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
|
# Structure of array (SoA) generation
The two header files [`SoALayout.h`](SoALayout.h) and [`SoAView.h`](SoAView.h) define preprocessor macros that
allow generating SoA classes. The SoA classes generate multiple, aligned column from a memory buffer. The memory
buffer is allocated separately by the user, and can be located in a memory space different from the local one (for
example, a SoA located in a GPU device memory can be fully pre-defined on the host and the resulting structure is
passed to the GPU kernel).
This columnar storage allows efficient memory access by GPU kernels (coalesced access on cache line aligned data)
and possibly vectorization.
Additionally, templation of the layout and view classes allows compile-time variations of accesses and checks:
verification of alignment and corresponding compiler hinting, cache strategy (non-coherent, streaming with immediate
invalidation), range checking.
Macro generation allows generating code that provides a clear and concise access of data when used. The code
generation uses the Boost Preprocessing library.
## Layout
`SoALayout` is a macro generated templated class that subdivides a provided buffer into a collection of columns,
Eigen columns and scalars. The buffer is expected to be aligned with a selectable alignment defaulting to the CUDA
GPU cache line (128 bytes). All columns and scalars within a `SoALayout` will be individually aligned, leaving
padding at the end of each if necessary. Eigen columns have each component of the vector or matrix properly aligned
in individual column (by defining the stride between components). Only compile-time sized Eigen vectors and matrices
are supported. Scalar members are members of layout with one element, irrespective of the size of the layout.
Static utility functions automatically compute the byte size of a layout, taking into account all its columns and
alignment.
## View
`SoAView` is a macro generated templated class allowing access to columns defined in one or multiple `SoALayout`s or
`SoAViews`. The view can be generated in a constant and non-constant flavors. All view flavors provide with the same
interface where scalar elements are accessed with an `operator()`: `soa.scalar()` while columns (Eigen or not) are
accessed via a array of structure (AoS) -like syntax: `soa[index].x()`. The "struct" object returned by `operator[]`
can be used as a shortcut: `auto si = soa[index]; si.z() = si.x() + zi.y();`
A view can be instanciated by being passed the layout(s) and view(s) it is defined against, or column by column.
Layout classes also define a `View` and `ConstView` subclass that provide access to each column and
scalar of the layout. In addition to those fully parametrized templates, two others levels of parametrization are
provided: `ViewTemplate`, `ViewViewTemplateFreeParams` and respectively `ConstViewTemplate`,
`ConstViewTemplateFreeParams`. The parametrization of those templates is explained in the [Template
parameters section](#template-parameters).
It is also possible to build a generic `View` or `ConstView` passing from the [Metarecords sublass](#metarecords-subclass). This
view can point to data belonging to different SoAs and thus not contiguous in memory.
## Metadata subclass
In order to no clutter the namespace of the generated class, a subclass name `Metadata` is generated. It is
instanciated with the `metadata()` member function and contains various utility functions, like `size()` (number
of elements in the SoA), `byteSize()`, `byteAlignment()`, `data()` (a pointer to the buffer). A `nextByte()`
function computes the first byte of a structure right after a layout, allowing using a single buffer for multiple
layouts.
## Metarecords subclass
The nested type `Metarecords` describes the elements of the SoA. It can be instantiated by the `records()` member
function of a `View` or `ConstView`. Every object contains the address of the first element of the column, the number
of elements per column, and the stride for the Eigen columns. These are used to validate the columns size at run time
and to build a generic `View` as described in [View](#view).
## ROOT serialization and de-serialization
Layouts can be serialized and de-serialized with ROOT. In order to generate the ROOT dictionary, separate
`clases_def.xml` and `classes.h` should be prepared. `classes.h` ensures the inclusion of the proper header files to
get the definition of the serialized classes, and `classes_def.xml` needs to define the fixed list of members that
ROOT should ignore, plus the list of all the columns. [An example is provided below.](#examples)
Serialization of Eigen data is not yet supported.
## Template parameters
The template shared by layouts and parameters are:
- Byte aligment (defaulting to the nVidia GPU cache line size (128 bytes))
- Alignment enforcement (`relaxed` or `enforced`). When enforced, the alignment will be checked at construction
time.~~, and the accesses are done with compiler hinting (using the widely supported `__builtin_assume_aligned`
intrinsic).~~ It turned out that hinting `nvcc` for alignement removed the benefit of more important `__restrict__`
hinting. The `__builtin_assume_aligned` is hence currently not use.
In addition, the views also provide access parameters:
- Restrict qualify: add restrict hints to read accesses, so that the compiler knows it can relax accesses to the
data and assume it will not change. On nVidia GPUs, this leads to the generation of instruction using the faster
non-coherent cache.
- Range checking: add index checking on each access. As this is a compile time parameter, the cost of the feature at
run time is null if turned off. When turned on, the accesses will be slowed down by checks. Uppon error detection,
an exception is launched (on the CPU side) or the kernel is made to crash (on the GPU side). This feature can help
the debugging of index issues at runtime, but of course requires a recompilation.
The trivial views subclasses come in a variety of parametrization levels: `View` uses the same byte
alignement and alignment enforcement as the layout, and defaults (off) for restrict qualifying and range checking.
`ViewTemplate` template allows setting of restrict qualifying and range checking, while
`ViewTemplateFreeParams` allows full re-customization of the template parameters.
## Using SoA layouts and views with GPUs
Instanciation of views and layouts is preferably done on the CPU side. The view object is lightweight, with only one
pointer per column, plus the global number of elements. Extra view class can be generated to restrict this number of
pointers to the strict minimum in scenarios where only a subset of columns are used in a given GPU kernel.
