| Message ID | 20210315100230.GF6064@kam.mff.cuni.cz |
|---|---|
| State | New |
| Headers | show |
| Series | znver3 tuning part 1 | expand |
[AMD Official Use Only - Internal Distribution Only] Hi Honza, Thank you for working on this. > -----Original Message----- > From: Gcc-patches <gcc-patches-bounces@gcc.gnu.org> On Behalf Of Jan > Hubicka > Sent: Monday, March 15, 2021 3:33 PM > To: gcc-patches@gcc.gnu.org; mjambor@suse.cz > Subject: znver3 tuning part 1 > > [CAUTION: External Email] > > Hi, > I plan to commit some retuning of znver3 codegen that is based on real > hardware benchmarks. It turns out that there are not too many changes > necessary sinze Zen3 is quite smooth upgrade to Zen2. In summary: > > - some instructions (like idiv) have shorter latencies. Adjusting > costs reduces code size a bit but seems within noise in benchmark > (since our cost calculation is quite off anyway because it does not > account register pressure and parallelism that does make huge > difference here) > - gather instructions are still microcoded but a lot faster than in > znver1/znver2 and it turns out they are now beneficial for few tsmc > benchmarks, so I plan to enable them. Can we get a copy of this benchmark to try ? we need to check on bigger benchmarks like SPEC also. > > It seems we missed revisiting this for znver2 tuning. > I think even for znver2 it may make sense to re-enable them, so I > will benchmark this as well. > - memcpy/memset expansion seems to work same way as for znver2, > so I am keeping same changes. > - instruction scheduler is already modified in trunk to some degree > reflecting new units. Problem with instruction scheduling is that > it treats zen as in-order CPU and is unlikely going to fill all > execution resources this way. > We may want to try to model the out-of-order nature similar way as > LLVM does, but at the other hand the current scheduling logic seems > to do mostly fine (i.e. not worse than llvm's). What matters is > to schedule for long latencies and just after branch boundaries > where simplified model seems to do just fine. So we can keep the existing model for znver3 for GCC 11 ? > - some move instruction latencies does not reflect reality > (at least the published latencies by Agner Fog or AMD optimization > manual that themseleves does not agree with each otehr). > Adjusting tables however triggers regressions in ImageMagick and > parest, so I am still looking if there is easy fix for this and if > not, I will wait for next stage1 with these. > Interesting property is that reg-reg moves are a zero latency. > Since costs are officially relative to reg-reg move it makes it bit > hard to define here :) > - fmadd was optimized and it is now 4 cycles (was 5 and 6 cycles on > znver2 and znver1 respectively) like on Intel. However there is still > problem with extending the critical chain in matrix multiplication > loop. The difference seems to be that Intel implementation needs the > accumulator value to be ready only 1 cycle after the execution > started processing the multiplication. > > So there is still a performance regression on matmul and thus I am > keeping the logic to break critical chains. My observation is also same here. > > This first patch is no-op and it only copies the cost tables. I will adjust them one- > by-one for easier hunting of possible regressions. > > Honza > > 2021-03-15 Jan Hubicka <hubicka@ucw.cz> > > * config/i386/i386-options.c (processor_cost_table): Add znver3_cost. > * config/i386/x86-tune-costs.h (znver3_cost): New gobal variable; copy > of znver2_cost. > > diff --git a/gcc/config/i386/i386-options.c b/gcc/config/i386/i386-options.c > index e93935f6f2c..7865bc110a3 100644 > --- a/gcc/config/i386/i386-options.c > +++ b/gcc/config/i386/i386-options.c > @@ -743,7 +743,7 @@ static const struct processor_costs > *processor_cost_table[] = > &btver2_cost, > &znver1_cost, > &znver2_cost, > - &znver2_cost > + &znver3_cost > }; > > /* Guarantee that the array is aligned with enum processor_type. */ diff --git > a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h > index cc27c7911e3..e655e668c7a 100644 > --- a/gcc/config/i386/x86-tune-costs.h > +++ b/gcc/config/i386/x86-tune-costs.h > @@ -1688,6 +1688,140 @@ struct processor_costs znver2_cost = { > "16", /* Func alignment. */ > }; > > +struct processor_costs znver3_cost = { > + { > + /* Start of register allocator costs. integer->integer move cost is > +2. */ > + > + /* reg-reg moves are done by renaming and thus they are even cheaper than > + 1 cycle. Because reg-reg move cost is 2 and following tables correspond > + to doubles of latencies, we do not model this correctly. It does not > + seem to make practical difference to bump prices up even more. */ > + 6, /* cost for loading QImode using > + movzbl. */ > + {6, 6, 6}, /* cost of loading integer registers > + in QImode, HImode and SImode. > + Relative to reg-reg move (2). */ > + {8, 8, 8}, /* cost of storing integer > + registers. */ > + 2, /* cost of reg,reg fld/fst. */ > + {6, 6, 16}, /* cost of loading fp registers > + in SFmode, DFmode and XFmode. */ > + {8, 8, 16}, /* cost of storing fp registers > + in