keeping values in cache

[C_ntua]

Is there a way to explicit tell the compiler to keep a value in the cache for the whole loop? Like

for (j=0; j<MAX; j++)
   for (i=0; i<MAX; i++)
      y[j] += a[i] * b[i]

...and tell the compiler to keep the value y[i] for the whole inner loop, gaining optimization. I know it is possible, but i don't know how or if the compiler would do this anyway.

[laserlight]

I am not sure if the compiler has control over what remains in the cache (and from what I understand modern compilers routinely ignore suggestions of what should be placed in a register), but in any case is this a bottle neck in your code to begin with?

[Elysia]

You can tell the compiler you want something placed in a register with the "register" keyword.
However, it's not advised, as the compiler can probably do better optimizations than you.
In fact, many compilers would put both i and j in registers for this loop.

What remains in the CPU cache cannot be controlled (for x86/x86-64).

[robwhit]

y[i] is not used in the loop except when i == j. Why would you want to cache it?

[C_ntua]

y[i] is not used in the loop except when i == j. Why would you want to cache it?

y[j] you mean. What do you mean exactly? Why isn't it used?

Yeah, that's more or less the code I want to optimize. I read a paper that said that they had the system write back in the main memory y[j] only after the inner loop was done, gaining a speed advantage, since it will be used in every inner loop. But I don't know how it is done. Only that it is done in a C program on a linux system.

[tabstop]

Do you see the symbols "y[i]" anywhere in your code?

And I would be extraordinarily surprised if any compiler at all doesn't do that optimatization as a matter of course.

(Edit: I mean adding a bunch of stuff to y[j], and only then writing it back. I'm not talking about cache.)

[Salem]

It sounds like some platform specific call to mess about with the cache control registers. But it would have to be a pretty large loop to offset the overheads of two system calls.

Unless there's some specific compiler flag (you'd have to look yourself).

[iMalc]

Is there a way to explicit tell the compiler to keep a value in the cache for the whole loop? Like

for (j=0; j<MAX; j++)
   for (i=0; i<MAX; i++)
      y[j] += a[i] * b[i]

...and tell the compiler to keep the value y[i] for the whole inner loop, gaining optimization. I know it is possible, but i don't know how or if the compiler would do this anyway.

It sounds like you want to do this (assuming you mean y[j], not y[i] in your description):

for (j=0; j<MAX; j++)
{
   double t=0;
   for (i=0; i<MAX; i++)
      t += a[i] * b[i];
   y[j] += t;
}

(Assuming a and b hold an array of doubles)
Don't bother unless you find that it is a bottleneck though.