This is just the skeleton of the implementation. Future commits will
fill it all in.
v2: Move to src/intel/compiler
v3 (idr): Use vecN instead of array[N] for slice type.
v4 (idr): Refactor lower_cooperative_matrix_load and
lower_cooperative_matrix_store into a single function.
v5 (idr): Remove old, verbose debug logging. Assert that entry is not
NULL in get_coop_type_for_slice. Use nir_component_mask(...) instead of
0xffff. s/cooperative_matrix/cmat/. All suggested by Caio.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
I put both R-b on this because, at this point, we've each done equal
parts authoring and reviewing.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25994>
The mechanism for selecting dispatch modes has changed from previous
platforms, add a new implementation brw_wm_state_simd_width_for_ksp()
using the new kernel dispatch controls.
[ Francisco Jerez: Split from a larger patch, handle multipolygon
dispatch, add additional comments. ]
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
Extend the pre-existing dual-SIMD8 compilation path in
brw_compile_fs() to attempt quad-SIMD8 and dual-SIMD16 compiles.
Instead of building every possible dispatch mode and then picking one
based on cycle-count heuristics, this attempts to only build a single
multipolygon kernel -- The different mulipolygon dispatch modes are
tried in the expected order of decreasing performance (quad-SIMD8,
dual-SIMD16 then dual-SIMD8), the first one that successfully compiles
without spills is taken as a simple heuristic, and no further
multipolygon builds are attempted.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
This is needed because the information stored on the ATTR file for
multipolygon fragment shaders isn't stored as a contiguous sequence in
the GRF, instead the ATTR source may be lowered by assign_urb_setup()
to use a <16;8,0> region, which reads 4 SIMD16 GRFs for a SIMD32
instruction, even though the result of fs_inst::size_read() is
expected to be 2 GRFs. Special case ATTR sources for multipolygon PS
shaders to calculate the number of physical GRFs that will actually be
read by the instruction after lowering, based on the number of
polygons processed by the instruction.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
This is based on a previous patch by Marcin Ślusarz addressing the
same issue, though it's largely rewritten, simplified and includes
additional fixes.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
This fixes a number of assumptions made by the multipolygon input
attribute handling code from assign_urb_setup() so it also works on
Xe2+, which has additional multipolygon dispatch modes (like SIMD4x8
and SIMD2x16) and uses a different more compact representation of the
plane parameters.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
The interpolation deltas of PS inputs now show up as a 12B vec3 (A0,
A1-A0, A2-A0) in the ATTR file, instead of the previously used 16B
format with an unused component.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
The X and Y barycentric vectors are no longer interleaved in SIMD8
chunks (yay), so this is mostly a matter of disabling the
lower_barycentrics() pass and switching to a simpler implementation of
fetch_barycentric_reg() that simply calls fetch_payload_reg() instead
of the SIMD8 shuffling we had to do in previous generations.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
This extends fetch_payload_reg() to support fetching vector registers
like barycentrics stored on the payload as a contiguous sequence of
SIMD-wide vectors. In the SIMD32 case, both halves of the SIMD16
vector registers specified as regs[0] and regs[1] are zipped to
construct a single SIMD32-wide vector.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
This includes the render target array index, viewport index, and
front/back facing fields, which are now replicated per pair of
subspans in order to support fixed-layout multi-polygon PS dispatch.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
Note from Caio: proper handling of brw_sample_mask_reg
will appear in later patches.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
The PS thread payload format has changed enough in Xe2 that it
probably doesn't make sense to share code with gfx6. See BSpec page
"PS Thread Payload for Normal Dispatch - 512 bit GRF" for the new
format.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
These are new variants of the existing brw_reg GRF constructors that
take registers numbers in the new 512b units. Mainly useful for
thread payload setup code to use register numbers in a format that
matches the BSpec.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26606>
These restrictions don't seem to be applicable anymore, and limiting
to SIMD8 wouldn't work since we're no longer building shaders with
that dispatch width.
[ Francisco: This one-liner change was squashed by Rohan Garg into a
previous version of my patch "Stop building SIMD8 programs", but it
makes more sense as a separate commit -- Formatted as a separate
patch. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26605>
SIMD8 kernels are no longer able to utilize the ALUs efficiently,
since they have twice the vector width as previous platforms. However
even though there aren't many reasons to use it, SIMD8 is still
supported by the instruction set technically, and it will still be
used for some SIMD-lowering sequences.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26605>
Similar to other FS dispatch modes, attempt to build a dual-SIMD8
program if the regular SIMD8 program didn't spill and doubling the
amount of space for varyings doesn't cause us to go over the thread
payload limit. Dual-SIMD8 builds in combination with coarse pixel
shading are currently not handled.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
It's not replicated per SIMD16 half of a SIMD32 thread on the PS
payload. Make fs_visitor::payload::depth_w_coef_reg a scalar rather
than an array.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
The copy would be discarded immediately. Until now we were relying on
DCE to eliminate these, but it seems like in some cases MOVs into the
null register emitted by lower_simd_width() are never eliminated,
likely because a lower_simd_width() call has been introduced close to
the bottom of optimize() which isn't follow by any additional DCE
passes.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
This fixes a number of regressions and hangs in multipolygon fragment
shaders that have FIND_LIVE_CHANNEL sequences which would otherwise
lead to access of a dead channel. Note that the failures don't seem
to be reproducible in simulation.
