07c8f163e4
Changes:
* Update lightrec to latest upstream
* Minimize logs when loading a cheevos-compatible content
* Cleanup retro_run() -
- move input query into separate functions
- move internal fps display to separate function
* Hide other inputs from core options -
- This adds a core option to hide some input options like
multitaps, player ports 3-8 and analog-related fine-tuning options.
- also combine dynarec-only options in one #define directive
* More core option fixes -
- This PR fixes core options and moves them to the related dynarec
modes where they are implemented.
LIGHTREC = relates to platforms that supports the new Lightrec mode
NEW_DYNAREC = relates to previous dynarec implementation that is
still used for some 32bit devices
- Dynarec Recompiler core option, both dynarec implementation can
be enabled or disabled
* Move guncon options to update_variables -
- This should stop unnecessary RETRO_ENVIRONMENT_GET_VARIABLE
callback and log spamming
* Fix some edge case where core can freeze upon loading content
* Automatically disable Lightrec when no BIOS is present, take 2
* cdriso: fix a disk switching deadlock when closing a CD image
* ARM NEON: Fixed bug where MSB of a 15-bit BGR color could corrupt
green value.
* cdriso: fix a disk switching deadlock
* unai: Add ARM-optimized lighting / blending functions
Addendum on UNAI ARM-optimized lighting/blending improvements -
"Looking at the generated ASM on 3DS, I thought I could squeeze out
some extra performance by moving the inner lighting and blending
functions to handwritten A32 assembly. This gives a medium improvement
generally (3-5fps faster on the beach in Crash 1) and a large
improvement when doing lots of blending (46-48fps before, 57-60fps
after, behind the waterfall in Water Dragon Isle in Chrono Cross).
Some other notes:
* I used the ARM11 MPCore (3DS CPU) timings for pipelining.
* I had a few stall cycles during lighting, so I used them to
preserve the MSB for lighting and blending, which saved a store,
load, and orr later on. ~3-6 cycles saved overall by doing that.
* I switched from u16 to uint_fast16_t, which is 32-bit on this
platform. This saved a few useless uxth instructions for another
few cycles. This shouldn't affect other platforms, but I don't know
for sure. Could typedef if necessary.
* A lot of the speed improvement in blending comes from not using two
instructions per and. For example, & 0x8000 -- the compiler
preferred to mask out bytes using bic 0x7F00 and bic 0x00FF. Both
slower and seemed less correct for what we're trying to do."
6 lines
636 B
Text
6 lines
636 B
Text
$NetBSD: distinfo,v 1.6 2020/05/30 14:49:00 nia Exp $
|
|
|
|
SHA1 (libretro-pcsx-rearmed-20200526-7973b25fe929f92e146a854ecaf4f3cea5b4ffb8.tar.gz) = dd07cdb90a432d3626147ae8901fdfd8489f883a
|
|
RMD160 (libretro-pcsx-rearmed-20200526-7973b25fe929f92e146a854ecaf4f3cea5b4ffb8.tar.gz) = dace80b932044463eb0cf6fc0c9eb5917f5c030b
|
|
SHA512 (libretro-pcsx-rearmed-20200526-7973b25fe929f92e146a854ecaf4f3cea5b4ffb8.tar.gz) = bd8c2e5ca9d434005790767d3a4a2d7bc86deb1ca272ef73676c99a0cacb1dd92e843732d4f3455f77d0bce4666953befef7a108a616e8a4b2ff752deeef4e69
|
|
Size (libretro-pcsx-rearmed-20200526-7973b25fe929f92e146a854ecaf4f3cea5b4ffb8.tar.gz) = 2278331 bytes
|