brw_swizzle_to_scs has been showing up in my CPU profiling, which is
rather silly - it's a tiny amount of code. It really should be inlined,
and can easily be implemented with fewer instructions.
The enum translation is as follows:
SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W, SWIZZLE_ZERO, SWIZZLE_ONE
0 1 2 3 4 5
4 5 6 7 0 1
SCS_RED, SCS_GREEN, SCS_BLUE, SCS_ALPHA, SCS_ZERO, SCS_ONE
which is simply (swizzle + 4) & 7.
Haswell needs extra textureGather workarounds to remap GREEN to BLUE,
but Broadwell and later do not.
This patch replicates swizzle_to_scs in gen7_wm_surface_state.c and
gen8_surface_state.c, since the Gen8+ code can be simplified to a mere
two instructions. Both copies can be marked static for easy inlining.
v2: Put the commit message in the code as comments (requested by
Jason Ekstrand). Also fix a typo.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Valve games use GL_SRGB8 textures. Instead of supporting that properly,
we fell back to MESA_FORMAT_R8G8B8A8_SRGB (with an alpha channel), which
meant that we had to use texture swizzling to override the alpha to 1.0
when sampling. This meant shader recompiles on Gen < 7.5 platforms.
By supporting MESA_FORMAT_R8G8B8X8_SRGB, the hardware just returns 1.0
for us, so we can just use SWIZZLE_XYZW, and avoid any recompiles. All
generations of hardware have supported the format for sampling and
filtering; we can easily support rendering by using the R8G8B8A8_SRGB
format and writing garbage to the X channel. (We do this already for
the non-SRGB version of this format.)
This removes all remaining shader recompiles in a time demo of "Counter
Strike: Global Offensive" (32 -> 0) on Sandybridge.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=87886
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
The logic in brw_blorp_surface_info::set uses brw_format_for_mesa_format
for source surfaces, and brw->render_target_format[] for destination
surfaces. We should do the same in the sRGB MSAA overrides.
Currently, this isn't a problem, since SRGB MSAA buffers are all RGBA.
The next commit will introduce RGBX SRGB MSAA buffers, at which point
we need to get the RGBX -> RGBA format overrides for rendering right.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
ir_to_mesa does this - apparently we just forgot or something.
Without this, we'll guess the wrong texture swizzle (XYZW for color
instead of XXX1 for depth) when doing precompiles.
This cuts 26 shader recompiles in a time demo of "Counter Strike:
Global Offensive" (58 -> 32) on Sandybridge. Haswell still has 0
recompiles.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=87886
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Gen7.5+ platforms that support the "Shader Channel Select" feature leave
key->tex.swizzles[i] as SWIZZLE_NOOP except when GL_DEPTH_TEXTURE_MODE
is GL_ALPHA (which is really uncommon). So, the precompile should leave
them as SWIZZLE_NOOP (aka SWIZZLE_XYZW) as well.
We didn't notice this because prog->ShadowSamplers is not set correctly.
The next patch will fix that problem.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=87886
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
According to the documentation, we need to do a CS stall on every fourth
PIPE_CONTROL command to avoid GPU hangs. The kernel does a CS stall
between batches, so we only need to count the PIPE_CONTROLs in our batches.
v2: Get the generation check right (caught by Chris Wilson),
combine the ++ with the check (suggested by Daniel Vetter).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
There are too many state flags to fit in one terminal screen, even with
a very tall terminal. Everything is flagged once, so a value of 1 means
that it hasn't ever happened again, and thus isn't terribly interesting.
Skipping those makes it easier to see the interesting values.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The mad instruction emitter already supported the saturate modifier,
but the ModifierFolding pass never tried folding cvt sat operations
in for NV50.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
This is a partial revert of c89306983c.
It split the {start,base}_vertex_location handling into several steps:
1. Set brw->draw.start_vertex_location = prim[i].start
and brw->draw.base_vertex_location = prim[i].basevertex.
(This happened once per _mesa_prim, in the main drawing loop.)
2. Add brw->vb.start_vertex_bias and brw->ib.start_vertex_offset
appropriately. (This happened in brw_prepare_shader_draw_parameters,
which was called just after brw_prepare_vertices, as part of state
upload, and only happened when BRW_NEW_VERTICES was flagged.)
3. Use those values when emitting 3DPRIMITIVE (once per _mesa_prim).
If we drew multiple _mesa_prims, but didn't flag BRW_NEW_VERTICES on
the second (or later) primitives, we would do step #1, but not #2.
The first _mesa_prim would get correct values, but subsequent ones
would only get the first half of the summation.
The reason I originally did this was because I needed the value of
gl_BaseVertexARB to exist in a buffer object prior to uploading
3DSTATE_VERTEX_BUFFERS. I believed I wanted to upload the value
of 3DPRIMITIVE's "Base Vertex Location" field, which was computed
as: (prims[i].indexed ? prims[i].start : prims[i].basevertex) +
brw->vb.start_vertex_bias. The latter value wasn't available until
after brw_prepare_vertices, and the former weren't available in the
state upload code at all. Hence the awkward split.
