README.md

SYCL_easy_tutorial

Overview

This is an easy SYCL tutorial written by jhson.
Feel free to read and use it. (but it is a little bit inefficient.)
SYCL device codes are inplemented as C++ template function in the directory "include/" The codes in the directory "test/" are host codes for using these device kernels.

0. How to Run

Environemnt

• clang 14.0.0 (https://github.com/intel/llvm.git 8c5b7017a925701ef4034056b5ed8e0fac2a0011)
• cmake 3.20.2
• cuda 10.2
• Intel CPU Runtime for OpenCL 18.1

Build steps

1. mkdir build && cd build
2. cmake ..
3. make

1. Device Querying Example

Very easy SYCL device querying example

2. Vector Addition Example

Vesy easy vector addition example implemented with SYCL DPC++.
No parallel optimization technique implemented.
A work item only performs a single element-wise addition.

3. Matrix Multiplication Example

Simple matrix multiplication example implemented with SYCL DPC++.
 C[M,N] = A[M,K] * B[K,N];
No parallel optimization technique implemented.
The work item with ID[m, n] only does dot product between the m-th row of A and the n-th colomn of B.

4. Matrix Stencil Operation Example

Simple stencil operation on matrix example implemented with SYCL DPC++.
This example performs correlation operation.
No parallel optimization technique implemented.

5. Parallel Reduction Example

Simple parallel reduction example implemented with SYCL DPC++.
Use the local memory and the sequential addressing optimization [1] technique

6. Memory Management Example

Sycl provide two memory management techniques: Unified shared memory (USM) and Buffer.
USM can use three different memory spaces, which are host, device, and shared. This example shows the performance of each technique.

7. Execution Model Example

References

[1] Harris, M. (2007). Optimizing parallel reduction in CUDA. NVIDIA Developer Technology.