Maxwell Example - continued

Problem:
- array dimensions are small even numbers, power of 2 and map to the same location in both 1st level and the 2nd level caches
- for the Maxwell example the print shows with NX=NY=NZ=64:
- Compiler is able to pad the arrays automatically. Compiling with the default optimizations: -mips4 -O3 gives for the performance 162 Mflop/s

In general:

primary cache 32 KB = 2(way-set-ass) * 4(size-real) * 4096

secondary cache 8 MB = 2(way-set-ass) * 4(size-real) * 1048576

C print position of arrays in memory with the code:

Integer*8 aEX

aEX = %LOC(EX(1,1,1))

print *,’Addr EX=‘,mod(aEX,4096), mod(aEX,1048576),’words’

Addr EX= 3720 470664

Addr EY= 3720 470664

…….. etc.

Addr HZ= 3720 470664

All arrays map to the same locations

in both caches

Problem: