The motivation is to kill tiling and pitch in struct intel_bo. That requires
us to make tiling and pitch not queryable, and be passed around as function
parameters.
See two commits ago for the rationale. This allows us to delete the
whole gen7_cc_state.c file.
This does move these commands before the depth stall flushes from
brw_emit_depthbuffer, which may be a problem. The documentation for
3DSTATE_DEPTH_BUFFER mentions that depth stall flushes are required
before changing any depth/stencil buffer state, but explicitly lists
3DSTATE_DEPTH_BUFFER, 3DSTATE_HIER_DEPTH_BUFFER, 3DSTATE_STENCIL_BUFFER,
and 3DSTATE_CLEAR_PARAMS. It does not mention this particular packet
(_3DSTATE_DEPTH_STENCIL_STATE_POINTERS).
No observed Piglit regressions on Sandybridge or Ivybridge.
Together with the last two commits, this makes a cairo-gl benchmark
faster by 0.324552% +/- 0.258355% on Ivybridge. No statistically
significant change on Sandybridge. (Thanks to Eric for the numbers.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, we would:
1. Emit the new indirect state.
2. Flag CACHE_NEW_BLEND_STATE.
3. Rely on later state atoms to notice CACHE_NEW_BLEND_STATE and emit a
pointer to the new indirect state.
This is rather cumbersome: it requires two state atoms instead of one,
and there's a strict ordering dependency in the list. Plus, the code
gets spread across two functions (or even files in the case of Gen7+).
Gen7+ has a packet to update just the blend state pointer, so it makes a
lot of sense to simply emit that right away. Gen6 has a combined packet
which updates blending, the color calculator, and depth/stencil state;
however, each can still be modified independently.
This drops the Gen6 micro-optimization where we tried to only emit one
packet that changed all three states. State updates are pretty cheap.
CACHE_NEW_BLEND_STATE is no longer necessary, so drop it.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Works similarly to clip distance. If the cull distance is negative
for all vertices against a specific plane then the primitive
is culled.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
cull distance is analogous to clip distance. If a register is
given this semantic, then the values in it are assumed to be a
float32 distance to a plane. Primitives will be completely
discarded if the plane distance for all of the vertices in
the primitive are < 0.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
We need to figure out the number of invocations of the clipper
before the emit, because in the emit we are after clipping
where the number of primitives will be equal to number of clipper
invocations minus the clipped primitives. So our computations
were always off by the number of clipped primitives.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Draw depended on clip_plane_enable being set in the rasterizer
to use clipdistance registers for clipping. That's really
unfriendly because it requires that rasterizer state to have
variants for every shader out there. Instead of depending on
the rasterizer lets extract the info from the available state:
if a shader writes clipdistance then we need to use it and we
need to clip using a number of planes equal to the number
of writen clipdistance components. This way clipdistances
just work.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
we were always fetching the info from the vertex shader, but if
geometry shader is present it should be used as the source of
that info.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Fixes piglit texture-packed-formats regression. We need to implement
more XBGR formats here eventually, but many are UINT/SINT formats
which swrast doesn't handle yet anyway (integer textures).
Bugzilla https://bugs.freedesktop.org/show_bug.cgi?id=64935
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Set env var RADEON_VA=0 to disable VM on Cayman/Trinity.
Useful for debugging.
Note: this is a candidate for the 9.1 branch.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Marek Olšák <maraeo@gmail.com>
Having figured out what was going on with piglit fbo-depth copypixels
GL_DEPTH_COMPONENT32F (falling all the way back to swrast on CopyPixels to
a float depth buffer), I'm not inclined to fix the problem currently but
it seems worth saving someone else the debug time.
Reviewed-by: Matt Turner <mattst88@gmail.com>
We do a lot of multiplies by 3 or 4 for skinning shaders, and we can avoid
the sequence if we just move them into the right argument of the MUL.
On pre-IVB, this means reliably putting a constant in a position where it
can't be constant folded, but that's still better than MUL/MACH/MOV.
Improves GLB 2.7 trex performance by 0.788648% +/- 0.23865% (n=29/30)
v2: Fix test for pre-sandybridge.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com> (v1)
This is a trivial port of 1d6ead3804 from
the FS.
No significant performance difference on trex (misplaced the data, but it
was about n=20).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This is different from how we do it in the FS - we are using MAD even when
some of the args are constants, because with the relatively unrestrained
ability to schedule a MOV to prepare a temporary with that data, we can
get lower latency for the sequence of instructions.
No significant performance difference on GLB2.7 trex (n=33/34), though it
doesn't have that many MADs. I noticed MAD opportunities while reading
the code for the DOTA2 bug.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
draw_vertex_buffer declared the size field to be a size_t, but the LLVM
code used an int32 instead. This caused problems on big-endian 64-bit
targets, because the first 32-bit chunk of the 64-bit size_t was always 0.
In one sense size_t seems like a good choice for a size, so one fix
would have been to try to get the LLVM code to use the equivalent of
size_t too. However, in practice, the size is taken from things like ~0
or width0, both of which are int-sized, so it seemed simpler to make the
size field int-sized as well.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <rsandifo@linux.vnet.ibm.com>
Without this, llvmpipe ends up giving a zero size to all uncompressed textures
on non-x86 systems, since align() cannot handle a 0 alignment.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <rsandifo@linux.vnet.ibm.com>
lp_build_add and lp_build_sub have fallback code for cases
that cannot be handled by known intrinsics. For UNORM formats,
this code was using modulo rather than saturating arithmetic.
This fixes some rendering issues for a gnome session on System z.
