This allows building tools for cross-compiling without building gallium
or vulkan drivers unnecessarily.
Backport-to: 25.1
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Signed-off-by: Janne Grunau <j@jannau.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34693>
(cherry picked from commit 007d7418f8)
This is the most serious bug we've had in a long time due to a fundamental
misunderstanding of the hardware (due to incomplete reverse-engineering). It
caught me off guard.
The texture descriptor has "mode" bits which configure two aspects of how the
address pointer is interpreted:
* whether it is indirected, pointing to a secondary page table for sparse
* whether it writes texture access counters (for Metal's idea of sparse).
...Neither of these is a "null texture" mode.
So why did I see Apple's blob using a non-normal mode for null textures, and why
did I copy those settings?
1. Because the hardware texture access counters provide a cheap way to detect
null texture accesses after the fact, which I think their GPU debug tools
use. I'm not sure why release builds of the driver do/did that, but whatever.
2. Because I assumed Cupertino knew best and I didn't bother looking too close.
We can't use them here (without doing extra memory allocations), since then
the GPU will increment access counters. And since our null texture address used
to just be a pointer in the command buffer, that mean the GPU will trash
whatever memory happened to be 0x400 bytes after the start of the null texture
descriptor. The symptom being random faults.
This bug was caught when trying to use the zero-page instead, which raised a
permission fault when the GPU tried to write counts. Then I remembered the
sparse mechanism and had a bit of a eureka moment. Immediately followed by an
"oh, f#$&" moment as I realized how many random bugs could potentially be root
caused to this.
The fix is two-fold:
1. Use normal layout instead.
2. Set the address to the zero-page (which is a fixed VA) and detect null
textures by checking the address, instead of the mode.
The latter is a good idea anyway, but both parts needs to be done atomically to
maintain bisectability.
Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34703>
(cherry picked from commit 3eb7575679)
rather than always early killing and then hitting pathological shuffle
situations, only early-kill when we can prove that we won't need to shuffle. it
turns out that's most of the time.
even with this heuristic, we still get hurt bad in shader-db due to extra moves.
but hopefully, the #s here are small enough that we can move on with our lives
and fix this source of known unsoundness.
this is tagged for backport as it's needed to avoid a perf regression with the
previous patch.
combined stats from this commit and the previous commit:
total instrs in shared programs: 2846065 -> 2852257 (0.22%)
instrs in affected programs: 618734 -> 624926 (1.00%)
total alu in shared programs: 2329477 -> 2335534 (0.26%)
alu in affected programs: 508119 -> 514176 (1.19%)
total gprs in shared programs: 894762 -> 901327 (0.73%)
gprs in affected programs: 36946 -> 43511 (17.77%)
Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34595>
(cherry picked from commit b1e86b3eae)
shader-db stats combined with next commit. this is the rip off the bandaid, next
is the optimize. split to enable bisecting.
the code we have to shuffle clobbered killed sources is broken and, after
thinking about that for a Long time, I don't see a reasonable way to fix it. But
if we late-kill sources - or model our calculations as-if we were late-killing
souces - we never have to shuffle onto a killed source and the problem goes away
entirely.
this is similar in spirit to what NAK does. it's not "optimal", but it's sane.
Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34595>
(cherry picked from commit b88fe9b0c5)
This hurts us in two ways:
* slightly more spilling (not actually a big problem)
* slightly worse occupancy (the shaders that are "helped" here are from trying
less hard to fit at higher occupancy levels)
However, in exchange we get a LOT more flexibility in the RA.
total instrs in shared programs: 2847015 -> 2846065 (-0.03%)
instrs in affected programs: 84134 -> 83184 (-1.13%)
total alu in shared programs: 2330406 -> 2329477 (-0.04%)
alu in affected programs: 62305 -> 61376 (-1.49%)
total code size in shared programs: 20497326 -> 20491690 (-0.03%)
code size in affected programs: 586664 -> 581028 (-0.96%)
total gprs in shared programs: 894202 -> 894762 (0.06%)
gprs in affected programs: 8900 -> 9460 (6.29%)
total scratch in shared programs: 13292 -> 13304 (0.09%)
scratch in affected programs: 2924 -> 2936 (0.41%)
total threads in shared programs: 27819712 -> 27814272 (-0.02%)
threads in affected programs: 55296 -> 49856 (-9.84%)
total spills in shared programs: 907 -> 914 (0.77%)
spills in affected programs: 419 -> 426 (1.67%)
total fills in shared programs: 857 -> 862 (0.58%)
fills in affected programs: 389 -> 394 (1.29%)
Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34595>
(cherry picked from commit 7fad96d194)
This patch does a couple of things to make CL integration with drivers
as seamless as possible:
- We pull in opencl-c.h and opencl-c-base.h to stop relying on system
headers.
- Parts of libcl.h are moved to new headers that are incomplete CL-safe
variants of libc headers.
- A couple of util headers are changed to remove now unnecessary
__OPENCL_VERSION__ guards and make more headers CL safe.
- Drivers now include src/compiler/libcl and use headers like
macros.h,u_math.h instead of libcl.h.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33576>
logicop_func=COPY is different from logicop_enable due to overriding blending.
maintain the info we need to implement properly. fixes
dEQP-VK.pipeline.shader_object_unlinked_binary.logic_op_na_formats.r32g32b32a32_sfloat.copy_blend
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34426>
../src/asahi/compiler/agx_nir_lower_address.c:82:30: runtime error: passing zero to ctz(), which is not a valid argument
#0 0x56175dca5684 in pass ../src/asahi/compiler/agx_nir_lower_address.c:82
#1 0x56175dca2eda in nir_function_intrinsics_pass ../src/compiler/nir/nir_builder.h:164
#2 0x56175dca308c in nir_shader_intrinsics_pass ../src/compiler/nir/nir_builder.h:191
#3 0x56175dca68ae in agx_nir_lower_address ../src/asahi/compiler/agx_nir_lower_address.c:167
#4 0x56175dc885c0 in agx_optimize_nir ../src/asahi/compiler/agx_compile.c:3063
#5 0x56175dc9650d in agx_compile_shader_nir ../src/asahi/compiler/agx_compile.c:3827
#6 0x56175dc52148 in main ../src/asahi/clc/asahi_clc.c:359
#7 0x7fdf1c343249 in __libc_start_call_main ../sysdeps/nptl/libc_start_call_main.h:58
#8 0x7fdf1c343304 in __libc_start_main_impl ../csu/libc-start.c:360
#9 0x56175dc44760 in _start (/builds/mesa/mesa/_build/src/asahi/clc/asahi_clc+0xbd0760)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33685>
This currently treats coarse and fine derivatives the same, but Qualcomm
needs to know whether just coarse derivatives are used or fine
derivatives/quad ops are also used. Rename this to
needs_coarse_quad_helper_invocations make clear the difference from the
new field, needs_full_quad_helper_invocations.
Reviewed-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Reviewed-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Fixes: 264d8a6766 ("ir3: Set need_full_quad depending on info.fs.require_full_quads")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33862>
