Hardware day: capturing a real RDNA2 GPU corpus

The GPU probe was built to run on a Steam Deck (RDNA2, gfx1033) and diff what real silicon computes against the reference and against emulation. Everything below the capture step works today against llvmpipe; this is the recipe for the day a Deck (or, later, a PS5) is in hand.

Why the Deck needs no platform GPU API

The Deck is x86-64 Linux with an RDNA2 GPU and no vendor libraries. obSCEne's GPU backend there is ordinary public Vulkan (gpu_vulkan.c, the same one llvmpipe uses in the build VM) - so the Deck is reached with the host build, not the hardware module, and needs nothing from sceGnm. That is the whole reason the Deck is the near-term target: the machinery that runs on it is already built and proven on llvmpipe.

1. Build

make deck BUILD=/tmp/obs

Produces /tmp/obs/obscene-deck - the GPU=1 host build, renamed. Copy that one file to the Deck (scp, a USB stick, whatever reaches it); it is freestanding apart from libvulkan, which the Deck has.

2. Run and capture

On the Deck, either run it and keep stdout:

./obscene-deck > deck-report.txt

or serve it and drive from the dev machine (the Deck has no keyboard convenient for this):

# on the Deck
./obscene-deck --serve 9803
# on the dev machine, once, to pull the whole report over the socket
obscene-tool drive --address <deck-ip>:9803 --command 'hello|1' --command 'report' \
    --part target=deck --part gpu='AMD Custom GPU 0405 (gfx1033)' > deck-report.txt

Reduce the report to its GPU records - the corpus the analysis reads:

grep -E '^OBS\|(gpudev|gpu|gpuop)\|' deck-report.txt > deck-corpus.txt

The gpudev line will name the real device (... gfx1033 ... integrated), which is how every tool below knows this is silicon and not the cpu llvmpipe or the reference oracle.

3. Analyse

One command produces the reference for the corpus's inputs, the exact-match diff, and the ULP approximation ranking:

scripts/gpu-analyze.sh deck-corpus.txt

Read it as three answers:

exact ops must be empty in the diff - floor, divf, fma, the bit ops. A divergence there is a device bug, on RDNA2 as much as anywhere.
transcendentals appear in the diff and in the ULP ranking. The ranking is the finding: how far RDNA2's rcp, sin, rsq sit from correctly-rounded, worst first. This is the RDNA2 approximation map - the thing that was theoretical until the Deck produced it.

4. Diff against emulation

To find where an emulator disagrees with the Deck - the shader gap to close - keep both corpora and diff them directly:

obscene-tool gpudiff deck-corpus.txt emulator-corpus.txt   # exact divergences, by operand
obscene-tool gpustats deck-corpus.txt <(obscene-tool gpuref deck-corpus.txt)  # RDNA2's own error

gpudiff naming the operands where the two disagree is the actionable output: each is a lane an emulator computes differently from the hardware, which is exactly a shader-recompiler gap.

Regression, meanwhile

reports/archive/gpu-golden.txt is the blessed llvmpipe snapshot. That golden is llvmpipe's, not the Deck's - the check skips on a different device rather than failing. A Deck golden could be blessed on the Deck, but only once one exists to bless.