64_dpp v255, v255, v255, src_scc row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4864 // GFX12: [0xff,0x00,0x40,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x05,0x30]
4865
4866 v_xor_b16_e64_dpp v5, v1, v2 quad_perm:[3,2,1,0]
4867 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
4868
4869 v_xor_b16_e64_dpp v5, v1, v2 quad_perm:[0,1,2,3]
4870 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
4871
4872 v_xor_b16_e64_dpp v5, v1, v2 row_mirror
4873 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
4874
4875 v_xor_b16_e64_dpp v5, v1, v2 row_half_mirror
4876 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
4877
4878 v_xor_b16_e64_dpp v5, v1, v2 row_shl:1
4879 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
4880
4881 v_xor_b16_e64_dpp v5, v1, v2 row_shl:15
4882 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
4883
4884 v_xor_b16_e64_dpp v5, v1, v2 row_shr:1
4885 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
4886
4887 v_xor_b16_e64_dpp v5, v1, v2 row_shr:15
4888 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
4889
4890 v_xor_b16_e64_dpp v5, v1, v2 row_ror:1
4891 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
4892
4893 v_xor_b16_e64_dpp v5, v1, v2 row_ror:15
4894 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
4895
4896 v_xor_b16_e64_dpp v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
4897 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4898
4899 v_xor_b16_e64_dpp v5, v1, v2 row_share:15 row_mask:0x0 bank_mask:0x1
4900 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4901
4902 v_xor_b16_e64_dpp v5, v1, v2 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
4903 // GFX12: [0x05,0x00,0x64,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x09,0x13]
4904
4905 v_xor_b16_e64_dpp v255, v255, v255 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
4906 // GFX12: [0xff,0x00,0x64,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x05,0x30]
4907
4908 v_add_nc_i16_e64_dpp v5, v1, v2 op_sel:[1,1,1] row_share:0 row_mask:0xf bank_mask:0xf
4909 // GFX12: [0x05,0x58,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4910
4911 v_add_nc_i16_e64_dpp v5, v1, v2 op_sel:[1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
4912 // GFX12: [0x05,0x08,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4913
4914 v_add_nc_i16_e64_dpp v5, v1, v2 op_sel:[0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4915 // GFX12: [0x05,0x10,0x0d,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x01,0x13]
4916
4917 v_add_nc_i16_e64_dpp v255, v255, v255 op_sel:[0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4918 // GFX12: [0xff,0xc0,0x0d,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x0d,0x30]
4919
4920 v_add_nc_u16_e64_dpp v5, v1, v2 op_sel:[1,1,1] row_share:0 row_mask:0xf bank_mask:0xf
4921 // GFX12: [0x05,0x58,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
4922
4923 v_add_nc_u16_e64_dpp v5, v1, v2 op_sel:[1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
4924 // GFX12: [0x05,0x08,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
4925
4926 v_add_nc_u16_e64_dpp v5, v1, v2 op_sel:[0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4927 // GFX12: [0x05,0x10,0x03,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x01,0x13]
4928
4929 v_add_nc_u16_e64_dpp v255, v255, v255 op_sel:[0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4930 // GFX12: [0xff,0xc0,0x03,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x0d,0x30]
4931
4932 v_cvt_pk_norm_i16_f16_e64_dpp v5, -v1, |v2| op_sel:[1,0,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4933 // GFX12: [0x05,0x0a,0x12,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x01,0x13]
4934
4935 v_cvt_pk_norm_i16_f16_e64_dpp v255, -|v255|, -|v255| op_sel:[0,1,0] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4936 // GFX12: [0xff,0x13,0x12,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x0d,0x30]
4937
4938 v_cvt_pk_norm_u16_f16_e64_dpp v5, -v1, |v2| op_sel:[1,0,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4939 // GFX12: [0x05,0x0a,0x13,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x01,0x13]
4940
4941 v_cvt_pk_norm_u16_f16_e64_dpp v255, -|v255|, -|v255| op_sel:[0,1,0] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4942 // GFX12: [0xff,0x13,0x13,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x0d,0x30]
4943
4944 v_div_fixup_f16_e64_dpp v5, -v1, v2, |exec_lo| op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
4945 // GFX12: [0x05,0x7c,0x54,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
4946
4947 v_div_fixup_f16_e64_dpp v5, -|v1|, -|v2|, null op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
4948 // GFX12: [0x05,0x0b,0x54,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
4949
4950 v_div_fixup_f16_e64_dpp v5, -|v1|, v2, -|-1| op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
4951 // GFX12: [0x05,0x15,0x54,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
4952
4953 v_div_fixup_f16_e64_dpp v5, v1, -|v2|, -|0.5| op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4954 // GFX12: [0x05,0x26,0x54,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x01,0x13]
4955
4956 v_div_fixup_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4957 // GFX12: [0xff,0xc7,0x54,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x0d,0x30]
4958
4959 v_fma_f16_e64_dpp v5, -v1, v2, |exec_lo| op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
4960 // GFX12: [0x05,0x7c,0x48,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
4961
4962 v_fma_f16_e64_dpp v5, -|v1|, -|v2|, null op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
4963 // GFX12: [0x05,0x0b,0x48,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
4964