SFmode, DFmode and XFmode. */ > + 2, /* cost of moving MMX register. */ > + {6, 6}, /* cost of loading MMX registers > + in SImode and DImode. */ > + {8, 8}, /* cost of storing MMX registers > + in SImode and DImode. */ > + 2, 2, 3, /* cost of moving XMM,YMM,ZMM > + register. */ > + {6, 6, 6, 6, 12}, /* cost of loading SSE registers > + in 32,64,128,256 and 512-bit. */ > + {8, 8, 8, 8, 16}, /* cost of storing SSE registers > + in 32,64,128,256 and 512-bit. */ > + 6, 6, /* SSE->integer and integer->SSE > + moves. */ > + 8, 8, /* mask->integer and integer->mask moves */ > + {6, 6, 6}, /* cost of loading mask register > + in QImode, HImode, SImode. */ > + {8, 8, 8}, /* cost if storing mask register > + in QImode, HImode, SImode. */ > + 2, /* cost of moving mask register. */ > + /* End of register allocator costs. */ }, > + > + COSTS_N_INSNS (1), /* cost of an add instruction. */ > + COSTS_N_INSNS (1), /* cost of a lea instruction. */ > + COSTS_N_INSNS (1), /* variable shift costs. */ > + COSTS_N_INSNS (1), /* constant shift costs. */ > + {COSTS_N_INSNS (3), /* cost of starting multiply for QI. */ > + COSTS_N_INSNS (3), /* HI. */ > + COSTS_N_INSNS (3), /* SI. */ > + COSTS_N_INSNS (3), /* DI. */ > + COSTS_N_INSNS (3)}, /* other. */ > + 0, /* cost of multiply per each bit > + set. */ > + /* Depending on parameters, idiv can get faster on ryzen. This is upper > + bound. */ > + {COSTS_N_INSNS (16), /* cost of a divide/mod for QI. */ > + COSTS_N_INSNS (22), /* HI. */ > + COSTS_N_INSNS (30), /* SI. */ > + COSTS_N_INSNS (45), /* DI. */ > + COSTS_N_INSNS (45)}, /* other. */ > + COSTS_N_INSNS (1), /* cost of movsx. */ > + COSTS_N_INSNS (1), /* cost of movzx. */ > + 8, /* "large" insn. */ > + 9, /* MOVE_RATIO. */ > + 6, /* CLEAR_RATIO */ > + {6, 6, 6}, /* cost of loading integer registers > + in QImode, HImode and SImode. > + Relative to reg-reg move (2). */ > + {8, 8, 8}, /* cost of storing integer > + registers. */ > + {6, 6, 6, 6, 12}, /* cost of loading SSE registers > + in 32bit, 64bit, 128bit, 256bit and 512bit */ > + {8, 8, 8, 8, 16}, /* cost of storing SSE register > + in 32bit, 64bit, 128bit, 256bit and 512bit */ > + {6, 6, 6, 6, 12}, /* cost of unaligned loads. */ > + {8, 8, 8, 8, 16}, /* cost of unaligned stores. */ > + 2, 2, 3, /* cost of moving XMM,YMM,ZMM > + register. */ > + 6, /* cost of moving SSE register to integer. */ > + /* VGATHERDPD is 23 uops and throughput is 9, VGATHERDPD is 35 uops, > + throughput 12. Approx 9 uops do not depend on vector size and every load > + is 7 uops. */ > + 18, 8, /* Gather load static, per_elt. */ > + 18, 10, /* Gather store static, per_elt. */ > + 32, /* size of l1 cache. */ > + 512, /* size of l2 cache. */ > + 64, /* size of prefetch block. */ > + /* New AMD processors never drop prefetches; if they cannot be performed > + immediately, they are queued. We set number of simultaneous prefetches > + to a large constant to reflect this (it probably is not a good idea not > + to limit number of prefetches at all, as their execution also takes some > + time). */ > + 100, /* number of parallel prefetches. */ > + 3, /* Branch cost. */ > + COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */ > + COSTS_N_INSNS (5), /* cost of FMUL instruction. */ > + /* Latency of fdiv is 8-15. */ > + COSTS_N_INSNS (15), /* cost of FDIV instruction. */ > + COSTS_N_INSNS (1), /* cost of FABS instruction. */ > + COSTS_N_INSNS (1), /* cost of FCHS instruction. */ > + /* Latency of fsqrt is 4-10. */ > + COSTS_N_INSNS (10), /* cost of FSQRT instruction. */ > + > + COSTS_N_INSNS (1), /* cost of cheap SSE instruction. */ > + COSTS_N_INSNS (3), /* cost of ADDSS/SD SUBSS/SD insns. */ > + COSTS_N_INSNS (3), /* cost of MULSS instruction. */ > + COSTS_N_INSNS (3), /* cost of MULSD instruction. */ > + COSTS_N_INSNS (5), /* cost of FMA SS instruction. */ > + COSTS_N_INSNS (5), /* cost of FMA SD instruction. */ > + COSTS_N_INSNS (10), /* cost of DIVSS instruction. */ > + /* 9-13. */ > + COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ > + COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ > + COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ > + /* Zen can execute 4 integer operations per cycle. FP operations > + take 3 cycles and it can execute 2 integer additions and 2 > + multiplications thus reassociation may make sense up to with of 6. > + SPEC2k6 bencharks suggests > + that 4 works better than 6 probably due to register pressure. > + > + Integer vector operations are taken by FP unit and execute 3 vector > + plus/minus operations per cycle but only one multiply. This is adjusted > + in ix86_reassociation_width. */ > + 4, 4, 3, 6, /* reassoc int, fp, vec_int, vec_fp. */ > + znver2_memcpy, > + znver2_memset, > + COSTS_N_INSNS (4), /* cond_taken_branch_cost. */ > + COSTS_N_INSNS (2), /* cond_not_taken_branch_cost. */ > + "16", /* Loop alignment. */ > + "16", /* Jump alignment. */ > + "0:0:8", /* Label alignment. */ > + "16", /* Func alignment. */ > +}; > + > /* skylake_cost should produce code tuned for Skylake familly of CPUs. */ > static stringop_algs skylake_memcpy[2] = { > {libcall, {{1024, rep_prefix_4_byte, true}, {-1, libcall, false}}}, Regards, Venkat.
