Performance is not the main concern here, but cycle count measurement accuracy. E.g. Intel rdtsc cycle counter read latency is few cycles where as clock_gettime() seems to take over a thousand. System call latencies also vary since those may launch other kernel services, which adds jitter to cycle count measurements.
Could we organize arch specific stuff better (no #ifdefs, but configure/build mechanism) and continue to support e.g. rdtsc. -Petri From: [email protected] [mailto:[email protected]] On Behalf Of ext Bill Fischofer Sent: Friday, March 06, 2015 6:42 PM To: Mike Holmes Cc: lng-odp Subject: Re: [lng-odp] [PATCH] linux-generic: time: remove platform specific acceleration I agree with Mike. Now that we have an official LNG performance implementation, linux-generic should focus on being a simple and clean functional implementation that elaborates the intended behavior of the ODP APIs without being overly concerned with performance. That doesn't mean we don't do straightforward functional optimizations that are algorithmic in nature, just that there shouldn't be any platform-specific tuning paths in linux-generic. On Fri, Mar 6, 2015 at 10:35 AM, Mike Holmes <[email protected]<mailto:[email protected]>> wrote: The DPDK implementation will take the X86 optimizations, presumably Cavium already have Octeon acceleration in their own ODP implementation. The linux generic implementation should be a model for the behavior of ODP, it happens to run in Linux but that should not be confused with it being used in any real application. If we fill it with optimizations it will be harder to maintain as a pure model. Mike On 6 March 2015 at 11:26, Maxim Uvarov <[email protected]<mailto:[email protected]>> wrote: Where should be that implementations placed if we want other platform to reuse them? Maxim. On 03/06/15 19:22, Mike Holmes wrote: Linux-generic should be a reference implementation and does not need archetecture specific acceleration. Signed-off-by: Mike Holmes <[email protected]<mailto:[email protected]>> --- platform/linux-generic/odp_time.c | 47 +++------------------------------------ 1 file changed, 3 insertions(+), 44 deletions(-) diff --git a/platform/linux-generic/odp_time.c b/platform/linux-generic/odp_time.c index 31d6656..6f40361 100644 --- a/platform/linux-generic/odp_time.c +++ b/platform/linux-generic/odp_time.c @@ -6,52 +6,13 @@ #define _POSIX_C_SOURCE 200809L +#include <time.h> #include <odp/time.h> -#include <odp/hints.h> #include <odp/system_info.h> - -#define GIGA 1000000000 - -#if defined __x86_64__ || defined __i386__ - -uint64_t odp_time_cycles(void) -{ - union { - uint64_t tsc_64; - struct { - uint32_t lo_32; - uint32_t hi_32; - }; - } tsc; - - __asm__ __volatile__ ("rdtsc" : - "=a" (tsc.lo_32), - "=d" (tsc.hi_32) : : "memory"); - - return tsc.tsc_64; -} - - -#elif defined __OCTEON__ - -uint64_t odp_time_cycles(void) -{ - #define CVMX_TMP_STR(x) CVMX_TMP_STR2(x) - #define CVMX_TMP_STR2(x) #x - uint64_t cycle; - - __asm__ __volatile__ ("rdhwr %[rt],$" CVMX_TMP_STR(31) : - [rt] "=d" (cycle) : : "memory"); - - return cycle; -} - -#else - -#include <time.h> -#include <stdlib.h> #include <odp_debug_internal.h> +#define GIGA 1000000000 + uint64_t odp_time_cycles(void) { struct timespec time; @@ -74,8 +35,6 @@ uint64_t odp_time_cycles(void) return cycles; } -#endif - uint64_t odp_time_diff_cycles(uint64_t t1, uint64_t t2) { if (odp_likely(t2 > t1)) _______________________________________________ lng-odp mailing list [email protected]<mailto:[email protected]> http://lists.linaro.org/mailman/listinfo/lng-odp -- Mike Holmes Technical Manager - Linaro Networking Group Linaro.org<http://www.linaro.org/> │ Open source software for ARM SoCs _______________________________________________ lng-odp mailing list [email protected]<mailto:[email protected]> http://lists.linaro.org/mailman/listinfo/lng-odp
_______________________________________________ lng-odp mailing list [email protected] http://lists.linaro.org/mailman/listinfo/lng-odp
