For example:
s2: %688:s[32-33] = p_linear_phi %3:s[10-11], %688:s[32-33]
would have been considered trivial.
This might happen due to parallelcopies when assigning phi registers.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 69b6069dd2 ("aco: refactor try_remove_trivial_phi() in RA")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8645>
Currently, these cannot be vectorized as in NIR
shift operands are 32bit while for 16bit-vectorization
they need to be 16bit.
No fossildb changes.
Fixes: fcd2ef23e5 ('radv: vectorize 16bit instructions')
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8612>
This will allow to propagate and emit sub-register constants
on all hardware generations.
Also fixes GFX8 constant emission to not use SDWA.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8260>
It could happen that due to inconsistent copy-propagation
v1 = p_parallelcopy v2b
instructions were left after optimization on GFX8.
Cc: 20.3
Cc: 21.0
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8260>
When NGG is used, the hw can't know the number of geometry shader
primitives. To fix that, the NGG geometry shader accumulates itself
the number of primitives by using an atomic operation directly to GDS.
Then, begin/query copy the start/stop values from GDS to the
query pool buffer using a PS_DONE event. This was actually wrong
because PS_DONE is completely asynchronous to everything and executed
when the preceding draws finish pixel shaders.
Fix this by using a COPY_DATA packet which is synced with CP. This
fixes random failures on Sienna Cichlid with
dEQP-VK.query_pool.statistics_query.*.geometry_shader_primitives.*.
Cc: <mesa-stable@lists.freedesktop.org>
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/8590>
Fixes several dEQP-VK.robustness.robustness2.* tests on GFX8. Generations
other than GFX8 don't fail the tests because bounds-checking is done using
the index (making it per-vertex).
fossil-db (Polaris):
Totals from 1387 (0.99% of 140385) affected shaders:
(no statistics affected)
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Fixes: 03a0d39366 ("aco: use MUBUF in some situations instead of splitting vertex fetches")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7834>
We check below if HTILE is in compressed state, so checking if
the image has HTILE is useless because radv_layout_is_htile_compressed()
will return FALSE if no HTILE.
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/8579>
This is already checked in radv_handle_depth_image_transition().
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/8579>
From comment in emit_mimg():
We don't need the bias, sample index, compare value or offset to be
computed in WQM but if the p_create_vector copies the coordinates, then it
needs to be in WQM.
fossil-db (GFX10.3):
Totals from 1778 (1.28% of 139391) affected shaders:
SGPRs: 105080 -> 105072 (-0.01%); split: -0.02%, +0.01%
VGPRs: 96800 -> 96776 (-0.02%); split: -0.07%, +0.05%
CodeSize: 10001120 -> 10001384 (+0.00%); split: -0.04%, +0.04%
MaxWaves: 18164 -> 18163 (-0.01%)
Instrs: 1883750 -> 1883598 (-0.01%); split: -0.06%, +0.05%
Cycles: 34800176 -> 34767840 (-0.09%); split: -0.10%, +0.01%
We don't have a p_create_vector if we use NSA.
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/8523>
In some rare cases, L2 needs to be flushed if an image is affected
by the pipe misaligned issue. This is roughly based on AMDVLK.
I confirmed that disabling TC-compat HTILE, and respectively DCC,
for the relevant images also fixes the regressions below.
This fixes some regressions introduced with L2 coherency for
dEQP-VK.renderpass2.depth_stencil_resolve.image_2d_* and for
dEQP-VK.renderpass2.suballocation.multisample_resolve.*.
Fixes: 4a783a3c78 ("radv: Use L2 coherency on GFX9+.")
Co-Authored-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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/8557>
The driver used to invalidate the vector cache for meta operations
but this has been removed and I think it should be restored to fix
a bunch of regressions on GFX8.
This probably needs to be cleaned up but this is a hotfix.
This fixes a bunch of regressions and flakes on GFX8 like
dEQP-VK.pipeline.multisample.sample_locations_ext.draw.color.samples_4.*.
Fixes: 8f8d72af55 ("radv: Use access helpers for flushing with meta operations.")
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/8573>
I don't know why this wasn't enabled but I think it should be.
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/8562>
v_or_b32 with a v2b definition should use SDWA if is_partial=true.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 56345b8c61 ("aco: allow reading/writing upper halves/bytes when possible")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8577>
This restores the previous logic because L2 coherency was fully
implemented. It appears that flushing L2 metadata with a CS_DONE
event hangs.
This fixes GPU hangs with Monster Hunter World.
Fixes: 4a783a3c ("radv: Use L2 coherency on GFX9+.")
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/8566>
The flush VA space was only allocated for command buffers on the
graphics queue. Also, the ZPASS_DONE event should never be emitted
on compute queues because it hangs.
Invalidating the L2 metadata cache is only required for coherency
between the RBs and L2, so only on the graphics queue.
The L2 cache is invalidated at beginning of any IBs and that should
also invalidate the L2 metadata cache for compute anyways.
Fixes: 4a783a3c ("radv: Use L2 coherency on GFX9+.")
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/8494>
