This introduces new vec8 and vec16 instructions (which are the only
instructions taking more than 4 sources), in order to construct 8 and 16
component vectors.
In order to avoid fixing up the non-autogenerated nir_build_alu() sites
and making them pass 16 src args for the benefit of the two instructions
that take more than 4 srcs (ie vec8 and vec16), nir_build_alu() is has
nir_build_alu_tail() split out and re-used by nir_build_alu2() (which is
used for the > 4 src args case).
v2 (Karol Herbst):
use nir_build_alu2 for vec8 and vec16
use python's array multiplication syntax
add nir_op_vec helper
simplify nir_vec
nir_build_alu_tail -> nir_builder_alu_instr_finish_and_insert
use nir_build_alu for opcodes with <= 4 sources
v3 (Karol Herbst):
fix nir_serialize
v4 (Dave Airlie):
fix serialization of glsl_type
handle vec8/16 in lowering of bools
v5 (Karol Herbst):
fix load store vectorizer
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Currently when lowering mod() we add an extra instruction so if
mod(a,b) == b then 0 is returned instead of b, as mathematically
mod(a,b) is in the interval [0, b).
But Vulkan spec has relaxed this restriction, and allows the result to
be in the interval [0, b].
This commit takes this in account to remove the extra instruction
required to return 0 instead.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2922>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2922>
When geometry shaders write a value to gl_Layer that doesn't correspond to
an existing layer in the target framebuffer the rendering behavior is
undefined according to the spec, however, there are CTS tests that trigger
this scenario on purpose, probably to ensure that nothing terrible happens.
For V3D, this situation is problematic because the binner uses the layer
index to select the offset to write into the tile state data, and we only
allocate tile state for MAX2(num_layers, 1), so we want to make sure we
don't produce values that would lead to out of bounds writes. The simulator
has an assert to catch this, although we haven't observed issues in actual
hardware it is probably best to play safe.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
It's undefined behavior UBSAN complains about, so fixing this will
reduce the noise a bit.
../src/compiler/nir/nir_clone.c:710:4: runtime error: null pointer passed as argument 2, which is declared to never be null"}
#0 0xac781be4 in clone_function ../src/compiler/nir/nir_clone.c:710"}
#1 0xac781be4 in nir_shader_clone ../src/compiler/nir/nir_clone.c:740"}
#2 0xacf99442 in panfrost_shader_compile ../src/gallium/drivers/panfrost/pan_assemble.c:54"}
#3 0xacf6b268 in panfrost_bind_shader_state ../src/gallium/drivers/panfrost/pan_context.c:1960"}
#4 0xaae326bc in set_fragment_shader ../src/mesa/state_tracker/st_cb_clear.c:135"}
#5 0xaae326bc in clear_with_quad ../src/mesa/state_tracker/st_cb_clear.c:335"}
#6 0xaae326bc in st_Clear ../src/mesa/state_tracker/st_cb_clear.c:518"}
#7 0x494d0e in deqp::gles2::TestCaseWrapper::iterate(tcu::TestCase*) (/deqp/modules/gles2/deqp-gles2+0x2ad0e)"}
#8 0x7f9cf2 in tcu::TestSessionExecutor::iterateTestCase(tcu::TestCase*) (/deqp/modules/gles2/deqp-gles2+0x38fcf2)"}
#9 0x7fa5f0 in tcu::TestSessionExecutor::iterate() (/deqp/modules/gles2/deqp-gles2+0x3905f0)"}
#10 0x7e1aac in tcu::App::iterate() (/deqp/modules/gles2/deqp-gles2+0x377aac)"}
#11 0x492d4c in main (/deqp/modules/gles2/deqp-gles2+0x28d4c)"}
#12 0xb64b9aa8 in __libc_start_main (/lib/arm-linux-gnueabihf/libc.so.6+0x1aaa8)"}
Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
I forgot why that was required, but it still is the correct thing to do.
Hit it at some point when working on implementing more CL features.
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Nir serializes uses nir_ssa_alu_instr_src_components in a few places to
determine how many components a src has, but that's not what this function
returns. It simply returns how many channels are used, which is still fine
for most of the code.
This was breaking code like this:
vec16 32 ssa_1 = intrinsic load_global
vec1 32 ssa_2 = fmax ssa_1.a, ssa_2.b
v2: make the 16bit encoding work for identify swizzles again
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Somehow adjusting maxloc based on existing outputs got lost, resulting
in the clipdist varying clobbering the position varying. Causing a
shader that had no position output in freedreno/ir3, which triggers GPU
hangs in neverball.
