| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
| |
There are a variety of places where we use dst.width / 8 to compute the
size of a single logical channel. Instead, we should be using exec_size.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
|
| |
The backend_shader class really is a representation of a shader. The fact
that it inherits from ir_visitor is somewhat immaterial.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
| |
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
| |
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, the vec4 backend labeled shaders as "vec4" - now it uses the
specific names "VS" and "GS".
The FS backend now correctly prints "VS" for vertex shaders (rather than
"fs"). It also prints "FS" instead of "fs" for fragment shaders;
preserving that behavior didn't seem essential.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
|
|
|
|
| |
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
| |
multiple registers.
v2: Avoid nested ternary operators in vec4_instruction::regs_read(). (Matt)
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
| |
The fs_visitor argument of fs_inst::regs_read() wasn't used at all.
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Right now virtual GRF book-keeping and allocation is performed in each
visitor class separately (among other hundred different things),
leading to duplicated logic in each visitor and preventing layering as
it forces any code that manipulates i965 IR and needs to allocate
virtual registers to depend on the specific visitor that happens to be
used to translate from GLSL IR.
v2: Use realloc()/free() to allocate VGRF book-keeping arrays (Connor).
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Now that we properly track accumulator dependencies, the scheduler is
able to schedule instructions between the mach and mov in the common
the integer multiplication pattern:
mul acc0, x, y
mach null, x, y
mov dest, acc0
Since a null destination implies no dependency on the destination, we
can also safely schedule instructions (that don't write the accumulator)
between the mul and mach.
GAINED: 103
LOST: 43
Causes one program to spill (643 -> 1076 instructions).
I committed this patch last year (commit 42a26cb5) but reverted it
(commit 0d3f83f4) after inexplicable artifacts in Kerbal Space Program
(bug 78648). Tapani reapplied this patch and could not reproduce the bug
with current Mesa.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
|
|
|
|
| |
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
|
|
|
|
|
| |
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is actually the squash of a bunch of different changes. Individual
commit titles follow:
i965/fs: Always 2-align registers SIMD16 for gen <= 5
i965/fs: Use the register width when applying offsets
This reworks both byte_offset() and offset() to be more intelligent.
The byte_offset() function now supports offsets bigger than 32. The
offset() function uses the byte_offset() function together with the
register width and the type size to offset the register by the correct
amount.
i965/fs: Change regs_read to be in hardware registers
i965/fs: Change regs_written to be actual hardware registers
i965/fs: Properly handle register widths in LOAD_PAYLOAD
The LOAD_PAYLOAD instruction is a bit special because it collects a
bunch of registers (with possibly different widths) into a single
payload block. Once the payload is constructed, it's treated as a
single block of data and most of the information such as register widths
doesn't matter anymore. In particular, the offset of any particular
source register is the accumulation of the sizes of the previous source
registers.
i965/fs: Properly set writemasks in LOAD_PAYLOAD
i965/fs: Handle register widths in demote_pull_constants
i965/fs: Get rid of implicit register doubling in the allocator
i965/fs: Reserve enough registers for PLN instructions
i965/fs: Make sources and destinations interfere in 16-wide
i965/fs: Properly handle register widths in CSE
i965/fs: Properly handle register widths in register_coalesce
i965/fs: Properly handle widths in copy propagation
i965/fs: Properly handle register widths in VARYING_PULL_CONSTANT_LOAD
i965/fs: Properly handle register widths and odd register sizes in spilling
i965/fs: Don't waste a register on texture lookups for gen >= 7
Previously, we were waisting a register in SIMD16 mode because we could
only allocate registers in pairs. Now that we can allocate and address
odd-sized registers, let's get rid of this special-case.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
|
| |
Now that nothing invalidates the CFG, we can calculate_cfg() immediately
after emit_fb_writes()/emit_thread_end() and never again.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
| |
Everything has been converted to preserve the CFG.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
| |
... on Gen6+. I'm not actually sure which class Gen6 fits into.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously instruction scheduling tracked dependencies on a per-register
basis. This meant that there was an artificial dependency between
interpolation instructions writing into the same virtual register.
Instruction scheduling would insert a number of instructions between the
two instructions in this example, when they are actually independent.
linterp vgrf8+0.0:F, hw_reg2:F, hw_reg3:F, hw_reg6:F
linterp vgrf8+1.0:F, hw_reg2:F, hw_reg3:F, hw_reg6+16:F
This lead to cases where the first texture coordinate is interpolated at
the beginning of the shader, but the second is done immediately before
the texture operation that uses it as a source.
