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
|
#include <stdio.h>
#include <assert.h>
#include <errno.h>
#include <string.h> // strerror_r
#include <set>
#include <string>
#include "trace.h"
// define TRACING to compile these functions, e.g.,
// gcc -DTRACING -c trace.cpp
#ifdef TRACING
#include <cuda_runtime.h>
#include "magma_internal.h"
#include "magmablas_v1.h"
// set TRACE_METHOD = 2 to record start time as
// later of CPU time and previous event's end time.
// set TRACE_METHOD = 1 to record start time using CUDA event.
#define TRACE_METHOD 2
/******************************************************************************/
struct event_log
{
int ncore;
int cpu_id [ MAX_CORES ];
double cpu_first;
double cpu_start[ MAX_CORES ][ MAX_EVENTS ];
double cpu_end [ MAX_CORES ][ MAX_EVENTS ];
char cpu_tag [ MAX_CORES ][ MAX_EVENTS ][ MAX_LABEL_LEN ];
char cpu_label[ MAX_CORES ][ MAX_EVENTS ][ MAX_LABEL_LEN ];
int ngpu;
int nqueue;
magma_queue_t queues [ MAX_GPU_QUEUES ];
int gpu_id [ MAX_GPU_QUEUES ];
magma_event_t gpu_first[ MAX_GPU_QUEUES ];
#if TRACE_METHOD == 2
double gpu_start[ MAX_GPU_QUEUES ][ MAX_EVENTS ];
#else
magma_event_t gpu_start[ MAX_GPU_QUEUES ][ MAX_EVENTS ];
#endif
magma_event_t gpu_end [ MAX_GPU_QUEUES ][ MAX_EVENTS ];
char gpu_tag [ MAX_GPU_QUEUES ][ MAX_EVENTS ][ MAX_LABEL_LEN ];
char gpu_label[ MAX_GPU_QUEUES ][ MAX_EVENTS ][ MAX_LABEL_LEN ];
};
// global log object
struct event_log glog;
/******************************************************************************/
void trace_init( int ncore, int ngpu, int nqueue, magma_queue_t* queues )
{
if ( ncore > MAX_CORES ) {
fprintf( stderr, "Error in trace_init: ncore %d > MAX_CORES %d\n",
ncore, MAX_CORES );
exit(1);
}
if ( ngpu*nqueue > MAX_GPU_QUEUES ) {
fprintf( stderr, "Error in trace_init: (ngpu=%d)*(nqueue=%d) > MAX_GPU_QUEUES=%d\n",
ngpu, nqueue, MAX_GPU_QUEUES );
exit(1);
}
glog.ncore = ncore;
glog.ngpu = ngpu;
glog.nqueue = nqueue;
// initialize ID = 0
for( int core = 0; core < ncore; ++core ) {
glog.cpu_id[core] = 0;
}
for( int dev = 0; dev < ngpu; ++dev ) {
for( int s = 0; s < nqueue; ++s ) {
int t = dev*glog.nqueue + s;
glog.gpu_id[t] = 0;
glog.queues[t] = queues[t];
}
magma_setdevice( dev );
magma_device_sync();
}
// now that all GPUs are sync'd, record start time
for( int dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
for( int s = 0; s < nqueue; ++s ) {
int t = dev*glog.nqueue + s;
magma_event_create( &glog.gpu_first[t] );
magma_event_record( glog.gpu_first[t], glog.queues[t] );
}
}
// sync again
for( int dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
glog.cpu_first = magma_wtime();
}
/******************************************************************************/
void trace_cpu_start( int core, const char* tag, const char* lbl )
{
int id = glog.cpu_id[core];
glog.cpu_start[core][id] = magma_wtime();
magma_strlcpy( glog.cpu_tag [core][id], tag, MAX_LABEL_LEN );
magma_strlcpy( glog.cpu_label[core][id], lbl, MAX_LABEL_LEN );
}
/******************************************************************************/
void trace_cpu_end( int core )
{
int id = glog.cpu_id[core];
glog.cpu_end[core][id] = magma_wtime();
if ( id+1 < MAX_EVENTS ) {
glog.cpu_id[core] = id+1;
}
else {
fprintf( stderr, "Error in %s: not enough CPU events.\n", __func__ );
}
}
/******************************************************************************/
void trace_gpu_start( int dev, int s, const char* tag, const char* lbl )
{
int t = dev*glog.nqueue + s;
int id = glog.gpu_id[t];
#if TRACE_METHOD == 2