4965 v_fma_f16_e64_dpp v5, -|v1|, v2, -|-1| op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
4966 // GFX12: [0x05,0x15,0x48,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
4967
4968 v_fma_f16_e64_dpp v5, v1, -|v2|, -|0.5| op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4969 // GFX12: [0x05,0x26,0x48,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x01,0x13]
4970
4971 v_fma_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4972 // GFX12: [0xff,0xc7,0x48,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x0d,0x30]
4973
4974 v_mad_i16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
4975 // GFX12: [0x05,0x78,0x53,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
4976
4977 v_mad_i16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
4978 // GFX12: [0x05,0x08,0x53,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
4979
4980 v_mad_i16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
4981 // GFX12: [0x05,0x10,0x53,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
4982
4983 v_mad_i16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4984 // GFX12: [0x05,0x20,0x53,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
4985
4986 v_mad_i16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4987 // GFX12: [0xff,0xc0,0x53,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
4988
4989 v_mad_i32_i16_e64_dpp v5, v1, v2, 0.5 op_sel:[1,0,0,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
4990 // GFX12: [0x05,0x08,0x5a,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x01,0x13]
4991
4992 v_mad_i32_i16_e64_dpp v255, v255, v255, src_scc op_sel:[0,1,0,0] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
4993 // GFX12: [0xff,0x90,0x5a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
4994
4995 v_mad_u16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
4996 // GFX12: [0x05,0x78,0x41,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
4997
4998 v_mad_u16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
4999 // GFX12: [0x05,0x08,0x41,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5000
5001 v_mad_u16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5002 // GFX12: [0x05,0x10,0x41,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5003
5004 v_mad_u16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5005 // GFX12: [0x05,0x20,0x41,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5006
5007 v_mad_u16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5008 // GFX12: [0xff,0xc0,0x41,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5009
5010 v_mad_u32_u16_e64_dpp v5, v1, v2, 0.5 op_sel:[1,0,0,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5011 // GFX12: [0x05,0x08,0x59,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x60,0x01,0x13]
5012
5013 v_mad_u32_u16_e64_dpp v255, v255, v255, src_scc op_sel:[0,1,0,0] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5014 // GFX12: [0xff,0x90,0x59,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5015
5016 v_max3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5017 // GFX12: [0x05,0x7c,0x2c,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5018
5019 v_max3_num_f16_e64_dpp v5, -|v1|, -|v2|, null op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5020 // GFX12: [0x05,0x0b,0x2c,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5021
5022 v_max3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5023 // GFX12: [0x05,0x15,0x2c,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5024
5025 v_max3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5026 // GFX12: [0x05,0x26,0x2c,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x01,0x13]
5027
5028 v_max3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5029 // GFX12: [0xff,0xc7,0x2c,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x0d,0x30]
5030
5031 v_max3_i16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5032 // GFX12: [0x05,0x78,0x4d,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5033
5034 v_max3_i16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5035 // GFX12: [0x05,0x08,0x4d,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5036
5037 v_max3_i16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5038 // GFX12: [0x05,0x10,0x4d,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5039
5040 v_max3_i16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5041 // GFX12: [0x05,0x20,0x4d,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5042
5043 v_max3_i16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5044 // GFX12: [0xff,0x40,0x4d,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5045
5046 v_max3_u16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5047 // GFX12: [0x05,0x78,0x4e,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5048
5049 v_max3_u16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5050 // GFX12: [0x05,0x08,0x4e,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5051
5052 v_max3_u16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5053 // GFX12: [0x05,0x10,0x4e,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5054
5055 v_max3_u16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5056 // GFX12: [0x05,0x20,0x4e,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5057
5058 v_max3_u16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5059 // GFX12: [0xff,0x40,0x4e,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5060
5061 v_med3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5062 // GFX12: [0x05,0x7c,0x32,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5063
5064 v_med3_num_f16_e64_dpp v5, -|v1|, -|v2|, null op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5065 // GFX12: [0x05,0x0b,0x32,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5066
