-
Notifications
You must be signed in to change notification settings - Fork 102
/
index.html
447 lines (442 loc) · 14.5 KB
/
index.html
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
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<link rel="stylesheet" href="../common-revealjs/css/reveal.css">
<link rel="stylesheet" href="../common-revealjs/css/theme/white.css">
<link rel="stylesheet" href="../common-revealjs/css/custom.css">
<script>
// This is needed when printing the slides to pdf
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match( /print-pdf/gi ) ? '../common-revealjs/css/print/pdf.css' : '../common-revealjs/css/print/paper.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
<script>
// This is used to display the static images on each slide,
// See global-images in this html file and custom.css
(function() {
if(window.addEventListener) {
window.addEventListener('load', () => {
let slides = document.getElementsByClassName("slide-background");
if (slides.length === 0) {
slides = document.getElementsByClassName("pdf-page")
}
// Insert global images on each slide
for(let i = 0, max = slides.length; i < max; i++) {
let cln = document.getElementById("global-images").cloneNode(true);
cln.removeAttribute("id");
slides[i].appendChild(cln);
}
// Remove top level global images
let elem = document.getElementById("global-images");
elem.parentElement.removeChild(elem);
}, false);
}
})();
</script>
</head>
<body>
<div class="reveal">
<div class="slides">
<div id="global-images" class="global-images">
<img src="../common-revealjs/images/sycl_academy.png" />
<img src="../common-revealjs/images/sycl_logo.png" />
<img src="../common-revealjs/images/trademarks.png" />
<img src="../common-revealjs/images/codeplay.png" />
</div>
<!--Slide 1-->
<section class="hbox">
<div class="hbox" data-markdown>
## GPU Programming Principles
</div>
</section>
<!--Slide 2-->
<section class="hbox" data-markdown>
## Learning Objectives
* Learn about coalesced global memory access for performance
* Learn about local memory and how to use it.
</section>
<!--Slide 3-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
* Reading from and writing to global memory is generally very expensive.
* It often involves copying data across an off-chip bus.
* This means you generally want to avoid unnecessary accesses.
* Memory access operations are done in chunks.
* This means accessing data that is physically closer together in memory is more efficient.
</div>
</section>
<!--Slide 4-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 5-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 6-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 7-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 8-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 9-->
<section>
<div class="hbox" data-markdown>
#### Coalesced global memory
</div>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 10-->
<section>
<div class="hbox" data-markdown>
#### Row-major vs Column-major
</div>
<div class="container" data-markdown>
* Coalescing global memory access is particularly important when working in multiple dimensions.
* This is because when doing so you have to convert from a position in 2d space to a linear memory space.
* There are two ways to do this, generally referred to as row-major and column-major.
</div>
</section>
<!--Slide 11-->
<section>
<div class="hbox" data-markdown>
#### Row-major vs Column-major
</div>
</section>
<section>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 12-->
<section>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 13-->
<section>
<div class="container" data-markdown>
</div>
</section>
<!--Slide 14-->
<section>
<div class="hbox" data-markdown>
#### Cost of accessing global memory
</div>
<div class="container" data-markdown>
* Global memory is very expensive to access.
* Even with coalesced global memory access if you are accessing the same elements multiple times that can be expensive.
* Instead you want to cache those values in a lower latency memory.
</div>
</section>
<!--Slide 15-->
<section>
<div class="hbox" data-markdown>
#### Using local memory
</div>
<div class="container" data-markdown>
</div>
<div class="container" data-markdown>
* Local memory is a "manually managed cache", often referred to as scratchpad.
* Local memory is a dedicated on-chip cache, shared per work-group.
* Local memory can be accessed in an uncoalesced fashion without much performance degradation.
</div>
</section>
<!--Slide 16-->
<section>
<div class="hbox" data-markdown>
#### Tiling
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* The iteration space of the kernel function is mapped across multiple work-groups.
* Each work-group has its own allocation of local memory.
* You want to split the input image data into tiles, one for each work-group.
</div>
</div>
</section>
<!--Slide 17-->
<section>
<div class="hbox" data-markdown>
#### Local accessors
</div>
<div class="container">
<code class="code-100pc"><pre>
auto scratchpad = sycl::local_accessor<int, dims>(sycl::range{workGroupSize}, cgh);
</code></pre>
</div>
<div class="container" data-markdown>
* Local memory is allocated via a `local_accessor`.