glog.gpu_start[t][id] = magma_wtime();
#else
magma_event_create( &glog.gpu_start[t][id] );
magma_event_record( glog.gpu_start[t][id], glog.queues[t] );
#endif
magma_strlcpy( glog.gpu_tag [t][id], tag, MAX_LABEL_LEN );
magma_strlcpy( glog.gpu_label[t][id], lbl, MAX_LABEL_LEN );
}
/******************************************************************************/
void trace_gpu_end( int dev, int s )
{
int t = dev*glog.nqueue + s;
int id = glog.gpu_id[t];
magma_event_create( &glog.gpu_end[t][id] );
magma_event_record( glog.gpu_end[t][id], glog.queues[t] );
if ( id+1 < MAX_EVENTS ) {
glog.gpu_id[t] = id+1;
}
else {
fprintf( stderr, "Error in %s: not enough GPU events (dev %d, queue %d).\n", __func__, dev, s );
}
}
/******************************************************************************/
void trace_finalize( const char* filename, const char* cssfile )
{
// these are all in SVG "pixels"
double xscale = 200.; // pixels per second
double height = 20.; // of each row
double margin = 5.; // page margin and between some elements
double space = 2.; // between rows
double pad = 5.; // around text
double label = 75.; // width of "CPU:", "GPU:" labels
double left = 2*margin + label;
double xtick = 0.5; // interval of xticks (in seconds)
char buf[ 1024 ];
// sync devices
for( int dev = 0; dev < glog.ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
double time = magma_wtime() - glog.cpu_first;
FILE* trace_file = fopen( filename, "w" );
if ( trace_file == NULL ) {
strerror_r( errno, buf, sizeof(buf) );
fprintf( stderr, "Can't open file '%s': %s (%d)\n", filename, buf, errno );
return;
}
fprintf( stderr, "writing trace to '%s'\n", filename );
// row for each CPU and GPU/queue (with space between), time scale, legend
// 4 margins: at top, above time scale, above legend, at bottom
int h = (int)( (glog.ncore + glog.ngpu*glog.nqueue)*(height + space) - space + 2*height + 4*margin );
int w = (int)( left + time*xscale + margin );
fprintf( trace_file,
"<?xml version=\"1.0\" standalone=\"no\"?>\n"
"<!DOCTYPE svg PUBLIC \"-//W3C//DTD SVG 1.1//EN\"\n"
" \"http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd\">\n\n"
"<svg version=\"1.1\" baseProfile=\"full\"\n"
" xmlns=\"http://www.w3.org/2000/svg\"\n"
" xmlns:inkscape=\"http://www.inkscape.org/namespaces/inkscape\"\n"
" viewBox=\"0 0 %d %d\" width=\"%d\" height=\"%d\" preserveAspectRatio=\"none\">\n\n",
w, h, w, h );
// Inkscape does not currently (Jan 2012) support external CSS;
// see http://wiki.inkscape.org/wiki/index.php/CSS_Support
// So embed CSS file here
FILE* css_file = fopen( cssfile, "r" );
if ( css_file == NULL ) {
strerror_r( errno, buf, sizeof(buf) );
fprintf( stderr, "Can't open file '%s': %s; skipping CSS\n", cssfile, buf );
}
else {
fprintf( trace_file, "<style type=\"text/css\">\n" );
while( fgets( buf, sizeof(buf), css_file ) != NULL ) {
fputs( buf, trace_file );
}
fclose( css_file );
fprintf( trace_file, "</style>\n\n" );
}
// format takes: x, y, width, height, class (tag), id (label)
const char* format =
"<rect x=\"%8.3f\" y=\"%4.0f\" width=\"%8.3f\" height=\"%2.0f\" class=\"%-8s\" inkscape:label=\"%s\"/>\n";
// accumulate unique legend entries
std::set< std::string > legend;
// output CPU events
double top = margin;
for( int core = 0; core < glog.ncore; ++core ) {
if ( glog.cpu_id[core] > MAX_EVENTS ) {
glog.cpu_id[core] += 1; // count last event
fprintf( stderr, "WARNING: trace on core %d, reached limit of %d events; output will be truncated.\n",
core, glog.cpu_id[core] );
}
fprintf( trace_file, "<!-- core %d, nevents %d -->\n", core, glog.cpu_id[core] );
