llvm/test/MC/AMDGPU/vop3p.s

   1 // RUN: llvm-mc -triple=amdgcn -mcpu=gfx900 -show-encoding %s | FileCheck -check-prefix=GFX9 %s
   2 //
   3 // Test op_sel/op_sel_hi
   4 //
   5
   6 v_pk_add_u16 v1, v2, v3
   7 // GFX9: v_pk_add_u16 v1, v2, v3 ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x18]
   8
   9 v_pk_add_u16 v1, v2, v3 op_sel:[0,0]
  10 // GFX9: v_pk_add_u16 v1, v2, v3 ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x18]
  11
  12 v_pk_add_u16 v1, v2, v3 op_sel_hi:[1,1]
  13 // GFX9: v_pk_add_u16 v1, v2, v3 ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x18]
  14
  15 v_pk_add_u16 v1, v2, v3 op_sel:[0,0] op_sel_hi:[1,1]
  16 // GFX9: v_pk_add_u16 v1, v2, v3 ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x18]
  17
  18 v_pk_add_u16 v1, v2, v3 op_sel_hi:[0,0]
  19 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel_hi:[0,0] ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x00]
  20
  21 v_pk_add_u16 v1, v2, v3 op_sel:[0,0] op_sel_hi:[0,0]
  22 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel_hi:[0,0] ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x00]
  23
  24 v_pk_add_u16 v1, v2, v3 op_sel:[1,0]
  25 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[1,0] ; encoding: [0x01,0x48,0x8a,0xd3,0x02,0x07,0x02,0x18]
  26
  27 v_pk_add_u16 v1, v2, v3 op_sel:[0,1]
  28 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[0,1] ; encoding: [0x01,0x50,0x8a,0xd3,0x02,0x07,0x02,0x18]
  29
  30 v_pk_add_u16 v1, v2, v3 op_sel:[1,1]
  31 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[1,1] ; encoding: [0x01,0x58,0x8a,0xd3,0x02,0x07,0x02,0x18]
  32
  33 v_pk_add_u16 v1, v2, v3 op_sel_hi:[0,1]
  34 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel_hi:[0,1] ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x10]
  35
  36 v_pk_add_u16 v1, v2, v3 op_sel_hi:[1,0]
  37 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel_hi:[1,0] ; encoding: [0x01,0x40,0x8a,0xd3,0x02,0x07,0x02,0x08]
  38
  39 v_pk_add_u16 v1, v2, v3 op_sel:[1,1] op_sel_hi:[1,1]
  40 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[1,1] ; encoding: [0x01,0x58,0x8a,0xd3,0x02,0x07,0x02,0x18]
  41
  42 v_pk_add_u16 v1, v2, v3 op_sel:[1,0] op_sel_hi:[1,0]
  43 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[1,0] op_sel_hi:[1,0] ; encoding: [0x01,0x48,0x8a,0xd3,0x02,0x07,0x02,0x08]
  44
  45 v_pk_add_u16 v1, v2, v3 op_sel:[0,1] op_sel_hi:[0,1]
  46 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[0,1] op_sel_hi:[0,1] ; encoding: [0x01,0x50,0x8a,0xd3,0x02,0x07,0x02,0x10]
  47
  48 v_pk_add_u16 v1, v2, v3 op_sel:[1,0] op_sel_hi:[0,1]
  49 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[1,0] op_sel_hi:[0,1] ; encoding: [0x01,0x48,0x8a,0xd3,0x02,0x07,0x02,0x10]
  50
  51 v_pk_add_u16 v1, v2, v3 op_sel:[0,1] op_sel_hi:[1,0]
  52 // GFX9: v_pk_add_u16 v1, v2, v3 op_sel:[0,1] op_sel_hi:[1,0] ; encoding: [0x01,0x50,0x8a,0xd3,0x02,0x07,0x02,0x08]
  53
  54 //
  55 // Test src2 op_sel/op_sel_hi
  56 //
  57
  58 v_pk_fma_f16 v8, v0, s0, v1
  59 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x1c]
  60
  61 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[0,0,0] neg_hi:[0,0,0]
  62 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x1c]
  63
  64 v_pk_fma_f16 v8, v0, s0, v1 op_sel:[0,0,0] op_sel_hi:[1,1,1] neg_lo:[0,0,0] neg_hi:[0,0,0]
  65 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x1c]
  66
  67 v_pk_fma_f16 v8, v0, s0, v1 op_sel:[0,0,0] op_sel_hi:[1,1,1]
  68 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x1c]
  69
  70 v_pk_fma_f16 v8, v0, s0, v1 op_sel:[0,0,0] op_sel_hi:[0,0,0]
  71 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 op_sel_hi:[0,0,0] ; encoding: [0x08,0x00,0x8e,0xd3,0x00,0x01,0x04,0x04]
  72
  73 v_pk_fma_f16 v8, v0, s0, v1 op_sel:[0,0,1] op_sel_hi:[0,0,1]
  74 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 op_sel:[0,0,1] op_sel_hi:[0,0,1] ; encoding: [0x08,0x60,0x8e,0xd3,0x00,0x01,0x04,0x04]
  75
  76 //
  77 // Test neg_lo/neg_hi
  78 //
  79
  80 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,1,1]
  81 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,1,1] ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0xfc]
