1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
|
Copyright 2000, 2001 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 the GNU Lesser General Public License as published by
the Free Software Foundation; either version 2.1 of the License, or (at your
option) any later version.
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 Lesser General Public
License for more details.
You should have received a copy of the GNU Lesser General Public License
along with the GNU MP Library; see the file COPYING.LIB. If not, write to
the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
02111-1307, USA.
The IA-64 ISA is strange and the Itanium pipeline is bizarre! How
come everybody else is moving from static to dynamic OOO pipelines,
and Intel moves in the opposite direction? Compiler writers failed to
generate reasonable code for the static RISC pipelines of the mid
1990'ies, and they sure won't do any better with the IA-64 ISA and its
Itanium implementation!
The IA-64 ISA keeps instructions three and three in 128 bit bundles.
Programmers/compilers need to put explicit breaks `;;' between WAW or
RAW dependencies. Such breaks can be anywhere in a bundle, or between
bundles.
The Itanium implementation can under ideal conditions execute two
bundles per cycle. But to reach that rate for integer code, one needs
to sprinkle the code with `nop.f' instructions, one per bundle. With
good scheduling, one can thus reach a peak execution rate of 4
instructions/cycle.
Taken cloop branches seem to insert a bubble into the pipeline most of
the time.
Loads to the fp registers bypass the L1 cache and thus get extremely
long latencies, 9 cycles.
The software pipeline stuff using br.ctop instruction causes delays,
since many issue slots are taken up by instructions with zero
predicates, and since about 15 extra instructions are needed to set
things up. These features are designed for code density, not maximum
speed.
Misc pipeline limitations:
* The getf.sig instruction can only execute in M0.
* At most four integer instructions/cycle.
* Nops take up resources like any plain instructions.
================================================================
mpn_add_n, mpn_sub_n:
The current code runs at 3 cycles/limb. Unrolling could clearly bring
down the time to 2 cycles/limb.
================================================================
mpn_addmul_1:
The current code runs at 3.7 cycles/limb, but that somewhat odd timing
is reached only for huge operands. It uses the mod-scheduled software
pipelining feature. The reason for the poor speed for small operands
is that mod-scheduled loops have a very long start-up overhead.
For optimal speed, we need to load two 64-bit limbs with the ldfp8
instruction, and stay away from mod-scheduled loops. Since rp and up
might be mutually aligned in two ways, we will need two loop variants,
with the same basic structure:
{ .mfi getf.sig
xma.l
(p6) cmp.leu p6, p7 =
} { .mfi stf8
xma.hu
(p7) cmp.ltu p6, p7 =
;;
} { .mib getf.sig
(p6) add 1
nop.b
} { .mib ldfp8 = [up], 16
(p7) add
nop.b
;;
{ .mfi getf.sig
xma.l
(p6) cmp.leu p6, p7 =
} { .mfi stf8
xma.hu
(p7) cmp.ltu p6, p7 =
;;
} { .mib getf.sig
(p6) add 1
nop.b
} { .mib ldfp8 = [rp], 16
(p7) add
br.cloop
;;
}
2 limbs/20 instructions
20 insn/max 6 insn/cycle: 3.3 cycles/2limb
8 memops/max 2 memops/cycle: 4.0 cycles/2limb
8 intops/max 2 intops/cycle: 4.0 cycles/2limb
4 fpops/max 2 fpops/cycle: 2.0 cycles/2limb
================================================================
mpn_submul_1:
The current code just calls mpn_mul_1 and mpn_sub_n and thus needs
about 7 cycles/limb.
We could implement this much like mpn_addmul_1, if we first complement
v. When v is complemented, the low product limb becomes complement of
true product. This should allow us to use the accumulation of xma.
