It seems a lot of the lowerings being run the second time were
unnecessary. In addition, when const_state is moved to the variant,
then it will become impossible to know ahead of time whether a variant
needs additional optimizing, which means that ir3_key_lowers_nir() needs
to go away. The new approach should have the same effect, since it skips
running lowerings that are unnecessary and then skips the opt loop if no
optimizations made progress, but it will work better when we move
ir3_nir_analyze_ubo_ranges() to be after variant creation.
The one maybe controversial thing I did is to make
nir_opt_algebraic_late() always happen during variant lowering. I wanted
to avoid code duplication, and it seems to me that we should push the
_late variants as far back as possible so that later opt_algebraic runs
don't miss out on optimization opportunities.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5508>
Add some helpers to properly maintain src/dst types, and in the cases
where opcode depends on src or dst type, maintain that as well.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5048>
We have to fixup the meta:split half flag, because `ir3_split_dest()` is
called before we fixup the dest type. But we should fixup both the
split src and dest, as well as the thing it is splitting.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5048>
It seems for short frag shaders, too much prefetch can be detrimental.
I think what we *really* want to do is decide after pre-RA sched, when
we also know about nop's and what the actual ir3 instruction count is.
But that will require re-working how prefetch lowering works. For now
this is a super crude heuristic to attempt to approximate a good
solution.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
If use NIR's 1-bit bool representation , we get exactly the bool behavior
the hardware provides: CMPS produces true or false, AND/OR/XOR work as
intended without extra absnegs, and we can pass those half values directly
to other CMPS. We emit an absneg for b2b1 ("turn a memory load into a
1-bit NIR boolean"), but we would have done so for the ir3_n2b() on the
use of that value anyway. The most awkward bit is that inot(a@1) is now a
sub(1, a), but we can encode the 1 as an immediate so it's fine.
No significant changes to GL_TIME_ELAPSED on my set of traces (n=21).
instructions in affected programs: 1570638 -> 1548702 (-1.40%)
nops in affected programs: 624053 -> 611381 (-2.03%)
non-nops in affected programs: 959061 -> 949797 (-0.97%)
mov in affected programs: 5258 -> 5252 (-0.11%)
cov in affected programs: 15099 -> 15902 (5.32%)
dwords in affected programs: 469600 -> 452768 (-3.58%)
last-baryf in affected programs: 162211 -> 154726 (-4.61%)
full in affected programs: 4881 -> 4797 (-1.72%)
sstall in affected programs: 173953 -> 174545 (0.34%)
(ss) in affected programs: 10922 -> 10934 (0.11%)
(sy) in affected programs: 728 -> 745 (2.34%)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4518>
So far we only handle full regs of arrays during pre-allocation.
This patch is to handle half regs of arrays and also consider the size
of half regs when finding out conflicts.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>
In some cases we need to split components out from what was already a
collect. That was making it hard to DCE unused components of the
collect. (Ie. unused components of fragcoord, etc)
So just detect this case and skip the chained collect+split.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
If I'm going to refactor a bit to use these meta instructions to also
handle input/output, then might as well cleanup the names first.
Nouveau also uses collect/split for names of these meta instructions,
and I like those names better.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We did this in some places before, but not consistantly. But it will be
useful for two-pass RA, to identify which registers have already been
assigned.
While we are cleaning this up, use __ssa_src() and new __ssa_dst()
helper more consistently. (If nothing else, this reduces the # of
callers of ir3_reg_create() to audit that we didn't miss something)
Signed-off-by: Rob Clark <robdclark@chromium.org>
We did this properly already for split/fanout. But collect was missed.
Extract out a helper to share.
This way we avoid copy propagating a mov from high or half reg into an
instruction which cannot consume a high/half reg.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
1) deduplicate IR3_SHADER_DEBUG=disasm versus fs/vs/etc handling
2) standardize shader stage name prints, in particular VERT vs BVERT
3) don't mix stderr and stdout
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
fishgl.com has a shader which does roughly:
foo = texture(...);
if (bar)
foo = texture(...);
after lowering phi webs to regs we end up w/ a vec4 reg (array). But
since it was not an indirect access, we try to skip the extra mov. This
results that the per-component fanout (split) meta instructions store
directly to the reg (array). Which doesn't work out in RA.