Fixes: d0f746b645 ("nir: Save nir_variable pointers in nir_lower_clip_vs rather than locs.")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
v2: add test
Fixes: ce9205c03b ('nir: add a load/store vectorization pass')
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> (v1)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v2)
Since a is non-negative, neither fsqrt nor frsq should return NaN. frsq
should only return Inf when fsqrt returns 0.
The changes are pretty small, but this turns a few hundred hurt shaders
in the next patch into helped shaders.
An alternative to the intBitsToFloat is to import numpy and do
np.finfo(np.float32).max. That's more explicit, but we may also want to
have specific bit encodings of float values later. I could be convinced
either way, but intBitsToFloat(0x7f7fffff) was what I implemented first.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
All Gen7+ platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 14661140 -> 14661104 (<.01%)
instructions in affected programs: 7520 -> 7484 (-0.48%)
helped: 36
HURT: 0
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.32% max: 0.61% x̄: 0.49% x̃: 0.52%
95% mean confidence interval for instructions value: -1.00 -1.00
95% mean confidence interval for instructions %-change: -0.52% -0.47%
Instructions are helped.
total cycles in shared programs: 228585416 -> 228584806 (<.01%)
cycles in affected programs: 56321 -> 55711 (-1.08%)
helped: 32
HURT: 0
helped stats (abs) min: 2 max: 98 x̄: 19.06 x̃: 10
helped stats (rel) min: 0.08% max: 6.41% x̄: 1.09% x̃: 0.65%
95% mean confidence interval for cycles value: -28.32 -9.80
95% mean confidence interval for cycles %-change: -1.63% -0.54%
Cycles are helped.
Sandy Bridge
total cycles in shared programs: 152991077 -> 152991075 (<.01%)
cycles in affected programs: 11525 -> 11523 (-0.02%)
helped: 2
HURT: 2
helped stats (abs) min: 2 max: 4 x̄: 3.00 x̃: 3
helped stats (rel) min: 0.07% max: 0.11% x̄: 0.09% x̃: 0.09%
HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2
HURT stats (rel) min: 0.08% max: 0.08% x̄: 0.08% x̃: 0.08%
95% mean confidence interval for cycles value: -5.27 4.27
95% mean confidence interval for cycles %-change: -0.16% 0.15%
Inconclusive result (value mean confidence interval includes 0).
No changes on Iron Lake or GM45.
In many cases, fsat, fneg, fabs, ineg, and iabs will get folded into
another instruction as either source or destination modifiers.
Counting them as instructions means that some if-statements won't get
converted to selects. For example,
vec1 32 ssa_25 = flt32 ssa_0, ssa_23.x
/* succs: block_1 block_2 */
if ssa_25 {
block block_1:
/* preds: block_0 */
vec1 32 ssa_26 = fabs ssa_24
vec1 32 ssa_27 = fneg ssa_26
vec1 32 ssa_28 = fabs ssa_20
vec1 32 ssa_29 = fneg ssa_28
vec1 32 ssa_30 = fmul ssa_27, ssa_29
vec1 32 ssa_31 = fsat ssa_30
/* succs: block_3 */
} else {
block block_2:
/* preds: block_0 */
/* succs: block_3 */
}
block block_3:
/* preds: block_1 block_2 */
block_1 isn't really 6 instructions, but it will be counted that way.
Most callers of the peephole_select pass use either 1 or 8. It's very
easy to blow way past either of these limits with things that are really
only one or two actual instructions.
I also tried some fancier things like making sure the fsat was of
another SSA def from the same block, but the simple test was actually
better.
The i965 back-end SEL peephole pass still helps ~700 shaders in
shader-db with this change.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
All Gen6+ platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 14743694 -> 14738910 (-0.03%)
instructions in affected programs: 156575 -> 151791 (-3.06%)
helped: 1204
HURT: 0
helped stats (abs) min: 1 max: 27 x̄: 3.97 x̃: 3
helped stats (rel) min: 0.15% max: 19.57% x̄: 5.15% x̃: 4.55%
95% mean confidence interval for instructions value: -4.12 -3.82
95% mean confidence interval for instructions %-change: -5.35% -4.95%
Instructions are helped.
total cycles in shared programs: 231749141 -> 231602916 (-0.06%)
cycles in affected programs: 2818975 -> 2672750 (-5.19%)
helped: 876
HURT: 322
helped stats (abs) min: 2 max: 788 x̄: 180.99 x̃: 220
helped stats (rel) min: <.01% max: 43.82% x̄: 20.75% x̃: 19.44%
HURT stats (abs) min: 1 max: 1188 x̄: 38.27 x̃: 20
HURT stats (rel) min: 0.09% max: 102.67% x̄: 5.17% x̃: 1.70%
95% mean confidence interval for cycles value: -130.47 -113.64
95% mean confidence interval for cycles %-change: -14.85% -12.72%
Cycles are helped.
