| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
| |
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>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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>
|
|
|
|
|
|
|
|
|
|
|
| |
I do appreciate the cleverness, but unfortunately it prevents a lot more
cleverness in the form of additional compiler optimizations brought on
by -fstrict-aliasing.
No difference in OglBatch7 (n=20).
Co-authored-by: Davin McCall <davmac@davmac.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.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>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reverts commit c1107cec44ab030c7fcc97c67baa12df1cc9d7b5.
Apparently the hardware spec text I quoted in the commit message was
outright lying about scalar source math being supported on SNB, the
hardware seems to load 32 contiguous bits of data for each channel
regardless of the regioning mode. Fixes regressions in the following
CTS tests (which we didn't catch early due to CTS being temporarily
disabled in our CI system):
es2-cts.gtf.gl.atan.atan_vec3_frag_xvary
es2-cts.gtf.gl.cos.cos_vec2_frag_xvary
es2-cts.gtf.gl.atan.atan_vec2_frag_xvary
es2-cts.gtf.gl.pow.pow_vec2_frag_xvary_yconsthalf
es2-cts.gtf.gl.cos.cos_float_frag_xvary
es2-cts.gtf.gl.pow.pow_float_frag_xvary_yconsthalf
es2-cts.gtf.gl.atan.atan_vec3_frag_xvaryyvary
es2-cts.gtf.gl.pow.pow_vec3_frag_xvary_yconsthalf
es2-cts.gtf.gl.cos.cos_vec3_frag_xvary
es2-cts.gtf.gl.atan.atan_vec2_frag_xvaryyvary
Cc: mesa-stable@lists.freedesktop.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96346
Reported-by: Mark Janes <mark.a.janes@intel.com>
Acked-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I haven't found any evidence that this isn't supported by the
hardware, in fact according to the SNB hardware spec:
"The supported regioning modes for math instructions are align16,
align1 with the following restrictions:
- Scalar source is supported.
[...]
- Source and destination offset must be the same, except the case of
scalar source."
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
| |
This doesn't actually handle the FS case, just add an assertion for
the moment so I don't forget to update it later on for SIMD32 fragment
shader dispatch.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
brw_null_vec() cannot handle widths over 16 but it doesn't really
matter what width we specify for null registers because destination
regions have no width field at the hardware level.
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>
|
|
|
|
|
|
|
|
| |
The benefit is we will be able to use the SIMD lowering pass to unroll
math instructions of unsupported width and then remove some cruft from
the generator.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
| |
Probably not needed since we fix the dst type of comparisons
automatically, but for consistency with the rest of null_reg_*
functions.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
| |
v2: Account for the stride of the dst (Iago)
Signed-off-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
| |
And use it in brw_fs_nir.cpp.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Gen4/5's SEL instruction cannot use conditional modifiers, so min/max
are implemented as CMP + SEL. Handling that after optimization lets us
CSE more.
On Ironlake:
total instructions in shared programs: 6426035 -> 6422753 (-0.05%)
instructions in affected programs: 326604 -> 323322 (-1.00%)
helped: 1411
total cycles in shared programs: 129184700 -> 129101586 (-0.06%)
cycles in affected programs: 18950290 -> 18867176 (-0.44%)
helped: 2419
HURT: 328
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Cuts 10k of .text, of which only 776 bytes are the fs_reg constructor
implementations themselves.
text data bss dec hex filename
5204535 214112 27784 5446431 531b1f i965_dri.so before
5193977 214112 27784 5435873 52f1e1 i965_dri.so after
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
| |
The 2-bit hardware register file field is ARF, GRF, MRF, IMM.
Rename GRF to VGRF (virtual GRF) so that we can reuse the GRF name to
mean an assigned general purpose register.
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since the types of the expression were
bool ? src_reg : (bool ? brw_reg : brw_reg)
the result of the second (nested) ternary would be implicitly
converted to a src_reg by the src_reg(struct brw_reg) constructor. I.e.,
bool ? src_reg : src_reg(bool ? brw_reg : brw_reg)
In the next patch, I make backend_reg (the parent of src_reg) inherit
from brw_reg, which changes this expression to return brw_reg, which
throws away any fields that exist in the classes derived from brw_reg.
I.e.,
src_reg(bool ? brw_reg(src_reg) : bool ? brw_reg : brw_reg)
Generally this code was gross, and wasn't actually shorter or easier to
read than an if ladder.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
|
|
|
|
|
|
|
|
|
| |
The scalar destination registers break copy propagation. Instead compute
the results to a regular register and then reference a component when we
later use the result as a source.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
|
|
|
|
|
|
|
|
|
|
| |
The docs specifically call out SEL with .l and .ge as the
implementations of MIN and MAX respectively. Among other things,
SEL with these conditional mods are commutative.
