Prefechind data from main memory into cache

I have a huge program which do a lot of different tasks. There are hundreds, and one of them is Transposition Table access. This TT is very big, and I access it from time to time to retrieve a cell of around 8 or 10 ints size.

As my program has to be tuned for speed I have used a _mm_prefetch intrinsic instruction for trying to prefetch data a few instruction before real call.

I have trying to measuring time doing differents test but I have not get any improvement, so I ask myselft if I have programmed the code well.

Is there any way of knowing if the data have move into cache after the prefetch? .... or is there any way of knowing if _mm_prefecth instructions has been sucessfull?

My code is near like this:

Code:

---

        unsigned int nt = n->hash & total_hash_entries;
        t_entrada_hash *p = &ptrTT[nt];

        _mm_prefetch((char*) p, _MM_HINT_NTA)

        .... a few lines of code here.....

        use of p.... i.e.:
        if (p->x = z...)

---

Well a long read of chapter 7 of this would be good.
http://www.intel.com/design/processo...als/248966.pdf
In particular, I think you need a different constant, and more code between prefetch and 'use' to let it happen.

This might be as well.
http://vast.uccs.edu/~tboult/FRAME/Cutler/frame99.htm

Have you actually used a profiler on your code? What makes you think a change in CPU cache usage will provide any speed-up?

gg

The prefetch instruction is "advisory" to the processor, so depending on what is going on, the processor may not perform a prefetch at all.

To determine if it "helps", you will have to benchmark the program with and without the prefetch instruction.

Edit: And unless you have several dozen instructions between the prefetch and the actual use of the data, it's unlikely that you will gain very much. If you do fetches in a loop, the recommended "advance" is that you fetch some 256 or 512 bytes ahead of where you are processing. But that only really helps if you have loops that access in a predictable pattern. This looks like a hash-table access, so you (probably) can't determine where the next access will be.

--
Mats

Quote:

---

Originally Posted by Codeplug

What makes you think a change in CPU cache usage will provide any speed-up?
gg

---

well, commmon sense: if can prefetch data into cache a hundreds instructions before use them, I will have a few cycles gain.

Use a profiler to tell you where to spend your time, otherwise your wasting it.

gg