Consider an innocuous instruction like:
and(1) v250:UD, g0.3<0,1,0>:UD, 4294967264u NoMask group0
If register allocation decides to spill v250, it will see this
instruction and say, "Oh no! The other components of v250 aren't set, so
I'd better add a fill before that instruction!"
But it gets even worse than that... if register coalesce decided to
merge two of these, the live range gets massively extended because the
writes don't fully initialize the value. This causes the need to spill
these registers in the first place.
Changing that instruction to SIMD16 on Xe2 or SIMD8 on other platforms
alleviates these issues.
shader-db:
Lunar Lake
total instructions in shared programs: 17118324 -> 17113191 (-0.03%)
instructions in affected programs: 93701 -> 88568 (-5.48%)
helped: 42 / HURT: 6
total cycles in shared programs: 895422566 -> 895079488 (-0.04%)
cycles in affected programs: 30111338 -> 29768260 (-1.14%)
helped: 35 / HURT: 40
total spills in shared programs: 3588 -> 3304 (-7.92%)
spills in affected programs: 285 -> 1 (-99.65%)
helped: 10 / HURT: 0
total fills in shared programs: 2218 -> 1663 (-25.02%)
fills in affected programs: 556 -> 1 (-99.82%)
helped: 10 / HURT: 0
Meteor Lake, DG2, Tiger Lake, and Ice Lake had similar results. (Meteor Lake shown)
total instructions in shared programs: 20059218 -> 20053563 (-0.03%)
instructions in affected programs: 96938 -> 91283 (-5.83%)
helped: 43 / HURT: 6
total cycles in shared programs: 884174588 -> 883536475 (-0.07%)
cycles in affected programs: 22105268 -> 21467155 (-2.89%)
helped: 35 / HURT: 27
total spills in shared programs: 5032 -> 4679 (-7.02%)
spills in affected programs: 355 -> 2 (-99.44%)
helped: 12 / HURT: 0
total fills in shared programs: 4782 -> 4113 (-13.99%)
fills in affected programs: 671 -> 2 (-99.70%)
helped: 12 / HURT: 0
Skylake
total instructions in shared programs: 19097658 -> 19097665 (<.01%)
instructions in affected programs: 14202 -> 14209 (0.05%)
helped: 0 / HURT: 5
total cycles in shared programs: 862058109 -> 862058267 (<.01%)
cycles in affected programs: 3450244 -> 3450402 (<.01%)
helped: 7 / HURT: 11
fossil-db:
Lunar Lake
Totals:
Cycle count: 31439652246 -> 31439652272 (+0.00%)
Totals from 2 (0.00% of 707091) affected shaders:
Cycle count: 2602 -> 2628 (+1.00%)
No other Intel platforms had any fossil-db changes.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35721>
`getopt_long()` returns an `int`, not a `char`; putting the value in
a `char` before comparing it to `-1` was making the comparison always
fail, resulting in the invalid codepath taken that then fails with:
option `-' is invalid: ignored
cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34756>
`subprocess.Popen()` returns immediately, and the subprocess might not
have finished by the time `stdout` is read on the next line, spuriously
failing the tests.
`subprocess.check_output()` makes sure the output is available before
returning, solving this issue; it additionally raises an error if the
subprocess failed, giving a better error than a failed diff later in the
script.
cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34756>
e4m3fn: 8bit floating point format with 4bit exponent, 3bit mantissa
and no infinities (finite only)
e5m2: 8bit floating point format with 5bit exponent, 2bit mantissa
and with infinities.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35434>
This will later be encoded by the backend into the
LSC extended descriptor message.
Reworks:
* Sagar: Add nir_intrinsic_ssbo_atomic_swap
Signed-off-by: Rohan Garg <rohan.garg@intel.com>
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35252>
We're not currently using RANGE_BASE and we'll use BASE for offset
optimizations on Xe2+.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35252>
Given that the intrinsic will be CSEed at the NIR level, we don't need to
preemptively set it up at the top of the shader. No change in HSW shader-db.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
Unless I've seriously missed something, we have the Z in the payload
(which we can always request if we need access to it and it's not already
passed to us due other WM IZ settings).
