This is for Gfx12.5 with the COMPUTE_WALKER::Inline Data payload. We
do this in a similar way to the compute kernels.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13171>
We only have a single prog_data::total_scratch for all shader variants
(SIMD 8, 16, 32). Therefore we should always max the total_scratch on
top of existing variant.
We probably haven't run into that issue before because we compile by
increasing SIMD size and higher SIMD size is more likely to spill. But
for bindless shaders with return shaders, if the last return part
doesn't spill, we completely ignore the previous parts' scratch
computation.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15193>
If we say that per-primitive attributes are flat (which is communicated by
3DSTATE_SBE.ConstantInterpolationEnable), GPU freaks out and applies it
to other (non-flat) attributes.
Fixes: be89ea3231 ("intel/compiler: Handle per-primitive inputs in FS")
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15169>
Bit is flipped compared to all the other packets.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 705395344d ("intel/fs: Add support for compiling bindless shaders with resume shaders")
Fixes: c3ac9afca3 ("anv: Create and return ray-tracing pipeline SBT handles")
Acked-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15078>
This bit in the compiler looks like it was added by accident on one of
the latest versions of the original commit, but it clearly doesn't
belong there.
Fixes: 03e1e19246 ("anv: Refactor descriptor copy")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15016>
This (sort of) matches the behavior of nir_opt_algebraic. This ensures
that subnormal values are properly flushed to zero.
With the aid of "nir/search: Float sources of texture instructions are
float users" and "nir/search: Transitively apply is_only_used_as_float",
there would have been no shader-db regressions on Intel platforms.
However, those caused a significant increase in compile time. Since the
instruction regressions were so small, I just dropped those commits
rather than improve them.
All Haswell and newer platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 20125042 -> 20125094 (<.01%)
instructions in affected programs: 7184 -> 7236 (0.72%)
helped: 0
HURT: 32
HURT stats (abs) min: 1 max: 4 x̄: 1.62 x̃: 2
HURT stats (rel) min: 0.11% max: 1.49% x̄: 0.85% x̃: 0.78%
95% mean confidence interval for instructions value: 1.39 1.86
95% mean confidence interval for instructions %-change: 0.74% 0.96%
Instructions are HURT.
total cycles in shared programs: 862745586 -> 862746551 (<.01%)
cycles in affected programs: 109872 -> 110837 (0.88%)
helped: 12
HURT: 23
helped stats (abs) min: 2 max: 774 x̄: 90.83 x̃: 19
helped stats (rel) min: 0.07% max: 25.23% x̄: 3.06% x̃: 0.40%
HURT stats (abs) min: 2 max: 1106 x̄: 89.35 x̃: 12
HURT stats (rel) min: 0.08% max: 45.40% x̄: 3.01% x̃: 0.47%
95% mean confidence interval for cycles value: -60.09 115.23
95% mean confidence interval for cycles %-change: -2.21% 4.07%
Inconclusive result (value mean confidence interval includes 0).
All of the shaders hurt are in either UE4 shooter-game or shooter_demo.
Tiger Lake
Instructions in all programs: 159893213 -> 159893290 (+0.0%)
SENDs in all programs: 6936431 -> 6936431 (+0.0%)
Loops in all programs: 38385 -> 38385 (+0.0%)
Cycles in all programs: 7019259514 -> 7019260087 (+0.0%)
Spills in all programs: 101389 -> 101389 (+0.0%)
Fills in all programs: 131532 -> 131532 (+0.0%)
Ice Lake
Instructions in all programs: 143624164 -> 143624235 (+0.0%)
SENDs in all programs: 6980289 -> 6980289 (+0.0%)
Loops in all programs: 38383 -> 38383 (+0.0%)
Cycles in all programs: 8440082767 -> 8440083238 (+0.0%)
Spills in all programs: 102246 -> 102246 (+0.0%)
Fills in all programs: 131908 -> 131908 (+0.0%)
Skylake
Instructions in all programs: 134185424 -> 134185495 (+0.0%)
SENDs in all programs: 6938790 -> 6938790 (+0.0%)
Loops in all programs: 38356 -> 38356 (+0.0%)
Cycles in all programs: 8222366529 -> 8222366923 (+0.0%)
Spills in all programs: 98821 -> 98821 (+0.0%)
Fills in all programs: 125218 -> 125218 (+0.0%)
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Fixes: f5dd6dfe01 ("anv: enable VK_KHR_shader_float_controls and SPV_KHR_float_controls")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13999>
Refactor descriptor copies to use the existing helper functions instead
of rolling our own. In order to facilitate this, we need to store the
appropriate buffer views for the relevant descriptors internally and
reuse them in the helpers.
