/** * Author......: See docs/credits.txt * License.....: MIT */ /** * device type */ #define DEVICE_TYPE_CPU 2 #define DEVICE_TYPE_GPU 4 #define DEVICE_TYPE_ACCEL 8 #if DEVICE_TYPE == DEVICE_TYPE_CPU #define IS_CPU #elif DEVICE_TYPE == DEVICE_TYPE_GPU #define IS_GPU #elif DEVICE_TYPE == DEVICE_TYPE_ACCEL #define IS_ACCEL #endif #if DEVICE_TYPE == DEVICE_TYPE_CPU #elif DEVICE_TYPE == DEVICE_TYPE_GPU #define REAL_SHM #elif DEVICE_TYPE == DEVICE_TYPE_ACCEL #endif #ifdef REAL_SHM #define SHM_TYPE __local #define SCR_TYPE __local #else #define SHM_TYPE __constant #define SCR_TYPE #endif /** * vendor specific */ #if VENDOR_ID == (1 << 0) #if AMD_ROCM == 0 #define IS_AMD #define AMD_GCN 0 #else #define IS_AMD #if defined __gfx600__ || defined __gfx601__ #define AMD_GCN 1 #elif defined __gfx700__ || defined __gfx701__ || defined __gfx702__ || defined __gfx703__ #define AMD_GCN 2 #elif defined __gfx800__ || defined __gfx801__ || defined __gfx802__ || defined __gfx803__ || defined __gfx804__ || defined __gfx810__ #define AMD_GCN 3 // According to AMD docs, GCN 3 and 4 are the same #elif defined __gfx900__ || defined __gfx901__ || defined __gfx902__ || defined __gfx903__ #define AMD_GCN 5 #else #define AMD_GCN 0 #endif #endif #elif VENDOR_ID == (1 << 1) #define IS_APPLE #define IS_GENERIC #elif VENDOR_ID == (1 << 2) #define IS_INTEL_BEIGNET #define IS_GENERIC #elif VENDOR_ID == (1 << 3) #define IS_INTEL_SDK #define IS_GENERIC #elif VENDOR_ID == (1 << 4) #define IS_MESA #define IS_GENERIC #elif VENDOR_ID == (1 << 5) #define IS_NV #elif VENDOR_ID == (1 << 6) #define IS_POCL #define IS_GENERIC #else #define IS_GENERIC #endif /** * function declarations can have a large influence depending on the opencl runtime */ #ifdef IS_NV #define DECLSPEC #else #define DECLSPEC inline #endif /** * AMD specific */ #ifdef IS_AMD #pragma OPENCL EXTENSION cl_amd_media_ops : enable #pragma OPENCL EXTENSION cl_amd_media_ops2 : enable #endif /** * Unrolling is generally enabled, for all device types and hash modes * There's a few exception when it's better not to unroll * Some algorithms run into too much register pressure due to loop unrolling */ // generic vendors: those algos have shown that they produce better results on both amd and nv when not unrolled // so we can assume they will produce better results on other vendors as well #if KERN_TYPE == 1420 #undef _unroll #endif #if KERN_TYPE == 1450 #undef _unroll #endif #if KERN_TYPE == 1460 #undef _unroll #endif #if KERN_TYPE == 1720 #undef _unroll #endif #if KERN_TYPE == 1750 #undef _unroll #endif #if KERN_TYPE == 1760 #undef _unroll #endif #if KERN_TYPE == 1800 #undef _unroll #endif #if KERN_TYPE == 6221 #undef _unroll #endif #if KERN_TYPE == 6222 #undef _unroll #endif #if KERN_TYPE == 6223 #undef _unroll #endif #if KERN_TYPE == 6400 #undef _unroll #endif #if KERN_TYPE == 6500 #undef _unroll #endif #if KERN_TYPE == 6800 #undef _unroll #endif #if KERN_TYPE == 7100 #undef _unroll #endif #if KERN_TYPE == 7400 #undef _unroll #endif #if KERN_TYPE == 7900 #undef _unroll #endif #if KERN_TYPE == 8900 #undef _unroll #endif #if KERN_TYPE == 10700 #undef _unroll #endif #if KERN_TYPE == 13721 #undef _unroll #endif #if KERN_TYPE == 13722 #undef _unroll #endif #if KERN_TYPE == 13723 #undef _unroll #endif #if KERN_TYPE == 13751 #undef _unroll #endif #if KERN_TYPE == 13752 #undef _unroll #endif #if KERN_TYPE == 13753 #undef _unroll #endif #if KERN_TYPE == 13800 #undef _unroll #endif #if KERN_TYPE == 15700 #undef _unroll #endif // nvidia specific #ifdef IS_NV #ifdef IS_GPU #if KERN_TYPE == 1500 #undef _unroll #endif #if KERN_TYPE == 3000 #undef _unroll #endif #if KERN_TYPE == 14000 #undef _unroll #endif #endif #endif // amd specific #ifdef IS_AMD #ifdef IS_GPU #if KERN_TYPE == 8000 #undef _unroll #endif #if KERN_TYPE == 8200 #undef _unroll #endif #if KERN_TYPE == 12300 #undef _unroll #endif #if KERN_TYPE == 14100 #undef _unroll #endif #if KERN_TYPE == 15300 #undef _unroll #endif #if KERN_TYPE == 15900 #undef _unroll #endif #endif #endif // apple specific #ifdef IS_APPLE #if KERN_TYPE == 5000 #undef _unroll #endif #endif