diff options
Diffstat (limited to 'gmp-6.3.0/mpn/powerpc64/README')
| -rw-r--r-- | gmp-6.3.0/mpn/powerpc64/README | 166 |
1 files changed, 166 insertions, 0 deletions
diff --git a/gmp-6.3.0/mpn/powerpc64/README b/gmp-6.3.0/mpn/powerpc64/README new file mode 100644 index 0000000..50dd399 --- /dev/null +++ b/gmp-6.3.0/mpn/powerpc64/README @@ -0,0 +1,166 @@ +Copyright 1999-2001, 2003-2005 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + POWERPC-64 MPN SUBROUTINES + + +This directory contains mpn functions for 64-bit PowerPC chips. + + +CODE ORGANIZATION + + mpn/powerpc64 mode-neutral code + mpn/powerpc64/mode32 code for mode32 + mpn/powerpc64/mode64 code for mode64 + + +The mode32 and mode64 sub-directories contain code which is for use in the +respective chip mode, 32 or 64. The top-level directory is code that's +unaffected by the mode. + +The "adde" instruction is the main difference between mode32 and mode64. It +operates on either on a 32-bit or 64-bit quantity according to the chip mode. +Other instructions have an operand size in their opcode and hence don't vary. + + + +POWER3/PPC630 pipeline information: + +Decoding is 4-way + branch and issue is 8-way with some out-of-order +capability. + +Functional units: +LS1 - ld/st unit 1 +LS2 - ld/st unit 2 +FXU1 - integer unit 1, handles any simple integer instruction +FXU2 - integer unit 2, handles any simple integer instruction +FXU3 - integer unit 3, handles integer multiply and divide +FPU1 - floating-point unit 1 +FPU2 - floating-point unit 2 + +Memory: Any two memory operations can issue, but memory subsystem + can sustain just one store per cycle. No need for data + prefetch; the hardware has very sophisticated prefetch logic. +Simple integer: 2 operations (such as add, rl*) +Integer multiply: 1 operation every 9th cycle worst case; exact timing depends + on 2nd operand's most significant bit position (10 bits per + cycle). Multiply unit is not pipelined, only one multiply + operation in progress is allowed. +Integer divide: ? +Floating-point: Any plain 2 arithmetic instructions (such as fmul, fadd, and + fmadd), latency 4 cycles. +Floating-point divide: + ? +Floating-point square root: + ? + +POWER3/PPC630 best possible times for the main loops: +shift: 1.5 cycles limited by integer unit contention. + With 63 special loops, one for each shift count, we could + reduce the needed integer instructions to 2, which would + reduce the best possible time to 1 cycle. +add/sub: 1.5 cycles, limited by ld/st unit contention. +mul: 18 cycles (average) unless floating-point operations are used, + but that would only help for multiplies of perhaps 10 and more + limbs. +addmul/submul:Same situation as for mul. + + +POWER4/PPC970 and POWER5 pipeline information: + +This is a very odd pipeline, it is basically a VLIW masquerading as a plain +architecture. Its issue rules are not made public, and since it is so weird, +it is very hard to figure out any useful information from experimentation. +An example: + + A well-aligned loop with nop's take 3, 4, 6, 7, ... cycles. + 3 cycles for 0, 1, 2, 3, 4, 5, 6, 7 nop's + 4 cycles for 8, 9, 10, 11, 12, 13, 14, 15 nop's + 6 cycles for 16, 17, 18, 19, 20, 21, 22, 23 nop's + 7 cycles for 24, 25, 26, 27 nop's + 8 cycles for 28, 29, 30, 31 nop's + ... continues regularly + + +Functional units: +LS1 - ld/st unit 1 +LS2 - ld/st unit 2 +FXU1 - integer unit 1, handles any integer instruction +FXU2 - integer unit 2, handles any integer instruction +FPU1 - floating-point unit 1 +FPU2 - floating-point unit 2 + +While this is one integer unit less than POWER3/PPC630, the remaining units +are more powerful; here they handle multiply and divide. + +Memory: 2 ld/st. Stores go to the L2 cache, which can sustain just + one store per cycle. + L1 load latency: to gregs 3-4 cycles, to fregs 5-6 cycles. + Operations that modify the address register might be split + to use also an integer issue slot. +Simple integer: 2 operations every cycle, latency 2. +Integer multiply: 2 operations every 6th cycle, latency 7 cycles. +Integer divide: ? +Floating-point: Any plain 2 arithmetic instructions (such as fmul, fadd, and + fmadd), latency 6 cycles. +Floating-point divide: + ? +Floating-point square root: + ? + + +IDEAS + +*mul_1: Handling one limb using mulld/mulhdu and two limbs using floating- +point operations should give performance of about 20 cycles for 3 limbs, or 7 +cycles/limb. + +We should probably split the single-limb operand in 32-bit chunks, and the +multi-limb operand in 16-bit chunks, allowing us to accumulate well in fp +registers. + +Problem is to get 32-bit or 16-bit words to the fp registers. Only 64-bit fp +memops copies bits without fiddling with them. We might therefore need to +load to integer registers with zero extension, store as 64 bits into temp +space, and then load to fp regs. Alternatively, load directly to fp space +and add well-chosen constants to get cancellation. (Other part after given by +subsequent subtraction.) + +Possible code mix for load-via-intregs variant: + +lwz,std,lfd +fmadd,fmadd,fmul,fmul +fctidz,stfd,ld,fctidz,stfd,ld +add,adde +lwz,std,lfd +fmadd,fmadd,fmul,fmul +fctidz,stfd,ld,fctidz,stfd,ld +add,adde +srd,sld,add,adde,add,adde |