Signed-off-by: Rohan Garg <rohan.garg@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14909>
In the future we'll want to reuse this intrinsic to deal with ray
queries. Ray queries will use a different global pointer and
programmatically change the control/level arguments of the trace send
instruction.
v2: Comment on barrier after sync trace instruction (Caio)
Generalize lsc helper (Caio)
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/13719>
VARYING_SLOT_{VIEWPORT,LAYER,PSIZ} all live in the same VUE header slot,
and the FS is already set up to read the x/y/z/w component of that vec4.
However, we were setting up the SBE to pass each of those items as a
separate FS input, so hypothetically if a shader read all three, we
would burn 3 FS inputs with redundant data. Not only was this passing
extra data to the FS, but it would count as extra input slots for the
"Do we have 16 or fewer attributes?" check for using SBE swizzling to
rearrange them in a convenient manner.
Now we make them share a single FS attribute and only count them once.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14210>
As indicated by
VkPhysicalDeviceFragmentShadingRatePropertiesKHR::fragmentShadingRateWithShaderSampleMask
our implementation will clamp to 1x1 when reading samplemask or
writing to samplemask.
This fixes vkd3d-proton tests test_sample_mask_dxbc & test_sample_mask_dxil
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: b6332fc4a8 ("intel/compiler: handle coarse pixel in render target writes descriptors")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14553>
Because 0 is no longer a recognizable value (it's NEVER, which isn't a
good default), we add an emit_alpha_test bool to tell the back-end when
to bother alpha testing. This lets us only touch crocus with the
change.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14157>
There's nothing in bspec that would suggest this is still needed.
It only affected gfx 9 and 10.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14280>
While our LIFO scheduling mode attempts to optimize for register
pressure, it's often hard for a scheduling algorithm to do better than
the instruction order provided by the shader author. Shader authors
often do perfectly reasonable things like using texture results
immediately after fetching them or constructing texture coordinates
immediately before the texture op. When we throw all the instruction
ordering information away, we loose any help the author may have given
us. By attempting NONE before we fall back to the worst case LIFO mode.
And, yes, I tried this with NONE both before and after LIFO and doing
NONE before LIFO is substantially better, according to shader-db.
total instructions in shared programs: 19673152 -> 19665202 (-0.04%)
instructions in affected programs: 33669 -> 25719 (-23.61%)
helped: 20
HURT: 0
helped stats (abs) min: 15 max: 4609 x̄: 397.50 x̃: 107
helped stats (rel) min: 2.33% max: 67.50% x̄: 14.60% x̃: 9.12%
95% mean confidence interval for instructions value: -867.61 72.61
95% mean confidence interval for instructions %-change: -21.74% -7.46%
Inconclusive result (value mean confidence interval includes 0).
total cycles in shared programs: 935562500 -> 935020920 (-0.06%)
cycles in affected programs: 18620349 -> 18078769 (-2.91%)
helped: 104
HURT: 48
helped stats (abs) min: 88 max: 60986 x̄: 8031.48 x̃: 3680
helped stats (rel) min: 0.61% max: 51.44% x̄: 14.95% x̃: 8.87%
HURT stats (abs) min: 10 max: 54724 x̄: 6118.62 x̃: 1530
HURT stats (rel) min: 0.13% max: 46.45% x̄: 10.28% x̃: 6.46%
95% mean confidence interval for cycles value: -5724.34 -1401.71
95% mean confidence interval for cycles %-change: -9.86% -4.10%
Cycles are helped.
total spills in shared programs: 12158 -> 10327 (-15.06%)
spills in affected programs: 1831 -> 0
helped: 20
HURT: 0
total fills in shared programs: 14749 -> 12635 (-14.33%)
fills in affected programs: 2114 -> 0
helped: 20
HURT: 0
LOST: 8
GAINED: 649
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>
The way the current scheduler loop is implemented, each scheduling pass
starts with what the previous pass had. This means that, if PRE screwed
everything up majorly, PRE_NON_LIFO would have to try to fix it. It
also meant that tiny changes to one pass would affect every later pass.
Instead, reset the order of the instructions before each scheduling
pass. This makes the passes entirely independent of each other.
Shader-db results on Ice Lake:
total instructions in shared programs: 19670486 -> 19670648 (<.01%)
instructions in affected programs: 25317 -> 25479 (0.64%)
helped: 2
HURT: 7
helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4
helped stats (rel) min: 0.07% max: 0.07% x̄: 0.07% x̃: 0.07%
HURT stats (abs) min: 8 max: 70 x̄: 24.29 x̃: 12
HURT stats (rel) min: 0.41% max: 4.95% x̄: 1.47% x̃: 0.87%
95% mean confidence interval for instructions value: -1.28 37.28
95% mean confidence interval for instructions %-change: -0.04% 2.30%
Inconclusive result (value mean confidence interval includes 0).