On 3/22/21 9:43 AM, Kumar, Venkataramanan via Gcc-patches wrote: > [AMD Official Use Only - Internal Distribution Only] Just brief note about this. You CC a public mailing list, thus the disclaimer does not make much sense. Please take a look here: https://gcc.gnu.org/lists.html#policies > > Hi Honza, > > Thank you for working on this. > >> -----Original Message----- >> From: Gcc-patches <gcc-patches-bounces@gcc.gnu.org> On Behalf Of Jan >> Hubicka >> Sent: Monday, March 15, 2021 3:33 PM >> To: gcc-patches@gcc.gnu.org; mjambor@suse.cz >> Subject: znver3 tuning part 1 >> >> [CAUTION: External Email] >> >> Hi, >> I plan to commit some retuning of znver3 codegen that is based on real >> hardware benchmarks. It turns out that there are not too many changes >> necessary sinze Zen3 is quite smooth upgrade to Zen2. In summary: >> >> - some instructions (like idiv) have shorter latencies. Adjusting >> costs reduces code size a bit but seems within noise in benchmark >> (since our cost calculation is quite off anyway because it does not >> account register pressure and parallelism that does make huge >> difference here) >> - gather instructions are still microcoded but a lot faster than in >> znver1/znver2 and it turns out they are now beneficial for few tsmc >> benchmarks, so I plan to enable them. > > Can we get a copy of this benchmark to try ? Sure: https://github.com/UoB-HPC/TSVC_2 Cheers, Martin > we need to check on bigger benchmarks like SPEC also. > >> >> It seems we missed revisiting this for znver2 tuning. >> I think even for znver2 it may make sense to re-enable them, so I >> will benchmark this as well. >> - memcpy/memset expansion seems to work same way as for znver2, >> so I am keeping same changes. >> - instruction scheduler is already modified in trunk to some degree >> reflecting new units. Problem with instruction scheduling is that >> it treats zen as in-order CPU and is unlikely going to fill all >> execution resources this way. >> We may want to try to model the out-of-order nature similar way as >> LLVM does, but at the other hand the current scheduling logic seems >> to do mostly fine (i.e. not worse than llvm's). What matters is >> to schedule for long latencies and just after branch boundaries >> where simplified model seems to do just fine. > > So we can keep the existing model for znver3 for GCC 11 ? > >> - some move instruction latencies does not reflect reality >> (at least the published latencies by Agner Fog or AMD optimization >> manual that themseleves does not agree with each otehr). >> Adjusting tables however triggers regressions in ImageMagick and >> parest, so I am still looking if there is easy fix for this and if >> not, I will wait for next stage1 with these. >> Interesting property is that reg-reg moves are a zero latency. >> Since costs are officially relative to reg-reg move it makes it bit >> hard to define here :) >> - fmadd was optimized and it is now 4 cycles (was 5 and 6 cycles on >> znver2 and znver1 respectively) like on Intel. However there is still >> problem with extending the critical chain in matrix multiplication >> loop. The difference seems to be that Intel implementation needs the >> accumulator value to be ready only 1 cycle after the execution >> started processing the multiplication. >> >> So there is still a performance regression on matmul and thus I am >> keeping the logic to break critical chains. > > My observation is also same here. > >> >> This first patch is no-op and it only copies the cost tables. I will adjust them one- >> by-one for easier hunting of possible regressions. >> >> Honza >> >> 2021-03-15 Jan Hubicka <hubicka@ucw.cz> >> >> * config/i386/i386-options.c (processor_cost_table): Add znver3_cost. >> * config/i386/x86-tune-costs.h (znver3_cost): New gobal variable; copy >> of znver2_cost. >> >> diff --git a/gcc/config/i386/i386-options.c b/gcc/config/i386/i386-options.c >> index e93935f6f2c..7865bc110a3 100644 >> --- a/gcc/config/i386/i386-options.c >> +++ b/gcc/config/i386/i386-options.c >> @@ -743,7 +743,7 @@ static const struct processor_costs >> *processor_cost_table[] = >> &btver2_cost, >> &znver1_cost, >> &znver2_cost, >> - &znver2_cost >> + &znver3_cost >> }; >> >> /* Guarantee that the array is aligned with enum processor_type. */ diff --git >> a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h >> index cc27c7911e3..e655e668c7a 100644 >> --- a/gcc/config/i386/x86-tune-costs.h >> +++ b/gcc/config/i386/x86-tune-costs.h >> @@ -1688,6 +1688,140 @@ struct processor_costs znver2_cost = { >> "16", /* Func alignment. */ >> }; >> >> +struct processor_costs znver3_cost = { >> + { >> + /* Start of register allocator costs. integer->integer move cost is >> +2. */ >> + >> + /* reg-reg moves are done by renaming and thus they are even cheaper than >> + 1 cycle. Because reg-reg move cost is 2 and following tables correspond >> + to doubles of latencies, we do not model this correctly. It does not >> + seem to make practical difference to bump prices up even more. */ >> + 6, /* cost for loading QImode using >> + movzbl. */ >> + {6, 6, 6}, /* cost of loading integer registers >> + in QImode, HImode and SImode. >> + Relative to reg-reg move (2). */ >> + {8, 8, 8}, /* cost of storing integer >> + registers. */ >> + 2, /* cost of reg,reg fld/fst. */ >> + {6, 6, 16}, /* cost of loading fp registers >> + in SFmode, DFmode and XFmode. */ >> + {8, 8, 16}, /* cost of storing fp registers >> + in SFmode, DFmode and XFmode. */ >> + 2, /* cost of