// copy back single float result from GPU and recycle buffer
func copyback(buf unsafe.Pointer) float32 {
var result_ [1]float32
result := result_[:]
cu.MemcpyDtoH(unsafe.Pointer(&result[0]), cu.DevicePtr(buf), 1*cu.SIZEOF_FLOAT32)
reduceBuffers <- buf
return result_[0]
}
func (src *slice) copyDtoH(dst unsafe.Pointer, dstlen int, elemsize int) {
if dstlen != src.Len() {
panic(fmt.Errorf("cuda4/safe: len mismatch: src.Len()=%v (device), len(dst)=%v (host)", src.Len(), dstlen))
}
cu.MemcpyDtoH(dst, src.Pointer(), int64(elemsize)*int64(dstlen))
}
func memCpyDtoH(dst, src unsafe.Pointer, bytes int64) { cu.MemcpyDtoH(dst, cu.DevicePtr(src), bytes) }