However, I believe that including brw->vb.start_vertex_bias was a
mistake. It's an extra bias we apply when uploading vertex data into
VBOs, to move [min_index, max_index] to [0, max_index - min_index].
>From the GL_ARB_shader_draw_parameters specification:
"<gl_BaseVertexARB> holds the integer value passed to the <baseVertex>
parameter to the command that resulted in the current shader
invocation. In the case where the command has no <baseVertex>
parameter, the value of <gl_BaseVertexARB> is zero."
I conclude that gl_BaseVertexARB should only include the baseVertex
parameter from glDraw*Elements*, not any internal biases we add for
optimization purposes.
With that in mind, gl_BaseVertexARB only needs prim[i].start or
prim[i].basevertex. We can simply store that, and go back to computing
start_vertex_location and base_vertex_location in brw_emit_prim(), like
we used to. This is much simpler, and should actually fix two bugs.
Fixes missing geometry in Unvanquished.
Cc: "10.4 10.3" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=85529
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
v2: Conditionalize it on having done any uploads (Turns out
u_upload_destroy() isn't safe with a NULL arg).
Reviewed-by: Dave Airlie <airlied@redhat.com> (v1)
Fixes the piglits which check that gl_VertexID includes the base vertex
offset:
arb_draw_indirect-vertexid elements
gl-3.2-basevertex-vertexid
Note that this leaves out the original G80, for which this will continue
to fail. It could be fixed by passing a driver constbuf value in, but
that's beyond the scope of this change.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.3 10.4" <mesa-stable@lists.freedesktop.org>
LAST_LINE_PIXEL has actually been renamed to PIXEL_CENTER_INTEGER in
rnndb; use that method to implement the rasterizer setting, used for
st/nine.
Signed-off-by: Tiziano Bacocco <tizbac2@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
total instructions in shared programs: 5877012 -> 5876617 (-0.01%)
instructions in affected programs: 33140 -> 32745 (-1.19%)
From before the commit that allows VF constant propagation (which hurt
some programs) to here, the results are:
total instructions in shared programs: 5877951 -> 5876617 (-0.02%)
instructions in affected programs: 123444 -> 122110 (-1.08%)
with no programs hurt.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
total instructions in shared programs: 5877951 -> 5877012 (-0.02%)
instructions in affected programs: 155923 -> 154984 (-0.60%)
Helps 1233, hurts 156 shaders. The hurt shaders are addressed in the
next commit.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
After CSEing some MOV ..., VF instructions we have code like
mov tmp, [1F, 2F, 3F, 4F]VF
mov r10, tmp
mov r11, tmp
...
use r10
use r11
We want to copy propagate tmp into the uses of r10 and r11, but *not*
constant propagate the VF immediate into the uses of tmp.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Port of commit a28ad9d4 from the fs backend.
No shader-db changes since we don't emit MOV ..., VF instructions yet.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Currently only handles consecutive instructions with the same
destination that collectively write all channels.
total instructions in shared programs: 5879798 -> 5869011 (-0.18%)
instructions in affected programs: 465236 -> 454449 (-2.32%)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
I don't feel great about assert(!"unimplemented: ...") but these
cases do only seem possible under some currently impossible circumstances.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Sometimes it's easier to generate 4x values into an array, and the
memcpy is 1 instruction, rather than 11 to piece 4 arguments together.
I'd forgotten to remove the prototype from fs_reg from a previous patch,
so it's already there for us here.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
As of 229bf4475f we started getting SIBGUS
from unaligned accesses on the hardware, for reasons I haven't figured
out. However, we should be avoiding unaligned accesses anyway, and our CL
setup certainly would have produced them.
E.g. this could happen on older kernels which don't support the
RADEON_INFO_SI_BACKEND_ENABLED_MASK query yet. The code in
si_write_harvested_raster_configs() doesn't deal with this correctly and
would probably mangle the value badly.
Cc: "10.4 10.3" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
The optimizer obviously doesn't have the ability to rewrite these to skip
the size checks per call, so we have to do it manually.
Improves a norast benchmark on simulation by 0.779706% +/- 0.405838%
(n=6087).
Our ability to perform register writes depends on the hardware and
kernel version. It shouldn't ever change on a per-context basis,
so we only need to check once.
Checking introduces a synchronization point between the CPU and GPU:
even though we submit very few GPU commands, the GPU might be busy doing
other work, which could cause us to stall for a while.
On an idle i7 4750HQ, this improves performance in OglDrvCtx (a context
creation microbenchmark) by 6.14748% +/- 1.6837% (n=20). With Unigine
Valley running in the background (to keep the GPU busy), it improves
performance in OglDrvCtx by 2290.92% +/- 29.5274% (n=5).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Kenneth Graunke