This hardware bug is the result of a control flow optimization present
in Gfx8-9 meant to prevent the ELSE instruction from disabling all
channels and update the control flow stack only to have them
re-enabled at the ENDIF instruction executed immediately after it.
Instead, on Gfx8-9 an ELSE instruction that would normally have ended
up with all channels disabled would pop off the last element of the
stack and jump directly to JIP+1 instead of to the ENDIF at JIP,
skipping over the ENDIF instruction. In simple cases this would work
okay (though it's actual performance benefit is questionable), but in
cases where a branch instruction within the IF block (e.g. BREAK or
CONTINUE) caused all active channels to jump outside the IF
conditional, the optimization would break the JIP chain of "join"
instructions by skipping the ENDIF, causing the block of instructions
immediately after the ENDIF to execute with all channels disabled
until execution reaches the reconvergence point.
This issue was observed on SKL in the
dEQP-VK.reconvergence.subgroup_uniform_control_flow_elect.compute.nesting4.0.38
test in combination with some Vulkan binding model changes Lionel is
working on. In such cases the execution with all channels disabled
was leading to corruption of an indirect message descriptor, causing a
hang.
Unfortunately the hardware bug doesn't provide a recommended
workaround. In order to fix the problem we point the JIP of an ELSE
instruction to the instruction immediately before the ENDIF -- However
that's not expected to work due to the restriction that JIP and UIP
must be equal if and only if BranchCtrl is disabled -- So this patch
also enables BranchCtrl, which is intended to support join
instructions within the "ELSE" block, which in turn disables the
optimization described above, which in turn causes us to execute the
instruction immediately *before* the ENDIF with all channels disabled
-- So in order to avoid further fallout from executing code with all
channels disabled we need to insert a NOP before ENDIF instructions
that have a matching ELSE instruction.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20921>
This allows us to communicate to the back-end that we don't actually
know if the framebuffer is multisampled or not. No drivers set anything
but ALWAYS/NEVER and we still have a few ALWAYS/NEVER assumptions but
those should be asserted.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21094>
This allows for the possibility that we may not know at compile time if
sample shading is enabled through the API. While we're here, also
document exactly what this bit means so we don't confuse ourselves.
v2: Fixup coarse pixel values (Lionel)
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21094>
Whenever one of them is BRW_SOMETIMES, we depend on dynamic flag pushed
in as a push constant. In this case, we have to often have to do the
calculation both ways and SEL the result. It's a bit more code but
decouples MSAA from the shader key.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21094>
Builds on the work of !15121. This gets to delete even more code
because many drivers shared a lot of code for i2b and f2b.
No shader-db or fossil-db changes on any Intel platform.
v2: Rebase on 1a35acd8d9.
v3: Update a comment in nir_opcodes_c.py. Suggested by Konstantin.
v4: Another rebase. Remove f2b stuff from Midgard.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20509>
Allowing longer writes reduces the number of send messages needed
to support unaligned 4-component writes.
Note: nothing currently generates 8-component writes, so this change
makes "second_mask" code path in emit_urb_direct_writes and
emit_urb_indirect_writes_mod dead.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20858>
Our hardware requires that we write to URB using full vec4s at aligned
addresses. It gives us an ability to mask-off dwords within vec4 we don't
want to write, but we have to know their positions at compile time.
Let's assume that:
- V represents one dword we want to write
- ? is an unitinitialized value
- "|" is a vec4 boundary.
When we want to write 2-dword value at offset 0 we generate 1 write message:
| V1 V2 ? ? |
with mask:
| 1 1 0 0 |
When we want to write 4-dword value at offset 2 we generate 2 write messages:
| ? ? V1 V2 | V3 V4 ? ? |
with mask:
| 0 0 1 1 | 1 1 0 0 |
However if we don't know the offset within vec4 at *compile time* we
currently generate 4 write messages:
| V1 V1 V1 V1 |
| 0 0 1 0 |
| V2 V2 V2 V2 |
| 0 0 0 1 |
| V3 V3 V3 V3 |
| 1 0 0 0 |
| V4 V4 V4 V4 |
| 0 1 0 0 |
where masks are determined at *run time*.
This is quite wasteful and slow.
However, if we could determine the offset modulo 4 statically at compile time,
we could generate only 1 or 2 write messages (1 if modulo is 0) instead of 4.
This is what this patch does: it analyzes the addressing expression for
modulo 4 value and if it can determine it at compile time, we generate
1 or 2 writes, and if it can't we fallback to the old 4 writes method.
In mesh shader, the value of offset modulo 4 should be known for all outputs,
with an exception of primitive indices.
The modulo value should be known because of MUE layout restrictions, which
require that user per-primitive and per-vertex data start at address aligned
to 8 dwords and we should statically always know the offset from this base.
There can be some cases where the offset from the base is more dynamic
(e.g. indirect array access inside a per-vertex value), so we always do
the analysis.
Primitive indices are an exception, because they form vec3s (for triangles),
which means that the offset will not be easy to analyse.
When U888X index format lands, primitive indices will use only one dword
per triangle, which means that we'll always write them using one message.
Task shaders don't have any predetermined structure of output memory, so
always do the analysis.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20050>
We can lower FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD into other more
generic sends and drop this internal opcode.