5067 v_med3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5068 // GFX12: [0x05,0x15,0x32,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5069
5070 v_med3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5071 // GFX12: [0x05,0x26,0x32,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x01,0x13]
5072
5073 v_med3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5074 // GFX12: [0xff,0xc7,0x32,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x0d,0x30]
5075
5076 v_med3_i16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5077 // GFX12: [0x05,0x78,0x50,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5078
5079 v_med3_i16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5080 // GFX12: [0x05,0x08,0x50,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5081
5082 v_med3_i16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5083 // GFX12: [0x05,0x10,0x50,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5084
5085 v_med3_i16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5086 // GFX12: [0x05,0x20,0x50,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5087
5088 v_med3_i16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5089 // GFX12: [0xff,0x40,0x50,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5090
5091 v_med3_u16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5092 // GFX12: [0x05,0x78,0x51,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5093
5094 v_med3_u16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5095 // GFX12: [0x05,0x08,0x51,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5096
5097 v_med3_u16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5098 // GFX12: [0x05,0x10,0x51,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5099
5100 v_med3_u16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5101 // GFX12: [0x05,0x20,0x51,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5102
5103 v_med3_u16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5104 // GFX12: [0xff,0x40,0x51,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5105
5106 v_min3_num_f16_e64_dpp v5, -v1, v2, |exec_lo| op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5107 // GFX12: [0x05,0x7c,0x2b,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5108
5109 v_min3_num_f16_e64_dpp v5, -|v1|, -|v2|, null op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5110 // GFX12: [0x05,0x0b,0x2b,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5111
5112 v_min3_num_f16_e64_dpp v5, -|v1|, v2, -|-1| op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5113 // GFX12: [0x05,0x15,0x2b,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5114
5115 v_min3_num_f16_e64_dpp v5, v1, -|v2|, -|0.5| op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5116 // GFX12: [0x05,0x26,0x2b,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x01,0x13]
5117
5118 v_min3_num_f16_e64_dpp v255, -|v255|, -|v255|, -|src_scc| op_sel:[0,0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5119 // GFX12: [0xff,0xc7,0x2b,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x0d,0x30]
5120
5121 v_min3_i16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5122 // GFX12: [0x05,0x78,0x4a,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5123
5124 v_min3_i16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5125 // GFX12: [0x05,0x08,0x4a,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5126
5127 v_min3_i16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5128 // GFX12: [0x05,0x10,0x4a,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5129
5130 v_min3_i16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5131 // GFX12: [0x05,0x20,0x4a,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5132
5133 v_min3_i16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5134 // GFX12: [0xff,0x40,0x4a,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5135
5136 v_min3_u16_e64_dpp v5, v1, v2, exec_hi op_sel:[1,1,1,1] row_ror:15 row_mask:0xf bank_mask:0xf
5137 // GFX12: [0x05,0x78,0x4b,0xd6,0xfa,0x04,0xfe,0x01,0x01,0x2f,0x01,0xff]
5138
5139 v_min3_u16_e64_dpp v5, v1, v2, exec_lo op_sel:[1,0,0,0] row_share:0 row_mask:0xf bank_mask:0xf
5140 // GFX12: [0x05,0x08,0x4b,0xd6,0xfa,0x04,0xfa,0x01,0x01,0x50,0x01,0xff]
5141
5142 v_min3_u16_e64_dpp v5, v1, v2, null op_sel:[0,1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5143 // GFX12: [0x05,0x10,0x4b,0xd6,0xfa,0x04,0xf2,0x01,0x01,0x5f,0x01,0x01]
5144
5145 v_min3_u16_e64_dpp v5, v1, v2, -1 op_sel:[0,0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5146 // GFX12: [0x05,0x20,0x4b,0xd6,0xfa,0x04,0x06,0x03,0x01,0x60,0x01,0x13]
5147
5148 v_min3_u16_e64_dpp v255, v255, v255, src_scc op_sel:[0,0,0,1] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5149 // GFX12: [0xff,0x40,0x4b,0xd6,0xfa,0xfe,0xf7,0x03,0xff,0x6f,0x0d,0x30]
5150
5151 v_pack_b32_f16_e64_dpp v5, -v1, |v2| op_sel:[1,0,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5152 // GFX12: [0x05,0x0a,0x11,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x01,0x13]
5153
5154 v_pack_b32_f16_e64_dpp v255, -|v255|, -|v255| op_sel:[0,1,0] row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5155 // GFX12: [0xff,0x13,0x11,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x0d,0x30]
5156
5157 v_sub_nc_i16_e64_dpp v5, v1, v2 op_sel:[1,1,1] row_share:0 row_mask:0xf bank_mask:0xf
5158 // GFX12: [0x05,0x58,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5159
5160 v_sub_nc_i16_e64_dpp v5, v1, v2 op_sel:[1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5161 // GFX12: [0x05,0x08,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
5162
5163 v_sub_nc_i16_e64_dpp v5, v1, v2 op_sel:[0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5164 // GFX12: [0x05,0x10,0x0e,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x01,0x13]
