A live range [0, 0) counts as empty. For function inputs this can
be a problem, so insert a nop at the beginning to make it [0, 1).
This is a bit of a hack but also the most simple solution.
Currently works on a220. Others in the a2xx family look pretty similar
and should be pretty straightforward to support with the same driver.
The a3xx has a new shader ISA, and while many registers appear similar,
the register addresses have been completely shuffled around. I am not
sure yet whether it is best to support with the same driver, but
different compiler, or whether it should be split into a different
driver.
v1: original
v2: build file updates from review comments, and remove GPL licensed
header files from msm kernel
v3: smarter temp/pred register assignment, fix clear and depth/stencil
format issues, resource_transfer fixes, scissor fixes
Signed-off-by: Rob Clark <robdclark@gmail.com>
Previously, the derivatives were calculated and passed in a packed form
to the sample code (for implicit derivatives, explicit derivatives were
packed to the same format).
There's several reasons why this wasn't such a good idea:
1) the derivatives may not even be needed (not as bad as it sounds since
llvm will just throw the calculations needed for them away but still)
2) the special packing format really shouldn't be part of the sampler
interface
3) depending what the sample code actually does the derivatives will
be processed differently, hence there is no "ideal" packing. For cube
maps with explicit derivatives (which we don't do yet) for instance the
packing looked downright useless, and for non-isotropic filtering we'd
need different calculations too.
So, instead just pass the derivatives as is (for explicit derivatives),
or let the rho calculating sample code calculate them itself. This still
does exactly the same packing stuff for implicit derivatives for now,
though explicit ones are handled in a more straightforward manner (quick
estimates show performance should be quite similar, though it is much
easier to follow and also does the rho calculation per-pixel until the
end, which we eventually need for spec compliance anyway).
No piglit changes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
After the previous fix that almost removes an allocation of 4*n^2
bytes, we can use a bitset to reduce another allocation from n^2 bytes
to n^2/8 bytes.
Between the previous commit and this one, the peak heap size for an
oglconform ARB_fragment_program max instructions test on i965 goes from
4GB to 255MB.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=55825
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We were allocating an adjacency_list entry for every possible
interference that could get created, but that usually doesn't happen.
We can save a lot of memory by resizing the array on demand.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We're already walking the list, and we can easily know when something
has no reason to be in the list any longer, so take a brief extra step
to reduce our worst-case runtime (an oglconform test that emits the
maximum instructions in a fragment program). I don't actually know what
the worst-case runtime was, because it was too long and I got bored.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We can execute way fewer instructions by doing our boolean manipulation
on an "int" of bits at a time, while also reducing our working set size.
Reduces compile time of L4D2's slowest shader from 4s to 1.1s
(-72.4% +/- 0.2%, n=10)
v2: Remove redundant masking (noted by Ken)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We were handling the the dependency workaround for the first written reg
of a send preceding the one we're fixing up, but didn't consider the other
regs. Thus if you had two sampler calls that got allocated to the same
set of regs, one might, rarely, ovewrite the other. This was occurring in
XBMC's GLSL shaders.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=44567
NOTE: This is a candidate for the stable branches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When forcing the compiler to always generate pull constants instead of
push constants (in order to have an easy to use testcase), improves
performance of my old GLSL demo 23.3553% +/- 1.42968% (n=7).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=60866
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The lowering process creates a new vgrf on gen7 that should be represented
in live interval analysis. As-is, it was getting a conflicting allocation
with gl_FragDepth in the dolphin emulator, producing broken rendering.
NOTE: This is a candidate for the 9.1 branch.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61317
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I was going to fix the code above like the previous commit, but we already
had that covered (otherwise all our uniform access would have been broken,
unlike just pull constants).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We were allowing a compressed instruction to write a register that
contained the last use of a uniform pull constant (either UBO load or push
constant spillover), so it would get half its values smashed.
Since we need to see the actual instruction to decide this, move the
pre-gen6 pixel_x/y logic here, which should improve the performance of
register allocation since virtual_grf_interferes() is called more than
once per instruction.
NOTE: This is a candidate for the stable branches.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61317
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I was looking at the list to see what might be interesting to document for
application developers, and it turns out some are completely dead.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We were assuming that each emitted primitive had the same
number of vertices. That is incorrect. Emitted primitives
can have arbirtrary number of vertices. Simply increment
index on iteration to fix it.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Whenever we're binding the shaders we're incrementing NumOutputs,
assuming the parser spots an output decleration, but we were never
reseting the variable. That means that each subsequent bind of
a geometry shader would add its number of output to the number
of output bound by all previously ran shaders and our indexes
would get completely messed up.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
This is needed for handling the dx10-style sample opcodes.
This also simplifies the logic by getting rid of sampler variants
completely (sampler_views though OTOH have sort of variants because
some of their state is different depending on the shader stage they
are bound to).
No significant performance difference (openarena run:
840 frames in 459.8 seconds vs. 840 frames in 460.5 seconds).
v2: fix reference counting bug spotted by Jose.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Can handle them since the single sampler interface was introduced.
v2: simplify txf/sample_i handling a bit according to Brian's feedback.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