total sends in shared programs: 730495 -> 730491 (<.01%)
sends in affected programs: 46 -> 42 (-8.70%)
helped: 2
HURT: 0
Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8122757 -> 8122617 (<.01%)
instructions in affected programs: 14716 -> 14576 (-0.95%)
helped: 46
HURT: 1
helped stats (abs) min: 1 max: 8 x̄: 3.07 x̃: 3
helped stats (rel) min: 0.36% max: 10.00% x̄: 2.54% x̃: 1.06%
HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel) min: 1.59% max: 1.59% x̄: 1.59% x̃: 1.59%
95% mean confidence interval for instructions value: -3.42 -2.54
95% mean confidence interval for instructions %-change: -3.28% -1.62%
Instructions are helped.
total cycles in shared programs: 188510100 -> 188509780 (<.01%)
cycles in affected programs: 58994 -> 58674 (-0.54%)
helped: 32
HURT: 1
helped stats (abs) min: 2 max: 96 x̄: 10.06 x̃: 6
helped stats (rel) min: 0.05% max: 15.29% x̄: 1.37% x̃: 0.31%
HURT stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2
HURT stats (rel) min: 0.68% max: 0.68% x̄: 0.68% x̃: 0.68%
95% mean confidence interval for cycles value: -16.34 -3.06
95% mean confidence interval for cycles %-change: -2.46% -0.15%
Cycles are helped.
Some backends require that there are no array varyings.
If there were no arrays in the input shader, the pass shouldn't have to
create new ones.
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2103
Fixes: bcd14756ee ('nir/lower_io_to_vector: add flat mode')
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The algebraic pass was exhibiting O(n^2) behavior in
dEQP-GLES2.functional.uniform_api.random.3 and
dEQP-GLES31.functional.ubo.random.all_per_block_buffers.13 (along with
other code-generated tests, and likely real-world loop-unroll cases).
In the process of using fmul(b2f(x), b2f(x)) -> b2f(iand(x, y)) to
transform:
result = b2f(a == b);
result *= b2f(c == d);
...
result *= b2f(z == w);
->
temp = (a == b)
temp = temp && (c == d)
...
temp = temp && (z == w)
result = b2f(temp);
nir_opt_algebraic, proceeding bottom-to-top, would match and convert
the top-most fmul(b2f(), b2f()) case each time, leaving the new b2f to
be matched by the next fmul down on the next time algebraic got run by
the optimization loop.
Back in 2016 in 7be8d07732 ("nir: Do opt_algebraic in reverse
order."), Matt changed algebraic to go bottom-to-top so that we would
match the biggest patterns first. This helped his cases, but I
believe introduced this failure mode. Instead of reverting that, now
that we've got the automaton, we can update the automaton's state
recursively and just re-process any instructions whose state has
changed (indicating that they might match new things). There's a
small chance that the state will hash to the same value and miss out
on this round of algebraic, but this seems to be good enough to fix
dEQP.
Effects with NIR_VALIDATE=0 (improvement is better with validation enabled):
Intel shader-db runtime -0.954712% +/- 0.333844% (n=44/46, obvious throttling
outliers removed)
dEQP-GLES2.functional.uniform_api.random.3 runtime
-65.3512% +/- 4.22369% (n=21, was 1.4s)
dEQP-GLES31.functional.ubo.random.all_per_block_buffers.13 runtime
-68.8066% +/- 6.49523% (was 4.8s)
v2: Use two worklists, suggested by @cwabbott, to cut out a bunch of
tricky code. Runtime of uniform_api.random.3 down -0.790299% +/-
0.244213% compred to v1.
v3: Re-add the nir_instr_remove() that I accidentally dropped in v2,
fixing infinite loops.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
My motivation was to clarify the changes in the following commit, but
incidentally, it reduces runtime of
dEQP-GLES2.functional.uniform_api.random.3 (an algebraic-heavy
testcase) by -5.39524% +/- 2.21179% (n=15)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
In order to have nir_opt_algebraic be able to do further algebraic
work on the output of a replacement, we need to maintain the
automaton's state.
Reviewed-by: Eric Anholt <eric@anholt.net>
This is similar to a scheduler I've written for vc4 and i965, but this
time written at the NIR level so that hopefully it's reusable. A notable
new feature it has is Goodman/Hsu's heuristic of "once we've started
processing the uses of a value, prioritize processing the rest of their
uses", which should help avoid the heuristic otherwise making such
systematically bad choices around getting texture results consumed.
