| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The address immediate field is only 9 bits and, since the value is in
bytes, the highest GRF we can point to with it is g15. This makes it
pretty close to useless for MOV_INDIRECT. There were already piles of
restrictions preventing us from using it prior to Broadwell, so let's get
rid of the gen8+ code path entirely.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97779
Cc: "12.0 13.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
(cherry picked from commit 2a4a86862c949055c71637429f6d5f2e725d07d8)
|
|
|
|
|
|
|
| |
Similar to brw_context(...), intel_texture_object(...), and so on.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arcero@collabora.com>
|
|
|
|
|
|
|
|
|
| |
dispatch case.
This avoids emitting a few extra instructions required to take the
dispatch mask into account when it's known to be tightly packed.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
On at least Sky Lake, ce0 does not contain the full story as far as enabled
channels goes. It is possible to have completely disabled channels where
the corresponding bits in ce0 are 1. In order to get the correct execution
mask, you have to mask off those channels which were disabled from the
beginning by taking the AND of ce0 with either sr0.2 or sr0.3 depending on
the shader stage. Failure to do so can result in FIND_LIVE_CHANNEL
returning a completely dead channel.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: Francisco Jerez <currojerez@riseup.net>
[ Francisco Jerez: Fix a couple of typos, add mask register type
assertion, clarify reason why ce0 can have bits set for disabled
channels, clarify that this may only be a problem when thread
dispatch doesn't pack channels tightly in the SIMD thread. Apply
same treatment to Align16 path. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
| |
This restriction seemed rather artificial... Removing it actually
simplifies things slightly.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The previous regs_written field can be recovered by rewriting each
rvalue reference of regs_written like 'x = i.regs_written' to 'x =
DIV_ROUND_UP(i.size_written, reg_unit)', and each lvalue reference
like 'i.regs_written = x' to 'i.size_written = x * reg_unit'.
For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible. I'll come
back later to clean up any ugliness introduced here.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The fs_reg::subreg_offset and ::offset fields are now redundant, the
sub-GRF offset can just be added to the single ::offset field
expressed in byte units. The current subreg_offset value can be
recovered by applying the following rule: Replace each rvalue
reference of subreg_offset like 'x = r.subreg_offset' with 'x =
r.offset % reg_unit', and each lvalue reference like 'r.subreg_offset
= x' with 'r.offset = ROUND_DOWN_TO(r.offset, reg_unit) + x'.
For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible. I'll come
back later to clean up any ugliness introduced here.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The fs_reg::offset field in byte units introduced in this patch is a
more straightforward alternative to the current register offset
representation split between fs_reg::reg_offset and ::subreg_offset.
The split representation makes it too easy to forget about one of the
offsets while dealing with the other, which has led to multiple
back-end bugs in the past. To make the matter worse the unit
reg_offset was expressed in was rather inconsistent, for uniforms it
would be expressed in either 4B or 16B units depending on the
back-end, and for most other things it would be expressed in 32B
units.
This encodes reg_offset as a new offset field expressed consistently
in byte units. Each rvalue reference of reg_offset in existing code
like 'x = r.reg_offset' is rewritten to 'x = r.offset / reg_unit', and
each lvalue reference like 'r.reg_offset = x' is rewritten to
'r.offset = r.offset % reg_unit + x * reg_unit'.
Because the change affects a lot of places and is rather non-trivial
to verify due to the inconsistent value of reg_unit, I've tried to
avoid making any additional changes other than applying the rewrite
rule above in order to keep the patch as simple as possible, sometimes
at the cost of introducing obvious stupidity (e.g. algebraic
expressions that could be simplified given some knowledge of the
context) -- I'll clean those up later on in a second pass.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Without this, we would pass over the instructions in the SIMD8 program
(which is located earlier in the buffer) when brw_set_uip_jip() is
called to handle the SIMD16 program.
The assertion about compacted control flow was bogus: halt, cont, break
cannot be compacted because they have both JIP and UIP. Instead, we
should never see a compacted instruction in this code at all.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
| |
This may have been the reason people ran into problems with
non-uniform HALT instructions and ended up using the inefficient
ANY16H/ANY8H predicates instead of ANY4H or NORMAL in order to prevent
non-uniform discard. The HALT instruction is able to handle
non-uniform execution masks just fine.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
| |
We no longer use this message. As far as I can tell, it's fairly
useless - the equivalent information is provided in the payload.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
| |
This uses one less instruction.
v2: Move emit_find_msb_using_lzd out of the visitor classes. Suggested
by Curro.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
| |
v2 (Matt):
- Take a DF source argument for the DIM instruction emission
in the visitors.