5165
5166 v_sub_nc_i16_e64_dpp v255, v255, v255 op_sel:[0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5167 // GFX12: [0xff,0xc0,0x0e,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x0d,0x30]
5168
5169 v_sub_nc_u16_e64_dpp v5, v1, v2 op_sel:[1,1,1] row_share:0 row_mask:0xf bank_mask:0xf
5170 // GFX12: [0x05,0x58,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5171
5172 v_sub_nc_u16_e64_dpp v5, v1, v2 op_sel:[1,0,0] row_share:15 row_mask:0x0 bank_mask:0x1
5173 // GFX12: [0x05,0x08,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x5f,0x01,0x01]
5174
5175 v_sub_nc_u16_e64_dpp v5, v1, v2 op_sel:[0,1,0] row_xmask:0 row_mask:0x1 bank_mask:0x3
5176 // GFX12: [0x05,0x10,0x04,0xd7,0xfa,0x04,0x02,0x00,0x01,0x60,0x01,0x13]
5177
5178 v_sub_nc_u16_e64_dpp v255, v255, v255 op_sel:[0,0,1] clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:1 fi:1
5179 // GFX12: [0xff,0xc0,0x04,0xd7,0xfa,0xfe,0x03,0x00,0xff,0x6f,0x0d,0x30]
5180
5181 v_dot2_f16_f16_e64_dpp v0, v1, v2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5182 // GFX12: [0x00,0x00,0x66,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x04,0x00]
5183
5184 v_dot2_f16_f16_e64_dpp v0, v1, v2, v3 op_sel:[1,1,0,0] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5185 // GFX12-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid op_sel operand
5186
5187 v_dot2_f16_f16_e64_dpp v0, s1, v2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5188 // GFX12-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
5189
5190 v_dot2_f16_f16_e64_dpp v0, v1, s2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5191 // GFX12: [0x00,0x00,0x66,0xd6,0xfa,0x04,0x0c,0x04,0x01,0xe4,0x04,0x00]
5192
5193 v_dot2_f16_f16_e64_dpp v0, v1, v2, v3 op_sel:[0,0,1,1] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5194 // GFX12: [0x00,0x60,0x66,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x04,0x00]
5195
5196 v_dot2_f16_f16_e64_dpp v0, |v1|, -v2, -|s3| op_sel:[0,0,1,1] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5197 // GFX12: [0x00,0x65,0x66,0xd6,0xfa,0x04,0x0e,0xc0,0x01,0xe4,0x04,0x00]
5198
5199 v_dot2_f16_f16_e64_dpp v5, v1, v2, 0.5 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf
5200 // GFX12: [0x05,0x00,0x66,0xd6,0xfa,0x04,0xc2,0x03,0x01,0x1b,0x00,0xff]
5201
5202 v_dot2_bf16_bf16_e64_dpp v0, v1, v2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5203 // GFX12: [0x00,0x00,0x67,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x04,0x00]
5204
5205 v_dot2_bf16_bf16_e64_dpp v0, v1, v2, v3 op_sel:[1,1,0,0] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5206 // GFX12-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid op_sel operand
5207
5208 v_dot2_bf16_bf16_e64_dpp v0, s1, v2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
5209 // GFX12-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
5210
5211 v_dot2_bf16_bf16_e64_dpp v0, v1, s2, v3 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
5212 // GFX12: [0x00,0x00,0x67,0xd6,0xfa,0x04,0x0c,0x04,0x01,0xe4,0x00,0x00]
5213
5214 v_dot2_bf16_bf16_e64_dpp v0, v1, v2, v3 op_sel:[0,0,1,1] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5215 // GFX12: [0x00,0x60,0x67,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x04,0x00]
5216
5217 v_dot2_bf16_bf16_e64_dpp v0, |v1|, -v2, -|s3| op_sel:[0,0,1,1] quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 fi:1
5218 // GFX12: [0x00,0x65,0x67,0xd6,0xfa,0x04,0x0e,0xc0,0x01,0xe4,0x04,0x00]
5219
5220 v_dot2_bf16_bf16_e64_dpp v5, v1, v2, 0 quad_perm:[3,2,1,0] row_mask:0xf bank_mask:0xf
5221 // GFX12: [0x05,0x00,0x67,0xd6,0xfa,0x04,0x02,0x02,0x01,0x1b,0x00,0xff]
5222
5223 v_minimum_f32 v5, v1, v2 quad_perm:[3,2,1,0]
5224 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
5225
5226 v_minimum_f32 v5, v1, v2 quad_perm:[0,1,2,3]
5227 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
5228
5229 v_minimum_f32 v5, v1, v2 row_mirror
5230 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
5231
5232 v_minimum_f32 v5, v1, v2 row_half_mirror
5233 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
5234
5235 v_minimum_f32 v5, v1, v2 row_shl:1
5236 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
5237
5238 v_minimum_f32 v5, v1, v2 row_shl:15
5239 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
5240
5241 v_minimum_f32 v5, v1, v2 row_shr:1
5242 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
5243
5244 v_minimum_f32 v5, v1, v2 row_shr:15
5245 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
5246
5247 v_minimum_f32 v5, v1, v2 row_ror:1
5248 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
5249
5250 v_minimum_f32 v5, v1, v2 row_ror:15
5251 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
5252
5253 v_minimum_f32 v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
5254 // GFX12: [0x05,0x00,0x65,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5255
5256 v_minimum_f32 v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
5257 // GFX12: [0x05,0x01,0x65,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
5258
5259 v_minimum_f32 v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5260 // GFX12: [0x05,0x02,0x65,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
5261
5262 v_minimum_f32 v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5263 // GFX12: [0xff,0x03,0x65,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
5264
5265 v_maximum_f32 v5, v1, v2 quad_perm:[3,2,1,0]
5266 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
5267
5268 v_maximum_f32 v5, v1, v2 quad_perm:[0,1,2,3]
5269 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
5270
5271 v_maximum_f32 v5, v1, v2 row_mirror
5272 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
5273
