Commit graph

63386 commits

Author SHA1 Message Date
Alyssa Rosenzweig
7a7a3b8847 asahi: defer texture packing to draw-time
needed for correctness with layout transitions.

fixes arb_shader_image_load_store-host-mem-barrier

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
b74d2dcd57 asahi: use 2D descriptors for cubes
fixes arb_shader_image_load_store-invalid case imageLoad/address bounds test/imageCube/rgba32f

this is also better codegen since it avoids the wacko division by 6. although it
creates a div by 6 in imageSize, that's better because that one is much more
likely to hoist to the preamble. probably should've done this from the start.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
86c4a72767 asahi: rm compact image atomic descriptors
these cause robustness problems -- since the target type might not match the
shader for invalid apps -- and are a dubious microoptimization. can revisit
later. for now, fixes imageAtomic*/target mismatch test.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
0cd032e4bb asahi: Be robust with arrays of images
fixes imageLoad/index bounds test subcase.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
b0f4b0a9ef asahi: handle null PBE
fixes some subcases of arb_shader_image_load_store-invalid

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
c73d7d042c asahi: fix dirty tracking fail with point sprites
fixes arb_point_sprite-checkerboard

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
683dac24b2 asahi: be robust about null xfb
fixes arb_enhanced_layouts-transform-feedback-layout-qualifiers tests.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
3ac44d8c5e asahi: remove bogus assertion
replace with optional debug flag for retaining debugability but getting
conformant behaviour by default. fixes piles of piglit crashes.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
e3c528068f asahi: bump limits
fixes gl-3.3-minmax.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
84a8324c55 asahi: handle self blits
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
de6336381b asahi: clamp clear colours
fixes arb_texture_view-rendering-formats_gles3

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
acd095f03d asahi: init clear colour between batches
should be a noop, but safer

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
c507f4a330 asahi: Identify bicubic filtering mode
Officially undocumented but supported by MSL.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
58d6374b4c asahi: Identify Primitive ID frag input
With a name from powervr :)

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Alyssa Rosenzweig
3da3b18d0c asahi: Quiet clang warning
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Asahi Lina
1b1a0e0ce7 asahi: Move USC cache flush to agx_batch_init_state
This shouldn't make a difference, but it's cleaner to do it here, since
then we don't have to check whether we have any draws.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Asahi Lina
5a9136a0ef asahi: Add extra CDM barrier bit for G13X
I don't know if we need the other bits yet. I'm assuming the logic here
is the same as we used to use for VDM...

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Asahi Lina
c20210c643 asahi: Fix CDM Launch/Barrier naming
"Launch" is actually just a barrier, and it seems likely to use the same
bit assignments as in VDM...

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26614>
2023-12-09 10:56:17 -04:00
Qiang Yu
43d83cfcf9 radeonsi: fix legacy merged LS/ES workgroup size for aco compilation
This is for ACO handling of nir_load_local_invocation_index, which will not
add (wave_id * wave_size) if workgroup size <= wave size. So merged shader
LS/ES need to set workgroup size too.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26560>
2023-12-09 06:19:49 +00:00
Marek Olšák
57e658d041 radeonsi: rework how guardband registers are updated to decrease overhead
See the code comments.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:28 +00:00
Pierre-Eric Pelloux-Prayer
2e50b6677b radeonsi: update guardband if vs_disables_clipping_viewport changes
Because the guardband state depends on this value.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:28 +00:00
Marek Olšák
11fcd58957 radeonsi: convert rasterizer state to tracked registers
Most of these registers don't change, so we should not set them when they
don't. This reworks the rasterizer state to use a custom emit function and
eliminate redundant register changes. This required merging the poly_offset
state into the rasterizer state and change how the poly offset state is
updated.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
cabf927765 radeonsi: convert depth-stencil-alpha state to tracked registers
Some of these registers don't change, so we should not set them when they
don't. This reworks the DSA state to use a custom emit function and
eliminate redundant register changes.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
17e01a9a9b radeonsi: merge context_reg_saved_mask and other_reg_saved_mask into a BITSET
There will be more than 64 context registers that we'll need to track,
so use BITSET for all of them.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
6d2a7f53ac radeonsi: decrease NUM_INTERP if export formats/colormask eliminated PS inputs
This adds a pass that removes output stores. It's called before NIR
optimizations that are before PS lowering, which will cause it to remove
PS inputs from register settings.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
bd700bace5 radeonsi: update comments about uniform inlining
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
dbea0f51a3 radeonsi: decrease NUM_INTERP if uniform inlining eliminated PS inputs
This should improve performance when that happens.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
1d4402b02a radeonsi: don't include the PARAM_GEN input in si_shader_info
It's only produced by lowering point smoothing.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
197af03698 radeonsi: add PS input info into si_shader_binary_info
It will be modified to reflect PS inputs after uniform inlining.
For now, it's just a copy of selector->info.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
3faa9ba06b radeonsi: gather lowered color inputs for monolithic PS
This also adds missing io_semantics to the input loads that the gathering
expects.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
94e0a1267e radeonsi: change the signature of si_nir_lower_ps_color_input
This will be needed later.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
7d2faa88ab nir,radeonsi: add FLAGS into load_vector_arg_amd to record color input usage
This will be needed for gathering color usage from lowered PS.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
00dd4d400e ac,radeonsi: rename pos_inputs -> fragcoord_components
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
4f2b794e98 radeonsi: track NIR progress properly for optimizations in si_get_nir_shader
Just a small code size decrease in 12 shaders.

