Skip to content

Latest commit

 

History

History
80 lines (64 loc) · 2.94 KB

README.md

File metadata and controls

80 lines (64 loc) · 2.94 KB

SYCL Academy

Exercise 10: Managing Dependencies

In this exercise you will learn how to create a data dependency data flow graph in both the buffer/accessor and USM models.

1.) Define a data flow graph with the buffer/accessor model

Using everything you have learned in previous exercises create an application using the buffer/accessor model which has four kernel functions. These kernel functions can do any computation you like, but they should follow the following dependencies.

      (kernel A)
     /          \
(kernel B)  (kernel C)
     \          /
      (kernel D)

The important thing here is that kernels B and C must depend on kernel A, kernel D must depend on kernels B and C and kernels B and C can be executed in any order and even concurrently if the device permits. Note that in the buffer/accessor these dependencies are created implicitly using the creation of accessors.

Feel free to use any method of synchronization and copy back you like, but remember to handle errors.

2.) Define a data flow graph with the USM model

Now do the same again but using the USM model. Note that in the USM model dependencies are defined explicitly by chaining commands via events.

Again feel free to use any method of synchronization and copy back you like, but remember to handle errors.

Build And Execution Hints

For DPC++: Using CMake to configure then build the exercise:

mkdir build
cd build
cmake .. "-GUnix Makefiles" -DSYCL_ACADEMY_USE_DPCPP=ON -DSYCL_ACADEMY_ENABLE_SOLUTIONS=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
make exercise_10

Alternatively from a terminal at the command line:

icpx -fsycl -o sycl-ex-10 -I../External/Catch2/single_include ../Code_Exercises/Exercise_10_Managing_Dependencies/source.cpp
./sycl-ex-10

In Intel DevCloud, to run computational applications, you will submit jobs to a queue for execution on compute nodes, especially some features like longer walltime and multi-node computation is only available through the job queue. Please refer to the guide.

So wrap the binary into a script job_submission and run:

qsub job_submission

For AdaptiveCpp:

# <target specification> is a list of backends and devices to target, for example
# "omp;generic" compiles for CPUs with the OpenMP backend and GPUs using the generic single-pass compiler.
# The simplest target specification is "omp" which compiles for CPUs using the OpenMP backend.
cmake -DSYCL_ACADEMY_USE_ADAPTIVECPP=ON -DSYCL_ACADEMY_INSTALL_ROOT=/insert/path/to/adaptivecpp -DACPP_TARGETS="<target specification>" ..
make exercise_10

alternatively, without CMake:

cd Code_Exercises/Exercise_10_Managing_Dependencies
/path/to/adaptivecpp/bin/acpp -o sycl-ex-10 -I../../External/Catch2/single_include --acpp-targets="<target specification>" source.cpp
./sycl-ex-10