prefetch/noprefetch

Invites the compiler to issue or disable requests to prefetch data from memory. This pragma only applies to Intel® MIC Architecture.

Syntax

#pragma prefetch

#pragma prefetch *:hint[:distance]

#pragma prefetch [var1 [: hint1 [: distance1]] [, var2 [: hint2 [: distance2]]]...]

#pragma noprefetch [var1 [, var2]...]

Arguments

var

An optional memory reference (data to be prefetched)

hint

An optional hint to the compiler to specify the type of prefetch. Possible values are the constants defined in the header xmmintrin.h:

  • _MM_HINT_T0 - for integer data that will be reused

  • _MM_HINT_T1 - for integer and floating point data that will be reused from L2 cache

  • _MM_HINT_T2 - for data that will be reused from L3 cache

  • _MM_HINT_NTA - for data that will not be reused

To use this argument, you must also specify var.

distance

An optional integer argument with a value greater than 0. It indicates the number of loop iterations ahead of which a prefetch is issued, before the corresponding load or store instruction. To use this argument, you must also specify var and hint.

Description

This pragma hints to the compiler to generate data prefetches for some memory references. These hints affect the heuristics used in the compiler. Prefetching data can minimize the effects of memory latency.

If you specify the prefetch pragma with no arguments, all arrays accessed in the immediately following loop are prefetched.

If the loop includes the expression A(j), placing #pragma prefetch A in front of the loop instructs the compiler to insert prefetches for A(j + d) within the loop. Here, d is the number of iterations ahead of which to prefetch the data, and is determined by the compiler.

If you specify #pragma prefetch *, then hint and distance prefetches all array accesses in the loop.

To use these pragmas, compiler option [Q]opt-prefetch must be set (it is turned on by default if the compiler general optimization level is set at option O2 or higher).

The noprefetch pragma hints to the compiler not to generate data prefetches for some memory references. This affects the heuristics used in the compiler.

Examples

Example: Using the prefetch pragma

#pragma prefetch htab_p:1:30 
#pragma prefetch htab_p:0:6

// Issue vprefetch1 for htab_p with a distance of 30 vectorized iterations ahead
// Issue vprefetch0 for htab_p with a distance of 6 vectorized iterations ahead
// If pragmas are not present, compiler chooses both distance values 

for (j=0; j<2*N; j++) { htab_p[i*m1 + j] = -1; }

Example: Using noprefetch and prefetch pragmas together

#pragma noprefetch b
#pragma prefetch a 
for(i=0; i<m; i++) { a[i]=b[i]+1; }

Example: Using noprefetch and prefetch pragmas together

for (i=i0; i!=i1; i+=is) {
 
float sum = b[i]; 
int ip = srow[i]; 
int c = col[ip];

#pragma noprefetch col 
#pragma prefetch value:1:80 
#pragma prefetch x:1:40

for(; ip<srow[i+1]; c=col[++ip])
  sum -= value[ip] * x[c];
  y[i] = sum; 
}
For more complete information about compiler optimizations, see our Optimization Notice.