This commit sets up the infrastructure to introduce the new
descriptor set approach while keeping the old paths so the
hard coded apps are still operational. The old paths will be
removed once the compiler can compiler shaders for those apps
and the driver-compiler interface is fully flushed out.
Signed-off-by: Karmjit Mahil <Karmjit.Mahil@imgtec.com>
Reviewed-by: Frank Binns <frank.binns@imgtec.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21331>
This commit changes the pipeline layout, desc. set layout,
and desc. set layout binding to keep track of shader stage usage
with a mask of enum pvr_stage_allocation instead of
VkShaderStageFlags.
This commit also makes renames the relevant fields to
'shader_stage_mask' to make the naming uniform across stucts.
Signed-off-by: Karmjit Mahil <Karmjit.Mahil@imgtec.com>
Reviewed-by: Frank Binns <frank.binns@imgtec.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21331>
agx_preprocess_nir runs once per shader, whereas agx_optimize_nir runs once per
variant. That means we want to do as much work as possible in agx_preprocess_nir
to make shader variants as cheap as possible to compiler. So, move our standard
suite of lowering and optimizing to the preprocess loop, leaving just a single
(easy) trip through the optimizer for simple variant processing.
Plus, we can remove variables when preprocessing, since we no longer use
variables anywhere. We remove them to reduce the RAM and disk cache footprint of
shader variants.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21104>
This pass works either early or late, so run it late. It creates some
nir_variables as a side effect, which is weird, but it doesn't matter because
the AGX backend doesn't look at variables and the metadata and lowered I/O
intrinsics are all correct.
This is the last step to moving I/O lowering (and hence shader preprocessing) to
CSO create time.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21104>
Rob added these new helpers a while back, which freedreno and radeonsi
both share. We should use them too. The new helpers use variables and
system value intrinsics, so we can drop the explicit binding table
creation and just use the normal paths.
Because we have to rewrite the system value uploading anyway, we drop
the scrambling of the default tessellation levels on upload, and instead
let the compiler go ahead and remap components like any normal shader.
In theory, this results in more shuffling in the shader. In practice,
we already do MOVs for message setup. In the passthrough shaders I
looked at, this resulted in no extra instructions on Icelake (SIMD8
SINGLE_PATCH) and Tigerlake (8_PATCH). On Haswell, one shader grew by
a single instruction for a pittance of cycles in a stage that isn't a
performance bottleneck anyway. Avoiding remapping wasn't so much of an
optimization as just the way that I originally wrote it. Not worth it.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20809>
When a shader has a comparison with the subgroup invocation id,
we can use a constant instead, saving a VALU instruction.
When the constant can't be represented as a 64-bit literal,
use the s_bfm_b64 instruction to generate it instead, which
is still a win.
Fossil DB stats on GFX11:
Totals from 300 (0.22% of 134913) affected shaders:
CodeSize: 2223052 -> 2214336 (-0.39%); split: -0.43%, +0.04%
Instrs: 430216 -> 429882 (-0.08%); split: -0.14%, +0.06%
Latency: 5881180 -> 5878181 (-0.05%); split: -0.05%, +0.00%
InvThroughput: 731846 -> 729293 (-0.35%)
Copies: 31662 -> 31847 (+0.58%); split: -0.03%, +0.61%
Branches: 8241 -> 8100 (-1.71%)
PreVGPRs: 15788 -> 15786 (-0.01%)
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20843>
glthread takes care of all uploads, so it's OK to leave uploaded VBOs
bound. The only thing that will be wrong is the bound vertex buffer
returned by glGet, but the only case when that would be wrong is when
an app that doesn't use VBOs queries the current VBO. That never happens.
However, this adds code to unbind all internal VBOs for the case when
glthread is abruptly disabled (e.g. for GL_DEBUG_OUTPUT_SYNCHRONOUS).
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
This will be needed for lowering DrawIndirect in glthread, which is
needed if non-VBO vertex arrays are present.
This only adds the drawid parameter in glthread's draw_arrays and
draw_elements functions, and implements where needed.
New GL API functions are added because we want to use separate
DISPATCH_CMD_* enums for draws with DrawID, so that we don't increase
the memory footprint of draws in glthread batches if drawid == 0.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
u_vbuf does this too. This is the last big missing piece to stop using
u_vbuf.
If the vertex range to upload is much larger than the draw vertex count and
if all attribs are not in VBOs, convert glDrawElements to glBegin/End.
This is a path that makes the Cogs game go from 1 FPS to ~197 FPS. There is
no change in FPS because u_vbuf does this, but it will be disabled.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
MultiDrawElementsUserBuf is changed to mean the same thing as
glMultiDrawElementsBaseVertex, but "gl_buffer_object *index_buffer" is
passed via a parameter instead of using the bound GL_ELEMENT_ARRAY_BUFFER.
This skips binding and unbinding the index buffer around every draw
where glthread uploads indices.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
DrawElementsUserBuf is changed to mean the same thing as
glDrawElementsInstancedBaseVertexBaseInstance, but "gl_buffer_object *
index_buffer" is passed via a parameter instead of using the bound
GL_ELEMENT_ARRAY_BUFFER.
This skips binding and unbinding the index buffer around every draw
where glthread uploads indices.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
The main goal is to never fail to upload non-VBO vertex arrays.
When glthread synchronized, it didn't upload vertices, expecting st/mesa
to do that. This keeps the required sync, and then upload vertices
in glthread.
Also, reorder the code and remove goto statements. This is pretty much
a rewrite.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20624>
Since the dirty range started out as 0..0, you would have 0..VBend as the
new dirty range on the first draw, and if your VB was >32b then you'd
flush every time you used it. Instead, if there's no existing dirty range
then just set it to our new VB's range.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21370>
Since the dirty range started out as 0..0, you would have 0..VBend as the
new dirty range on the first draw, and if your VB was >32b then you'd
flush every time you used it. Instead, if there's no existing dirty range
then just set it to our new VB's range.
Perf results with zink+anv on my CFL:
sauerbraten: +24.8182% +/- 0.602077% (n=5)
portal-2-v2.trace: +4.64289% +/- 0.285285% (n=5)
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21370>
