In a future patch, generate_ddy will want to inspect inst->exec_size.
Change generate_ddx as well for consistency.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The PLN instruction is no more. Its functionality is now implemented
using two MAD instructions with the new native-float type. Instead of
pln(16) r20.0<1>:F r10.4<0;1,0>:F r4.0<8;8,1>:F
we now have
mad(8) acc0<1>:NF r10.7<0;1,0>:F r4.0<8;8,1>:F r10.4<0;1,0>:F
mad(8) r20.0<1>:F acc0<8;8,1>:NF r5.0<8;8,1>:F r10.5<0;1,0>:F
mad(8) acc0<1>:NF r10.7<0;1,0>:F r6.0<8;8,1>:F r10.4<0;1,0>:F
mad(8) r21.0<1>:F acc0<8;8,1>:NF r7.0<8;8,1>:F r10.5<0;1,0>:F
... and in the case of SIMD8 only the first pair of MAD instructions is
used.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
If multiple instructions are emitted, special handling of things like
conditional mod and NoDDClr/NoDDChk need to be performed.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This isn't technically broken, but the next patch will make this
function report whether it generated multiple instructions, and that
information will be used to disable the application of conditional mod
by the generic code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This new type exposes the additional precision offered by the
accumulator register and will be used in the next patch to implement the
functionality of the PLN instruction using a pair of MAD instructions.
One weird thing to note: align1 ternary instructions may only have an
accumulator in the dst or src1 normally, but when src0's type is :NF
the accumulator is read.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The hardware register types' encodings have changed on Gen11. Good thing
we have that superfluous looking brw_reg_type abstraction lying around!
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
test_fuzz_compact_instruction() was attempting to modify the uint64_t
data array of a brw_inst through a pointer to uint32_t, which has
undefined behavior. This was causing the test_eu_compact unit test to
fail mysteriously for me on GCC 7 with some additional
harmless-looking changes I had applied to my tree, which happened to
affect the order instructions are emitted by GCC causing the bit
twiddling to be done after the clear_pad_bits() call which is supposed
to overwrite the same data through a pointer of different type,
leading to data corruption. A similar failure has been reported by
Vinson Lee on the master branch built with GCC 8.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105052
Tested-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This looks like it should be protected by the assume() about
nr_color_regions, but my compiler warns anyway.
Reviewed-by: Matt Turner <mattst88@gmail.com>
This code to re-set the type of the source and destination is not
necessary since we never manipulate the types. Looks like a
left over from a time where we had to retype to float temporarily
to handle 64-bit inputs.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Divide it by two as we do for other stages. This is because the
component layout qualifier is always in 32-bit units.
Fixes issues in a new CTS test (still WIP):
KHR-GL45.enhanced_layouts.varying_double_components
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Fix the following:
warning: format ‘%lx’ expects argument of type ‘long unsigned int’, but
argument 3 has type ‘uint64_t {aka long long unsigned int}.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Otherwise loop unrolling will fail to see the actual cost of
the unrolling operations when the loop body contains 64-bit integer
instructions, and very specially when the divmod64 lowering applies,
since its lowering is quite expensive.
Without this change, some in-development CTS tests for int64
get stuck forever trying to register allocate a shader with
over 50K SSA values. The large number of SSA values is the result
of NIR first unrolling multiple seemingly simple loops that involve
int64 instructions, only to then lower these instructions to produce
a massive pile of code (due to the divmod64 lowering in the unrolled
instructions).
With this change, loop unrolling will see the loops with the int64
code already lowered and will realize that it is too expensive to
unroll.
v2: Run nir_algebraic first so we can hopefully get rid of some of
the int64 instructions before we even attempt to lower them.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Ken called this out in review, but it seems I forgot to make the change.
I noticed that the control flow annotations in the fragment shader
disassembly of tests/shaders/glsl-fs-loop-continue.shader_test were not
correct, and moving this line to the correct place fixes it.
Allows nir drivers to either use a single or dual locations for
vs double inputs.
i965 uses dual locations for both OpenGL and Vulkan drivers, for
now gallium OpenGL drivers only use a single location.
The following patch will also make use of this option when
calling nir_shader_gather_info().
Reviewed-by: Karol Herbst <kherbst@redhat.com>
First we move double_inputs_read into a vs struct in the union,
double_inputs_read is only used for vs inputs so this will
save space and also allows us to add a new double_inputs field.