The idea behind this change is to allow bindless surfaces to be used
for UBO pulls and why it's interesting to be able to reuse
setup_surface_descriptors(). But that will come in a later change.
No shader-db changes on TGL & DG2.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20416>
The ACP entries created by copy propagation to track the implied
copies of LOAD_PAYLOAD instructions don't model the behavior of
LOAD_PAYLOAD correctly, since (as of 41868bb682) header
moves are implicitly retyped to UD and the destination of non-header
copies implicitly uses the same type as the corresponding source, even
though the ACP entries created for such copies could incorrectly
represent a type conversion, which can lead to mis-optimization of the
program.
According to Marcin, this fixes the func.mesh.ext.workgroup_id.task.q0
crucible test.
Fixes: 41868bb682 ("i965/fs: Rework the fs_visitor LOAD_PAYLOAD instruction")
Reported-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Tested-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18980>
I stumbled on this when I inserted some suboptimal lowering code after all
optimizations. Adding certain subset of optimizations after my lowering code
actually avoided this bug, so I think it's not possible to hit this on upstream.
Let's fix this for the next person generating suboptimal code...
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20857>
I've been running into failures with tests like :
dEQP-VK.robustness.robustness2.bind.notemplate.rgba32i.unroll.nonvolatile.uniform_buffer_dynamic.no_fmt_qual.len_4.samples_1.1d.frag
With the load_global_const_block_intel NIR intrinsic, you can load a
vec8/vec16 with a predicate. The predicate is correctly uniformized to
feed into the SEND instruction's flag register.
The problem is that a series of optimization first remove the
find_live_channel and then changes the broadcast into a simple MOV
instruction, on the assumption that the first channel is always active
if there is not control flow. This is correct.
But after that the cmod optimzation will remove this instruction :
mov.nz.f0.0(16) null:D, vgrf16+0.0<0>:D NoMask
because it seems to be equivalent to :
cmp.g.f0.0(16) vgrf16:D, vgrf12:D, 63d
In this case vgrf16 is the predicate to the load block SEND
instruction. Since the execution mask is different between both, some
of the channels of the SEND instruction end up not being loaded or
loaded with the wrong predication and we end up with incorrect UBO
data.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20852>
These are handled identically in almost all cases. There is one place
in the legacy surface lowering that was obtaining the bitsize from the
opcode, but the LSC-based lowering uses (type_sz(inst->dst.type) * 8)
for that and works just fine. If we just do that in the legacy lowering
too, then we don't need this plethora of opcodes.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20604>
These are basically identical save for:
- shared has surface hardcoded to SLM rather than an SSBO index
- shared has to handle adding the 'base' const_index (SSBO have none)
- the NIR source index for data is shifted by one
It's not worth copy and pasting the entire function for this.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20604>
These are now basically identical to their non-float counterparts. The
only thing that differed was the opcode checking to determine which
operands existed. Now that we have a unified opcode enum and a helper
for the number of data operands, we can just use that.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20604>
The only reason for the separate opcode was because of the overlapping
BRW_AOP_* enums, making it impossible to tell whether a particular AOP
was the integer or float operation. Now that we use the lsc_opcode
enums, we can just have the legacy lowering inspect the opcode and
select the right descriptor. No need for a separate opcode.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20604>
This gets our logical atomic messages using the lsc_opcode enum rather
than the legacy BRW_AOP_* defines. We have to translate one way or
another, and using the modern set makes sense going forward.
One advantage is that the lsc_opcode encoding has opcodes for both
integer and floating point atomics in the same enum, whereas the legacy
encoding used overlapping values (BRW_AOP_AND == 1 == BRW_AOP_FMAX),
which made it impossible to handle both sensibly in common code.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20604>
This avoids a violation of the Vulkan memory model that was leading to
intermittent failures of at least 8k test-cases of the Vulkan CTS
(within the group dEQP-VK.memory_model.*) on TGL and DG2 platforms.
In theory the issue may be reproducible on earlier platforms like IVB
and ICL, but the SYNC.ALLWR instruction is not available on those
platforms so a different (likely costlier) fix will be needed.
The issue occurs within the sequence we emit for a NIR memory barrier
with acquire semantics requiring the synchronization of multiple
caches, e.g. in pseudocode for a barrier involving the TGM and UGM
caches on DG2:
x <- load.ugm // Atomic read sequenced-before the barrier
y <- fence.ugm
z <- fence.tgm
wait(y, z)
w <- load.tgm // Read sequenced-after the barrier
In the example we must provide the guarantee that the memory load for
x is completed before the one for w, however this ordering can be
reversed with the intervention of a concurrent thread, since the UGM
fence will block on the prior UGM load and potentially take a long
time, while the TGM fence may complete and invalidate the TGM cache
immediately, so a concurrent thread could pollute the TGM cache with
stale contents for the w location *before* the UGM load has completed,
leading to an inversion of the expected memory ordering.
v2: Apply the workaround regardless of whether the NIR barrier
intrinsic specifies multiple storage classes or a single one,
since an acquire barrier is required to order subsequent requests
relative to previous atomic requests of unknown storage class not
necessarily specified by the memory scope information of the
intrinsic.
Cc: mesa-stable
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20690>
