Radeon ROCm 2.1 Bringing RocTracer Preview, Improved DGEMM Performance
Just prior to closing out 2018 they released the big ROCm 2.0 update while now ROCm 2.1 is being prepared for release based upon public Git activity by AMD developers.
This next update to the Radeon open-source/Linux compute stack includes a preview release of RocTracer 1.0 for HSA API and HSA async GPU activity tracing. In addition to the RocTracer 1.0 preview is also support for monitoring the real-time PCI Express bandwidth via the ROCM-SMI tool (there is the change coming in Linux 5.1 to expose AMDGPU PCIe bandwidth via sysfs for upstream), DGEMM optimizations, and other changes. The improved DGEMM performance is said to be for large square and reduced matrix sizes. ROCm 2.1 is also timed quite nicely for the new Radeon VII.
There doesn't appear to be any notable changes on the ROCm OpenCL front, such as allowing SPIR-V support. It's also not mentioned if they have addressed any of the performance shortcomings in select cases compared to their Radeon PAL OpenCL driver.
Look for the ROCm 2.1 binaries to become available shortly and sources via GitHub.
8 Comments