We add the new field because c2acf97fcc changed the behaviour
of double_inputs_read, and while it's no longer used to track
actual reads in i965 we do still want to track this for gallium
drivers.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
18fde36ced changed the way temporary
registers were allocated in lower_integer_multiplication so that we
allocate regs_written(inst) space and keep the stride of the original
destination register. This was to ensure that any MUL which originally
followed the CHV/BXT integer multiply regioning restrictions would
continue to follow those restrictions even after lowering. This works
fine except that I forgot to reset the register file to VGRF so, even
though they were assigned a number from alloc.allocate(), they had the
wrong register file. This caused some GLES 3.0 CTS tests to start
failing on Sandy Bridge due to attempted reads from the MRF:
ES3-CTS.functional.shaders.precision.int.highp_mul_fragment.snbm64
ES3-CTS.functional.shaders.precision.int.mediump_mul_fragment.snbm64
ES3-CTS.functional.shaders.precision.int.lowp_mul_fragment.snbm64
ES3-CTS.functional.shaders.precision.uint.highp_mul_fragment.snbm64
ES3-CTS.functional.shaders.precision.uint.mediump_mul_fragment.snbm64
ES3-CTS.functional.shaders.precision.uint.lowp_mul_fragment.snbm64
This commit remedies this problem by, instead of copying inst->dst and
overwriting nr, just make a new register and set the region to match
inst->dst.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103626
Fixes: 18fde36ced
Cc: "17.3" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We have to start render targets at binding table index 0 in order to use
headerless FB write messages, and in fact already assume this in a bunch
of places in the code. Let's finish that off, and not bother storing 0
in a struct to pretend to add it in a few places.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Previously the dataflow propagation algorithm would calculate the ACP
live-in and -out sets in a two-pass fixed-point algorithm. The first
pass would update the live-out sets of all basic blocks of the program
based on their live-in sets, while the second pass would update the
live-in sets based on the live-out sets. This is incredibly
inefficient in the typical case where the CFG of the program is
approximately acyclic, because it can take up to 2*n passes for an ACP
entry introduced at the top of the program to reach the bottom (where
n is the number of basic blocks in the program), until which point the
algorithm won't be able to reach a fixed point.
The same effect can be achieved in a single pass by computing the
live-in and -out sets in lock-step, because that makes sure that
processing of any basic block will pick up the updated live-out sets
of the lexically preceding blocks. This gives the dataflow
propagation algorithm effectively O(n) run-time instead of O(n^2) in
the acyclic case.
The time spent in dataflow propagation is reduced by 30x in the
GLES31.functional.ssbo.layout.random.all_shared_buffer.5 dEQP
test-case on my CHV system (the improvement is likely to be of the
same order of magnitude on other platforms). This more than reverses
an apparent run-time regression in this test-case from my previous
copy-propagation undefined-value handling patch, which was ultimately
caused by the additional work introduced in that commit to account for
undefined values being multiplied by a huge quadratic factor.
According to Chad this test was failing on CHV due to a 30s time-out
imposed by the Android CTS (this was the case regardless of my
undefined-value handling patch, even though my patch substantially
exacerbated the issue). On my CHV system this patch reduces the
overall run-time of the test by approximately 12x, getting us to
around 13s, well below the time-out.
v2: Initialize live-out set to the universal set to avoid rather
pessimistic dataflow estimation in shaders with cycles (Addresses
performance regression reported by Eero in GpuTest Piano).
Performance numbers given above still apply. No shader-db changes
with respect to master.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104271
Reported-by: Chad Versace <chadversary@chromium.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This creates two new internal dependencies, idep_nir_headers and
idep_nir. The former encapsulates the generation of nir_opcodes.h and
nir_builder_opcodes.h and adding src/compiler/nir as an include path.
This ensures that any target that needs nir headers will have the
includes and that the generated headers will be generated before the
target is build. The second, idep_nir, includes the first and
additionally links to libnir.
This is intended to make it easier to avoid race conditions in the build
when using nir, since the number of consumers for libnir and it's
headers are quite high.
Acked-by: Eric Engestrom <eric.engestrom@imgtec.com>
Signed-off-by: Dylan Baker <dylan.c.baker@intel.com>
Currently the meosn build has a mix of two styles:
arg : [foo, ...
bar],
and
arg : [
foo, ...,
bar,
]
For consistency let's pick one. I've picked the later style, which I
think is more readable, and is more common in the mesa code base.
v2: - fix commit message
Acked-by: Eric Engestrom <eric.engestrom@imgtec.com>
Signed-off-by: Dylan Baker <dylan.c.baker@intel.com>
We already had to switch all of the W types to UW to prevent issues
with vector immediates on gen10. We may as well use unsigned types
everywhere.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Gen 10 has a strange hardware bug involving V immediates with W types.