  82
  83 v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[1,1,1]
  84 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[1,1,1] ; encoding: [0x08,0x47,0x8e,0xd3,0x00,0x01,0x04,0x1c]
  85
  86 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,1,1] neg_hi:[1,1,1]
  87 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,1,1] neg_hi:[1,1,1] ; encoding: [0x08,0x47,0x8e,0xd3,0x00,0x01,0x04,0xfc]
  88
  89 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,0,0]
  90 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[1,0,0] ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x3c]
  91
  92 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[0,1,0]
  93 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[0,1,0] ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x5c]
  94
  95 v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[0,0,1]
  96 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_lo:[0,0,1] ; encoding: [0x08,0x40,0x8e,0xd3,0x00,0x01,0x04,0x9c]
  97
  98 v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[1,0,0]
  99 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[1,0,0] ; encoding: [0x08,0x41,0x8e,0xd3,0x00,0x01,0x04,0x1c]
 100
 101 v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[0,1,0]
 102 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[0,1,0] ; encoding: [0x08,0x42,0x8e,0xd3,0x00,0x01,0x04,0x1c]
 103
 104 v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[0,0,1]
 105 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 neg_hi:[0,0,1] ; encoding: [0x08,0x44,0x8e,0xd3,0x00,0x01,0x04,0x1c]
 106
 107
 108 // Test clamp
 109 v_pk_fma_f16 v8, v0, s0, v1 clamp
 110 // GFX9: v_pk_fma_f16 v8, v0, s0, v1 clamp ; encoding: [0x08,0xc0,0x8e,0xd3,0x00,0x01,0x04,0x1c]
 111
 112 v_pk_add_u16 v1, v2, v3 clamp
 113 // GFX9: v_pk_add_u16 v1, v2, v3 clamp ; encoding: [0x01,0xc0,0x8a,0xd3,0x02,0x07,0x02,0x18]
 114
 115 v_pk_min_i16 v0, v1, v2 clamp
 116 // GFX9: v_pk_min_i16 v0, v1, v2 clamp ; encoding: [0x00,0xc0,0x88,0xd3,0x01,0x05,0x02,0x18]
 117
 118 //
 119 // Instruction tests:
 120 //
 121
 122 v_pk_mul_lo_u16 v0, v1, v2
 123 // GFX9: v_pk_mul_lo_u16 v0, v1, v2 ; encoding: [0x00,0x40,0x81,0xd3,0x01,0x05,0x02,0x18]
 124
 125 v_pk_add_i16 v0, v1, v2
 126 // GFX9: v_pk_add_i16 v0, v1, v2 ; encoding: [0x00,0x40,0x82,0xd3,0x01,0x05,0x02,0x18]
 127
 128 v_pk_sub_i16 v0, v1, v2
 129 // GFX9: v_pk_sub_i16 v0, v1, v2 ; encoding: [0x00,0x40,0x83,0xd3,0x01,0x05,0x02,0x18]
 130
 131 v_pk_lshlrev_b16 v0, v1, v2
 132 // GFX9: v_pk_lshlrev_b16 v0, v1, v2 ; encoding: [0x00,0x40,0x84,0xd3,0x01,0x05,0x02,0x18]
 133
 134 v_pk_lshrrev_b16 v0, v1, v2
 135 // GFX9: v_pk_lshrrev_b16 v0, v1, v2 ; encoding: [0x00,0x40,0x85,0xd3,0x01,0x05,0x02,0x18]
 136
 137 v_pk_ashrrev_i16 v0, v1, v2
 138 // GFX9: v_pk_ashrrev_i16 v0, v1, v2 ; encoding: [0x00,0x40,0x86,0xd3,0x01,0x05,0x02,0x18]
 139
 140 v_pk_max_i16 v0, v1, v2
 141 // GFX9: v_pk_max_i16 v0, v1, v2 ; encoding: [0x00,0x40,0x87,0xd3,0x01,0x05,0x02,0x18]
 142
 143 v_pk_min_i16 v0, v1, v2
 144 // GFX9: v_pk_min_i16 v0, v1, v2 ; encoding: [0x00,0x40,0x88,0xd3,0x01,0x05,0x02,0x18]
 145
 146 v_pk_add_u16 v0, v1, v2
 147 // GFX9: v_pk_add_u16 v0, v1, v2 ; encoding: [0x00,0x40,0x8a,0xd3,0x01,0x05,0x02,0x18]
 148
 149 v_pk_max_u16 v0, v1, v2
 150 // GFX9: v_pk_max_u16 v0, v1, v2 ; encoding: [0x00,0x40,0x8c,0xd3,0x01,0x05,0x02,0x18]
 151
 152 v_pk_min_u16 v0, v1, v2
 153 // GFX9: v_pk_min_u16 v0, v1, v2 ; encoding: [0x00,0x40,0x8d,0xd3,0x01,0x05,0x02,0x18]
 154
 155 v_pk_fma_f16 v0, v1, v2, v3
 156 // GFX9: v_pk_fma_f16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0x8e,0xd3,0x01,0x05,0x0e,0x1c]
 157
 158 v_pk_add_f16 v0, v1, v2
 159 // GFX9: v_pk_add_f16 v0, v1, v2 ; encoding: [0x00,0x40,0x8f,0xd3,0x01,0x05,0x02,0x18]
 160
 161 v_pk_mul_f16 v0, v1, v2
 162 // GFX9: v_pk_mul_f16 v0, v1, v2 ; encoding: [0x00,0x40,0x90,0xd3,0x01,0x05,0x02,0x18]
 163
 164 v_pk_min_f16 v0, v1, v2
 165 // GFX9: v_pk_min_f16 v0, v1, v2 ; encoding: [0x00,0x40,0x91,0xd3,0x01,0x05,0x02,0x18]
 166
 167 v_pk_max_f16 v0, v1, v2
 168 // GFX9: v_pk_max_f16 v0, v1, v2 ; encoding: [0x00,0x40,0x92,0xd3,0x01,0x05,0x02,0x18]