sh0dan writes:
It seems movntq is only suited for direct memory copies - otherwise the penalty is simply too big.
The movntq was a big surprise for me too. It is definately faster in direct copying (BitBlt) for instance, but not very useful in routines that actually do some processing.
The data is not read again, and it is 8-byte aligned. But I guess the problem lies in the fast, that movntq cannot be defered. A movq to memory can be stored in the data cache for later storage, whereas movntq must be dispatched directly. When doing some processing the processor can take time to do the write.
I saw some really big penalties in the AMD Pipeline analysis tool (most movntq's took an average of 60 cycles!) It showed penalties for Data Cache miss, and the Load/Store Queue being full. Using movq's, the average cycle/instruction was about 2-4, with the load-store queue doing ok.
It should be noted that the system I tested on were Athlon Tbird 1200, and an Athlon XP 2200+ - both with DDR RAM.
The conclusion can only be to benchmark your code, to be sure if you are actually gaining something from using movntq.
phaeron writes:
MOVNTQ can be a win even in non-trivial routines if used carefully -- VirtualDub's YCbCr?-to-RGB conversions use it for a significant gain on large MPEG-1 files (around 640x360). However, it is definitely not to be used lightly.
The purpose of MOVNTQ is to defeat a cache optimization known as write-back. The early caches introduced into x86 systems were initially write-through, meaning that writes to memory not present in the cache went directly to main memory. This led to some interesting behavior on Pentium systems where doing dummy reads periodically during heavy write loops would boost throughput, because the reads would pull memory into cache, the writes would then be combined in the cache, and a lot more burst transfers would occur over the bus.
Beginning with the Pentium Pro, Intel started using write-back as the L1 cache strategy. Whenever a write occurs to an uncached location, the CPU pulls an entire cache line of data from main memory, pushes the writes into the L1 cache, and then later writes the line back to main memory. This is typically a win as burst transfers to and from the cache are generally faster than piecemeal 32-bit accesses. Where write-back hurts is when code is only writing to memory -- in this case, the initial read is a waste as all of the cache line is going to be overwritten anyway.
The MOVNTQ instruction tells the CPU to use write combining (WC) instead of write-back (WB) as the mode for memory. Writes are then collected in write combining buffers, which are bursted out to memory once full. This can be much, much faster than traditional methods large memcpy() operations in particular can be more than twice as fast but there are a few caveats, as you noticed. The biggest is that MOVNTQ is only efficient when data must be bounced back to main memory anyway; it is a loss when data can stay within the caches, as otherwise without MOVNTQ the data would stay within the much faster L1 and L2 caches. It is never a good idea to use MOVNTQ when you know the target is in the cache. Another is that you must write all bytes in a cache line. Piecemeal writes with MOVNTQ result in partial bus transactions, which means that instead of a nice long burst, you get a bunch of small writes. In a WB scenario the CPU already has all the data in the cache line and can rewrite the unwritten parts with the same data, but it can't do this when streaming writes are involved. Processing vertical columns of single pixels at a time with MOVNTQ involved, for example, would be a bad idea.
Profiling MOVNTQ can be tricky. A memcpy() rewritten to use block prefetching and streaming stores can execute more than twice as fast as REP MOVSD. What the profile won't tell you is that by converting your memcpy() to streaming, you slowed down the routine right after it that now suffers lots of cache misses pulling the data back into memory. MOVNTQ is thus most useful when you are already suffering cache misses because your buffers are way too big, sit too long between the write and the next read, or need to bounce between CPUs. Sometimes an even better idea than streaming stores is to block the various pipeline stages into strips, thus keeping each strip in cache -- Intel calls this "strip mining." However, it is very difficult to do when many different modules are involved, such as in an Avisynth filter graph.