It appears that a mov(8) g2<1>W 0x76543210V will actually result in g2
getting the value {3, 2, 1, 0, 3, 2, 1, 0}. In particular, the bottom
four nibbles are repeated instead of the top four being taken. (A mov
of 0x00003210V yields the same result.) This bug does not appear in any
hardware documentation as far as we can tell and the simulator does not
implement the bug either.
Commit 6132992cdb was mostly a no-op
except that it changed the type of the subgroup invocation from UW to W
and caused us to tickle this bug with basically every compute shader
that uses any sort of invocation ID (which is most of them). This is
also potentially an issue for geometry shader input pulls and SampleID
setup. The easy solution is just to change the few places where we use
a vector integer immediate with a W type to use a UW type.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
Fixes: 6132992cdb
Some cases weren't handled, such as stride 4 which is needed for 64-bit
operations. Presumably fixes the assertion failure mentioned in commit
2d04572038 (Revert "i965/fs: Use align1 mode on ternary instructions
on Gen10+") but who can really say since the commit neglected to list
any of them!
Reviewed-by: Scott D Phillips <scott.d.phillips@intel.com>
v2: do not try to handle it as a system value directly for the SPIR-V
path. In GL we rather handle it as a uniform like we do for the
GLSL path (Jason).
v3:
- Remove the uniform variable, it is alwats -1 now (Jason)
- Also do the lowering for the TessEval stage (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Older OpenGL defines two equations for converting from signed-normalized
to floating point data. These are:
f = (2c + 1)/(2^b - 1) (equation 2.2)
f = max{c/2^(b-1) - 1), -1.0} (equation 2.3)
Both OpenGL 4.2+ and OpenGL ES 3.0+ mandate that equation 2.3 is to be
used in all scenarios, and remove equation 2.2. DirectX uses equation
2.3 as well. Intel hardware only supports equation 2.3, so Gen7.5+
systems that use the vertex fetcher hardware to do the conversions
always get formula 2.3.
This can make a big difference for 10-10-10-2 formats - the 2-bit value
can represent 0 with equation 2.3, and cannot with equation 2.2.
Ivybridge and older were using equation 2.2 for OpenGL, and 2.3 for ES.
Now that Ivybridge supports OpenGL 4.2, this is wrong - we need to use
the new rules, at least in core profile. That would leave Gen4-6 doing
something different than all other hardware, which seems...lame.
With context version promotion, applications that requested a pre-4.2
context may get promoted to 4.2, and thus get the new rules. Zero cases
have been reported of this being a problem. However, we've received a
report that following the old rules breaks expectations. SuperTuxKart
apparently renders the cars red when following equation 2.2, and works
correctly when following equation 2.3:
https://github.com/supertuxkart/stk-code/issues/2885#issuecomment-353858405
So, this patch deletes the legacy equation 2.2 support entirely, making
all hardware and APIs consistently use the new equation 2.3 rules.
If we ever find an application that truly requires the old formula, then
we'd likely want that application to work on modern hardware, too. We'd
likely restore this support as a driconf option. Until then, drop it.
This commit will regress Piglit's draw-vertices-2101010 test on
pre-Haswell without the corresponding Piglit patch to accept either
formula (commit 35daaa1695ea01eb85bc02f9be9b6ebd1a7113a1):
draw-vertices-2101010: Accept either SNORM conversion formula.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
This should shut up some Valgrind errors during pre-regalloc
scheduling. The errors were harmless since they could only have led
to the estimation of the bank conflict penalty of an instruction
pre-regalloc, which is inaccurate at that point of the program
compilation, but no less accurate than the intended "return 0"
fall-back path. The scheduling pass is normally re-run after regalloc
with a well-defined grf_used value and accurate bank conflict
information.
Fixes: acf98ff933 "intel/fs: Teach instruction scheduler about GRF bank conflict cycles."
Reported-by: Eero Tamminen <eero.t.tamminen@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The weight_vector_type constructor was inadvertently assuming C++17
semantics of the new operator applied on a type with alignment
requirement greater than the largest fundamental alignment.
