angrylion's unfiltered views (Unfiltered, Depth, Coverage) draw into the
frontend's software framebuffer when it lends one, as the filtered view
already does, rather than into a buffer of their own to be copied every
frame; the interlaced field they drop now shows the frame before again
without presenting anything. Its upscaled resolve to RDRAM averages
samples with a shift and works out each row's addresses once.
parallel-RDP reuses the nodes it queues to its timeline worker, handed
back through a lock-free stack, instead of allocating and freeing one
for every coherency operation. GLideN64 checks a framebuffer the CPU
has not written with one memcmp before counting differing pixels. Rice
uploads a texture once after converting and expanding it, rather than
up to three times, and no longer flushes GL every frame. glN64 decodes
textures and backgrounds into buffers it keeps rather than allocating
two per load.
glN64 no longer reads its texture cache's dummy after freeing it when
the context goes, and Rice deletes what it made with new rather than
freeing it. The GL lane runs again under AddressSanitizer, which fails
on both, and a new lane (tools/ci/angrylion_lend.py) holds angrylion's
frames in a lent buffer to its own and requires the unfiltered views to
draw into it.
libretroadmin
9dc453242e