Using HALT to immediately jump to the end of the shader is required to
implement GL_EXT_gpu_shader4 and OpenGL 3.0. However, vanilla OpenGL
1.2 doesn't forbid it and it likely makes something somewhere faster.
We should be consistent and implement the same discard behavior on all
hardware if we can.
The rules for HALT on Gen4-5 are a bit different from Gen6+. On the
older hardware, there is no stack for HALT; instead it's up to software
to save and restore mask registers. However, there's no real saving
needed since we only use HALT to jump to the end of the program where
we're about about to do our FB writes. All we need to do is reset AMask
to DMask, the value it was initialized to at the start of the thread.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5244>
Instead of emitting the stall MOV "inside" the
SHADER_OPCODE_MEMORY_FENCE generation, use the scheduling fences when
creating the IR.
For IvyBridge, every (data cache) fence is accompained by a render
cache fence, that now is explicit in the IR, two
SHADER_OPCODE_MEMORY_FENCEs are emitted (with different SFIDs).
Because Begin and End interlock intrinsics are effectively memory
barriers, move its handling alongside the other memory barrier
intrinsics. The SHADER_OPCODE_INTERLOCK is still used to distinguish
if we are going to use a SENDC (for Begin) or regular SEND (for End).
This change is a preparation to allow emitting both SENDs in Gen11+
before we can stall on them.
Shader-db results for IVB (i965):
total instructions in shared programs: 11971190 -> 11971200 (<.01%)
instructions in affected programs: 11482 -> 11492 (0.09%)
helped: 0
HURT: 8
HURT stats (abs) min: 1 max: 3 x̄: 1.25 x̃: 1
HURT stats (rel) min: 0.03% max: 0.50% x̄: 0.14% x̃: 0.10%
95% mean confidence interval for instructions value: 0.66 1.84
95% mean confidence interval for instructions %-change: 0.01% 0.27%
Instructions are HURT.
Unlike the previous code, that used the `mov g1 g2` trick to force
both `g1` and `g2` to stall, the scheduling fence will generate `mov
null g1` and `mov null g2`. During review it was decided it was not
worth keeping the special codepath for the small effect will have.
Shader-db results for HSW (i965), BDW and SKL don't have a change
on instruction count, but do report changes in cycles count, showing
SKL results below
total cycles in shared programs: 341738444 -> 341710570 (<.01%)
cycles in affected programs: 7240002 -> 7212128 (-0.38%)
helped: 46
HURT: 5
helped stats (abs) min: 14 max: 1940 x̄: 676.22 x̃: 154
helped stats (rel) min: <.01% max: 2.62% x̄: 1.28% x̃: 0.95%
HURT stats (abs) min: 2 max: 1768 x̄: 646.40 x̃: 362
HURT stats (rel) min: <.01% max: 0.83% x̄: 0.28% x̃: 0.08%
95% mean confidence interval for cycles value: -777.71 -315.38
95% mean confidence interval for cycles %-change: -1.42% -0.83%
Cycles are helped.
This seems to be the effect of allocating two registers separatedly
instead of a single one with size 2, which causes different register
allocation, affecting the cycle estimates.
while ICL also has not change on instruction count but report changes
negative changes in cycles
total cycles in shared programs: 352665369 -> 352707484 (0.01%)
cycles in affected programs: 9608288 -> 9650403 (0.44%)
helped: 4
HURT: 104
helped stats (abs) min: 24 max: 128 x̄: 88.50 x̃: 101
helped stats (rel) min: <.01% max: 0.85% x̄: 0.46% x̃: 0.49%
HURT stats (abs) min: 2 max: 2016 x̄: 408.36 x̃: 48
HURT stats (rel) min: <.01% max: 3.31% x̄: 0.88% x̃: 0.45%
95% mean confidence interval for cycles value: 256.67 523.24
95% mean confidence interval for cycles %-change: 0.63% 1.03%
Cycles are HURT.
AFAICT this is the result of the case above.
Shader-db results for TGL have similar cycles result as ICL, but also
affect instructions
total instructions in shared programs: 17690586 -> 17690597 (<.01%)
instructions in affected programs: 64617 -> 64628 (0.02%)
helped: 55
HURT: 32
helped stats (abs) min: 1 max: 16 x̄: 4.13 x̃: 3
helped stats (rel) min: 0.05% max: 2.78% x̄: 0.86% x̃: 0.74%
HURT stats (abs) min: 1 max: 65 x̄: 7.44 x̃: 2
HURT stats (rel) min: 0.05% max: 4.58% x̄: 1.13% x̃: 0.69%
95% mean confidence interval for instructions value: -2.03 2.28
95% mean confidence interval for instructions %-change: -0.41% 0.15%
Inconclusive result (value mean confidence interval includes 0).
