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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
|
/* Optimized version of the standard memset() function.
Copyright (c) 2002 Hewlett-Packard Co/CERN
Sverre Jarp <Sverre.Jarp@cern.ch>
Return: dest
Inputs:
in0: dest
in1: value
in2: count
The algorithm is fairly straightforward: set byte by byte until we
we get to a 16B-aligned address, then loop on 128 B chunks using an
early store as prefetching, then loop on 32B chucks, then clear remaining
words, finally clear remaining bytes.
Since a stf.spill f0 can store 16B in one go, we use this instruction
to get peak speed when value = 0. */
#include <asm/asmmacro.h>
#include <asm/export.h>
#undef ret
#define dest in0
#define value in1
#define cnt in2
#define tmp r31
#define save_lc r30
#define ptr0 r29
#define ptr1 r28
#define ptr2 r27
#define ptr3 r26
#define ptr9 r24
#define loopcnt r23
#define linecnt r22
#define bytecnt r21
#define fvalue f6
// This routine uses only scratch predicate registers (p6 - p15)
#define p_scr p6 // default register for same-cycle branches
#define p_nz p7
#define p_zr p8
#define p_unalgn p9
#define p_y p11
#define p_n p12
#define p_yy p13
#define p_nn p14
#define MIN1 15
#define MIN1P1HALF 8
#define LINE_SIZE 128
#define LSIZE_SH 7 // shift amount
#define PREF_AHEAD 8
GLOBAL_ENTRY(memset)
{ .mmi
.prologue
alloc tmp = ar.pfs, 3, 0, 0, 0
lfetch.nt1 [dest] //
.save ar.lc, save_lc
mov.i save_lc = ar.lc
.body
} { .mmi
mov ret0 = dest // return value
cmp.ne p_nz, p_zr = value, r0 // use stf.spill if value is zero
cmp.eq p_scr, p0 = cnt, r0
;; }
{ .mmi
and ptr2 = -(MIN1+1), dest // aligned address
and tmp = MIN1, dest // prepare to check for correct alignment
tbit.nz p_y, p_n = dest, 0 // Do we have an odd address? (M_B_U)
} { .mib
mov ptr1 = dest
mux1 value = value, @brcst // create 8 identical bytes in word
(p_scr) br.ret.dpnt.many rp // return immediately if count = 0
;; }
{ .mib
cmp.ne p_unalgn, p0 = tmp, r0 //
} { .mib
sub bytecnt = (MIN1+1), tmp // NB: # of bytes to move is 1 higher than loopcnt
cmp.gt p_scr, p0 = 16, cnt // is it a minimalistic task?
(p_scr) br.cond.dptk.many .move_bytes_unaligned // go move just a few (M_B_U)
;; }
{ .mmi
(p_unalgn) add ptr1 = (MIN1+1), ptr2 // after alignment
(p_unalgn) add ptr2 = MIN1P1HALF, ptr2 // after alignment
(p_unalgn) tbit.nz.unc p_y, p_n = bytecnt, 3 // should we do a st8 ?
;; }
{ .mib
(p_y) add cnt = -8, cnt //
(p_unalgn) tbit.nz.unc p_yy, p_nn = bytecnt, 2 // should we do a st4 ?
} { .mib
(p_y) st8 [ptr2] = value,-4 //
(p_n) add ptr2 = 4, ptr2 //
;; }
{ .mib
(p_yy) add cnt = -4, cnt //
(p_unalgn) tbit.nz.unc p_y, p_n = bytecnt, 1 // should we do a st2 ?
} { .mib
(p_yy) st4 [ptr2] = value,-2 //
(p_nn) add ptr2 = 2, ptr2 //
;; }
{ .mmi
mov tmp = LINE_SIZE+1 // for compare
(p_y) add cnt = -2, cnt //
(p_unalgn) tbit.nz.unc p_yy, p_nn = bytecnt, 0 // should we do a st1 ?
} { .mmi
setf.sig fvalue=value // transfer value to FLP side
(p_y) st2 [ptr2] = value,-1 //
(p_n) add ptr2 = 1, ptr2 //
;; }
{ .mmi
(p_yy) st1 [ptr2] = value //
cmp.gt p_scr, p0 = tmp, cnt // is it a minimalistic task?
