rocprof studioSee what your GPU actually did

A Rust + Tauri desktop app and headless core for reading rocprofv3, ROCPD, ATT and RADV captures — with an MCP server and kernel/shader tuning when you ask for them.

Get started

MCP mode

rocprof studio — trace workspace (bundled synthetic example)

The rocprof studio trace workspace showing the bundled synthetic example

Built withRust + Tauri 2ROCm · rocprofv3MCP

The trace workspace timeline (synthetic example)

Timeline → wave → instruction → source

Follow a wave from hardware execution to the ISA that caused it and back to the source line.

Instruction hotspot histograms (synthetic example)

Instruction hotspots

Which instructions dominated, by hits and by attributed cost.

Occupancy

Per-SE and per-kernel wave residency over time, exportable as SVG.

Source-to-ISA flamegraph (synthetic example)

Source flamegraph

ISA cost rolled up through the source hierarchy.

Counter plots and expressions (synthetic example)

Counters

Scalar expressions and named-axis tensor definitions over recorded series.

Automatable over MCP

A loopback, bearer-token server with 31 tools for import, analysis, profiling and tuning.

Kernel tuning

Capture a kernel’s inputs, replay it in isolation, compare variants. Opt-in, needs a GPU.

Timeline detail aggregation menu (desktop app)

Detail on your terms

Exact events, upstream mip reduction or pixel aggregation — never silent sampling.

These screenshots use the bundled synthetic example

Most screenshots on this site were captured from public/demo.json, a synthetic trace shipped for exploring the interface. The app labels it "Example capture · Synthetic trace for exploring the interface. These are not GPU measurements." Treat every number in those images as illustrative, never as a hardware measurement. Screenshots taken from real recorded captures (ROCPD, ATT) are captioned accordingly.

Read this first

rocprof studio (package rocprof-studio, version 0.3.0, badged PREVIEW in the app) says no more than it can back up. Three things first:

Importing a capture is read-only. The import path never writes into your capture directories or databases.
Two surfaces are not read-only, and only run when you invoke them. The MCP tool profile_run (and profile_start) launches a program under a profiler; the kernel- and shader-tuning workers compile candidate variants and can write replacement source back to disk. See MCP mode, Kernel tuning and Shader tuning.
This is not full ROCprof Compute Viewer parity. The project tracks ROCprof Compute Viewer as its baseline and is not complete against it. It is also not a pixel-identical Qt reimplementation, and it has no LLM source-iteration feature despite the "Kernel experiments" placeholder in the sidebar.

What it is

It inspects ROCm GPU profiling captures. A single Rust analysis core (crates/profile-core) does the importing and analysis; a Tauri 2 shell (src-tauri) wraps it with a React/TypeScript UI (src), and the same core also runs as a headless importer CLI and behind the MCP server. You point it at the artifacts a profiler already produced and it reconstructs the timeline, the per-wave execution, the ISA, the source linkage, occupancy, counters and sampling, then lets you export evidence about what you saw.

It reads decoded rocprofv3 directories, kernel-trace and counter CSVs, SPM and PC-sampling JSON, ROCPD/legacy databases, raw ATT and RADV .rgp — all through one analysis core shared by the GUI and the headless CLI. Synthetic values are labelled, "not collected" is shown distinctly from an observed zero, and summed-duration vs interval-union accounting is kept separate.

If you are new, start with Installation, then Running & modes, Input formats and the UI tour. If you came for automation, go straight to MCP mode.