Some hardware, like i965, doesn't support group sizes greater than 32.
In that case, we can reduce the destination size of the ballot
intrinsic, which will simplify our code generation.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The implementation of ballotARB() will start by zeroing the flags
register. So, a doing something like
if (gl_SubGroupInvocationARB % 2u == 0u) {
... = ballotARB(true);
[...]
} else {
... = ballotARB(true);
[...]
}
(like fs-ballot-if-else.shader_test does) would generate identical MOVs
to the same destination (the flag register!), and we definitely do not
want to pull that out of the control flow.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The implementations of the ARB_shader_ballot intrinsics will explicitly
read the flag as a source register.
Reviewed-by: Matt Turner <mattst88@gmail.com>
We already had a channel_num system value, which I'm renaming to
subgroup_invocation to match the rest of the new system values.
Note that while ballotARB(true) will return zeros in the high 32-bits on
systems where gl_SubGroupSizeARB <= 32, the gl_SubGroup??MaskARB
variables do not consider whether channels are enabled. See issue (1) of
ARB_shader_ballot.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The implementations of the ARB_shader_group_vote intrinsics will
explicitly write the flag as the destination register.
Reviewed-by: Matt Turner <mattst88@gmail.com>
I don't expect anyone is going to care about using this in vec4 programs
(vertex/tessellation/geometry on Gen6/7), no one has come up with a good
way to implement it much less test it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Specifically, constant fold intrinsics from ARB_shader_group_vote, but I
suspect it'll be useful for other things in the future.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This actually takes advantage of the newly pushed UBO data, avoiding
pull loads.
Improves performance in GLBenchmark Manhattan 3.1 by:
HSW: ~1%, BDW/SKL/KBL GT2: 3-4%, SKL GT4: 7-8%, APL: 4-5%.
(thanks to Eero Tamminen for these numbers)
shader-db results on Skylake, ignoring programs with spill/fill changes:
total instructions in shared programs: 13963994 -> 13651893 (-2.24%)
instructions in affected programs: 4250328 -> 3938227 (-7.34%)
helped: 28527
HURT: 0
total cycles in shared programs: 179808608 -> 172535170 (-4.05%)
cycles in affected programs: 79720410 -> 72446972 (-9.12%)
helped: 26951
HURT: 1248
LOST: 46
GAINED: 21
Many "Deus Ex: Mankind Divided" shaders which already spilled end up
spill a lot more (about 240 programs hurt, 9 helped). The cycle
estimator suggests this is still overall a win (-0.23% in cycle counts)
presumably because we trade pull loads for fills.
v2: Drop "PULL" environment variable left in for initial debugging
(caught by Matt).
Reviewed-by: Matt Turner <mattst88@gmail.com>
With UBOs, the answer of "have we decided to push this uniform" gets
a bit more complicated - for one, we have multiple surfaces. This
patch refactors things so we can add the new code in a single place.
Reviewed-by: Matt Turner <mattst88@gmail.com>
This patch starts uploading UBO data via 3DSTATE_CONSTANT_* packets,
and updates the compiler to know that there's extra payload data, so
things continue working. However, it still issues pull loads for all
data. I wanted to separate the two aspects for greater bisectability.
v2: Update for new intel_bufferobj_buffer parameter.
Reviewed-by: Matt Turner <mattst88@gmail.com>
This adds a NIR pass that decides which portions of UBOS we should
upload as push constants, rather than pull constants.
v2: Switch to uint16_t for the UBO block number, because we may
have a lot of them in Vulkan (suggested by Jason). Add more
comments about bitfield trickery (requested by Matt).
v3: Skip vec4 stages for now...I haven't finished wiring up support
in the vec4 backend, and so pushing the data but not using it
will just be wasteful.
Reviewed-by: Matt Turner <mattst88@gmail.com>
By default, 3DSTATE_CONSTANT_* Constant Buffer 0 is relative to dynamic
state base address. This makes it unusable for pushing UBOs. I'd like
to be able to use all four push buffers.
There is a bit in the INSTPM register (or CS_DEBUG_MODE2 on Skylake)
which controls whether buffer 0 is relative to dynamic state base
address, or simply a normal pointer. Setting that gives us full
flexibility.
