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
|
Requirements:
- automake, autoconf, libtool
(not needed when compiling a release)
- pkg-config (http://www.freedesktop.org/wiki/Software/pkg-config)
(not needed when compiling a release using the included isl and pet)
- gmp (http://gmplib.org/)
- libyaml (http://pyyaml.org/wiki/LibYAML)
(only needed if you want to compile the pet executable)
- LLVM/clang libraries, 2.9 or higher (http://clang.llvm.org/get_started.html)
Unless you have some other reasons for wanting to use the svn version,
it is best to install the latest release (3.9).
For more details, see pet/README.
If you are installing on Ubuntu, then you can install the following packages:
automake autoconf libtool pkg-config libgmp3-dev libyaml-dev libclang-dev llvm
Note that you need at least version 3.2 of libclang-dev (ubuntu raring).
Older versions of this package did not include the required libraries.
If you are using an older version of ubuntu, then you need to compile and
install LLVM/clang from source.
Preparing:
Grab the latest release and extract it or get the source from
the git repository as follows. This process requires autoconf,
automake, libtool and pkg-config.
git clone git://repo.or.cz/ppcg.git
cd ppcg
./get_submodules.sh
./autogen.sh
Compilation:
./configure
make
make check
If you have installed any of the required libraries in a non-standard
location, then you may need to use the --with-gmp-prefix,
--with-libyaml-prefix and/or --with-clang-prefix options
when calling "./configure".
Using PPCG to generate CUDA or OpenCL code
To convert a fragment of a C program to CUDA, insert a line containing
#pragma scop
before the fragment and add a line containing
#pragma endscop
after the fragment. To generate CUDA code run
ppcg --target=cuda file.c
where file.c is the file containing the fragment. The generated
code is stored in file_host.cu and file_kernel.cu.
To generate OpenCL code run
ppcg --target=opencl file.c
where file.c is the file containing the fragment. The generated code
is stored in file_host.c and file_kernel.cl.
Specifying tile, grid and block sizes
The iterations space tile size, grid size and block size can
be specified using the --sizes option. The argument is a union map
in isl notation mapping kernels identified by their sequence number
in a "kernel" space to singleton sets in the "tile", "grid" and "block"
spaces. The sizes are specified outermost to innermost.
The dimension of the "tile" space indicates the (maximal) number of loop
dimensions to tile. The elements of the single integer tuple
specify the tile sizes in each dimension.
In case of hybrid tiling, the first element is half the size of
the tile in the time (sequential) dimension. The second element
specifies the number of elements in the base of the hexagon.
The remaining elements specify the tile sizes in the remaining space
dimensions.
The dimension of the "grid" space indicates the (maximal) number of block
dimensions in the grid. The elements of the single integer tuple
specify the number of blocks in each dimension.
The dimension of the "block" space indicates the (maximal) number of thread
dimensions in the grid. The elements of the single integer tuple
specify the number of threads in each dimension.
For example,
{ kernel[0] -> tile[64,64]; kernel[i] -> block[16] : i != 4 }
specifies that in kernel 0, two loops should be tiled with a tile
size of 64 in both dimensions and that all kernels except kernel 4
should be run using a block of 16 threads.
Since PPCG performs some scheduling, it can be difficult to predict
what exactly will end up in a kernel. If you want to specify
tile, grid or block sizes, you may want to run PPCG first with the defaults,
examine the kernels and then run PPCG again with the desired sizes.
Instead of examining the kernels, you can also specify the option
--dump-sizes on the first run to obtain the effectively used default sizes.
Compiling the generated CUDA code with nvcc
To get optimal performance from nvcc, it is important to choose --arch
according to your target GPU. Specifically, use the flag "--arch sm_20"
for fermi, "--arch sm_30" for GK10x Kepler and "--arch sm_35" for
GK110 Kepler. We discourage the use of older cards as we have seen
correctness issues with compilation for older architectures.
Note that in the absence of any --arch flag, nvcc defaults to
"--arch sm_13". This will not only be slower, but can also cause
correctness issues.
If you want to obtain results that are identical to those obtained
by the original code, then you may need to disable some optimizations
by passing the "--fmad=false" option.
