| Message ID | 1491959850-30756-11-git-send-email-cota@braap.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 12/04/2017 09:17, Emilio G. Cota wrote: > + > +/* In user-mode we can get better hashing because we do not have a TLB */ > +static inline unsigned int tb_jmp_cache_hash_func(target_ulong pc) > +{ > + return (pc ^ (pc >> TB_JMP_CACHE_BITS)) & (TB_JMP_CACHE_SIZE - 1); > +} What about multiplicative hashing? return (uint64_t) (pc * 2654435761) >> 32; Paolo
On Wed, Apr 12, 2017 at 11:46:47 +0800, Paolo Bonzini wrote: > > > On 12/04/2017 09:17, Emilio G. Cota wrote: > > + > > +/* In user-mode we can get better hashing because we do not have a TLB */ > > +static inline unsigned int tb_jmp_cache_hash_func(target_ulong pc) > > +{ > > + return (pc ^ (pc >> TB_JMP_CACHE_BITS)) & (TB_JMP_CACHE_SIZE - 1); > > +} > > What about multiplicative hashing? > > return (uint64_t) (pc * 2654435761) >> 32; I tested this one, taking the TB_JMP_CACHE_SIZE-1 lower bits of the result: http://imgur.com/QIhm875 In terms of quality it's good (I profile hit rates and they're all pretty good), but shift+xor are just so hard to beat: shift+xor take 1 cycle each; the multiplication takes on my machine 3 or 4 cycles (source: Fog's tables). Thanks, E.
diff --git a/include/exec/tb-hash.h b/include/exec/tb-hash.h index 2c27490..b1fe2d0 100644 --- a/include/exec/tb-hash.h +++ b/include/exec/tb-hash.h @@ -22,6 +22,8 @@ #include "exec/tb-hash-xx.h" +#ifdef CONFIG_SOFTMMU + /* Only the bottom TB_JMP_PAGE_BITS of the jump cache hash bits vary for addresses on the same page. The top bits are the same. This allows TLB invalidation to quickly clear a subset of the hash table. */ @@ -45,6 +47,16 @@ static inline unsigned int tb_jmp_cache_hash_func(target_ulong pc) | (tmp & TB_JMP_ADDR_MASK)); } +#else + +/* In user-mode we can get better hashing because we do not have a TLB */ +static inline unsigned int tb_jmp_cache_hash_func(target_ulong pc) +{ + return (pc ^ (pc >> TB_JMP_CACHE_BITS)) & (TB_JMP_CACHE_SIZE - 1); +} + +#endif /* CONFIG_SOFTMMU */ + static inline uint32_t tb_hash_func(tb_page_addr_t phys_pc, target_ulong pc, uint32_t flags) {
Optimizations to cross-page chaining and indirect jumps make performance more sensitive to the hit rate of tb_jmp_cache. The constraint of reserving some bits for the page number lowers the achievable quality of the hashing function. However, user-mode does not have this requirement. Thus, with this change we use for user-mode a hashing function that is both faster and of better quality than the previous one. Measurements: - specINT 2006 (test set), x86_64-linux-user. Host: Intel i7-4790K @ 4.00GHz Y axis: Speedup over 95b31d70 1.3x+-+-------------------------------------------------------------+-+ | jr $$ | 1.25x+-+.... jr+xxhash %% ....................................