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>
The new option replaces the two other _split lowering options, since
there's no need for separate options.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Rob Clark <robdclark@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4738>
The hardware is capable of automatically filling in certain values in
the VPC without writing them from the last geometry stage, like
gl_PointCoord or gl_PrimitiveID when there is no GS. However, we *do*
have to enable these outputs (i.e. set the VPC_VAR_DISABLE bit to 0) as
VPC_VAR_DISABLE is really about FS inputs rather than VS outputs. To do
this, we move the computation of the enable bits to ir3_link_add(),
which is also a nice refactor anyway. In addition we detect the PrimID
case specifically so that the driver can program the location.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4704>
If you have a struct, the var's base driver location is not the last
driver location that will be accessed in that var. We have a shader
struct member with this number for us, already. Fixes overflows in:
dEQP-GLES31.functional.program_interface_query.program_output.type.interface_blocks.out.named_block_explicit_location.struct.mat3x2
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4670>
It was ignoring the sizes of the output variables and assuming
single-slot, and failing to update num_outputs.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4670>
discard is supposed to be a terminator, killing the thread, so that it's
possible to exit main solely by a discard e.g. inside of an infinite
loop. However, it currently isn't treated as a terminator in NIR due to
workarounds turning it into demote (d3d-style kill) and even if that
were fixed, we probably wouldn't want to treat discard_if as a jump
since otherwise the scheduler wouldn't be able to schedule things around
it. So, add this workaround which inserts jump instructions as
necessary to guarantee that the program always terminates.
This fixes a hang in dEQP-VK.graphicsfuzz.while-inside-switch, which
conditionally does a discard inside an infinite loop.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4658>
The first block was being added to the list twice, once here and once in
emit_block(), leading to list corruption and infinite loops when trying
to traverse the list of blocks backwards.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4658>
Since we consume NIR, we get FRAG_RESULT_DEPTH in .x. Something must have
been working out for this code to not be trying to get an undefined value,
but go ahead and drop it now.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4668>
The intersects() function assumes that inside each instruction values
always die before they are defined, so that if the end of one range is
the same instruction as the beginning of the next then they don't
intersect. However, this isn't the case for values that become live at
the beginning of a basic block, which become live *before* the first
instruction, or instructions that die at the end of a basic block which
die after the last instruction.
For example, imagine that we have two values, A which is defined earlier
in the block and B which is defined in the last instruction of the block
and both die at the end of the basic block (e.g. are used in the next
iteration of a loop). We would compute a range for A of, say, (10, 20)
and for B of (20, 20) since each block's end_ip is the same as the ip of
the last instruction, and RA would consider them to not interfere.
There's a similar problem with values that become live at the beginning.
The fix is to offset the block's start_ip and end_ip by one so that they
don't correspond to any actual instruction. One way to think about this
is that we're adding fake instructions at the beginning and end of a
block where values become live & die. We could invert the order, so that
values consumed by each instruction are considered dead at the end of
the previous instruction, but then values that become dead at the
beginning of the basic block would incorrectly have an empty live range,
with a similar problem at the end of the basic block if we try to say
that values are defined at the beginning of the next instruction. So
the extra padding instructions are unavoidable.
This fixes an accidental infinite loop in the shader for
dEQP-VK.spirv_assembly.type.scalar.u32.switch_vert.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4614>
I had missed that LDC actually uses vec4 units for its offset. This
means that we have to create a new instruction, and lower it in
ir3_nir_lower_io_offsets, similar to the existing SSBO instructions.
Unfortunately we can't assume that loads are always vec4-aligned, so we
have to use the alignment information that NIR gives us. Unfortunately,
it's currently woefully inadequate, and will have to be fixed to give us
good codegen in the future.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4568>
Fixes (with other patches to allow these tests to run):
dEQP-VK.ycbcr.query.size_lod.vertex.*
Suggested-by: Rob Clark <robclark@gmail.com>
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4557>
I missed that this had a micro-optimization to assume that there was
only ever one source, which is no longer valid for the bindless model
since we now have a bindless handle source. Remove the optimization to
fix assertion failures with turnip.
Fixes e.g.
dEQP-VK.glsl.texture_functions.query.texturesize.sampler2d_fixed_vertex
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4548>
Do a better job of pruning when removing unused instructions, including
cleaning up dangling false-deps.
This introduces a new ssa src pointer iterator, which makes it easy to
clear links without having to think about whether it is a normal ssa
src or a false-dep.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
The src of the SFU instruction could also be array/reg (non-SSA).
Handle this case too.
The postsched cp pass makes this scenario more likely.
Fixes: cc82521de4 ("freedreno/ir3: round-robin RA")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
Realize that certain instructions make a vecN live, and account for
this, in hopes of scheduling the remaining components of the vecN
sooner.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
This replaces the depth-first search scheduler with a more traditional
ready-list scheduler. It primarily tries to reduce register pressure
(number of live values), with the exception of trying to schedule kills
as early as possible. (Earlier iterations of this scheduler had a
tendency to push kills later, and in particular moving texture fetches
which may not be necessary ahead of kills.)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
If the group pass must insert a mov to resolve conflicts, avoid the mov
appearing *after* the meta:collect whose src it is.
The current pre-RA scheduler doesn't really care about the initial
instruction order, but the new one will in some cases.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
for verbose mode, print also the instruction "cycle" (which takes into
account (rptN) and (nopN)) in addition to instruction offset.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4440>
SEL_B32 (and presumably B16) checks for 0 or nonzero in the condition
(tested by just stuffing a uniform's value into it), so there's no need to
do ir3_b2n() on it, or any preceding ir3_n2b().
instructions in affected programs: 664444 -> 659927 (-0.68%)
nops in affected programs: 267898 -> 266312 (-0.59%)
non-nops in affected programs: 420260 -> 417329 (-0.70%)
dwords in affected programs: 144032 -> 137568 (-4.49%)
last-baryf in affected programs: 10801 -> 10321 (-4.44%)
full in affected programs: 2003 -> 2002 (-0.05%)
sstall in affected programs: 76670 -> 77405 (0.96%)
(ss) in affected programs: 4515 -> 4525 (0.22%)
(sy) in affected programs: 612 -> 604 (-1.31%)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4516>
The v->binning variant is never added to shader->variants, so just free
each one as we free the nonbinning variant.
Noticed from drm-shim mode running out of open fds, since each bo ends up
with an fd.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4502>
Carve out some space at the beginning for push constants, and push them
directly, rather than remapping them to a UBO and then relying on the
UBO pushing code. Remapping to a UBO is easy now, where there's a single
table of UBO's, but with the bindless model it'll be a lot harder. I
haven't removed all the code to move the remaining UBO's over by 1,
though, because it's going to all get rewritten with bindless anyways.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4358>
This commit updates tu6_emit_vpc to selectively emit GS-specifc
configuration. Most of this is repurposed from fd6_program.c.
This also refactors `link_geometry_stages` to ir3_nir_lower_tess.c
so it can be shared between fd and tu.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4436>
We don't really track these as the ir is transformed, but it would be a
useful thing for some passes to have. So add a pass to collect this
information. It uses instr->data (generic per-pass ptr), with the
hashsets hanging under a mem_ctx for easy disposal at the end of the
pass.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4423>
Don't fold conversions into array (incl phi lowered to regs/array).
These aren't SSA. Avoids crashes in particular in frag shaders with
flow control, which would leave a dangling array write disconnect from
the original cov src.
Possibly this could be slightly relaxed, if there is no other consumer
of the src, and it were in the same block. But it would require
updating block->keeps, and taking care of barrier state. Which isn't a
thing the cf pass does currently.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4423>
Immediate offsets are currently collapsed for ldlw, but ldlw does
behave correctly with immediate values. For example,
`ldlw.u32 r0.x, l[4], 1` actually means to use the value of
regid 4 (r1.x) as the offset when we actually want it to use the
imm value of 4 as the offset.
This commit disables copy prop for ldlw offsets so the same
intrinsic gets compiled to:
mov.u32u32 r0.y, 0x00000004
ldlw.u32 r0.x, l[r0.y], 1
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4439>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4439>
Let's make it clear what includes are being added everywhere, so that
they can be cleaned up.
Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4360>
Split out the construction of registers and classes (which is the same
on all gens) from setting up conflicts. Prep to re-work how we setup
conflicts on a6xx+ which merged half/full register file.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
Since we are re-assigning the scalars anyways in the second pass, assign
them to the highest free reg in the first pass (rather than lowest) to
allow packing vecN regs as low as possible.
Note this required some changes specifically for tex instructions with a
single component writemask that is not necessarily .x, as previously
these would get assigned in the first RA pass, and since they are still
scalar, we'd end up w/ some r47.* and other similarly way-to-high
assignments after the 2nd pass.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
Using the output of the first pass isn't ideal, as it can bake in the
losses from fragmentation which the scalar pass is intended to fill in.
This gets worse when we start using "vectorish" instructions, due to
higher use of vecN values.
Instead, we can just use the outputs of the liveness analysis to get a
more accurate # of maximum live values at any point.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
Decouple the messy logic of figuring out vreg names defined/used by an
instruction from the logic of what to do about it by introducing
iterators. There is still *some* array vs ssa special casing in
ra_block_compute_live_ranges(), but less than before. And this will
avoid introducing a second copy of the def/use logic in a following
patch which uses the liveranges to calculate the maximum # of live
values (which is the optimal target for max physical register window
to round-robin within).
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
Account for the # of regs an instruction writes, and fix an off-by-one.
(We are about to replace this with calculating the register target using
the live-ranges, but in debugging that it was useful to assert() if it
chose a higher target.)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
Extract out a helper from the select_reg callback. And include all the
instructions in the hashtable, not just SFU. This will be useful in the
following commits.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
In particular, make sure we see all the shader-db stats. The format
(order) is the sameish, except split across multiple lines to make it
easier to read.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
When we have a tess or gs stage, VS outputs aren't normal varyings, so
regid is r63.x.. we shouldn't extend our registerfootprint to 64!
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
This way RA doesn't have to special case it in use/def accounting..
This gets rid of an extra level of split/collect, which shouldn't be
needed. And interferes with scheduler trying to put tex-prefetches
after inputs but before other instructions. (Otherwise it would have
to figure out which split/collects need to go before the tex-prefetch)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4272>
If precision lowering happens at GLSL IR, loop_analysis at IR doesn't
work as expected since it can't handle things like:
"(expression bool < (expression float16_t f2fmp (var_ref ndx) ) (constant float16_t (1.000000)) )"
So we'd rather do this optimization at the NIR stage.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3885>
This is what the HW provides us. If we need integer pixel centers, we
want the state tracker to do the lowering pass so that it gets to optimize
on the subtract. This is also the shader instructions that the blob is
doing on GLES, and is what Vulkan wants too, as was noted in MR !4172.
shader-db on a630:
total instructions in shared programs: 186689 -> 186168 (-0.28%)
total nops in shared programs: 66253 -> 66139 (-0.17%)
total non-nops in shared programs: 120436 -> 120029 (-0.34%)
total dwords in shared programs: 292192 -> 291168 (-0.35%)
total last-baryf in shared programs: 4810 -> 4734 (-1.58%)
total full in shared programs: 10176 -> 10195 (0.19%)
total constlen in shared programs: 54589 -> 54575 (-0.03%)
total sstall in shared programs: 24582 -> 24802 (0.89%)
total (ss) in shared programs: 3921 -> 3925 (0.10%)
total (sy) in shared programs: 1934 -> 1923 (-0.57%)
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4223>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4223>
- Add one member to the existed ir3_stream_output so that we could
assign location information from nir_xfb_info, rather than defining
new struct.
- Redefine maximum of so buffers, streams and outputs, which will be
used for turnip.
- Also enable caps for transform feedback for spirv_to_nir.
v2. Remove redefined maximums and use IR3_MAX_SO_* and add
IR3_MAX_SO_STREAMS.
v3. Remove the newly added location field so that we could keep aligned
with 32 bytes. Instead we create an array mapping between the location
and consecutive index, which is GL driver is doing.
Signed-off-by: Hyunjun Ko <zzoon@igalia.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3942>
In the second (scalar pass) use the information about # of registers
used in the first pass as the target max, and round-robin within that
range. This generally gives the post-RA sched pass more opportunities
to re-order instructions to remove nop's.
Also, we can be a bit clever when assigning dest registers for SFU
instructions, by picking the register used for it's src (if available
and already assigned). This avoids some (ss) syncs caused by write
after read hazards. (Ie. the SFU instruction will read it's own src
before writing dest.)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4071>
Doesn't take into account stalls that result from a register written in
a different block, etc. But this should be more useful than just using
number of (ss)'s by trying to estimate how costly a given sync is.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4071>
They were inserting a nop between back to back SFU instrucions. But
that doesn't actually appear to be required. And they get stripped out
later anyways before legalize.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4071>
Need to list_delinit() before we clone the instruction to split it into
individual samgpN instructions, otherwise we get list corruption.
Tested-by: Eduardo Lima Mitev <elima@igalia.com>
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3989>
1) emperically, 10 seems like a more accurate # than 4
2) push "soft" delay handling into ir3_delayslots(), as
we should also be using it to calculate the costs
that the schedulers use
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3989>
In schedule live value tracking, differentiate between half vs full
precision. Half-precision live values are less costly than full
precision.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3989>
Current scheduler thresholds try to ensure there are warps available to
switch to when hiding texture fetch latency. But if there is none to
hide, we should allow scheduler to use more registers to reduce nops.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3989>
To avoid spurious sync flags, we want to, for a6xx+, operate in terms of
half-regs, with a full precision register testing the corresponding two
half-regs that it conflicts with.
And while we are at it, stop open-coding BITSET
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3989>
This way we can also use ir3_print from computerator, which mostly
bypasses the ir3_block construct (since it doesn't need to do
scheduling, etc)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3926>
Updates for differences between fdre-a3xx's early version of ir3, and
what we have now in mesa. And updates for instruction name and syntax
changes.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3926>
Now that we have access to the interior switch statement not going through
the txs special case for coord_components, we can just use it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3728>
This lowers mediump FS outputs to fp16 in the ir3 backend. For now
this is a modest improvement, which mostly helps us whittle down the
full mediump work. Once the GLSL level support lands, then right hand
side of the store output intrinsics will be fp16 expressions and we'll
cancel out the fp16 -> fp32 -> fp 16 round trip here.
We've had different attempts at implementing this: rewriting stores in
the GLSL IR, lowering GLSL IR outputs to temporaries and inserting
conversions when writing the temporaries to the outputs. In the end,
GLSL ends up getting in the way a lot and doing it at the nir level is
easier and still possible since we have the output var precisions.
This part of the fp16 work is more of a step on the way towards full
fp16 support and will add a few extra conversion instructions:
total instructions in shared programs: 8151 -> 8163 (0.15%)
instructions in affected programs: 1187 -> 1199 (1.01%)
helped: 4
HURT: 10
total nops in shared programs: 3146 -> 3152 (0.19%)
nops in affected programs: 563 -> 569 (1.07%)
helped: 5
HURT: 10
total non-nops in shared programs: 5005 -> 5011 (0.12%)
non-nops in affected programs: 92 -> 98 (6.52%)
helped: 0
HURT: 3
total dwords in shared programs: 12832 -> 12800 (-0.25%)
dwords in affected programs: 96 -> 64 (-33.33%)
helped: 1
HURT: 0
total last-baryf in shared programs: 118 -> 115 (-2.54%)
last-baryf in affected programs: 21 -> 18 (-14.29%)
helped: 1
HURT: 0
total full in shared programs: 424 -> 417 (-1.65%)
full in affected programs: 15 -> 8 (-46.67%)
helped: 7
HURT: 0
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>
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>
This pass tries to fold f2f16 conversion into alu instructions.
This will be useful to help reduce the number of instructions once
mesa starts supporting precision lowering. For example:
add.f r0.w, r0.w, c0.x
cov.f32f16 hr2.x, r0.w
to
add.f hr2.x, r0.w, c0.x
Additionally this pass also tries to fold f2f16 conversion into load_input
instruction:
bary.f r0.x, 3, r0.w
cov.f32f16 hr0.x, r0.x
to
bary.f hr1.x, 3, r0.x
v2: Edit to not fold OPC_MAX_F and OPC_MIN_F, since that's not valid.
v3: Add OPC_ABSNEG_F to the blacklist as well.
v4: Don't remove dead cov instructions, DCE will do that later; don't
iterate through sources when a cov only has one; remove special
handling of IR3_REG_ARRAY and IR3_REG_RELATIV.
v5: Handle folding into u32.u32 movs of floats correctly, don't bail
out on IR3_REG_RELATIV or IR3_REG_ARRAY movs.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>
The previous commit leads to match immed values unexpectedly.
This makes constlen for each shader including bvert wrong.
Also fixes atan2 for mediump deqp tests.
Fixes: cbd1f47433 ("freedreno/ir3: convert back to 32-bit values for half constant registers.")
v2: Move conversion up above fabs/fneg modifier handling as well.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3737>
This lets is_same_type_reg() recognize that the dst and src of the
immediate MOV are the same and unblocks fp16 constant propagation.
Signed-off-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3737>
This means you can directly use format utils on it without having to have
your own GL enum to number-of-components switch statement (or whatever) in
your vulkan backend.
Thanks to imirkin for fixing up the nouveau driver (and a couple of core
details).
This fixes the computed qualifiers for EXT_shader_image_load_store's
non-integer sizeNxM qualifiers, which we don't have tests for.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com> (v3d)
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3355>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3355>
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>
This was somehow working to create the instructions in a random block,
and use the value in other blocks, by dumb luck. But two-pass-RA's
better choice of register assignment causes a couple dEQPs to start
failing without this fix:
dEQP-GLES3.functional.shaders.metamorphic.bubblesort_flag.variant_1
dEQP-GLES3.functional.shaders.metamorphic.bubblesort_flag.variant_2
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
Some of the aspects of tex prefetch are in common with normal tex
instructions, such as having a wrmask to control which components
are written. Add a helper for this.
This should result in actually using the prefetch wrmask to avoid
fetching unneeded components.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
ra_block_compute_live_ranges() treats zero as "not yet defined", so
probably best to not let this be a valid instruction #
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
After RA, we can schedule to increase parallelism (reduce nop's) without
worrying about increasing register pressure. This pass lets us cut down
the instruction count ~10%, and prioritize bary.f, kill, etc, which
would tend to increase register pressure if we tried to do that before
RA.
It should be more useful if RA round-robin'd register choices.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
kill (and other cat0/flow instructions) do not have a dst register.
Which was mostly harmless before, other than RA thinking it would need
a free register to write. (But nothing consumed it, so the value would
be immediately dead.) But this would cause more problems with postsched
which would see a bogus dependency.
Also, post-RA sched *does* need to see the dependency on the predicate
register.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
Originally these were nested functions, which worked nicely, giving us
the function of a local macro that was actual 'c' syntax (ie. not token
pasted macro). But these were converted to macros because clang doesn't
let us have nice gcc extensions.
Extract these back out into functions, before adding more things and
making the macros even more cumbersome.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
A post-RA sched pass will move the extra mov's to the wrong place, so
rework the fixup so it can run after RA (and therefore after postsched)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
We want to do this only once. If we have post-RA sched pass, then we
don't want to do it pre-RA. Since legalize is where we resolve the
branch/jumps, we might as well move this into legalize.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
This way we can deal with it in one place, *after* all the blocks have
been scheduled. Which will simplify life for a post-RA sched pass.
This has the benefit of already taking into account nop's that legalize
has to insert for non-delay related reasons.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
This scenario can come up with block-sched and nop-sched moved to after
RA. So lets fix it first to keep things bisectable.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>
We need the LINEAR versions for AMD_shader_explicit_vertex_parameter.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3578>
legalize is computing a lot of state that goes in the variant, let's just
store it directly instead of passing pointers around. This leaves
max_bary in place, which is doing some surprising work (overwriting the
original total_in in some cases).
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3494>
turnip can have multiple inputs with the same location, and different
location_frac.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3109>
Just make it be all SSBOs then all storage images. The remapping table
was there to make it so that the big gap present from gallium's atomic
lowering would get cleaned up, but that's no longer case. The table has
made it very hard to support Vulkan storage images, so it's time for it to
go.
This does mean that an SSBO/IBO that is only loaded (or size-queried) will
now occupy a slot in the table where it wouldn't before. This seems like
a minor cost compared to being able to drop this much logic.
With the remapping table gone, SSBO array handling for turnip just falls
out.
Fixes many array cases of
dEQP-VK.binding_model.shader_access.primary_cmd_buf.storage_buffer.*
Reviewed-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Jonathan Marek <jonathan@marek.ca> (turnip)
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3240>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3240>
This is a more explicit name now that we don't want it to be doing any
memory barrier stuff for us.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3307>
Right now, it's implemented as a no-op for everyone. For most drivers,
it's a switch case in the NIR -> whatever which just breaks. For ir3,
they already have code to delete tessellation barriers so we just add a
case to also delete memory_barrier_tcs_patch.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3307>
Setting up transitive conflicts between a full register and its two
half registers (eg r0.x and hr0.x and hr0.y) will make the half
registers conflict. They don't actually conflict and this prevents us
from using both at the same time.
Add and use a new ra helper that sets up transitive conflicts between
a register and its subregisters, except it carefully avoids the
subregister conflict.
Signed-off-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Rob Clark <robdclark@chromium.org>
These days `ctx->inputs` is the split scalar input components and
`ir->inputs` is the full vecN. This got fixed in the load_input case,
but the load_interpolated_input case was missed.
Fixes: bdf6b7018c ("freedreno/ir3: re-work shader inputs/outputs")
Signed-off-by: Rob Clark <robdclark@chromium.org>
The first pre_assign_inputs loop doesn't pre-assign sysvals, so skip the
second part for sysvals.
The sysvals don't need to be pre-assigned since the state for those isn't
shared between binning / nonbinning shaders.
Fixes assert failures in cases where the sysvals didn't end up in the same
registers for binning / nonbinning.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@gmail.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3168>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3168>
lower_mul_2x32_64 generates mul_high opcodes, and lower_mul_high is done by
nir_lower_alu, so call nir_lower_alu after nir_opt_algebraic.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Eric Anholt <eric@anholt.net>
Add some infrastructure to trace scheduler decisions. The next patch
will add some more traces, just splitting this out to reduce clutter.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Sometimes sched changes that are a win in terms of instruction count
and/or register pressure, are worse in real life, due to keeping varying
storage locked for too long. Add a shader-db stat to give this more
visibility.
Signed-off-by: Rob Clark <robdclark@chromium.org>
We were iterating over the entire 32-entry array each time, when we
can just use a bitset to know that we're only uploading from the first
entry normally.
Knocks ir3_emit_user_consts down from ~.5% of CPU to .1% on WebGL
fishtank.
Reviewed-by: Rob Clark <robdclark@chromium.org>
The logic to ensure VS and BS inputs are aligned wasn't accounting for
unused inputs in VS. This *usually* doesn't happen, but it seems it
can in the case of ARB programs?
Fixes assert:
```
fd6_program_create: Assertion `bs->inputs[i].regid == vs->inputs[i].regid' failed.
```
Fixes: 882d53d8e3 ("freedreno/ir3+a6xx: same VBO state for draw/binning")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Fixes crashes that were unnoticed in CI because debug_assert() was not
enabled (but become real crashes after the next patch):
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.ivec2_highp_geometry
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.ivec2_lowp_geometry
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.ivec2_mediump_geometry
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.uvec2_highp_geometry
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.uvec2_lowp_geometry
dEQP-GLES31.functional.shaders.builtin_functions.integer.bitfieldextract.uvec2_mediump_geometry
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
v2. [Hyunjun Ko (zzoon@igalia.com)]
Avoid using too much open code like "instr->regs[n]->flags |= FOO"
v3. [Hyunjun Ko (zzoon@igalia.com)]
Remove redundant code for both 16b and 32b operations.
Reviewed-by: Rob Clark <robdclark@gmail.com>
Prefetch only supports the basic 2D texture case, checking is_array is
needed because 1d array textures pass the coord num_components==2 test.
Fixes: 2a0d45ae ("freedreno/ir3: Add a NIR pass to select tex instructions eligible for pre-fetch")
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@gmail.com>
To make PIPE_FORMATs usable from non-gallium parts of Mesa, I want to
move their helpers out of gallium. Since u_format used
util_copy_rect(), I moved that in there, too.
I've put it in a separate directory in util/ because it's a big chunk
of related code, and it's not clear to me whether we might want it as
a separate library from libmesa_util at some point.
Closes: #1905
Acked-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We can clear the "needs" flags once we emit a flag. And also, don't
open-code the opcode name.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
For pre-fs-dispatch texture fetch, we need to assign bary_ij to r0.x,
even if it is not used in the shader (ie. only varying use is for tex
coords). But if, for example, gl_FragCoord is used, it could get
assigned on top of bary_ij, resulting in a GPU hang.
The solution to this is two-fold: (1) the inputs/outputs rework has the
benefit of making RA realize bary_ij is a vec2, even if there are no
split/collect instructions (due to no varying fetches in the shader
itself). And (2) extend the live ranges of meta:input instructions to
the first non-input, to prevent RA from assigning the same register to
multiple inputs.
Backport note: because of (1) above, a better solution for 19.3 would be
to revert f30c256ec0.
Fixes: f30c256ec0 ("freedreno/ir3: enable pre-fs texture fetch for a6xx")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
At the ir3 level, we would assume that we could use wrmask to mask
off other components of an instruction returning a vecN when they are
not used. Which would let RA use components not written for other live
values. But this is only true for tex instructions.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Allow inputs/outputs to be vecN (ie. whatever their actual size is), and
use split to get scalar components of inputs, and collect to gather up
scalar components of outputs.
The main motivation is to simplify RA, by only having to consider split/
collect to figure out where values need to land in consecutive scalar
registers, rather than having to also deal with left/right neighbors.
Because of varying packing, and the resulting fractional location
(location_frac), to implement load_input/store_output, it is still
convenient to have a table of scalar inputs/outputs. We move this to
the compile ctx (since it is only needed for nir->ir3).
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
In almost all places, the add_sysval_input() is paired directly with a
create_input(). (The one exception is frag shader ij bary coord, and
this exception will go away in a later patch.)
So go ahead and clean this up before reworking input/output handling.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This is a driver-param (loaded from uniform), not a sysval (populated by
hw into a register). So it has no value to having a sysval slot.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Currently it is always 0x1 (scalar), but that will change in a later
patch.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We can at least get rid of the if-not-NULL check in a bunch of places.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We keep kill's alive w/ keeps these days, rather than a fake output.
This condition was left over from prior to that change.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
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>
This doesn't really work, we can't necessarily just change the outputs
to half-precision like this in anything but simple cases.
Keep the shader key entry around though, eventually with proper mediump
support we could use this with a nir pass to use lower precision frag
shader outputs when the render target format has <= 16b/component.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
The instruction has 3 src regs, so `instr->regs[0..3]` are valid, but
`instr->regs[4]` is not.
```
Test case 'dEQP-GLES31.functional.shaders.linkage.es31.tessellation.varying.rules.output_superfluous_declaration'..
==29239== Invalid read of size 8
==29239== at 0x5BE9CDC: emit_cat6 (ir3.c:841)
==29239== by 0x5BEA1BF: ir3_assemble (ir3.c:921)
==29239== by 0x5BDF0A7: ir3_shader_assemble (ir3_shader.c:133)
==29239== by 0x5BDF193: assemble_variant (ir3_shader.c:162)
==29239== by 0x5BDF407: create_variant (ir3_shader.c:215)
==29239== by 0x5BDF4DB: shader_variant (ir3_shader.c:241)
==29239== by 0x5BDF553: ir3_shader_get_variant (ir3_shader.c:257)
==29239== by 0x5BA85F7: ir3_shader_variant (ir3_gallium.c:80)
==29239== by 0x5BA7703: ir3_cache_lookup (ir3_cache.c:96)
==29239== by 0x5B8B8B3: fd6_emit_get_prog (fd6_emit.h:119)
==29239== by 0x5B8C137: fd6_draw_vbo (fd6_draw.c:186)
==29239== by 0x5BB1FBB: fd_draw_vbo (freedreno_draw.c:290)
==29239== Address 0xb97f2d0 is 0 bytes after a block of size 240 alloc'd
==29239== at 0x4848D54: malloc (in /usr/lib/aarch64-linux-gnu/valgrind/vgpreload_memcheck-arm64-linux.so)
==29239== by 0x61BD35B: ralloc_size (ralloc.c:119)
==29239== by 0x61BD41B: rzalloc_size (ralloc.c:151)
==29239== by 0x5BE599B: ir3_alloc (ir3.c:45)
==29239== by 0x5BEA583: instr_create (ir3.c:984)
==29239== by 0x5BEA5DF: ir3_instr_create2 (ir3.c:1000)
==29239== by 0x5BEE317: ir3_STLW (ir3.h:1431)
==29239== by 0x5BF12D3: emit_intrinsic_store_shared_ir3 (ir3_compiler_nir.c:903)
==29239== by 0x5BF418B: emit_intrinsic (ir3_compiler_nir.c:1802)
==29239== by 0x5BF5D07: emit_instr (ir3_compiler_nir.c:2339)
==29239== by 0x5BF603F: emit_block (ir3_compiler_nir.c:2426)
==29239== by 0x5BF624B: emit_cf_list (ir3_compiler_nir.c:2474)
==29239==
```
Probably this only triggers in non-optimized builds?
Fixes: 1f3b52ce50 ("freedreno/a6xx: Add register offset for STG/LDG")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
The instruction count is (mostly) a measure of what optimization passes
can do, while # of nops is more an indication of how effectively the
scheduler is balancing register pressure vs instruction count. So track
these independently.
(There could be opportunities to rematerialize values to reduce register
pressure, swapping some nop's with other alu instructions, so nothing is
truely independent.. but it is still useful to break these stats out.)
Signed-off-by: Rob Clark <robdclark@chromium.org>