TOTALS FROM AFFECTED SHADERS (12/58918)
  SGPRS: 600.00 -> 600.00 (0.00 %)
  VGPRS: 528.00 -> 520.00 (-1.52 %)
  Spilled SGPRs: 0.00 -> 0.00 (0.00 %)
  Spilled VGPRs: 0.00 -> 0.00 (0.00 %)
  Private memory VGPRs: 0.00 -> 0.00 (0.00 %)
  Scratch size: 0.00 -> 0.00 (0.00 %) dwords per thread
  Code Size: 39772.00 -> 39688.00 (-0.21 %) bytes
  Max Waves: 180.00 -> 180.00 (0.00 %)
  Outputs: 0.00 -> 0.00 (0.00 %)
  Patch Outputs: 0.00 -> 0.00 (0.00 %)

Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
e1e35112c8 winsys/amdgpu: bypass GL2 for command buffers
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
c77bcf00a3 radeonsi/gfx11: prefer Wave64 for VS/TCS/TES/GS because it's slightly faster
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
257f07f499 radeonsi: clean up how debug flags and shader profiles determine the wave size
- remove DBG_W32_PS_DISCARD
- just return the wave size instead of setting local variables dbg_wave_size
  and profile_wave_size

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
716b521515 radeonsi/gfx11: disable the shader profile for Medical that disables binning
GFX11 performs better with the default behavior.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
f85488824e radeonsi/gfx11: disable the shader profile for Medical that forces Wave64
GFX10 should keep using it, but not GFX11.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
65b3b0b355 radeonsi/gfx11: prefer Wave64 for PS without inputs for better VALU perf
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
48ce5fbaa1 radeonsi: group most vertex element fields
fix_fetch and vertex_buffer_index can't be grouped because we do
memcmp on those arrays

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
7fa0ee15df radeonsi: set OOB_SELECT for VBOs in si_create_vertex_elements
we can do this since the stride is in the CSO now

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
55d81214c9 radeonsi: replace gl_FrontFacing with a constant if one side is always culled
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
1afe6f3321 radeonsi: don't print the preamble state separately for GALLIUM_DDEBUG
because it's always printed as part of command buffers.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
9e76459616 radeonsi: execute streamout_begin after cache flushes
so that si_emit_streamout_begin can assume that cache flushes have
finished.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
2022854360 radeonsi/gfx11: skip si_set_streamout_enable because it has no effect
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
bf7debee82 radeonsi: in bind_{blend,rs}_state, only call 1 update function per if
Also don't use "key.ps.part.prolog.color_two_side" during updates
because it would depend on the order the update functions are called,
which is not a problem now, but it's a trap for the future.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
53aa36772a radeonsi: rewrite si_get_total_colormask as si_any_colorbuffer_written
The result is only used as bool.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
e2b817b948 radeonsi: rewrite how shader key bits dependent on current_rast_prim are updated
Don't set do_update_shaders every time current_rast_prim changes, which can
be EVERY DRAW. Instead, just update the shader key bits and set
do_update_shaders only if any bits are different.

When we bind a new rasterizer state, do the same.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00
Marek Olšák
4ab5374ec3 radeonsi: clean up setting poly/line/stipple shader key bits
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26307>
2023-12-09 00:05:27 +00:00