We can't currently write this on Haswell and earlier, and will need
to update the kernel command parser, and then do the whole version
checking song and dance.
Reviewed-by: Matt Turner <mattst88@gmail.com>
This optimization has been removed on gen10+.
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
In 5f2fe9302c is_geminilake was introduced for the differenciate
broxton from geminilake. Unfortunately I failed as verifying that
is_broxton is throughout the code base to mean Gen9lp.
Fixes: 5f2fe9302c ("intel: common: add flag to identify platforms by name")
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v1: By Ben Widawsky <benjamin.widawsky@intel.com>
v2: v1 had an assert only for VS. Add the restriction for GS, HS and
DS as well and make sure the allocated sizes are not multiple of 3.
v3: Move the entry_size checks in to compiler code (Ken)
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
V2: Start using gen10 functions isl_gen10*(), gen10_blorp_exec()
gen10_init_atoms() (Jason)
Remove Vulkan changes. Do them later in a separate patch.
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Simple search for a backslash followed by two newlines.
If one of the newlines were to be removed, this would cause issues, so
let's just remove these trailing backslashes.
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We moved to INTEL_SCALAR_* when we added more than a single stage, but
never went back and converted the VS to work that way. Be consistent.
Also update the documentation to actually mention these debug variables.
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
We can just update the gl_transform_feedback_info fields at link time
to make the VUE header fields have the right location and component.
Then we don't need to handle them specially at draw time, which is
expensive.
Reviewed-by: Rafael Antognolli <rafael.antognolli@intel.com>
Scalar mode has been default since Broadwell, and vector mode is getting
increasingly unmaintained. There are a few things that don't even fully
work in vector mode on Skylake, but we've never cared because nobody
uses it. There's no point in porting it forward to new platforms.
So, just ignore the debug options to force it on.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reorder the uniforms to load first the dvec4-aligned variables in the
push constant buffer and then push the vec4-aligned ones. It takes
into account that the relocated uniforms should be aligned to their
channel size.
This fixes a bug were the dvec3/4 might be loaded one part on a GRF and
the rest in next GRF, so the region parameters to read that could break
the HW rules.
v2:
- Fix broken logic.
- Add a comment to explain what should be needed to optimise the usage
of the push constant buffer slots, as this patch does not pack the
uniforms.
v3:
- Implemented the push constant buffer usage optimization.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Acked-by: Francisco Jerez <currojerez@riseup.net>
It was setting XYWZ swizzle and writemask to all uniforms, no matter if they
were a vector or scalar, so this can lead to problems when loading them
to the push constant buffer.
Moreover, 'shift' calculation was designed to calculate the offset in
DWORDS, but it doesn't take into account DFs, so the calculated swizzle
for the later ones was wrong.
The indirect case is not changed because MOV INDIRECT will write
to all components. Added an assert to verify that these uniforms
are aligned.
v2:
- Fix 'shift' calculation (Curro)
- Set both swizzle and writemask.
- Add assert(shift == 0) for the indirect case.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
We are going to add a packing feature to reduce the usage of the push
constant buffer. One of the consequences is that 'nr_params' would be
modified by vec4_visitor's run call, so we need to restore it if one of
them failed before executing the fallback ones. Same thing happens to the
uniforms values that would be reordered afterwards.
Fixes GL45-CTS.arrays_of_arrays_gl.InteractionFunctionCalls2 when
the dvec4 alignment and packing patch is applied.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Acked-by: Francisco Jerez <currojerez@riseup.net>
intel_asm_annotation.c is part of libintel_compiler.la, which contains
code for disassembling and validating shaders that we want to call in
aubinator_error_decode.
dump_assembly() calls nir_print_instr() to print annotations, and
although dump_assembly() is not called by aubinator_error_decode (nor is
any function in intel_asm_annotation.c) it causes undefined references
to nir_print_instr().
To work around, provide a no-op weak symbol to resolve against.
This will allow the validator to run on shader programs we find in the
GPU hang error state.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This will allow us to more easily run brw_validate_instructions() on
shader programs we find in GPU hang error states.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
The only thing still using it is INVOCATION_ID for geometry shaders.
That's easily enough inlined into the nir_intrinsic_load_invocation_id
handling code.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We're already doing this in the FS back-end. This just does the same
thing in the vec4 back-end.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The NIR pass already handles remapping system values to attributes for
us so we delete the system value code as part of the conversion.