- Indentation.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
So far we only used instructions with this size in situations where we
did not operate per-channel and we wanted to ignore the execution mask,
but gen7 fp64 will need to emit code with a width of 4 that needs
normal execution masking.
v2:
- Modify the assert instead of deleting it (Curro)
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Pulling DF uniforms from pull constant buffer generates messages like:
send(4) g12<1>DF g12<0,1,0>F
sampler ld SIMD4x2 Surface = 1 Sampler = 0 mlen 1 rlen 1
which produces GPU hangs in Cherryview/Braswell:
"For 64-bit Align1 operation or multiplication of dwords in CHV,
source horizontal stride must be aligned to qword."
This seems to be documented in the Cherryview PRM, Volume 7, Page 843:
"When source or destination datatype is 64b or operation is integer
DWord multiply, regioning in Align1 must follow these rules:
1. Source and Destination horizontal stride must be aligned to the
same qword."
We should set the destination type to UD, D, or F so that
the register stride checker doesn't notice. The destination type of
send messages is basically irrelevant anyway.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95462
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Pulling DF inputs from the URB generates messages like:
send(8) g23<1>DF g1<8,8,1>UD
urb 3 SIMD8 read mlen 1 rlen 2 { align1 1Q };
which makes the simulator angry:
"For 64-bit Align1 operation or multiplication of dwords in CHV,
source horizontal stride must be aligned to qword."
This seems to be documented in the Cherryview PRM, Volume 7, Page 823:
"When source or destination datatype is 64b or operation is integer
DWord multiply, regioning in Align1 must follow these rules:
1. Source and Destination horizontal stride must be aligned to the
same qword."
Setting the source horizontal stride to QWord is insane, as it's the
message header containing 8 URB handles in a single 32-bit DWord.
Instead, we should whack the destination type to UD, D, or F so that
the register stride checker doesn't notice. The destination type of
send messages is basically irrelevant anyway.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95462
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
|
|
| |
These are completely ignored by the conversion to brw_reg, so they
better be zero.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
| |
This generalizes the current fs_inst::force_sechalf flag to allow
specifying channel enable groups other than 0 or 8. At some point it
will likely make sense to fix the vec4 generator to support arbitrary
execution groups and then move the definition of fs_inst::group into
backend_instruction (e.g. so we can do FP64 in the VEC4 back-end).
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Alternatively we could have extended the current semantics to 32-wide
mode by changing brw_broadcast() to emit multiple indexed MOV
instructions in the generator copying the selected value to all
destination registers, but it seemed rather silly to waste EU cycles
unnecessarily copying the exact same value 32 times in the GRF.
The vstride change in the Align16 path is required to avoid assertions
in validate_reg() since the change causes the execution size of the
MOV and SEL instructions to be equal to the source region width.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
| |
The hardware has messages that can write 32 32bit components at once
but the channel enable mask gets messed up. We need to split them
into several 16-wide scratch writes for the channel enables to be
applied correctly. The SIMD lowering pass cannot be used for this
because scratch writes are emitted rather late during register
allocation long after SIMD lowering has been done.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
| |
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
Most of these are bugs because the intended execution size of an
instruction and the dispatch width of the shader aren't necessarily
the same (especially in SIMD32 programs).
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
| |
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
| |
Trivial clean-up.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
| |
This gets IF and DO instructions working in SIMD32 programs. brw_IF()
and brw_DO() should probably behave in the same way as other generator
functions that emit control flow instructions and just figure out the
right execution size by themselves from the current execution controls
specified through the brw_codegen argument. Changing that will
require updating lots of Gen4-5 clipper code though, so for the moment
just pass the current value redundantly from the FS generator.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Instead of just halving the execution size when the instruction is
compressed hoping that it will give a legal source region width, we
can calculate the maximum legal width value in closed form from the
component size and stride. This makes sure that brw_reg_from_fs_reg()
always returns a valid hardware region even for virtual 32-wide
instructions (e.g. send-like instructions) that would seem to exceed
the hardware region width limit after halving.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
| |
Curro is planning to eliminate p->compressed, so let's avoid using it
here and just pass in the value directly.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
[ Francisco Jerez: Pass boolean flag instead of brw_compression enum. ]
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
By using the new compression/group control interface. This will allow
easier extension to support arbitrary channel enable groups at the IR
level.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
| |
The right value is dependent on the specific IR instruction being
generated so it has to be reset in every iteration of the loop anyway.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
| |
Most of these were resetting quarter control to zero incorrectly even
though everything they needed to do was disable instruction
compression -- The brw_SAMPLE() case was doing the right thing but it
can be simplified slightly by using the new compression control
interface.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
| |
It's just a byte MOV with strided source.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
| |
These can be easily represented in the IR as a MOV instruction with
strided source so they seem rather redundant.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
| |
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
| |
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
| |
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
| |
generator.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
generator.
