Lavender/flower plants draw white with a cyan tint instead of purple (NVIDIA) #674

Closed

opened 2026-11-08 16:37:08 +00:00 by xophsy · 2 comments

xophsy commented

2026-11-08 16:37:08 +00:00

(Migrated from github.com)

Builder version

v2.0.0 (source 5a5c7e6, with PR #477 merged locally). Also seen on the first-day build (cdb4f1a, no local changes).

How you built

From source: python -I nx2/make_game.py --nup ... with the v2.0.0 HookshotBuilder tools (llvm-mingw, RX_CPM, WGPU_NATIVE set).

Game files given to it

A .nup (build e0f3ccfa)

The error

Lavender / flower plants draw white with a cyan tint instead of purple. The leaves keep their shape but have no colour, as if the texture is not bound. Screenshots (broken / expected) will be added to this issue.

From hookshot_log.txt (v2.0.0 run):

0 skipped draws, no draws left out once the cache is warm.
prepared 31985 shader programs (39835 new, 7 failed, ...): 7 stage-5 programs do not lower, unsupported native opcode (native op 0x3a9, ...) (an ATOMG.CAS-like form). The plants were already white on the first-day build with 0 failed, so these are probably unrelated.

Settings: Ultra preset, frame generation off, resolution 100%, Lanczos, sharpness 40, RX_VK_SYNC_PIPELINES=1.

Operating system and free space on the output drive

Windows 11 (26300), RTX 4080 SUPER 16 GB, 2560x1440 at 359 Hz. About 700 GB free.

Confirmations

On the newest release, and it still happens there.
Open issues searched before filing (related: #35, #203, #11).
No game files or anything built from them is attached.

### Builder version v2.0.0 (source `5a5c7e6`, with PR #477 merged locally). Also seen on the first-day build (`cdb4f1a`, no local changes). ### How you built From source: `python -I nx2/make_game.py --nup ...` with the v2.0.0 HookshotBuilder tools (llvm-mingw, `RX_CPM`, `WGPU_NATIVE` set). ### Game files given to it A .nup (build `e0f3ccfa`) ### The error Lavender / flower plants draw white with a cyan tint instead of purple. The leaves keep their shape but have no colour, as if the texture is not bound. Screenshots (broken / expected) will be added to this issue. From `hookshot_log.txt` (v2.0.0 run): - `0 skipped` draws, no draws left out once the cache is warm. - `prepared 31985 shader programs (39835 new, 7 failed, ...)`: 7 stage-5 programs do not lower, `unsupported native opcode (native op 0x3a9, ...)` (an ATOMG.CAS-like form). The plants were already white on the first-day build with `0 failed`, so these are probably unrelated. Settings: Ultra preset, frame generation off, resolution 100%, Lanczos, sharpness 40, `RX_VK_SYNC_PIPELINES=1`. ### Operating system and free space on the output drive Windows 11 (26300), RTX 4080 SUPER 16 GB, 2560x1440 at 359 Hz. About 700 GB free. ### Confirmations - On the newest release, and it still happens there. - Open issues searched before filing (related: #35, #203, #11). - No game files or anything built from them is attached.

Mitchdl93 commented

2026-11-09 00:53:42 +00:00

(Migrated from github.com)

Hi! I investigated this issue on Hookshot 2.0.5 and found a register-mapping collision in the Ampere-to-Maxwell shader lowering.

In LowerToMaxwellWith, marker registers are allocated while scanning native operands sequentially. A marker chosen for an earlier destination can overlap a native source register encountered later, causing distinct values to share a register mapping.

The attached patch reserves all native operand register ranges before allocating markers. It fixes the white/cyan flowers on my NVIDIA RTX 5070 Ti Laptop GPU without changing textures or overriding their colors.

Validation:

Two synthetic FADD regression cases fail with the original code and pass with the correction.
The existing DEPBAR lowering test still passes.
Flowers render purple at multiple distances and camera angles, including after regenerating and reloading the shader cache.

An enemy with a similar visual issue also appears fixed during further play, although that observation needs separate confirmation.

The attachment includes the source patch, synthetic regression test, documentation, and a before/after screenshot. It contains no game binaries/assets, shader caches, save files, logs, or local installation metadata.

Could you review the approach? This is a local fix and broader testing would be helpful.

Hookshot-register-fix.zip

Hi! I investigated this issue on Hookshot 2.0.5 and found a register-mapping collision in the Ampere-to-Maxwell shader lowering. In `LowerToMaxwellWith`, marker registers are allocated while scanning native operands sequentially. A marker chosen for an earlier destination can overlap a native source register encountered later, causing distinct values to share a register mapping. The attached patch reserves all native operand register ranges before allocating markers. It fixes the white/cyan flowers on my NVIDIA RTX 5070 Ti Laptop GPU without changing textures or overriding their colors. Validation: - Two synthetic FADD regression cases fail with the original code and pass with the correction. - The existing DEPBAR lowering test still passes. - Flowers render purple at multiple distances and camera angles, including after regenerating and reloading the shader cache. An enemy with a similar visual issue also appears fixed during further play, although that observation needs separate confirmation. The attachment includes the source patch, synthetic regression test, documentation, and a before/after screenshot. It contains no game binaries/assets, shader caches, save files, logs, or local installation metadata. Could you review the approach? This is a local fix and broader testing would be helpful. [Hookshot-register-fix.zip](https://github.com/user-attachments/files/34558794/Hookshot-register-fix.zip)

NotRexed commented

2026-11-09 09:18:17 +00:00

(Migrated from github.com)

Fixed in v2.0.6 (and in v2.0.7) by PR #922: a register renaming collision in the Ampere-to-Maxwell lowering left the flowers' colour undefined (the same cause @Mitchdl93 found, thanks for the analysis). Translated shaders are rebuilt on the next start after the update. If the flowers are still white on v2.0.7, comment with a screenshot and log and we'll reopen.

Commenting is not possible because the repository is archived.