This commit adds a NIR pass for lowering away returns in functions. If the
return is in a loop, it is lowered to a break. If it is not in a loop,
it's lowered away by moving/deleting code as needed.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
This can happen if a function ends in a return instruction and you remove
the return.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The efficiency should be approximately the same. We do a little more work
per phi node because we have to sort the predecessors. However, we no
longer have to walk the blocks a second time to pop things off the stack.
The bigger advantage, however, is that we can now re-use the phi placement
and per-block SSA value tracking in other passes.
As a side-benifit, the phi builder actually handles unreachable blocks
correctly. The original vars_to_ssa code, because of the way it iterated
the blocks and added phi sources, didn't add sources corresponding to
predecessors of unreachable blocks. The new strategy employed by the phi
builder creates a phi source for each predecessor and should correctly
handle unreachable blocks by setting those sources to SSA undefs.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Previously, nir_dominance.c didn't properly handle unreachable blocks.
This can happen if, for instance, you have something like this:
loop {
if (...) {
break;
} else {
break;
}
}
In this case, the block right after the if statement will be unreachable.
This commit makes two changes to handle this. First, it removes an assert
and allows block->imm_dom to be null if the block is unreachable. Second,
it properly skips unreachable blocks in calc_dom_frontier_cb.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Right now, we have phi placement code in two places and there are other
places where it would be nice to be able to do this analysis. Instead of
repeating it all over the place, this commit adds a helper for placing all
of the needed phi nodes for a value.
v2: Add better documentation
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
In many places, the convention is to pass an existing ssadef name ptr
when construction/initializing a new nir_ssa_def. But that goes badly
(as noticed by garbage in nir_print output) when the original string
gets freed.
Just use ralloc_strdup() instead, and add ralloc_free() in the two
places that would care (not that the strings wouldn't eventually get
freed anyways).
Also fixup the nir_search code which was directly setting ssadef->name
to use the parent instruction as memctx.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Many of our optimizations, while great for cutting shaders down to size,
aren't really precision-safe. This commit tries to flag all of the
inexact floating-point optimizations so they don't get run on values that
are flagged "exact". It's a bit conservative and maybe flags some safe
optimizations as unsafe but that's better than missing one.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The previous transformation got the arguments to fmin backwards. When NaNs
are involved, the GLSL min/max aren't commutative so it matters.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The fxor opcode is required to return 1.0f or 0.0f but the input variable
may not be 1.0f or 0.0f.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
No shader-db changes, but this is symmetric with the previous commit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This also prevented some regressions with other patches in my local
tree.
Broadwell / Skylake
total instructions in shared programs: 8980835 -> 8980833 (-0.00%)
instructions in affected programs: 45 -> 43 (-4.44%)
helped: 1
HURT: 0
total cycles in shared programs: 70077904 -> 70077900 (-0.00%)
cycles in affected programs: 122 -> 118 (-3.28%)
helped: 1
HURT: 0
No changes on earlier platforms.
v2: Modify the comments to look more like a proof.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
On Intel platforms that don't set lower_flrp, using bcsel instead of
flrp seems to be a small amount worse. On those platforms, the use of
flrp, bcsel, and multiply of b2f is still an active area of research.
In review, Matt suggested this is because bcsel turns into CMP+SEL, and
because of the flag register we can't schedule instructions well.
shader-db results:
G4X / Ironlake
total instructions in shared programs: 4016538 -> 4012279 (-0.11%)
instructions in affected programs: 161556 -> 157297 (-2.64%)
helped: 1077
HURT: 1
total cycles in shared programs: 84328296 -> 84315862 (-0.01%)
cycles in affected programs: 4174570 -> 4162136 (-0.30%)
helped: 926
HURT: 53
Unsurprisingly, no changes on later platforms.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
When we replace an expresion we have to compute bitsize information for the
replacement. We do this in two passes to validate that bitsize information
is consistent and correct: first we propagate bitsize from child nodes to
parent, then we do it the other way around, starting from the original's
instruction destination bitsize.
v2 (Iago):
- Always use nir_type_bool32 instead of nir_type_bool when generating
algebraic optimizations. Before we used nir_type_bool32 with constants
and nir_type_bool with variables.
- Fix bool comparisons in nir_search.c to account for bitsized types.
v3 (Sam):
- Unpack the double constant value as unsigned long long (8 bytes) in
nir_algrebraic.py.
v4 (Sam):
- Use helpers to get type size and base type from nir_alu_type.
Signed-off-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
v2: Squash multiple commits addressing the new parameter in different
files so we don't break the build (Iago)
v3: Fix tgsi (Samuel)
v4: Fix nir_clone.c (Samuel)
v5: Fix vc4 and freedreno (Iago)
v6 (Sam)
- Fix build errors in nir_lower_indirect_derefs
- Use helper to get type size from nir_alu_type.
