It is now to the point where we have no regressing piglit tests. It
also fixes Yo Frankie! and Humus DynamicBranching, probably due to the
piglit bias tests that work that didn't on the Mesa IR backend.
As a downside, performance takes about a 5-10% performance hit at the
moment (e.g. nexuiz 19.8fps -> 18.8fps), which I plan to resolve by
reintroducing 16-wide fragment shaders where possible. It is a win,
though, for fragment shaders using flow control.
Simply using RNDU, RNDZ, or RNDE does not produce the desired result.
Rather, the RND* instructions place a value in the destination register
that may be 1 less than the correct answer. They can also set per-channel
"increment bits" in a flag register, which, if set, mean dest needs to
be incremented by 1. A second instruction - a predicated add -
completes the job.
Notably, RNDD always produces the correct answer in a single
instruction.
Fixes piglit test glsl-fs-trunc.
This cuts usually 2 out of 3 instructions for flag reg generation (if
statements, conditional assignment) by producing the conditional mod
in the expression representing the boolean value.
Fixes glsl-fs-vec4-indexing-temp-dst-in-nested-loop-combined (register
allocation no longer fails for the conditional generation
proliferation)
This will be a place to peephole comparisions directly to the flag
regs, and for now avoids using MOV with conditional mod on gen6, which
is now illegal.
GLES1 and GLES2 install their own exec pointers and don't need the
Save table. Also, the SET_* macros use different indices for the different
APIs so the offsets used in vtxfmt.c are actually wrong for the ES APIs.
Just always check for FLUSH_UPDATE_CURRENT and call Driver.BeginVertices
when necessary. By using the unlikely() macros, this ends up as
a 10% performance improvement (for isosurf, anyway) over the old,
complicated function pointer swapping.
Don't use r0 for FF_SYNC dest reg on Sandybridge, which would
smash FFID field in GS payload, that cause later URB write fail.
Also not use r0 in any URB write requiring allocate.
Actually validate that the implementation supports the particular
shader target as well. Previously if a driver only supported vertex
shaders, for example, glCreateShaderObjectARB would gladly create a
fragment shader.
NOTE: this is a candidate for the 7.9 branch.
This really amounts to just using the return value from
link_function_calls. All the work was being done, but the result was
being ignored.
Fixes piglit test link-unresolved-funciton.
NOTE: this is a candidate for the 7.9 branch.
Together with the previous commit, this generalize the benefits of
d2cf757f44 to all depth formats, in
particular:
- simpler float -> 24unorm conversion
- avoid unsigned comparisons (not directly supported on SSE) by aligning
to the least significant bit
- avoid unecessary/repeated mask ANDing
Verified with trivial/tri-z that the exact same assembly is produced for
X8Z24.
Z32_FLOAT uses <4 x float> as intermediate/destination type,
instead of <4 x i32>.
The necessary bitcasts got removed with commit
5b7eb868fd
Also use depth/stencil type and build contexts consistently, and
make the depth pointer argument a ordinary <i8 *>, to catch this
sort of issues in the future (and also to pave way for Z16 and
Z32_FLOAT_S8_X24 support).
__GLcontextModes is always only used as an implementation internal struct
at this point and we shouldn't install glcore.h anymore. Anything that
needs __GLcontextModes should just include the struct in its headers files
directly.
This is relying on lp_build_pack2 using the sse2 pack intrinsics which
handle clamping.
(Alternatively could have make it use lp_build_packs2 but it might
not even produce more efficient code than not using the fastpath
in the first place.)
Fixes this GCC warning.
r300_state.c: In function 'r300InvalidateState':
r300_state.c:2247: warning: 'hw_format' may be used uninitialized in this function
r300_state.c:2247: note: 'hw_format' was declared here
This adds proper support for the GL_ARB_shader_stencil_export extension
to the GLSL compiler. Thanks to Ian for pointing out where I need to add things.
If the pipe driver has shader stencil export we can accelerate DrawPixels
using it. It tries to pick an S8 texture and works its way to X24S8 and S8X24
if that isn't supported.
We need a texture to put the drawpixels stuff into, an S8 texture is less
memory/bandwidth than the 32-bit X24S8, but we might not be able to render
directly to an S8, so this lets us specify we won't be rendering to this
texture.
this improves mesa texstore for 8/24 so it can create S24X8/X24S8 variants
by keeping the depth bits static.
it also adds a texstore for S8 so we can write out an S8 texture to use
in the sampler for accel draw pixels to save memory bw.
