3 - automake, autoconf, libtool
4 (not needed when compiling a release)
5 - pkg-config (http://www.freedesktop.org/wiki/Software/pkg-config)
6 (not needed when compiling a release using the included isl and pet)
7 - gmp (http://gmplib.org/)
8 - libyaml (http://pyyaml.org/wiki/LibYAML)
9 (only needed if you want to compile the pet executable)
10 - LLVM/clang libraries, 2.9 or higher (http://clang.llvm.org/get_started.html)
11 Unless you have some other reasons for wanting to use the svn version,
12 it is best to install the latest release (3.3).
13 For more details, see pet/README.
15 If you are installing on Ubuntu, then you can install the following packages:
17 automake autoconf libtool pkg-config libgmp3-dev libyaml-dev libclang-dev llvm
19 Note that you need at least version 3.2 of libclang-dev (ubuntu raring).
20 Older versions of this package did not include the required libraries.
21 If you are using an older version of ubuntu, then you need to compile and
22 install LLVM/clang from source.
27 Grab the latest release and extract it or get the source from
28 the git repository as follows. This process requires autoconf,
29 automake, libtool and pkg-config.
31 git clone git://repo.or.cz/ppcg.git
44 If you have installed any of the required libraries in a non-standard
45 location, then you may need to use the --with-gmp-prefix,
46 --with-libyaml-prefix and/or --with-clang-prefix options
47 when calling "./configure".
50 Using PPCG to generate CUDA or OpenCL code
52 To convert a fragment of a C program to CUDA, insert a line containing
56 before the fragment and add a line containing
60 after the fragment. To generate CUDA code run
62 ppcg --target=cuda file.c
64 where file.c is the file containing the fragment. The generated
65 code is stored in file_host.cu and file_kernel.cu.
67 To generate OpenCL code run
69 ppcg --target=opencl file.c
71 where file.c is the file containing the fragment. The generated code
72 is stored in file_host.c and file_kernel.cl.
75 Specifying tile, grid and block sizes
77 The iterations space tile size, grid size and block size can
78 be specified using the --sizes option. The argument is a union map
79 in isl notation mapping kernels identified by their sequence number
80 in a "kernel" space to singleton sets in the "tile", "grid" and "block"
81 spaces. The sizes are specified outermost to innermost.
83 The dimension of the "tile" space indicates the (maximal) number of loop
84 dimensions to tile. The elements of the single integer tuple
85 specify the tile sizes in each dimension.
87 The dimension of the "grid" space indicates the (maximal) number of block
88 dimensions in the grid. The elements of the single integer tuple
89 specify the number of blocks in each dimension.
91 The dimension of the "block" space indicates the (maximal) number of thread
92 dimensions in the grid. The elements of the single integer tuple
93 specify the number of threads in each dimension.
97 { kernel[0] -> tile[64,64]; kernel[i] -> block[16] : i != 4 }
99 specifies that in kernel 0, two loops should be tiled with a tile
100 size of 64 in both dimensions and that all kernels except kernel 4
101 should be run using a block of 16 threads.
103 Since PPCG performs some scheduling, it can be difficult to predict
104 what exactly will end up in a kernel. If you want to specify
105 tile, grid or block sizes, you may want to run PPCG first with the defaults,
106 examine the kernels and then run PPCG again with the desired sizes.
107 Instead of examining the kernels, you can also specify the option
108 --dump-sizes on the first run to obtain the effectively used default sizes.
111 Compiling the generated CUDA code with nvcc
113 To get optimal performance from nvcc, it is important to choose --arch
114 according to your target GPU. Specifically, use the flag "--arch sm_20"
115 for fermi, "--arch sm_30" for GK10x Kepler and "--arch sm_35" for
116 GK110 Kepler. We discourage the use of older cards as we have seen
117 correctness issues with compilation for older architectures.
118 Note that in the absence of any --arch flag, nvcc defaults to
119 "--arch sm_13". This will not only be slower, but can also cause
121 If you want to obtain results that are identical to those obtained
122 by the original code, then you may need to disable some optimizations
123 by passing the "--fmad=false" option.
126 Compiling the generated OpenCL code with gcc
128 To compile the host code you need to link against the file
129 ocl_utilities.c which contains utility functions used by the generated
130 OpenCL host code. To compile the host code with gcc, run
132 gcc -std=c99 file_host.c ocl_utilities.c -lOpenCL
134 Note that we have experienced the generated OpenCL code freezing
135 on some inputs (e.g., the PolyBench symm benchmark) when using
136 at least some version of the Nvidia OpenCL library, while the
137 corresponding CUDA code runs fine.
138 We have experienced no such freezes when using AMD, ARM or Intel
144 When processing a PolyBench/C 3.2 benchmark, you should always specify
145 -DPOLYBENCH_USE_C99_PROTO on the ppcg command line. Otherwise, the source
146 files are inconsistent, having fixed size arrays but parametrically
147 bounded loops iterating over them.
150 CUDA and function overloading
152 While CUDA supports function overloading based on the arguments types,
153 no such function overloading exists in the input language C. Since PPCG
154 simply prints out the same function name as in the original code, this
155 may result in a different function being called based on the types
156 of the arguments. For example, if the original code contains a call
157 to the function sqrt() with a float argument, then the argument will
158 be promoted to a double and the sqrt() function will be called.
159 In the transformed (CUDA) code, however, overloading will cause the
160 function sqrtf() to be called. Until this issue has been resolved in PPCG,
161 we recommend that users either explicitly call the function sqrtf() or
162 explicitly cast the argument to double in the input code.
165 Additional variables in generated code
167 The generated code may contain additional variables with names
168 that match /^[bthsgc][0-d]+$/. These variables may shadow or
169 conflict with variables in the input program. Until this issue
170 has been resolved in PPCG, you should avoid such variable names
171 in your input program.
176 If you use PPCG for your research, you are invited to cite
179 @article{Verdoolaege2013PPCG,
180 author = {Verdoolaege, Sven and Juega, Juan Carlos and Cohen, Albert and
181 G\'{o}mez, Jos{\'e} Ignacio and Tenllado, Christian and
183 title = {Polyhedral parallel code generation for CUDA},
184 journal = {ACM Trans. Archit. Code Optim.},
185 issue_date = {January 2013},
191 pages = {54:1--54:23},
192 doi = {10.1145/2400682.2400713},
195 address = {New York, NY, USA},