Signed-off-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Samuel Iglesias Gonsalvez <siglesias@igalia.com>
Tested-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Some opcodes need explicit bitsizes, and sometimes we need to use the
double version when constant folding.
v2: fix output type for u2f (Iago)
v3: do not change vecN opcodes to be float. The next commit will add
infrastructure to enable 64-bit integer constant folding so this is isn't
really necessary. Also, that created problems with source modifiers in
some cases (Iago)
v4 (Jason):
- do not change bcsel to work in terms of floats
- leave ldexp generic
Squashed changes to handle different bit sizes when constant
folding since otherwise we would break the build.
v2:
- Use the bit-size information from the opcode information if defined (Iago)
- Use helpers to get type size and base type of nir_alu_type enum (Sam)
- Do not fallback to sized types to guess bit-size information. (Jason)
Squashed changes in i965 and gallium/nir drivers to support sized types.
These functions should only see sized types, but we can't make that change
until we make sure that nir uses the sized versions in all the relevant places.
A later commit will address this.
Signed-off-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
v2: use a ternary (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This really hacky commit adds a bit size to registers and SSA values. It
also adds rules in the validator to validate that they do the right things.
It's still an open question as to whether or not we want a bit_size in
nir_alu_instr or if we just want to let it inherit from the destination.
I'm inclined to just let it inherit from the destination. A similar
question needs to be asked about intrinsics.
v2 (Connor):
- Relax validation: comparisons have explicit destination sizes
and implicit source sizes.
v3 (Sam):
- Use helpers to get size and base types of nir_alu_type enum.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
v2: Fix size/type mask to properly handle 8-bit types.
v3: Add helpers to get the bitsize and base type of a
nir_alu_type enum.
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This allows us to first generate atomic operations for shared
variables using these opcodes, and then later we can lower those to
the shared atomics intrinsics with nir_lower_io.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Previously we were receiving shared variable accesses via a lowered
intrinsic function from glsl. This change allows us to send in
variables instead. For example, when converting from SPIR-V.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Since we aren't going to put the function parameters or the return variable
in the list of locals, it won't get a proper declaration. This changes
nir_print to print the type along with each parameter or return variable.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Otherwise, we have a problem when we go to print functions with arguments
because their names get added to the hash table during declaration which
happens after we print the prototype.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
NIR has never been used on IR where we haven't already done function
inlining so this code has been dead from the beginning. Let's just get rid
of it for now. We can always put it back in if we decide to use NIR for
function inlining at some point in the future.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
This new pass lowers load/store_var intrinsics that act on indirect derefs
to if-ladder of direct load/store_var intrinsics. The if-ladders perform a
simple binary search on the indirect.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
No shader-db changes, but does recognize some extract_u16 which enables
the next patch to optimize some code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Two shaders that appear in Unigine benchmarks (Heaven and Valley) unpack
three bytes from an integer and convert each into a float:
float((val >> 16u) & 0xffu)
float((val >> 8u) & 0xffu)
float((val >> 0u) & 0xffu)
Instead of shifting, masking, and type converting like this:
shr(8) g15<1>UD g25<8,8,1>UD 0x00000010UD
and(8) g16<1>UD g15<8,8,1>UD 0x000000ffUD
mov(8) g17<1>F g16<8,8,1>UD
shr(8) g18<1>UD g25<8,8,1>UD 0x00000008UD
and(8) g19<1>UD g18<8,8,1>UD 0x000000ffUD
mov(8) g20<1>F g19<8,8,1>UD
and(8) g21<1>UD g25<8,8,1>UD 0x000000ffUD
mov(8) g22<1>F g21<8,8,1>UD
i965 can simply extract a byte and convert to float in a single
instruction:
mov(8) g17<1>F g25.2<32,8,4>UB
mov(8) g20<1>F g25.1<32,8,4>UB
mov(8) g22<1>F g25.0<32,8,4>UB
This patch implements the first step: recognizing byte extraction. A
later patch will optimize out the conversion to float.
instructions in affected programs: 28568 -> 27450 (-3.91%)
helped: 7
cycles in affected programs: 210076 -> 203144 (-3.30%)
helped: 7
This patch decreases the number of instructions in the two Unigine
programs by:
#1721: 4520 -> 4374 instructions (-3.23%)
#1706: 3752 -> 3582 instructions (-4.53%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Unfortunately, glslang gives us cull/clip distance and GS streams even if
the shader doesn't use it whenever a shader is declared as version 450.
This is a glslang bug, but we can easily enough ignore it for now.
"Result is the same as computing the sum of the absolute values of
OpDPdx and OpDPdy on P."
We were doing sum of absolute values of OpDPdx of P and OpDPdx of NULL.
These were originally added to reduce compiler warnings but aren't really
needed. Getting rid of them reduces the diff between the Vulkan branch and
master, so we might as well.