moving MMX register. */ >> + {6, 6}, /* cost of loading MMX registers >> + in SImode and DImode. */ >> + {8, 8}, /* cost of storing MMX registers >> + in SImode and DImode. */ >> + 2, 2, 3, /* cost of moving XMM,YMM,ZMM >> + register. */ >> + {6, 6, 6, 6, 12}, /* cost of loading SSE registers >> + in 32,64,128,256 and 512-bit. */ >> + {8, 8, 8, 8, 16}, /* cost of storing SSE registers >> + in 32,64,128,256 and 512-bit. */ >> + 6, 6, /* SSE->integer and integer->SSE >> + moves. */ >> + 8, 8, /* mask->integer and integer->mask moves */ >> + {6, 6, 6}, /* cost of loading mask register >> + in QImode, HImode, SImode. */ >> + {8, 8, 8}, /* cost if storing mask register >> + in QImode, HImode, SImode. */ >> + 2, /* cost of moving mask register. */ >> + /* End of register allocator costs. */ }, >> + >> + COSTS_N_INSNS (1), /* cost of an add instruction. */ >> + COSTS_N_INSNS (1), /* cost of a lea instruction. */ >> + COSTS_N_INSNS (1), /* variable shift costs. */ >> + COSTS_N_INSNS (1), /* constant shift costs. */ >> + {COSTS_N_INSNS (3), /* cost of starting multiply for QI. */ >> + COSTS_N_INSNS (3), /* HI. */ >> + COSTS_N_INSNS (3), /* SI. */ >> + COSTS_N_INSNS (3), /* DI. */ >> + COSTS_N_INSNS (3)}, /* other. */ >> + 0, /* cost of multiply per each bit >> + set. */ >> + /* Depending on parameters, idiv can get faster on ryzen. This is upper >> + bound. */ >> + {COSTS_N_INSNS (16), /* cost of a divide/mod for QI. */ >> + COSTS_N_INSNS (22), /* HI. */ >> + COSTS_N_INSNS (30), /* SI. */ >> + COSTS_N_INSNS (45), /* DI. */ >> + COSTS_N_INSNS (45)}, /* other. */ >> + COSTS_N_INSNS (1), /* cost of movsx. */ >> + COSTS_N_INSNS (1), /* cost of movzx. */ >> + 8, /* "large" insn. */ >> + 9, /* MOVE_RATIO. */ >> + 6, /* CLEAR_RATIO */ >> + {6, 6, 6}, /* cost of loading integer registers >> + in QImode, HImode and SImode. >> + Relative to reg-reg move (2). */ >> + {8, 8, 8}, /* cost of storing integer >> + registers. */ >> + {6, 6, 6, 6, 12}, /* cost of loading SSE registers >> + in 32bit, 64bit, 128bit, 256bit and 512bit */ >> + {8, 8, 8, 8, 16}, /* cost of storing SSE register >> + in 32bit, 64bit, 128bit, 256bit and 512bit */ >> + {6, 6, 6, 6, 12}, /* cost of unaligned loads. */ >> + {8, 8, 8, 8, 16}, /* cost of unaligned stores. */ >> + 2, 2, 3, /* cost of moving XMM,YMM,ZMM >> + register. */ >> + 6, /* cost of moving SSE register to integer. */ >> + /* VGATHERDPD is 23 uops and throughput is 9, VGATHERDPD is 35 uops, >> + throughput 12. Approx 9 uops do not depend on vector size and every load >> + is 7 uops. */ >> + 18, 8, /* Gather load static, per_elt. */ >> + 18, 10, /* Gather store static, per_elt. */ >> + 32, /* size of l1 cache. */ >> + 512, /* size of l2 cache. */ >> + 64, /* size of prefetch block. */ >> + /* New AMD processors never drop prefetches; if they cannot be performed >> + immediately, they are queued. We set number of simultaneous prefetches >> + to a large constant to reflect this (it probably is not a good idea not >> + to limit number of prefetches at all, as their execution also takes some >> + time). */ >> + 100, /* number of parallel prefetches. */ >> + 3, /* Branch cost. */ >> + COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */ >> + COSTS_N_INSNS (5), /* cost of FMUL instruction. */ >> + /* Latency of fdiv is 8-15. */ >> + COSTS_N_INSNS (15), /* cost of FDIV instruction. */ >> + COSTS_N_INSNS (1), /* cost of FABS instruction. */ >> + COSTS_N_INSNS (1), /* cost of FCHS instruction. */ >> + /* Latency of fsqrt is 4-10. */ >> + COSTS_N_INSNS (10), /* cost of FSQRT instruction. */ >> + >> + COSTS_N_INSNS (1), /* cost of cheap SSE instruction. */ >> + COSTS_N_INSNS (3), /* cost of ADDSS/SD SUBSS/SD insns. */ >> + COSTS_N_INSNS (3), /* cost of MULSS instruction. */ >> + COSTS_N_INSNS (3), /* cost of MULSD instruction. */ >> + COSTS_N_INSNS (5), /* cost of FMA SS instruction. */ >> + COSTS_N_INSNS (5), /* cost of FMA SD instruction. */ >> + COSTS_N_INSNS (10), /* cost of DIVSS instruction. */ >> + /* 9-13. */ >> + COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ >> + COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ >> + COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ >> + /* Zen can execute 4 integer operations per cycle. FP operations >> + take 3 cycles and it can execute 2 integer additions and 2 >> + multiplications thus reassociation may make sense up to with of 6. >> + SPEC2k6 bencharks suggests >> + that 4 works better than 6 probably due to register pressure. >> + >> + Integer vector operations are taken by FP unit and execute 3 vector >> + plus/minus operations per cycle but only one multiply. This is adjusted >> + in ix86_reassociation_width. */ >> + 4, 4, 3, 6, /* reassoc int, fp, vec_int, vec_fp. */ >> + znver2_memcpy, >> + znver2_memset, >> + COSTS_N_INSNS (4), /* cond_taken_branch_cost. */ >> + COSTS_N_INSNS (2), /* cond_not_taken_branch_cost. */ >> + "16", /* Loop alignment. */ >> + "16", /* Jump alignment. */ >> + "0:0:8", /* Label alignment. */ >> + "16", /* Func alignment. */ >> +}; >> + >> /* skylake_cost should produce code tuned for Skylake familly of CPUs. */ >> static stringop_algs skylake_memcpy[2] = { >> {libcall, {{1024, rep_prefix_4_byte, true}, {-1, libcall, false}}}, > > > Regards, > Venkat. >