Now that more is done in the IR, more dependencies are visible and
more SWSB annotations are emitted. Mixed with different register
allocation decisions like above, some shaders will see more `sync
nops` while others able to avoid them.
Most of the new `sync nops` are also redundant and could be dropped,
which will be fixed in a separate change.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3278>
So we can eventually remove the cycle count estimates from the CFG
data structure and consolidate performance information in the
brw::performance object.
It would be cleaner to pass the brw::performance object directly to
the disassembler but that isn't straightforward since the disassembler
is built as a plain C file unlike the rest of the compiler back-end.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
These should be more accurate than the current cycle counts, since
among other things they consider the effect of post-scheduling passes
like the software scoreboard on TGL. In addition it will enable us to
clean up some of the now redundant cycle-count estimation
functionality in the instruction scheduler.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Change brw_memory_fence to return the number of messages emitted, and
use that to update the send_count statistic in code generation.
This will fix the book-keeping for IVB since the memory fences will
result in two SEND messages.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4646>
The current workaround for this hardware bug involved marking the ADD
instruction used to initialize the address register as NoMask on
Gen12, which was based on the assumption that the problem was caused
by a hardware bug affecting the application of the execution mask to
the address register write.
However that doesn't seem to be the case: The address register write
was working correctly, the real problem leading to hangs on TGL is
that the indirect addressing logic is unable to deal with garbage
values in the address register (e.g. misaligned offsets), even for
channels which are currently inactive due to non-uniform control flow.
The current workaround isn't able to avoid that situation in general,
since the result of the NoMask ADD instruction for a dead channel is
calculated based on the corresponding (dead) component of the
indirect_byte_offset source, which would still be undefined in the
likely case that the source was initialized under control flow itself.
This would lead to hangs whenever MOV_INDIRECT was used under
non-uniform control flow in some scenarios like a tessellation shader
from GFXBench5/gl_4 (AKA Car Chase) on TGL. In addition I've managed
to reproduce the same issue on earlier platforms by initializing the
whole address register with garbage before the ADD instruction, so
this seems to be a long-standing issue we have avoided mostly by luck.
This patch fixes the problem and applies the workaround to all
platforms, since even when the hardware is able to deal with garbage
address values without hanging there might be a significant
performance cost from reading random GRF registers due to the useless
extra EU cycles spent fetching registers for dead channels and due to
the potential for unintended serialization with respect to other
random instructions that could be executed in parallel, which may have
had a cost of the order of hundreds of cycles in the worst case
scenario.
Fixes: f93dfb509c "intel/fs: Write the address register with NoMask for MOV_INDIRECT"
Tested-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Passing shader_stats to the fs_generator constructor means that the
SIMD8 shader stats from the visitor (such as the scheduler mode) will be
reported out for the SIMD16/SIMD32 versions as well.
As you can see, we are now passing 'shader_stats' and 'stats' to
generate_code(), which is obviously odd looking. Ian rebased and
committed an old patch of mine which added the shader_stats struct on
July 30 in commit dabb5d4bee (i965/fs: Add a shader_stats struct.) and
shortly after on August 12 Jason added the brw_compile_stats struct in
commit 134607760a (intel/compiler: Fill a compiler statistics struct).
I'd like to combine the two, but I'm not sure how. shader_stats is an
input to generate_code() while brw_compile_stats is an output and is
only used by the Vulkan driver. Leave it as is for now...
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4093>
Found by inspection. Existing code was trying to avoid assuming that
an SBID had been assigned to the virtual instruction, but
synchronizing the header setup with respect to the previous SIMD16
SEND by using SYNC.ALLRD doesn't really seem possible unless the SEND
instruction had been assigned an SBID. Assert-fail instead if no SBID
has been allocated.
Fixes: 15e3a0d9d2 "intel/eu/gen12: Set SWSB annotations in hand-crafted assembly."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Cc: 20.0 <mesa-stable@lists.freedesktop.org>
v2: (Francisco Jerez)
- Drop vec4 changes.
- Handle explicit acc0 operand and implicit one.
- Make sure instruction is SIMD16, prediction is off and default mask
control set to true.
v3: (Francisco Jerez)
- Clear accumulator only when it's written.
- Use BRW_MASK_DISABLE instead of true.
