This tutorial builds on the Subgroups concepts demonstrated in the Subgroups Visualization tutorial. In this tutorial you will use Subgroup functions to improve the performance of a simple 8-bit image copy.
This tutorial requires support for the cl_intel_subgroups extension. Because the
application compiles for an OpenCL 2.0 device (using -cl-std=CL2.0) by default, some
modifications may be needed to run it on an OpenCL 1.2 device.
Intel Subgroup capabilities are exposed by the Intel cl_intel_subgroups
extension. The Intel Subgroups extension builds on the core Subgroup capabilities by adding
generalized data interchange functions (intel_sub_group_shuffle and others), and block
reads and writes, for both buffers and images (intel_sub_group_block_read and intel_sub_group_block_write).
The goal of this tutorial is improve the performance of a simple kernel that copies a
single-channel 8-bit image using Subgroup block reads and writes. You will do this by
editing subgroup_imagecopy_kernel.cl.
By default, the kernel in this file loads an image using the standard OpenCL read_image
and write_image built-in functions. Four pixels are copied per work item, so four
calls are needed to read the data, and four calls are needed to write the data.
Change the kernel in this file to copy the image using Subgroup block reads and writes instead. Note that the Subgroup block reads and writes process 32-bits of data at a time. Since the image is a single-channel 8-bit image, this means that only one call is required to read the block of data, and only one call is needed to write the data.
By default, the application copies the image 1000 times, and prints how long it took to execute. Pay most attention to the Image Copy Execution time, not the Total time, because the Total time, because the Total time also includes the time spent reading the input image from a file and writing the output image to a file. Do you see a performance improvement using Subgroup block reads and writes?
Here are some additional ideas to experiment with:
- Use the flexible block reads and writes from cl_intel_media_block_io to perform the copy more flexibly. Hint: It is easier to use the flexible block reads and writes if you use the "required Subgroup size" kernel attribute to specify a fixed Subgroup size.
- Modify the kernel and host code to process multiple image rows per work item. Does this improve performance even more?
- Don't just copy the image, instead flip it vertically (easier) or horizontally (harder). Hint: To flip the image horizontally, you'll likely want to use Subgroup shuffles to flip the data within the Subgroup.
subgroups_imagecopy
Intel and the Intel logo are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries.
*Other names and brands may be claimed as the property of others.
OpenCL and the OpenCL logo are trademarks of Apple Inc. used by permission by Khronos.