1
Copyright 2000, 2001 Free Software Foundation, Inc.
3
This file is part of the GNU MP Library.
5
The GNU MP Library is free software; you can redistribute it and/or modify
6
it under the terms of the GNU Lesser General Public License as published by
7
the Free Software Foundation; either version 2.1 of the License, or (at your
8
option) any later version.
10
The GNU MP Library is distributed in the hope that it will be useful, but
11
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
12
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
13
License for more details.
15
You should have received a copy of the GNU Lesser General Public License
16
along with the GNU MP Library; see the file COPYING.LIB. If not, write to
17
the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
23
AMD K7 MPN SUBROUTINES
26
This directory contains code optimized for the AMD Athlon CPU.
28
The mmx subdirectory has routines using MMX instructions. All Athlons have
29
MMX, the separate directory is just so that configure can omit it if the
30
assembler doesn't support MMX.
36
Times for the loops, with all code and data in L1 cache.
41
mpn_copyi 0.75 or 1.0 \ varying with data alignment
42
mpn_copyd 0.75 or 1.0 /
44
mpn_divrem_1 17.0 integer part, 15.0 fractional part
51
mpn_addmul/submul_1 3.9
53
mpn_mul_basecase 4.42 cycles/crossproduct (approx)
54
mpn_sqr_basecase 2.3 cycles/crossproduct (approx)
55
or 4.55 cycles/triangleproduct (approx)
57
Prefetching of sources hasn't yet been tried.
63
cmov, MMX, 3DNow and some extensions to MMX and 3DNow are available.
65
Write-allocate L1 data cache means prefetching of destinations is unnecessary.
67
Floating point multiplications can be done in parallel with integer
68
multiplications, but there doesn't seem to be any way to make use of this.
70
Unsigned "mul"s can be issued every 3 cycles. This suggests 3 is a limit on
71
the speed of the multiplication routines. The documentation shows mul
72
executing in IEU0 (or maybe in IEU0 and IEU1 together), so it might be that,
73
to get near 3 cycles code has to be arranged so that nothing else is issued
74
to IEU0. A busy IEU0 could explain why some code takes 4 cycles and other
75
apparently equivalent code takes 5.
81
Unrolled loops are used to reduce looping overhead. The unrolling is
82
configurable up to 32 limbs/loop for most routines and up to 64 for some.
83
The K7 has 64k L1 code cache so quite big unrolling is allowable.
85
Computed jumps into the unrolling are used to handle sizes not a multiple of
86
the unrolling. An attractive feature of this is that times increase
87
smoothly with operand size, but it may be that some routines should just
88
have simple loops to finish up, especially when PIC adds between 2 and 16
91
Position independent code is implemented using a call to get %eip for the
92
computed jumps and a ret is always done, rather than an addl $4,%esp or a
93
popl, so the CPU return address branch prediction stack stays synchronised
94
with the actual stack in memory.
96
Branch prediction, in absence of any history, will guess forward jumps are
97
not taken and backward jumps are taken. Where possible it's arranged that
98
the less likely or less important case is under a taken forward jump.
104
Instructions in general code have been shown grouped if they can execute
105
together, which means up to three direct-path instructions which have no
106
successive dependencies. K7 always decodes three and has out-of-order
107
execution, but the groupings show what slots might be available and what
108
dependency chains exist.
110
When there's vector-path instructions an effort is made to get triplets of
111
direct-path instructions in between them, even if there's dependencies,
112
since this maximizes decoding throughput and might save a cycle or two if
113
decoding is the limiting factor.
120
divl 39 cycles back-to-back
122
loop 1 cycle vector (decl/jnz opens up one decode slot)
125
mull issue every 3 cycles, latency 4 cycles low word, 6 cycles high word
126
popl vector (use movl for more than one pop)
127
pushl direct, will pair with a load
128
shrdl %cl vector, 3 cycles, seems to be 3 decode too
129
xorl r,r false read dependency recognised
135
"AMD Athlon Processor X86 Code Optimization Guide", AMD publication number
136
22007, revision E, November 1999. Available on-line,
138
http://www.amd.com/products/cpg/athlon/techdocs/pdf/22007.pdf
140
"3DNow Technology Manual", AMD publication number 21928F/0-August 1999.
141
This describes the femms and prefetch instructions. Available on-line,
143
http://www.amd.com/K6/k6docs/pdf/21928.pdf
145
"AMD Extensions to the 3DNow and MMX Instruction Sets Manual", AMD
146
publication number 22466, revision B, August 1999. This describes
147
instructions added in the Athlon processor, such as pswapd and the extra
148
prefetch forms. Available on-line,
150
http://www.amd.com/products/cpg/athlon/techdocs/pdf/22466.pdf
152
"3DNow Instruction Porting Guide", AMD publication number 22621, revision B,
153
August 1999. This has some notes on general Athlon optimizations as well as
154
3DNow. Available on-line,
156
http://www.amd.com/products/cpg/athlon/techdocs/pdf/22621.pdf