> > Hi, > > I plan to commit some retuning of znver3 codegen that is based on real > > hardware benchmarks. It turns out that there are not too many changes > > necessary sinze Zen3 is quite smooth upgrade to Zen2. In summary: > > > > - some instructions (like idiv) have shorter latencies. Adjusting > > costs reduces code size a bit but seems within noise in benchmark > > (since our cost calculation is quite off anyway because it does not > > account register pressure and parallelism that does make huge > > difference here) > > - gather instructions are still microcoded but a lot faster than in > > znver1/znver2 and it turns out they are now beneficial for few tsmc > > benchmarks, so I plan to enable them. > > Can we get a copy of this benchmark to try ? > we need to check on bigger benchmarks like SPEC also. Yes, I am also running specs. However for basic instruction selection tuning smaller benchmarks are doing quite well. In general if there are relatively natural loops where gather helps, i think we should enable it and try to fix possible regressions (I did not see one in spec runs, but I plan to do more benhcmarking this week). I did some work on TSVC mostly because zen3 seems very smooth update to zen2 for instruction selection (which is already happy with almost everything especially for scalar code) and vectorizer costs seems to be place where we seem to have most room for improvement. I briefly analyzed all tsvc kernels where we regress compared to clang, aocc and icc. You can search tsvc in bugzilla. Richard also wrote some observations there. These are related to missing features rather than cost model however. One problem of tsvc is that it is FP only. I hacked it for integer but it would be nice to have someting else as well. > > > > > It seems we missed revisiting this for znver2 tuning. > > I think even for znver2 it may make sense to re-enable them, so I > > will benchmark this as well. > > - memcpy/memset expansion seems to work same way as for znver2, > > so I am keeping same changes. > > - instruction scheduler is already modified in trunk to some degree > > reflecting new units. Problem with instruction scheduling is that > > it treats zen as in-order CPU and is unlikely going to fill all > > execution resources this way. > > We may want to try to model the out-of-order nature similar way as > > LLVM does, but at the other hand the current scheduling logic seems > > to do mostly fine (i.e. not worse than llvm's). What matters is > > to schedule for long latencies and just after branch boundaries > > where simplified model seems to do just fine. > > So we can keep the existing model for znver3 for GCC 11 ? I think so - I experimented with making the model bit more precise and it does not seem to add any performance improvements and makes the automaton a lot bigger. The existing model already handles the updated zen3 latencies... I think the only possible iprovment here would be to start modelling explicitly the out of order nature but even then I am not sure how much benefits that can bring (given that we are limited to relatively small basic blocks and do not have a lot of information needed to model the execution precisely). Do you have some options on this? Honza
On Mon, Mar 22, 2021 at 12:02 PM Jan Hubicka <hubicka@ucw.cz> wrote: > > > > Hi, > > > I plan to commit some retuning of znver3 codegen that is based on real > > > hardware benchmarks. It turns out that there are not too many changes > > > necessary sinze Zen3 is quite smooth upgrade to Zen2. In summary: > > > > > > - some instructions (like idiv) have shorter latencies. Adjusting > > > costs reduces code size a bit but seems within noise in benchmark > > > (since our cost calculation is quite off anyway because it does not > > > account register pressure and parallelism that does make huge > > > difference here) > > > - gather instructions are still microcoded but a lot faster than in > > > znver1/znver2 and it turns out they are now beneficial for few tsmc > > > benchmarks, so I plan to enable them. > > > > Can we get a copy of this benchmark to try ? > > we need to check on bigger benchmarks like SPEC also. > > Yes, I am also running specs. However for basic instruction selection > tuning smaller benchmarks are doing quite well. In general if there are > relatively natural loops where gather helps, i think we should enable it > and try to fix possible regressions (I did not see one in spec runs, but > I plan to do more benhcmarking this week). > > I did some work on TSVC mostly because zen3 seems very smooth update to > zen2 for instruction selection (which is already happy with almost > everything especially for scalar code) and vectorizer costs seems to be > place where we seem to have most room for improvement. > > I briefly analyzed all tsvc kernels where we regress compared to clang, > aocc and icc. You can search tsvc in bugzilla. Richard also wrote some > observations there. These are related to missing features rather than > cost model however. > > One problem of tsvc is that it is FP only. I hacked it for integer but > it would be nice to have someting else as well. > > > > > > > > It seems we missed revisiting this for znver2 tuning. > > > I think even for znver2 it may make sense to re-enable them, so I > > > will benchmark this as well. > > > - memcpy/memset expansion seems to work same way as for znver2, > > > so I am keeping same changes. > > > - instruction scheduler is already modified in trunk to some degree > > > reflecting new units. Problem with instruction scheduling is that > > > it treats zen as in-order CPU and is unlikely going to fill all > > > execution resources this way. > > > We may want to try to model the out-of-order nature similar