func Init() {
if cudaCtx != 0 {
return // already inited
}
var flag uint
switch *Flag_sched {
default:
panic("sched flag: expecting auto,spin,yield or sync: " + *Flag_sched)
case "auto":
flag = cu.CTX_SCHED_AUTO
case "spin":
flag = cu.CTX_SCHED_SPIN
case "yield":
flag = cu.CTX_SCHED_YIELD
case "sync":
flag = cu.CTX_BLOCKING_SYNC
}
tryCuInit()
dev := cu.Device(*Flag_gpu)
cudaCtx = cu.CtxCreate(flag, dev)
M, m := dev.ComputeCapability()
concurrent := dev.Attribute(cu.CONCURRENT_KERNELS)
log.Print("CUDA ", float32(cu.Version())/1000, " ",
dev.Name(), "(", (dev.TotalMem())/(1024*1024), "MB) ",
"compute ", M, ".", m,
" concurrent: ", concurrent == 1, "\n")
cudaCC = 10*M + m
log.Println("set preference for device L1 cache")
cuda.DeviceSetCacheConfig(cuda.FUNC_CACHE_PREFER_L1)
initStreampool()
}
func InitCuda() {
runtime.LockOSThread()
cu.Init(0)
cu.CtxCreate(cu.CTX_SCHED_AUTO, 0).SetCurrent()
}
// needed for all other tests.
func init() {
cu.Init(0)
ctx := cu.CtxCreate(cu.CTX_SCHED_AUTO, 0)
cu.CtxSetCurrent(ctx)
fmt.Println("Created CUDA context")
}
