opt_sampler_eot() was relying on the default builder to have the same
width as the sampler and FB write opcodes it was eliminating, the
channel selects didn't matter because the builder was only being used
to allocate registers, no new instructions were being emitted with it.
A future commit will change the width of the default builder what will
break this assumption, so initialize it explicitly here.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This wasn't taking into account the execution controls of the original
instruction, but it was most likely not a bug because control flow
instructions are typically full width.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Emit the SELs and MOVs with the same execution controls as the
original MOVs, and the CMP with the same execution controls as the IF.
Also explicitly check that the execution controls of any pair of MOVs
being folded into a SEL are compatible (which is almost always going
to be the case), since otherwise it would seem wrong to initialize the
builder object below from the then_mov instruction only.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
lower_integer_multiplication() was ignoring the execution controls of
the original MUL instruction. Fix it by using the new fs_builder
constructor.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
demote_pull_constants() was ignoring the execution size and channel
selects of the instruction that wanted the constant, which doesn't
matter for uniform pull constant loads because all channels get the
same scalar value, but it might for varying pull constant loads. Fix
it by using the new fs_builder() constructor that takes care of
setting execution controls compatible with the instruction passed as
argument.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The execution size was being left equal to the default of 8/16, which
AFAICT would have overwritten components other than the one we wanted
to initialize and could potentially have corrupted other registers.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We have a number of optimization passes that repeat the same pattern
before inserting new instructions into the program based on some
previous instruction: They point the default builder at the original
instruction, then call exec_all() and group() to select the same
execution controls the original instruction had, and then maybe call
annotate() to clone the debug annotation from the original
instruction.
In fact an optimization pass missing any of these steps is likely to
be broken if the intention was to emit new code based on a preexisting
instruction, so let's make it easy for passes to do the right thing by
having an fs_builder constructor that automates the task of setting up
a builder to emit a given instruction provided as argument.
The following patches fix all cases I've found in which we weren't
explicitly initializing the execution controls of the emitted
instructions, and clean-up optimization passes which were already
doing the right thing to use the new constructor.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Images take up zero uniform slots in the nir_shader::num_uniforms
calculation, but nir_setup_uniforms needs to be executed even if the
program has no non-image uniforms so the driver-specific image
parameters are uploaded. nir_setup_uniforms is a no-op if there are
really no uniforms, so checking the num_uniform count is useless in
any case.
The nir_setup_inputs and _outputs changes shouldn't lead to any
functional change, they are just meant to preserve the symmetry
between them and nir_setup_uniforms.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Image variables need to allocate additional uniform slots over
nir_shader::num_uniforms. nir_setup_uniforms() overwrites the values
imported from the SIMD8 visitor and then exits early before entering
the nir_shader::uniforms loop, so image uniforms are never re-created.
Instead leave the imported values alone, they *must* be the same for
the uniform layout of both runs to be compatible.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Rewrite the NIR atomic counter intrinsics translation code making use
of the recently introduced surface builder. This will allow the
removal of some of the functionality duplicated between the visitor
and surface builder.
v2: Drop VEC4 suport.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Implement helper functions that can be used to construct and send
untyped and typed surface read, write and atomic messages to the
shared dataport unit easily.
v2: Drop VEC4 suport.
v3: Reimplement in terms of logical send opcodes.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This will be handy to avoid some ugly ternary operators in the next
patch, like:
fs_reg reg = (size == 0 ? null_reg_ud() : vgrf(..., size));
Because a zero-size register allocation is guaranteed not to ever be
read or written we can just return the null register. Another
possibility would be to actually allocate a zero-size VGRF what would
involve defining a zero-size register class in the register allocator
and a considerable amount of churn.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Each logical variant is largely equivalent to the original opcode but
instead of taking a single payload source it expects its arguments
separately as individual sources, like:
typed_surface_write_logical null, coordinates, source, surface,
num_coordinates, num_components
This patch defines the opcodes and usual instruction boilerplate,
including a placeholder lowering function provided mainly as
documentation for their source registers.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This cleans up the VEC4 implementation of setup_uniform_values()
somewhat and will avoid duplication of the image uniform upload code
by having a common interface to upload a vector of uniforms on either
back-end.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This should match the set of cases in which we currently call fail()
or no16() from the emit_texture_*() methods and the ones in which
emit_texture_gen4() enables the SIMD16 workaround.
