AMDGPU DC Gets "PERF_TRACE" To Help With Performance Profiling
Published on Wednesday was the latest batch of AMDGPU DC display code changes for its eventual inclusion into the AMDGPU DRM driver for mainline past the 4.20~5.0 cycle with that feature merge window being over. The most notable change with this latest AMDGPU DC haul is a new "PERF_TRACE" addition.
The 26 patches sent out on Wednesday refactor the DCE clock code as well as the DC to SMU interface. Most interesting to us though is this PERF_TRACE feature on Linux. This PERF_TRACE functionality isn't to be confused with the perf subsystem nor the perf-trace user-space utility.
This AMDGPU DC PERF_TRACE feature is for exposing performance data from the AMDGPU driver for helping its developers make performance improvements.
PERF_TRACE is basically a debug call to place before and after the AMDGPU specific code being tested. Calls to AMDGPU's PERF_TRACE will then post the time (in nanoseconds) to a new amdgpu_perf_trace debugfs file as well as the number of GPU reads/writes that happened in that span. It's a basic approach for allowing the driver developers to help understand their driver performance in a basic level but this feature isn't of any relevance to end-users.
Those wishing to test out the latest AMDGPU DC code in general for ensuring no display regressions can find the patches for now on the amd-gfx list.
The 26 patches sent out on Wednesday refactor the DCE clock code as well as the DC to SMU interface. Most interesting to us though is this PERF_TRACE feature on Linux. This PERF_TRACE functionality isn't to be confused with the perf subsystem nor the perf-trace user-space utility.
This AMDGPU DC PERF_TRACE feature is for exposing performance data from the AMDGPU driver for helping its developers make performance improvements.
PERF_TRACE is basically a debug call to place before and after the AMDGPU specific code being tested. Calls to AMDGPU's PERF_TRACE will then post the time (in nanoseconds) to a new amdgpu_perf_trace debugfs file as well as the number of GPU reads/writes that happened in that span. It's a basic approach for allowing the driver developers to help understand their driver performance in a basic level but this feature isn't of any relevance to end-users.
Those wishing to test out the latest AMDGPU DC code in general for ensuring no display regressions can find the patches for now on the amd-gfx list.
3 Comments