Write is twice the speed as read?

[Ioi Lam <Ioi.Lam@xxxxxxx>]

Folks,

I am trying to optimize large block memcpy speed on my Versatile A/B 
board (210 Mhz ARM926EJ-S with VFP, Linux 2.6.9). So I wrote a few 
timing loops to determine the speed of moving memory. For my 
requirement, I need to copy a block of memory that's 2x the size of the 
dcache.

So I wrote two loops to time the speed of reading or writing (but not 
reading *and* writing at the same time). With a 32KB dcache, I try to 
read or write 4KB or 64KB of data using load/store multiple 
instructions. The strange results are that large blocks of writes are 
about 2x the speed of reads. The speed of memcpy is measured using the 
standard GNU C library.

rd_ldmia(4)     read 4KB of data (ldmia)      134217728 bytes in     220864 us
                                               607.694 MByte/sec
rd_ldmia(64)    read 64KB of data (ldmia)     134217728 bytes in    1030162 us
                                               130.287 MByte/sec
wd_stmia(4)     write 4KB of data (stmia)     134217728 bytes in     544041 us
                                               246.705 MByte/sec
wd_stmia(64)    write 64KB of data (stmia)    134217728 bytes in     539586 us
                                               248.742 MByte/sec
memcpy(4)       memcpy 4KB of data            134217728 bytes in     575432 us
                                               233.246 MByte/sec
memcpy(64)      memcpy 64KB of data           134217728 bytes in    1510152 us
                                                88.876 MByte/sec


Does anyone know why the reads are just half the speed of the writes? 
Maybe I am not doing the right thing to use the full bandwidth between 
main_mem -> dcache? I am hoping that if I can get the reads to be as 
fast as the writes, memcpy can achieve around 120MB/sec.

--------------------------------------------------------------------------------------------------

        @void rd_ldmia(int *addr, int num_bytes, int num_loops);
.global    rd_ldmia
rd_ldmia:
       stmdb   sp!, {r4, r5, r6, r7, r8}
       mov     r8, r0
rd_ldmia_0:
       mov     r3, r1
rd_ldmia_1:
       ldmia   r0!, {r4, r5, r6, r7}
       ldmia   r0!, {r4, r5, r6, r7}
       ldmia   r0!, {r4, r5, r6, r7}
       ldmia   r0!, {r4, r5, r6, r7}
       subs    r3, r3, #64
       bne     rd_ldmia_1
       subs    r2, r2, #1
       movne   r0, r8
       bne     rd_ldmia_0

       ldmia   sp!, {r4, r5, r6, r7, r8}
       mov     pc, lr

        @void wd_stmia(int *addr, int num_bytes, int num_loops);
.global    wd_stmia
wd_stmia:
       stmdb   sp!, {r4, r5, r6, r7, r8}
       mov     r8, r0
wd_stmia_0:
       mov     r3, r1
wd_stmia_1:
       stmia   r0!, {r4, r5, r6, r7}
       stmia   r0!, {r4, r5, r6, r7}
       stmia   r0!, {r4, r5, r6, r7}
       stmia   r0!, {r4, r5, r6, r7}
       subs    r3, r3, #64
       bne     wd_stmia_1
       subs    r2, r2, #1
       movne   r0, r8
       bne     wd_stmia_0

       stmia   sp!, {r4, r5, r6, r7, r8}
       mov     pc, lr



-------------------------------------------------------------------
List admin: http://lists.arm.linux.org.uk/mailman/listinfo/linux-arm
FAQ:        http://www.arm.linux.org.uk/mailinglists/faq.php
Etiquette:  http://www.arm.linux.org.uk/mailinglists/etiquette.php