| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
C.f. 'i965/fs: Print fs_reg::offset field consistently for all
register files.'.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
The offset printing code in fs_visitor::dump_instruction() was doing
things differently for sources and destinations and for each register
file -- In some cases it would be added to the base register number
fs_reg::nr, in other cases it would follow the base register separated
with a plus sign, in other cases (uniforms) it would do both (!). The
sub-register offset was also being printed or not rather
inconsistently. Fix it.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
Get rid of some leftover redundant arithmetic introduced during the
conversion to byte offsets and sizes that can be simplified easily.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
component() was generally a better alternative because of several
issues set_smear() had:
- It wouldn't take the original stride and offset of the register
into account, which means that set_smear() on the result of
e.g. another set_smear() call or an offset() call would give a
bogus region as result.
- It was an inherently destructive operation. See the
'nir_intrinsic_shader_clock' hunk below for how this could lead to
subtle bugs in cases where set_smear() was called multiple times on
the same register like 'r.set_smear(0), r.set_smear(1)' with the
expectation that each call would return a separate value instead of
a reference to the same subsequently mutated object.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
Also changed the argument names since 'src' and 'dst' don't make that
much sense outside of the context of copy propagation.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This makes the function less annoying to use and more accurate -- We
shouldn't propagate a copy into a register region that wasn't fully
contained in the destination of the copy (IOW, a source region that
wasn't fully defined by the copy) just because the number of registers
written and read by each instruction happened to get rounded up to the
same GRF multiple.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
| |
By keeping track of 'offset' in byte units.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
| |
component_size().
Using component_size() is easier and generally more correct because it
takes into account the register type and stride for you.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
| |
No need to unroll the first iteration of the loop manually.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These were bashing the 'offset' and 'stride' values of several
registers without taking the previous value into account, which
probably didn't matter in practice for optimize_frontfacing_ternary()
because the 'tmp' register already had a known region, but it would
have given the wrong region as result in the other cases in
lower_integer_multiplication(). subscript(..., i) is a more
straightforward way to take the i-th field of a given type from each
channel of a register which should give the right answer as result
regardless of the original 'offset' and 'stride' parameters of the
register region.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
| |
of rounding.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
| |
In the most common case this can now be implemented as a simple
addition because the offset is already encoded as a single scalar
value in bytes.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
This makes the helper function less annoying to use and somewhat more
accurate.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
The 'scan_inst->dst.offset % REG_SIZE' term in the final
'scan_inst->dst.offset' calculation is obviously bogus. The offset
from the start of the copy destination register 'inst->dst' where the
destination of the generating instruction 'scan_inst' would be written
to (before compute-to-mrf runs) is just the offset of 'scan_inst->dst'
relative to the source of the copy instruction (AKA rel_offset in the
code below).
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This was dropping 'inst->dst.offset' on the floor. Nothing in the
code above seems to guarantee that it's zero and in that case the
offset of the register being coalesced into wouldn't be taken into
account while rewriting the generating instruction.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
This makes sure that overlap checks are done correctly throughout the
back-end when the '*this' register starts before the register/size
pair provided as argument, and is actually less annoying to use than
in_range() at this point since regions_overlap() takes its size
arguments in bytes.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
This is copy-pasted almost line by line from the FS back-end. The
only reason it cannot be implemented in terms of backend_reg is that
the backend_reg::nr field doesn't have the same meaning for uniforms
on both back-ends. It could be easily deduplicated by using a
template function.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
Its only use left in the FS back-end should be using regions_overlap()
instead to avoid getting a false negative result in cases where source
and destination overlap but the former starts before the latter in the
VGRF file.
v2: Put back lost components factor (Iago).
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
fs_inst::overwrites_reg is rather easy to misuse because it cannot
tell how large the register region starting at 'reg' is, so in cases
where the destination region starts after 'reg' it may give a
misleading result. regions_overlap() is somewhat more verbose to use
but handles arbitrary overlap correctly so it should generally be used
instead.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
is_nop_mov() was broken for LOAD_PAYLOAD instructions in two ways: On
the one hand the original destination register offset wasn't being
taken into account which would give incorrect results if it was
already non-zero, and on the other hand all source registers were
being treated as if they had a size of 32B, which is almost never the
case in SIMD16 programs for non-header sources.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The previous overlap condition only made sure that the VGRF numbers or
GRF-aligned offsets were different without taking the amount of data
written and read by the instruction into consideration. Use the
regions_overlap() helper instead.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
This could potentially have misoptimized a program in cases where
inst->src[0] had a non-zero sub-GRF offset.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
| |
raw copy.
