-
Notifications
You must be signed in to change notification settings - Fork 102
/
index.html
323 lines (318 loc) · 11.2 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
<!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>
## Local memory
</div>
</section>
<!--Slide 2-->
<section class="hbox" data-markdown>
## Learning Objectives
* Learn about tiling using local memory
* Learn about how to synchronize work-groups
</section>
<!--Slide 3-->
<section>
<div class="hbox" data-markdown>
#### Cost of accessing global memory
</div>
<div class="container" data-markdown>
* As we covered earlier 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 4-->
<section>
<div class="hbox" data-markdown>
#### Why are image convolutions good on a GPU?
</div>
<div class="container">
<div class="col" data-markdown>
</div>
<div class="col" data-markdown>
* Looking at the image convolution example.
* For each output pixel we are reading up to NxM pixels from the input image, where N and M are the dimensions of the filter.
* This means each input pixel is being read up to NxM times:
* 3x3 filter: up to 9 ops.
* 5x5 filter: up to 25 ops.
* 7x7 filter: up to 49 ops.
* If each of these operations is a separate load from global memory this becomes very expensive.
</div>
</div>
</section>
<!--Slide 5-->
<section>
<div class="hbox" data-markdown>
#### Using local memory
</div>
<div class="container" data-markdown>
</div>
<div class="container" data-markdown>
* The solution is local memory.
* Local memory is generally on-chip and doesn't have a cache as it's managed manually so is much lower latency.
* Local memory is a smaller dedicated region of memory per work-group.
* Local memory can be used to cache, allowing us to read from global memory just once and then read from local memory instead, often referred to as a scratchpad.
</div>
</section>
<!--Slide 6-->
<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 it's own region of local memory.
* You want to split the input image data into tiles, one for each work-group.
</div>
</div>
</section>
<!--Slide 7-->
<section>
<div class="hbox" data-markdown>
#### Local accessors
</div>
<div class="container">
<code class="code-100pc"><pre>
auto scratchpad = sycl::accessor<int, 1, sycl::access::target::local>(sycl::range{workGroupSize}, cgh);
</code></pre>
</div>
<div class="container" data-markdown>
* Local memory is allocated via an `accessor` with the `access::target::local` access target.
* 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 8-->
<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 9-->
<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 are not guaranteed to all execute at the same time (in parallel).
</div>
</div>
</section>
<!--Slide 10-->
<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 11-->
<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 12-->
<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 13-->
<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 14-->
<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 15-->
<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 16-->
<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 17-->
<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 18-->
<section>
<div class="hbox" data-markdown>
#### Local memory image convolution performance
</div>
<div class="container"data-markdown>
</div>
</section>
<!--Slide 11-->
<section>
<div class="hbox" data-markdown>
## Questions
</div>
</section>
<!--Slide 12-->
<section>
<div class="hbox" data-markdown>
#### Exercise
</div>
<div class="container" data-markdown>
Code_Exercises/Local_Memory_Tiling/source
</div>
<div class="container" data-markdown>
Use local memory to cache a tile of the input image data per work-group.
</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>