After this change, the artificial dependency is removed and the
interpolation instructions are scheduled together.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
| |
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
|
|
|
|
| |
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
|
|
|
|
|
|
|
| |
Will simplify the automated conversion if we want to allow compiling the
driver for a single generation.
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
|
|
|
|
|
|
| |
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
| |
This reverts commit 42a26cb5e441a01d5288b299980f23affaad53fe.
Cc: "10.2" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=78648
|
|
|
|
|
|
|
|
| |
We had a handful of cases where we'd used brw_imm_*() to generate an
immediate, rather than fs_reg(). We shouldn't do that but we shouldn't
limit scheduling flexibility on account of immediate arguments either.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
| |
Many instructions implicitly update the accumulator on Gen < 6. The instruction
scheduling code just calls add_barrier_deps() for each accumulator access on
these platforms, but a large class of operations don't actually update the
accumulator -- mostly move and logical instructions. Teaching the scheduling
code about this would allow more flexibility to schedule instructions.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77740
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Now that we properly track accumulator dependencies, the scheduler is
able to schedule instructions between the mach and mov in the common
the integer multiplication pattern:
mul acc0, x, y
mach null, x, y
mov dest, acc0
Since a null destination implies no dependency on the destination, we
can also safely schedule instructions (that don't write the accumulator)
between the mul and mach.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Our hardware has an "accumulator" register, which can be used to store
intermediate results across multiple instructions. Many instructions
can implicitly write a value to the accumulator in addition to their
normal destination register. This is enabled by the "AccWrEn" flag.
This patch introduces a new flag, inst->writes_accumulator, which
allows us to express the AccWrEn notion in the IR. It also creates a
n ALU2_ACC macro to easily define emitters for instructions that
implicitly write the accumulator.
Previously, we only supported implicit accumulator writes from the
ADDC, SUBB, and MACH instructions. We always enabled them on those
instructions, and left them disabled for other instructions.
To take advantage of the MAC (multiply-accumulate) instruction, we
need to be able to set AccWrEn on other types of instructions.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
|
|
|
|
|
|
| |
To be consistent with the fs backend.
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
| |
To be consistent with the fs backend. Also the instruction scheduler
incorrectly considered SEL with a conditional modifier to read the flag
register.
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
| |
<4,1,1> isn't a real thing. We meant <4,4,1>, i.e., each component of
the whole register.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
| |
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
| |
Those are the terms used in the docs, and think "n-wide" was something I
just happened to say. Note that shader-db needs updating for the
INTEL_DEBUG=fs parsing.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I recently made us try two different things that tried to reduce register
pressure so that we would be more likely to allocate successfully. But
now that we have the logic for trying two, we can make the first thing we
try be the normal, not-prioritizing-register-pressure heuristic.
This means one less scheduling pass in the common case of that heuristic
not producing spills, plus the best schedule we know how to produce, if
that one happens to succeed. This is important, because our register
allocation produces a lot of possibly avoidable dependencies for the
post-register-allocation schedule, despite ra_set_allocate_round_robin().
GLB2.7: 1.04127% +/- 0.732461% fps improvement (n=31)
nexuiz: No difference (n=5)
lightsmark: 0.838512% +/- 0.300147% fps improvement (n=86)
minecraft apitrace: No difference (n=15)
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
| |
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
| |
Fixes "Missing break in switch" defect reported by Coverity.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
|
|
|
|
|
|
|
| |
Fixes "Uninitialized scalar field" defect reported by Coverity.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since LIFO fails on some shaders in one particular way, and non-LIFO
systematically fails in another way on different kinds of shaders, try
them both, and pick whichever one successfully register allocates first.
Slightly prefer non-LIFO in case we produce extra dependencies in register
allocation, since it should start out with fewer stalls than LIFO.
This is madness, but I haven't come up with another way to get unigine
tropics to not spill while keeping other programs from not spilling and
retaining the non-unigine performance wins from texture-grf.
total instructions in shared programs: 1626728 -> 1626288 (-0.03%)
instructions in affected programs: 1015 -> 575 (-43.35%)
GAINED: 50
LOST: 0
Improves Unigine Tropics performance by 14.5257% +/- 0.241838% (n=38)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70445
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We care about depth-until-program-end, as a proxy for "make sure I
schedule those early instructions that open up the other things that can
make progress while keeping register pressure low", not actual latency
(since we're relying on the post-register-alloc scheduling to actually
schedule for the hardware).
total instructions in shared programs: 1609931 -> 1609931 (0.00%)
instructions in affected programs: 0 -> 0
GAINED: 55
LOST: 43
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, the best thing we had was to schedule the things unblocked by
the last chosen instruction, on the hope that it would be consuming two
values at the end of their live intervals while only producing one new
value. But that's just a guess, and we can do counting of usage of
registers to know when an instruction would (almost surely) reduce
register pressure.