Acked-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Flat-shaded inputs and other per-primitive values can no longer be
considered to be uniform across fragment shader subgroups due to
multipolygon dispatch.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Since the <8;8,0> regions they use in multipolygon mode could violate
regioning restrictions in some cases, depending on the execution type
of the instruction. Note that the assertion is removed from
try_copy_propagate() since a more accurate check is used within that
function than what fs_inst::can_change_types() can do.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
The updated layout includes one copy of each plane parameter per
channel of the SIMD thread, in order to allow channels to process
different polygons.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Main motivation is that for multipolygon PS shaders the i-th plane
parameter for the j-th input attribute will no longer necessarily be a
scalar, since different channels may be processing different polygons
with different input plane parameters, so simply taking a component()
of the result of interp_reg() will no longer work. Instead of
duplicating the multipolygon handling logic in every caller of
interp_reg(), fold the component() call into interp_reg() so we can
replace it with multipolygon-correct code more easily.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Instead of treating fs_reg::nr as an offset for ATTR registers simply
consider different indices as denoting disjoint spaces that can never
be accessed simultaneously by a single region. From now on geometry
stages will just use ATTR #0 for everything and select specific
attributes via offset() with the native dispatch width of the program,
which should work on current platforms as well as on Xe2+. See
"intel/fs: Map all GS input attributes to ATTR register number 0." for
the rationale.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Instead of treating fs_reg::nr as an offset for ATTR registers simply
consider different indices as denoting disjoint spaces that can never
be accessed simultaneously by a single region. From now on geometry
stages will just use ATTR #0 for everything and select specific
attributes via offset() with the native dispatch width of the program,
which should work on current platforms as well as on Xe2+. See
"intel/fs: Map all GS input attributes to ATTR register number 0." for
the rationale.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Instead of treating fs_reg::nr as an offset for ATTR registers simply
consider different indices as denoting disjoint spaces that can never
be accessed simultaneously by a single region. From now on geometry
stages will just use ATTR #0 for everything and select specific
attributes via offset() with the native dispatch width of the program,
which should work on current platforms as well as on Xe2+. See
"intel/fs: Map all GS input attributes to ATTR register number 0." for
the rationale.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Instead of treating fs_reg::nr as an offset for ATTR registers simply
consider different indices as denoting disjoint spaces that can never
be accessed simultaneously by a single region. From now on geometry
stages will just use ATTR #0 for everything and select specific
attributes via offset() with the native dispatch width of the program,
which should work on current platforms as well as on Xe2+. See
"intel/fs: Map all GS input attributes to ATTR register number 0." for
the rationale.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
The fs_reg::nr field currently has a somewhat inconsistent meaning for
ATTR registers depending on the shader stage. In geometry stages it
has a similar effect as fs_reg::offset except it's expressed in 32B
units instead of B units. In the PS however it's expressed in units
of logical scalar attributes (16B on present platforms), which isn't
currently handled correctly throughout the back-end since some places
assume 32B units in all cases.
The different format of the PS setup data in multi-polygon dispatch
modes would make its behavior even more irregular, which would be
worsened further (for both geometry and pixel stages) by the register
size changes coming up on Xe2, particularly in brw_ir_fs.h helpers
where neither the devinfo struct nor the shader stage are available.
Instead of treating it as an offset simply consider different
fs_reg::nr indices as denoting disjoint spaces that can never be
accessed simultaneously by a single region. From now on geometry
stages will just use ATTR #0 for everything and select specific
attributes via offset() with the native dispatch width of the program,
which should work on current platforms as well as on Xe2+.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
To allow specifying the number of polygons that will be processed per
SIMD thread.
Rework:
* Jordan: Add needs_register_pressure following
09cdb77a92 ("intel/fs: report max register pressure in shader stats")
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
And use it in ANV in order to return a "SIMDNxM" name from
vkGetPipelineExecutablePropertiesKHR.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Add fields that track the number of polygons processed per PS SIMD
thread (note that this might be lower than the value that was
specified to the compiler via brw_compile_fs_params if compilation at
the desired polygon count wasn't possible), and the dispatch width of
the multi-polygon PS kernel.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
Add a brw_compile_fs_params parameter that specifies to the compiler
the maximum number of polygons that may be processed in parallel per
PS SIMD thread.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26585>
In the past, multiple writes to a single register were pretty common,
but since we've transitioned to NIR, and leave the IR in SSA form for
everything not captured in a phi-web, the pattern of generating new
temporary registers at each step is a lot more common.
This pass isn't nearly as useful now. Across fossil-db on Alchemist,
this affects only 0.55% of shaders, which fall into two cases:
- Coarse pixel shading pixel-X/Y setup. There are a few cases where
we write a partial calculation into a register, then have a second
instruction read that as a source and overwrite it as a destination.
While we could use a temporary here, it doesn't actually help with
register pressure at all, since there's the same amount of values
live at both instructions regardless. So while this pass kicks in,
it doesn't do anything useful.
- Geometry shader control data bits (5 shaders total). We track masks
for handling EndPrimitive in a single register across the program,
and apparently in some cases can split the live range. However, it's
a single register...only in geometry shaders...which use EndPrimitive.
None of them appear to be in danger of spilling, either. So this tiny
benefit doesn't seem to justify the cost of running the pass.
So, just throw it out. It's not worth keeping.
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/26343>