Signed-off-by: Rob Clark <robdclark@chromium.org>
We only ever return the shader we were passed in (but internally
modified).
Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Removes a compiler warning about uninitialized variable.
Fixes: c02ffd2700 "ir3: Use the new NIR lowering pass for integer multiplication"
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Rob Clark <robclark@gmail.com>
Reviewed-by: Eduardo Lima <elima@igalia.com>
For a6xx, we construct/emit a single VS const state used for both
binning pass and draw pass. So far we were mostly getting lucky that
there were not (obvious) mismatches between the const_state (like
different lowered immediates) between the binning and draw pass
VS ir3_shader_variant.
And I guess this situation will come up more as GS and tess is added
into the equation.
Since really everything about the const state is not specific to the
variant, move this. The main exception is lowered immediates, but these
are the last to appear in the layout, and it doesn't hurt for each new
shader variant to just append any immed's it lowers to the end of the
immediate state.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Next patch moves const_state to ir3_shader, before the compile context
is created. So move the code around in prep to call it earlier.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Combine the offsets of differenet parts of the constant space with (what
was formerly known as) ir3_driver_const_layout. Bunch of churn, but no
functional change.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Move to ir3_compiler so it doesn't depend on the compile context. Prep
work for moving constant state from variant (where we have compile
context) to shader (where we do not).
Signed-off-by: Rob Clark <robdclark@chromium.org>
And a comment.. since we are mixing units of bytes/dwords/vec4,
hopefully this will avoid some unit confusion.
Signed-off-by: Rob Clark <robdclark@chromium.org>
We have a rather big constant file and it seems that the best way to
use it is to upload all UBOs and lower UBO access the load_uniform.
Signed-off-by: Kristian H. Kristensen <hoegsberg@chromium.org>
Reviewed-by: Rob Clark <robdclark@gmail.com>
Fixes
dEQP-GLES3.functional.shaders.indexing.varying_array.vec3_dynamic_write_dynamic_loop_read
regression.
Fixes: c1a27ba9ba freedreno/ir3: HIGH reg w/a for a6xx
Signed-off-by: Rob Clark <robdclark@gmail.com>
The wrmask is handled in regmask_get()/regmask_set(), but it wasn't
being propagated from SSA src to dst. So for example, an SSBO read
value that is passed in as src2.y component to atomic op, wasn't
getting the (sy) flag set. Causing lots of fail.
Signed-off-by: Rob Clark <robdclark@gmail.com>
It seems like some instructions (noticed this w/ cat3), cannot read HIGH
regs.. cat1 (mov/cov) can, and possibly some/all of cat2.
The blob seems to stick w/ an extra mov into low regs. So lets do the
same.
This fixes WGID on a6xx, which unsurprisingly is related to a lot of
deqp compute fails.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Images and SSBOs don't map directly to the hw. They end up being part
texture and part something else. Starting with a6xx, the hack used for
a5xx to smash the image tex state into hw texture state starting from
MAX counting down won't work, because we start using tex state also for
SSBO read.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Note that image/ssbo support is currently only implemented for a5xx.
But the instruction encoding is the same for a4xx.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Fixes the following crash that happened after d6110d4d
The problem happens if we first compile a "vanilla" shader with nothing
lowered in NIR, which perform the final lowering passes on so->shader->
nir (including nir_lower_locals_to_regs()), and then later we have
compile a shader with some lowering. The second time through we would
have already done nir_lower_locals_to_regs().
Arguably this was already a bug, just one we hadn't noticed yet.
Fixes: d6110d4d54 intel/compiler: move nir_lower_bool_to_int32 before nir_lower_locals_to_regs
Signed-off-by: Rob Clark <robdclark@gmail.com>
We also enable it in all of the NIR drivers.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>