Curiosity of performance between host_memory_bandwidth and host_memory_bandwidth_xrt

[anonymous1782]

Hi,

I have a question about the performance between host_memory_bandwidth and host_memory_bandwidth_xrt.

It seems that they have same kernel and similar host code, but the performance results are quite different.

For buffer size 32.00 KB example,

OpenCL host example is

Concurrent Read and Write Throughput = 12.1802 (GB/sec) for buffer size 32.00 KB
Read Throughput = 8.17289 (GB/sec) for buffer size 32.00 KB
Write Throughput = 8.26587 (GB/sec) for buffer size 32.00 KB

,while XRT host example is

Concurrent Read and Write Throughput = 12.4713 (GB/sec) for buffer size 32.00 KB
Read Throughput = 0.00397313 (GB/sec) for buffer size 32.00 KB
Write Throughput = 0.0040426 (GB/sec) for buffer size 32.00 KB

What is the main factor contributing to this difference?
Do XRT and OpenCL affect the performance significantly?

[vishnuchebrolu]

Hi @anonymous1782 ,

Can you please share the Vitis and XRT version that are being used along with Platform that is targetted.

Thanks,
Vishnu

[anonymous1782]

Hi @vishnuchebrolu

Hmm.. both results shown in the post are not my own. The results are from the README.md file in this repository. (OpenCL, XRT)

By the way, I'm using vitis v2021.2 and xrt v 2.12.427.

Thanks,