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
|
/*
* McKinley-optimized version of copy_page().
*
* Copyright (C) 2002 Hewlett-Packard Co
* David Mosberger <davidm@hpl.hp.com>
*
* Inputs:
* in0: address of target page
* in1: address of source page
* Output:
* no return value
*
* General idea:
* - use regular loads and stores to prefetch data to avoid consuming M-slot just for
* lfetches => good for in-cache performance
* - avoid l2 bank-conflicts by not storing into the same 16-byte bank within a single
* cycle
*
* Principle of operation:
* First, note that L1 has a line-size of 64 bytes and L2 a line-size of 128 bytes.
* To avoid secondary misses in L2, we prefetch both source and destination with a line-size
* of 128 bytes. When both of these lines are in the L2 and the first half of the
* source line is in L1, we start copying the remaining words. The second half of the
* source line is prefetched in an earlier iteration, so that by the time we start
* accessing it, it's also present in the L1.
*
* We use a software-pipelined loop to control the overall operation. The pipeline
* has 2*PREFETCH_DIST+K stages. The first PREFETCH_DIST stages are used for prefetching
* source cache-lines. The second PREFETCH_DIST stages are used for prefetching destination
* cache-lines, the last K stages are used to copy the cache-line words not copied by
* the prefetches. The four relevant points in the pipelined are called A, B, C, D:
* p[A] is TRUE if a source-line should be prefetched, p[B] is TRUE if a destination-line
* should be prefetched, p[C] is TRUE if the second half of an L2 line should be brought
* into L1D and p[D] is TRUE if a cacheline needs to be copied.
*
* This all sounds very complicated, but thanks to the modulo-scheduled loop support,
* the resulting code is very regular and quite easy to follow (once you get the idea).
*
* As a secondary optimization, the first 2*PREFETCH_DIST iterations are implemented
* as the separate .prefetch_loop. Logically, this loop performs exactly like the
* main-loop (.line_copy), but has all known-to-be-predicated-off instructions removed,
* so that each loop iteration is faster (again, good for cached case).
*
* When reading the code, it helps to keep the following picture in mind:
*
* word 0 word 1
* +------+------+---
* | v[x] | t1 | ^
* | t2 | t3 | |
* | t4 | t5 | |
* | t6 | t7 | | 128 bytes
* | n[y] | t9 | | (L2 cache line)
* | t10 | t11 | |
* | t12 | t13 | |
* | t14 | t15 | v
* +------+------+---
*
* Here, v[x] is copied by the (memory) prefetch. n[y] is loaded at p[C]
* to fetch the second-half of the L2 cache line into L1, and the tX words are copied in
* an order that avoids bank conflicts.
*/
#include <asm/asmmacro.h>
#include <asm/page.h>
#define PREFETCH_DIST 8 // McKinley sustains 16 outstanding L2 misses (8 ld, 8 st)
#define src0 r2
#define src1 r3
#define dst0 r9
#define dst1 r10
#define src_pre_mem r11
#define dst_pre_mem r14
#define src_pre_l2 r15
#define dst_pre_l2 r16
#define t1 r17
#define t2 r18
#define t3 r19
#define t4 r20
#define t5 t1 // alias!
#define t6 t2 // alias!
#define t7 t3 // alias!
#define t9 t5 // alias!
#define t10 t4 // alias!
#define t11 t7 // alias!
#define t12 t6 // alias!
#define t14 t10 // alias!
#define t13 r21
#define t15 r22
#define saved_lc r23
#define saved_pr r24
#define A 0
#define B (PREFETCH_DIST)
#define C (B + PREFETCH_DIST)
#define D (C + 3)
#define N (D + 1)
#define Nrot ((N + 7) & ~7)
GLOBAL_ENTRY(copy_page)
.prologue
alloc r8 = ar.pfs, 2, Nrot-2, 0, Nrot
.rotr v[2*PREFETCH_DIST], n[D-C+1]
.rotp p[N]
.save ar.lc, saved_lc
mov saved_lc = ar.lc
.save pr, saved_pr
mov saved_pr = pr
.body
mov src_pre_mem = in1
mov pr.rot = 0x10000
mov ar.ec = 1 // special unrolled loop
mov dst_pre_mem = in0
mov ar.lc = 2*PREFETCH_DIST - 1
add src_pre_l2 = 8*8, in1
add dst_pre_l2 = 8*8, in0
add src0 = 8, in1 // first t1 src
add src1 = 3*8, in1 // first t3 src
add dst0 = 8, in0 // first t1 dst
add dst1 = 3*8, in0 // first t3 dst
mov t1 = (PAGE_SIZE/128) - (2*PREFETCH_DIST) - 1
nop.m 0
nop.i 0
;;
// same as .line_copy loop, but with all predicated-off instructions removed:
.prefetch_loop:
(p[A]) ld8 v[A] = [src_pre_mem], 128 // M0
(p[B]) st8 [dst_pre_mem] = v[B], 128 // M2
br.ctop.sptk .prefetch_loop
;;
cmp.eq p16, p0 = r0, r0 // reset p16 to 1 (br.ctop cleared it to zero)
mov ar.lc = t1 // with 64KB pages, t1 is too big to fit in 8 bits!
mov ar.ec = N // # of stages in pipeline
;;
.line_copy:
(p[D]) ld8 t2 = [src0], 3*8 // M0
(p[D]) ld8 t4 = [src1], 3*8 // M1
(p[B]) st8 [dst_pre_mem] = v[B], 128 // M2 prefetch dst from memory
(p[D]) st8 [dst_pre_l2] = n[D-C], 128 // M3 prefetch dst from L2
;;
(p[A]) ld8 v[A] = [src_pre_mem], 128 // M0 prefetch src from memory
(p[C]) ld8 n[0] = [src_pre_l2], 128 // M1 prefetch src from L2
(p[D]) st8 [dst0] = t1, 8 // M2
(p[D]) st8 [dst1] = t3, 8 // M3
;;
(p[D]) ld8 t5 = [src0], 8
(p[D]) ld8 t7 = [src1], 3*8
(p[D]) st8 [dst0] = t2, 3*8
(p[D]) st8 [dst1] = t4, 3*8
;;
(p[D]) ld8 t6 = [src0], 3*8
(p[D]) ld8 t10 = [src1], 8
(p[D]) st8 [dst0] = t5, 8
(p[D]) st8 [dst1] = t7, 3*8
;;
(p[D]) ld8 t9 = [src0], 3*8
(p[D]) ld8 t11 = [src1], 3*8
(p[D]) st8 [dst0] = t6, 3*8
(p[D]) st8 [dst1] = t10, 8
;;
(p[D]) ld8 t12 = [src0], 8
(p[D]) ld8 t14 = [src1], 8
(p[D]) st8 [dst0] = t9, 3*8
(p[D]) st8 [dst1] = t11, 3*8
;;
(p[D]) ld8 t13 = [src0], 4*8
(p[D]) ld8 t15 = [src1], 4*8
(p[D]) st8 [dst0] = t12, 8
(p[D]) st8 [dst1] = t14, 8
;;
(p[D-1])ld8 t1 = [src0], 8
(p[D-1])ld8 t3 = [src1], 8
(p[D]) st8 [dst0] = t13, 4*8
(p[D]) st8 [dst1] = t15, 4*8
br.ctop.sptk .line_copy
;;
mov ar.lc = saved_lc
mov pr = saved_pr, -1
br.ret.sptk.many rp
END(copy_page)
|
