Move fetched texture preparation to fetch worker

[LisaScheers]

Stack

Depends on #6031. This is a stacked follow-up; the follow-up change is the commit Move fetched texture preparation to fetch worker. The diff against develop will shrink when #6031 merges.

What changed

• Queue alpha histogram and pick-mask preparation on the texture fetch worker after decode.
• Carry preparation beside the exact decoded raw buffer through the fetch handoff and skip the redundant General-queue job.
• Validate or invalidate preparation for stale, scaled, and local buffers while retaining the existing fallback.
• Request early preparation only for eligible upload textures.

Why

analyzeAlpha() can take multiple milliseconds. #6031 removes it from the frame-critical GL path; this follow-up performs the work earlier on the fetch worker so main-thread upload scheduling does not need another preparation hop.

Validation

• ReleaseOS universal viewer configured with Xcode and LL_TESTS=ON.
• INTEGRATION_TEST_llimagegl_prepare: 4/4 passed.
• Full secondlife-bin universal Release target: passed.

[Copilot]

Pull request overview

This PR advances the texture pipeline optimization stack by moving CPU-side upload preparation (alpha classification and pick-mask generation) earlier into the texture fetch worker flow, reducing main-thread scheduling overhead and avoiding redundant “General” queue preparation hops.

Changes:

• Threads LLImageGL::TextureUploadPreparation alongside decoded raw buffers through LLTextureFetch → LLViewerFetchedTexture to apply preparation directly before GL texture creation.
• Adds LLImageGL::prepareForUpload() plus plumbing to apply/discard prepared metadata and skip re-running analyzeAlpha() / updatePickMask() when prepared.
• Introduces an integration test validating preparation behavior (including odd dimensions) and wires it into llrender when LL_TESTS is enabled.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
indra/newview/llviewertexture.h	Extends fetched-texture creation scheduling APIs to accept precomputed upload preparation metadata.
indra/newview/llviewertexture.cpp	Applies/invalidates preparation during fetch-result processing and texture creation scheduling; retains a fallback worker path when needed.
indra/newview/lltexturefetch.h	Extends fetch APIs to return and optionally compute upload preparation on the fetch side.
indra/newview/lltexturefetch.cpp	Stores preparation in fetch workers, schedules preparation on the fetch request queue after decode, and hands results back to consumers.
indra/llrender/llimagegl.h	Introduces TextureUploadPreparation struct and new APIs for preparation/apply/discard.
indra/llrender/llimagegl.cpp	Implements preparation logic, consumes prepared results in setImage(), and updates pick-mask sizing math.
indra/llrender/tests/llimagegl_prepare_test.cpp	Adds coverage for alpha classification and pick-mask packing, including odd width/height cases.
indra/llrender/CMakeLists.txt	Registers the new integration test under LL_TESTS.

Comments suppressed due to low confidence (1)

indra/newview/lltexturefetch.cpp:2842

• Same implicit-conversion concern as above: pass raw_image.get() explicitly when calling callbackUploadPreparation() in the fallback path.

        LLTextureFetchWorker* worker = getWorker(id);
        if (worker)
        {
            worker->callbackUploadPreparation(raw_image, {});
        }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[akleshchev]

This doesn't seem to work right. I tried to upload a texture, scheduleUploadPreparation didn't get hit. I rezzed an abject and that triggered scheduleUploadPreparation. I'm not doing any uploading, yet I'm hitting upload preparation in scheduleCreateTexture.

At a minimum the naming is wrong and confusing (TextureUploadPreparation is no longer about upload), but from a 'mechanics' stadpoint, fetcher is already a thread with a bunch of workers, it should be issuing jobs to its own workers, not to the general pool. Or calculations can be done by/after decoder.