| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
| |
Just a small code reduction from shared infrastructure.
|
|
|
|
| |
Just a small code reduction from shared infrastructure.
|
|
|
|
|
|
|
|
|
|
| |
This ended up confusing the scheduler for things like fabs (implemented as
fmaxabs x, x) or squaring a number, and it would try to avoid scheduling
them because it appeared more expensive than other instructions.
Fixes failure to register allocate in
dEQP-GLES2.functional.uniform_api.random.3 with almost no shader-db
effects (+.35% max temps)
|
|
|
|
|
|
|
|
|
| |
This helps us get the delay slots between SFU writes and reads filled.
total instructions in shared programs: 94494 -> 93970 (-0.55%)
instructions in affected programs: 59206 -> 58682 (-0.89%)
3DMMES performance +1.89967% +/- 0.157611% (n=10,9)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The latency_between was trying to handle the delay between the coordinate
write ("before") and the corresponding sample read ("after"), but we were
handing in the two instructions swapped.
This meant that we tried to fit things between a tex_s and its *preceding*
tex_result. This made us only interleave normal texture coordinates by
accident, and pessimized UBO reads by pushing the tex_result collection
earlier until there was nothing but it (and then its preceding coordinate
setup) left.
In addition to latency reduction, things end up packing better (probably
due to reduced live ranges of the texture results):
total instructions in shared programs: 98121 -> 94775 (-3.41%)
instructions in affected programs: 91196 -> 87850 (-3.67%)
3DMMES performance +1.15569% +/- 0.124714% (n=8,10)
|
|
|
|
|
| |
For now we're still just generating MOVs, but this will let us fold into
other ops in the future. No difference on shader-db.
|
|
|
|
|
|
| |
Every caller was dereffing the qinst, and this will let us make the number
of sources vary depending on the destination of the qinst so that we can
have general ALU ops that store to tex_[strb] and get an implicit uniform.
|
|
|
|
|
|
|
|
| |
I dropped the first hunk of this change last minute when I decided it
wasn't actually needed, and apparently failed to piglit it in simulation.
The simulator threw an an assertion in gl-1.0-drawpixels-color-index,
which queued up 5 coordinates (3 before a switch, two after) before
loading the result.
|
|
|
|
|
| |
This is a suboptimal implementation, but Jonas Pfeil found that it was
still a massive performance gain.
|
| |
|
|
|
|
|
|
|
|
| |
If a block might be entered from multiple locations, then the uniform
stream will (probably) be at different points, and we need to make sure
that it's pointing where we expect it to be. The kernel also enforces
that any block reading a uniform resets uniforms, to prevent reading
outside of the uniform stream by using looping.
|
|
|
|
|
|
| |
This uses the branch condition code in inst->cond to jump to either
successor[0] (condition matches) or successor[0] (condition doesn't
match).
|
|
|
|
|
|
|
|
| |
Basically we just treat each block independently. The only inter-block
scheduling I can think of that would be be interesting would be to move
texture result collection to after a short loop/if block that doesn't do
texturing. However, the kernel disallows that as part of its security
validation.
|
|
|
|
|
|
|
| |
The optimization passes and scheduling aren't actually ready for multiple
blocks with control flow yet (as seen by the "cur_block" references in
them instead of iterating over blocks), but this creates the structures
necessary for converting them.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clean up misrepetitions ('if if', 'the the' etc) found throughout the
comments. This has been done manually, after grepping
case-insensitively for duplicate if, is, the, then, do, for, an,
plus a few other typos corrected in fly-by
v2:
* proper commit message and non-joke title;
* replace two 'as is' followed by 'is' to 'as-is'.
v3:
* 'a integer' => 'an integer' and similar (originally spotted by
Jason Ekstrand, I fixed a few other similar ones while at it)
Signed-off-by: Giuseppe Bilotta <[email protected]>
Reviewed-by: Chad Versace <[email protected]>
|
|
|
|
|
|
|
|
| |
Now that vc4 automated code documentation can be generated with
doxygen, fix the warnings issued by Doxygen 1.8.11.
Signed-off-by: Rhys Kidd <[email protected]>
Reviewed-by: Emil Velikov <[email protected]>
|
|
|
|
|
|
|
|
| |
This lets us write the Z directly from the FTOI for computed Z, and may
let us coalesce color writes in the future.
No change in my shader-db, but clearly drops an instruction in piglit's
early-z test.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Now that the field exists in the instruction, we can make discards less
special. As a bonus, that means that we should be able to merge some more
.sf instructions together when we get around to that.
This causes some scheduling changes, as it allows tlb_color_reads to be
delayed past the discard condition setup. Since the tlb_color_read ends
up later, this may mean performance improvements, but I haven't tested.
total instructions in shared programs: 78114 -> 78035 (-0.10%)
instructions in affected programs: 1922 -> 1843 (-4.11%)
total estimated cycles in shared programs: 234318 -> 234329 (0.00%)
estimated cycles in affected programs: 8200 -> 8211 (0.13%)
|
|
|
|
|
|
|
|
|
|
|
|
| |
Silences the following GCC warning:
mesa/src/gallium/drivers/vc4/vc4_qir_schedule.c: In function 'qir_schedule_instructions':
mesa/src/gallium/drivers/vc4/vc4_qir_schedule.c:578:16: warning: missing braces around initializer [-Wmissing-braces]
struct schedule_state state = { 0 };
^
Signed-off-by: Rhys Kidd <[email protected]>
Signed-off-by: Eric Anholt <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I'm moving away from QIR being SSA (since NIR is doing lots of SSA
optimization for us now) and instead having QIR just be QPU operations
with virtual registers. By making our SELs be composed of two MOVs, we
could potentially coalesce the registers for the MOV's src and dst and
eliminate the MOV.
total instructions in shared programs: 88448 -> 88028 (-0.47%)
instructions in affected programs: 39845 -> 39425 (-1.05%)
total estimated cycles in shared programs: 246306 -> 245762 (-0.22%)
estimated cycles in affected programs: 162887 -> 162343 (-0.33%)
|
|
This is a rewrite of vc4_opt_qpu_schedule.c to operate on QIR. Texture
fetch can probably take as much as the rest of the cycles of the program,
so it's important to hide our other cycles during it (which is hard to do
after register allocation). Also, we can queue up multiple texture
requests before collecting the resulting samples, so that we keep the
texture unit busy more of the time.
High-settings openarena performance +2.35849% +/- 0.221154% (n=7). Also
about 2-3% on the multiarb demo. 8 piglit tests
(ext_framebuffer_multisample accuracy depthstencil) go from failing in
rendering to failing in register allocation, but hopefully I can fix that
up with some better register pressure handling here.
total instructions in shared programs: 87723 -> 88448 (0.83%)
instructions in affected programs: 78411 -> 79136 (0.92%)
total estimated cycles in shared programs: 276583 -> 246306 (-10.95%)
estimated cycles in affected programs: 265691 -> 235414 (-11.40%)
|