way as > > > LLVM does, but at the other hand the current scheduling logic seems > > > to do mostly fine (i.e. not worse than llvm's). What matters is > > > to schedule for long latencies and just after branch boundaries > > > where simplified model seems to do just fine. > > > > So we can keep the existing model for znver3 for GCC 11 ? > > I think so - I experimented with making the model bit more precise and > it does not seem to add any performance improvements and makes the > automaton a lot bigger. The existing model already handles the updated > zen3 latencies... > > I think the only possible iprovment here would be to start modelling > explicitly the out of order nature but even then I am not sure how much > benefits that can bring (given that we are limited to relatively small > basic blocks and do not have a lot of information needed to model the > execution precisely). Do you have some options on this? I think it makes sense to model instruction fetch quite precisely (including, or rather either/or fetch from the uop cache) up to where OOO starts. From there on backwards only very long latency insns and of course insn dependences should be a factor to maximise issue width per fetch block. Not sure if it makes sense to model the uop cache at all or whehter we should switch between L1 fetch and uop cache assumption based on loop depth? That said, for loops scheduling is somewhat moot but for cold (in terms of the OOO window size) serial code it makes sense to optimize for uop issue. I also note that this seems to work out quite well with the existing automata - if only as side-effect. Richard. > Honza
[AMD Public Use] Hi Honza, > -----Original Message----- > From: Jan Hubicka <hubicka@ucw.cz> > Sent: Monday, March 22, 2021 4:31 PM > To: Kumar, Venkataramanan <Venkataramanan.Kumar@amd.com> > Cc: gcc-patches@gcc.gnu.org; mjambor@suse.cz > Subject: Re: znver3 tuning part 1 > > [CAUTION: External Email] > > > > Hi, > > > I plan to commit some retuning of znver3 codegen that is based on > > > real hardware benchmarks. It turns out that there are not too many > > > changes necessary sinze Zen3 is quite smooth upgrade to Zen2. In summary: > > > > > > - some instructions (like idiv) have shorter latencies. Adjusting > > > costs reduces code size a bit but seems within noise in benchmark > > > (since our cost calculation is quite off anyway because it does not > > > account register pressure and parallelism that does make huge > > > difference here) > > > - gather instructions are still microcoded but a lot faster than in > > > znver1/znver2 and it turns out they are now beneficial for few tsmc > > > benchmarks, so I plan to enable them. > > > > Can we get a copy of this benchmark to try ? > > we need to check on bigger benchmarks like SPEC also. > > Yes, I am also running specs. However for basic instruction selection tuning > smaller benchmarks are doing quite well. In general if there are relatively > natural loops where gather helps, i think we should enable it and try to fix > possible regressions (I did not see one in spec runs, but I plan to do more > benhcmarking this week). Okay Thank you. > > I did some work on TSVC mostly because zen3 seems very smooth update to > zen2 for instruction selection (which is already happy with almost everything > especially for scalar code) and vectorizer costs seems to be place where we > seem to have most room for improvement. > > I briefly analyzed all tsvc kernels where we regress compared to clang, aocc and > icc. You can search tsvc in bugzilla. Richard also wrote some observations there. > These are related to missing features rather than cost model however. > > One problem of tsvc is that it is FP only. I hacked it for integer but it would be > nice to have someting else as well. > > > > > > > > It seems we missed revisiting this for znver2 tuning. > > > I think even for znver2 it may make sense to re-enable them, so I > > > will benchmark this as well. > > > - memcpy/memset expansion seems to work same way as for znver2, > > > so I am keeping same changes. > > > - instruction scheduler is already modified in trunk to some degree > > > reflecting new units. Problem with instruction scheduling is that > > > it treats zen as in-order CPU and is unlikely going to fill all > > > execution resources this way. > > > We may want to try to model the out-of-order nature similar way as > > > LLVM does, but at the other hand the current scheduling logic seems > > > to do mostly fine (i.e. not worse than llvm's). What matters is > > > to schedule for long latencies and just after branch boundaries > > > where simplified model seems to do just fine. > > > > So we can keep the existing model for znver3 for GCC 11 ? > > I think so - I experimented with making the model bit more precise and it does > not seem to add any performance improvements and makes the automaton a > lot bigger. The existing model already handles the updated > zen3 latencies... > > I think the only possible iprovment here would be to start modelling explicitly the > out of order nature but even then I am not sure how much benefits that can > bring (given that we are limited to relatively small basic blocks and do not have a > lot of information needed to model the execution precisely). Do you have some > options on this? Given that basic blocks are small and hardware itself reorders the instructions, I don't think precisely modelling the scheduler will give much benefit. > > Honza Regards, Venkat.