We also change nir_lower_vs_inputs to take an explicit inputs_read
bitmask and pass in the inputs_read from prog_data instead from pulling
it out of NIR. This is because the version in prog_data may get
EDGEFLAG added to it on some old platforms.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We also add a nice little comment to make it more clear exactly what
happens with the edge flag copy.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
NIR calls these system values but they come in from the VF unit as
vertex data. It's terribly convenient to just be able to treat them as
such in the back-end.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The vec4 backend will want to count in units of vec4s, not scalar
components. The simplest solution is to move the multiplication by 4
into the scalar backend. This also improves consistency with how we
count varyings.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Now that we have nice block iterators, there's no good reason for this
to be off on it's own. While we're here, we convert to using the NIR
const index getters/setters instead of whacking const_index values
directly.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Commit e1af20f18a changed the shader_info
from being embedded into being just a pointer. The idea was that
sharing the shader_info between NIR and GLSL would be easier if it were
a pointer pointing to the same shader_info struct. This, however, has
caused a few problems:
1) There are many things which generate NIR without GLSL. This means
we have to support both NIR shaders which come from GLSL and ones
that don't and need to have an info elsewhere.
2) The solution to (1) raises all sorts of ownership issues which have
to be resolved with ralloc_parent checks.
3) Ever since 00620782c9, we've been
using nir_gather_info to fill out the final shader_info. Thanks to
cloning and the above ownership issues, the nir_shader::info may not
point back to the gl_shader anymore and so we have to do a copy of
the shader_info from NIR back to GLSL anyway.
All of these issues go away if we just embed the shader_info in the
nir_shader. There's a little downside of having to copy it back after
calling nir_gather_info but, as explained above, we have to do that
anyway.
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
These enums live inside struct brw_wm_prog_data, so it makes sense to
keep them in the same header. It also allows to use them without
including brw_eu_defines.h.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The regioning parameters are now properly set by convert_to_hw_regs()
and we don't need to fix them in the generator. That latter fix
previously done in the generator was strictly speaking wrong for any
non-identity regions.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
On gen7, the swizzles used in DF align16 instructions works for element
size of 32 bits, so we can address only 2 consecutive DFs. As we assumed that
in the rest of the code and prepare the instructions for this (scalarize_df()),
we need to set it to two again.
However, for DF align1 instructions, a width of 2 is wrong as we are not
reading the data we want. For example, an uniform would have a region of
<0, 2, 1> so it would repeat the first 2 DFs, when we wanted to access
to the first 4.
This patch sets the default one to 4 and then modifies the width of
align16 instruction's DF sources when we translate the logical swizzle
to the physical one.
v2:
- Remove conditional (Curro).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
From IVB PRM, vol4, part3, "General Restrictions on Regioning
Parameters":
"If ExecSize = Width and HorzStride ≠ 0, VertStride must
be set to Width * HorzStride."
In next patch, we are going to modify the region parameter for
uniforms and vgrf. For uniforms that are the source of
DF align1 instructions, they will have <0, 4, 1> regioning and
the execsize for those instructions will be 4, so they will break
the regioning rule. This will be the same for VGRF sources where
we use the vstride == 0 exploit.
As we know we are not going to cross the GRF boundary with that
execsize and parameters (not even with the exploit), we just fix
the vstride here.
v2:
- Move is_align1_df() (Curro)
- Refactor exec_size == width calculation (Curro)
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Until now the spilling cost calculation was neglecting the amount of
data read from the register during the spilling cost calculation.
This caused it to make suboptimal decisions in some cases leading to
higher memory bandwidth usage than necessary.
Improves Unigine Heaven performance by ~4% on BDW, reversing an
unintended FPS regression from my previous commit
147e71242c with n=12 and statistical
significance 5%. In addition SynMark2 OglCSDof performance is
improved by an additional ~5% on SKL, and a Kerbal Space Program
apitrace around the Moho planet I can provide on request improves by
~20%.
Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Plamena Manolova <plamena.manolova@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This is what we use later on to compute the number of registers that
will actually get spilled to memory, so it's more likely to match
reality than the current open-coded approximation.
Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Plamena Manolova <plamena.manolova@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Curro pointed out that I should not just check for MACH, but use
the reads_accumulator_implicitly() helper, which would also prevent
the same bug with MAC and SADA2 (if we ever decide to use them).
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This shuffles constants down in the reverse of what the previous
patch does and applies some simpilifications that may be made
possible from doing so.
Shader-db results BDW:
total instructions in shared programs: 12980814 -> 12977822 (-0.02%)
instructions in affected programs: 281889 -> 278897 (-1.06%)
helped: 1231
HURT: 128
total cycles in shared programs: 246562852 -> 246567288 (0.00%)
cycles in affected programs: 11271524 -> 11275960 (0.04%)
helped: 1630
HURT: 1378
V2: mark float opts as inexact
Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
opt_register_coalesce() was optimizing sequences such as:
mul(8) acc0:D, attr18.xyyy:D, attr19.xyyy:D
mach(8) vgrf5.xy:D, attr18.xyyy:D, attr19.xyyy:D
mov(8) m4.zw:F, vgrf5.xxxy:F
into:
mul(8) acc0:D, attr18.xyyy:D, attr19.xyyy:D
mach(8) m4.zw:D, attr18.xxxy:D, attr19.xxxy:D
This doesn't work - if we're going to reswizzle MACH, we'd need to
reswizzle the MUL as well. Here, the MUL fills the accumulator's .zw
components with attr18.yy * attr19.yy. But the MACH instruction expects
.z to contain attr18.x * attr19.x. Bogus results ensue.
No change in shader-db on Haswell. Prevents regressions in Timothy's
patches to use enhanced layouts for varying packing (which rearrange
code just enough to trigger this pre-existing bug, but were fine
themselves).
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
In commit c35fa7a, we changed the "width" of DF source registers to 2,
which is conceptually fine. Unfortunately a VertStride of 2 is not
allowed by align16 instructions on IVB/BYT, and the regular VertStride
of 4 works fine in any case.
See generated_tests/spec/arb_gpu_shader_fp64/execution/built-in-functions/vs-round-double.shader_test
for example:
cmp.ge.f0(8) g18<1>DF g1<0>.xyxyDF -g8<2>DF { align16 1Q };
ERROR: In Align16 mode, only VertStride of 0 or 4 is allowed
cmp.ge.f0(8) g19<1>DF g1<0>.xyxyDF -g9<2>DF { align16 2N };
ERROR: In Align16 mode, only VertStride of 0 or 4 is allowed
v2:
- Add spec quote (Curro).
- Change the condition to only BRW_VERTICAL_STRIDE_2 (Curro)
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This is required for correctness in presence of multiple 4-wide flag
writes (e.g. 4-wide instructions with a conditional mod set) which
update a different portion of the same 8-bit flag subregister.
Right now we keep track of flag dataflow with 8-bit granularity and
consider flag writes to have killed any previous definition of the
same subregister even if the write was less than 8 channels wide,
which can cause live flag register updates to be dead
code-eliminated incorrectly.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
horiz_offset() shouldn't be doing anything for scalar registers,
because all channels of any SIMD instructions will end up reading or
writing the same component of the register, so shifting the register
offset would be wrong.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Re-implement in terms of is_uniform() for
simplicity. Pass argument by const reference. Clarify commit
message. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Otherwise for a pack_double_2x32_split opcode, we emit:
vec1 64 ssa_135 = pack_double_2x32_split ssa_133, ssa_134
mov(8) g5<1>UD g5<4>.xUD { align16 1Q compacted };
mov(8) g7<2>UD g5<4,4,1>UD { align1 1Q };
ERROR: When the destination spans two registers, the source must span two registers
(exceptions for scalar source and packed-word to packed-dword expansion)
mov(8) g8<2>UD g5.4<4,4,1>UD { align1 2N };
ERROR: The offset from the two source registers must be the same
mov(8) g5<1>UD g6<4>.xUD { align16 1Q compacted };
mov(8) g7.1<2>UD g5<4,4,1>UD { align1 1Q };
ERROR: When the destination spans two registers, the source must span two registers
(exceptions for scalar source and packed-word to packed-dword expansion)
mov(8) g8.1<2>UD g5.4<4,4,1>UD { align1 2N };
ERROR: The offset from the two source registers must be the same
The intention was to emit mov(4)s for the instructions that have ERROR
annotations.