5274 v_maximum_f32 v5, v1, v2 row_half_mirror
5275 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
5276
5277 v_maximum_f32 v5, v1, v2 row_shl:1
5278 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
5279
5280 v_maximum_f32 v5, v1, v2 row_shl:15
5281 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
5282
5283 v_maximum_f32 v5, v1, v2 row_shr:1
5284 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
5285
5286 v_maximum_f32 v5, v1, v2 row_shr:15
5287 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
5288
5289 v_maximum_f32 v5, v1, v2 row_ror:1
5290 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
5291
5292 v_maximum_f32 v5, v1, v2 row_ror:15
5293 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
5294
5295 v_maximum_f32 v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
5296 // GFX12: [0x05,0x00,0x66,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5297
5298 v_maximum_f32 v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
5299 // GFX12: [0x05,0x01,0x66,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
5300
5301 v_maximum_f32 v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5302 // GFX12: [0x05,0x02,0x66,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
5303
5304 v_maximum_f32 v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5305 // GFX12: [0xff,0x03,0x66,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
5306
5307 v_minimum_f16 v5, v1, v2 quad_perm:[3,2,1,0]
5308 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
5309
5310 v_minimum_f16 v5, v1, v2 quad_perm:[0,1,2,3]
5311 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
5312
5313 v_minimum_f16 v5, v1, v2 row_mirror
5314 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
5315
5316 v_minimum_f16 v5, v1, v2 row_half_mirror
5317 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
5318
5319 v_minimum_f16 v5, v1, v2 row_shl:1
5320 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
5321
5322 v_minimum_f16 v5, v1, v2 row_shl:15
5323 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
5324
5325 v_minimum_f16 v5, v1, v2 row_shr:1
5326 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
5327
5328 v_minimum_f16 v5, v1, v2 row_shr:15
5329 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
5330
5331 v_minimum_f16 v5, v1, v2 row_ror:1
5332 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
5333
5334 v_minimum_f16 v5, v1, v2 row_ror:15
5335 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
5336
5337 v_minimum_f16 v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
5338 // GFX12: [0x05,0x00,0x67,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5339
5340 v_minimum_f16 v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
5341 // GFX12: [0x05,0x01,0x67,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
5342
5343 v_minimum_f16 v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5344 // GFX12: [0x05,0x02,0x67,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
5345
5346 v_minimum_f16 v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5347 // GFX12: [0xff,0x03,0x67,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
5348
5349 v_maximum_f16 v5, v1, v2 quad_perm:[3,2,1,0]
5350 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1b,0x00,0xff]
5351
5352 v_maximum_f16 v5, v1, v2 quad_perm:[0,1,2,3]
5353 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0xe4,0x00,0xff]
5354
5355 v_maximum_f16 v5, v1, v2 row_mirror
5356 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x40,0x01,0xff]
5357
5358 v_maximum_f16 v5, v1, v2 row_half_mirror
5359 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x41,0x01,0xff]
5360
5361 v_maximum_f16 v5, v1, v2 row_shl:1
5362 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x01,0x01,0xff]
5363
5364 v_maximum_f16 v5, v1, v2 row_shl:15
5365 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x0f,0x01,0xff]
5366
5367 v_maximum_f16 v5, v1, v2 row_shr:1
5368 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x11,0x01,0xff]
5369
5370 v_maximum_f16 v5, v1, v2 row_shr:15
5371 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x1f,0x01,0xff]
5372
5373 v_maximum_f16 v5, v1, v2 row_ror:1
5374 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x21,0x01,0xff]
5375
5376 v_maximum_f16 v5, v1, v2 row_ror:15
5377 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x2f,0x01,0xff]
5378
5379 v_maximum_f16 v5, v1, v2 row_share:0 row_mask:0xf bank_mask:0xf
5380 // GFX12: [0x05,0x00,0x68,0xd7,0xfa,0x04,0x02,0x00,0x01,0x50,0x01,0xff]
5381
5382 v_maximum_f16 v5, |v1|, -v2 row_share:15 row_mask:0x0 bank_mask:0x1
5383 // GFX12: [0x05,0x01,0x68,0xd7,0xfa,0x04,0x02,0x40,0x01,0x5f,0x01,0x01]
5384
5385 v_maximum_f16 v5, -v1, |v2| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5386 // GFX12: [0x05,0x02,0x68,0xd7,0xfa,0x04,0x02,0x20,0x01,0x60,0x09,0x13]
5387
5388 v_maximum_f16 v255, -|v255|, -|v255| row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5389 // GFX12: [0xff,0x03,0x68,0xd7,0xfa,0xfe,0x03,0x60,0xff,0x6f,0x05,0x30]
5390
5391 v_minimum3_f32 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5392 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5393
5394 v_minimum3_f32 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5395 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5396
5397 v_minimum3_f32 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5398 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5399
5400 v_minimum3_f32 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5401 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5402
5403 v_minimum3_f32 v5, v1, v2, v3 row_mirror
5404 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5405
5406 v_minimum3_f32 v5, v1, v2, v255 row_half_mirror
5407 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5408
