Repository navigation
master-964: automatic direct convolution selection (#2134) is ~7x slower end-to-end on RDNA2 (RX 6950 XT, Vulkan) #2137
Description
Machine
- AMD Radeon RX 6950 XT 16 GB (RDNA2 / Navi 21), Windows 11, official
sd-master-1d0a58a-bin-win-vulkan-x64.ziprelease binaries - Same-session A/B, idle machine (load gate checked), model file RAM-cached both legs (tensor load 0.2 s on both)
Repro
sd-cli.exe -m juggernautXL_ragnarok.safetensors --steps 20 --cfg-scale 7 \
--sampling-method euler_a --seed 42 -W 512 -H 512 \
-o out.png --prompt "a majestic gothic cathedral on a cliff at sunset, highly detailed"
| Binary | Flags | generate_image |
|---|---|---|
| master-959 (d637cfb) | default | 10.9 s |
| master-964 (1d0a58a) | default (auto selections) | 73.4 s (6.7x) |
| master-964 (1d0a58a) | --conv-direct off |
9.5 s (faster than 959) |
Diagnosis
master-964 logs, on this GPU:
[I] flash attention auto: Vulkan0 -> off (probe: normal 0.618 ms, flash 5.565 ms) --- flash_attn.cpp:95
[I] conv2d direct auto: Vulkan0 -> on (implicit GEMM) --- conv_direct.cpp:42
[I] conv3d direct auto: Vulkan0 -> on (implicit GEMM) --- conv_direct.cpp:42
The auto-probe for #2134 selects implicit-GEMM conv2d, but end-to-end that path is
catastrophically slower on RDNA2: forcing --conv-direct off restores (and beats)
master-959 throughput. The isolated per-op probe latency does not predict the full
pipeline behavior on this hardware.
Note: --flash-attn auto -> off seems fine here (9.5 s leg ran with flash off); the
regression is entirely the conv2d/conv3d direct selection.
Suggestion
Consider making the conv-direct auto probe more representative (e.g. timed over a few
layers with realistic tensor sizes, or a short end-to-end warmup), and/or documenting
--conv-direct off as the fallback for RDNA2 in the README while the probe is tuned.
Happy to test candidate fixes on this GPU.
Activity
laurentvv commentedon Oct 15, 2026
Draft comment for leejet/stable-diffusion.cpp #2137 (conv-direct RDNA2 regression)
Fixed on master-968 (8cd731e) — #2141 benchmarks the direct convolution in auto mode and now rejects it on RDNA2. Thank you.
Datapoint (same machine as the issue: RX 6950 XT, Windows, official
sd-master-*-bin-win-vulkan-x64.zipbinaries, SDXL 512×512 20 steps cfg 7 seed 42, same-session A/B, model RAM-cached):conv2d direct auto: Vulkan0 -> on(implicit GEMM — ~7× slower end-to-end)generate_image completed in 9.52s)conv2d direct auto: Vulkan0 -> off (probe output check failed)+ same for conv3dThe output-check added by #2141 catches the bad probe on this GPU and auto falls back to the correct path — no
--conv-direct offoverride needed anymore. Flash-attn auto→off remains correct (probe: normal 0.673 ms vs flash 5.600 ms).Qwen-Image-2.1-Turbo real path (workflow, 1152×640, 8 steps) also re-validated on 968: exit 0, 123 s wall (959 reference ~112 s gen — parity), render visually clean.
From our side #2137 can be closed as fixed by #2141.