ui: GpuCompute: correct the Summary tab and rename Workload Analysis - #7242
Open
zytyz wants to merge 4 commits into
Open
ui: GpuCompute: correct the Summary tab and rename Workload Analysis#7242zytyz wants to merge 4 commits into
zytyz wants to merge 4 commits into
Conversation
The section holds per-pipeline compute utilization only. The plain name "Workload Analysis" does not say which part of the workload it covers, and it collides with a memory counterpart that vendor plugins register alongside it. The file name, the section id and the registration function follow the title. The metric rows and the analysis prompt do not change.
The GpuCompute details view calls this lens Compute Workload Analysis. The agent skill still called it Workload Analysis, so an agent and a user read two names for one section. The generic workflow, the NVIDIA extraction and its query file take the new name, and every link and prose reference follows. The interpretation rules and the SQL do not change.
The Summary tab builds its query from the well-known metric roles. It runs that query at the end of onTraceLoad. A plugin that depends on GpuComputePlugin registers its own metric ids in its own onTraceLoad, which the plugin manager runs later. Those ids are therefore absent from the query, and the compute-throughput and memory-throughput columns stay empty for that plugin's hardware, even though the details tables show the values. The query now runs on onTraceReady, after every plugin has registered. The cached rows start as undefined, so a tab that opens before the query ends runs its own query instead of keeping an empty table.
The Summary tab draws every column against the largest value in the table. That suits duration, register count and grid size, which have no upper bound. It does not suit the two throughput columns, which are already a percentage of a hardware peak. The error is invisible on a busy workload and severe on an idle one. On an MTIA capture whose memory throughput spans 0.03 to 0.93 percent of peak, a kernel at the trace mean of 0.47 percent fills 51 percent of its bar, because it is measured against the busiest kernel in the trace rather than against the hardware. The two throughput columns now render against 100, and label the bar with the value and a percent sign, so that kernel reads "0.47%" on a bar filled to 0.47 percent.
Contributor
Author
|
@dreveman ptal. Thanks! |
dreveman
approved these changes
Aug 27, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Small fixes so a vendor plugin can register its own metrics into the
GpuCompute tab.
Fix the Summary tab so a dependent plugin's metric ids reach it. The
prefetch ran before those plugins had registered, so the columns came out
empty and stayed empty.
Scale the compute and memory throughput bars against the hardware peak
rather than the largest value in the table, and label them with a percent
sign. On a capture spanning 0.03% to 0.93% of peak, a kernel at the mean
0.47% filled 51% of its bar, and the busiest always pinned at 100%.
Rename "Workload Analysis" to "Compute Workload Analysis". It covers
per-pipeline compute utilization only, and the plain name collides with the
memory counterpart a vendor plugin registers alongside it.
Rename the matching workflow in the Perfetto agent skill, so an agent and a
user read one name for one section.