Post a Comment On: cbloom rants

How to do an atomic load/store of a 128 bit value on x64 is an ugly question. The only guaranteed way is via cmpxchg16b. But that's an awfully slow way to do a load or store. movdqa appears to be an atomic way to move 128 bits - on most chips. Not all. And Intel/AMD don't want to clearly identify the cases where it is atomic or not. (they specifically don't guarantee it) At the moment, shipping code needs to use cmpx16 to be safe. (my tests indicate that the x64 chips in the modern game consoles *do* have atomic movdqa, so it seems safe to use there) My main curiousity is whether there exist any modern ("Core 2" or newer) x64 chips that *do not* provide an atomic movdqa. Anyhoo, here's a test to check if movdqa is atomic on your machine. If you like, run it and send me the results : (Windows, x64 only) cb_x64_atomics_test.7z The atomic test will just run for a while. If it has a failure it will break. (you will get some errors about not having a v: or r: drive; you can just ignore them.) Copy the output, or should be able to get the log file here : "c:\oodlelogs\oodle_cb_test_x64_atomics.log" "c:\oodlelogs\oodle_cb_test_x64_atomics.prev" email me at cb at my domain. For the record, an atomic 128 bit load/store for x64 Windows using cmpx16 : #include intrin.h> void AtomicLoad128(__int64 * out_Result, __int64 volatile * in_LoadFrom) { // do a swap of the value in out_Result to itself // if it matches, it stays the same // it it doesn't match, we get a load _InterlockedCompareExchange128(in_LoadFrom,out_Result[1],out_Result[0],out_Result); } void AtomicStore128(__int64 volatile * out_StoreTo,const __int64 * in_Value) { // do an initial non-atomic load of StoreTo : __int64 check_StoreTo[2]; check_StoreTo[0] = out_StoreTo[0]; check_StoreTo[1] = out_StoreTo[1]; // store with cmpx16 : while( ! _InterlockedCompareExchange128(out_StoreTo,in_Value[1],in_Value[0],check_StoreTo) ) { // check_StoreTo was reloaded with the value in out_StoreTo _mm_pause(); } }