Emulator

Thank you! You were right, that 5.0-15289 doesn't have that issue with the bloom cutting off on the bottom half of the screen.

I've attached a few different fifo logs for you. I don't know if changing Dolphin's settings affects it, so I just included them anyway.

Changing dolphin settings sometimes can be baked into the fifolog, but it seems like that isn't the case here, which is nice because it means I can try out different configurations easily.

I've used goldeneye_efb_to_texture_only for everything as it has a good angle. I also used the hardware fifoplayer to take screenshots (both on real hardware and on dolphin for consistency), and tried out a mix of settings (Nvidia GPU+Vulkan, Intel GPU+OpenGL+framebuffer fetch, and the software renderer; and also toggling Manual Texture Sampling, store EFB copies to texture only, and safe/medium texture cache). All of these are attached.

My main observation is that manual texture sampling mostly improves things, and the fbfetch path (which is not finished) also helps a lot, but even with that it's still a bit too strong. The software render is also brighter than it should be (across the whole image).

Additionally, the texture cache setting only affects things when "store EFB copies to texture only" is enabled (i.e. it only matters when bloom is drawn over only half of the screen). The "safe" version in this case is different from what's drawn in every other case, and also doesn't seem to match real hardware. It's odd that "safe" gives less accurate results, but I'm guessing it's some kind of weird texture cache issue.

One other oddity is that "VideoCommon\BPStructs.cpp:274 W[Video]: Oversized EFB copy: 160x112 (offset 0,448 stride 2560)" is repeatedly logged. This means that the game's trying to make an EFB copy from y=448 to 560 when the EFB itself is from y=0 to y=528. This is out of bounds, but not down the middle of the screen. I also notice that the game draws the bloom effect in the region from t=448 to y=528 (but it actually splits it into two sections in some cases, though not all, and it also seems to be doing trickery to merge the two textures together where they're just located after each other in memory; that's probably what's breaking the store to texture only path).