Currently the generator code for most opcodes honours the default
access mode (which should typically be Align1 in the scalar back-end),
but generate_code() doesn't set it explicitly which means that the
access mode from a previous instruction could leak into the following
ones if you did something special and weren't careful enough to save
and restore the previous access mode.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
Most of this wouldn't have worked for SIMD32 and had various
dispatch_width and compression control bugs. It's mostly dead now
with SIMD lowering of math instructions turned on in the compiler.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
| |
For consistency with the Gen7 variant. I'm not doing the same to the
uniform pull constant message at this point because the non-GEN7 one
is still overloaded to be either an expression-like logical
instruction or a Gen4-specific physical send message.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
This will allow the SIMD lowering pass to split 32-wide varying pull
constant loads (not natively supported by the hardware) into 16-wide
instructions.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
| |
This lets the rest of the backend know that the uniform pull constant
load opcodes don't respect channel enables -- Without this the
register allocator has no way to know that the return payload of a
pull constant load is not per-channel and spills of the destination
will be broken under non-uniform control flow.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
| |
Rob's nir_lower_wpos_ytransform() pass flips dFdy in the opposite case
of what I expected, so we always take the negate_value case. It doesn't
really matter.
v2: Write src0 before src1 in ADD instructions (requested by Matt).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This handles gl_FragCoord transformations and other window system vs.
user FBO coordinate system flipping by multiplying/adding uniform
values, rather than recompiles.
This is much better because we have no decent way to guess whether
the application is going to use a shader with the window system FBO
or a user FBO, much less the drawable height. This led to a lot of
recompiles in many applications.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In a5d7e144eaf43fee37e6ff9e2de194407087632b, Connor generalized the
exec_size halving code to handle more cases. As part of this, he made
it not halve anything if the region accessed falls completely in a
single register.
Unfortunately, it started producing some invalid regions:
-add(16) g6<1>F g10<8,8,1>UW -g1<0,1,0>F { align1 compr };
-add(16) g8<1>F g12<8,8,1>UW -g1.1<0,1,0>F { align1 compr };
+add(16) g6<1>F g10<16,16,1>UW -g1<0,1,0>F { align1 compr };
+add(16) g8<1>F g12<16,16,1>UW -g1.1<0,1,0>F { align1 compr };
Here, the UW source region completely fits within a register. However,
we have to use instruction compression because the destination region
spans two registers. <16,16,1> is invalid because it's compressed.
To handle this, skip the "everything fits in one register" case and
fall through to the exec_size halving case when compressed.
Fixes hundreds of Piglit regressions on GM965.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95370
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
| |
brw_reg_from_fs_reg() needs to know whether the instruction will be
compressed or not.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95370
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Commit 5310bca024f77da40ea6f4c275455f9cb0528f9e added a new "double df"
field to the brw_reg struct, adding an extra 4 bytes of data that isn't
usually initialized (or may contain irrelevant garbage if the struct is
mutated). This means that it's no longer safe to memcmp().
Instead, add a brw_regs_equal() function which ignores the extra df bits
unless they matter. To keep the implementation cheap, we wrap the first
set of fields in a union/struct so that we can use a single DWord
comparison.
v2: Drop unnecessary casts (caught by Francisco Jerez).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The HW has a restriction that only vertical stride may cross register
boundaries. Previously, this only mattered for SIMD16 instructions where
we needed to use the same regioning parameters as the equivalent SIMD8
instruction but double the exec size. But we need to do the same
splitting for 64-bit instructions as well as instructions with a stride
of 2 (which effectively consume 64 bits per element). Fix up the code to
do the right thing instead of special-casing SIMD16.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|