Guided auto-parallelization is a diagnostic tool to help you locate portions in your serial code that can be parallelized. When you invoke the guided auto-parallelization tool, the Intel® compiler guides you to make changes in your code that efficiently increases automatic parallelization/vectorization of your code.
To invoke this tool, use the compiler option –guide[=n] (on Linux* operating systems) or /Qguide[:n] (on Windows* operating systems). Using this option causes the compiler to generate messages suggesting ways to improve optimization. You can also use more specific compiler options such as –guide-vec , –guide-par, and –guide-data-trans (on Linux* OS), or the /Qguide-vec, /Qguide-par, and /Qguide-data-trans (on Windows* OS) to obtain individual guided optimizations for vectorizing, parallelizing, and data transformation.
When any guided auto-parallelization option is used, the compiler provides only diagnostic advice. Object files or executables are not created in this mode. See the table below for descriptions of the options.
Linux* |
Windows* |
Description |
|---|---|---|
-guide(=n) |
/Qguide[:n] |
Lets you to set a level of guidance for auto-vectorization and data transformation. To obtain guidance for auto-parallelization, you must use the -parallel (Linux*) or /Qparallel (Windows*) option along with the -guide or /Qguide option. |
-guide-par(=n) |
/Qguide-par[:n] |
Lets you to set a level of guidance only for auto-parallelization.
|
-guide-vec(=n) |
/Qguide-vec[:n] |
Lets you to set a level of guidance only for auto-vectorization. |
-guide-data-trans(=n) |
/Qguide-data-trans[:n] |
Lets you to set a level of guidance only for data transformation. |
For all of the above options, the argument n is optional and it specifies the level of guidance. The argument n takes the values 1-4. When n is not specified, the default is 4. If you specify n=1 or 2, a standard level of guidance is provided.
When you use n=3 or n=4, you may get advanced messages. For example, you may get messages about how to optimize a particular loop-nest or get a message on how exception-handling inside a loop-nest affects optimizations for that loop-nest. Or you may get a message on how to provide extra information to the compiler on cost-modeling (expected values of trip-counts, and so on).
If you specify a level of guidance for the -guide or /Qguide option along with one of the other guide options, and set a different level of guidance (values for n) for the other guide options, the value set for the individual guide optimization overrides the setting specified for -guide or /Qguide option.
If you do not specify a level of guidance for the -guide or /Qguide option, but specify one of the other guide options with a level of guidance, the -guide or /Qguide option is enabled with the greatest value passed to any of the individual guide optimizations specified.
Capturing Guidance Messages
The guided auto-parallelization tool analyzes all of your serial code or individual parts of your code and generates advisory messages. By default, messages that are generated by the guided auto-parallelization tool are output to stderr.
To capture messages into a file, use the options listed in the following table.
Note
The options listed in the table must be used along with the -guide or /Qguide, or -guide-par or /Qguide-par, or -guide-vec or /Qguide-vec, or -guide-data-trans or /Qguide-data-trans options. If they are not, the guided auto-parallelization tool ignores them.
Linux* |
Windows* |
Description |
|---|---|---|
-guide-file[=file_name] |
/Qguide-file[:file_name] |
Gathers all messages generated during a guided auto-parallization compile into a specified file. |
-guide-file-append[=file_name] |
/Qguide-file-append[:file_name] |
Lets you specify the file into which all messages generated during a guided auto-parallelization compile should be appended. |
For the -guide-file or /Qguide-file and -guide-file-append or /Qguide-file-append options, the file_name argument can also include a path. If the path is not specified, the file is created in the current working directory. If there is already a file with the file_name, it is overwritten. If you don’t provide an extension with file_name, the extension is .guide.
Guided Auto-parallelization on Known Hotspots
When you know the hotspots in your application and want advice only for those hotspots, use the –guide-opts (Linux*) or /Qguide-opts (Windows*) option along with any of the guided auto-parallelization options like -guide, -guide-vec, -guide-par, -guide-data-trans (on Linux*) or /Qguide, /Qguide-vec, /Qguide-par, /Qguide-data-trans (Windows*) options. The output message indicates which portions of the input files were analyzed. Complete routine name is generated in the message if a routine is selected to be analyzed.
Use the string parameter to provide information about known hotspots. The string parameter takes one or more of the following variables: filename, routine, range. The compiler parses the string parameter and generates syntax errors if there are any.
Syntax
-guide-opts=string
Follow these guidelines when using the string parameter:
- Use only valid file names, routine names and line numbers. The guided auto-parallelization tool ignores invalid values and issues a diagnostic message stating what was ignored.
- Enclose routine names within single-quotes. Specify original source names (demangled names) as routine names. A routine name alone may not always be sufficient to uniquely identify a routine. You may need to specify additional parameter information to uniquely identify the routine.
For example:
-guide-opts="foo.cpp,'CLHEP::StaticRandomSta::restore(std::basic_istream<char,std::char_traits<char>>&)'"
-guide-opts="bar.f90,'module_1::routine_name`"
-guide-opts="baz.c,'c_routine_name'"
For any specified routine name, the GAP tool first tries to uniquely identify the routine using specified routine information. If that is not possible, then it selects all routines with the specified routine name. The GAP tool uses the parameter information, if specified, to narrow down the selection.
- When inlining is involved, use the callee line numbers. The generated messages also use the callee line-numbers.