total instructions in shared programs: 4408303 -> 4408186 (<.01%)
instructions in affected programs: 1164 -> 1047 (-10.05%)
total cycles in shared programs: 142485036 -> 142484566 (<.01%)
cycles in affected programs: 26820 -> 26350 (-1.75%)
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
NIR catches that if you're just doing something like adding two smooth
inputs, we can do the multiply once on the result instead of on each
input. BRW shader-db results:
total instructions in shared programs: 4409146 -> 4408303 (-0.02%)
instructions in affected programs: 800761 -> 799918 (-0.11%)
total cycles in shared programs: 143203198 -> 142485036 (-0.50%)
cycles in affected programs: 79081682 -> 78363520 (-0.91%)
total sends in shared programs: 363044 -> 363042 (<.01%)
sends in affected programs: 33 -> 31 (-6.06%)
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
This moves some conversions to NIR that may get eliminated, and also
distinguishes gl_FragCoord.z/w loads at the shader info level so we don't
need to flag uses_src_depth/uses_src_w when only gl_FragCoord.xy get used
(as is typical). This reduces thread payload setup on many shaders.
Also, interestingly, blorp shaders stop reserving space for z/w despite
not putting them in the payload (since PS_EXTRA isn't filled out for z/w).
HSW shader-db is noise:
total instructions in shared programs: 9942649 -> 9942997 (<.01%)
instructions in affected programs: 143167 -> 143515 (0.24%)
total cycles in shared programs: 314768862 -> 314299112 (-0.15%)
cycles in affected programs: 62951452 -> 62481702 (-0.75%)
LOST: 44
GAINED: 26
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
This will be used for representing gl_FragCoord in NIR and reducing
payload registers pushed.
HSW results:
total instructions in shared programs: 9940636 -> 9948574 (0.08%)
instructions in affected programs: 852560 -> 860498 (0.93%)
total cycles in shared programs: 314804525 -> 314900080 (0.03%)
cycles in affected programs: 39786599 -> 39882154 (0.24%)
LOST: 5
GAINED: 11
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
This was leftover dead code from 4bb6e6817e ("intel: Use a system value
for gl_FragCoord") -- the sysval doesn't do any interpolation and doesn't
have sources that could use a barycentric.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25190>
This generates one extra instruction to set the rounding mode to RTE due
to f2f16_rtne in the lowering. This changes the result for
fquantize2f16(65505.0) from 65536 to 65504, which fixes SPIR-V
conformance for this value:
If Value is positive with a magnitude too large to represent as a
16-bit floating-point value, the result is positive infinity. If Value
is negative with a magnitude too large to represent as a 16-bit
floating-point value, the result is negative infinity.
SPIR-V doesn't specify whether this overflow check is before or after
rounding, but IEEE specifies rounding first, which is what produces our
65504.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25552>
Source: BSpec "Instruction Fields" page (56701), SWSB field.
Credits to Caio Oliveira here, since he was helping me while we found
this issue together.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35395>
This would cause Mesa to print this message even if an Intel GPU is just
being enumerated by a Vulkan application. For example, `vulkaninfo
--summary`.
Fixes: 52f73db5b7 ("brw: implement read without format lowering")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35396>
The intention here is to build a SIMD8 value, that will be expanded
as needed -- just like the SHL/ADD case, but with a single instruction.
Found when the was triggering invalid MAD with SIMD32 (that gets compressed)
*and* with overlapping destination and source *and* which would cause
conflict when divided into two SIMD16.
Fixes: 338273dedd ("brw/reg_allocate: Optimize spill offset calculation using integer MAD")
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35302>
Load the format enum and then just go through a series of :
if format == R16G16B16A16_UNORM
color = lower_r32g32_uint_tor_r16g16b16a16_unorm(color)
else if format == R16G16B16A16_SNORM
...
For Gfx12.5, there is no in-shader conversion.