The logic seems sound here, I've worked it out a few times on paper, though
it would be good to have some review.
Signed-off-by: Dave Airlie <airlied@redhat.com>
There was a check to only do the rebase if we didn't have everything
in VBOs, but nexuiz apparently hands us a mix of VBOs and arrays,
resulting in blocking on the GPU to do a rebase.
Improves nexuiz 800x600, high-settings performance on my Ironlake 41%
(+/- 1.3%), from 14.0fps to 19.7fps.
The format selection of the CopyTexSubImage is pretty bogus still, but
this at least avoids software fallbacks in nexuiz, bringing
performance from 7.5fps to 12.8fps on my machine.
This assertion was added in commit f1c1ee11, but it did not notice
that the array is accessed with 'size-1' instead of 'size'. As a
result, the assertion was off by one. This caused failures in at
least glsl-orangebook-ch06-bump.
If an GLSL shader is used that does not provide all stages and
assembly shaders are provided for the missing stages, validate the
assembly shaders.
Fixes bugzilla #30787 and piglit tests glsl-invalid-asm0[12].
NOTE: this is a candidate for the 7.9 branch.
There was actually a large quantity of scalar code in these functions
previously. This tries to move more into intrinsics.
Introduce an sse2 mm_mullo_epi32 replacement to avoid sse4 dependency
in the new rasterization code.
It's now much more correct for gen6 than the old backend, with just 2
regressions I've found (one of which is common with pre-gen6 and will
be fixed by an array splitting IR pass).
This does leave the old Mesa IR backend getting used still when we
don't have GLSL IR, but the plan is to get GLSL IR input to the driver
for the ARB programs and fixed function by the next release.
Pre-gen6, you could mix int and float just fine. Now, you get goofy
results.
Fixes:
glsl-arb-fragment-coord-conventions
glsl-fs-fragcoord
glsl-fs-if-greater
glsl-fs-if-greater-equal
glsl-fs-if-less
glsl-fs-if-less-equal
This is a hw requirement in math args. This also is inefficient, as
we're calculating the same result 8 times, but then we've been doing
that on pre-gen6 as well. If we're doing math on uniforms, though,
we'd probably be better served by having some sort of mechanism for
precalculating those results into another uniform value to use.
Fixes 7 piglit math tests.
Add ability to set the GLSL version used by the GLcontext by setting the
environment variable INTEL_GLSL_VERSION. For example,
env INTEL_GLSL_VERSION=130 prog args
If the environment variable is missing, the GLSL versions defaults to 120.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
By calling radeon_draw_buffers (which sets the necessary flags
in radeon->NewGLState) and revalidating if NewGLState is non-zero
in r200TclPrimitive. This fixes an assert in libdrm (the color-/
depthbuffer was changed but not yet validated) and and stops the
kernel cs checker from complaining about them (when they're too
small).
Thanks to Mario Kleiner for the hint to call radeon_draw_buffer
(instead of my half-broken hack).
v2: Also fix the swtcl r200 path.
Cc: Mario Kleiner <mario.kleiner@tuebingen.mpg.de>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This didn't produce a statistically significant performance difference
in my demo (n=4) or nexuiz (n=3), but it still seems like a good idea
and is recommended by the HW team.
Having the single opcode write then read the reg meant that single
instruction opcodes had to consider their source regs to interfere
with their dest regs.
SSE support for 32bit and 16bit unsigned arithmetic is not complete, and
can easily result in inefficient code.
In most cases signed/unsigned doesn't make a difference, such as for
integer texture coordinates.
So remove uint_coord_type and uint_coord_bld to avoid inefficient
operations to sneak in the future.
We need to move the texture sampler resources out of the range of the vertex attribs.
We could probably improve this using an allocator but this is the simple answer for now.
makes mesa-demos/src/glsl/vert-tex work.
We can't patch true-block at end-if time, as there is no guarantee that
the block at the beginning of the true stanza is the same at the end of
the true stanza -- other control flow elements may have been emitted half
way the true stanza.
Although this bug surfaced recently with the commit to skip mip filtering
when lod is an integer the bug was always there, although probably it
was avoided until now: e.g., cubemap selection nests if-then-else on the
else stanza, which does not suffer from the same problem.
