In theory this wait is required for correct behaviour of discarded threads with
ATEST. Mesa usually waits before the instruction after ATEST, so this wait will
get optimized out by va_merge_flow, but as our scheduler gets more sophisticated
this could become an issue.
Let's stay on the safe side and insert the recommended wait.
No shader-db changes.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17428>
In theory, ATEST can take any combination of registers for inputs.
Experimentally, however, ATEST requires the coverage mask in R60. This avoids
regressing the following dEQP tests, which write their coverage mask with
pixel-frequency-shading but without writing to the depth/stencil buffer.
dEQP-GLES31.functional.shaders.sample_variables.sample_mask.discard_half_per_pixel.*
This issue is known to affect both Mali-G52 (v7) and Mali-G57 (v9). I am unsure
if this is a silicon bug or just an obscure implementation detail.
No shader-db changes.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17428>
va_pack asserts a large number of invariants about the instruction being packed.
If one of these fails (due to an invalid instruction), it's helpful to inspect
the failing instruction, as it may not be apparent in a large shader. Pass the
instruction through with all the assertions in va_pack for easier debugging.
Now assertion failures when packing are easier to debug:
Invalid invariant lo.type == hi.type:
= STORE.i32.flow2.wls br5, br2, wls_ptr[1], byte_offset:0
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17224>
When we assert out due to certain invalid encoding, it's helpful to know what
instruction is causing the failure, since it may not be obvious from the
assembly for large shaders. Now we get nice errors when failing:
Invalid opcode:
br0 = VAR_TEX.f32.flow8.store.skip.lod_mode.center , texture_index:0, varying_index:0
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17224>
Otherwise the list 'next' changing will cause the assertion in
list_for_each_entry to be hit.
This was not hit before because list_assert is defined for debug
builds but not debugoptimized.
Fixes: 5067a26f44 ("pan/bi: Use flow control lowering on Valhall")
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17371>
Whereas the compiler needs to know the warp size for lowering divergent
indirects, the driver needs to know it to report the subgroup size. Move the
Bifrost-specific helper to common and add the trivial implementation for
Midgard.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17265>
Float conversions with explicit rounding modes are required for OpenCL,
as well as for Vulkan with the VK_KHR_16bit_storage extension (mandatory
in Vulkan 1.1). Since the hardware conversion instructions allow
configuring the round mode, this is easy to support :-)
Fixes test_half.vstore_half_rtz.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17262>
This got messed up when scalarizing the IR. Fix the definition of the opcode to
return (instead of break, asserting out) and to respect the swizzle (instead of
failing validation). Noticed when bringing up OpenCL on Valhall.
Fixes: 5febeae58e ("pan/bi: Emit collect and split")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17222>
If a shader ends with a workgroup barrier, it must wait for slot #7 at the end
to finish the barrier. After inserting flow control, we get:
BARRIER
NOP.wait
NOP.end
Currently, the flow control pass assumes that .end implies all other control
flow, and will merge this down to
BARRIER.end
However, this is incorrect. Slot #7 is no longer waited on. In theory, this
cannot affect the correctness of the shader. In practice, the hardware checks
that all barriers are reached. Terminating without waiting on slot #7 first
raises an INSTR_BARRIER_FAULT. We need to weaken the flow control merging
slightly to avoid this incorrect merge, instead emitting:
BARRIER.wait
NOP.end
Of course, all of these cases are inefficient: terminal barriers shouldn't be
emitted in the first place. I wrote out an optimization for this. We can merge
it if we find a workload that it actually helps.
Fixes test_half.vstore_half.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17264>
This is the instruction that the hardware actually supports. Do the rename, use
the more specific accurate model in the IR, and rework the Valhall texturing
code to emit MKVEC.v2i8 instead of MKVEC.v4i8.
Will fix:
dEQP-GLES31.functional.texture.gather.offset_dynamic.implementation_offset.*
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
They are in a different place, but the encoding is otherwise as usual. This will
be required for texture gathers with dynamic offsets.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
Valhall does not have Bifrost's 4-source MKVEC.v4i8. Instead, it has a (somewhat
limtied) 3-source MKVEC.v2i8. The full MKVEC.v4i8 may be lowered to a pair of
MKVEC.v2i8 instructions.
For good code quality on both Bifrost and Valhall, we need to model both
instructions in their full generality.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
This avoids needless variation from Bifrost. While at it, fix the opcode
definition: there are no abs/neg/swizzle modifiers on the signed integer source,
and there's no clamp. However, there are round and infinity modes, like on
Bifrost.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
This will fix:
dEQP-GLES31.functional.shaders.multisample_interpolation.interpolate_at_offset.at_sample_position.default_framebuffer
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
We generate FADD_RSCALE.f32 in our sample variables implementations. Valhall
doesn't have a dedicated FADD_RSCALE.f32 implementation, it should be aliased to
FMA_RSCALE.f32. Handle that alias in isel lowering. This will fix:
dEQP-GLES31.functional.shaders.multisample_interpolation.interpolate_at_offset.*
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
When lowering vars to scratch, we need to be careful with alignment on Valhall,
where packed TLS access must not straddle a 16-byte boundary. Fixes regressions
when enabling indirect access to temps on Valhall.
Fixes: 6761dbf891 ("panfrost: Use packed TLS on Valhall")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
This was missing the message, breaking UBO-to-push and who-knows-what-else, when
enabling fp16 const buffers.
Fixes: 3dc2095b07 ("pan/bi: Model LD_BUFFER instructions")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>
This pass is super easy to unit test, so we have no excuse not to test
thoroughly. va_mark_last only inserts annotations in a shader without any
annotations, so our test cases are simply annotated shaders. The CASE macro just
has to compare the case against the case with the annotations stripped and added
back with va_mark_last.
In retrospect, I should have used that technique for the flow control insertion
tests too.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
On Valhall, register reads may be marked as "last" [1]. Setting the last flag
promises the hardware that the value of the register is no longer required. This
may enable hardware optimizations. In particular, it may permit the hardware to
avoid register file writes if a write to the marked register is still in the
forwarding buffer. This may improve power efficiency.
In principle, this is trivial: run liveness analysis and mark killed sources,
like we would in an SSA-based register allocator. In practice, there are a few
wrinkles to avoid hazards around staging registers and 64-bit register pairs,
requiring some additional data flow analysis and fix ups. However, nothing here
is particularly "hard", and all the ideas are already in use for the Bifrost
scheduler and the Bifrost/Valhall scoreboard analyses.
[1] In Mesa's compiler, this is called discard for historical reasons.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
This helps "contain the crazy" and avoids special casing BLEND in compiler
passes. The Valhall instruction is roughly the same as its Bifrost counterpart,
as long as we fix up the source order (as we already do for bitwise operations)
everything works out.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
Post-RA liveness relies on the caller updating the live variable with the
results of bi_postra_liveness_ins. It is not automatic, as with regular
liveness. This means ignoring the result of bi_postra_liveness_ins is surely an
error. Mark it as MUST_CHECK to catch that error at compile time.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
Add a unit test for the quirk discovered in the previos commit, because this
will cause flakes (instead of fails) if we get it wrong. Better have a
deterministic fail mode.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
For some unknown reason, waiting for general slots (at least for memory stores)
doesn't work properly on a BARRIER instruction. We need to wait for all general
slots right before issuing the BARRIER in addition to the general wait on the
BARRIER itself. I don't know if this is a hardware bug or some hideous
gate-saving quirk, but I observe the Mali-G78 DDK using the same workaround,
which implies this really is necessary.
Fixes rare flakes in:
dEQP-GLES31.functional.compute.shared_var.work_group_size.float_128_1_1
Note that the flakes from that test are extremely timing dependent. Without this
change, that test is racy but we almost always win the race. Reproducing the
issue reliably requires high system load (e.g. running the CTS in the
background) and simultaneously running that test a large number of times.
Minimal shader-db impact. In particular, no cycle count regressions.
total instructions in shared programs: 2699419 -> 2699458 (<.01%)
instructions in affected programs: 22014 -> 22053 (0.18%)
helped: 2
HURT: 25
helped stats (abs) min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.12% max: 0.12% x̄: 0.12% x̃: 0.12%
HURT stats (abs) min: 1.0 max: 3.0 x̄: 1.64 x̃: 1
HURT stats (rel) min: 0.07% max: 2.82% x̄: 0.69% x̃: 0.49%
95% mean confidence interval for instructions value: 1.01 1.87
95% mean confidence interval for instructions %-change: 0.38% 0.88%
Instructions are HURT.
total cvt in shared programs: 14468.81 -> 14469.42 (<.01%)
cvt in affected programs: 221.33 -> 221.94 (0.28%)
helped: 2
HURT: 25
helped stats (abs) min: 0.015625 max: 0.015625 x̄: 0.02 x̃: 0
helped stats (rel) min: 0.18% max: 0.18% x̄: 0.18% x̃: 0.18%
HURT stats (abs) min: 0.015625 max: 0.046875 x̄: 0.03 x̃: 0
HURT stats (rel) min: 0.10% max: 4.44% x̄: 1.06% x̃: 0.79%
95% mean confidence interval for cvt value: 0.02 0.03
95% mean confidence interval for cvt %-change: 0.57% 1.36%
Cvt are HURT.
total quadwords in shared programs: 1462496 -> 1462528 (<.01%)
quadwords in affected programs: 4632 -> 4664 (0.69%)
helped: 0
HURT: 4
HURT stats (abs) min: 8.0 max: 8.0 x̄: 8.00 x̃: 8
HURT stats (rel) min: 0.35% max: 7.69% x̄: 4.03% x̃: 4.03%
95% mean confidence interval for quadwords value: 8.00 8.00
95% mean confidence interval for quadwords %-change: -2.71% 10.76%
Inconclusive result (%-change mean confidence interval includes 0).
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
Test cases for insert flow are necessarily the reference test cases with the
NOPs stripped out. That means we don't need to duplicate the test bodies.
Deduplicate.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
This definition is a hardware property. It's not specific to the flow control
insertion pass, so move it to common code where other passes can use it.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17091>
Otherwise, load_push_constant won't work properly. This could probably be made
to work if we tried hard enough, but we still don't want reordering for internal
(meta) shaders which are layed out deliberately.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16916>
Bifrost supports "fast access uniforms" loaded from a single contiguous buffer.
This maps directly to Vulkan push constants, with some caveats:
* No indirect access. Indirects need to be lowered to a UBO pull.
* Strict alignment requirements. These will be met in practice.
Implement the NIR intrinsic and map it to the native hardware construct.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16916>
Like other optimizations, breaking this pass may not affect functional
correctness. It's also dead simple to unit test the pass, so we have no excuse
not to. Add unit tests for the functionality we currently support, since we just
extended it and want to make sure everything still works.
This includes tests for use of modifiers to get more small constants. There are
lots of subtle gotchas there, so let's add lots of unit tests to make sure we
got it right.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16862>
We need to distinguish signed integer instructions from unsigned integer
instructions, to distinguish sign-extension and zero-extension of sources.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16862>
CSEL executes on the conversion unit (CVT), while MUX executes on the special
function unit (SFU). Throughput on CVT is 4x higher than SFU, so this is
(almost) always an optimization.
The "real" MUX is still used for unusual cases, like 8-bit and bitselect.
Note that it's easier for us to use MUX everywhere for the IR. This is an easy
fixup to get better codegen on Valhall without touching the core Bifrost code.
shader-db is a bit of a toss up: register pressure and instruction count are
hurt in some cases due to restrictions on FAU access. In particular, a shader
that muxes between two uniforms needs an extra move due to extra constant
(zero). However, in terms of throughput this is still a win: 2 CVT instructions
(MOV + CSEL) have 2x throughput to 1 SFU instruction (MUX). The MOV has
opportunities for CSE, but that can hurt pressure in turn. Overall, cycles are
helped substantially.
total instructions in shared programs: 2728438 -> 2731597 (0.12%)
instructions in affected programs: 414391 -> 417550 (0.76%)
helped: 87
HURT: 1063
helped stats (abs) min: 1.0 max: 6.0 x̄: 5.17 x̃: 6
helped stats (rel) min: 0.19% max: 15.79% x̄: 4.12% x̃: 4.11%
HURT stats (abs) min: 1.0 max: 56.0 x̄: 3.40 x̃: 2
HURT stats (rel) min: 0.11% max: 23.43% x̄: 1.15% x̃: 0.63%
95% mean confidence interval for instructions value: 2.47 3.03
95% mean confidence interval for instructions %-change: 0.61% 0.90%
Instructions are HURT.
total cycles in shared programs: 142103 -> 142015.75 (-0.06%)
cycles in affected programs: 1263.45 -> 1176.20 (-6.91%)
helped: 281
HURT: 176
helped stats (abs) min: 0.015625 max: 2.234375 x̄: 0.50 x̃: 0
helped stats (rel) min: 0.71% max: 54.17% x̄: 16.93% x̃: 15.31%
HURT stats (abs) min: 0.015625 max: 30.0 x̄: 0.30 x̃: 0
HURT stats (rel) min: 0.84% max: 120.00% x̄: 7.16% x̃: 5.00%
95% mean confidence interval for cycles value: -0.33 -0.05
95% mean confidence interval for cycles %-change: -9.08% -6.22%
Cycles are helped.
total cvt in shared programs: 13983.34 -> 14891.70 (6.50%)
cvt in affected programs: 7498.36 -> 8406.72 (12.11%)
helped: 71
HURT: 4711
helped stats (abs) min: 0.0625 max: 0.0625 x̄: 0.06 x̃: 0
helped stats (rel) min: 5.41% max: 40.00% x̄: 10.23% x̃: 9.30%
HURT stats (abs) min: 0.015625 max: 2.640625 x̄: 0.19 x̃: 0
HURT stats (rel) min: 0.18% max: 141.18% x̄: 16.21% x̃: 9.52%
95% mean confidence interval for cvt value: 0.18 0.20
95% mean confidence interval for cvt %-change: 15.21% 16.42%
Cvt are HURT.
total sfu in shared programs: 11320.44 -> 7882.56 (-30.37%)
sfu in affected programs: 7618.50 -> 4180.62 (-45.13%)
helped: 4782
HURT: 0
helped stats (abs) min: 0.0625 max: 10.5625 x̄: 0.72 x̃: 0
helped stats (rel) min: 1.34% max: 100.00% x̄: 41.91% x̃: 37.50%
95% mean confidence interval for sfu value: -0.75 -0.68
95% mean confidence interval for sfu %-change: -42.68% -41.14%
Sfu are helped.
total ls in shared programs: 129660 -> 129690 (0.02%)
ls in affected programs: 25 -> 55 (120.00%)
helped: 0
HURT: 1
total quadwords in shared programs: 1482728 -> 1484128 (0.09%)
quadwords in affected programs: 58624 -> 60024 (2.39%)
helped: 24
HURT: 195
helped stats (abs) min: 8.0 max: 8.0 x̄: 8.00 x̃: 8
helped stats (rel) min: 3.70% max: 20.00% x̄: 10.34% x̃: 10.00%
HURT stats (abs) min: 8.0 max: 24.0 x̄: 8.16 x̃: 8
HURT stats (rel) min: 1.41% max: 50.00% x̄: 4.84% x̃: 2.56%
95% mean confidence interval for quadwords value: 5.70 7.09
95% mean confidence interval for quadwords %-change: 2.22% 4.14%
Quadwords are HURT.
total spills in shared programs: 125 -> 127 (1.60%)
spills in affected programs: 0 -> 2
helped: 0
HURT: 1
total fills in shared programs: 800 -> 828 (3.50%)
fills in affected programs: 0 -> 28
helped: 0
HURT: 1
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16857>
It's portable, and useful to both Bifrost and Valhall, in the clause scheduler
and in an instruction selection respectively. Move it from the Bifrost clause
scheduler to common code so we can share the benefits.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16857>
It's unnecessary and breaks the empty shader optimizations. Noticed while
inspecting a trace from dEQP-GLES3.functional.color_clear.masked_scissored_rgb,
which does not produce any varyings other than gl_Position in its vertex shader
and hence should omit the varying shader.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16868>
We now have infrastructure in place to generate variants of vertex shaders
specialized for transform feedback. All that's left is launching these
compute-like kernels before the IDVS job, implementing both the
transform feedback and the regular rasterization pipeline. This implements
transform feedback on Valhall, passing the relevant GLES3.1 tests.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15720>
Translate the intrinsics we introduced to lower away transform feedback into
Panfrost system values which the GL driver can handle.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15720>