Implement Thumb-2 optimized memcpy/memset
jserv opened this issue · comments
Directory kernel/lib contains the implementation of memcpy and memset, but it is too generic. We can utilize several ARM Cortex-M3/M4 specific features to optimize:
- Thumb-2
- apply 32-bit aligned data copy in inner loop, which is not necessary to Cortex-M3/M4, but it could be better for the external memory access depending on memory controller.
- unaligned memory access
- PLD instruction to preload cache with memory source
lk implements arm-m optimized memcpy and memset routines in git commit littlekernel/lk@33b94d9
It looks so weird. Can you explain?
@jserv The implementation is the branch.
https://github.com/gapry/f9-kernel/blob/benchmark_memcpy/benchmark/benchmark.c
My approach is that measure the case, alignment and unalignment, five times and take the avg time. Assume my approach is correct, the data imply the conclusion is the unalignment case is better than alignment after the optimized on the stm32F407.
@gapry In order to clarify the performance gain, please compare the optimized memcpy routines with plain byte-oriented C version.
The simplest and inefficient implementation of memcpy:
void memcpy(void* src, void* dst, size_t len)
{
char* p = (char*)src;
char* q = (char*)dst;
while(len--) *p++ = *q++;
}
@jserv For now, I use DWT to measure the elapsed clock cycles. You can check the commit: gapry@33e58df
and the completed Implementation: https://github.com/gapry/f9-kernel/blob/benchmark_memcpy/benchmark/benchmark.c
The profile result:
unalignment:
alignment:
