We originally had a single lower_fmod option. In commit 2ab2d2e5, Sam
split 32 and 64-bit lowering into separate flags, with the rationale
that some drivers might want different options there. This left 16-bit
unhandled, so Iago added a lower_fmod16 option in commit ca31df6f.
Now that lower_fmod64 is gone (in favor of nir_lower_doubles and
nir_lower_dmod), we re-combine lower_fmod16 and lower_fmod32 into a
single lower_fmod flag again. I'm not aware of any hardware which
need lowering for one bitsize and not the other.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
nir_lower_doubles offers a wide variety of fp64 lowering, including
lowering fmod@64. The version there also better handles imprecisions
due to lowered frcp@64. Let's consolidate on one version.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Since NIR_PASS no longer swaps out the NIR pointer when NIR_TEST_* is
enabled, we can just take a single pointer and not a pointer to pointer.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Now that NIR_TEST_* doesn't swap the shader out from under us, it's
sufficient to just modify the shader rather than having to return in
case we're testing serialization or cloning.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
With 8 and 16-bit types and anything where we have to use non-trivial
strides registersto deal with restrictions, we end up with things that
look like partial writes even though we don't care about any values in
the register except those written by that instruction. This is
particularly important when dealing with loops because liveness sees
is_partial_write and the fact that an old version from a previous loop
iteration may be valid at that point and extends all purely partially
written values to the entire loop.
This commit adds a new UNDEF instruction which does nothing (the
generator doesn't emit anything) but which does a fake write to the
register. This informs liveness that we don't care about any values
before that point so it won't consider those registers to be falsely
live. We can safely emit UNDEF instructions for all SSA values that
come in from NIR and nearly all temporaries generated by various stages
of the compiler. In particular, we need to insert UNDEF instructions
when we handle region restrictions because the newly allocated registers
are almost guaranteed to be partially written.
No shader-db changes.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110432
Reviewed-by: Matt Turner <mattst88@gmail.com>
This might be slightly faster since we're doing one read rather than
two before we decide to skip. The more important reason, however, is
because no_spill prevents us from re-spilling spill registers. In the
new world in which we don't re-calculate liveness every spill, we may
not have valid liveness for spill registers so we shouldn't even look
their live ranges up.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110825
Fixes: e99081e76d "intel/fs/ra: Spill without destroying the..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Tested-by: Tapani Pälli <tapani.palli@intel.com>
We set header_present but then pass it some random garbage. Give it g0
instead. I'm not actually sure this does anything but g0 is the usual
header data and this is what the windows driver does so it seems like a
good idea.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
(Technically this is common code, but it doesn't affect i965 or anv.)
Improves performance of GFXBench5/gl_tess_off on Skylake GT4e at 1080p
by 9.3933% +/- 0.0305157% by eliminating all spilling in the GS.
Improves performance of GFXBench5/gl_4_off (Car Chase) on Skylake GT4e
at 1080p by 0.325208% +/- 0.0842233% (n=18).
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The difference between imov and fmov has been a constant source of
confusion in NIR for years. No one really knows why we have two or when
to use one vs. the other. The real reason is that they do different
things in the presence of source and destination modifiers. However,
without modifiers (which many back-ends don't have), they are identical.
Now that we've reworked nir_lower_to_source_mods to leave one abs/neg
instruction in place rather than replacing them with imov or fmov
instructions, we don't need two different instructions at all anymore.
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Acked-by: Rob Clark <robdclark@chromium.org>
A few of our very late passes can end up generating vectors accidentally
so we need to get rid of them. The only known case of this is the ffma
peephole which generates fneg and fabs as vectors. Currently, they're
not a problem because they get turned into fmov which the back-end
compiler knows how to handle as a vector. That's about to change.
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This flag has caused more confusion than good in most cases. You can
validly use imov for floats or fmov for integers because, without source
modifiers, neither modify their input in any way. Using imov for floats
is more reliable so we go that direction.
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
For a block with a contiguous chunk of 32 vars that don't need updating,
this lets us skip 32 vars at a time. Also, by using bitscan, we only
iterate for each set bit rather than testing them all one at a time.
Looking at perf (with -O0 which is unfortunately necessary to get
reasonable back-traces), this seems to cuts about 50-60% of the time
spent in compute_start_end() which is, itself about 4-6% of the
run-time. In the real world, with a release driver build, this cuts
1.34% off a full shader-db run. (I ran shader-db 5 times in each
configuration).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Otherwise, we get an effectively random spill reg because we no longer
have the information from RA to guide us. Also, a completely clean
graph has undefined data in in_stack which is used for choosing the
spill reg so it really is non-deterministic.
Fixes: e99081e76d "intel/fs/ra: Spill without destroying the..."
Tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In the future I want to expand this to 128-bits, for vec16 support, so
lets just put the code in place to use bitset ranges now.
v2: just declare the bitset to be the max of what we should ever see
and change assert to reflect it.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Our tessellation control shaders can be dispatched in several modes.
- SINGLE_PATCH (Gen7+) processes a single patch per thread, with each
channel corresponding to a different patch vertex. PATCHLIST_N will
launch (N / 8) threads. If N is less than 8, some channels will be
disabled, leaving some untapped hardware capabilities. Conditionals
based on gl_InvocationID are non-uniform, which means that they'll
often have to execute both paths. However, if there are fewer than
8 vertices, all invocations will happen within a single thread, so
barriers can become no-ops, which is nice. We also burn a maximum
of 4 registers for ICP handles, so we can compile without regard for
the value of N. It also works in all cases.
- DUAL_PATCH mode processes up to two patches at a time, where the first
four channels come from patch 1, and the second group of four come
from patch 2. This tries to provide better EU utilization for small
patches (N <= 4). It cannot be used in all cases.
- 8_PATCH mode processes 8 patches at a time, with a thread launched per
vertex in the patch. Each channel corresponds to the same vertex, but
in each of the 8 patches. This utilizes all channels even for small
patches. It also makes conditions on gl_InvocationID uniform, leading
to proper jumps. Barriers, unfortunately, become real. Worse, for
PATCHLIST_N, the thread payload burns N registers for ICP handles.
This can burn up to 32 registers, or 1/4 of our register file, for
URB handles. For Vulkan (and DX), we know the number of vertices at
compile time, so we can limit the amount of waste. In GL, the patch
dimension is dynamic state, so we either would have to waste all 32
(not reasonable) or guess (badly) and recompile. This is unfortunate.
Because we can only spawn 16 thread instances, we can only use this
mode for PATCHLIST_16 and smaller. The rest must use SINGLE_PATCH.
This patch implements the new 8_PATCH TCS mode, but leaves us using
SINGLE_PATCH by default. A new INTEL_DEBUG=tcs8 flag will switch to
using 8_PATCH mode for testing and benchmarking purposes. We may
want to consider using 8_PATCH mode in Vulkan in some cases.
The data I've seen shows that 8_PATCH mode can be more efficient in
some cases, but SINGLE_PATCH mode (the one we use today) is faster
in other cases. Ultimately, the TES matters much more than the TCS
for performance, so the decision may not matter much.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The payload field is actually "instance" (thread number), which is used
to calculate the invocation ID.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
When we add 8_PATCH mode, this will get a bit more complex, so we may
as well start by putting it in a helper function.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Instead of re-building the interference graph every time we spill, we
modify it in place so we can avoid recalculating liveness and the whole
O(n^2) interference graph building process. We make a simplifying
assumption in order to do so which is that all spill/fill temporary
registers live for the entire duration of the instruction around which
we're spilling. This isn't quite true because a spill into the source
of an instruction doesn't need to interfere with its destination, for
instance. Not re-calculating liveness also means that we aren't
adjusting spill costs based on the new liveness. The combination of
these things results in a bit of churn in spilling. It takes a large
cut out of the run-time of shader-db on my laptop.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15311224 -> 15311360 (<.01%)
instructions in affected programs: 77027 -> 77163 (0.18%)
helped: 11
HURT: 18
total cycles in shared programs: 355544739 -> 355830749 (0.08%)
cycles in affected programs: 203273745 -> 203559755 (0.14%)
helped: 234
HURT: 190
total spills in shared programs: 12049 -> 12042 (-0.06%)
spills in affected programs: 2465 -> 2458 (-0.28%)
helped: 9
HURT: 16
total fills in shared programs: 25112 -> 25165 (0.21%)
fills in affected programs: 6819 -> 6872 (0.78%)
helped: 11
HURT: 16
Total CPU time (seconds): 2469.68 -> 2360.22 (-4.43%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This is slightly less convenient in some places but it will make it much
easier when we want to start adding nodes dynamically.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The old code was arranged by the type of interference being added. It
would set up payload registers and then add payload interference for all
VGRFs. It would set up MRFs and add MRF interference for all VGRFs.
This commit re-arranges things to be organized differently. It first
creates and sets up all RA nodes and then groups interference into two
new categories: live range and instruction interference. Once all the
RA nodes have been set up, it walks the list of VGRFs and sets up their
live range interference and then walks the list of instructions and sets
up instruction interference. This new arrangement will be advantageous
for a future patch but, at the moment, it cuts 2% off the run-time of
shader-db on my laptop.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15311224 -> 15311224 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
total cycles in shared programs: 355544739 -> 355544739 (0.00%)
cycles in affected programs: 0 -> 0
helped: 0
HURT: 0
Total CPU time (seconds): 2523.45 -> 2469.68 (-2.13%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The only use of the MRF hack these days is for spilling and there we
don't need the precise MRF usage information. If we're spilling then we
know pretty well how many MRFs are going to be used. It is possible if
the only things that are spilled have fewer SIMD channels than the
dispatch width of the shader that this may be more MRFs than needed.
That's a risk we're willing to takd.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15311100 -> 15311224 (<.01%)
instructions in affected programs: 16664 -> 16788 (0.74%)
helped: 1
HURT: 5
total cycles in shared programs: 355543197 -> 355544739 (<.01%)
cycles in affected programs: 731864 -> 733406 (0.21%)
helped: 3
HURT: 6
The hurt shaders are all SIMD32 compute shaders where we reserve enough
space for a 32-wide spill/fill but don't need it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This accomplishes two things. First, it makes interfaces which are
really private to RA private to RA. Second, it gives us a place to
store some common stuff as we go through the algorithm.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
There's no reason why we need to use the calculated payload_node_count
value which is just first_non_payload_grf aligned up. The grf_used
value will be aligned up to 16 anyway (which is a much bigger alignment)
before being handed off to hardware.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We only have one node per VGRF so this was adding way too much
interference. No idea how we didn't catch this before.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15311100 -> 15311100 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
total cycles in shared programs: 355468050 -> 355543197 (0.02%)
cycles in affected programs: 2472492 -> 2547639 (3.04%)
helped: 17
HURT: 20
Fixes: 014edff0d2 "intel/fs: Add interference between SENDS sources"
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In the last phase of the schedule and RA loop, the RA call is redundant
if we spill. Immediately afterwards, we're going to see that we
couldn't allocate without spilling and call back into RA and tell it to
go ahead and spill. We've known about it for a while but we've always
brushed over it on the theory that, if you're going to spill, you'll be
calling RA a bunch anyway and what does one extra RA hurt? As it turns
out, it hurts more than you'd expect. Because the RA interference graph
gets sparser with each spill and the RA algorithm is more efficient on
sparser graphs, the RA call that we're duplicating is actually the most
expensive call in the RA-and-spill loop.
There's another extra RA call we do that's a bit harder to see which
this also removes. If we try to compile a shader that isn't the minimum
dispatch width and it fails to allocate without spilling we call fail()
to set an error but then go ahead and do the first spilling RA pass and
only after that's complete do we detect the fail and bail out. By
making minimum dispatch widths part of the spill condition, we side-step
this problem.
Getting rid of these extra spills takes the compile time of a nasty
Aztec Ruins shader from about 28 seconds to about 26 seconds on my
laptop. It also makes shader-db 1.5% faster
Shader-db results on Kaby Lake:
total instructions in shared programs: 15311100 -> 15311100 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
total cycles in shared programs: 355468050 -> 355468050 (0.00%)
cycles in affected programs: 0 -> 0
helped: 0
HURT: 0
Total CPU time (seconds): 2524.31 -> 2486.63 (-1.49%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This can be used by both etnaviv and freedreno/a2xx as they are both vec4
architectures with some instructions being scalar-only.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
In order to set up KILL sets, the dataflow code was walking the entire
array of ACPs for every instruction. If you assume the number of ACPs
increases roughly with the number of instructions, this is O(n^2). As
it turns out, regions_overlap() is not nearly as cheap as one would like
and shows up as a significant chunk on perf traces.
This commit changes things around and instead first builds an array of
exec_lists which it uses like a hash table (keyed off ACP source or
destination) similar to what's done in the rest of the copy-prop code.
By first walking the list of ACPs and populating the table and then
walking instructions and only looking at ACPs which probably have the
same VGRF number, we can reduce the complexity to O(n). This takes the
execution time of the piglit vs-isnan-dvec test from about 56.4 seconds
on an unoptimized debug build (what we run in CI) with NIR_VALIDATE=0 to
about 38.7 seconds.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
If the destination of an ACP entry exists only within this block, then
there's no need to keep it for dataflow analysis. We can delete it from
the out_acp table and avoid growing the bitsets any bigger than we
absolutely have to. This reduces the maximum number of global ACP
entries in the vs-isnan-dvec with software fp64 on Kaby Lake from 8630
to 3942 and takes the execution time of the piglit vs-isnan-dvec test
from about 1:16.2 on an unoptimized debug build (what we run in CI) with
NIR_VALIDATE=0 to about 56.4 seconds.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
While the number of ACPs is generally not huge compared to the number of
blocks, 16 does seem a bit small. Bumping it to 64 takes the execution
time of the piglit vs-isnan-dvec test from about 1:18.1 on an unoptimized
debug build (what we run in CI) with NIR_VALIDATE=0 to about 1:16.2.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
In the FS IR we pretend that the instruction is predicated with (+f0.1)
just for flag dependency tracking purposes. Since the instruction
doesn't support predication before Haswell, we unset the predicate so we
should also unset the flag register so that we can round-trip the
disassembly.
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
On haswell, for dim instruction we encode immediate float value operand
into double float,
v2: Fix comment (Matt Turner)
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
For the W or UW (signed or unsigned word) source types, the 16-bit value
must be replicated in both the low and high words of the 32-bit
immediate value.
v2: Fix replication in other places as well
V3: fix a few nits (Matt Turner)
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Print quad value same as unsigned quad so that we can distinguish in
between quater control disassembled values for e.g 1/2/3[Q] and
immediate quad value for e.g 1Q. This allows round-tripping through the
assembler/disassembler.
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
