func MemCpy(dst, src unsafe.Pointer, bytes int64) {
Sync()
timer.Start("memcpy")
cu.MemcpyAsync(cu.DevicePtr(uintptr(dst)), cu.DevicePtr(uintptr(src)), bytes, stream0)
Sync()
timer.Stop("memcpy")
}
func MemCpyHtoD(dst, src unsafe.Pointer, bytes int64) {
Sync() // sync previous kernels
timer.Start("memcpyHtoD")
cu.MemcpyHtoD(cu.DevicePtr(uintptr(dst)), src, bytes)
Sync() // sync copy
timer.Stop("memcpyHtoD")
}
func MemCpyDtoH(dst, src unsafe.Pointer, bytes int64) {
Sync() // sync previous kernels
timer.Start("memcpyDtoH")
cu.MemcpyDtoH(dst, cu.DevicePtr(uintptr(src)), bytes)
Sync() // sync copy
timer.Stop("memcpyDtoH")
}
func newSlice(nComp int, size [3]int, alloc func(int64) unsafe.Pointer, memType int8) *data.Slice {
data.EnableGPU(memFree, cu.MemFreeHost, MemCpy, MemCpyDtoH, MemCpyHtoD)
length := prod(size)
bytes := int64(length) * cu.SIZEOF_FLOAT32
ptrs := make([]unsafe.Pointer, nComp)
for c := range ptrs {
ptrs[c] = unsafe.Pointer(alloc(bytes))
cu.MemsetD32(cu.DevicePtr(uintptr(ptrs[c])), 0, int64(length))
}
return data.SliceFromPtrs(size, memType, ptrs)
}
// Memset sets the Slice's components to the specified values.
// To be carefully used on unified slice (need sync)
func Memset(s *data.Slice, val ...float32) {
if Synchronous { // debug
Sync()
timer.Start("memset")
}
util.Argument(len(val) == s.NComp())
for c, v := range val {
cu.MemsetD32Async(cu.DevicePtr(uintptr(s.DevPtr(c))), math.Float32bits(v), int64(s.Len()), stream0)
}
if Synchronous { //debug
Sync()
timer.Stop("memset")
}
}
func memFree(ptr unsafe.Pointer) { cu.MemFree(cu.DevicePtr(uintptr(ptr))) }