+-+ | jr+hash+inline @@ +++ | 1.2x+-+.............................................................+-+ | @@@ | | +++@@ ++@:@ +++ @@+ | 1.15x+-+..................$$$@@...............$$@.@.......@@...@@....+-+ | $ $@@ $$@ @ %%@ @@ | 1.1x+-+..................$.$@@...............$$@.@......%%@.$$@@....+-+ | +++@@+ $ $@@ $$@ @ ++%%@+$$@@ +++| 1.05x+-+.........$$@@.....$.$@@...@@..........$$@.@..@@@.%%@.$$@@...@@-+ | $$@@ $ $@@$$$@@ $$% @$$@+@$$%@ $$@@+$$@@ | |+$$++++++++$$@@+++@@$ $@@$+$@@+++@@$$+@@$$% @$$@+@$$%@ $$%@ $$@@ | 1x+-$$@@A$$%@R$$@@R$$@@$_$%@$_$%@$$s@@$$%%@$$%.@$$%.@$$%@.$$%@.$$%@-+ | $$@@+$$%@ $$%@ $$@@$+$%@$ $%@$$%%@$$+%@$$% @$$% @$$%@ $$%@ $$%@ | 0.95x+-$$%@.$$%@.$$%@.$$%@$.$%@$.$%@$$.%@$$.%@$$%.@$$%.@$$%@.$$%@.$$%@-+ | $$%@ $$%@ $$%@ $$%@$ $%@$ $%@$$ %@$$ %@$$% @$$% @$$%@ $$%@ $$%@ | 0.9x+-$$%@-$$%@-$$%@-$$%@$$$%@$$$%@$$%%@$$%%@$$%@@$$%@@$$%@-$$%@-$$%@-+ astabzip2 gcc gobmh264rehmlibquantumcfomneperlbensjexalanchmean png: http://imgur.com/RiaBuIi That is, a 6.45% hmean improvement for this commit. Note that this is the test set, so some benchmarks take almost no time (and therefore aren't that sensitive to changes here). See "train" results below. Note also that hashing quality is not the only requirement: xxhash gives on average the highest hit rates. However, the time spent computing the hash negates the performance gains coming from the increase in hit rate. Given these results, I dropped xxhash from subsequent experiments. - specINT 2006 (train set), x86_64-linux-user. Host: Intel i7-4790K @ 4.00GHz Y axis: Speedup over 95b31d70 1.4x+-+--------------------------------------------------------------+-+ | jr $$ +++ | | jr+hash %% : | 1.3x+-+.......................................................%%%....+-+ | +++ +++ %:% | | +++ %%% : %+% | 1.2x+-+.....................%%......................%.%..%%%.$$.%....+-+ | ++%% %%% $$+% %:% $$+% | | +++ $$$% $$+% $$ % %:% $$ % | 1.1x+-+...........%%......$.$%................$$.%.$$.%.$$.%.$$.%..%%%-+ | +++ %% $ $% +++ $$ % $$ % $$ % $$ % +%+% | | ++%% +++ ++%% ++%% $ $% $$$+ +++ %%% $$ % $$ % $$ % $$ % $$+% | 1x+-$$$%RGR%%R$$$%H$$$%P$j$%h$s$%.$$%%..%.%.$$.%.$$.%.$$.%.$$.%.$$.%-+ | $+$% $$$% $ $% $+$% $ $% $ $% $$+% % % $$ % $$ % $$ % $$ % $$ % | | $ $% $ $% $ $% $ $% $ $% $ $% $$ % % % $$ % $$ % $$ % $$ % $$ % | 0.9x+-$.$%.$.$%.$.$%.$.$%.$.$%.$.$%.$$.%..%.%.$$.%.$$.%.$$.%.$$.%.$$.%-+ | $ $% $ $% $ $% $ $% $ $% $ $% $$ % $$+% $$ % $$ % $$ % $$ % $$ % | | $ $% $ $% $ $% $ $% $ $% $ $% $$ % $$ % $$ % $$ % $$ % $$ % $$ % | 0.8x+-$$$%-$$$%-$$$%-$$$%-$$$%-$$$%-$$%%-$$%%-$$%%-$$%%-$$%%-$$%%-$$%%-+ astarbzip2 gcc gobmh264rehlibquantumcfomneperlbensjexalancbhmean png: http://imgur.com/55iJJgD That is, a 10.19% hmean improvement for jr+hash (this commit). - NBench, arm-linux-user. Host: Intel i7-4790K @ 4.00GHz Y axis: Speedup over 95b31d70 1.35x+-+-------------------------------------------------------------+-+ | @@@ jr $$ | 1.3x+-+.............