The final release has replaced the August release candidate. RISC-V conformance is checkable against named configurations, while device selection and workload performance remain separate questions.
The release candidate is no longer the current stopping point
PoCL published 7.2 final on September 4, after the August 7 release candidate. The project reports OpenCL 3.0 CPU conformance and adds extensions including integer dot product and suggested local work size. The former advice to wait for final 7.2 is therefore obsolete.
The Khronos register independently lists Star64 with a StarFive JH7110 CPU and Milk-V Jupiter with a SpacemiT K1 CPU, using a named PoCL 7.2-pre revision. Entries include the device type, driver version, CTS version and compilation path. Conformance is tied to those configurations; it is not a blanket certification of every build for every RISC-V board.
The CPU is the compute device in these entries
An OpenCL application chooses a platform and a device. PoCL’s CPU driver executes kernels on processor cores; choosing it does not select a graphics processor merely because the board also contains one. Our diagram traces this selection boundary rather than depicting a GPU acceleration result.
In an environment with several OpenCL implementations installed, record the platform name, device name, device type and driver version before comparing runs. Otherwise a faster result might simply come from a different selected device.
The OpenCL specification also defines optional capabilities. Query the features and extensions the application actually requires instead of treating the version string as a promise of every operation. A suggested work-group size is a starting point to evaluate, not proof of the optimal setting for all inputs.
Measure the full workload after checking correctness
Suppose a small job takes little time to execute but spends most of its first run compiling a kernel. A cold-start result and a repeated run with cached compilation answer different questions. Report both when startup matters, and compare against the same reference output with the relevant numerical tolerance.
For a long-running workload, separate initialization, compilation, data movement and kernel execution. Do not subtract inconvenient stages from one implementation while retaining them for another. No timing or board test is claimed here: this is a method for deciding whether the conformant CPU path is useful for your application, beyond the existence of the badge.
September 8: update to the September 4 final release, verify named RISC-V conformance entries, and distinguish the OpenCL device from the board GPU and workload performance.