ROCprof Compute Viewer at a glance

2026-08-24

3 min read time

Applies to Linux

ROCprof Compute Viewer (RCV) is the visualization front end for Advanced Thread Trace (ATT) data: it renders decoded, instruction-level wavefront execution traces so you can see exactly which instruction stalled, for how long, and why. Use RCV once a coarser-grained tool (kernel-level counters or API traces) has identified which kernel is slow; RCV narrows the analysis down to the specific instruction responsible.

This topic orients you to how RCV fits into the profiling pipeline and which view to open for a given question. For step-by-step instructions, follow the links to the relevant how-to and reference topics.

How it works

RCV sits at the end of a three-stage pipeline:

This output path is independent of the general --output-format setting used for other rocprofv3 tracing. For how to collect the trace itself, see Using thread trace.

Input formats at a glance

RCV auto-detects which of the following you’re pointing it at; you don’t need to specify the format.

Format	When to use it	Details
rocprofv3 UI output (JSON)	Standard path — rocprofv3 --att already decoded the trace for you. Identified by a filenames.json file in the directory.	Import a rocprofv3 UI output directory
Raw .att/.out	Trace captured directly through the rocprofiler-sdk API; requires a decoder-enabled build.	Import raw .att and .out files

Views at a glance

Each view answers a different performance question. Full descriptions and controls for every view are in Views and controls.

View	Question it answers	Reference
Compute Unit	Which wavefronts ran, stalled, or idled, and when?	Compute Unit and Utilization view
Utilization	How busy was each hardware pipe (VALU, MFMA, VMEM, LDS)?	Compute Unit and Utilization view
Instructions (ISA)	Which instruction cost the most cycles, and is it waiting on memory?	Instructions view
Hotspot	What are the most expensive instructions, without scrolling the full ISA list?	Hotspot view
Flamegraph	Which code path dominates accumulated latency?	Flamegraph view
Global view	How were waves distributed across all CUs and selected shader engines over time?	Global view
Occupancy / Kernel dispatch	How many waves were resident, per shader engine or per kernel?	Occupancy and Kernel dispatch views
Counters	How did a specific SQ hardware counter trend over the run?	Counters view
Summary (MI200/MI300 only)	What’s the aggregate utilization and instruction cost across the whole trace?	Summary view

Getting started

RCV depends on AQLprofile, ROCprofiler-SDK, and Qt. The ROCprof Trace Decoder is an additional dependency only when importing raw .att/.out files. Prebuilt binaries are available in the RCV releases, so a build from source isn’t required to get started.

For supported versions and build instructions, see Building ROCprof Compute Viewer from source. For rocprofv3’s own prerequisites (AQLprofile, ROCprofiler-SDK, and the decoder it bundles), see the Requirements section.

Navigation

RCV’s mouse and keyboard controls (zoom, pan, measure cycle ranges, jump between a token and its ISA line) are listed in Views and controls.

See also

• Building ROCprof Compute Viewer from source — building RCV from source

• Visualize and analyze GPU thread trace data — importing traces and generating ISA/source correlation

• Views and controls — every view, its controls, and its shortcuts

• Hidden latency — how RCV separates hidden from non-hidden latency

• Troubleshooting — common issues