Hint for reviewers: It's not a big deal if I happen to have missed
some case here, it will just lead to an assertion failure down the
road which is easily fixable, however being stricter than necessary
won't cause any visible breakage, it would just decrease performance
silently due to the unnecessary message splitting, so feel free to
double-check that all cases listed here already cause a SIMD8/16
fall-back with the current texturing code -- You may want to skip over
the Gen5-6 cases though if you don't have pencil and paper at hand.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Unlike its Gen5 and Gen7 counterparts this patch isn't a plain
refactor of the previous Gen4 texturing code, it's more of a rewrite
largely based on emit_texture_gen4_simd16(). The reason is that on
the one hand the original emit_texture_gen4() code didn't seem easily
fixable to be SIMD width-invariant and had plenty of clutter to
support SIMD-width workarounds which are no longer required. On the
other hand emit_texture_gen4_simd16() was missing a number of
SIMD8-only opcodes. This should generalize both and roughly match
their current behaviour where there is overlap.
Incidentally this will fix the following piglits on Gen4:
arb_shader_texture_lod.execution.arb_shader_texture_lod-texgrad
arb_shader_texture_lod.execution.tex-miplevel-selection *gradarb 2d
arb_shader_texture_lod.execution.tex-miplevel-selection *gradarb 3d
arb_shader_texture_lod.execution.tex-miplevel-selection *projgradarb 2d
arb_shader_texture_lod.execution.tex-miplevel-selection *projgradarb 2d_projvec4
arb_shader_texture_lod.execution.tex-miplevel-selection *projgradarb 3d
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This should be largely equivalent to emit_texture_gen5() except for
slight codestyle changes and the use i965 opcodes instead of the
ir_texture_opcode enum, see "i965/fs: Implement lowering of logical
texturing opcodes on Gen7+." for the mapping between them.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
These weren't being handled by emit_texture_gen7() but we can easily
lower them here for consistency with other texturing opcodes.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This should be largely equivalent to emit_texture_gen7() except that
we now get i965 sampling opcodes directly rather than
ir_texture_opcode enum values. The mapping is as follows:
- ir_tex -> SHADER_OPCODE_TEX
- ir_txb -> FS_OPCODE_TXB
- ir_txl -> SHADER_OPCODE_TXL
- ir_txd -> SHADER_OPCODE_TXD
- ir_txf -> SHADER_OPCODE_TXF
- ir_txf_ms -> SHADER_OPCODE_TXF_CMS
- ir_txs -> SHADER_OPCODE_TXS
- ir_query_levels -> SHADER_OPCODE_TXS too, the visitor will make
sure that the provided lod value is zero in this
case.
- ir_lod -> SHADER_OPCODE_LOD
- ir_tg4 -> SHADER_OPCODE_TG4_OFFSET if the offset value is not
immediate, SHADER_OPCODE_TG4 otherwise.
Other than that there are only minor changes and style fixes like the
implementation now being factored out in static functions to improve
encapsulation.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This hasn't been overallocating space for the header for a long time.
It still leaves the header uninitialized though until the generator
fixes it.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
So that it's left uninitialized by LOAD_PAYLOAD, we only need to
reserve space for it in the message since it will be initialized
implicitly by the generator.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
dispatch_width is global for a single compilation and doesn't
necessarily match the desired execution width if we had to lower the
original full-width instruction due to hardware limitations. These
were all inside a Gen4-specific branch so this patch shouldn't have
any effect on more recent hardware.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Each logical variant is largely equivalent to the original opcode but
instead of taking a single payload source it expects the arguments
separately as individual sources, like:
tex_logical dst, coordinates, shadow_c, lod, lod2,
sample_index, mcs, sampler, offset,
num_coordinate_components, num_grad_components
This patch defines the opcodes and usual instruction boilerplate,
including a placeholder lowering function provided mostly as
documentation for their source registers.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The only non-trivial thing it still has to do is figure out where to
take the src/dst depth values from and predicate the instruction if
discard is in use. The manual SIMD unrolling logic in the dual-source
case goes away because this is now handled transparently by the SIMD
lowering pass.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
This does essentially the same thing as
fs_visitor::emit_single_fb_write(), with some slight differences:
- We don't have to worry about exec_size and use_2nd_half anymore,
16-wide sources have already been lowered to 8-wide thanks to the
previous commit and the manual argument unzipping is no longer
required.
- The src/dst_depth and sample_mask values are now explicit sources
of the instruction instead of being taken from the visitor state
directly. The same goes for the kill-pixel mask that will be
passed to the instruction explicitly as predicate.
- Everything is now done in static functions to improve
encapsulation.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
This shouldn't have any effect because we don't emit logical
framebuffer writes yet.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
There's no need to initialize the wrong half of oMask in the payload
when we're doing an 8-wide framebuffer write because it will be
ignored by the hardware anyway. By doing it this way we can let the
SIMD lowering pass split the sample_mask source as a regular
per-channel source, otherwise we would have to introduce some sort of
per-instruction source query or use fs_inst::header_size for the
lowering pass to be able to find out whether some source is
header-like, and leave the source untouched in that case.
