Before rebasing on top of Ken's split-SEND optimization (see !17018),
this commit just caused some scheduling changes in various tessellation
and geometry shaders. These changes were caused by the addition of real
latency information for the URB messages.
With the addition of the split-SEND optimization, the changes
are... staggering. All of the shaders helped for spills and fills are
vertex shaders from Batman Arkham Origins. What surprises me is that
these shaders account for such a high percentage of the spills and fills
in fossil-db. 85%?!?
v2: Use FIXED_GRF instead of BRW_GENERAL_REGISTER_FILE in an assertion.
Suggested by Ken.
Tiger Lake, Ice Lake, and Skylake had similar results. (Ice Lake shown)
total instructions in shared programs: 20013625 -> 19954020 (-0.30%)
instructions in affected programs: 4007157 -> 3947552 (-1.49%)
helped: 31161
HURT: 0
helped stats (abs) min: 1 max: 400 x̄: 1.91 x̃: 2
helped stats (rel) min: 0.08% max: 59.70% x̄: 2.20% x̃: 1.83%
95% mean confidence interval for instructions value: -1.97 -1.86
95% mean confidence interval for instructions %-change: -2.22% -2.18%
Instructions are helped.
total cycles in shared programs: 859337569 -> 858636788 (-0.08%)
cycles in affected programs: 74168298 -> 73467517 (-0.94%)
helped: 13812
HURT: 16846
helped stats (abs) min: 1 max: 291078 x̄: 82.83 x̃: 4
helped stats (rel) min: <.01% max: 37.09% x̄: 3.47% x̃: 2.02%
HURT stats (abs) min: 1 max: 1543 x̄: 26.31 x̃: 14
HURT stats (rel) min: <.01% max: 77.97% x̄: 4.11% x̃: 2.58%
95% mean confidence interval for cycles value: -55.10 9.39
95% mean confidence interval for cycles %-change: 0.62% 0.77%
Inconclusive result (value mean confidence interval includes 0).
Broadwell
total cycles in shared programs: 904844939 -> 904832320 (<.01%)
cycles in affected programs: 525360 -> 512741 (-2.40%)
helped: 215
HURT: 4
helped stats (abs) min: 4 max: 1018 x̄: 60.16 x̃: 39
helped stats (rel) min: 0.14% max: 15.85% x̄: 2.16% x̃: 2.04%
HURT stats (abs) min: 79 max: 79 x̄: 79.00 x̃: 79
HURT stats (rel) min: 1.31% max: 1.57% x̄: 1.43% x̃: 1.43%
95% mean confidence interval for cycles value: -75.02 -40.22
95% mean confidence interval for cycles %-change: -2.37% -1.81%
Cycles are helped.
No shader-db changes on any older Intel platforms.
Tiger Lake, Ice Lake, and Skylake had similar results. (Ice Lake shown)
Instructions in all programs: 142622800 -> 141461114 (-0.8%)
Instructions helped: 197186
Cycles in all programs: 9101223846 -> 9099440025 (-0.0%)
Cycles helped: 37963
Cycles hurt: 151233
Spills in all programs: 98829 -> 13695 (-86.1%)
Spills helped: 2159
Fills in all programs: 128142 -> 18400 (-85.6%)
Fills helped: 2159
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17379>
The lowering is currently fake. It just changes the opcode from the
_LOGICAL version to the non-_LOGICAL version.
v2: Remove some rebase cruft. 's/gfx8_//;s/simd8_/' in
brw_instruction_name. Both suggested by Ken.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17379>
If these checks had been in place previously, some bugs
that... eh-hem... practically took down the Intel CI would have been
caught earlier. *blush*
v2: Update to account for split sends.
v3: Add some more Gfx version checks. Remove the redundant "src0 is a
GRF" check. Both suggested by Ken.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17379>
An argument could be made that all stage-specific opcodes for vec4
stages should be prefixed with VEC4_ like the stage-agnostic opcodes.
