Remove constraint that virtual framebuffers have to represent VRAM. #11553
Conversation
hrydgard
changed the title
hrydgard
changed the title
|
This seems to work now. |
GPU/Common/FramebufferCommon.cpp
Outdated
| @@ -253,10 +253,10 @@ void FramebufferManagerCommon::EstimateDrawingSize(u32 fb_address, GEBufferForma | |||
|
|
|||
| if (viewport_width != region_width) { | |||
| // The majority of the time, these are equal. If not, let's check what we know. | |||
| const u32 fb_normalized_address = fb_address | 0x44000000; | |||
| const u32 fb_normalized_address = fb_address & 0x3FFFFFFF; | |||
There was a problem hiding this comment.
Hm. This heuristic probably isn't valid outside VRAM anyway - since we're setting fb_address in GetFramebufferHeuristicInputs, maybe the simpler thing is just not to mask this at all (removing this line entirely)?
If fb_address needs masking, then there's a bug in GetFramebufferHeuristicInputs. vfbs_[i]->fb_address possibly still needs masking, of course.
-[Unknown]
There was a problem hiding this comment.
Yeah, you're right - GetFramebufferHeuristic should only be called when we're actually rendering which means it must be VRAM because the GPU is incapable of rendering ot RAM.
I'll get rid of the unnecessary variable here.
| @@ -700,7 +700,7 @@ void FramebufferManagerCommon::UpdateFromMemory(u32 addr, int size, bool safe) { | |||
| // If we're not rendering to it, format may be wrong. Use displayFormat_ instead. | |||
| fmt = displayFormat_; | |||
| } | |||
| DrawPixels(vfb, 0, 0, Memory::GetPointer(addr | 0x04000000), fmt, vfb->fb_stride, vfb->width, vfb->height); | |||
| DrawPixels(vfb, 0, 0, Memory::GetPointer(addr), fmt, vfb->fb_stride, vfb->width, vfb->height); | |||
There was a problem hiding this comment.
This is inside a MaskedEqual() check which is probably dangerous and needs to be changed. I think MaskedEqual needs removing entirely, or it needs to check the VRAM bit and not mask it out (probably better to just ensure we never put anything in with kernel/cached bits.)
-[Unknown]
There was a problem hiding this comment.
Removing MaskedEqual entirely. I think this should be fine as-is now.
GPU/Common/TextureCacheCommon.cpp
Outdated
| // These checks are mainly to reduce scanning all textures. | ||
| const u32 addr = (address | 0x04000000) & 0x3F9FFFFF; | ||
| const u32 addr = address & 0x3F9FFFFF; |
There was a problem hiding this comment.
We might need to only do this mask when it's in VRAM now, right?
-[Unknown]
GPU/Common/TextureCacheCommon.cpp
Outdated
| @@ -750,8 +750,7 @@ bool TextureCacheCommon::AttachFramebuffer(TexCacheEntry *entry, u32 address, Vi | |||
| AttachedFramebufferInfo fbInfo = { 0 }; | |||
|
|
|||
| const u64 mirrorMask = 0x00600000; | |||
| // Must be in VRAM so | 0x04000000 it is. Also, ignore memory mirrors. | |||
| const u32 addr = (address | 0x04000000) & 0x3FFFFFFF & ~mirrorMask; | |||
| const u32 addr = (address & 0x3FFFFFFF) & ~mirrorMask; | |||
There was a problem hiding this comment.
Same here.
-[Unknown]
|
It's a little bit more sane now. Generally we're probably still doing some unneeded masking here and there since some inputs are guaranteed to not have the kernel/uncached bits but I don't think it matters too much, and it's best to defend against inputs where we don't have that guarantee, like the HLE memcpy etc. |
hrydgard
force-pushed
the
framebuffers-outside-vram
branch
from
7ee8a66
to
cf653b7
Compare
|
does this addition work also on android devices?? |
|
@owengque This particular change is just preparation for #11531, but yes, it will work and drastically speed up Digimon on all devices including Android. |
|
I just tried Digimon Adventure, but it's still slow when walking near his friends. if away from many characters it is stable 30fps. |
|
SD 625. |
|
First of all, the change hasn't been applied to the current builds yet. Are you building yourself and if so are you including #11531? Secondly it seems you have turned off buffer effects. The point of this is to allow them, making the outlines visible, while still not murdering performance. It's possible the game still isn't fast enough on your device, but this is still going to be a huge improvement with the buffer effects on. |
Accidentally broken in hrydgard#11553.
Prerequisite for #11531, virtual readbacks.
Seems to work fine now.