total cycles in shared programs: 935535948 -> 935490243 (<.01%)
cycles in affected programs: 421994824 -> 421949119 (-0.01%)
helped: 1269
HURT: 879
helped stats (abs) min: 1 max: 12008 x̄: 259.38 x̃: 52
helped stats (rel) min: <.01% max: 28.02% x̄: 1.12% x̃: 0.14%
HURT stats (abs) min: 1 max: 29931 x̄: 322.46 x̃: 20
HURT stats (rel) min: <.01% max: 32.17% x̄: 1.74% x̃: 0.22%
95% mean confidence interval for cycles value: -71.37 28.81
95% mean confidence interval for cycles %-change: -0.11% 0.21%
Inconclusive result (value mean confidence interval includes 0).
total spills in shared programs: 12403 -> 12430 (0.22%)
spills in affected programs: 1355 -> 1382 (1.99%)
helped: 2
HURT: 7
total fills in shared programs: 15128 -> 15182 (0.36%)
fills in affected programs: 3294 -> 3348 (1.64%)
helped: 2
HURT: 7
LOST: 21
GAINED: 28
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>
This reverts commit ba2fa1ceaf. Doing
optimizations after scheduling but before RA means doing them in the
middle of the scheduling loop which introduces additional dependencies
between one scheduling iteration and the next. That won't work if we
want to make the scheduling modes independent, at least not unless we
have some way of fully cloning the IR.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>
Now that we're being conservative in the pass, it's easy to tell when it
makes progress and we can put it in the OPT() macro. This way, we get
nice INTEL_DEBUG=optimizer dumps for it. While we're here, fix the
header comment which is massively out-of-date.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>
Instead of modifying every single instruction, keep track of which VGRFs
are actually split in a bit-set, and only modify the instructions that
actually touch split regs.
This cuts the runtime of the following Vulkan CTS on my SKL box by 45%
from 3:21 to 1:51: dEQP-VK.ssbo.phys.layout.random.16bit.scalar.13
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>
This rolls compute_sample_position into emit_samplepos_setup, its only
caller, by using a loop instead of calling it twice. We also
early-return for the !persample_dispatch case instead of doing it as
part of the sample calculation. This means that we don't call
fetch_payload_reg() to get sample_pos_reg unless we're actually going to
use it so the function is safe to call even if we haven't set up
sample_pos_reg. This will be important for the next commit.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14198>
Use the same URB access helpers that were added for Task Output. The
Arrayed I/O (per-primitive and per-vertex) is handled by applying the
pitch from the MUE layout into the NIR intrinsics and including the
non-arrayed offset on top of it. After that, the index src can be
used directly for lowering.
Because we keep around the non-arrayed offset AND the pitch is
aligned, we can identify cases where the access is indirect but
guaranteed to be aligned, and dispatch a single message. Added a TODO
to explore that later.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13661>
Task/Mesh stages are CS-like stages, and include many
builtins (e.g. workgroup ID/index) and intrinsics (e.g. workgroup
memory primitives) originally present only in CS.
This commit add two new stages (task and mesh) that 'inherit' from CS
by embedding a brw_cs_prog_data in their own prog_data structure, so
that CS functionality can be easily reused. They also currently use
the same helpers to select the SIMD variant to use -- that was
recently added for CS.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13661>
Allows to assert its existence for per-primitive variables and will
later be useful to implement the "more than 16 attributes" case for
Mesh.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13661>
In Fragment Shader, regular inputs are laid out in the thread payload
in a one dword per each half-GRF, that gives room for having the two
delta dwords needed for interpolation.
Per-primitive inputs are laid out before the regular inputs, and since
there's no need to have delta information, they are packed. So
half-GRF will be fully filled with 4 dwords of input.
When num_per_primitive_inputs is zero (the default case), behavior
should be the same as before.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13661>
This allows, for example, fs_inst::components_read() without passing
devinfo as extra argument.
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11766>
For SIMD8 half float payload, each component takes a full register, so
we can use existing LOAD_PAYLOAD infrastruture for required padding by
alternating plain 8-wide half float vector and null vector.
Also this patch removes an unwanted assertion from
opt_copy_propagation_local for LOAD_PAYLOAD.
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11766>
To support SIMD8 half float payloads, each component takes one full
32bit wide register in both SIMD8H and SIMD16H mode. So we can make use
of existing LOAD_PAYLOAD infrastructure alternating a half float vector
and a null vector, in order to handle required padding.
v2: (Francisco)
- Skip header sources
- Fix comparision units
- Don't allocate VGRF for padded source
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Suggested-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11766>
We can use LOAD_PAYLOAD infrastructure in order to handle 16bit float
payload. Let's rely on source type for padding sources, if not set
previously then default one would be 32-bit.
This patch will be used later in the series to handle 16-bit float
payloads.
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11766>