fprintf( trace_file, "<g inkscape:groupmode=\"layer\" inkscape:label=\"core %d\">\n", core );
fprintf( trace_file, "<text x=\"%8.3f\" y=\"%4.0f\" width=\"%4.0f\" height=\"%2.0f\">CPU %d:</text>\n",
margin,
top + height - pad,
label, height,
core );
for( int i = 0; i < glog.cpu_id[core]; ++i ) {
double start = glog.cpu_start[core][i] - glog.cpu_first;
double end = glog.cpu_end [core][i] - glog.cpu_first;
fprintf( trace_file, format,
left + start*xscale,
top,
(end - start)*xscale,
height,
glog.cpu_tag[core][i],
glog.cpu_label[core][i] );
legend.insert( glog.cpu_tag[core][i] );
}
top += (height + space);
fprintf( trace_file, "</g>\n\n" );
}
// output GPU events
for( int dev = 0; dev < glog.ngpu; ++dev ) {
for( int s = 0; s < glog.nqueue; ++s ) {
int t = dev*glog.nqueue + s;
if ( glog.gpu_id[t] >= MAX_EVENTS-1 ) {
glog.gpu_id[t] += 1; // count last event
fprintf( stderr, "WARNING: trace on gpu %d/queue %d reached limit of %d events; output will be truncated.\n",
dev, s, glog.gpu_id[t] );
}
fprintf( trace_file, "<!-- gpu %d, queue %d (t %d), nevents %d -->\n", dev, s, t, glog.gpu_id[t] );
fprintf( trace_file, "<g inkscape:groupmode=\"layer\" inkscape:label=\"gpu %d queue %d\">\n", dev, s );
fprintf( trace_file, "<text x=\"%8.3f\" y=\"%4.0f\" width=\"%4.0f\" height=\"%2.0f\">GPU %d (s%d):</text>\n",
margin,
margin + (dev*glog.nqueue + s + glog.ncore)*(height + space) + height - pad,
label, height,
dev, s );
magma_setdevice( dev );
float start, end;
for( int i = 0; i < glog.gpu_id[t]; ++i ) {
#if TRACE_METHOD == 2
start = glog.gpu_start[t][i] - glog.cpu_first;
if ( i > 0 ) {
// later of task's CPU start time and previous task's end time
start = max( start, end );
}
#else
cudaEventElapsedTime( &start, glog.gpu_first[t], glog.gpu_start[t][i] );
start *= 1e-3; // ms to seconds
#endif
cudaEventElapsedTime( &end, glog.gpu_first[t], glog.gpu_end [t][i] );
end *= 1e-3; // ms to seconds
fprintf( trace_file, format,
left + start*xscale,
top,
(end - start)*xscale,
height,
glog.gpu_tag[t][i],
glog.gpu_label[t][i] );
legend.insert( glog.gpu_tag[t][i] );
}
top += (height + space);
fprintf( trace_file, "</g>\n\n" );
}
}
// output time scale
top += (-space + margin);
fprintf( trace_file, "<g inkscape:groupmode=\"layer\" inkscape:label=\"scale\">\n" );
fprintf( trace_file, "<text x=\"%8.1f\" y=\"%4.0f\" width=\"%2.0f\" height=\"%2.0f\">Time (sec):</text>\n",
margin, top + height - pad, label, height );
for( double s=0; s < time; s += xtick ) {
fprintf( trace_file,
"<line x1=\"%8.1f\" y1=\"0\" x2=\"%8.1f\" y2=\"%4.0f\"/>"
"<text x=\"%8.1f\" y=\"%4.0f\">%4.1f</text>\n",
left + s*xscale,
left + s*xscale,
top,
left + s*xscale,
top + height - pad,
s );
}
fprintf( trace_file, "</g>\n\n" );
top += (height + margin);
// output legend
fprintf( trace_file, "<g inkscape:groupmode=\"layer\" inkscape:label=\"legend\">\n" );
fprintf( trace_file, "<text x=\"%8.1f\" y=\"%4.0f\" width=\"%2.0f\" height=\"%2.0f\">Legend:</text>\n",
margin, top + height - pad, label, height );
double x=left;
for( std::set<std::string>::iterator it=legend.begin(); it != legend.end(); ++it ) {
fprintf( trace_file,
"<rect x=\"%8.1f\" y=\"%4.0f\" width=\"%2.0f\" height=\"%2.0f\" class=\"%s\"/>"
"<text x=\"%8.1f\" y=\"%4.0f\" width=\"%2.0f\" height=\"%2.0f\">%s</text>",
x, top, label, height, (*it).c_str(),
x + pad, top + height - pad, label, height, (*it).c_str() );
x += label + margin;
}
fprintf( trace_file, "</g>\n\n" );
fprintf( trace_file, "</svg>\n" );
fclose( trace_file );
}
#endif // TRACING
|