@.@. jr+inline %% ...@@@................+-+ | @ @ jr+inline+hash @@ @ @ | | @ @ @ @ | 1.25x+-+.............@.@..........................@.@................+-+ | @ @ @@@ @ @ | 1.2x+-+.............@.@..................$$%.@...@.@................+-+ | @ @ $$% @ @ @ | | @ @ %%@ $$% @ %% @ | 1.15x+-+.............@.@........%%@.......$$%.@$$$%.@................+-+ | @ @ %%@ $$% @$ $% @ | 1.1x+-+.............@.@......$$$%@.......$$%.@$.$%.@...............@@-+ | @ @ $ $%@ $$% @$ $% @ @@ | | @ @ $ $%@ $$%%@ $$% @$ $% @ $$%%@ | 1.05x+-+...........$$%.@$$$%@@$.$%@.$$.%@.$$%.@$.$%.@.........@@.$$.%@-+ | $$%%@ $$% @$ $% @$ $%@ $$ %@ $$% @$ $% @ %%%@ $$ %@ | 1x+-$$.%@AR%%%@R$$%B@$G$%P@$T$%@_$$+%@l$$%+@$s$%.@$$$%@.$$.%@.$$.%@-+ +-$$%%@-$$%%@-$$%@@$$$%@@$$$%@-$$%%@-$$%@@$$$%@@$$$%@-$$%%@-$$%%@-+ ASSIGNMBITFIELFOFP_EMULATHUFFMANLU_DECOMPNEURNUMERICSTRING_SOhmean png: http://imgur.com/i5e1gdY That is, a 11% hmean perf gain--it almost doubles the perf gain from implementing the jr optimization. - NBench, x86_64-linux-user. Host: Intel i7-4790K @ 4.00GHz 1.1x+-+-------------------------------------------------------------+-+ | jr $$ | 1.08x+-+..... jr+inline %% ...................................+-+ | jr+inline+hash @@ | | $$ @@ | 1.06x+-$$.@@.........................%%%.............................+-+ | $$%%@ % % | 1.04x+-$$.%@.........................%.%.............................+-+ | $$ %@ @@@ $$ % $$ | | $$ %@ @ @ %% $$ % $$%%@ | 1.02x+-$$.%@........%%.@$$$%@@......$$.%@..%%@@..%%........$$.%@.$$%%@-+ | $$ %@ @@ %% @$ $% @$$$ $$ %@ $$% @ %%@@$$$% $$ %@ $$ %@ | 1x+-$$.%@A$$R@@RG%%B@$G$%P@$T$%P_$$T%@h$$%+@$$$%e@$.$%@.$$.%@.$$.%@-+ | $$ %@ $$%%@ $$% @$ $% @$ $% $$ %@ $$% @$ $% @$ $%@ $$ %@ $$ %@ | 0.98x+-$$.%@.$$.%@.$$%.@$.$%.@$.$%@.$$.%@.$$%.@$.$%.@$.$%@.$$.%@.$$.%@-+ | $$ %@ $$ %@ $$% @$ $% @$ $%@ $$ %@ $$% @$ $% @$ $%@ $$ %@ $$ %@ | | $$ %@ $$ %@ $$% @$ $% @$ $%@ $$ %@ $$% @$ $% @$ $%@ $$ %@ $$ %@ | 0.96x+-$$.%@.$$.%@.$$%.@$.$%.@$.$%@.$$.%@.$$%.@$.$%.@$.$%@.$$.%@.$$.%@-+ +-$$%%@-$$%%@-$$%@@$$$%@@$$$%@-$$%%@-$$%@@$$$%@@$$$%@-$$%%@-$$%%@-+ ASSIGNMBITFIELFOFP_EMULATHUFFMANLU_DECOMPNEURNUMERICSTRING_SOhmean png: http://imgur.com/Xu0Owgu The fact that NBench is not very sensitive to changes here was mentioned in the previous commit's log. We get a very slight overall decrease in hmean performance, although some workloads improve as well. Note that there are no error bars: NBench re-runs itself until confidence on the stability of the average is >= 95%, and it doesn't report the resulting stddev. Signed-off-by: Emilio G. Cota <cota@braap.org> --- include/exec/tb-hash.h | 12 ++++++++++++ 1 file changed, 12 insertions(+)