We had five copies of the same "walk the cache and look for an
existing shader variant for this program" code. Now we have one
helper function that returns the key.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
A number of games have large arrays of constants, which we promote to
uniforms. This introduces copies from the uniform array to the original
temporary array. Normally, copy propagation eliminates those copies,
making everything refer to the uniform array directly.
A number of shaders in "Deus Ex: Mankind Divided" recently exposed a
limitation of copy propagation - if we had any intrinsics (i.e. image
access in a compute shader), we weren't able to get rid of these copies.
That meant that any variable indexing remained on the temporary array
rather being moved to the uniform array. i965's scalar backend
currently doesn't support indirect addressing of temporary arrays,
which meant lowering it to if-ladders. This was horrible.
According to Marek, on radeonsi/GCN, "F1 2015" uses 64% less
spilled-temp-array memory.
On i965/Skylake:
total instructions in shared programs: 13362954 -> 13329878 (-0.25%)
instructions in affected programs: 43745 -> 10669 (-75.61%)
helped: 12
HURT: 0
total cycles in shared programs: 248081010 -> 245949178 (-0.86%)
cycles in affected programs: 4597930 -> 2466098 (-46.37%)
helped: 12
HURT: 0
total spills in shared programs: 9493 -> 9507 (0.15%)
spills in affected programs: 25 -> 39 (56.00%)
helped: 0
HURT: 1
total fills in shared programs: 12127 -> 12197 (0.58%)
fills in affected programs: 110 -> 180 (63.64%)
helped: 0
HURT: 1
Helps Deus Ex: Mankind Divided. The one shader with hurt spills/fills
is from Tomb Raider at Ultra settings, but that same shader has a
-39.55% reduction in instructions and -14.09% reduction in cycle counts,
so it seems like a win there as well.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
(Co-authored by Matt Turner.)
Image atomics, for example, return a value - but the shader may not
want to use it. We assigned a useless VGRF destination. This seemed
harmless, but it can actually be quite harmful. The register allocator
has to assign that VGRF to a real register. It may assign the same
actual GRF to the destination of an instruction that follows soon after.
This results in a write-after-write (WAW) dependency, and stall.
A number of "Deus Ex: Mankind Divided" shaders use image atomics, but
don't use the return value. Several of these were hitting WAW stalls
for nearly 14,000 (poorly estimated) cycles a pop. Making dead code
elimination null out the destination avoids this issue.
This patch cuts one shader's estimated cycles by -98.39%! Removing the
message response should also help with data cluster bandwidth.
On Skylake:
(instruction counts remain identical)
total cycles in shared programs: 255413890 -> 248081010 (-2.87%)
cycles in affected programs: 12019948 -> 4687068 (-61.01%)
helped: 24
HURT: 10
v2: Make can_omit_write independent of can_eliminate (Curro).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
In theory we might have incorrectly NOP'd instructions that write the
flag, but where that flag value isn't used, and yet the instruction
either writes the accumulator or has side effects.
I don't believe any such instructions exist, so this is mostly a
code cleanup.
Curro pointed out that FS_OPCODE_FB_WRITE has a null destination and
actually writes the flag on Gen4-5 to dynamically decide whether to
write some payload data. The hunk removed in this patch might have
NOP'd it, except that we don't actually mark flags_written() in the
IR, so it doesn't think the flag is touched at all. That's sketchy,
but it means it wouldn't hit this today (though there are likely other
problems!).
v2: Properly replace the inst->regs_written() check in the second
hunk with the flag being live (mistake caught by Curro).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
According to Matt, the dead code pass explicitly avoided IF and WHILE
because on Sandybridge, these could have conditional modifiers and
null destination registers. Normally, those instructions use BAD_FILE
for the destination register. Nowadays, we don't do that anymore, so
we could technically drop these checks.
However, it's clearer to explicitly leave control flow instructions
alone, so change it to the more generic !inst->is_control_flow().
This should have no actual change.
[This patch implements review feedback from Curro and Matt.]