} { .mbb
(p_yy) add cnt = -1, cnt //
(p_scr) br.cond.dpnt.many .fraction_of_line // go move just a few
;; }
{ .mib
nop.m 0
shr.u linecnt = cnt, LSIZE_SH
(p_zr) br.cond.dptk.many .l1b // Jump to use stf.spill
;; }
TEXT_ALIGN(32) // --------------------- // L1A: store ahead into cache lines; fill later
{ .mmi
and tmp = -(LINE_SIZE), cnt // compute end of range
mov ptr9 = ptr1 // used for prefetching
and cnt = (LINE_SIZE-1), cnt // remainder
} { .mmi
mov loopcnt = PREF_AHEAD-1 // default prefetch loop
cmp.gt p_scr, p0 = PREF_AHEAD, linecnt // check against actual value
;; }
{ .mmi
(p_scr) add loopcnt = -1, linecnt //
add ptr2 = 8, ptr1 // start of stores (beyond prefetch stores)
add ptr1 = tmp, ptr1 // first address beyond total range
;; }
{ .mmi
add tmp = -1, linecnt // next loop count
mov.i ar.lc = loopcnt //
;; }
.pref_l1a:
{ .mib
stf8 [ptr9] = fvalue, 128 // Do stores one cache line apart
nop.i 0
br.cloop.dptk.few .pref_l1a
;; }
{ .mmi
add ptr0 = 16, ptr2 // Two stores in parallel
mov.i ar.lc = tmp //
;; }
.l1ax:
{ .mmi
stf8 [ptr2] = fvalue, 8
stf8 [ptr0] = fvalue, 8
;; }
{ .mmi
stf8 [ptr2] = fvalue, 24
stf8 [ptr0] = fvalue, 24
;; }
{ .mmi
stf8 [ptr2] = fvalue, 8
stf8 [ptr0] = fvalue, 8
;; }
{ .mmi
stf8 [ptr2] = fvalue, 24
stf8 [ptr0] = fvalue, 24
;; }
{ .mmi
stf8 [ptr2] = fvalue, 8
stf8 [ptr0] = fvalue, 8
;; }
{ .mmi
stf8 [ptr2] = fvalue, 24
stf8 [ptr0] = fvalue, 24
;; }
{ .mmi
stf8 [ptr2] = fvalue, 8
stf8 [ptr0] = fvalue, 32
cmp.lt p_scr, p0 = ptr9, ptr1 // do we need more prefetching?
;; }
{ .mmb
stf8 [ptr2] = fvalue, 24
(p_scr) stf8 [ptr9] = fvalue, 128
br.cloop.dptk.few .l1ax
;; }
{ .mbb
cmp.le p_scr, p0 = 8, cnt // just a few bytes left ?
(p_scr) br.cond.dpnt.many .fraction_of_line // Branch no. 2
br.cond.dpnt.many .move_bytes_from_alignment // Branch no. 3
;; }
TEXT_ALIGN(32)
.l1b: // ------------------------------------ // L1B: store ahead into cache lines; fill later
{ .mmi
and tmp = -(LINE_SIZE), cnt // compute end of range
mov ptr9 = ptr1 // used for prefetching
and cnt = (LINE_SIZE-1), cnt // remainder
} { .mmi
mov loopcnt = PREF_AHEAD-1 // default prefetch loop
cmp.gt p_scr, p0 = PREF_AHEAD, linecnt // check against actual value
;; }
{ .mmi
(p_scr) add loopcnt = -1, linecnt
add ptr2 = 16, ptr1 // start of stores (beyond prefetch stores)
add ptr1 = tmp, ptr1 // first address beyond total range
;; }
{ .mmi
add tmp = -1, linecnt // next loop count
mov.i ar.lc = loopcnt
;; }
.pref_l1b:
{ .mib
stf.spill [ptr9] = f0, 128 // Do stores one cache line apart
nop.i 0
br.cloop.dptk.few .pref_l1b
;; }
{ .mmi
add ptr0 = 16, ptr2 // Two stores in parallel
mov.i ar.lc = tmp
;; }
.l1bx:
{ .mmi
stf.spill [ptr2] = f0, 32
stf.spill [ptr0] = f0, 32
;; }
{ .mmi
stf.spill [ptr2] = f0, 32
stf.spill [ptr0] = f0, 32
;; }
{ .mmi
stf.spill [ptr2] = f0, 32
stf.spill [ptr0] = f0, 64
cmp.lt p_scr, p0 = ptr9, ptr1 // do we need more prefetching?
;; }
{ .mmb
stf.spill [ptr2] = f0, 32
(p_scr) stf.spill [ptr9] = f0, 128
br.cloop.dptk.few .l1bx
;; }
{ .mib
cmp.gt p_scr, p0 = 8, cnt // just a few bytes left ?
(p_scr) br.cond.dpnt.many .move_bytes_from_alignment //
;; }
.fraction_of_line:
{ .mib
add ptr2 = 16, ptr1
|