> > 2021-03-15 Jan Hubicka <hubicka@ucw.cz> > > * config/i386/i386-options.c (processor_cost_table): Add znver3_cost. > * config/i386/x86-tune-costs.h (znver3_cost): New gobal variable; copy > of znver2_cost. I have backported the patch to gcc10 branch as g:aa99212489545c6c970a8f91b3d37ea6466cb988. Honza
On 3/31/21 1:08 PM, Jan Hubicka wrote: >> >> 2021-03-15 Jan Hubicka <hubicka@ucw.cz> >> >> * config/i386/i386-options.c (processor_cost_table): Add znver3_cost. >> * config/i386/x86-tune-costs.h (znver3_cost): New gobal variable; copy >> of znver2_cost. > > I have backported the patch to gcc10 branch as > g:aa99212489545c6c970a8f91b3d37ea6466cb988. > > Honza > Looking at the backport, it has likely enabled PR99753. Please consider backporting 4f00c4d40a539360938607561460904663c64cda. Cheers, Martin
> On 3/31/21 1:08 PM, Jan Hubicka wrote: > > > > > > 2021-03-15 Jan Hubicka <hubicka@ucw.cz> > > > > > > * config/i386/i386-options.c (processor_cost_table): Add znver3_cost. > > > * config/i386/x86-tune-costs.h (znver3_cost): New gobal variable; copy > > > of znver2_cost. > > > > I have backported the patch to gcc10 branch as > > g:aa99212489545c6c970a8f91b3d37ea6466cb988. > > > > Honza > > > > Looking at the backport, it has likely enabled PR99753. > Please consider backporting 4f00c4d40a539360938607561460904663c64cda. Not this patch but backport of the original Venkat's change indeed triggers PR99763. I am now testing backport if this revision. Thanks, Honza > > Cheers, > Martin
diff --git a/gcc/config/i386/i386-options.c b/gcc/config/i386/i386-options.c index e93935f6f2c..7865bc110a3 100644 --- a/gcc/config/i386/i386-options.c +++ b/gcc/config/i386/i386-options.c @@ -743,7 +743,7 @@ static const struct processor_costs *processor_cost_table[] = &btver2_cost, &znver1_cost, &znver2_cost, - &znver2_cost + &znver3_cost }; /* Guarantee that the array is aligned with enum processor_type. */ diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h index cc27c7911e3..e655e668c7a 100644 --- a/gcc/config/i386/x86-tune-costs.h +++ b/gcc/config/i386/x86-tune-costs.h @@ -1688,6 +1688,140 @@ struct processor_costs znver2_cost = { "16", /* Func alignment. */ }; +struct processor_costs znver3_cost = { + { + /* Start of register allocator costs. integer->integer move cost is 2. */ + + /* reg-reg moves are done by renaming and thus they are even cheaper than + 1 cycle. Because reg-reg move cost is 2 and following tables correspond + to doubles of latencies, we do not model this correctly. It does not + seem to make practical difference to bump prices up even more. */ + 6, /* cost for loading QImode using + movzbl. */ + {6, 6, 6}, /* cost of loading integer registers + in QImode, HImode and SImode. + Relative to reg-reg move (2). */ + {8, 8, 8}, /* cost of storing integer + registers. */ + 2, /* cost of reg,reg fld/fst. */ + {6, 6, 16}, /* cost of loading fp registers + in SFmode, DFmode and XFmode. */ + {8, 8, 16}, /* cost of storing fp registers + in SFmode, DFmode and XFmode. */ + 2, /* cost of moving MMX register. */ + {6, 6}, /* cost of loading MMX registers + in SImode and DImode. */ + {8, 8}, /* cost of storing MMX registers + in SImode and DImode. */ + 2, 2, 3, /* cost of moving XMM,YMM,ZMM + register. */ + {6, 6, 6, 6, 12}, /* cost of loading SSE registers + in 32,64,128,256 and 512-bit. */ + {8, 8, 8, 8, 16}, /* cost of storing SSE registers + in 32,64,128,256 and 512-bit. */ + 6, 6, /* SSE->integer and integer->SSE + moves. */ + 8, 8, /* mask->integer and integer->mask moves */ + {6, 6, 6}, /* cost of loading mask register + in QImode, HImode, SImode. */ + {8, 8, 8}, /* cost if storing mask register + in QImode, HImode, SImode. */ + 2, /* cost of moving mask register. */ + /* End of register allocator costs. */ + }, + + COSTS_N_INSNS (1), /* cost of an add instruction. */ + COSTS_N_INSNS (1), /* cost of a lea instruction. */ + COSTS_N_INSNS (1), /* variable shift costs. */ + COSTS_N_INSNS (1), /* constant shift costs. */ + {COSTS_N_INSNS (3), /* cost of starting multiply for QI. */ + COSTS_N_INSNS (3), /* HI. */ + COSTS_N_INSNS (3), /* SI. */ + COSTS_N_INSNS (3), /* DI. */ + COSTS_N_INSNS (3)}, /* other. */ + 0, /* cost of multiply per each bit + set. */ + /* Depending on parameters, idiv can get faster on ryzen. This is upper + bound. */ + {COSTS_N_INSNS (16), /* cost of a divide/mod for QI. */ + COSTS_N_INSNS (22), /* HI. */ + COSTS_N_INSNS (30), /* SI. */ + COSTS_N_INSNS (45), /* DI. */ + COSTS_N_INSNS (45)}, /* other. */ + COSTS_N_INSNS (1), /* cost of movsx. */ + COSTS_N_INSNS (1), /* cost of movzx. */ + 8, /* "large" insn. */ + 9, /* MOVE_RATIO. */ + 6, /* CLEAR_RATIO */ + {6, 6, 6}, /* cost of loading integer registers + in QImode, HImode and SImode. + Relative to reg-reg move (2). */ + {8, 8, 8}, /* cost of storing integer + registers. */ + {6, 6, 6, 6, 12}, /* cost of loading SSE registers + in 32bit, 64bit, 128bit, 256bit and 512bit */ + {8, 8, 8, 8, 16}, /* cost of storing SSE register + in 32bit, 64bit, 128bit, 256bit and 512bit */ + {6, 6, 6, 6, 12}, /* cost of unaligned loads. */ + {8, 8, 8, 8, 16}, /* cost of unaligned stores. */ + 2, 2, 3, /* cost of moving XMM,YMM,ZMM + register. */ + 6, /* cost of moving SSE register to integer. */ + /* VGATHERDPD is 23 uops and throughput is 9, VGATHERDPD is 35 uops, + throughput 12. Approx 9 uops do not depend on vector size and every load + is 7 uops. */ + 18, 8, /* Gather load static, per_elt. */ + 18, 10, /* Gather store static, per_elt. */ + 32, /* size of l1 cache. */ + 512, /* size of l2 cache. */ + 64, /* size of prefetch block. */ + /* New AMD processors never drop prefetches; if they cannot be performed + immediately, they are queued. We set number of simultaneous prefetches + to a large constant to reflect this (it probably is not a good idea not + to limit number of prefetches at all, as their execution also takes some + time). */ + 100, /* number of parallel prefetches. */ + 3, /* Branch cost. */ + COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */ + COSTS_N_INSNS (5), /* cost of FMUL instruction. */ + /* Latency of fdiv is 8-15. */ + COSTS_N_INSNS (15), /* cost of FDIV instruction. */ + COSTS_N_INSNS (1), /* cost of FABS instruction. */ + COSTS_N_INSNS (1), /* cost of FCHS instruction. */ + /* Latency of fsqrt is 4-10. */ + COSTS_N_INSNS (10), /* cost of FSQRT instruction. */ + + COSTS_N_INSNS (1), /* cost of cheap SSE instruction. */ + COSTS_N_INSNS (3), /* cost of ADDSS/SD SUBSS/SD insns. */ + COSTS_N_INSNS (3), /* cost of MULSS instruction. */ + COSTS_N_INSNS (3), /* cost of MULSD instruction. */ + COSTS_N_INSNS (5), /* cost of FMA SS instruction. */ + COSTS_N_INSNS (5), /* cost of FMA SD instruction. */ + COSTS_N_INSNS (10), /* cost of DIVSS instruction. */ + /* 9-13. */ + COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ + COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ + COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ + /* Zen can execute 4 integer operations per cycle. FP operations + take 3 cycles and it can execute 2 integer additions and 2 + multiplications thus reassociation may make sense up to with of 6. + SPEC2k6 bencharks suggests + that 4 works better than 6 probably due to register pressure. + + Integer vector operations are taken by FP unit and execute 3 vector + plus/minus operations per cycle but only one multiply. This is adjusted + in ix86_reassociation_width. */ + 4, 4, 3, 6, /* reassoc int, fp, vec_int, vec_fp. */ + znver2_memcpy, + znver2_memset, + COSTS_N_INSNS (4), /* cond_taken_branch_cost. */ + COSTS_N_INSNS (2), /* cond_not_taken_branch_cost. */ + "16", /* Loop alignment. */ + "16", /* Jump alignment. */ + "0:0:8", /* Label alignment. */ + "16", /* Func alignment. */ +}; + /* skylake_cost should produce code tuned for Skylake familly of CPUs. */ static stringop_algs skylake_memcpy[2] = { {libcall, {{1024, rep_prefix_4_byte, true}, {-1, libcall, false}}},