As a bonus this achieves the same purpose as the previous code without
making use of the SET_OMASK pseudo-instruction, which will be removed
in a future commit.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Flatten the if ladder to match the way that the ordering of these
fields is specified in the hardware documentation a bit more closely.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In cases where the color0 argument wasn't being provided,
emit_single_fb_writes() would take the alpha channel directly from the
visitor state instead of taking it from its arguments. This sort of
hack didn't fit nicely into the logical send-message approach because
all parameters of the instruction have to be visible to the SIMD
lowering pass for it to be able to split them into halves at all.
Fix it by using LOAD_PAYLOAD in fs_visitor::emit_fb_writes() to
provide an actual color0 vector with undefined contents except for the
alpha component to match the previous behavior when no color buffers
are enabled.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It's surprising that we weren't checking for this already. A future
patch will cause code like the following to be emitted:
MOV(16) tmp<1>:uw, src
MOV(8) dst<1>:ud, tmp<8,8,1>:ud
The second MOV comes from the expansion of a LOAD_PAYLOAD header copy,
so I don't have control over its types. Copy propagation will happily
turn this into:
MOV(8) dst<1>:ud, src
Which has different semantics. Fix it by preventing propagation in
cases where a single channel of the instruction would span several
channels of the copy (this requirement could in fact be relaxed if the
copy is just a trivial memcpy, but this case is unusual enough that I
don't think it matters in practice).
I'm deliberately only checking if the type of the instruction is
larger than the original, because the converse case seems to be
handled correctly already in the code below.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We were previously guessing the half based on the EOT flag which seems
rather gross.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The logical variant is largely equivalent to the original opcode but
instead of taking a single payload source it expects its arguments
that make up the payload separately as individual sources, like:
fb_write_logical null, color0, color1, src0_alpha,
src_depth, dst_depth, sample_mask, num_components
This patch defines the opcode and usual instruction boilerplate,
including a placeholder lowering function provided mainly as
self-documentation.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This lowering pass implements an algorithm to expand SIMDN
instructions into a sequence of SIMDM instructions in cases where the
hardware doesn't support the original execution size natively for some
particular instruction. The most important use-cases are:
- Lowering send message instructions that don't support SIMD16
natively into SIMD8 (several texturing, framebuffer write and typed
surface operations).
- Lowering messages that don't support SIMD8 natively into SIMD16
(*cough*gen4*cough*).
- 64-bit precision operations (e.g. FP64 and 64-bit integer
multiplication).
- SIMD32.
The algorithm works by splitting the sources of the original
instruction into chunks of width appropriate for the lowered
instructions, and then interleaving the results component-wise into
the destination of the original instruction. The pass is controlled
by the get_lowered_simd_width() function that currently just returns
the original execution size making the whole pass a no-op for the
moment until some user is introduced.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
v2: Reverse order of the source transformations and split_inst emit
call to make the code a bit easier to understand.
Typically BAD_FILE sources are used to mark a source as not present
what implies that no registers are read. This will become much more
frequent with logical send opcodes which have a large number of
sources, many of them optionally used and marked as BAD_FILE when they
aren't applicable. It will prove to be useful to be able to rely on
the value of regs_read() regardless of whether a source is present or
not.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
And start using it in fs_builder::LOAD_PAYLOAD(). This will be used
to emit logical send message opcodes which have an unusually large
number of arguments.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This pass will house ad-hoc lowering code for several send
message-like virtual opcodes that will represent their logically
independent arguments as separate instruction sources rather than as a
single payload blob. This pass will basically just take the separate
arguments that are supposed to be part of the payload and concatenate
them to construct a message in the form required by the hardware.
Virtual instructions in separate-source form will eventually allow
some simplification of the visitor code and make several
transformations easier like lowering SIMD16 instructions to SIMD8
algorithmically in cases where the hardware doesn't support the former
natively.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This cleans up fs_inst::regs_read() slightly by disentangling the
calculation of "components" from the handling of message payload
arguments. This will also simplify the SIMD lowering and logical send
message lowering passes, because it will avoid expressions like
'regs_read * REG_SIZE / component_size' which are not only ugly, they
may be inaccurate because regs_read rounds up the result to the
closest register multiple so they could give incorrect results when
the component size is lower than one register (e.g. uniforms). This
didn't seem to be a problem right now because all such expressions
happen to be dealing with per-channel GRFs only currently, but that's
by no means obvious so better be safe than sorry.
v2: Split PIXEL_X/Y and LINTERP into separate case blocks.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