Unfortunately on earlier C++ dialects the implementation was allowed
to raise an allocation failure when the alignment requirement of the
allocated type was unsupported, in an implementation-defined fashion.
It's expected that a C++ implementation recent enough to implement
P0035R4 would have honored allocation requests for such over-aligned
types even if the C++17 dialect wasn't active, which is likely the
reason why this problem wasn't caught by our CI system.
A more elegant fix would involve wrapping the __SSE2__ block in a
'__cpp_aligned_new >= 201606' preprocessor conditional and continue
taking advantage of the language feature, but that would yield lower
compile-time performance on old compilers not implementing it
(e.g. GCC versions older than 7.0).
Fixes: af2c320190 "intel/fs: Implement GRF bank conflict mitigation pass."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104226
Reported-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We still have gpu hangs on Cannonlake when using push constants, so
disable them for now until we have a proper fix for these hangs.
v2: Add warning message when creating context too.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Cc: Ben Widawsky <ben@bwidawsk.net>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
In Vulkan, we don't support classic pull constants and everything the
client asks us to push, we push. However, for pushed UBOs, we still
want to fall back to conventional pulls if we run out of space.
This rewires the logic for assigning uniform locations to work in terms
of "complex alignments". The basic idea is that, as we walk the list of
instructions, we keep track of the alignment and continuity requirements
of each slot and assert that the alignments all match up. We then use
those alignments in the compaction stage to ensure that everything gets
placed at a properly aligned register. The old mechanism handled
alignments by special-casing each of the bit sizes and placing 64-bit
values first followed by 32-bit values.
The old scheme had the advantage of never leaving a hole since all the
64-bit values could be tightly packed and so could the 32-bit values.
However, the new scheme has no type size special cases so it handles not
only 32 and 64-bit types but should gracefully extend to 16 and 8-bit
types as the need arises.
Tested-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This addresses a long-standing back-end compiler bug that could lead
to cross-channel data corruption in loops executed non-uniformly. In
some cases live variables extending through a loop divergence point
(e.g. a non-uniform break) into a convergence point (e.g. the end of
the loop) wouldn't be considered live along all physical control flow
paths the SIMD thread could possibly have taken in between due to some
channels remaining in the loop for additional iterations.
This patch fixes the problem by extending the CFG with physical edges
that don't exist in the idealized non-vectorized program, but
represent valid control flow paths the SIMD EU may take due to the
divergence of logical threads. This makes sense because the i965 IR
is explicitly SIMD, and it's not uncommon for instructions to have an
influence on neighboring channels (e.g. a force_writemask_all header
setup), so the behavior of the SIMD thread as a whole needs to be
considered.
No changes in shader-db.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This makes the dataflow propagation logic of the copy propagation pass
more intelligent in cases where the destination of a copy is known to
be undefined for some incoming CFG edges, building upon the
definedness information provided by the last patch. Helps a few
programs, and avoids a handful shader-db regressions from the next
patch.
shader-db results on ILK:
total instructions in shared programs: 6541547 -> 6541523 (-0.00%)
instructions in affected programs: 360 -> 336 (-6.67%)
helped: 8
HURT: 0
LOST: 0
GAINED: 10
shader-db results on BDW:
total instructions in shared programs: 8174323 -> 8173882 (-0.01%)
instructions in affected programs: 7730 -> 7289 (-5.71%)
helped: 5
HURT: 2
LOST: 0
GAINED: 4
shader-db results on SKL:
total instructions in shared programs: 8185669 -> 8184598 (-0.01%)
instructions in affected programs: 10364 -> 9293 (-10.33%)
helped: 5
HURT: 2
LOST: 0
GAINED: 2
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Currently the liveness analysis pass would extend a live interval up
to the top of the program when no unconditional and complete
definition of the variable is found that dominates all of its uses.
This can lead to a serious performance problem in shaders containing
many partial writes, like scalar arithmetic, FP64 and soon FP16
operations. The number of oversize live intervals in such workloads
can cause the compilation time of the shader to explode because of the
worse than quadratic behavior of the register allocator and scheduler
when running out of registers, and it can also cause the running time
of the shader to explode due to the amount of spilling it leads to,
which is orders of magnitude slower than GRF memory.
This patch fixes it by computing the intersection of our current live
intervals with the subset of the program that can possibly be reached
from any definition of the variable. Extending the storage allocation
of the variable beyond that is pretty useless because its value is
guaranteed to be undefined at a point that cannot be reached from any
definition.
