- OpenGL display views now use a shared fetch object to fetch the emulated GPU framebuffers and store them in shared textures within a shared context. In conjunction with the new double-buffering support from the last commit, this eliminates the copying between the framebuffers and each display view.
- OpenGL display views now use shared HQnx LUT textures, rather than having to initialize and maintain a copy of the LUT textures for each display view.
- OpenGL display views no longer perform any rendering while their associated NSView is hidden, improving the performance of creating new display views.
- OpenGL display views can now DMA directly from pinned-memory both custom-sized framebuffers and CPU-pixel-scaled native-sized framebuffers at the same time.
- Framebuffers are now page-aligned on 4KB boundaries. This is to
improve performance when using the framebuffers directly as pinned AGP
memory.
- Framebuffers are now double-buffered. The target buffer index is now
tracked using the bufferIndex field of NDSDisplayInfo.
- Clients may no longer supply their own buffers to
SetCustomFramebufferSize(). Clients must use the pointers supplied by
NDSDisplayInfo.
- The frameskip flag is now set only on line 0 and remains consistent
for all 192 lines of rendering.
- GPUSubsystem no longer needs a special allocater/deallocator for
itself, so it has been reverted back to a standard C++ new/delete.
- Add a GPUClientFetchObject helper class as an aid to clients that
need to read out the framebuffers. (Should probably move to its own
file at some later date.)
This reverts commit 53c4a27aef.
I forgot that these functions are based on element count, not based on
byte count. Rename “length” to “elementCount” for better clarification.
- Fetching and loading of GPU frame data is now performed as two
separate operations.
- Display windows no longer draw concurrently on backgrounds threads;
instead they are updated synchronously.
- Associate the CALayer after the .xib completely loads the NSView for
better compatibility.
- MacOGLDisplayView now creates an NSOpenGLContext instead of a
CGLContextObj, bringing back compatibility with macOS 10.5 Leopard.
- Fix building with the Xcode 3 project.
- Most notably, HandleGPUFrameEndEvent() now sends the entirety of the
NDSDisplayInfo struct to the client.
- The OpenGL blitter now skips the loading, processing and rendering of
disabled screens.
- Begin preparing DisplayView to handle the upcoming Apple Metal
blitter.
- Do some misc. code cleanup and simplification.
- Also disable the 4xBRZ shader for low-tier GPUs. (Testing has shown
that low-tier GPUs have no chance at running this shader in real-time
anyways.)
- Also do some misc. tweaks to other various shaders.
- Move Mac-specific OpenGL code to its own file.
- Eliminate the CocoaDSDisplayDelegate and CocoaDSDisplayVideoDelegate
protocols. Instead, call ClientDisplay3DView class methods directly.
- Fix bug where restored windows would fail to update properly if the
window size would be the same as the one set in user defaults.
(Regression from commit cffc343.)
- Fix bug where changing the rotation to be exactly 180 degrees
different from the old rotation would cause the view to render the
screens with a vertical offset. (Regression from commit cffc343.)
- OGLDisplayLayer respects its own _needUpdateRotationScale and
_needUpdateVertices flags once again, preventing it from repeatedly
uploading already established data to the GPU. (Regression from commit
cffc343.)
- The new Hybrid display orientations show three NDS screens — the top
and bottom screens on the left side, and a larger major screen on the
right side. This feature is intended to better use the widescreen
resolution of most users’ host displays, which are usually 16:9 or
16:10. Three different Hybrid orientations are provided (3:2, 16:9, and
16:10) so that users can choose the display size ratio that they prefer.
- Horizontal display orientation no longer uses the display separation
setting.
- Do some additional code cleanup.
Do note that we need to do this for SoftRasterizer as well, but
SoftRasterizer will need some additional rework on shadow polygon
handling to get all the test cases to work.
Fixes issue #21.
Now that framebuffer sizes can be greater than 256x192, using MSAAx16
is excessive and consumes too much bandwidth. We’re reducing the
maximum sample size of MSAA to 8 since it will significantly reduce
bandwidth consumption at the larger framebuffer sizes while remaining
mostly visually equivalent to MSAAx16.
rogerman