=================================================================== RCS file: /home/cvs/OpenXM_contrib/gmp/mpn/alpha/Attic/README,v retrieving revision 1.1.1.1 retrieving revision 1.1.1.2 diff -u -p -r1.1.1.1 -r1.1.1.2 --- OpenXM_contrib/gmp/mpn/alpha/Attic/README 2000/01/10 15:35:22 1.1.1.1 +++ OpenXM_contrib/gmp/mpn/alpha/Attic/README 2000/09/09 14:12:21 1.1.1.2 @@ -1,53 +1,224 @@ This directory contains mpn functions optimized for DEC Alpha processors. +ALPHA ASSEMBLY RULES AND REGULATIONS + +The `.prologue N' pseudo op marks the end of instruction that needs +special handling by unwinding. It also says whether $27 is really +needed for computing the gp. The `.mask M' pseudo op says which +registers are saved on the stack, and at what offset in the frame. + +Cray code is very very different... + + RELEVANT OPTIMIZATION ISSUES EV4 1. This chip has very limited store bandwidth. The on-chip L1 cache is -write-through, and a cache line is transfered from the store buffer to the -off-chip L2 in as much 15 cycles on most systems. This delay hurts -mpn_add_n, mpn_sub_n, mpn_lshift, and mpn_rshift. + write-through, and a cache line is transfered from the store buffer to + the off-chip L2 in as much 15 cycles on most systems. This delay hurts + mpn_add_n, mpn_sub_n, mpn_lshift, and mpn_rshift. 2. Pairing is possible between memory instructions and integer arithmetic -instructions. + instructions. -3. mulq and umulh is documented to have a latency of 23 cycles, but 2 of -these cycles are pipelined. Thus, multiply instructions can be issued at a -rate of one each 21nd cycle. +3. mulq and umulh are documented to have a latency of 23 cycles, but 2 of + these cycles are pipelined. Thus, multiply instructions can be issued at + a rate of one each 21st cycle. EV5 1. The memory bandwidth of this chip seems excellent, both for loads and -stores. Even when the working set is larger than the on-chip L1 and L2 -caches, the perfromance remain almost unaffected. + stores. Even when the working set is larger than the on-chip L1 and L2 + caches, the performance remain almost unaffected. -2. mulq has a measured latency of 13 cycles and an issue rate of 1 each 8th -cycle. umulh has a measured latency of 15 cycles and an issue rate of 1 -each 10th cycle. But the exact timing is somewhat confusing. +2. mulq has a latency of 12 cycles and an issue rate of 1 each 8th cycle. + umulh has a measured latency of 14 cycles and an issue rate of 1 each + 10th cycle. But the exact timing is somewhat confusing. 3. mpn_add_n. With 4-fold unrolling, we need 37 instructions, whereof 12 are memory operations. This will take at least - ceil(37/2) [dual issue] + 1 [taken branch] = 20 cycles + ceil(37/2) [dual issue] + 1 [taken branch] = 19 cycles We have 12 memory cycles, plus 4 after-store conflict cycles, or 16 data - cache cycles, which should be completely hidden in the 20 issue cycles. + cache cycles, which should be completely hidden in the 19 issue cycles. The computation is inherently serial, with these dependencies: + + ldq ldq + \ /\ + (or) addq | + |\ / \ | + | addq cmpult + \ | | + cmpult | + \ / + or + + I.e., 3 operations are needed between carry-in and carry-out, making 12 + cycles the absolute minimum for the 4 limbs. We could replace the `or' + with a cmoveq/cmovne, which could issue one cycle earlier that the `or', + but that might waste a cycle on EV4. The total depth remain unaffected, + since cmov has a latency of 2 cycles. + addq / \ addq cmpult - | | - cmpult | - \ / - or - I.e., there is a 4 cycle path for each limb, making 16 cycles the absolute - minimum. We could replace the `or' with a cmoveq/cmovne, which would save - a cycle on EV5, but that might waste a cycle on EV4. Also, cmov takes 2 - cycles. - addq - / \ - addq cmpult | \ cmpult -> cmovne -STATUS +Montgomery has a slightly different way of computing carry that requires one +less instruction, but has depth 4 (instead of the current 3). Since the +code is currently instruction issue bound, Montgomery's idea should save us +1/2 cycle per limb, or bring us down to a total of 17 cycles or 4.25 +cycles/limb. Unfortunately, this method will not be good for the EV6. +EV6 + +Here we have a really parallel pipeline, capable of issuing up to 4 integer +instructions per cycle. One integer multiply instruction can issue each +cycle. To get optimal speed, we need to pretend we are vectorizing the code, +i.e., minimize the iterative dependencies. + +There