- Use correct width for brw_acc_reg().
- Fix last_inst_offset.
v4: (Francisco Jerez)
- Don't check for last instruction for accummulator write.
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3376>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3376>
Like a SHADER_OPCODE_MEMORY_FENCE but doesn't doesn't generate any
assembly code.
Will be used when the compiler shouldn't reorder certain instructions
but there's no need to generate code for the HW to do it -- as the
ordering will be guaranteed by other means.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3226>
This commit fills in a number of different pieces:
1. We add support to brw_nir_lower_mem_access_bit_sizes to handle the
new intrinsics. This involves simple plumbing work as well as a
tiny bit of extra logic to always scalarize scratch intrinsics
2. Add code to brw_fs_nir.cpp to turn nir_load/store_scratch intrinsics
into byte/dword scattered read/write messages which use the A32
stateless model.
3. Add code to lower_surface_logical_send to handle dword scattered
messages and the A32 stateless model.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
This can be useful to measure whether memory access optimizations are
having the desired effect. For example, we might see a reduction in
image loads/stores, or constant buffer loads. We can already see this
in cycle estimates to some extent, but this is a more direct approach,
minus a lot of the noise of random scheduler shuffling.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Apparently the ts_request_type and ts_resource_select thread spawner
message descriptor bits were removed from the hardware at least since
ICL. Drop them in order to avoid assertion failures on Gen12+
platforms which don't have any encoding for this. On Gen9+ these are
probably just ignored by the hardware, so this is unlikely to have had
any functional implications prior to Gen12.
v2: Mark TS message fields as non-existing in brw_inst.h on ICL. (Caio)
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
The WAIT instruction has been removed, but SYNC.bar can be used
instead to wait for a notification on n0.0.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewers are encouraged to audit the code generation pass
independently for the case I missed some potential data hazard or new
code has been added in the meantime.
v2: Add SYNC instruction to cr0 workaround in brw_float_controls_mode().
v3: Drop likely redundant (and potentially harmful) RegDist SWSB
annotation from ce0 read in brw_find_live_channel() (Caio).
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
A future lowering pass will simulate the same behavior originally
provided by NoDDChk/NoDDClr at the IR level by using appropriate SWSB
annotations.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The new SEND instruction behaves like the former SENDS instruction.
The original single-payload SEND instruction is gone.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The SEND instruction is now four-source. The descriptor is no longer
part of source 1, so avoid touching it to avoid corruption while
initializing the descriptor.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We can't really handle it in the little-core 64-bit case but it's not
really needed there. Where we really want this is for when we need to
do 16 -> 8-bit conversions.
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
During generate_shuffle(), when we use byte sized registers we end up
with a destination stride of 2. We don't take the stride into
consideration when selecting the group offset for the last MOV
operation, which means we end up moving things to the wrong place,
leaving the last few channels untouched. Take the destination stride
in consideration so we don't miss the last channels.
v2: Assert this is not necessary for the IVB special case (Jason).
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
The current code can create functions with a width of 32, which is not
supported by our hardware. Add some code to simplify how we express
what we want and prevent such cases.
For some unknown reason, all the tests I could run seem to work even
with these unsupported MOVs.
Fixes: b0858c1cc6 "intel/fs: Add a couple of simple helper opcodes"
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
There are cases where we try to generate registers with a stride of
32, while the hardware maximum is just 16. This happens, for example,
when using 8 bit integers on SIMD32. This results in a crash because
the variable 'width' has a value of 32:
../../src/intel/compiler/brw_reg.h:550: brw_reg brw_vecn_reg(unsigned
int, brw_reg_file, unsigned int, unsigned int): Assertion `!"Invalid
register width"' failed.
This change prevents the crash and makes the tests pass.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Before this commit, we had only FPRoundingMode decoration (the per
instruction one) that is applied during the SPIR-V handling. In
vtn_alu we find out the rounding mode, and generate the code
accordingly that later will be used to look for the respective
nir_op_f2f16_{rtz,rtne}.
Per-instruction gets prioritized because we make them explicit
conversions (with RTZ or RTNE nir opcodes) and they will override the
default execution mode defined with float controls. However, we need
to come back to the mode defined by float controls after the execution
of the FP Rounding instruction.