## Examples
A layout can be defined as:
```C++
#include "DataFormats/SoALayout.h"
GENERATE_SOA_LAYOUT(SoA1LayoutTemplate,
// predefined static scalars
// size_t size;
// size_t alignment;
// columns: one value per element
SOA_COLUMN(double, x),
SOA_COLUMN(double, y),
SOA_COLUMN(double, z),
SOA_EIGEN_COLUMN(Eigen::Vector3d, a),
SOA_EIGEN_COLUMN(Eigen::Vector3d, b),
SOA_EIGEN_COLUMN(Eigen::Vector3d, r),
SOA_COLUMN(uint16_t, color),
SOA_COLUMN(int32_t, value),
SOA_COLUMN(double *, py),
SOA_COLUMN(uint32_t, count),
SOA_COLUMN(uint32_t, anotherCount),
// scalars: one value for the whole structure
SOA_SCALAR(const char *, description),
SOA_SCALAR(uint32_t, someNumber)
);
// Default template parameters are <
// size_t ALIGNMENT = cms::soa::CacheLineSize::defaultSize,
// bool ALIGNMENT_ENFORCEMENT = cms::soa::AlignmentEnforcement::relaxed
// >
using SoA1Layout = SoA1LayoutTemplate<>;
using SoA1LayoutAligned = SoA1LayoutTemplate<cms::soa::CacheLineSize::defaultSize, cms::soa::AlignmentEnforcement::enforced>;
```
The buffer of the proper size is allocated, and the layout is populated with:
```C++
// Allocation of aligned
size_t elements = 100;
using AlignedBuffer = std::unique_ptr<std::byte, decltype(std::free) *>;
AlignedBuffer h_buf (reinterpret_cast<std::byte*>(aligned_alloc(SoA1LayoutAligned::alignment, SoA1LayoutAligned::computeDataSize(elements))), std::free);
SoA1LayoutAligned soaLayout(h_buf.get(), elements);
```
A view will derive its column types from one or multiple layouts. The macro generating the view takes a list of layouts or views it
gets is data from as a first parameter, and the selection of the columns the view will give access to as a second parameter.
```C++
// A 1 to 1 view of the layout (except for unsupported types).
GENERATE_SOA_VIEW(SoA1ViewTemplate,
SOA_VIEW_LAYOUT_LIST(
SOA_VIEW_LAYOUT(SoA1Layout, soa1)
),
SOA_VIEW_VALUE_LIST(
SOA_VIEW_VALUE(soa1, x),
SOA_VIEW_VALUE(soa1, y),
SOA_VIEW_VALUE(soa1, z),
SOA_VIEW_VALUE(soa1, color),
SOA_VIEW_VALUE(soa1, value),
SOA_VIEW_VALUE(soa1, py),
SOA_VIEW_VALUE(soa1, count),
SOA_VIEW_VALUE(soa1, anotherCount),
SOA_VIEW_VALUE(soa1, description),
SOA_VIEW_VALUE(soa1, someNumber)
)
);
using SoA1View = SoA1ViewTemplate<>;
SoA1View soaView(soaLayout);
for (size_t i=0; i < soaLayout.metadata().size(); ++i) {
auto si = soaView[i];
si.x() = si.y() = i;
soaView.someNumber() += i;
}
```
The mutable and const views with the exact same set of columns and their parametrized variants are provided from the layout as:
```C++
// (Pseudo-code)
struct SoA1Layout::View;
template<bool RESTRICT_QUALIFY = cms::soa::RestrictQualify::enabled,
bool RANGE_CHECKING = cms::soa::RangeChecking::disabled>
struct SoA1Layout::ViewTemplate;
template<size_t ALIGNMENT = cms::soa::CacheLineSize::defaultSize,
bool ALIGNMENT_ENFORCEMENT = cms::soa::AlignmentEnforcement::relaxed,
bool RESTRICT_QUALIFY = cms::soa::RestrictQualify::enabled,
bool RANGE_CHECKING = cms::soa::RangeChecking::disabled>
struct SoA1Layout::ViewTemplateFreeParams;
struct SoA1Layout::ConstView;
template<bool RESTRICT_QUALIFY = cms::soa::RestrictQualify::enabled,
bool RANGE_CHECKING = cms::soa::RangeChecking::disabled>
struct SoA1Layout::ConstViewTemplate;
template<size_t ALIGNMENT = cms::soa::CacheLineSize::defaultSize,
bool ALIGNMENT_ENFORCEMENT = cms::soa::AlignmentEnforcement::relaxed,
bool RESTRICT_QUALIFY = cms::soa::RestrictQualify::enabled,
bool RANGE_CHECKING = cms::soa::RangeChecking::disabled>
struct SoA1Layout::ConstViewTemplateFreeParams;
```
## Current status and further improvements
### Available features
- The layout and views support scalars and columns, alignment and alignment enforcement and hinting (linked).
- Automatic `__restrict__` compiler hinting is supported and can be enabled where appropriate.
- Automatic creation of trivial views and const views derived from a single layout.
- Cache access style, which was explored, was abandoned as this not-yet-used feature interferes with `__restrict__`
support (which is already in used in existing code). It could be made available as a separate tool that can be used
directly by the module developer, orthogonally from SoA.
- Optional (compile time) range checking validates the index of every column access, throwing an exception on the
CPU side and forcing a segmentation fault to halt kernels. When not enabled, it has no impact on performance (code
not compiled)
- Eigen columns are also suported, with both const and non-const flavors.
- ROOT serialization and deserialization is supported. In CMSSW, it is planned to be used through the memory
managing `PortableCollection` family of classes.
- An `operator<<()` is provided to print the layout of an SoA to standard streams.
|