Significantly improves performance of a Control compute shader. Also seems
to increase FPS at the very start of the game by ~9% (RX 580, 1080p,
medium settings, no MSAA).
fossil-db (Navi):
Totals from 315 (0.23% of 135946) affected shaders:
SGPRs: 18296 -> 18336 (+0.22%); split: -0.26%, +0.48%
VGPRs: 11856 -> 11844 (-0.10%); split: -0.81%, +0.71%
CodeSize: 2233800 -> 2457508 (+10.01%)
MaxWaves: 4506 -> 4497 (-0.20%); split: +0.04%, -0.24%
Instrs: 438766 -> 486215 (+10.81%); split: -0.00%, +10.81%
Cycles: 7880180 -> 8963340 (+13.75%); split: -0.00%, +13.75%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6558>
SPIRV->NIR emits nir_op_unpack_half_2x16_flush_to_zero instead of
nir_op_unpack_half_2x16 if the shader enables denorm flush to zero
for 16-bit floating point.
This doesn't fix anything known and CTS doesn't have tests.
Fixes: 56d9bcdded ("radv: enable more float_controls features")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6939>
Fixes: 7568c97df1 ("radv: Use atomics to read query results.")
Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7050>
With arrays we really have to use the correct size for the base
mipmap to get the right array pitch. In particular, using
surf_pitch results in pitch that is bigger than the base mipmap
and hence results in wrong pitches computed by the HW.
It seems that on GFX9 this has mostly been hidden by the epitch
provided in the descriptor but this is not something we do on
GFX10 anymore.
Now this has some draw-backs:
1. normalized coordinates don't work
2. Bounds checking uses slightly bigger bounds.
2 mostly is not an issue as we still ensure that they're within
the texture memory and not overlapping other layers/mips, but
we can't properly ignore writes.
1 is kinda dead in the water ... On the other hand I'd argue that
using normalized coords & a filter for sampling a block view of
a compressed format is extraordinarily useless.
The old method we employed already had these drawbacks for everything
except the base miplevel of the imageview.
AFAICT this is the same tradeoff AMDVLK makes and no CTS test hits
this. (once it does I think the HW is dead in the water ... Only
workaround I can think of is shader processing which is hard because
we don't know texture formats at compile time.)
I also removed the extra calculations when the image has only 1 mip
level because they ended up being a no-op in that case.
CC: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2292
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2266
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2483
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2906
Gitlab: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3607
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7090>
AC_FETCH_FORMAT_NONE is not zero... Oops.
Fixes: b0829c6af7 ("radv: replace RADV_ALPHA_ADJUST by AC_FETCH_FORMAT")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7103>
nir_lower_io_arrays_to_elements lowers arrays or matrices to elements,
which ends up to vectors for matrices, but a bunch of IO optimizations
only work for scalars.
Calling it before lower_io_to_scalar_early allows nir_link_opt_varyings
to remove duplicated inputs and replace constant inputs.
fossils-db (Navi10):
Totals from 294 (0.22% of 136546) affected shaders:
CodeSize: 861356 -> 860224 (-0.13%); split: -0.13%, +0.00%
Instrs: 161972 -> 161832 (-0.09%); split: -0.09%, +0.00%
Cycles: 1185680 -> 1185120 (-0.05%); split: -0.05%, +0.00%
SMEM: 31422 -> 31424 (+0.01%)
Copies: 9065 -> 9068 (+0.03%)
Only Talos and Dark Souls 3 are affected.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7041>
While we are it, give it a name.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7065>
The game appears to use HLSL, so this workaround now lives in
SPIR-V -> NIR.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7062>
With us not creating a bo_list anymore, there is a problem if we
delete a buffer between enumerating all buffers and doing the submission.
Also changes this to a rwlock given the wider scope of the things under
lock. (especialy some of the syncobj stuff is now under the lock)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7091>
To match ACO.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7022>
There is no longer gaps in the ESGS ring.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7022>
It's actually simpler for the backend to know the variable location.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7022>
It's no longer needed to do that.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7010>
This results in at least one less radv_optimize_nir() iteration.
No fossil-db changes.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6891>
The fce metadata can always be set to false as we don't care about
the compressed clear color.
Avoiding useless fast clear eliminates improves basemark performance by
1%-1.5%.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7005>
The LLVM backend has a trick which helps reduce LDS bank conflicts
by swizzling the LDS address where each vertex is emitted.
This commit implements the same thing for ACO.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
ACO NGG GS now supports everything we need except streamout
(aka. transform feedback), but we don't use NGG anyway when
streamout is needed.
Also add a note to the new features txt.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
In each GS thread, we calculate the number of "real" primitives that
were emitted (points, lines, triangles, not strips). Then we
accumulate the number of "real" primitives emitted by the
entire threadgroup in GDS.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
Merged shaders have a workgroup barrier which makes sure that
the first half is completed in every wave before the 2nd half
is started.
This barrier is located in divergent control flow, so that waves
that don't have any invocations in the 2nd half can finish as early
as possible. This is problematic for NGG GS because it has more
workgroup barriers after the 2nd half.
So, for NGG GS we need to put the barrier outside
control flow because otherwise the waves that have 0 GS threads
won't be able to wait for the waves which have non-zero GS threads.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
We store emitted GS vertices in LDS.
Then, at the end of the shader, the emitted vertices are compacted
and each thread loads a single vertex from LDS in order to export
a primitive as needed, and the vertex attributes.
The reason this is done is because there is an impedance mismatch
between how API GS and the NGG HW works. API GS can emit an arbitrary
number of vertices and primites in each thread, but NGG HW can only
export one vertex per thread.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
This function calculates two things at once:
1. The total number of vertices emitted by the threadgroup.
2. Exclusive scan of emitted vertex count accross the threadgroup.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
For NGG GS, we need to store the following in LDS:
1. The ESGS ring, similarly to legacy ESGS.
2. Emitted vertices from the GS threads.
3. Temporary space used by the workgroup scan.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
Make it possible for ACO to recognize when to use HW NGG GS.
Also add a few notes about the various GS stages in the comments.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
NGG GS need to use the same instructions to export vertex
attributes at the end.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
Make the NGG VS/TES code easier to follow, give better names to
some functions and make ngg_nogs_early_prim_export a variable.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
Use lshl_or instead of lshl_add, which makes it more robust in
handling -1 and -2 indices which will now just become null
exports, which is what we want.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
Previously, this function inferred the vertex and primitive counts
from the gs_tg_info shader argument, but in case of NGG GS, it will
need to be calculated in runtime.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6964>