See tests/spec/arb_gpu_shader_fp64/execution/vs-isinf-dvec.shader_test
for example.
v2 (Samuel):
- Instead of setting the exec size to a fixed value, don't double it
(Curro).
- Add PICK_{HIGH,LOW}_32BIT to the condition.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Trivial rebase changes. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Drop useless vec4_visitor dependencies. Demote to
static stand-alone function. Don't write unused components in the
result. Use vec4_builder interface for register allocation. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Take into account offset values less than a full register (32 bytes)
when getting the var from register.
This is required when dealing with an operation that writes half of the
register (like one d2x in IVB/BYT, which uses exec_size == 4).
v2:
- Take in account this offset < 32 in liveness analysis too (Curro)
v3:
- Change formula in var_from_reg() (Curro)
- Remove useless changes (Curro)
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
On IVB, DF instructions have lowered the SIMD width to 4 but the
exec_size will be later doubled. Fix the assert to avoid crashing in
this case.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Simplify assert. Except for the 'inst->group % 4
== 0' part the assertion was redundant with the previous assertion. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This way we can set the destination type as double to all these new opcodes,
avoiding any optimizer's confusion that was happening before.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Drop no_spill workaround originally needed due to
the bogus destination type of VEC4_OPCODE_FROM_DOUBLE. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
When doing a 64-bit to a smaller data type size conversion, the destination should
be aligned to 64-bits. Because of that, we need to gather the data after the
actual conversion.
Until now, these two operations were done by VEC4_OPCODE_FROM_DOUBLE but
now we split them explicitely in two different instructions:
VEC4_OPCODE_FROM_DOUBLE just do the conversion and
VEC4_OPCODE_PICK_LOW_32BIT will gather the data.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
In the generator we must generate slightly different code for
Ivybridge/Baytrail, because of the way the stride works in
this hardware.
v2:
- Use stride and don't need to fix dst (Curro)
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Keep the original type when dealing with null registers. Especially
because we do no want to introduce an implicit conversion between
types that could affect the conditional flags.
This affects especially when the original type is DF, and we are working
on Ivybridge/Baytrail.
v2 (Curro)
- Fix typo.
- Use retype() instead of applying the type directly.
- Remove unneeded retype.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
We need to split DF instructions in two on IVB/BYT as it needs an
execsize 8 to process 4 DF values (one GRF in total).
v2:
- Rename helper and make it static inline function (Matt).
- Fix indention and add braces (Matt).
v3:
- Don't edit IR instruction when doubling exec_size (Curro)
- Add comment into the code (Curro).
- Manage ARF registers like the others (Curro)
v4:
- Add get_exec_type() function and use it to calculate the execution
size.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Fix bogus 'type != BAD_FILE' check. Take
destination type as execution type where there is no valid source.
Assert-fail if the deduced execution type is byte. Clarify comment
in get_lowered_simd_width(). Move SIMD width workaround outside of
'if (...inst->size_written > REG_SIZE)' conditional block, since the
problem should be independent of whether the amount of data written
by the instruction is greater or lower than a GRF. Drop redundant
is_ivb_df definition. Drop bogus inst->exec_size < 8 check.
Simplify channel group assertion. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The hardware applies the same channel enable signals to both halves of
the compressed instruction which will be just wrong under non-uniform
control flow. Fix this by splitting those instructions to SIMD4.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
According to the IVB and HSW PRMs:
"2.When the destination requires two registers and the sources are
indirect, the sources must use 1x1 regioning mode."
So for DF instructions the execution size is not limited by the number
of address registers that are available, but by the EU decompression
logic not handling VxH indirect addressing correctly.
This patch limits the SIMD width to 4 in this case.
v2:
- Fix typo (Matt).
- Fix condition (Curro)
v3:
- Add spec quote (Curro)
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
When converting a DF to 32-bit conversions, we set dst stride to 2,
to fulfill alignment restrictions because the upper Dword of every
Qword will be written with undefined value.
But in IVB/BYT, this is not necessary, as each DF conversion already
writes 2, the first one the real value, and the second one a 0.