According to Jason, this improves performance of the subgroup Vulkan
CTS tests significantly (e.g. the runtime of the dvec4 broadcast test
improves by nearly 50x).
No significant change in the running time of shader-db (with 5%
statistical significance).
shader-db results on IVB:
total cycles in shared programs: 61108780 -> 60932856 (-0.29%)
cycles in affected programs: 16335482 -> 16159558 (-1.08%)
helped: 5121
HURT: 4347
total spills in shared programs: 1309 -> 1288 (-1.60%)
spills in affected programs: 249 -> 228 (-8.43%)
helped: 3
HURT: 0
total fills in shared programs: 1652 -> 1597 (-3.33%)
fills in affected programs: 262 -> 207 (-20.99%)
helped: 4
HURT: 0
LOST: 2
GAINED: 209
shader-db results on BDW:
total cycles in shared programs: 67617262 -> 67361220 (-0.38%)
cycles in affected programs: 23397142 -> 23141100 (-1.09%)
helped: 8045
HURT: 6488
total spills in shared programs: 1456 -> 1252 (-14.01%)
spills in affected programs: 465 -> 261 (-43.87%)
helped: 3
HURT: 0
total fills in shared programs: 1720 -> 1465 (-14.83%)
fills in affected programs: 471 -> 216 (-54.14%)
helped: 4
HURT: 0
LOST: 2
GAINED: 162
shader-db results on SKL:
total cycles in shared programs: 65436248 -> 65245186 (-0.29%)
cycles in affected programs: 22560936 -> 22369874 (-0.85%)
helped: 8457
HURT: 6247
total spills in shared programs: 437 -> 437 (0.00%)
spills in affected programs: 0 -> 0
helped: 0
HURT: 0
total fills in shared programs: 870 -> 854 (-1.84%)
fills in affected programs: 16 -> 0
helped: 1
HURT: 0
LOST: 0
GAINED: 107
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This should allow the post-RA scheduler to do a slightly better job at
hiding latency in presence of instructions incurring bank conflicts.
The main purpuse of this patch is not to improve performance though,
but to get conflict cycles to show up in shader-db statistics in order
to make sure that regressions in the bank conflict mitigation pass
don't go unnoticed.
Acked-by: Matt Turner <mattst88@gmail.com>
Unnecessary GRF bank conflicts increase the issue time of ternary
instructions (the overwhelmingly most common of which is MAD) by
roughly 50%, leading to reduced ALU throughput. This pass attempts to
minimize the number of bank conflicts by rearranging the layout of the
GRF space post-register allocation. It's in general not possible to
eliminate all of them without introducing extra copies, which are
typically more expensive than the bank conflict itself.
In a shader-db run on SKL this helps roughly 46k shaders:
total conflicts in shared programs: 1008981 -> 600461 (-40.49%)
conflicts in affected programs: 816222 -> 407702 (-50.05%)
helped: 46234
HURT: 72
The running time of shader-db itself on SKL seems to be increased by
roughly 2.52%±1.13% with n=20 due to the additional work done by the
compiler back-end.
On earlier generations the pass is somewhat less effective in relative
terms because the hardware incurs a bank conflict anytime the last two
sources of the instruction are duplicate (e.g. while trying to square
a value using MAD), which is impossible to avoid without introducing
copies. E.g. for a shader-db run on SNB:
total conflicts in shared programs: 944636 -> 623185 (-34.03%)
conflicts in affected programs: 853258 -> 531807 (-37.67%)
helped: 31052
HURT: 19
And on BDW:
total conflicts in shared programs: 1418393 -> 987539 (-30.38%)
conflicts in affected programs: 1179787 -> 748933 (-36.52%)
helped: 47592
HURT: 70
On SKL GT4e this improves performance of GpuTest Volplosion by 3.64%
±0.33% with n=16.
NOTE: This patch intentionally disregards some i965 coding conventions
for the sake of reviewability. This is addressed by the next
squash patch which introduces an amount of (for the most part
boring) boilerplate that might distract reviewers from the
non-trivial algorithmic details of the pass.
The following patch is squashed in:
SQUASH: intel/fs/bank_conflicts: Roll back to the nineties.
Acked-by: Matt Turner <mattst88@gmail.com>
SSBO loads were using byte_scattered read messages as they allow
reading 16-bit size components. byte_scattered messages can only
operate one component at a time so we needed to emit as many messages
as components.