5409 v_minimum3_f32 v5, v1, v2, s105 row_shl:1
5410 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5411
5412 v_minimum3_f32 v5, v1, v2, vcc_hi row_shl:15
5413 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5414
5415 v_minimum3_f32 v5, v1, v2, vcc_lo row_shr:1
5416 // GFX12: [0x05,0x00,0x2d,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5417
5418 v_minimum3_f32 v5, |v1|, v2, -ttmp15 row_shr:15
5419 // GFX12: [0x05,0x01,0x2d,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5420
5421 v_minimum3_f32 v5, v1, -|v2|, exec_hi row_ror:1
5422 // GFX12: [0x05,0x02,0x2d,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5423
5424 v_minimum3_f32 v5, -v1, v2, |exec_lo| row_ror:15
5425 // GFX12: [0x05,0x04,0x2d,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5426
5427 v_minimum3_f32 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5428 // GFX12: [0x05,0x03,0x2d,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5429
5430 v_minimum3_f32 v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
5431 // GFX12: [0x05,0x05,0x2d,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
5432
5433 v_minimum3_f32 v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5434 // GFX12: [0x05,0x06,0x2d,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
5435
5436 v_minimum3_f32 v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5437 // GFX12: [0xff,0x87,0x2d,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
5438
5439 v_maximum3_f32 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5440 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5441
5442 v_maximum3_f32 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5443 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5444
5445 v_maximum3_f32 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5446 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5447
5448 v_maximum3_f32 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5449 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5450
5451 v_maximum3_f32 v5, v1, v2, v3 row_mirror
5452 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5453
5454 v_maximum3_f32 v5, v1, v2, v255 row_half_mirror
5455 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5456
5457 v_maximum3_f32 v5, v1, v2, s105 row_shl:1
5458 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5459
5460 v_maximum3_f32 v5, v1, v2, vcc_hi row_shl:15
5461 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5462
5463 v_maximum3_f32 v5, v1, v2, vcc_lo row_shr:1
5464 // GFX12: [0x05,0x00,0x2e,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5465
5466 v_maximum3_f32 v5, |v1|, v2, -ttmp15 row_shr:15
5467 // GFX12: [0x05,0x01,0x2e,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5468
5469 v_maximum3_f32 v5, v1, -|v2|, exec_hi row_ror:1
5470 // GFX12: [0x05,0x02,0x2e,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5471
5472 v_maximum3_f32 v5, -v1, v2, |exec_lo| row_ror:15
5473 // GFX12: [0x05,0x04,0x2e,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5474
5475 v_maximum3_f32 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5476 // GFX12: [0x05,0x03,0x2e,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5477
5478 v_maximum3_f32 v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
5479 // GFX12: [0x05,0x05,0x2e,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
5480
5481 v_maximum3_f32 v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5482 // GFX12: [0x05,0x06,0x2e,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
5483
5484 v_maximum3_f32 v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5485 // GFX12: [0xff,0x87,0x2e,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
5486
5487 v_minimum3_f16 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5488 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5489
5490 v_minimum3_f16 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5491 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5492
5493 v_minimum3_f16 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5494 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5495
5496 v_minimum3_f16 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5497 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5498
5499 v_minimum3_f16 v5, v1, v2, v3 row_mirror
5500 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5501
5502 v_minimum3_f16 v5, v1, v2, v255 row_half_mirror
5503 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5504
5505 v_minimum3_f16 v5, v1, v2, s105 row_shl:1
5506 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5507
5508 v_minimum3_f16 v5, v1, v2, vcc_hi row_shl:15
5509 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5510
5511 v_minimum3_f16 v5, v1, v2, vcc_lo row_shr:1
5512 // GFX12: [0x05,0x00,0x2f,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5513
5514 v_minimum3_f16 v5, |v1|, v2, -ttmp15 row_shr:15
5515 // GFX12: [0x05,0x01,0x2f,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5516
5517 v_minimum3_f16 v5, v1, -|v2|, exec_hi row_ror:1
5518 // GFX12: [0x05,0x02,0x2f,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5519
5520 v_minimum3_f16 v5, -v1, v2, |exec_lo| row_ror:15
5521 // GFX12: [0x05,0x04,0x2f,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5522
5523 v_minimum3_f16 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5524 // GFX12: [0x05,0x03,0x2f,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5525
5526 v_minimum3_f16 v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
5527 // GFX12: [0x05,0x05,0x2f,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5528
5529 v_minimum3_f16 v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5530 // GFX12: [0x05,0x06,0x2f,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