That is, IVB/BYT already set stride = 2 implicitly, so we must set it to
1 explicitly to avoid ending up with stride = 4.
v2:
- Fix typo (Matt)
v3:
- Fix stride in the destination's brw_reg, don't modity IR (Curro)
v4:
- Remove 'is_dst' argument of brw_reg_from_fs_reg() (Curro)
- Fix comment (Curro).
- Relax hstride assert (Curro)
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Minor spelling fixes. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
v2:
- Change the name to lower_conversions.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This reverts commit 7dccd38b40.
d2x pass fixes SEL instructions when there is a type conversion
by doing a SEL without type conversion and then convert the result.
This pass also takes into account the non-uniform control flow.
Then, 7dccd38b40 is not needed anymore.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Generalize it to lower any unsupported narrower conversion.
v2 (Curro):
- Add supports_type_conversion()
- Reuse existing intruction instead of cloning it.
- Generalize d2x to narrower and equal size conversions.
v3 (Curro):
- Make supports_type_conversion() const and improve it.
- Use foreach_block_and_inst to process added instructions.
- Simplify code.
- Add assert and improve comments.
- Remove redundant mov.
- Remove useless comment.
- Remove saturate == false assert and add support for saturation
when fixing the conversion.
- Add get_exec_type() function.
v4 (Curro):
- Use get_exec_type() function to get sources' type.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
On HSW+, scalar DF sources can be accessed using the normal <0,1,0>
region, but on IVB and BYT DF regions must be programmed in terms of
floats. A <0,2,1> region accomplishes this.
v2:
- Apply region <0,2,1> in brw_reg_from_fs_reg() (Curro).
v3:
- Added comment explaining the reason (Curro).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Then the SIMD lowering pass will get rid of any compressed instructions with scalar
source (whether force_writemask_all or not) and we avoid hitting the Gen7 region
decompression bug.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Suggested-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
In IVB and BYT, both regioning parameters and execution sizes are measured as
32-bits element size.
So when we have something like:
mov(8) g2<1>DF g3<4,4,1>DF
We are not actually moving 8 doubles (our intention), but 4 doubles.
We need to double the parameters to cope with this issue. However,
horizontal strides don't behave as they're supposed to on IVB
for DF regions, they will cause each 32-bit half of DF sources to be
strided individually, and doubling the value won't make any difference.
v2:
- Use devinfo directly (Matt).
- Use Baytrail instead of Valleview (Matt).
- Use IvyBridge instead of Ivy (Matt)
- Double the exec_size in code emission (Curro)
v3:
- Change hstride doubling by an assert and fix commit log (Curro).
- Substitute remaining compiler->devinfo by devinfo (Curro).
v4:
- Fix comment (Curro).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The execution data size is the biggest type size of any instruction
operand.
We will use it to know if the instruction deals with DF, because in Ivy
we need to double the execution size and regioning parameters.
v2:
- Fix typo in commit log (Matt)
- Use static inline function instead of fs_inst's method (Curro).
- Define the result as a constant (Curro).
- Fix indentation (Matt).
- Add braces to nested control flow (Matt).
v3 (Curro):
- Add get_exec_type() and other auxiliary functions and use them to
calculate its size.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
[ Francisco Jerez: Fix bogus 'type != BAD_FILE' check. Fix deduced
execution type for integer vector types. Take destination type as
execution type where there is no valid source. Assert-fail if the
deduced execution type is byte. Move into brw_ir_fs.h header for
consistency with the VEC4 back-end. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
On IVB/BYT, region parameters and execution size for DF are in terms of
32-bit elements, so they are doubled. For evaluating the validity of an
instruction, we halve them.
v2 (Sam):
- Add comments.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
4-wide DF operations where NibCtrl applies require and execsize of 8
in IvyBridge/BayTrail.
v2:
- Refactor NibCtrl printing (Matt)
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The individual branches of an if/else/endif construct will be executed
some unknown number of times between 0 and 1 relative to the parent
block. Use some factor in between as weight while approximating the
cost of spill/fill instructions within a conditional if-else branch.
This favors spilling registers used within conditional branches which
are likely to be executed less frequently than registers used at the
top level.