Results for v3d:
total instructions in shared programs: 6497588 -> 6518242 (0.32%)
total threads in shared programs: 154000 -> 152828 (-0.76%)
total uniforms in shared programs: 2119629 -> 2068681 (-2.40%)
total spills in shared programs: 4984 -> 472 (-90.53%)
total fills in shared programs: 6418 -> 1546 (-75.91%)
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> (v1)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> (v2)
v2: Use the DAG datastructure, fold in the scheduling-for-parallelism
patch, include SSA defs in live values so we can switch to bottom-up
if we want.
v3: Squash in improvements from Alejandro Piñeiro for getting V3D to
successfully register allocate on GLES3.1 dEQP. Make sure that
discards don't move after store_output. Comment spelling fix.
v7: run nir_opt_algebraic
v9: rework the callback function
v9: update alignment on all loads/stores, even if they're not vectorized
v10: add tests for 64-bit offsets
v10: add tests for signed offsets
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v9)
This pass combines intersecting, adjacent and identical loads/stores into
potentially larger ones and will be used by ACO to greatly reduce the
number of memory operations.
v2: handle nir_deref_type_ptr_as_array
v3: assume explicitly laid out types for derefs
v4: create less deref casts
v4: fix shared boolean vectorization
v4: fix copy+paste error in resources_different
v4: fix extract_subvector() to pass
nir_load_store_vectorize_test.ssbo_load_intersecting_32_32_64
v4: rebase
v5: subtract from deref/offset instead of scheduling offset calculations
v5: various non-functional changes/cleanups
v5: require less metadata and preserve more
v5: rebase
v6: cleanup and improve dependency handling
v6: emit less deref casts
v6: pass undef to components not set in the write_mask for new stores
v7: fix 8-bit extract_vector() with 64-bit input
v7: cleanup creation of store write data
v7: update align correctly for when the bit size of load/store increases
v7: rename extract_vector to extract_component and update comment
v8: prevent combining of row-major matrix column acceses
v9: rework process_block() to be able to vectorize more
v9: rework the callback function
v9: update alignment on all loads/stores, even if they're not vectorized
v9: remove entry::store_value, since it will not be updated if it's was
from a vectorized load
v9: fix bug in subtract_deref(), causing artifacts in Dishonored 2
v9: handle nir_intrinsic_scoped_memory_barrier
v10: use nir_ssa_scalar
v10: handle non-32-bit offsets
v10: use signed offsets for comparison
v10: improve create_entry_key_from_offset()
v10: support load_shared/store_shared
v10: remove strip_deref_casts()
v10: don't ever pass NULL to memcmp
v10: remove recursion in gcd()
v10: fix outdated comment
v11: use the new nir_extract_bits()
v12: remove use of nir_src_as_const_value in resources_different
v13: make entry key hash function deterministic
v13: simplify mask_sign_extend()
v14: add comment in hash_entry_key() about hashing pointers
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v9)
These will be useful in the upcoming load/store vectorizer.
v11: rebase
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This fixes a build failure on MSVC.
BTW, it looks like clang supports _Pragma() but I don't know if it
understands the "gcc unroll N" directive.
Signed-off-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Only NPOT vectors greater than vec4 use the extra uint32.
This is for instructions that share the dest code.
load_const and undef already support 1-16 in the header.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
vec4 scalarized ALUs typically have 4 equal instruction headers, so remove
the last 3.
There are no bits left in the ALU header for more flags, so future
extensions of NIR will have to use something like instr_type == 15
to describe more complex ALU instructions.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
It can be derived from src and var. This frees 10 bits in the header
that will be used later.
"mode" is moved in the structure, because those bits will be used for
something else later.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
- type_cast: deduplicate types if the last one is the same
- derive the type from the parent for other derefs
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The majority of constants can be packed like this.
v2: - use enum for the packing encoding,
- trim packed_value to 20 bits add 1 bit to last_component,
which simplifies a later commit
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
All of the tables are static const, so they only need to be validated
once. As noted in the previous commit, the compiler should be able to
eliminate all of this code when the assertions would pass. Even with
the help of the previous commit, this does not always occur.
-Og: -95.688 +/- 3.91935 (-24.9562% +/- 1.0222%) N=5
-O1: No difference proven at 95.0% confidence. N=5
-O2: -1.962 +/- 0.85001 (-0.860013% +/- 0.372589%) N=5
Reviewed-by: Eric Anholt <eric@anholt.net>