5531
5532 v_minimum3_f16 v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5533 // GFX12: [0xff,0x87,0x2f,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
5534
5535 v_maximum3_f16 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5536 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5537
5538 v_maximum3_f16 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5539 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5540
5541 v_maximum3_f16 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5542 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5543
5544 v_maximum3_f16 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5545 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5546
5547 v_maximum3_f16 v5, v1, v2, v3 row_mirror
5548 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5549
5550 v_maximum3_f16 v5, v1, v2, v255 row_half_mirror
5551 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5552
5553 v_maximum3_f16 v5, v1, v2, s105 row_shl:1
5554 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5555
5556 v_maximum3_f16 v5, v1, v2, vcc_hi row_shl:15
5557 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5558
5559 v_maximum3_f16 v5, v1, v2, vcc_lo row_shr:1
5560 // GFX12: [0x05,0x00,0x30,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5561
5562 v_maximum3_f16 v5, |v1|, v2, -ttmp15 row_shr:15
5563 // GFX12: [0x05,0x01,0x30,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5564
5565 v_maximum3_f16 v5, v1, -|v2|, exec_hi row_ror:1
5566 // GFX12: [0x05,0x02,0x30,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5567
5568 v_maximum3_f16 v5, -v1, v2, |exec_lo| row_ror:15
5569 // GFX12: [0x05,0x04,0x30,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5570
5571 v_maximum3_f16 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5572 // GFX12: [0x05,0x03,0x30,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5573
5574 v_maximum3_f16 v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
5575 // GFX12: [0x05,0x05,0x30,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5576
5577 v_maximum3_f16 v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5578 // GFX12: [0x05,0x06,0x30,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
5579
5580 v_maximum3_f16 v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5581 // GFX12: [0xff,0x87,0x30,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
5582
5583 v_maximumminimum_f32 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5584 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5585
5586 v_maximumminimum_f32 v5, v1, v2, v3 row_mirror
5587 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5588
5589 v_maximumminimum_f32 v5, v1, v2, v255 row_half_mirror
5590 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5591
5592 v_maximumminimum_f32 v5, v1, v2, s105 row_shl:1
5593 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5594
5595 v_maximumminimum_f32 v5, v1, v2, vcc_hi row_shl:15
5596 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5597
5598 v_maximumminimum_f32 v5, v1, v2, vcc_lo row_shr:1
5599 // GFX12: [0x05,0x00,0x6d,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5600
5601 v_maximumminimum_f32 v5, |v1|, v2, -ttmp15 row_shr:15
5602 // GFX12: [0x05,0x01,0x6d,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5603
5604 v_maximumminimum_f32 v5, v1, -|v2|, exec_hi row_ror:1
5605 // GFX12: [0x05,0x02,0x6d,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5606
5607 v_maximumminimum_f32 v5, -v1, v2, |exec_lo| row_ror:15
5608 // GFX12: [0x05,0x04,0x6d,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5609
5610 v_maximumminimum_f32 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5611 // GFX12: [0x05,0x03,0x6d,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5612
5613 v_maximumminimum_f32 v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
5614 // GFX12: [0x05,0x05,0x6d,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
5615
5616 v_maximumminimum_f32 v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5617 // GFX12: [0x05,0x06,0x6d,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
5618
5619 v_maximumminimum_f32 v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5620 // GFX12: [0xff,0x87,0x6d,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
5621
5622 v_minimummaximum_f32 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5623 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5624
5625 v_minimummaximum_f32 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5626 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5627
5628 v_minimummaximum_f32 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5629 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5630
5631 v_minimummaximum_f32 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5632 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5633
5634 v_minimummaximum_f32 v5, v1, v2, v3 row_mirror
5635 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5636
5637 v_minimummaximum_f32 v5, v1, v2, v255 row_half_mirror
5638 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5639
5640 v_minimummaximum_f32 v5, v1, v2, s105 row_shl:1
5641 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5642
5643 v_minimummaximum_f32 v5, v1, v2, vcc_hi row_shl:15
5644 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5645
5646 v_minimummaximum_f32 v5, v1, v2, vcc_lo row_shr:1
5647 // GFX12: [0x05,0x00,0x6c,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5648
5649 v_minimummaximum_f32 v5, |v1|, v2, -ttmp15 row_shr:15
5650 // GFX12: [0x05,0x01,0x6c,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5651
5652 v_minimummaximum_f32 v5, v1, -|v2|, exec_hi row_ror:1
5653 // GFX12: [0x05,0x02,0x6c,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5654