Therefore, the new SHADER_OPCODE_FLOAT_CONTROL_MODE opcode will be
used to set the default rounding mode and denorms treatment in the
whole shader while the pre-existent SHADER_OPCODE_RND_MODE, will be
used as prioritized rounding mode in a per-instruction basis.
v2:
- Fix bug in defining BRW_CR0_FP_MODE_MASK.
v3:
- Update comment (Caio).
v4:
- Split the patch into the helper and the new opcode (this
one) (Caio).
v5:
- Add an explanation on the actual purpose and priority of the newly
introduced opcode in the commit log (Caio).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
v2:
- Fix bug in defining BRW_CR0_FP_MODE_MASK.
v3:
- Update comment (Caio).
v4:
- Split the patch into the helper (this one) and the new
opcode (Caio).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
This reverts commit c0504569ea. Now that
we're doing interpolation lowering in NIR, we can continue to stride the
FS input registers directly in the brw_fs_nir code like we did before.
This fixes SIMD32 fragment shaders which broke because lower_simd_width
depended on the 0 stride to split PLN instructions correctly.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This commit is all annoying plumbing work which just adds support for a
new brw_compile_stats struct. This struct provides a binary driver
readable form of the same statistics we dump out to stderr when we
INTEL_DEBUG is set with a shader stage.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
When dumping shader's assembly with INTEL_DEBUG=vs,tcs,...
sha1 of the resulting assembly is also printed, having environment
variable INTEL_SHADER_ASM_READ_PATH present driver will try to
load a "%sha1%.bin" file from the path and substitute current
assembly with the one from the file.
Signed-off-by: Danylo Piliaiev <danylo.piliaiev@globallogic.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The issue here was discovered by a set of Vulkan CTS tests:
dEQP-VK.glsl.derivate.*.dynamic_*
These tests use ballot ops to construct a branch condition that takes
the same path for each 2x2 quad but may not be uniform across the whole
subgroup. They then tests that derivatives work and give the correct
value even when executed inside such a branch. Because the derivative
isn't executed in uniform control-flow and the values coming into the
derivative aren't smooth (or worse, linear), they nicely catch bugs that
aren't uncovered by simpler derivative tests.
Unfortunately, these tests require Vulkan and the equivalent GL test
would require the GL_ARB_shader_ballot extension which requires int64.
Because the requirements for these tests are so high, it's not easy to
test on older hardware and the bug is only proven to exist on gen7;
gen4-6 are a conjecture.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Matt Turner <mattst88@gmail.com>
Line wrap some awfully long lines while we are here.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
It'll grow further, and we'd like to avoid adding an additional
parameter to fs_generator() for each new piece of data.
v2 (idr): Rebase on 17 months. Track a visitor instead of a cfg.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Gen11 SLM is not on L3 anymore, so now the hardware has two separate
fences. Add a way to control which fence types to use.
At this time, we don't have enough information in NIR to control the
visibility of the memory being fenced, so for now be conservative and
assume that fences will need a stall. With more information later
we'll be able to reduce those.
Fixes Vulkan CTS tests in ICL:
dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.device.payload_nonlocal.workgroup.guard_local.buffer.comp
dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.device.payload_local.buffer.guard_nonlocal.workgroup.comp
dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.device.payload_local.image.guard_nonlocal.workgroup.comp
dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.workgroup.payload_local.buffer.guard_nonlocal.workgroup.comp
dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.workgroup.payload_local.image.guard_nonlocal.workgroup.comp
The whole set of supported tests in dEQP-VK.memory_model.* group
should be passing in ICL now.
v2: Pass BTI around instead of having an enum. (Jason)
Emit two SHADER_OPCODE_MEMORY_FENCE instead of one that gets
transformed into two. (Jason)
List tests fixed. (Lionel)
v3: For clarity, split the decision of which fences to emit from the
emission code. (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
v2: 1) Drop changes for vec4 backend as on Gen11+ we don't support
align16 mode (Matt Turner)
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This rewrites the ddy in EXECUTE_4 mode with a loop to make it more
obvious what is going on and also sets the group each of the 4 threads
in the groups are supposed to execute.
Fixes the following CTS tests :
dEQP-VK.glsl.derivate.dfdyfine.dynamic_*
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Co-Authored-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Fixes: 2134ea3800 ("intel/compiler/fs: Implement ddy without using align16 for Gen11+")
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>
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>
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>
We now have a lowering pass that will do this at the fs_visitor level,
so we can remove this code from gen11+.
v2: Reduce size of the "i" array from 4 to 2 (Matt).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Move the scalar-region conversion from the IR to the generator, so it
doesn't affect the Gen11 path. We need the non-scalar regioning
for a later lowering pass that we are adding.
v2: Better commit message (Matt)
Reviewed-by: Matt Turner <mattst88@gmail.com>