are two dependencies to watch out for. 1) Address arithmetic +dependencies, and 2) carry propagation dependencies. + +We can avoid serializing due to address arithmetic by unrolling the loop, so +that addresses don't depend heavily on an index variable. Avoiding +serializing because of carry propagation is trickier; the ultimate performance +of the code will be determined of the number of latency cycles it takes from +accepting carry-in to a vector point until we can generate carry-out. + +Most integer instructions can execute in either the L0, U0, L1, or U1 +pipelines. Shifts only execute in U0 and U1, and multiply only in U1. + +CMOV instructions split into two internal instructions, CMOV1 and CMOV2, but +the execute efficiently. But CMOV split the mapping process (see pg 2-26 in +cmpwrgd.pdf), suggesting the CMOV should always be placed as the last +instruction of an aligned 4 instruction block (?). + +Perhaps the most important issue is the latency between the L0/U0 and L1/U1 +clusters; a result obtained on either cluster has an extra cycle of latency +for consumers in the opposite cluster. Because of the dynamic nature of the +implementation, it is hard to predict where an instruction will execute. + +The shift loops need (per limb): + 1 load (Lx pipes) + 1 store (Lx pipes) + 2 shift (Ux pipes) + 1 iaddlog (Lx pipes, Ux pipes) +Obviously, since the pipes are very equally loaded, we should get 4 insn/cycle, or 1.25 cycles/limb. + +For mpn_add_n, we currently have + 2 load (Lx pipes) + 1 store (Lx pipes) + 5 iaddlog (Lx pipes, Ux pipes) + +Again, we have a perfect balance and will be limited by carry propagation +delays, currently three cycles. The superoptimizer indicates that ther +might be sequences that--using a final cmov--have a carry propagation delay +of just two. Montgomery's subtraction sequence could perhaps be used, by +complementing some operands. All in all, we should get down to 2 cycles +without much problems. + +For mpn_mul_1, we could do, just like for mpn_add_n: + not newlo,notnewlo + addq cylimb,newlo,newlo || cmpult cylimb,notnewlo,cyout + addq cyout,newhi,cylimb +and get 2-cycle carry propagation. The instructions needed will be + 1 ld (Lx pipes) + 1 st (Lx pipes) + 2 mul (U1 pipe) + 4 iaddlog (Lx pipes, Ux pipes) +issue1: addq not mul ld +issue2: cmpult addq mul st +Conclusion: no cluster delays and 2-cycle carry delays will give us 2 cycles/limb! + +Last, we have mpn_addmul_1. Almost certainly, we will get down to 3 +cycles/limb, which would be absolutely awesome. + +Old, perhaps obsolete addmul_1 dependency diagram (needs 175 columns wide screen): + + i + s + s i + u n + e s + d t + r + i u +l n c +i s t +v t i +e r o + u n +v c +a t t +l i y +u o p +e n e +s s s + issue + in + cycle + -1 ldq + / \ + 0 | \ + | \ + 1 | | + | | + 2 | | ldq + | | / \ + 3 | mulq | \ + | \ | \ + 4 umulh \ | | + | | | | + 5 | | | | ldq + | | | | / \ + 4calm 6 | | ldq | mulq | \ + | | / | \ | \ + 4casm 7 | | / umulh \ | | +6 | || | | | | + 3aal 8 | || | | | | ldq +7 | || | | | | / \ + 4calm 9 | || | | ldq | mulq | \ +9 | || | | / | \ | \ + 4casm 10 | || | | / umulh \ | | +9 | || | || | | | | + 3aal 11 | addq | || | | | | ldq +9 | // \ | || | | | | / \ + 4calm 12 \ cmpult addq<-cy | || | | ldq | mulq | \ +13 \ / // \ | || | | / | \ | \ + 4casm 13 addq cmpult stq | || | | / umulh \ | | +11 \ / | || | || | | | | + 3aal 14 addq | addq | || | | | | ldq +10 \ | // \ | || | | | | / \ + 4calm 15 cy ----> \ cmpult addq<-cy | || | | ldq | mulq | \ +13 \ / // \ | || | | / | \ | \ + 4casm 16 addq cmpult stq | || | | / umulh \ | | +11 \ / | || | || | | | | + 3aal 17 addq | addq | || | | | | +10 \ | // \ | || | | | | + 4calm 18 cy ----> \ cmpult addq<-cy | || | | ldq | mulq +13 \ / // \ | || | | / | \ + 4casm 19 addq cmpult stq | || | | / umulh \ +11 \ / | || | || | | + 3aal 20 addq | addq | || | | +10 \ | // \ | || | | + 4calm 21 cy ----> \ cmpult addq<-cy | || | | ldq + \ / // \ | || | | / + 22 addq cmpult stq | || | | / + \ / | || | || + 23 addq | addq | || + \ | // \ | || + 24 cy ----> \ cmpult addq<-cy | || + \ / // \ | || + 25 addq cmpult stq | || + \ / | || + 26 addq | addq + \ | // \ + 27 cy ----> \ cmpult addq<-cy + \ / // \ + 28 addq cmpult stq + \ / +As many as 6 consecutive points will be under execution simultaneously, or if we addq +schedule loads even further away, maybe 7 or 8. But the number of live quantities \ +is reasonable, and can easily be satisfied. cy ---->