// called by init()
func initGPUProf(OD string) {
if *Flag_gpuprof {
util.PanicErr(os.Setenv("CUDA_PROFILE", "1"))
util.PanicErr(os.Setenv("CUDA_PROFILE_CSV", "1"))
out := OD + "gpuprofile.csv"
log.Println("writing GPU profile to", out)
util.PanicErr(os.Setenv("CUDA_PROFILE_LOG", out))
cfgfile := OD + "cudaprof.cfg"
util.PanicErr(os.Setenv("CUDA_PROFILE_CONFIG", cfgfile))
util.FatalErr(ioutil.WriteFile(cfgfile, []byte(CUDA_PROFILE_CONFIG), 0666), "gpuprof")
//AtExit(cuda.DeviceReset)
}
}
// generate wrapper code from template
func wrapgen(filename, funcname string, argt, argn []string) {
kernel := &Kernel{funcname, argt, argn, make(map[int]string)}
for _, s := range cc {
kernel.PTX[s] = filterptx(fmt.Sprint(util.NoExt(filename), "_", s, ".ptx"))
}
wrapfname := util.NoExt(filename) + "_wrapper.go"
wrapout, err := os.OpenFile(wrapfname, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0666)
util.PanicErr(err)
defer wrapout.Close()
util.PanicErr(templ.Execute(wrapout, kernel))
}
// Filter comments and ".file" entries from ptx code.
// They spoil the git history
func filterptx(fname string) string {
f, err := os.Open(fname)
util.PanicErr(err)
defer f.Close()
in := bufio.NewReader(f)
var out bytes.Buffer
out.Write(([]byte)("`"))
line, err := in.ReadBytes('\n')
for err != io.EOF {
util.PanicErr(err)
if !bytes.HasPrefix(line, []byte("//")) && !bytes.HasPrefix(line, []byte(" .file")) {
out.Write(line)
}
line, err = in.ReadBytes('\n')
}
out.Write(([]byte)("`"))
return out.String()
}
// generate cuda wrapper for file.
func cuda2go(fname string) {
// open cuda file
f, err := os.Open(fname)
util.PanicErr(err)
defer f.Close()
// read tokens
var token []string
var s scanner.Scanner
s.Init(f)
tok := s.Scan()
for tok != scanner.EOF {
if !filter(s.TokenText()) {
token = append(token, s.TokenText())
}
tok = s.Scan()
}
// find function name and arguments
funcname := ""
argstart, argstop := -1, -1
for i := 0; i < len(token); i++ {
if token[i] == "__global__" {
funcname = token[i+2]
argstart = i + 4
}
if argstart > 0 && token[i] == ")" {
argstop = i + 1
break
}
}
argl := token[argstart:argstop]
// isolate individual arguments
var args [][]string
start := 0
for i, a := range argl {
if a == "," || a == ")" {
args = append(args, argl[start:i])
start = i + 1
}
}
// separate arg names/types and make pointers Go-style
argn := make([]string, len(args))
argt := make([]string, len(args))
for i := range args {
if args[i][1] == "*" {
args[i] = []string{args[i][0] + "*", args[i][2]}
}
argt[i] = typemap(args[i][0])
argn[i] = args[i][1]
}
wrapgen(fname, funcname, argt, argn)
}
// path to the executable.
func procSelfExe() string {
me, err := os.Readlink("/proc/self/exe")
util.PanicErr(err)
return me
}