We've been using these in the linear path for a while now. Based on
Chris's SSSE3 code, but using only sse2 opcodes. Speed seems to be
identical, but code is simpler & removes dependency on SSE3.
Should be easier to extend to other rgba8 formats.
Specifically, can do early-depth-test even when alpahtest or
kill-pixel are active, providing we defer the actual z write until the
final mask is avaialable.
Improves demos/fire.c especially in the case where you get close to
the trees.
The current interpolation schemes causes precision loss.
Changing the operation order helps, but does not completely avoid the
problem.
The only short term solution is to clamp z to 1.0.
This is unfortunate, but probably unavoidable until interpolation is
improved.
Operate simultanouesly on <width, height, depth> vector as much as possible,
instead of doing the operations on vectors with broadcasted scalars.
Also do the 24.8 fixed point scalar with integer shift of the texture size,
for unnormalized coordinates.
AoS path only for now -- the same thing can be done for SoA.
Fixes these GCC warnings.
brw_wm_fp.c: In function 'search_or_add_const4f':
brw_wm_fp.c:92: warning: 'reg.Index2' is used uninitialized in this function
brw_wm_fp.c:84: note: 'reg.Index2' was declared here
brw_wm_fp.c:92: warning: 'reg.RelAddr2' is used uninitialized in this function
brw_wm_fp.c:84: note: 'reg.RelAddr2' was declared here
Clamp against 0 instead of -0.5, which simplifies things.
The former version would have resulted in both int coords being zero
(in case of coord being smaller than 0) and some "unused" weight value,
whereas now the int coords will be 0 and 1, but weight will be 0, hence the
lerp should produce the same value.
Still not happy about differences between normalized and non-normalized...
Haven't looked at what code this exactly generates but URem can't be fast.
Instead of using two URem only use one and replace the second one with
select/add (this is what the corresponding aos code already does).
Rearrange order of operations a bit to make some clamps easier.
All calculations should be equivalent.
Note there seems to be some inconsistency in the clamp to edge case
wrt normalized/non-normalized coords, could potentially simplify this too.
Sometimes coords are clamped to positive numbers before doing conversion
to int, or clamped to 0 afterwards, in this case can use itrunc
instead of ifloor which is easier. This is only the case for nearest
calculations unfortunately, except linear MIRROR_CLAMP_TO_EDGE which
for the same reason can use a unsigned float build context so the
ifloor_fract helper can reduce this to itrunc in the ifloor helper itself.
sse2 supports round to nearest directly (or rather, assuming default nearest
rounding mode in MXCSR). Use intrinsic to use this rather than round (sse41)
or bit manipulation whenever possible.
The constness of the function parameter gets inlined with the rest of
the function. However, there is also an assignment to the parameter.
If this occurs inside a loop the loop analysis code will get confused
by the assignment to a read-only variable.
Fixes bugzilla #30552.
NOTE: this is a candidate for the 7.9 branch.
Improves performance of my GLSL demo 14.3% (+/- 4%, n=4) by
eliminating the moves used in ir_assignment and ir_swizzle handling.
Still 16.5% to go to catch up to the Mesa IR backend, presumably
because instructions are almost perfectly mis-scheduled now.
We were trying to remap a fully-filled array down to only handing the
WM the components it uses. This is called attribute swizzling, and if
you don't enable it you just get 1:1 mappings of inputs to outputs.
This almost fixes glsl-routing, except for the highest gl_TexCoord[]
indices.
We would compute a new buffer, but never point the hardware at the new
buffer. This partially fixes glsl-routing, as now it get the updated
uniform for which attribute to draw.
Avoid multiplying fixed-point values. Calculate triangle area in
floating point use that for culling.
Lift area calculations up a level as we are already doing this in the
triangle_both() case.
Would like to share the calculated area with attribute interpolation,
but the way the code is structured makes this difficult.
interp data is stored in gpr0 so first interp overwrote it
and subsequent ones got wrong values
reserve register 0 so it's not used for attribs.
alternative is to interpolate attrib0 last (reverse, as r600c does)
We sensibly only provide it if the FS asks for it. We could actually
skip WPOS unless the FS needed WPOS.zw, but that's something for
later.
Fixes: glsl-texture2d and probably many others.