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This fixes some issues we'd hit later if using viewport
indexes.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Just always use the layer clear pipelines,
the overhead of emitting the layer shouldn't be
too large.
v2: Bas suggested we always use it.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is needed to have common code for gs copy shader emission.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
makes it easier to add other shader stages.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This just adds the infrastructure to allow writing layer
and viewport index. It's just a first patch out of the geom
shader tree, and doesn't do much on its own.
v2: add missing if statement change (Bas)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Otherwise we read past the end of the buffer.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
the hack was introduced to avoid an extra copying
but now with dri3 we don't need it anymore
v1.1: rebasing
Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
Acked-by: Christian König <christian.koenig@amd.com>
this avoids an extra copy which occurs in case of dri2
v1.1: fallback to dri2 if dri3 fails to initialize
v2: add PIPE_BIND_SCANOUT to output buffers as they will
be send to X server directly (Michel)
Suggested-by: Christian König <christian.koenig@amd.com>
Tested-by: Andy Furniss <adf.lists@gmail.com>
Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
dri3 allows us to send handle of a texture directly to X
so this patch allows a state tracker to directly send its
texture to X to be used as back buffer and avoids extra
copying
v2: use clip width/height to display a portion of the surface
v3: remove redundant variables, fix wrapping, rename variables
handle vaapi path
v3.1: we need clip_width/height for every frame so we don't need
to maintain it for each buffer instead use a global variable
v4: In case of single gpu we can cache the buffers as applications
use constant number of buffer and we can avoid calls to present
extension for every frame
Reviewed and Suggested-by: Leo Liu <leo.liu@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Tested-by: Andy Furniss <adf.lists@gmail.com>
Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
The same we do in the OpenGL driver (comment copied from there).
This is required to ensure that we execute the fragment shader stage when
side-effects (such as image or ssbo stores) are present but there are no
color writes.
I found this while writing a test to check rendering to a framebuffer
without attachments where the fragment shader does not produce any
color outputs but writes to an image via imageStore(). Without this patch
the fragment shader does not execute and the image is not written,
which is not correct.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Fixes crash in dEQP-VK.ubo.random.all_shared_buffer.48 due to a
fragment shader code bigger than 128 kB.
This patch increases the allocation size limit to 1 MB.
v2:
- Increase it to 1 MB (Jason)
- Increase device->instruction_block_pool allocation size in
anv_device.c (Jason)
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Address loading can often end up as shl + shr + shl combinations. The
latter two are equal shifts, which get converted into an and mask.
However if the previous shl is more than the mask is trying to remove
(in terms of low bits), we can just remove the and entirely. This
reduces some large shaders by as many as 3% of instructions (out of 2K).
total instructions in shared programs : 6495509 -> 6491076 (-0.07%)
total gprs used in shared programs : 954621 -> 954623 (0.00%)
local gpr inst bytes
helped 0 0 1014 1014
hurt 0 2 0 0
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
We don't need to support all the color buffers for advanced blend, just
cb0. For Fermi, we use the special binding slots so that we don't
overlap with user textures, while Kepler+ gets a dedicated position for
the fb handle in the driver constbuf.
This logic is only triggered when a FBFETCH is actually present so it
should be a no-op most of the time.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This implements support for emitting FBFETCH ops, using the existing
lowering pass for advanced blend logic, and disabling hw blend when
advanced blending is enabled.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This is so that we can differentiate between flushing any framebuffer
reading caches from regular sampler caches.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The extension spec is not currently published, so it's a bit premature
to require it for BlendBarrier usage.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This is just prep work for layered clears, it doesn't change
anything.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Pretty straightforward. Also deleted the big comment block as it
is a pretty standard pattern for filling in arrays.
Also removed the error message on non-existent devices, as getting
7 errors printed to the console each time you enumerate the
devices is pretty confusing.
v2: Add constant for number of DRM devices.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
This had been updated in one place but not the other.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Just noticed this while in the area.
v2: one replacement was incorrect.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We only need one per samples (maybe not even that), reduce
all the unneeded ones.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The existing lowering is in place to lower that to RCP + MUL, or fancier
things down the line if necessary.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Double-precision division, to allow more precision than a DRCP + DMUL
sequence.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This may fix GPU hangs on Gen8. I don't know if it does though.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Everything is in place and the test results look solid.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Renaming data sources was added in
e8bb97ce30
It was possible to use a new name longer than
the name array in hud_graph of 128. This
patch truncates the name to fit the array.
CC: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
It should be close to the GPU load, but it can be much lower if something
is stalling shader execution (e.g. CP DMA).
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
If begin_frame is called before setting intra_matrix and
non_intra_matrix it leads to segmentation faults when
vl_mpeg12_decoder.c is used.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92634
Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
In commit 7428e6f86a we switched the barrier SEND message's
destination type to UW to avoid problems in SIMD16 compute shaders.
Tessellation control shaders also use barriers, and in vec4 mode, we
were emitting them in align16 mode. The simulator warns that only UD,
D, F, and DF are valid destination types - UW is technically illegal.
So, switch to align1 mode. Either mode should work fine.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
We were using impl->num_blocks, but that isn't guaranteed to be
up-to-date until after the block_index metadata is required. If we were
unlucky, this could lead to overwriting memory.
Noticed by inspection.
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