It also fixes various piglit tests on z, such as
spec/ARB_color_buffer_float/GL_RGBA8-render.
The patch deliberately doesn't tackle the more complicated
SNORM case.
Tested against piglit on x86_64 and System z with no regressions.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <rsandifo@linux.vnet.ibm.com>
Now that Gen6+ relies on hardware contexts, we don't need to record an
occlusion query value at the end of each batch. That means we no longer
need to reserve space for the absurd number of PIPE_CONTROLs required to
do that on Sandybridge.
See commit 4e087de51a, which bumped this
up to 60 bytes. This is not quite a revert, as it uses 24 bytes instead
of 16, and saves the comments. As far as I can tell, the old value of
16 bytes was just wrong, so we shouldn't go back to that.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
We always allocate the maximum amount of space and never change it, so
it makes sense to do it once. Programming it on startup also lets us
skip re-programming it from BLORP.
This removes a tiny amount of overhead from our drawing loop.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
This removes a tiny bit of code from our drawing loop.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Now that we emit invariant state at startup (and never select the media
pipeline), the 3D pipeline will always already be selected, even if BLORP
is the first operation. So this is unnecessary.
v2: Fix unused variable warning (intel_context is no longer used).
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Now that we have hardware contexts, we can safely initialize our GPU
state once at startup, rather than needing a state atom with the
BRW_NEW_CONTEXT flag set.
This removes a tiny bit of code from our drawing loop.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
The existing code already returned a boolean; this just clarifies that.
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
We already implemented this for ES3, so we just need to turn it on.
Fixes 6 Piglit tests:
spec/glsl-1.50/compiler/built-in-functions/determinant-mat[234].{vert,frag}
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Page 17 of the GLSL 1.50.11 specification states:
"There is a built-in macro definition for each profile the
implementation supports. All implementations provide the following
macro:
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Previously we only supported "#version 150". This patch recognizes
"compatibility" to give the user a more descriptive error message.
Fixes Piglit's version-150-core-profile test.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
If we didn't successfully parse the #version line, there's no point in
continuing with parsing and compiling: it's already failed.
Furthermore, it can actually be harmful: right after handling #version,
we call _mesa_glsl_initialize_types(), which checks state->es_shader and
language_version. If it isn't valid, it hits an assertion failure.
Fixes Piglit's "invalid-version-es." When processing "#version 110 es",
our code set state->es_shader and state->language_version = 110. It
then properly determined that this was invalid and flagged an error.
Since we continued anyway, we hit the assertion mentioned above.
NOTE: This is a candidate for the 9.1 branch.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
This was building the temporary array to pass to
save_SamplerParameteriv, and then not passing it.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Vinson Lee <vlee@freedesktop.org>
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
The GPU (at least a3xx, but I think also a2xx) can render directly to
memory, bypassing tiling. Although it can't do this if blend, depth,
and a few other features of the pipeline are enabled. This direct
memory mode can be faster for some sorts of operations, such as simple
blits. In particular, this significantly speeds up XA by avoiding to
pull the entire dest pixmap into GMEM, render tiles, and write it all
back out again. This should also speed up resource copy-region and
blit.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The adreno a3xx GPU is found in newer snapdragon devices, such as the
nexus4. The a3xx is GLESv3 and OpenCL capable, although that is not
enabled yet in gallium.
Compared to a2xx, it introduces an entirely new unified shader ISA, and
re-shuffles all or nearly all of the registers. The good news is that
(for the most part) the registers are more orthogonal, not combining
unrelated state in a single register. And that there is a lot more
flexibility, so we don't need to patch and re-emit the shader like we
did on a2xx.
The shader compiler is currently quite dumb, there would be a lot of
room for improvement with an optimizing pass. Despite that, with the
a320 in my nexus4 it seems to be ~2-3x faster compared to the a220 in my
HP touchpad.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Split the parts that are specific to adreno a2xx series GPUs from the
parts that will be in common with a3xx, so that a3xx support can be
added more cleanly.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We use 128bit vector interleave for untwiddling in the blend code (with
256bit vectors). llvm generates terrible code for this for some reason,
so instead of generating a shuffle for 2 128bit vectors use a
extract/insert shuffle instead (it only seems to matter we're not using
128bit wide vectors for the shuffle). This decreases instruction count of
the blend code generated for a rgba8 render target without blending from
169 to 113 with llvm 3.1 and from 136 to 114 in llvm 3.2/3.3, and I got
a ~8% (llvm 3.1) and ~5% (3.2/3.3) performance improvement in gears.
(The generated code is still not terribly good as we could actually avoid
the interleaving completely but llvm can't know this.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Probably due to CRLF endings, the discovery of python import statements
was not working on Windows builds, causing incremental builds to often
fail unless one wiped out the build directory.
NOTE: This is a candidate for stable branches.
We flush pending rendering before running CopySubBuffer, which
ensures that the right bits get to the screen.
NOTE: This is a candidate for stable release branches.
Reviewed-by: Brian Paul <brianp@vmware.com>
The coordinates need to be inverted between glX and gallium.
NOTE: This is a candidate for stable release branches.
Reviewed-by: Brian Paul <brianp@vmware.com>
Just like we produce from inside the Intel driver, this can help provide
information quickly about FBO incompatibility problems (particularly when
using apitrace replay).
Currently, in driver-marked incompleteness cases, you'll get both the
driver message and the core message on Intel. Until the other drivers are
fixed to produce output, I think this is better than not putting in a
message for driver-marked incomplete.
Reviewed-by: Brian Paul <brianp@vmware.com>