Noticed the problem by inspection while typing in the previous commit.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
| |
raw copy.
This was likely the original intention, and at least register coalesce
relies on it.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
| |
Unlike the FS counterpart of this commit this was likely not (yet) a
bug, but let's fix it already in preparation for implementing support
for sub-GRF offsets in the VEC4 back-end.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
There was a workaround for this in fs_inst::size_read() for the
SHADER_OPCODE_MOV_INDIRECT instruction and FIXED_GRF register file
*only*. We should take this possibility into account for the sources
and destinations of all instructions on all optimization passes that
need to quantize dataflow in 32B increments by adding the amount of
misalignment to the size read or written from the regs_read() and
regs_written() helpers respectively.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This fixes regs_written() and regs_read() to return a more accurate
value when the padding left between components due to a stride value
greater than one causes the region bounds given by size_written or
size_read to overflow into the next register. This could become a
problem in optimization passes that keep track of dataflow using
fixed-size arrays with register granularity, because the overflow
register (not actually accessed by the region) may not have been
allocated at all which could lead to undefined memory access.
An alternative to this would be to subtract the trailing padding
already during the calculation of fs_inst::size_read and
::size_written, but that would break code that currently assumes that
::size_read and _written are whole multiples of the component size,
and would be hard to maintain looking forward because size_written is
assigned from a bunch of different places.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
This will be useful later on when we start using reg_offset() on fixed
hardware registers.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
| |
This restriction seemed rather artificial... Removing it actually
simplifies things slightly.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
| |
The LINTERP virtual instruction only reads three scalar components
from the first 16B of the second source, we can now teach size_read()
about it since its return value is represented with byte granularity.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
| |
UNIFORM files.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
units.
The previous regs_read value can be recovered by rewriting each
reference of regs_read() like 'x = i.regs_read(j)' to 'x =
DIV_ROUND_UP(i.size_read(j), 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 <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The previous regs_read value can be recovered by rewriting each
reference of regs_read() like 'x = i.regs_read(j)' to 'x =
DIV_ROUND_UP(i.size_read(j), 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 <[email protected]>
|
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
in bytes.
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 <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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 <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
::regs_written.
This is in preparation for dropping vec4_instruction::regs_read and
::regs_written in favor of more accurate alternatives expressed in
byte units. The main reason these wrappers are useful is that a
number of optimization passes implement dataflow analysis with
register granularity, so these helpers will come in handy once we've
switched register offsets and sizes to the byte representation. The
wrapper functions will also make sure that GRF misalignment (currently
neglected by most of the back-end) is taken into account correctly in
the calculation of regs_read and regs_written.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is in preparation for dropping fs_inst::regs_read and
::regs_written in favor of more accurate alternatives expressed in
byte units. The main reason these wrappers are useful is that a
number of optimization passes implement dataflow analysis with
register granularity, so these helpers will come in handy once we've
switched register offsets and sizes to the byte representation. The
wrapper functions will also make sure that GRF misalignment (currently
neglected by most of the back-end) is taken into account correctly in
the calculation of regs_read and regs_written.
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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 <[email protected]>
|
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
expressed in bytes.
The dst/src_reg::offset field in byte units introduced in the previous
patch is a more straightforward alternative to an offset
representation split between ::reg_offset and ::subreg_offset fields.
The split representation makes it too easy to forget about one of the
offsets while dealing with the other, which has led to multiple FS
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.
v2: Fix division by the wrong reg_unit in the UNIFORM case of
convert_to_hw_regs(). (Iago)
Reviewed-by: Iago Toral Quiroga <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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 <[email protected]>
|
|
|
|
|
|
| |
Signed-off-by: Eero Tamminen <[email protected]>
Reviewed-by: Eric Engestrom <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
AEP requires ASTC, which is currently only enabled on Skylake and later.
(It may be possible to extend this to Cherryview/Braswell in the future,
but earlier hardware doesn't have ASTC support.)
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]>
|
|
|
|
|
| |
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>
|