Skills

rdc-cli Skill

Overview

rdc-cli is a Unix-friendly command-line interface for RenderDoc GPU captures. It provides a daemon-backed architecture using JSON-RPC over TCP, a virtual filesystem (VFS) path namespace for navigating capture internals, and composable commands designed for shell pipelines, scripting, and CI assertions.

Install: pip install rdc-cli (requires a local RenderDoc build with Python bindings). Check setup: rdc doctor.

Core Workflow

Follow this session lifecycle for any capture analysis task:

1. Open a capture:
   • Local: rdc open path/to/capture.rdc
   • Remote replay (Proxy): rdc open capture.rdc --proxy host:port
   • Split thin-client: rdc open --connect host:port --token TOKEN
   • Android device: rdc open capture.rdc --android [--serial SERIAL]
2. Inspect metadata: rdc info, rdc stats, rdc events
3. Navigate the VFS: rdc ls /, rdc ls /textures, rdc cat /draws/0/pipeline/summary
4. Analyze specifics: rdc shaders, rdc pipeline, rdc resources, rdc bindings
5. Debug shaders: rdc debug pixel EID X Y, rdc debug vertex EID VTXID, rdc debug thread EID GX GY GZ
6. Export data: rdc texture ID -o out.png, rdc rt EID, rdc buffer ID -o buf.bin, rdc cbuffer EID --stage ps --binding 0, rdc log
7. Close the session: rdc close

Session Management

• Default session name: default (or value of $RDC_SESSION).
• Override per-command: rdc --session myname open capture.rdc.
• Check active session: rdc status.
• Navigate to a specific event: rdc goto EID.

Output Formats

All list/table commands default to TSV (tab-separated values) with a header row, suitable for cut, awk, and sort.

Flag	Format	Use Case
(default)	TSV with header	Human reading, shell pipelines
--no-header	TSV without header	Piping to awk/cut without stripping
--json	JSON array	Structured processing with jq
--jsonl	Newline-delimited JSON	Streaming processing, large datasets
-q / --quiet	Minimal (single column)	Extracting IDs for loops

Example -- get all draw call EIDs as a plain list:

rdc draws -q

Example -- JSON pipeline with jq:

rdc events --json | jq '.[] | select(.type == "DrawIndexed")'

Render Pass Analysis

List passes (Phase 8 columns)

rdc passes outputs 6 columns: NAME, DRAWS, DISPATCHES, TRIANGLES, BEGIN_EID, END_EID.

rdc passes                          # TSV table
rdc passes --json                   # includes load_ops/store_ops per pass
rdc passes --deps --table           # per-pass READS/WRITES/LOAD/STORE

Inspect a single pass

rdc pass <name> shows enriched attachments: resource name, format, dimensions, and load/store ops.

rdc pass GBuffer
rdc pass GBuffer --json
rdc pass 0                          # by 0-based index

Detect dead render targets

rdc unused-targets finds render targets written but never consumed by visible output. Columns: ID, NAME, WRITTEN_BY, WAVE.

rdc unused-targets                  # TSV
rdc unused-targets --json           # structured
rdc unused-targets -q               # one resource ID per line (for scripting)

Frame statistics

rdc stats outputs three sections: Per-Pass Breakdown, Top Draws by Triangle Count, and Largest Resources.

rdc stats                           # all three sections
rdc stats --json                    # includes largest_resources array

GL/GLES/D3D11 captures without native BeginPass/EndPass markers get synthetic pass inference automatically — no extra flags needed.

Common Tasks

Find all draw calls

rdc draws
rdc draws --pass "GBuffer" --json

Trace a pixel

rdc debug pixel 1024 512 384            # EID first, then X Y
rdc debug pixel 1024 512 384 --json     # structured output
rdc debug pixel 1024 512 384 --trace    # full step-by-step trace

Search shaders by name or source

rdc search "main"                  # regex search over shader disassembly
rdc search "Sample" --stage ps -C 2
rdc shaders --name "GBuffer*"

Export render targets

rdc rt EID -o output.png
rdc rt EID --depth -o depth.png      # export the raw depth attachment
rdc texture ID -o tex.png            # export a texture by resource ID (PNG)

Decode a constant buffer

rdc cbuffer EID --stage ps --binding 0           # decode to JSON
rdc cbuffer EID --stage vs --binding 0 --raw -o cbuffer.bin

Browse VFS paths

rdc ls /
rdc ls /textures -l
rdc tree /draws/42/pipeline --depth 2     # pipeline state lives under /draws/<eid>/pipeline/
rdc cat /draws/42/pipeline/summary
rdc cat /events/42

Inspect pipeline state at a draw call

rdc goto EID
rdc pipeline --json
rdc bindings --json

Compare state before/after a pass

rdc goto 100 && rdc pipeline --json > before.json
rdc goto 200 && rdc pipeline --json > after.json
diff before.json after.json

CI Assertions

rdc-cli provides assertion commands that exit non-zero on failure, designed for automated testing pipelines:

Command	Purpose
rdc assert-pixel EID X Y --expect "R G B A"	Assert pixel RGBA (4 space-separated floats) at EID/coordinates
rdc assert-clean	Assert no validation errors in capture
rdc assert-count <what> --expect N [--op CHOICE]	Assert a capture metric (e.g. draws) satisfies a comparison
rdc assert-state EID KEY-PATH --expect VALUE	Assert pipeline state value at EID matches expected
rdc assert-image EXPECTED ACTUAL [--threshold FLOAT]	Compare two image files pixel-by-pixel