Improves the framerate of the SynMark2 OglCSDof benchmark by ~1.9x on
my SKL GT4e. Should have a comparable effect on other platforms. No
significant regressions.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We already provide a default LOD for textureQueryLevels and texture() on
non-fragment stages. However, there are more cases where one is needed
such as textureSize(gsampler2DMS*) in SPIR-V. Instead of trying to list
out all of the cases one at a time, just provide the default for all TXS
and TXL operations. This fixes a shader validation error in the new
Sascha deferredmultisampling demo which uses textureSize(gsampler2DMS).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100391
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Cc: "13.0 17.0" <mesa-stable@lists.freedesktop.org>
Technically those hw operations are only available on gen7, as gen8+
support the conversion on the MOV. But, when using the builder to
implement nir operations (example: nir_op_fquantize2f16), it is not
needed to do the gen check. This check is done later, on the final
emission at brw_F32TO16 (brw_eu_emit), choosing between the MOV or the
specific operation accordingly.
So in the middle, during optimization phases those hw operations can
be around for gen8+ too.
Without this patch, several (at least 95) vulkan-cts quantize tests
crashes when using INTEL_DEBUG=optimizer. For example:
dEQP-VK.spirv_assembly.instruction.graphics.opquantize.too_small_vert
v2: simplify the code using GEN_GE (Ilia Mirkin)
v3: tweak brw_instruction_name instead of changing opcode_descs
table, that is used for validation (Matt Turner)
Reviewed-by: Matt Turner <mattst88@gmail.com>
SEL can only convert between a few integer types, which we basically
never do.
Fixes fs/vs-double-uniform-array-direct-indirect-non-uniform-control-flow
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
We need to know if sample shading has been requested during shader
compilation since that affects the way fragment coordinates are
computed.
Notice that the semantics of fragment coordinates only depend on
whether sample shading has been requested, not on whether more
than one sample will actually be produced (that is,
minSampleShading and rasterizationSamples do not affect this
behavior).
Because this setting affects the code we generate for the shader, we also
need to include it in the WM prog key. Notice we don't need to alter the
OpenGL code because it doesn't ever use this behavior, so they key's
value is always false (the default).
Fixes:
dEQP-VK.glsl.builtin_var.fragcoord_msaa.*
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
We want to be able to check the progress of each pass and dump the NIR
for debugging purposes if it changed.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The NIR story on conversion opcodes is a mess. We've had way too many
of them, naming is inconsistent, and which ones have explicit sizes was
sort-of random. This commit re-organizes things and makes them all
consistent:
- All non-bool conversion opcodes now have the explicit size in the
destination and are named <src_type>2<dst_type><size>.
- Integer <-> integer conversion opcodes now only come in i2i and u2u
forms (i2u and u2i have been removed) since the only difference
between the different integer conversions is whether or not they
sign-extend when up-converting.
- Boolean conversion opcodes all have the explicit size on the bool and
are named <src_type>2<dst_type>.
Making things consistent also allows nir_type_conversion_op to be moved
to nir_opcodes.c and auto-generated using mako. This will make adding
int8, int16, and float16 versions much easier when the time comes.
Reviewed-by: Eric Anholt <eric@anholt.net>
compiler/brw_vec4_gs_visitor.cpp:744:39: error:
‘GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES’ was not declared in this scope
output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
Fixes: d0d4a5f43b ("i965: split EU defines to brw_eu_defines.h")
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
At the moment all the tests but test_eu_compact are actual C++ gtests.
To simplify things, we can move the gtest.la to the common TEST_LIBS.
As we're here, we can rename change the test extension [to .cpp] to
avoid using the confusing dummy.cpp.
Add a nice comment in the makefile for posterity.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Mostly a dummy git mv with a couple of noticable parts:
- With the earlier header cleanups, nothing in src/intel depends
files from src/mesa/drivers/dri/i965/
- Both Autoconf and Android builds are addressed. Thanks to Mauro and
Tapani for the fixups in the latter
- brw_util.[ch] is not really compiler specific, so it's moved to i965.
v2:
- move brw_eu_defines.h instead of brw_defines.h
- remove no-longer applicable includes
- add missing vulkan/ prefix in the Android build (thanks Tapani)
v3:
- don't list brw_defines.h in src/intel/Makefile.sources (Jason)
- rebase on top of the oa patches
[Emil Velikov: commit message, various small fixes througout]
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>