See commit 3b7f683f.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
|
|
|
|
|
|
| |
Not a typo. Replace the default builder with one of bogus width to
catch cases in which optimization passes assume that the default
dispatch width is good enough. The execution controls of instructions
emitted during optimization should in general match the original code
that is being manipulated. Many of the problems fixed in this series
were caught by the assertions introduced in this patch.
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>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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>
|
|
|
|
|
|
|
|
| |
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 is essentially the same problem fixed in an earlier patch for
immediates. Setting the stride to zero will be particularly useful
for my future SIMD lowering pass, because we will be able to just
check whether the stride of a source register is zero and skip
emitting the copies required to unzip it in that case.
Instead of setting stride to zero in every caller of emit_uniformize()
I've changed the function to return the result as its return value
(previously it was being written into a caller-provided destination
register), because this way we can enforce that the result is used with
the correct regioning from the function itself.
The changes to the prototype of its VEC4 counterpart are mainly for
the sake of symmetry, VEC4 registers don't have stride.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
is on.
This assertion was meant to catch code inadvertently escaping the
control flow jail determined by the group of channel enable signals
selected by some caller, however it seems useful to be able to
increase the default execution size as long as force_writemask_all is
enabled, because force_writemask_all is an explicit indication that
there is no longer a one-to-one correspondence between channels and
SIMD components so the restriction doesn't apply.
In addition reorder the calls to fs_builder::group and ::exec_all in a
couple of places to make sure that we don't temporarily break this
invariant in the future for instructions with exec_size higher than
the dispatch width.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As of now, the width field is no longer used for anything. The width field
"seemed like a good idea at the time" but is actually entirely redundant
with the instruction's execution size. Initially, it gave us the ability
to easily set the instructions execution size based entirely on register
widths. With the builder, we can easiliy set the sizes explicitly and the
width field doesn't have as much purpose. At this point, it's just
redundant information that can get out of sync so it really needs to go.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
| |
Previously we used dst.width but the two *should* be the same.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
|
|
The purpose of this change is threefold: First, it improves the
modularity of the compiler back-end by separating the functionality
required to construct an i965 IR program from the rest of the visitor
god-object, what in turn will reduce the coupling between other
components and the visitor allowing a more modular design. This patch
doesn't yet remove the equivalent functionality from the visitor
classes, as it involves major back-end surgery.
Second, it improves consistency between the scalar and vector
back-ends. The FS and VEC4 builders can both be used to generate
scalar code with a compatible interface or they can be used to
generate natural vector width code -- 1 or 4 components respectively.
Third, the approach to IR construction is somewhat different to what
the visitor classes currently do. All parameters affecting code
generation (execution size, half control, point in the program where
new instructions are inserted, etc.) are encapsulated in a stand-alone
object rather than being quasi-global state (yes, anything defined in
one of the visitor classes is effectively global due to the tight
coupling with virtually everything else in the compiler back-end).
This object is lightweight and can be copied, mutated and passed
around, making helper IR-building functions more flexible because they
can now simply take a builder object as argument and will inherit its
IR generation properties in exactly the same way that a discrete
instruction would from the same builder object.
The emit_typed_write() function from my image-load-store branch is an
example that illustrates the usefulness of the latter point: Due to
hardware limitations the function may have to split the untyped
surface message in 8-wide chunks. That means that the several
functions called to help with the construction of the message payload
are themselves required to set the execution width and half control
correctly on the instructions they emit, and to allocate all registers
with half the default width. With the previous approach this would
require the used helper functions to be aware of the parameters that
might differ from the default state and explicitly set the instruction
bits accordingly. With the new approach they would get a modified
builder object as argument that would influence all instructions
emitted by the helper function as if it were the default state.
Another example is the fs_visitor::VARYING_PULL_CONSTANT_LOAD()
method. It doesn't actually emit any instructions, they are simply
created and inserted into an exec_list which is returned for the
caller to emit at some location of the program. This sort of two-step
emission becomes unnecessary with the builder interface because the
insertion point is one more of the code generation parameters which
are part of the builder object. The caller can simply pass
VARYING_PULL_CONSTANT_LOAD() a modified builder object pointing at the
location of the program where the effect of the constant load is
desired. This two-step emission (which pervades the compiler back-end
and is in most cases redundant) goes away: E.g. ADD() now actually
adds two registers rather than just creating an ADD instruction in
memory, emit(ADD()) is no longer necessary.
v2: Drop scalarizing VEC4 builder.
v3: Take a backend_shader as constructor argument. Improve handling
of debug annotations and execution control flags.
v4: Drop Gen6 IF with inline comparison. Rename "instr" variable.
Initialize cursor to NULL by default and add method to explicitly
point the builder at the end of the program.
Reviewed-by: Matt Turner <mattst88@gmail.com>
|