Compiling the generated OpenCL code with gcc
To compile the host code you need to link against the file
ocl_utilities.c which contains utility functions used by the generated
OpenCL host code. To compile the host code with gcc, run
gcc -std=c99 file_host.c ocl_utilities.c -lOpenCL
Note that we have experienced the generated OpenCL code freezing
on some inputs (e.g., the PolyBench symm benchmark) when using
at least some version of the Nvidia OpenCL library, while the
corresponding CUDA code runs fine.
We have experienced no such freezes when using AMD, ARM or Intel
OpenCL libraries.
By default, the compiled executable will need the _kernel.cl file at
run time. Alternatively, the option --opencl-embed-kernel-code may be
given to place the kernel code in a string literal. The kernel code is
then compiled into the host binary, such that the _kernel.cl file is no
longer needed at run time. Any kernel include files, in particular
those supplied using --opencl-include-file, will still be required at
run time.
Function calls
Function calls inside the analyzed fragment are reproduced
in the CUDA or OpenCL code, but for now it is left to the user
to make sure that the functions that are being called are
available from the generated kernels.
In the case of OpenCL code, the --opencl-include-file option
may be used to specify one or more files to be #include'd
from the generated code. These files may then contain
the definitions of the functions being called from the
program fragment. If the pathnames of the included files
are relative to the current directory, then you may need
to additionally specify the --opencl-compiler-options=-I.
to make sure that the files can be found by the OpenCL compiler.
The included files may contain definitions of types used by the
generated kernels. By default, PPCG generates definitions for
types as needed, but these definitions may collide with those in
the included files, as PPCG does not consider the contents of the
included files. The --no-opencl-print-kernel-types will prevent
PPCG from generating type definitions.
GNU extensions
By default, PPCG may print out macro definitions that involve
GNU extensions such as __typeof__ and statement expressions.
Some compilers may not support these extensions.
In particular, OpenCL 1.2 beignet 1.1.1 (git-6de6918)
has been reported not to support __typeof__.
The use of these extensions can be turned off with the
--no-allow-gnu-extensions option.
Processing PolyBench
When processing a PolyBench/C 3.2 benchmark, you should always specify
-DPOLYBENCH_USE_C99_PROTO on the ppcg command line. Otherwise, the source
files are inconsistent, having fixed size arrays but parametrically
bounded loops iterating over them.
However, you should not specify this define when compiling
the PPCG generated code using nvcc since CUDA does not support VLAs.
CUDA and function overloading
While CUDA supports function overloading based on the arguments types,
no such function overloading exists in the input language C. Since PPCG
simply prints out the same function name as in the original code, this
may result in a different function being called based on the types
of the arguments. For example, if the original code contains a call
to the function sqrt() with a float argument, then the argument will
be promoted to a double and the sqrt() function will be called.
In the transformed (CUDA) code, however, overloading will cause the
function sqrtf() to be called. Until this issue has been resolved in PPCG,
we recommend that users either explicitly call the function sqrtf() or
explicitly cast the argument to double in the input code.
Contact
For bug reports, feature requests and questions,
contact http://groups.google.com/group/isl-development
Whenever you report a bug, please mention the exact version of PPCG
that you are using (output of "./ppcg --version"). If you are unable
to compile PPCG, then report the git version (output of "git describe")
or the version number included in the name of the tarball.
Citing PPCG
If you use PPCG for your research, you are invited to cite
the following paper.
@article{Verdoolaege2013PPCG,
author = {Verdoolaege, Sven and Juega, Juan Carlos and Cohen, Albert and
G\'{o}mez, Jos{\'e} Ignacio and Tenllado, Christian and
Catthoor, Francky},
title = {Polyhedral parallel code generation for CUDA},
journal = {ACM Trans. Archit. Code Optim.},
issue_date = {January 2013},
volume = {9},
number = {4},
month = jan,
year = {2013},
issn = {1544-3566},
pages = {54:1--54:23},
doi = {10.1145/2400682.2400713},
acmid = {2400713},
publisher = {ACM},
address = {New York, NY, USA},
}
|