Here is how it works:
#define umul_ppmma(ph, pl, m0, m1, a) \
do { \
UDItype __m0 = (m0), __m1 = (m1), __a = (a); \
__asm__ ("xma.hu %0 = %1, %2, %3" \
: "=f" (ph) \
: "f" (m0), "f" (m1), "f" (__a)); \
(pl) = __m0 * __m1 + __a; \
} while (0)
mp_limb_t
mpn_submul_1 (mp_ptr rp, mp_srcptr up, mp_size_t n, mp_limb_t vl)
{
mp_limb_t cl, cy;
mp_size_t i;
mp_limb_t phi, plo;
mp_limb_t x;
mp_limb_t ul, vln;
vln = -vl;
cl = 0;
for (i = n; i != 0; i--)
{
ul = *up++; /* will need this in both fregs and gregs */
x = *rp;
umul_ppmma (phi, plo, ul, vln, x);
cy = plo < cl;
plo -= cl;
cl = ul - phi;
cl += cy;
*rp++ = plo;
}
return cl;
}
================================================================
mpn_mul_1:
The current code runs at 3.7 cycles/limb. The code is very similar to
the mpn_addmul_1 code. See comments above.
Faster code wouldn't be too hard to write. This is one possible
pattern:
{ .mfi getf.sig
xma.l
(p6) cmp.leu p6, p7 =
} { .mfi stf8
xma.hu
(p7) cmp.ltu p6, p7 =
;;
} { .mib getf.sig
(p6) add 1
nop.b
} { .mib ldf8 = [up], 8
(p7) add
br.cloop
;;
}
1 limb/12 instructions
12 insn/max 6 insn/cycle: 2.0 cycles/limb
4 memops/max 2 memops/cycle: 2.0 cycles/limb
4 intops/max 2 intops/cycle: 2.0 cycles/limb
2 fpops/max 2 fpops/cycle: 1.0 cycles/limb
================================================================
mpn_mul_8
The add+cmp+add we use on the other codes is optimal for shortening
recurrencies (2 cycles) but the sequence takes up 4 execution slots. When
recurrency depth is not critical, a more standard 3-cycle add+cmp+add is
better.
/* First load the 8 values from v */
ldfp8 v0, v1 = [r35], 16;;
ldfp8 v2, v3 = [r35], 16;;
ldfp8 v4, v5 = [r35], 16;;
ldfp8 v6, v7 = [r35], 16;;
/* In the inner loop, get a new U limb and store a result limb. */
mov lc = un
Loop: ldf8 u0 = [r33], 8
xma.l lp0 = v0, u0, hp0
xma.hu hp0 = v0, u0, hp0
xma.l lp1 = v1, u0, hp1
xma.hu hp1 = v1, u0, hp1
xma.l lp2 = v2, u0, hp2
xma.hu hp2 = v2, u0, hp2
xma.l lp3 = v3, u0, hp3
xma.hu hp3 = v3, u0, hp3
xma.l lp4 = v4, u0, hp4
xma.hu hp4 = v4, u0, hp4
xma.l lp5 = v5, u0, hp5
xma.hu hp5 = v5, u0, hp5
xma.l lp6 = v6, u0, hp6
xma.hu hp6 = v6, u0, hp6
xma.l lp7 = v7, u0, hp7
xma.hu hp7 = v7, u0, hp7
getf.sig l0 = lp0
getf.sig l1 = lp1
getf.sig l2 = lp2
getf.sig l3 = lp3
getf.sig l4 = lp4
getf.sig l5 = lp5
getf.sig l6 = lp6
getf.sig l7 = lp7
add+cmp+add l0, l0, h0
add+cmp+add l1, l1, h1
add+cmp+add l2, l2, h2
add+cmp+add l3, l3, h3
add+cmp+add l4, l4, h4
add+cmp+add l5, l5, h5
add+cmp+add l6, l6, h6
add+cmp+add l7, l7, h7
st8 [r32] = xx, 8
br.cloop Loop
50 insn at max 6 insn/cycle: 8.33 cycles/limb8
10 memops at max 2 memops/cycle: 5 cycles/limb8
16 fpops at max 2 fpops/cycle: 8 cycles/limb8
24 intops at max 4 intops/cycle: 6 cycles/limb8
10+24 memops+intops at max 4/cycle 8.5 cycles/limb8
1.0625 cycles/limb
================================================================
mpn_lshift, mpn_rshift
The current code runs at 2 cycles/limb, but has a too deep software
pipeline. The code suffers badly from the 4 cycle latency of the
variable shift instructions.
Using 63 separate loops, we could use the double-word SHRP
instruction. That instruction has a plain single-cycle latency. We
need 63 loops since this instruction only accept immediate count.
|