[Perf] (DirectGLES): upload the union dirty box instead of scatter rects when staging through the unpack ring (each PBO-sourced glTexSubImage is a GPU copy job on Mali; ~100 sprite rects per tick cost +6ms/frame of GPU time)

2026-09-07 19:58:32 +09:00 · 2026-08-26 05:52:31 -04:00
parent 56e5d13dc9
commit 33eabfc2ff
1 changed files with 9 additions and 0 deletions
@@ -4002,6 +4002,15 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                dirtyRectCount = textureMipmapObject->GetStorageDirtyRects(
                                    uploadTarget, level, dirtyRects,
                                    MG_State::GLState::MipmapStorage::kMaxDirtyRects);
+                                // The scatter refinement pays only on the client-pointer path,
+                                // where fewer bytes mean less driver-side copying. Through the
+                                // unpack ring every glTexSubImage is a GPU copy job (Mali), so
+                                // ~100 sprite rects become ~100 jobs whose fixed cost dwarfs
+                                // the union box's extra bytes - measured +6 ms/frame of GPU
+                                // time in MC's animated-atlas ticks. One box, one job.
+                                if (BufferImpl::UnpackRingAvailable()) {
+                                    dirtyRectCount = 0;
+                                }
                            }
                            const auto rectShadowPtr = [&](const MG_State::GLState::MipmapDirtyRegion& rect) {
                                return static_cast<const Uint8*>(uploadData) +