Now we support all the vertex formats! This means we don't hit u_vbuf for format
translation, which helps performance in lots of applications. By doing the
lowering in NIR, the vertex fetch code itself can be optimized by NIR (e.g.
nir_opt_algebraic) which can improve generated code quality.
In my first implementation of this, I had a big switch statement mapping format
enums to interchange formats and post-processing code. This ends up being really
unwieldly, the combinatorics of bit packing + conversion + swizzles is
enormous and for performance we want to support everything (no u_vbuf
fallbacks). To keep the combinatorics in check, we rely on parsing the
util_format_description to separate out the issues of bit packing, conversion,
and swizzling, allowing us to handle bizarro formats like B10G10R10A2_SNORM with
no special casing.
In an effort to support everything in one shot, this handles all the formats
needed for the extensions EXT_vertex_array_bgra, ARB_vertex_type_2_10_10_10_rev,
and ARB_vertex_type_10f_11f_11f_rev.
Passes dEQP-GLES3.functional.vertex_arrays.*
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19996>
Better occupancy, which is especially important when the background shader
does memory access (for reloads). On my 4K monitor, glmark2 -bdesktop fullscreen
from 95fps to 133fps.
At default settings, glmark2 -bterrain from 63fps to 71fps.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19997>
Massive performance gains, some fps before/after numbers from glmark2:
[shading] 1486 -> 2391
[refract] 87 -> 127
[terrain] 32 -> 56
...and it's basically for free with enough copy/paste, so thank you to Boris
Brezillon for an excellent Asahi patch, the LRU cache seems to work great on M1
:-p
There are a few minor changes I made from panfrost, notably adjusting the
constants to account for 16KiB pages and switching from pthread_mutex to
simple_mtx to be less weird in Mesa.
For context on the design, the following commits evolved it in Panfrost and
their commit messages may be useful... The logic in this module is the product
of years of mistakes and correcting course :-)
f06809cdca ("panfrost: Evict the BO cache when allocation fails")
77d0498913 ("panfrost: Fix major flaw in BO cache")
ee82f9f07e ("panfrost: Try to evict unused BOs from the cache")
2225383af8 ("panfrost: Make sure the BO is 'ready' when picked from the cache")
9af4aeaaf7 ("panfrost: Don't return imported/exported BOs to the cache")
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19971>
This defeats the point of specifying alignments and of packing allocations
together with the BO cache. We're a real driver now, let's allocate memory like
one.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19971>
We have these native. Passes the relevant piglits. Large reduction in memory
usage on Xonotic on higher settings (8x less memory per texture), which allows
Xonotic to run at high settings without OOMing.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Tested-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19903>
Flag day change to replace the previous hardcoded background/end-of-tile shaders
and the API-style load/store_output in fragment shaders with the generated
shaders and lowered *_agx intrinsics. This gets us working non-UNORM8 render
targets and working MRT. It's also a step in the direction of working MSAA but
that needs a lot more work, since the multisampling programming model on AGX is
quite different from any of the APIs (including Metal).
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19871>
With multiple render targets, it's not practical to generate all
variants of the background and end-of-tile programs at start up. Rather
than trying, add a hash table of meta program keys to background
programs, and compile variants as they're needed.
With the new infrastructure, it's sensible to handle clears with the
same code path as reloads. In addition to getting us closer to multiple
render target support, this gets us support for non-RGBA8 render
targets, as the u8norm tilebuffer format was baked into the hardcoded
clear shader and store shaders used before.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19871>
The compiler can't handle load/store_output directly for nontrivial tilebuffer
layouts. Add a NIR pass to lower these intrinsics, applying a given layout.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19871>
The track_alloc and track_free symbols are used, we need to link them in.
Depending on build flags / environment / etc, fixes the potential build error
hit by a CI job:
mold: error: undefined symbol: agxdecode_track_alloc
>>> referenced by agx_device.c
>>> src/asahi/lib/libasahi_lib.a(src/asahi/lib/libasahi_lib.a.p/agx_device.c.o):(agx_shmem_alloc)>>> referenced by agx_device.c
>>> src/asahi/lib/libasahi_lib.a(src/asahi/lib/libasahi_lib.a.p/agx_device.c.o):(agx_bo_create)
mold: error: undefined symbol: agxdecode_track_free
>>> referenced by agx_device.c
>>> src/asahi/lib/libasahi_lib.a(src/asahi/lib/libasahi_lib.a.p/agx_device.c.o):(agx_bo_unreference)
...when trying to link with libasahi_lib without libasahi_decode for unit tests.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19871>
This avoids exposing the ISA-internal agx_format to the driver, instead hiding
it behind a real PIPE_FORMAT. This lets us use real pipe formats in formatted
intrinsics in NIR, which is convenient; it will allow us to simplify the
compiler/driver ABI; and it lets us use common format helpers (e.g.
util_format_get_blocksize) for the internal formats in driver lowering.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19871>
Obviously there can't *actually* be memoryless render targets, because
how would partial renders work? The control stream with memoryless looks
like everything would if it went to memory (e.g. full 2D MSAA
attachments for the partial loads/stores even if only a resolved
2D image for the final store). Except the memoryless attachments all
load from the same address 0xeeee0000. Clearly that's not actually what
happens, so what gives? Unclear... but I see the magic bits mentioned
here set, and I assume there are some firmware (or kernel) shenanigans
used to JIT allocate the backing storage for partial renders.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19811>
Stencil texturing is easy: S8_UINT is textured like R8_UINT (with a
little swizzle fixup), and stencil is always S8_UINT thanks to
u_transfer_helper. So we just need to do some fixups to make
u_transfer_helper's seperate_stencil work and everything will work out.
Passes dEQP-GLES31.functional.stencil_texturing.*
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19811>
The big discovery is the "number of uniform registers" field. I learned
about this one accidentally when my preamble shaders weren't working
right, because we had inadvertently hardcoded "at most 32 registers" :-)
In the course of identifying that field, I found that the pipeline
address is used as a tagged pointer, with some unknown field in the
bottom bits and alignment demanded. The XML is updated to account for
this.
I later found that there's also a "number of general purpose registers
used by the preamble shader" field. I missed this one first, because the
encoding is slightly different from the usual "number of general purpose
registers in the main shader" field. The specification is slightly
coarser. I don't know why the hardware needs that
information anyway -- occupancy of the preamble shader should be
irrelevant -- but it's not a big deal.
Finally I found that the "more than 4 textures?" bit is... not that. I
do not yet know what it is, but it is... not that.
These all use the new groups() modifier for GenXML
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18813>
The start field in the Uniform USC word is only 8-bits, whereas 9-bits
are required to address the entire uniform register file. This other
word gets used for the high half, with start indexed from u128l in
the natural way.
Apparently spending the evening stuffing too many uniforms into Metal is
paying off.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18813>
This gets shader-db's runner working, in conjunction with a shader-db ./run
modified to set ASAHI_MESA_DEBUG=precompile. This flag triggers precompiles of
all shaders witha default key so we can exercise the compiler.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18813>
