The maximum number of DWORDS per IB is limited by the hardware. So,
when the number of sequences is too high, it would just hang.
The solution here is to implement IB chaining inside the DGC cmdbuf
itself, so that a sequence chains the next one basically.
In practice, games only use up to 4K sequences and they aren't affected
by this change.
This fixes dEQP-VK.dgc.ext.misc.properties.maxIndirectSequenceCount.
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/13536
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36062>
e.g. load X; load W; ==> load XYZW. Verified with a shader test.
This will be used by AMD drivers. See the code comments.
Reviewed-by: Simon Perretta <simon.perretta@imgtec.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36098>
A TF2 shader propagates 0 to the consumer, which eliminates 1 input
if we run algebraic opts and DCE before compaction.
This is a prerequisite for removing all IO var optimizations from the GLSL
linker that are redundant with nir_opt_varyings.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36091>
This workaround doesn't mitigate the issue reliably/completely. An
alternative (but complex) solution also exists.
This introduces a small option that allows to disable the current
workaround as preliminary work.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36110>
This exposes an experimental implementation of HIC with
RADV_PERFTEST=hic. It's passing 100% of VKCTS but it requires some
benchmarks first to verify if performance is acceptable or not.
No addrlib support for GFX6-9.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35974>
Because there is no surface<->surface helper in addrlib, this allocates
a temporary buffer on the CPU to do image->buffer->image. It's a naive
implementation which is probably not the best for performance, but it
works at least.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35974>
On GFX12, everything is compressed with DCC and it's completely
transparent to the userspace driver, so that should be optimal.
On older gens, using HIC disables compression which isn't optimal
for device access.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35974>
On VCN5 both distinct and coincide output/dpb are supported. Tier3
(coincide) requires tiling, Tier2 (distinct) also works with linear.
Application can decide which one to use.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35878>
All planes must have the same swizzle mode and no tile swizzle.
Only linear decode target requires the custom height alignment.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35878>
On SDMA4, when the pitch isn't aligned, the width needs to be scaled
by 3 for 96-bits formats.
On SDMA5+, the pitch is aligned and the driver doesn't need to fallback
to unaligned copies.
CC: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36067>
This will enable large code removal.
shader->config.lds_size is now always computed the same as ACO except for
compute shaders.
We have to add a new 8-bit user SGPR bitfield called
GS_STATE_GS_OUT_LDS_OFFSET_256B, which contains the offset
that was previously set by the relocation.
Since the offset must be a multiple of 256, we have to add padding
to the LDS size computation to make sure the alignment to 256 for the ESGS
LDS size doesn't cause us to exceed the maximum LDS size.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35529>
This increases primitive throughput when packing reduces the number
of pos exports due to holes in clip and cull distance arrays that could be
punched out by nir_opt_clip_cull_const. This applies to all chips.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
This increases primitive throughput for all hw with NGG if the shader
culls and the removal of cull distances reduces the number of position
exports.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
It will be different between NGG and legacy because NGG with culling
will not export cull distances.
The number of position exports could also be gathered from final NIR
to reduce logic duplication.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
It will be different between NGG and legacy because NGG with culling
will not export cull distances.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
This is very useful for increasing raw primitive throughput for GS
(mostly just RDNA 2), increasing raw primitive throughput with clip
and cull distance outputs when they actually cull anything (RDNA 1-4),
and reducing attribute store bandwidth usage (RDNA 3-4).
It will also replace fixed-func culling against cull distances when
culling in the shader is enabled, which will increase primitive throughput
even further.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
It receives all shaders and decides how to link them.
When culling is enabled for GS, we will need ES, GS, and FS in this
function at the same time.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
Clip and cull distance outputs decrease primitive throughput, so culling
against them in the shader has even more benefit than other culling
options.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
This simplifies shader info in drivers by returning GSVS emit sizes from
ac_nir_lower_legacy_gs. The pass knows the sizes, so drivers shouldn't
have to determine them independently.
This also makes the values more accurate because both drivers were
computing the GSVS emit sizes inaccurately and had redundant fields
in shader info. RADV had a lot of redudancy there.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35473>