For Gfx12/11, the in-shader conversion covers the following formats :
- ISL_FORMAT_R10G10B10A2_UNORM
- ISL_FORMAT_R10G10B10A2_UINT
- ISL_FORMAT_R11G11B10_FLOAT
For Gfx9, the following formats :
- ISL_FORMAT_R16G16B16A16_UNORM
- ISL_FORMAT_R16G16B16A16_SNORM
- ISL_FORMAT_R10G10B10A2_UNORM
- ISL_FORMAT_R10G10B10A2_UINT
- ISL_FORMAT_R8G8B8A8_UNORM
- ISL_FORMAT_R8G8B8A8_SNORM
- ISL_FORMAT_R16G16_UNORM
- ISL_FORMAT_R16G16_SNORM
- ISL_FORMAT_R11G11B10_FLOAT
- ISL_FORMAT_R8G8_UNORM
- ISL_FORMAT_R8G8_SNORM
- ISL_FORMAT_R16_UNORM
- ISL_FORMAT_R16_SNORM
- ISL_FORMAT_R8_UNORM
- ISL_FORMAT_R8_SNORM
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/22524>
0e3e5146cf ("intel/brw: Use correct instruction for value change check
when coalescing") enabled some new cases that exposed a pre-existing
bug that would turn something like this :
mul.sat(16) %789:F, %787:F, %788:F
mov.g.f0.0(16) %790:F, %789:F
(+f0.0) sel(16) %800:UD, %790:UD, 0u
into this :
mul.sat(16) %790:F, %787:F, %788:F
mov.g.f0.0(16) null:F, null<8,8,1>:F
(+f0.0) sel(16) %800:UD, %790:UD, 0u
The mov[] array can contain the same instruction because it's repeated
for each REG_SIZE writes and a SIMD16 instruction will write 2
REG_SIZE.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 0e3e5146cf ("intel/brw: Use correct instruction for value change check when coalescing")
Cc: mesa-stable
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35276>
There's always only the ARF scalar register source, so don't
bother printing other information that won't be used. Matches
the assembler code.
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35297>
When we have partial VGRF MOVs with offsets, we will reach
`channels_remaining == 0` with `inst` that is not writing the whole VGRF.
Currently, even though we check `can_coalesce_vars()` for each offset
separately, it will always check if the dst value is not changed only
for the offset from the instruction that satisfied the
`channels_remaining == 0` condition.
Instead, we should remember and use the correct instruction for each
written offset separately.
Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10916
Signed-off-by: Sviatoslav Peleshko <sviatoslav.peleshko@globallogic.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35062>
This fprintf() was added in commit cce3bea2a7 ("i965/disasm: Align send
instruction meta-information with dst.")) to align the human-readable
send message info (e.g. "render MsgDesc: RT write ...") with the
destination register on the previous line.
Two months later we disabled printing the instruction offset in commit
662f1ccc24 ("i965: Disable hex offset printing in disassembly."),
thereby unaligning the human-readable send message info for the next 11
years.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35077>
Previously, brw_try_override_assembly was only called when a debug flag was
enabled. However, during investigations involving workloads such as Steam
games, enabling the debug flag results in excessive NIR and ISA output to
stderr, making debugging more difficult.
This change ensures that brw_try_override_assembly is called when the
INTEL_SHADER_ASM_READ_PATH is set, regardless of the debug flag. This
improves usability in scenarios where minimal debug output is desired.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35115>
Those intrinsics have different semantics in particular with regards
to divergence. Turning one into the other without invalidating the
divergence information breaks NIR validation. But also the conversion
means we get artificially less convergent values in the shaders.
So just handle load_push_constants in the backend and stop changing
things in Hasvk.
Fixes a bunch of tests in
dEQP-VK.descriptor_indexing.*
dEQP-VK.pipeline.*.push_constant.graphics_pipeline.dynamic_index_*
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34546>
Those intrinsics have different semantics in particular with regards
to divergence. Turning one into the other without invalidating the
divergence information breaks NIR validation. But also the conversion
means we get artificially less convergent values in the shaders.
So just handle load_push_constants in the backend and stop changing
things in Anv.
Fixes a bunch of tests in
dEQP-VK.descriptor_indexing.*
dEQP-VK.pipeline.*.push_constant.graphics_pipeline.dynamic_index_*
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34546>
This patch improves code readability by centralizing the type stomping
logic for Gen12.5 region restrictions in `brw_lower_alu_restrictions`.
It removes redundant comments and ensures type consistency assertions
in `brw_broadcast`, `generate_mov_indirect`, and `generate_shuffle`.
Thank you Ken for guiding me on this.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35006>
This will be necessary to select the right value for flat inputs in
fragment shaders when fragment shader barycentrics are in use.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34445>