5655 v_minimummaximum_f32 v5, -v1, v2, |exec_lo| row_ror:15
5656 // GFX12: [0x05,0x04,0x6c,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5657
5658 v_minimummaximum_f32 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5659 // GFX12: [0x05,0x03,0x6c,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5660
5661 v_minimummaximum_f32 v5, -|v1|, v2, -|-1| mul:2 row_share:15 row_mask:0x0 bank_mask:0x1
5662 // GFX12: [0x05,0x05,0x6c,0xd6,0xfa,0x04,0x06,0xab,0x01,0x5f,0x01,0x01]
5663
5664 v_minimummaximum_f32 v5, v1, -|v2|, -|0.5| mul:4 row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5665 // GFX12: [0x05,0x06,0x6c,0xd6,0xfa,0x04,0xc2,0xd3,0x01,0x60,0x09,0x13]
5666
5667 v_minimummaximum_f32 v255, -|v255|, -|v255|, -|src_scc| clamp div:2 row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5668 // GFX12: [0xff,0x87,0x6c,0xd6,0xfa,0xfe,0xf7,0xfb,0xff,0x6f,0x05,0x30]
5669
5670 v_maximumminimum_f16 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5671 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5672
5673 v_maximumminimum_f16 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5674 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5675
5676 v_maximumminimum_f16 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5677 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5678
5679 v_maximumminimum_f16 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5680 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5681
5682 v_maximumminimum_f16 v5, v1, v2, v3 row_mirror
5683 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5684
5685 v_maximumminimum_f16 v5, v1, v2, v255 row_half_mirror
5686 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5687
5688 v_maximumminimum_f16 v5, v1, v2, s105 row_shl:1
5689 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5690
5691 v_maximumminimum_f16 v5, v1, v2, vcc_hi row_shl:15
5692 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5693
5694 v_maximumminimum_f16 v5, v1, v2, vcc_lo row_shr:1
5695 // GFX12: [0x05,0x00,0x6f,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5696
5697 v_maximumminimum_f16 v5, |v1|, v2, -ttmp15 row_shr:15
5698 // GFX12: [0x05,0x01,0x6f,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5699
5700 v_maximumminimum_f16 v5, v1, -|v2|, exec_hi row_ror:1
5701 // GFX12: [0x05,0x02,0x6f,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5702
5703 v_maximumminimum_f16 v5, -v1, v2, |exec_lo| row_ror:15
5704 // GFX12: [0x05,0x04,0x6f,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5705
5706 v_maximumminimum_f16 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5707 // GFX12: [0x05,0x03,0x6f,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5708
5709 v_maximumminimum_f16 v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
5710 // GFX12: [0x05,0x05,0x6f,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5711
5712 v_maximumminimum_f16 v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5713 // GFX12: [0x05,0x06,0x6f,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
5714
5715 v_maximumminimum_f16 v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5716 // GFX12: [0xff,0x87,0x6f,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]
5717
5718 v_minimummaximum_f16 v5, v1, v2, v3 quad_perm:[3,2,1,0]
5719 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x1b,0x00,0xff]
5720
5721 v_minimummaximum_f16 v5, v1, s2, v3 quad_perm:[3,2,1,0]
5722 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0x0c,0x04,0x01,0x1b,0x00,0xff]
5723
5724 v_minimummaximum_f16 v5, v1, 2.0, v3 quad_perm:[3,2,1,0]
5725 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0xe8,0x0d,0x04,0x01,0x1b,0x00,0xff]
5726
5727 v_minimummaximum_f16 v5, v1, v2, v3 quad_perm:[0,1,2,3]
5728 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0xe4,0x00,0xff]
5729
5730 v_minimummaximum_f16 v5, v1, v2, v3 row_mirror
5731 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0x0e,0x04,0x01,0x40,0x01,0xff]
5732
5733 v_minimummaximum_f16 v5, v1, v2, v255 row_half_mirror
5734 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0xfe,0x07,0x01,0x41,0x01,0xff]
5735
5736 v_minimummaximum_f16 v5, v1, v2, s105 row_shl:1
5737 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0xa6,0x01,0x01,0x01,0x01,0xff]
5738
5739 v_minimummaximum_f16 v5, v1, v2, vcc_hi row_shl:15
5740 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0xae,0x01,0x01,0x0f,0x01,0xff]
5741
5742 v_minimummaximum_f16 v5, v1, v2, vcc_lo row_shr:1
5743 // GFX12: [0x05,0x00,0x6e,0xd6,0xfa,0x04,0xaa,0x01,0x01,0x11,0x01,0xff]
5744
5745 v_minimummaximum_f16 v5, |v1|, v2, -ttmp15 row_shr:15
5746 // GFX12: [0x05,0x01,0x6e,0xd6,0xfa,0x04,0xee,0x81,0x01,0x1f,0x01,0xff]
5747
5748 v_minimummaximum_f16 v5, v1, -|v2|, exec_hi row_ror:1
5749 // GFX12: [0x05,0x02,0x6e,0xd6,0xfa,0x04,0xfe,0x41,0x01,0x21,0x01,0xff]
5750
5751 v_minimummaximum_f16 v5, -v1, v2, |exec_lo| row_ror:15
5752 // GFX12: [0x05,0x04,0x6e,0xd6,0xfa,0x04,0xfa,0x21,0x01,0x2f,0x01,0xff]
5753
5754 v_minimummaximum_f16 v5, -|v1|, -|v2|, null row_share:0 row_mask:0xf bank_mask:0xf
5755 // GFX12: [0x05,0x03,0x6e,0xd6,0xfa,0x04,0xf2,0x61,0x01,0x50,0x01,0xff]
5756
5757 v_minimummaximum_f16 v5, -|v1|, v2, -|-1| row_share:15 row_mask:0x0 bank_mask:0x1
5758 // GFX12: [0x05,0x05,0x6e,0xd6,0xfa,0x04,0x06,0xa3,0x01,0x5f,0x01,0x01]
5759
5760 v_minimummaximum_f16 v5, v1, -|v2|, -|0.5| row_xmask:0 row_mask:0x1 bank_mask:0x3 bound_ctrl:1 fi:0
5761 // GFX12: [0x05,0x06,0x6e,0xd6,0xfa,0x04,0xc2,0xc3,0x01,0x60,0x09,0x13]
5762
5763 v_minimummaximum_f16 v255, -|v255|, -|v255|, -|src_scc| clamp row_xmask:15 row_mask:0x3 bank_mask:0x0 bound_ctrl:0 fi:1
5764 // GFX12: [0xff,0x87,0x6e,0xd6,0xfa,0xfe,0xf7,0xe3,0xff,0x6f,0x05,0x30]