86782156e7
asahi: invert geometry shaders
...
instead of dumping GS outputs to memory and using a GS copy vertex shader,
invert the GS to get a hardware vertex shader to use directly. theoretically,
reduces mem bandwidth at the cost of repeated work if we don't optimze well.
perhaps more importantly, it should reduce heap usage.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:32 +00:00
1ac3e48fcc
asahi: fix vertex out size calc
...
expression underallocates if var31 is actually used!
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
141e084f55
asahi: use load_instance_id in gs lowering
...
it will lower differently in the prepass vs the rast shader
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
3641a37e8f
asahi: set gs_grid[0] even for direct draws
...
gs rast program needs to know the # of input primitives
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
200ca69086
asahi: support stage override in sysval lower
...
for gs rast program. should clean up later but for now this will do
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
3bd400a2b4
asahi: collapse indirection with GS
...
1. compile the gs rast program with the rest of the GS. this requires
duplicating some shader key stuff but this will be improved soon, and
actually this makes it easier to reason about caching imo.
2. don't recurse into draw_vbo, collapse the GS launches into the draw. again
this duplicates some code but makes it easier to reason about things and is
probably faster. and makes part #1 feasible without piling on hacks (my
original goal)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
2da426afe0
asahi: allow lowering bindings after lowering textures
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
17896f1699
nir: rm load_vert_id_in_prim_agx
...
now unused since we separate vs/gs
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
b26409246c
asahi: be a bit more methodical with shader stages
...
the mapping between descriptor stages and hardware stages is about to get even
more complicated, so clean up the driver ahead-of-time to stay ahead of the mess
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
6154d65714
asahi: infer stage in descriptor update
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
37fa7bff58
asahi: shrink GS key
...
no functional change, just dropping the unused index size field
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
f4a187bc89
asahi: allow bindful GS textures
...
no longer merging GS so it's fine
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
e13d001160
asahi: rm arrayed output lowering
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
d96fbd4618
asahi: separate GS from VS
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
af7084efa7
asahi: rework shader stage handling a bit
...
prep for separable vs/gs
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
0b30777028
asahi: add geometry parameters for separable GS
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
b2d547b361
asahi: add shader_info::outputs for gs lower
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
cdd1b012e9
asahi: simplify IA mode handling
...
we'll solve the IA variants problem a different way, for now make more
straightforward.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
e1783fc083
libagx: generalize vertex_id_for_topology
...
so we can use it in separable GS
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
c9f55166dc
libagx: static assert some sizes
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
aaa609f030
libagx: use real PACKED macro
...
hopefully end these alignment bugs for good
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
fc929d9318
libagx: use native static_assert on host
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
bfaa3691e1
asahi: make GS flatshade_first more dynamic
...
baby step
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
4036c710bd
libagx: fix buggy align macro
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
784323cf0f
asahi: rip out existing MDI+GS implementation
...
I don't love deleting working code without a working replacement. But this impl
is deeply questionable as-is, and will only balloon in complexity for shader
objects. We need to go back to the drawing board with MDI and lower to real
draws, possibly with control stream looping.
Good news is that this is GL so we can prototype all this.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
21403278ac
asahi: fix dirty tracking issue
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
beabd93c13
asahi,agx: use hw clip distance
...
rather than lowering in the fragment shader. this is what we're "supposed" to
do, what apple does, and it saves a byte of the fs key
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
9aeaf99259
agx: fix buffer overflow with varying slots
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
de31b28053
asahi: allow vertex/geom/tess side effects
...
seems to work ok now.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
de786a1c5c
asahi: drop xfb hack
...
now that we can spill it's unnecessary. it also breaks things (-:
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
5361b1f331
asahi: bump max threads per wg
...
gl4.3 min.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
8a2c1d337e
agx: add =spill debug option
...
to force spilling so we can exercise the new code.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
42084d50e5
agx: add parallel copy printing
...
useful for debugging.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
871f3d3f7d
agx: unit test spill/fill lowering
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
a5ff85e782
agx: unit test memory parallel copies
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
14e1bb78ae
agx: add spill/fill lowering pass
...
This simple pass turns moves of memory registers (outputted by RA from the
spiller produced moves/phis of memory variables) into concrete stack load/store
instructions. it's a lot more convenient to do this as a dedicated post-RA pass
than trying to fold this into the hairball that is RA proper.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
40da539d01
agx: introduce "memory variables"
...
in order to spill, we need to be able to assign spill slots. this is equivalent
to register allocation, so we borrow NAK's neat solution: add "memory variables"
into the IR which get RA'd to "memory registers" that correspond directly to
stack slots. then we can just spill by inserting moves.
this requires extensions to RA, parallel copy lowering, and the IR itself. but
it makes competent spilling so much more straightforward.
to exercise this code path, the commit also includes a trivial spiller that
simply moves every source/destination between memory variables and general
variables, which has the effect of spilling everything. to be replaced by
Braun-Hack eventually, this is just foundations.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
06eb552baa
agx: print register vectors
...
easier dbg
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
282f2ac1aa
agx: allow 16-bit immediate on stack load/store
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
485e17e01e
asahi: allocate preamble scratch
...
needed for preamble spilling
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
f97b041e87
agx: set nr_preamble_gprs for preamble scratch
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
94b9200042
asahi: add has_scratch to shader key
...
some stages are unspillable
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
9445005c87
agx: improve scratch size accounting
...
- prep for preamble scratch
- only include scratch actually used
- prep for spilling scratch
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
a2328820f7
agx: clamp register file based on workgroup size
...
once we allow workgroup sizes larger than 384 threads, we need to restrict the
register file to ensure we fit.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
bb37b072a5
asahi,agx: use intrinsics pass
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
e34a707d78
asahi: optimize memory_barrier
...
don't flush everything, only flush batches that could actually need a flush.
this eliminates memory barrier flushes in an apitrace of ryujinx, although perf
impact there seems neglible.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
745c9d8bdc
asahi: track batches with incoherent writes
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
98ffbf4cf6
asahi: rm blank
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:31 +00:00
53e820f20e
asahi: hoist layer id code
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:30 +00:00
bbe6689108
asahi: hoist xfb code
...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io >
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616 >
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig