nir_const_value_for_int asserts signed bounds on the input, but we pass in an
unsigned value that would be out-of-bounds for 32-bit channels, causing the
assert to fail for 32-bit channel formats.
Fixes dEQP-VK.pipeline.monolithic.logic_op.r32_uint.* on AGXV (and probably
PanVK).
Fixes: dbd0615e7a ("nir/lower_blend: Avoid useless iand with logic ops")
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Yonggang Luo <luoyonggang@gmail.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24252>
Consider the snippet of NIR:
div 32 %447 = @load_reg (%442) (base=0, legacy_fabs=0, legacy_fneg=0)
div 32 %463 = @load_reg (%442) (base=0, legacy_fabs=0, legacy_fneg=0)
con 32 %409 = iadd %17 (0x3), %447
@store_output (%182 (0x601), %463) (base=0, wrmask=x, component=0, src_type=invalid...
@store_reg (%409, %442) (base=0, wrmask=x, legacy_fsat=0)
The load_reg's are trivial, so the %442 read will get folded into store_output.
But under the old definition, the store_reg is also trivial so it gets folded
into the iadd... causing a read-after-write hazard and invalid code generation.
The fix is to amend our definition of store_reg triviality to account for loads
getting folded in. It's not good enough that there's no intervening load_reg,
there can also be no intervening source that gets chased to a load_reg. Handle
that case as well.
Identified in dEQP-VK.geometry.input.basic_primitive.triangles_adjacency on
V3DV.
Fixes: d313eba94e ("nir: Add pass for trivializing register access")
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reported-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24153>
In order for a register load to be trivial, it cannot be used in any
block other than the one in which it is loaded. We're not currently
explicitly doing anything to ensure this invariant holds. It may be
that it holds regardless but I couldn't find any documented reason why
it should so let's explicitly handle that case. Worst case, the newly
added code does nothing.
Fixes: d313eba94e ("nir: Add pass for trivializing register access")
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24153>
Because this pass is intended to be run after out-of-SSA and directly
before injesting the NIR into the back-end, it may come after divergence
analysis and needs to preserve the divergence information. Fortunately,
since all we ever do is insert nir_op_mov, this is easy.
Fixes: d313eba94e ("nir: Add pass for trivializing register access")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24153>
This commit makes three changes:
1. Default all newly created registers divergent because this is the
safer default.
2. Make divergence analysis do something sane with register divergence.
It's not perfect because divergence analysis isn't able to prove
registers divergent based on stores but at least if someone uses
registers a bit they'll end up with safe defaults. This matches
what they'd get with nir_ssa_def_init().
3. Make the load_reg() helper automatically propagate divergence from
the register. Because the defaults for both nir_ssa_def_init() and
nir_decl_reg() are to mark everything divergent, this only means
that nir_load_reg() of a uniform reg is now uniform.
Putting all these together, nir_from_ssa should now be producing
load_reg intrinsics with the proper uniform information.
Fixes: 7229bffcb1 ("nir: Add intrinsics for register access")
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24153>
In the case of:
halt
// succs: b9
if %618 {
block b3:// preds:
break
// succs: b6
} else {
block b4: // preds: , succs: b5
}
block b5: // preds: b4
32 %556 = iadd %617, %2 (0x1)
opt_constant_if() doesn't work because stitch_blocks() can't join blocks if the
before ends in a jump and the after isn't empty.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24235>
Most of the variables in the hash table will never actually be looked up
for any given block so cloning every possible value just creates a bunch
of unrequired memcpy calls.
Here we change the code to only clone the copies array once it is
actually looked up for the first time.
The Blender shader compile time in issue #9326 improves as folows:
151.09 seconds -> 21.11 seconds
The CTS test dEQP-GLES31.functional.ubo.random.all_per_block_buffers.20
improves as follows:
1.67 seconds -> 0.92 seconds
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24227>
If kill alias results in the hash table entry holding an empty
copies array then remove the hash entry and return the dynamic array
to the unused pool.
This helps avoid hash table size getting out of control in very large
shaders.
151.09 seconds -> 118.60 seconds
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24227>
Here we change things to simply clone the entire hash table. This
is much faster than trying to rebuild it and is needed to avoid
slow compilation of very large shaders.
The Blender shader compile time in issue #9326 improves as folows:
251.29 seconds -> 151.09 seconds
The CTS test dEQP-GLES31.functional.ubo.random.all_per_block_buffers.20
improves as follows:
2.38 seconds -> 1.67 seconds
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24227>
Consider code like:
32x4 %2 = @load_interpolated_input (%1, %0 (0x0)) (base=0, component=0, dest_type=float32, io location=VARYING_SLOT_VAR0 slots=1 mediump) // Color
32x4 %3 = fabs %2
32x4 %4 = fsat %3
32x4 %5 = fsin %4
The existing logic would incorrectly tell the backend that both fabs and fsat
could be folded, and then half the shader disappears. Whoops. Fix by stopping
the folding in this case. I choose to do this check in the fsat rather than the
fabs because it's more straightforward (1 source vs N uses) but it's somewhat
arbitrary.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24116>
Consider the IR:
%0 = load_reg
%1 = fneg %0
%2 = ffloor %1
%3 = bcsel .., .., %1
Because the fneg has both foldable and non-foldable users, nir/legacy does not
fold the fneg into the load_reg. This ensures that the backend correctly emits a
dedicated fneg instruction (with the load_reg folded in) for the bcsel to use.
However, because the chasing helpers did not previously take other uses of a
modifier into account, the helpers would fuse in the fneg to the ffloor. Except
that doesn't work, because the load_reg instruction is supposed to be
eliminated. So we end up with broken chased IR:
1 = fneg r0
2 = ffloor -NULL
3 = bcsel, ..., 1
The fix is easy: only fold modifiers into ALU instructions if the modifiers can
be folded away. If we can't eliminate the modifier instruction altogether, it's
not necessarily beneficial to fold it anyway from a register pressure
perspective. So this is probably ok. With that check in place we get correct IR
1 = fneg r0
2 = ffloor 1
3 = bcsel, ..., 1
Fixes carchase/230.shader_test under softpipe.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24116>
This avoids the silly compiler versions. Some bits are slightly more
complicated, because they have to account for inverted enum values (rather than
a separate invert bit), but this is a LOT friendlier to drivers using the pass
and it makes the pass itself more readable.
The conversion functions in panfrost/panvk will go away momentarily.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24076>
A number of passes lower SSA partially to registers, do work that would be
invalid in SSA, and then go back into SSA with nir_lower_regs_to_ssa. As a step
towards replacing nir_register with intrinsics,
the nir_lower_{phis,ssa_defs}_to_regs passes are changed to produce intrinsics
instead of nir_registers, and their callers are updated to call
nir_lower_reg_intrinsics_to_ssa instead of nir_lower_regs_to_ssa to compensate.
Jointly authored with Faith.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23089>
Note the writemask handling is chosen for consistency with the rest of NIR. In
every other instance, writemask=w requires a vec4 source. This is hardcoded into
nir_validate and nir_print as what it means to have a writemask.
More importantly, consistency with how register writemasks currently work.
nir_print hides it, but r0.w = fneg ssa_1.x is actually a vec4 instruction with
source ssa_1.xxxx. As a silly example nir_dest_num_components(that) = 4 in the
old model. I realize this is quite strange coming from a scalar ISA, but it's
perfectly natural for the class of vec4 hardware for which this was designed. In
that hardware, conceptually all instructions are vec4`, so the sequence "fneg
ssa_1 and write to channel w" is implemented as "fneg a vec4 with ssa_1.x in the
last component and write that vec4 out but mask to write only the w channel".
Isn't this inefficient? It can be. To save power, Midgard has scalar ALUs in
addition to vec4 ALUs. Those details are confined to the backend VLIW scheduler;
the instruction selection is still done as vec4. This mechanism has little in
common with AMD's SALUs. Midgard has a wave size of 1, with special hacks for
derivatives.
As a result, all backends consuming register writemasks are expecting this
pattern of code. Changing the store to take a vec1 instead of a vec4 would
require changing every backend to reswizzle the sources to resurrect the vec4. I
started typing a branch to do this yesterday, but it made a mess of both Midgard
and nir-to-tgsi. Without any good reason to think it'd actually help
performance, I abandoned the idea. Getting all 15 backends converted to the
helpers is enough of a challenge without forcing 10 backends to reswizzle their
sources too.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23089>
Alyssa Rosenzweig