I'll leave those additional sed jobs for another day.
egrep -lr '(VS|GS|TCS)_OPCODE_URB_WRITE' src |\
while read f; do
sed --in-place 's/\(VS\|GS\|TCS\)_OPCODE_URB_WRITE/VEC4_\1_OPCODE_URB_WRITE/g' $f
done
egrep -lr 'T.S_OPCODE[_A-Z]*URB_OFFSETS' src |\
while read f; do
sed --in-place 's/\(T.S_OPCODE[_A-Z]*URB_OFFSETS\)/VEC4_\1/g' $f
done
Suggested-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17379>
With stages dispatching with a mask, we can run into situations where
we don't have the global address in all lanes. The existing code
always assumed we had the addres in at least lane0.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: bb40e999d1 ("intel/nir: use a single intel intrinsic to deal with ray traversal")
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17330>
EOT messages need to use g112-g127 for their sources. With the new
opt_split_sends pass, we may be constructing an EOT message from two
different registers, and be able to copy propagate the original values
into those SENDs.
This can cause problems if we copy propagate from a large register
(say an RGBA value which is 4 GRFs in SIMD8 or 8 GRFs in SIMD16), in a
situation where the SEND only read a subset of that (say the alpha value
out of an RGBA texturing result). g112-127 can only hold 16 registers
worth of data, and sometimes we can only use g112-126. So, we can't
propagate if the GRFs in question are larger than 15 GRFs.
Fixes a shader validation failure in Alan Wake. Thanks to Ian Romanick
for catching this!
shader-db on Icelake shows that only SIMD32 programs are affected, and
the results are pretty negligable:
total instructions in shared programs: 19615228 -> 19615269 (<.01%)
instructions in affected programs: 10702 -> 10743 (0.38%)
helped: 1 / HURT: 43 / largest change: +/- 2 instructions
total cycles in shared programs: 852001706 -> 852001566 (<.01%)
cycles in affected programs: 767098 -> 766958 (-0.02%)
helped: 68 / HURT: 64 / largest change: +/- 774 cycles
GAINED: 2 / LOST: 0
Fixes: 589b03d02f ("intel/fs: Opportunistically split SEND message payloads")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6803
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17390>
Not had time to investiguate what is going is on but it's definitely a
contributor to failures.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16104>
The thread dispatch SEND instructions will dispatch new threads
immediately even before the caller of the SEND instruction has reached
EOT. So we really need to make sure all the memory writes are visible
to other threads within the DSS before the SEND.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15755>
While we've taken advantage of split-sends in select situations, there
are many other cases (such as sampler messages, framebuffer writes, and
URB writes) that have never received that treatment, and continued to
use monolithic send payloads.
This commit introduces a new optimization pass which detects SEND
messages with a single payload, finds an adjacent LOAD_PAYLOAD that
produces that payload, splits it two, and updates the SEND to use both
of the new smaller payloads.
In places where we manually used split SENDS, we rely on underlying
knowledge of the message to determine a natural split point. For
example, header and data, or address and value.
In this pass, we instead infer a natural split point by looking at the
source registers. Often times, consecutive LOAD_PAYLOAD sources may
already be grouped together in a contiguous block, such as a texture
coordinate. Then, there is another bit of data, such as a LOD, that
may come from elsewhere. We look for the point where the source list
switches VGRFs, and split it there. (If there is a message header, we
choose to split there, as it will naturally come from elsewhere.)
This not only reduces the payload sizes, alleviating register pressure,
but it means that we may be able to eliminate some payload construction
altogether, if we have a contiguous block already and some extra data
being tacked on to one side or the other.
shader-db results for Icelake are:
total instructions in shared programs: 19602513 -> 19369255 (-1.19%)
instructions in affected programs: 6085404 -> 5852146 (-3.83%)
helped: 23650 / HURT: 15
helped stats (abs) min: 1 max: 1344 x̄: 9.87 x̃: 3
helped stats (rel) min: 0.03% max: 35.71% x̄: 3.78% x̃: 2.15%
HURT stats (abs) min: 1 max: 44 x̄: 7.20 x̃: 2
HURT stats (rel) min: 1.04% max: 20.00% x̄: 4.13% x̃: 2.00%
95% mean confidence interval for instructions value: -10.16 -9.55
95% mean confidence interval for instructions %-change: -3.84% -3.72%
Instructions are helped.
total cycles in shared programs: 848180368 -> 842208063 (-0.70%)
cycles in affected programs: 599931746 -> 593959441 (-1.00%)
helped: 22114 / HURT: 13053
helped stats (abs) min: 1 max: 482486 x̄: 580.94 x̃: 22
helped stats (rel) min: <.01% max: 78.92% x̄: 4.76% x̃: 0.75%
HURT stats (abs) min: 1 max: 94022 x̄: 526.67 x̃: 22
HURT stats (rel) min: <.01% max: 188.99% x̄: 4.52% x̃: 0.61%
95% mean confidence interval for cycles value: -222.87 -116.79
95% mean confidence interval for cycles %-change: -1.44% -1.20%
Cycles are helped.
total spills in shared programs: 8387 -> 6569 (-21.68%)
spills in affected programs: 5110 -> 3292 (-35.58%)
helped: 359 / HURT: 3
total fills in shared programs: 11833 -> 8218 (-30.55%)
fills in affected programs: 8635 -> 5020 (-41.86%)
helped: 358 / HURT: 3
LOST: 1 SIMD16 shader, 659 SIMD32 shaders
GAINED: 65 SIMD16 shaders, 959 SIMD32 shaders
Total CPU time (seconds): 1505.48 -> 1474.08 (-2.09%)
Examining these results: the few shaders where spills/fills increased
were already spilling significantly, and were only slightly hurt. The
applications affected were also helped in countless other shaders, and
other shaders stopped spilling altogether or had 50% reductions. Many
SIMD16 shaders were gained, and overall we gain more SIMD32, though many
close to the register pressure line go back and forth.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17018>
SEND messages with EOT need to use g112-g127 for their sources so that
the hardware is able to launch new threads while old ones are finishing
without worrying about register overlap when pushing payloads. For the
newer split-send messages, this applies to both source registers.
Our special case for this in the register allocator was only considering
the first source. This wasn't a problem because we hadn't ever tried to
use split-sends with EOT before. However, my new optimization pass is
going to introduce some shortly, so we'll need to handle them properly.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17018>
Now that we've removed the thread_local lookup tables using
pointer-to-member C++ features, this can go back to being a standard
C file, like it was in the past. We just need to annotate a couple
of things with "struct".
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
We had been using thread_local index -> opcode_desc tables to avoid
plumbing through a storage location throughout all the code. But now
we have done so with the new brw_isa_info structure. So we can just
store the tables there, and initialize it with the compiler.
This fixes crashes in gtk4-demo on iris, and should help with some
programs on zink as well. Something was going wrong with the
thread_local variables not being set up correctly. While we might be
able to work around that issue, there's really no advantage to storing
these lookup tables in TLS (beyond it being simpler to do originally).
So let's simply stop doing so.
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6728
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6229
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
This structure will contain the opcode mapping tables in the next
commit. For now, this is the mechanical change to plumb it into all
the necessary places, and it continues simply holding devinfo.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
This patch creates a new header file, brw_isa_info.h, which will
contains all the functions related to opcode encoding on various
generations. Opcode numbers may have different meanings on different
hardware, so we remap them between an enum we can easily work with
and the hardware encoding.
We move the brw_inst setters and getters to brw_inst.h.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
This header file didn't include normal guards against being included
multiple times. It also defined a function in a header file without
marking it static inline.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
src/mesa/main includes are for Mesa's OpenGL implementation, and the
compiler is used in Vulkan drivers and other tools. We really only
needed one #define, which is that we offer 32 samplers. It probably
makes more sense to have our own defined limit for that rather than
importing a project-wide value which theoretically could be adjusted,
so swap MAX_SAMPLERS for a new BRW_MAX_SAMPLERS and call it a day.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17309>
Otherwise passes which expect offsets to be in bytes (like
brw_nir_lower_mem_access_bit_sizes, called from brw_postprocess_nir)
may produce incorrect results.
Fixes 64-bit load/stores in task/mesh shaders.
Fixes: c36ae42e4c ("intel/compiler: Use nir_var_mem_task_payload")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16196>
v2: Do the synchronization in the correct place. Noticed by Curro.