The only failure mode I know of in this new dominant heuristic is that
inside of a loop when scheduling the iterator (for example), choosing the
last use of the iterator doesn't actually reduce the live interval of the
iterator. But it doesn't seem to matter in shader-db:
total instructions in shared programs: 1618700 -> 1618700 (0.00%)
instructions in affected programs: 0 -> 0
GAINED: 13
LOST: 0
Note: The new functions are made virtual because I expect we'll soon lift
the pre-regalloc scheduling heuristic over to the vec4 backend.
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
| |
The latency information has been obtained empirically from
measurements taken on Haswell and Ivy Bridge.
Acked-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch fixes the three dead code elimination passes and the
VEC4/FS instruction scheduling passes so they leave instructions with
side effects alone.
At some point it might be interesting to have the instruction
scheduler calculate the exact memory dependencies between atomic ops,
but they're rare enough that it seems unlikely that it will make any
practical difference.
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
| |
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
|
| |
This avoids a lot of message setup we had to do otherwise. Improves
GLB2.7 performance with register spilling force enabled by 1.6442% +/-
0.553218% (n=4).
v2: Use BRW_PREDICATE_NONE, improve a comment (by Paul).
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When faced with a million instructions that all became candidates at the
same time (none of which individually reduce register pressure), the ones
on the critical path are more likely to be the ones that will free up some
candidates soon.
shader-db:
total instructions in shared programs: 1681070 -> 1681070 (0.00%)
instructions in affected programs: 0 -> 0
GAINED: 40
LOST: 74
Fixes indistinguishable-from-hanging behavior in GLES3conform's
uniform_buffer_object_max_uniform_block_size test, regressed by
c3c9a8c85758796a26b48e484286e6b6f5a5299a. Given that
93bd627d5a6c485948b94488e6cd53a06b7ebdcf was unlocked by that commit, the
net effect on 16-wide program count is still quite positive, and I think
this should give us more stable scheduling (less dependency on original
instruction emit order).
v2: Comment suggestions by Paul
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70943
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a step in doing scheduling as described in Muchnick (p538). A
difference is that our latency function is only specific to one
instruction (it doesn't describe, for example, the different latency
between WAR of a send's arguments and RAW of a send's destination), but
that's changeable later. We also don't separately compute the postorder
traversal of the graph, since we can use the setting of the delay field as
the "visited" flag.
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
| |
These ralloc contexts belong to a specific object and are being
deallocated manually from the class destructor. Now that we've hooked
up destructors to ralloc there's no reason for them to be children of
any other context, and doing so might to lead to double frees under
some circumstances. The class destructor has all the responsibility
of freeing class memory resources now.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This was a hack to avoid choosing to schedule all texturing before
consumption of any texture results due to the way dependency chains worked
out in the presence of MRFs. On gen7, we don't have MRFs, so the problem
doesn't apply, and this was just badly constraining our scheduling.
total instructions in shared programs: 1615306 -> 1612534 (-0.17%)
instructions in affected programs: 9958 -> 7186 (-27.84%)
GAINED: 259
LOST: 9
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The LIFO plan was simple: Take the most recently made available
instructions, and pick those first.
But because of the order we were pushing things onto our list of
available-to-schedule instructions, it meant that when a set of
instructions was made available at the same time (for example, everything
at the start of the program that didn't depend on other instructions) we'd
schedule them in reverse order.
If you had 10 texture calls in a row in your program, each with
independent argument setup, we'd set up the last texture call's args and
execute it first, even though we wouldn't be able to consume its results
until we'd finished the other 9 texture calls (assuming consumption of
texture results happens near each texture call, and combines it with
another texture result, which is normal for a convolution shader).
To fix this, walk the list for doing LIFO in the order that instructions
were originally generated in the program, but choose to push
newly-made-available instructions to the other end of the list instead.
total instructions in shared programs: 1587242 -> 1586290 (-0.06%)
instructions in affected programs: 7801 -> 6849 (-12.20%)
GAINED: 76
LOST: 67
Thanks to Chia-I Wu for pointing out the bug in my first version of the
patch that made it a huge loss.
Reviewed-by: Matt Turner <mattst88@gmail.com>
|