Splitting self-intersecting register swaps into
3 sections was unnecessary and only worked because
the middle section was always empty for full dword
swaps.
No fossil-db changes.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11823>
This allows more potential compiler optimizations if the value is a
constant or from a scalar load.
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/11579>
After disabling SMEM stores, nir_opt_access() now does the same analysis
and we don't need this anymore. Doing it in isel is also too late if we
want to lower descriptor loads in NIR.
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/11652>
Previously, it could happen that a parallelcopy of
a definition was inserted before the instruction.
Fixes Rage 2 with GFX7.
No fossil-db changes.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11571>
The Navi 1x NGG hardware can hang in certain conditions when
not every wave launched before s_sendmsg(GS_ALLOC_REQ).
As a workaround, to ensure this never happens, let's emit a
workgroup barrier at the beginning of NGG VS and TES.
Note that NGG GS already has a workgroup barrier so it doesn't
need this.
Cc: mesa-stable
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/10837>
Navi 1x GPUs have an issue: they can hang when the output vertex
and primitive counts are zero. The workaround is exporting a dummy
triangle.
This commit changes the dummy triangle's vertex so its positions
are all NaN. This should make sure the triangle is never rendered.
Cc: mesa-stable
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/10837>
This commit fixes two issues with it:
1. Prevent it from going into an infinite loop.
2. Check all uses, not just first use.
Closes: #4916
Fixes: b4e22eb482
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11361>
Instead of avoiding out-of-bounds access, avoid creating a load larger
than the original attribute. This should work just as well, since the only
situations expending a load helped was because we shrunk it first.
Also fixes a bug where a 3 component load (4 components with the first
component skipped) would be incorrectly expanded to 4 components because
the stride check would never be performed. Maybe we should avoid skipping
the first component in some situations, but I'm not sure if it's worth
the VGPR cost.
fossil-db (vega10):
Totals from 583 (0.39% of 149974) affected shaders:
CodeSize: 1496848 -> 1500868 (+0.27%); split: -0.03%, +0.30%
Instrs: 286155 -> 286575 (+0.15%); split: -0.07%, +0.22%
Latency: 2947101 -> 2946865 (-0.01%); split: -0.23%, +0.22%
InvThroughput: 797396 -> 797127 (-0.03%); split: -0.08%, +0.04%
fossil-db (polaris10):
Totals from 583 (0.39% of 151365) affected shaders:
SGPRs: 38880 -> 39216 (+0.86%)
VGPRs: 24440 -> 24356 (-0.34%)
CodeSize: 1506808 -> 1510876 (+0.27%); split: -0.01%, +0.28%
Instrs: 288735 -> 289167 (+0.15%); split: -0.06%, +0.21%
Latency: 2963263 -> 2961884 (-0.05%); split: -0.24%, +0.19%
InvThroughput: 802351 -> 801665 (-0.09%); split: -0.12%, +0.04%
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/9007>
Instead of assuming scalar alignment for an attribute, we can use the
required alignment of other attributes in a binding to expect a higher
one.
This uses the alignment of all attributes in the pipeline, not just the
ones loaded. This can create slightly better code, but could break
pipelines which relied on unused (and unaligned) attributes no being
loaded. I don't think such pipelines are allowed by the spec.
fossil-db (Sienna Cichlid):
Totals from 44350 (30.32% of 146267) affected shaders:
VGPRs: 1694464 -> 1700616 (+0.36%); split: -0.08%, +0.44%
CodeSize: 60207184 -> 58093836 (-3.51%); split: -3.51%, +0.00%
MaxWaves: 1175998 -> 1174948 (-0.09%); split: +0.02%, -0.11%
Instrs: 11763444 -> 11458952 (-2.59%); split: -2.60%, +0.01%
Latency: 70679612 -> 67062215 (-5.12%); split: -5.27%, +0.15%
InvThroughput: 11482495 -> 11362911 (-1.04%); split: -1.20%, +0.16%
VClause: 359459 -> 343248 (-4.51%); split: -6.36%, +1.85%
SClause: 422404 -> 419229 (-0.75%); split: -1.17%, +0.42%
Copies: 754384 -> 764368 (+1.32%); split: -1.74%, +3.06%
Branches: 197472 -> 197474 (+0.00%); split: -0.03%, +0.03%
PreVGPRs: 1215348 -> 1215503 (+0.01%)
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/9007>
The v_mbcnt instructions can take an extra source that they add to
the result. This is not exposed in SPIR-V but we now expose it in NIR.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11072>
When the result of these loads is only used by cross-lane instructions,
it is beneficial to use a VGPR destination. This is because this allows
to put the s_waitcnt further down, which decreases latency.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11072>
Previously there hasn't been any validation for these instructions,
but after shooting myself in the leg with it a few times, I decided
to add the validation now.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11072>
Fix up the operand size for v_sad instructions, and implement
the new NIR horizontal add. There is no viable way to do this
in SALU, so let's always use a VGPR destination.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11072>
We tried to use this instruction for a more optimal sequence,
but it turned out that it doesn't exactly work as it was
supposed to. This note is to help others who want to use it.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11072>
This allows partial writes and writes to the upper half of the destination.
fossil-db (Sienna Cichlid):
Totals from 135 (0.09% of 149839) affected shaders:
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/11113>
Doesn't seem to create incorrect code, but it is suboptimal.
No fossil-db changes.
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/11113>
When the offset is negative, reg() isn't 255. Fix this by splitting
SGPR and literal emission. While we are at it, adjust a comment
saying that literals are also accepted on GFX6 which is wrong.
Fixes another batch of robustness tests.
Cc: 21.1 mesa-stable
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/11247>
GFX6-7 are affected by a hw bug that prevents address clamping to work
correctly when the SGPR offset is used. Use the VGPR offset to fix it.
Fixes various hangs with dEQP-VK.robustness.robustness2.* on Bonaire.
Cc: 21.1 mesa-stable
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/11238>