* Unlike regular `accessor`s they are not created with a `buffer`, they allocate memory per work-group for the duration of the kernel function.
* The `range` provided is the number of elements of the specified type to allocate per work-group.
</div>
</section>
<!--Slide 18-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container" data-markdown>
* Local memory can be used to share partial results between work-items.
* When doing so it's important to synchronize between writes and read to memory to ensure all work-items have reached the same point in the program.
</div>
</section>
<!--Slide 19-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Remember that work-items within a workgroup are not guaranteed to execute in lockstep.
</div>
</div>
</section>
<!--Slide 20-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* A work-item can share results with other work-items via local (or global) memory.
</div>
</div>
</section>
<!--Slide 21-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* This means it's possible for a work-item to read a result that hasn't been written to yet.
* This creates a data race.
</div>
</div>
</section>
<!--Slide 22-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* This problem can be solved with a synchronization primitive called a work-group barrier.
</div>
</div>
</section>
<!--Slide 23-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* When a work-group barrier is inserted work-items will wait until all work-items in the work-group have reached that point.
</div>
</div>
</section>
<!--Slide 24-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Only then can any work-items in the work-group continue execution.
</div>
</div>
</section>
<!--Slide 25-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* So now you can be sure that all of the results that you want to read have been written to.
</div>
</div>
</section>
<!--Slide 26-->
<section>
<div class="hbox" data-markdown>
#### Synchronization
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* However note that this does not apply across work-group boundaries.
* So if you write in a work-item of one work-group and then read it in a work-item of another work-group you again have a data race.
* Furthermore, remember that work-items can only access their own local memory and not that of any other work-groups.
</div>
</div>
</section>
<!--Slide 27-->
<section>
<div class="hbox" data-markdown>
#### Group_barrier
</div>
<div class="container">
<code class="code-100pc"><pre>
sycl::group_barrier(item.get_group());
</code></pre>
</div>
<div class="container" data-markdown>
* Work-group barriers can be invoked by calling `group_barrier` and passing a `group` object.
* You can retrieve a `group` object representing the current work-group by calling `get_group` on an `nd_item`.
* Note this requires the `nd_range` variant of `parallel_for`.
</div>
</section>
<!--Slide 28-->
<section>
<div class="hbox" data-markdown>
#### Matrix Transpose
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* In the next exercise we will transpose a matrix.
</div>
</section>
<!--Slide 29-->
<section>
<div class="hbox" data-markdown>
#### Matrix Transpose
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Reading naively from global memory and writing to global memory will give poor performance.
* This is because at least one of our memory transactions will be uncoalesced.
* Adjacent work items are reading a contiguous block from memory, and writing in a strided fashion into the out array.
</div>
</section>
<!--Slide 30-->
<section>
<div class="hbox" data-markdown>
#### Matrix Transpose
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Using scratchpad memory can allow us to make uncoalesced loads or stores into local memory, not global memory.
* Uncoalesced local memory transactions are less detrimental to performance than uncoalesced global memory transactions.
</div>
</section>
<!--Slide 31-->
<section>
<div class="hbox" data-markdown>
#### Matrix Transpose
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Global memory loads and stores are now coalesced.
* Adjacent work items are reading and writing contiguous blocks.
</div>
</section>
<!--Slide 32-->
<section>
<div class="hbox" data-markdown>
## Questions
</div>
</section>
<!--Slide 33-->
<section>
<div class="container" data-markdown>
Code_Exercises/Matrix_Transpose
</div>
<div class="container" data-markdown>
Use good memory access patterns to transpose a matrix.
</div>
</section>
</div>
</div>
<script src="../common-revealjs/js/reveal.js"></script>
<script src="../common-revealjs/plugin/markdown/marked.js"></script>
<script src="../common-revealjs/plugin/markdown/markdown.js"></script>
<script src="../common-revealjs/plugin/notes/notes.js"></script>
<script>
Reveal.initialize({mouseWheel: true, defaultNotes: true});
Reveal.configure({ slideNumber: true });
</script>
</body>
</html>