include/kernel_vendor.h

   1 /**
   2  * Author......: Jens Steube <jens.steube@gmail.com>
   3  * License.....: MIT
   4  */
   5
   6 #pragma OPENCL EXTENSION cl_khr_int64_base_atomics     : enable
   7 #pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
   8
   9 /**
  10  * device type
  11  */
  12
  13 #define DEVICE_TYPE_CPU   2
  14 #define DEVICE_TYPE_GPU   4
  15 #define DEVICE_TYPE_ACCEL 8
  16
  17 #if   DEVICE_TYPE == DEVICE_TYPE_CPU
  18 #define IS_CPU
  19 #elif DEVICE_TYPE == DEVICE_TYPE_GPU
  20 #define IS_GPU
  21 #elif DEVICE_TYPE == DEVICE_TYPE_ACCEL
  22 #define IS_ACCEL
  23 #endif
  24
  25 /**
  26  * vendor specific
  27  */
  28
  29 #if   VENDOR_ID == (1 << 0)
  30 #define IS_AMD
  31 #elif VENDOR_ID == (1 << 6)
  32 #define IS_NV
  33 #else
  34 #define IS_GENERIC
  35 #endif
  36
  37 /**
  38  * AMD specific
  39  */
  40
  41 #ifdef IS_AMD
  42 #pragma OPENCL EXTENSION cl_amd_media_ops  : enable
  43 #pragma OPENCL EXTENSION cl_amd_media_ops2 : enable
  44 #endif
  45
  46 /**
  47  * Unrolling is generally enabled, for all device types and hash modes
  48  * There's a few exception when it's better not to unroll
  49  */
  50
  51 // Some algorithms run into too much register pressure due to loop unrolling
  52
  53 #ifdef IS_NV
  54 #ifdef IS_GPU
  55
  56 #if KERN_TYPE == 1500
  57 #undef _unroll
  58 #endif
  59 #if KERN_TYPE == 1800
  60 #undef _unroll
  61 #endif
  62 #if KERN_TYPE == 3000
  63 #undef _unroll
  64 #endif
  65 #if KERN_TYPE == 6221
  66 #undef _unroll
  67 #endif
  68 #if KERN_TYPE == 6222
  69 #undef _unroll
  70 #endif
  71 #if KERN_TYPE == 6223
  72 #undef _unroll
  73 #endif
  74 #if KERN_TYPE == 6500
  75 #undef _unroll
  76 #endif
  77 #if KERN_TYPE == 7100
  78 #undef _unroll
  79 #endif
  80 #if KERN_TYPE == 7400
  81 #undef _unroll
  82 #endif
  83 #if KERN_TYPE == 8200
  84 #undef _unroll
  85 #endif
  86 #if KERN_TYPE == 10400
  87 #undef _unroll
  88 #endif
  89 #if KERN_TYPE == 10500
  90 #undef _unroll
  91 #endif
  92 #if KERN_TYPE == 10700
  93 #undef _unroll
  94 #endif
  95 #if KERN_TYPE == 12300
  96 #undef _unroll
  97 #endif
  98 #if KERN_TYPE == 12400
  99 #undef _unroll
 100 #endif
 101
 102 #endif
 103 #endif
 104
 105 #ifdef IS_AMD
 106 #ifdef IS_GPU
 107
 108 #if KERN_TYPE == 3200
 109 #undef _unroll
 110 #endif
 111 #if KERN_TYPE == 5200
 112 #undef _unroll
 113 #endif
 114 #if KERN_TYPE == 6100
 115 #undef _unroll
 116 #endif
 117 #if KERN_TYPE == 6221
 118 #undef _unroll
 119 #endif
 120 #if KERN_TYPE == 6222
 121 #undef _unroll
 122 #endif
 123 #if KERN_TYPE == 6223
 124 #undef _unroll
 125 #endif
 126 #if KERN_TYPE == 6400
 127 #undef _unroll
 128 #endif
 129 #if KERN_TYPE == 6500
 130 #undef _unroll
 131 #endif
 132 #if KERN_TYPE == 6800
 133 #undef _unroll
 134 #endif
 135 #if KERN_TYPE == 7100
 136 #undef _unroll
 137 #endif
 138 #if KERN_TYPE == 7400
 139 #undef _unroll
 140 #endif
 141 #if KERN_TYPE == 8000
 142 #undef _unroll
 143 #endif
 144 #if KERN_TYPE == 8200
 145 #undef _unroll
 146 #endif
 147 #if KERN_TYPE == 10900
 148 #undef _unroll
 149 #endif
 150 #if KERN_TYPE == 11600
 151 #undef _unroll
 152 #endif
 153 #if KERN_TYPE == 12300
 154 #undef _unroll
 155 #endif
 156 #if KERN_TYPE == 12800
 157 #undef _unroll
 158 #endif
 159 #if KERN_TYPE == 12900
 160 #undef _unroll
 161 #endif
 162 #if KERN_TYPE == 13000
 163 #undef _unroll
 164 #endif
 165
 166 #endif
 167 #endif
 168
 169 // Some algorithms break due to loop unrolling, it's unknown why, probably compiler bugs
 170 // Can overlap with above cases
 171
 172 #ifdef IS_AMD
 173 #ifdef IS_GPU
 174
 175 #if KERN_TYPE == 1750
 176 #undef _unroll
 177 #endif
 178 #if KERN_TYPE == 1760
 179 #undef _unroll
 180 #endif
 181 #if KERN_TYPE == 6500
 182 #undef _unroll
 183 #endif
 184 #if KERN_TYPE == 7100
 185 #undef _unroll
 186 #endif
 187 #if KERN_TYPE == 9600
 188 #undef _unroll
 189 #endif
 190 #if KERN_TYPE == 12200
 191 #undef _unroll
 192 #endif
 193 #if KERN_TYPE == 12300
 194 #undef _unroll
 195 #endif
 196
 197 #endif
 198 #endif