Sycl Academy LogoSycl Academy Logo
listShow Lessons List
  1. What is SYCL
  2. Enqueueing a Kernel
  3. Managing Data
  4. Handling Errors
  5. Device Discovery
  6. Data Parallelism
  7. Introduction to USM
  8. Using USM
  9. Asynchronous Execution
  10. Data and Dependencies
  11. In Order Queue
  12. Advanced Data Flow
  13. Multiple Devices
  14. ND Range Kernels
  15. Image Convolution
  16. Coalesced Global Memory
  17. Vectors
  18. Local Memory Tiling
  19. Further Optimisations
  20. Matrix Transpose
  21. More SYCL Features
  22. Functors
Lesson
Source
Solution
GitHub Iconedit

Coalesced Global Memory

In this exercise you will learn how to apply row-major and column-major when linearizing the global id in order to compare the performance difference due to coalesced global memory access.

1.) Evaluate global memory access

Now that you have a working image convolution kernel you should evaluate whether the global memory access patterns in your kernel are coalesced.

Consider two alternative ways to linearize the global id:

auto rowMajorLinearId    = (idx[1] * width) + idx[0];  // row-major
auto columnMajorLinearId = (idx[0] * height) + idx[1];  // column-major

Try using both of these and compare the execution time of each.

Build and execution hints

From the syclacademy directory

cd build/Code_Exercises/Coalesced_Global_Memory

and execute: * make Coalesced_Global_Memory_source - to build source.cpp * make Coalesced_Global_Memory_solution - to build the solution provided * make - to build both

Alternatively from a terminal at the command line:

icpx -fsycl -o Coalesced_Global_Memory_source -I../../Utilities/include/ -I../../External/stb ../Code_Exercises/Coalesced_Global_Memory/source.cpp

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][devcloud-job-submission].

So wrap the binary into a script job_submission

#!/bin/bash
./Coalesced_Global_Memory_source

and run:

qsub -l nodes=1:gpu:ppn=2 -d . job_submission

The stdout will be stored in job_submission.o<job id> and stderr in job_submission.e<job id>.

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 Coalesced_Global_Memory_source

Alternatively from a terminal at the command line:

icpx -fsycl -o Coalesced_Global_Memory_source -I../External/Catch2/single_include ../Code_Exercises/Coalesced_Global_Memory/source.cpp
./Coalesced_Global_Memory_source

For AdaptiveCpp:

# <target specification> is a list of backends and devices to target, for example
# "generic" compiles for CPUs and GPUs using the generic single-pass compiler.
# When in doubt, use "generic" as it usually generates the fastest binaries.
#
# Recent, full installations of AdaptiveCpp may not need targets to be provided,
# compiling for "generic" by default.
cmake -DSYCL_ACADEMY_USE_ADAPTIVECPP=ON -DSYCL_ACADEMY_INSTALL_ROOT=/insert/path/to/adaptivecpp -DACPP_TARGETS="<target specification>" ..
make Coalesced_Global_Memory_source

alternatively, without CMake:

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