Example CI script:

#!/bin/bash
set -e
rdc open test_capture.rdc
rdc assert-clean
rdc assert-count draws --expect 10 --op ge
rdc assert-pixel 1024 256 256 --expect "1.0 0.0 0.0 1.0"
rdc close

Shader Edit-Replay

Modify and replay shaders without recompiling the application:

rdc shader-encodings                       # list available encodings
rdc shader EID ps --source > s.frag        # extract shader source
# ... edit s.frag ...
rdc shader-build s.frag --stage ps         # compile; prints the built shader ID
rdc shader-replace EID ps --with <ID>      # hot-swap built shader ID into capture
rdc shader-restore EID ps                  # revert single shader (STAGE required)
rdc shader-restore-all                     # revert all modifications

Remote Capture Workflow

rdc-cli wraps renderdoccmd remoteserver to support PC-to-PC remote captures.

• rdc serve [--port PORT] [--allow-ips CIDR] [--no-exec] [--daemon] — launch remoteserver on the target machine
• rdc remote connect <host:port> — save remote connection state
• rdc remote list — enumerate capturable apps on the remote
• rdc remote capture <app> -o frame.rdc [--args ...] [--frame N] [--keep-remote] — inject, capture, and transfer back. --keep-remote skips the transfer and prints the remote path; replay it with rdc open <path> --proxy host:port. (The CLI's own next: hint currently still references the deprecated --remote alias for --proxy.)
• rdc open frame.rdc --proxy host:port — remote-backed replay (daemon local, GPU remote)

remote_state.py persists the last connected host so subsequent rdc remote list can omit --url.

Split Mode (thin client)

Split mode decouples CLI and daemon — run the daemon where the GPU is and connect from a machine that doesn't need the renderdoc module. Useful when the analyst's laptop is macOS/Windows and the GPU is on a Linux server.

• Server side: rdc open capture.rdc --listen [ADDR[:PORT]]
  • Prints these four labeled lines to stdout (among other status output): host: ADDR, port: PORT, token: TOKEN, connect with: rdc open --connect ADDR:PORT --token TOKEN
• Client side: rdc open --connect HOST:PORT --token TOKEN

SSH tunnel tip (use the port from --listen, or rdc serve's default 39920): ssh -L 39920:localhost:39920 user@server, then connect to localhost:39920.

Every normal command (rdc draws, rdc rt, ...) works transparently in Split mode. Binary exports use file_read RPC with raw binary frames — no base64 overhead.

Android Workflow

• Prerequisite: the RenderDoc APK must already be installed on the host via rdc setup-renderdoc --android (upstream) or --android --arm (ARM PS fork for Mali). rdc android setup does not push the APK itself.
• rdc android setup [--serial SERIAL] — starts remoteserver on the device via RenderDoc's Device Protocol API (StartRemoteServer), sets adb forward, saves remote state.
• rdc android capture <activity> [--serial SERIAL] [--timeout N] [--port PORT] [-o out.rdc] — GPU debug layers based capture (works around EMUI/Mali injection limitations).
• rdc android stop [--serial SERIAL] — stops the remoteserver and cleans state.
• For remote replay: rdc open frame.rdc --android [--serial SERIAL] — this is the only form that rewrites the saved adb://SERIAL to the forwarded localhost:PORT. Passing --proxy adb://SERIAL directly bypasses the rewrite and is known to crash the daemon (see session.py:_resolve_android_url).

Hardware matrix: Adreno is the happy path; Mali may need the ARM Performance Studio fork (see rdc setup-renderdoc --android --arm).

Troubleshooting

Always run rdc doctor first. It reports status for renderdoc module, renderdoccmd, adb, Android APK, and platform-specific toolchains. Only the missing-renderdoc-module case emits a dedicated build-hint block; other checks surface inline hints in the detail column, so read each failing line rather than relying on a uniform next-step list.

Common failure categories (conceptual, not literal error strings — map from the text the tool actually emits):

• network / connect failed — remote host unreachable, firewall, wrong port. Verify rdc serve is running on the target.
• version mismatch — host and target RenderDoc versions differ. Re-run rdc setup-renderdoc or rdc setup-renderdoc --android to align.
• inject failed / ident=0 — injection blocked (Android EMUI, macOS SIP, Windows privilege). Run rdc doctor and check the platform-specific detail.
• OpenCapture unsupported — local GPU can't replay the capture's API surface; switch to --proxy or --android remote replay.
• duplicate Vulkan layer — rdc doctor warns when more than one VK_LAYER_RENDERDOC_Capture Vulkan layer is registered (a system-installed RenderDoc plus the rdc-managed one create a split-brain). The loader then picks one non-deterministically and capture can silently time out; unregister the extra manifest so only one remains.
• not loaded / no session — forgot rdc open; use rdc status to inspect.

For long operations (large capture transfers, remote replay init), the CLI has limited progress feedback — this is a known UX gap, not a hang. Wait up to the --timeout value before concluding failure.

Command Reference

For the complete list of all commands with their arguments, options, types, and defaults, see references/commands-quick-ref.md.