But for vec2 and vec4 of 16-bit, being multiple of 32-bit we can use the
untyped_surface_read message to read pairs of 16-bit components using only
one message. Once each pair is read it is unshuffled to return the proper
16-bit components. vec3 case is assimilated to vec4 but the 4th component
is ignored.
16-bit scalars are read using one byte_scattered_read message.
v2: Removed use of stride = 2 on sources (Jason Ekstrand)
Rework optimization using unshuffle 16 reads (Chema Casanova)
v3: Use W and D types insead of HF and F in shuffle to avoid rounding
erros (Jason Ekstrand)
Use untyped_surface_read for 16-bit vec3. (Jason Ekstrand)
v4: Use subscript insead of chaging type and stride (Jason Ekstrand)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Currently, we use byte-scattered write messages for storing 16-bit
into an SSBO. This is because untyped surface messages have a fixed
32-bit size.
This patch optimizes these 16-bit writes by combining 2 values (e.g,
two consecutive components aligned with 32-bits) into a 32-bit register,
packing the two 16-bit words.
16-bit single component values will continue to use byte-scattered
write messages. The same will happens when the first consecutive
component is not aligned 32-bits.
This optimization reduces the number of SEND messages used for storing
16-bit values potentially by 2 or 4, which cuts down execution time
significantly because byte-scattered writes are an expensive
operation as they only write a component for message.
v2: Removed use of stride = 2 on sources (Jason Ekstrand)
Rework optimization using shuffle 16 write and enable writes
of 16bit vec4 with only one message of 32-bits. (Chema Casanova)
v3: - Fix coding style (Eduardo Lima)
- Reorganize code to avoid duplication. (Jason Ekstrand)
- Include new comments to explain the length calculations to
fix alignment issues of components. (Jason Ekstrand)
- Fix issues with writemask yz with 16-bit writes. (Jason Ektrand)
v4: (Jason Ekstrand)
- Reorganize 64-bit ssbo-writes to avoid using slots_per_component.
- Comment about why suffle is needed when using byte_scattered_write.
Signed-off-by: Eduardo Lima <elima@igalia.com>
Signed-off-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
load_ubo is using 32-bit loads as uniforms surfaces have a 32-bit
surface format defined. So when reading 16-bit components with the
sampler we need to unshuffle two 16-bit components from each 32-bit
component.
Using the sampler avoids the use of the byte_scattered_read message
that needs one message for each component and is supposed to be
slower.
v2: (Jason Ekstrand)
- Simplify component selection and unshuffling for different bitsizes
- Remove SKL optimization of reading only two 32-bit components when
reading 16-bits types.
Reviewed-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
This helpers are used to load/store 16-bit types from/to 32-bit
components.
The functions shuffle_32bit_load_result_to_16bit_data and
shuffle_16bit_data_for_32bit_write are implemented in a similar
way than the analogous functions for handling 64-bit types.
v1: Explain need of temporary in shuffle operations. (Jason Ekstrand)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Used to enable 16-bit reads at do_untyped_vector_read, that is used on
the following intrinsics:
* nir_intrinsic_load_shared
* nir_intrinsic_load_ssbo
v2: Removed use of stride = 2 on 16-bit sources (Jason Ekstrand)
v3: - Add bitsize to scattered read operation (Jason Ekstrand)
- Remove implementation of 16-bit UBO read from this patch.
- Avoid assertion at opt_algebraic caused by ADD of two IMM with
offset with BRW_REGISTER_TYPE_UD type found on matrix tests.
(Jose Maria Casanova)
v4: (Jason Ekstrand)
- Put if case for 16-bits at the beginning of the if ladder.
- Use type_sz(dest.type) * 8 as bit_size parameter for scattered read.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
v2: Fix alignment style (Topi Pohjolainen)
(Jason Ekstrand)
- Enable bit_size parameter to scattered messages to enable different
bitsizes byte/word/dword.
- Remove use of brw_send_indirect_scattered_message in favor of
brw_send_indirect_surface_message.
- Move scattered messages to surface messages namespace.
- Assert align1 for scattered messages and assume Gen8+.
- Inline brw_set_dp_byte_scattered_read.
v3: (Jason Ekstrand)
- Use renamed brw_byte_scattered_data_element_from_bit_size method
- Assert scattered read for Gen8+ and Haswell.
- Use conditional expresion at components_read.
- Include comment about params for scattered opcodes.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