Fixes: b5fa43952a ("intel/fs: Better handle constant sources of FS_OPCODE_PACK_HALF_2x16_SPLIT")
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Tested-by: Felix DeGrood <felix.j.degrood@intel.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17037>
I don't know when this was added but it's really neat and we should use
it instead of NIR_PASS_V since NIR_DEBUG=print and a few validation
things will work better.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17014>
This fixes many tests in following groups on DG2:
dEQP-VK.memory_model.*
dEQP-VK.fragment_shader_interlock.*
v2: use memory scope and setup descriptor also
for barriers without defined scope (Curro),
use local scope and flush type none with
NIR_SCOPE_NONE scope, cleanups (Lionel)
v3: use LSC_FENCE_THREADGROUP for NIR_SCOPE_WORKGROUP,
remove default case (Curro), use eviction if scope
was not defined, use LSC_FENCE_GPU scope for vertex
stage
v4: use LSC_FENCE_TILE independent of stage for device
scope (Curro)
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15743>
fs_visitor::assign_curb_setup() maps UNIFORM registers to HW regs,
and contains the following assert:
assert(inst->src[i].stride == 0);
emit_a64_oword_block_header's striding tricks run afoul of this
restriction, by producing stride 1 values on a 64-bit UNIFORM source.
Work around this by copying the UNIFORM value to a VGRF first.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16938>
NIR has fdiv, and all the NIR backends have to have lower_fdiv set
appropriately already since various passes (format conversions,
tgsi_to_nir, nir_fast_normalize(), etc.) might generate one.
This causes softpipe and llvmpipe to now do actual divides, since
lower_fdiv is not set there. Note that llvmpipe's rcp implementation is a
divide of 1.0 by x, so now we're going to be just doing div(x, y) instead
of mul(x, div(1.0, y)).
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16823>
This will allow the use of static_assert here instead of our
compiler-specific implementation.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16670>
Call it once instead of calling the very same function for each source
and destination. This should make those ternary operators a little
easier to read, IMHO.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15835>
In opt_algebraic(), handle TYPE_DF in a different check than TYPE_Q. We have a
separate flag for each type, use separate checks so platforms where one is true
and the other is not can work properly.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15835>
Don't compute it based on devinfo->has_64bit_float. Othwerwise we may
end up emitting 64bit-int (Q) instructions on platforms with 64bit
floats but not 64bit integers.
Right now, the only platforms where has_64bit_int is different from
has_64bit_float are the platforms that use GFX7_FEATURES.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15835>
This expression accidentally performs a 32-bit sign-extension when
processing the second half of the expression (the low 16 bits).
Consider -7W, which is represented as 0xfff9fff9 in our encoding (the
16-bit word is replicated to both halves of the 32-bit dword).
Tigerlake's compaction stores the low 11-bits of an immediate as-is,
and replicates the 12th bit. So here, compacted_imm will be 0xff9.
( (int)(0xff9 << 20) >> 4) |
((short)(0xff9 << 4) >> 4))
0xfff90000 | (0xff90 >> 4)
0xfff90000 | 0xfffffff9 ...oops...
0xfffffff9
By casting the second line of the expression to unsigned short, we
prevent the sign-extension when it combines both parts, so we get:
0xfff90000 | 0x0000fff9
0xfff9fff9
Fixes: 12d3b11908 ("intel/compiler: Add instruction compaction support on Gen12")
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16833>
For cases with lots of very small primitives, this may improve
performance because we're not executing those dead channels all the
time.
Shader-db reports no instruction or cycle-count changes. However, by
hacking up the driver to report when this optimization triggers, it
appears to affect about 10% of shader-db.
v2 (Kenneth Graunke): Always enable VMask prior to XeHP for now,
because using VMask on those platforms allows us to perform the
eliminate_find_live_channel() optimization. However, XeHP doesn't
seem to have packed fragment shader dispatch, so we lose that
optimization regardless, and there's no reason not to avoid vmask.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/1054>
Originally, we had virtual opcodes for scratch access, and let the
generator count spills/fills separately from other sends. Later, we
started using the generic SHADER_OPCODE_SEND for spills/fills on some
generations of hardware, and simply detected stateless messages there.
But then we started using stateless messages for other things:
- anv uses stateless messages for the buffer device address feature.
- nir_opt_large_constants generates stateless messages.
- XeHP curbe setup can generate stateless messages.
So counting stateless messages is not accurate. Instead, we move the
spill/fill accounting to the register allocator, as it generates such
things, as well as the load/store_scratch intrinsic handling, as those
are basically spill/fills, just at a higher level.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16691>
If we saw a HALT instruction, we would forget to invalidate our analysis
pass information before returning progress.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16677>
This is set to true for all drivers that have a GLSL level
of support lower than 4.00. This matches the rule for setting the
GLSL IR option EmitNoIndirectSampler.
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16543>
