Advanced

Generic Pipeline

This program allows to build and to execute generic chains directly from the command line interface (CLI). Basically, there are 3 different types of task that can be instantiated:

• First task: To be correct, a chain should start with a task that have no input socket. As a consequence, there is only one first task and it is possible to choose between read and initialize tasks.
• Middle task: A task that have an input and an output socket. It is up to the user to decide the number and the combination of middle tasks he wants. It is possible to select between relay, relayf, increment and incrementf tasks.
• Last task: To be correct, a chain should always end with a task that does not have an output socket. As a consequence, there is only one last task and it is possible to choose between write or finalize tasks.

Here is a summary of the available tasks and their behavior:

• read: Reads data from a binary file and writes the read bytes on its output socket.
• initialize or init: Initializes the data in its output socket (useful for benchmark and validation).
• relay: Copies the data from its input socket into its output socket.
• relayf: Variant of the relay task that uses a forward socket, consequently, this task does NOTHING.
• increment or incr: Increments (+1) the data of its input socket and writes the result in its output socket.
• incrementf or incrf: Variant of the increment task that uses a forward socket to write the result in place.
• write: Writes contents of its input socket into a binary file. It expects 0 or 1 values in its input socket to work correctly.
• finalize or fin: Memorizes (= copies) the input data for further validation (if there is a validation).

There are three main ways of describing a processing chain:

1. Specification of homogeneous types of task per stage. This is performed with the combination of the -R (or --tsk-types-sta) and -n (or --tsk-per-sta) CLI parameters. -R gives the tasks type per stage (example of a 4-stage pipeline: -R (read,incr,relayf,write)) and -n gives how many tasks of the same type will be created per stage. For instance, the combination of -R (read,incr,relayf,write) and -n "1,2,3,1" will produce a 4-stage pipeline with the following sequence of tasks: read \(\rightarrow\) incr \(\rightarrow\) incr \(\rightarrow\) relayf \(\rightarrow\) relayf \(\rightarrow\) relayf \(\rightarrow\) write.

2. Specification of heterogeneous types of task per stage. This is achieved with -r (or --tsk-types) CLI parameter. For instance, -r ((init),(incrf,relay,incr),(fin)) will produce a 3-stage pipeline with the following sequence of tasks: init \(\rightarrow\) incrf \(\rightarrow\) relay \(\rightarrow\) incr \(\rightarrow\) fin.

3. Use of a scheduler to perform the pipeline decomposition in stages automatically. This is achieved with -C (or --chain) CLI parameter. For instance, -C "(init,relayf_15,incrementf_S_60,relay_15,fin)" defines a chain that starts with an initialize task, after that, a relayf task of 15 microseconds is executed, then an incrementf task of 60 microseconds is executed (note that the _S means that this task will be considered sequential and "non-replicable" for the scheduler). Finally, a 15 microseconds relay task and a finalize task are executed. By default the scheduler considers that the number of resources \(R\) is the number of CPU hardware threads but you can override this behavior by using the -t (or --n-threads) CLI parameter. It is also possible to choose the scheduler algorithm through the -S (or --sched) CLI parameter. For now, the only available scheduler is OTAC.

The first notation is a compressed way to describe chains of tasks. By default, the chain is split in pipeline stages according to the given decomposition (with -R and -r) and each stage is run on a separated thread. It is also possible to run the chain in a sequence (with the -q or --force-sequence CLI parameter). In this case, the given stage decomposition is ignored and all the tasks of the chain are run by the same thread.

Note

You cannot use -r and -R parameters at the same time, they are exclusive.

Note

If StreamPU has been compiled with the CMake -DSPU_LINK_HWLOC=ON option, then it is possible to specify an pinning policy with the -P or --pinning-policy CLI argument.

Tip

For the increment, incrementf, relay and relayf tasks it is possible to specify the duration. For instance, relay_12 means that the relay task will spend 12 microseconds in active waiting. This is different from using the -s CLI parameter. The -s parameter will set the same duration for all the previously mentioned tasks.

Moreover, for each stage it is possible to specify the number of replications (= number of threads that will execute the stage) with the -t (or --n-threads) CLI parameter. Here are some examples of generated pipelines:

3-stage pipeline with in/out sockets3-stage pipeline with forward sockets3-stage pipeline with hybrid socketsComplex 5-stage pipeline

test-generic-pipeline: input/output sockets & 3-stage pipeline.

test-generic-pipeline -e 100 -n "1,3,1" -t "3,1,3" -R "(init,increment,fin)"

test-generic-pipeline: forward sockets & 3-stage pipeline.

test-generic-pipeline -i INPUT_FILE -n "1,3,1" -t "1,3,1" -R "(read,relayf,write)"

test-generic-pipeline: hybrid in/out and forward sockets & 3-stage pipeline.

test-generic-pipeline -i INPUT_FILE -t "1,3,1" -r "((read),(relayf,incrementf,relay),(write))"

test-generic-pipeline: hybrid in/out and forward sockets & 5-stage pipeline.

test-generic-pipeline -e 100 -t "1,3,1,2,1" -r "((init,relayf,incr),(relayf,relay),(incrf),(relay),(relay,fin))"

---

Command Line Arguments

The following verbatim is a copy-paste from the -h stdout:

usage: ./bin/test-generic-pipeline [options]

  -t, --n-threads          Number of threads to run in parallel for each stage                   [empty]
  -f, --n-inter-frames     Number of frames to process in one task                               [1]
  -s, --sleep-time         Sleep time duration in one task (microseconds)                        [5]
  -d, --data-length        Size of data to process in one task (in bytes)                        [2048]
  -e, --n-exec             Number of executions (0 means -> never stop because of this counter)  [0]
  -l, --n-exec-pro         Number of executions during the scheduler profiling phase             [100]
  -u, --buffer-size        Size of the buffer between the different stages of the pipeline       [16]
  -o, --dot-filepath       Path to dot output file                                               [empty]
  -i, --in-filepath        Path to the input file (used to generate bits of the chain)           [empty]
  -j, --out-filepath       Path to the output file (written at the end of the chain)             ["file.out"]
  -c, --copy-mode          Enable to copy data in sequence (performance will be reduced)         [false]
  -b, --step-by-step       Enable step-by-step sequence execution (performance will be reduced)  [false]
  -p, --print-stats        Enable to print per task statistics (performance will be reduced)     [false]
  -g, --debug              Enable task debug mode (print socket data)                            [false]
  -q, --force-sequence     Force sequence instead of pipeline                                    [false]
  -w, --active-waiting     Enable active waiting in the pipeline synchronizations                [false]
  -n, --tsk-per-sta        The number of tasks on each stage of the pipeline                     [empty]
  -r, --tsk-types          The socket type of each task (SFWD or SIO)                            [empty]
  -R, --tsk-types-sta      The socket type of tasks on each stage (SFWD or SIO)                  [empty]
  -C, --chain              Description of the tasks chain (to be combined with '-S' param)       [empty]
  -S, --sched              Scheduler algorithm for the pipeline creation ('OTAC')                ["OTAC"]
  -v, --verbose            Show information about the scheduling choices                         [false]
  -h, --help               This help                                                             [false]