include/kernel_vendor.h

   1 /**
   2  * Author......: Jens Steube <jens.steube@gmail.com>
   3  * License.....: MIT
   4  */
   5
   6 //fails on intel opencl sdk
   7 //#pragma OPENCL EXTENSION cl_khr_int64_base_atomics     : enable
   8 //#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
   9
  10 /**
  11  * device type
  12  */
  13
  14 #define DEVICE_TYPE_CPU   2
  15 #define DEVICE_TYPE_GPU   4
  16 #define DEVICE_TYPE_ACCEL 8
  17
  18 #if   DEVICE_TYPE == DEVICE_TYPE_CPU
  19 #define IS_CPU
  20 #elif DEVICE_TYPE == DEVICE_TYPE_GPU
  21 #define IS_GPU
  22 #elif DEVICE_TYPE == DEVICE_TYPE_ACCEL
  23 #define IS_ACCEL
  24 #endif
  25
  26 /**
  27  * vendor specific
  28  */
  29
  30 #if   VENDOR_ID == (1 << 0)
  31 #define IS_AMD
  32 #elif VENDOR_ID == (1 << 6)
  33 #define IS_NV
  34 #else
  35 #define IS_GENERIC
  36 #endif
  37
  38 /**
  39  * AMD specific
  40  */
  41
  42 #ifdef IS_AMD
  43 #pragma OPENCL EXTENSION cl_amd_media_ops  : enable
  44 #pragma OPENCL EXTENSION cl_amd_media_ops2 : enable
  45 #endif
  46
  47 /**
  48  * Unrolling is generally enabled, for all device types and hash modes
  49  * There's a few exception when it's better not to unroll
  50  */
  51
  52 // Some algorithms run into too much register pressure due to loop unrolling
  53
  54 #ifdef IS_NV
  55 #ifdef IS_GPU
  56
  57 #if KERN_TYPE == 1500
  58 #undef _unroll
  59 #endif
  60 #if KERN_TYPE == 1800
  61 #undef _unroll
  62 #endif
  63 #if KERN_TYPE == 3000
  64 #undef _unroll
  65 #endif
  66 #if KERN_TYPE == 6221
  67 #undef _unroll
  68 #endif
  69 #if KERN_TYPE == 6222
  70 #undef _unroll
  71 #endif
  72 #if KERN_TYPE == 6223
  73 #undef _unroll
  74 #endif
  75 #if KERN_TYPE == 6500
  76 #undef _unroll
  77 #endif
  78 #if KERN_TYPE == 7100
  79 #undef _unroll
  80 #endif
  81 #if KERN_TYPE == 7400
  82 #undef _unroll
  83 #endif
  84 #if KERN_TYPE == 7900
  85 #undef _unroll
  86 #endif
  87 #if KERN_TYPE == 8200
  88 #undef _unroll
  89 #endif
  90 #if KERN_TYPE == 10400
  91 #undef _unroll
  92 #endif
  93 #if KERN_TYPE == 10500
  94 #undef _unroll
  95 #endif
  96 #if KERN_TYPE == 10700
  97 #undef _unroll
  98 #endif
  99 #if KERN_TYPE == 12300
 100 #undef _unroll
 101 #endif
 102 #if KERN_TYPE == 12400
 103 #undef _unroll
 104 #endif
 105
 106 #endif
 107 #endif
 108
 109 #ifdef IS_AMD
 110 #ifdef IS_GPU
 111
 112 #if KERN_TYPE == 3200
 113 #undef _unroll
 114 #endif
 115 #if KERN_TYPE == 5200
 116 #undef _unroll
 117 #endif
 118 #if KERN_TYPE == 6100
 119 #undef _unroll
 120 #endif
 121 #if KERN_TYPE == 6221
 122 #undef _unroll
 123 #endif
 124 #if KERN_TYPE == 6222
 125 #undef _unroll
 126 #endif
 127 #if KERN_TYPE == 6223
 128 #undef _unroll
 129 #endif
 130 #if KERN_TYPE == 6400
 131 #undef _unroll
 132 #endif
 133 #if KERN_TYPE == 6500
 134 #undef _unroll
 135 #endif
 136 #if KERN_TYPE == 6800
 137 #undef _unroll
 138 #endif
 139 #if KERN_TYPE == 7100
 140 #undef _unroll
 141 #endif
 142 #if KERN_TYPE == 7400
 143 #undef _unroll
 144 #endif
 145 #if KERN_TYPE == 8000
 146 #undef _unroll
 147 #endif
 148 #if KERN_TYPE == 8200
 149 #undef _unroll
 150 #endif
 151 #if KERN_TYPE == 10900
 152 #undef _unroll
 153 #endif
 154 #if KERN_TYPE == 11600
 155 #undef _unroll
 156 #endif
 157 #if KERN_TYPE == 12300
 158 #undef _unroll
 159 #endif
 160 #if KERN_TYPE == 12800
 161 #undef _unroll
 162 #endif
 163 #if KERN_TYPE == 12900
 164 #undef _unroll
 165 #endif
 166 #if KERN_TYPE == 13000
 167 #undef _unroll
 168 #endif
 169
 170 #endif
 171 #endif
 172
 173 // Some algorithms break due to loop unrolling, it's unknown why, probably compiler bugs
 174 // Can overlap with above cases
 175
 176 #ifdef IS_AMD
 177 #ifdef IS_GPU
 178
 179 #if KERN_TYPE == 1750
 180 #undef _unroll
 181 #endif
 182 #if KERN_TYPE == 1760
 183 #undef _unroll
 184 #endif
 185 #if KERN_TYPE == 6500
 186 #undef _unroll
 187 #endif
 188 #if KERN_TYPE == 7100
 189 #undef _unroll
 190 #endif
 191 #if KERN_TYPE == 9600
 192 #undef _unroll
 193 #endif
 194 #if KERN_TYPE == 12200
 195 #undef _unroll
 196 #endif
 197 #if KERN_TYPE == 12300
 198 #undef _unroll
 199 #endif
 200
 201 #endif
 202 #endif