When NIR was originally drafted, there was no easy way to determine if
something was constant or not. The result was that we had lots of
special-casing for constant values such as this. Now that load_const
instructions are SSA-only, it's really easy to find constants and this
isn't really needed anymore.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Rob Clark <robclark@gmail.com>
This fixes two issues. First, we had a use-after-free in the case where
the instruction got deleted and we tried to return mov->dest.write_mask.
Second, in the case where we are doing a self-mov of a register, we delete
those channels that are moved to themselves from the write-mask. This
means that those channels aren't reported as being handled even though they
are. We now stash off the write-mask before remove unneeded channels so
that they still get reported as handled.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94073
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
The sub-determinate implementation pattern fixed by
6a7e2904e0 has a second instance in the
same file.
With the previous algorithm, when row and j are both 3, the index
overruns the array. This only impacts the stack on 32 bit builds.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This commit adds the capability to NIR to support separate textures and
samplers. As it currently stands, glsl_to_nir only sets the texture deref
and leaves the sampler deref alone as it did before and nir_lower_samplers
assumes this. Backends can still assume that they are combined and only
look at only at the texture index. Or, if they wish, they can assume that
they are separate because nir_lower_samplers, tgsi_to_nir, and prog_to_nir
all set both texture and sampler index whenever a sampler is required (the
two indices are the same in this case).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We're about to separate the two concepts. When we do, the sampler will
become optional. Doing a rename first makes the separation a bit more
safe because drivers that depend on GLSL or TGSI behaviour will be fine to
just use the texture index all the time.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Direct access to intr->const_index[n], where different slots have
different meanings, is somewhat confusing.
Instead, let's put some extra info in nir_intrinsic_infos[] about which
slots map to what, and add some get/set helpers. The helpers validate
that the field being accessed (base/writemask/etc) is applicable for the
intrinsic opc, for some extra safety. And nir_print can use this to
dump out decoded const_index fields.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Previously, in order to get things working, we just always shadowed
variables. Now, we rewrite derefs whenever it's safe to do so and only
shadow if we have an in or out variable that we write or read to
respectively.
These are used in GLSL IR to removed unused varyings and match
transform feedback variables. There is no need to use these in NIR.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Helps 11 shaders in UnrealEngine4 demos.
I seriously hope they would have given us bitfieldReverse() if we
exposed GL 4.0 (but we do expose ARB_gpu_shader5, so why not use that
anyway?).
instructions in affected programs: 4875 -> 4633 (-4.96%)
cycles in affected programs: 270516 -> 244516 (-9.61%)
I suspect there's a *lot* of room to improve nir_search/opt_algebraic's
handling of this. We'd actually like to match, e.g., step2 by matching
step1 once and then doing a pointer comparison for the second instance
of step1, but unfortunately we generate an enormous tuple for instead.
The .text size increases by 6.5% and the .data by 17.5%.
text data bss dec hex filename
22957 45224 0 68181 10a55 nir_libnir_la-nir_opt_algebraic.o
24461 53160 0 77621 12f35 nir_libnir_la-nir_opt_algebraic.o
I'd be happy to remove this if Unreal4 uses bitfieldReverse() if it is
in a GL 4.0 context once we expose GL 4.0.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The next patch adds an algebraic rule that uses the constant 0xff00ff00.
Without this change, the build fails with
return hex(struct.unpack('I', struct.pack('i', self.value))[0])
struct.error: 'i' format requires -2147483648 <= number <= 2147483647
The hex() function handles integers of any size, and assigning a
negative value to an unsigned does what we want in C. The pack/unpack is
unnecessary (and as we see, buggy).
Reviewed-by: Dylan Baker <baker.dylan.c@gmail.com>
Walking the SSA definitions in order means that we consider the smallest
algebraic optimizations before larger optimizations. So if a smaller
rule is part of a larger rule, the smaller one will happen first,
preventing the larger one from happening.
instructions in affected programs: 32721 -> 32611 (-0.34%)
helped: 106
In programs whose nir_optimize loop count changes (129 of them):
before: 1164 optimization loops
after: 1071 optimization loops
Of the 129 affected, 16 programs' optimization loop counts increased.
Prevents regressions and annoyances in the next commits.
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Parameterize build_asin() on the fit coefficients so the
implementation can be shared while still using different polynomials
for asin and acos. Also switch back to implementing acos in terms of
asin -- The improvement obtained from cancelling out the pi/2 terms
was negligible compared to the approximation error.
vtn_handle_type() creates a signed type regardless of the value of the
signedness flag, which usually doesn't make much of a difference
except when the type is used as base sampled type of an image type,
what will cause the base type of the NIR image variable to be
inconsistent with its format and cause an assertion failure in the
back-end (most likely only reproducible on Gen7), and may change the
semantics of the image intrinsic subtly (e.g. UMIN may become IMIN).