Instead of doing a add and then mask out the upper bits, we can
simply do a add with a half wide type (this, of course, assumes
the hw can actually do it...), so we'll get the required zero
in the upper bits automatically.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
There were complaints from a mingw build:
u_draw.h:134:14: error: invalid conversion from ‘uint {aka unsigned int}’
to ‘pipe_prim_type’ [-fpermissive]
Reviewed-by: Brian Paul <brianp@vmware.com>
For a load locked, we might not use the first result but the second
result is the predicate result of the locking. In that case the load
splitting logic doesn't apply (which is designed for splitting 128-bit
loads). Instead we take the predicate and move it into the first
position (as having a dead result in first def's position upsets all
sorts of things including RA). Update the emitters to deal with this as
well.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Tested-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
For user-supplied constbufs, fileIndex is 0. In that case, when we
subtract 1, we'll end up loading from constbuf offset -16. This is
illegal, and there are asserts to avoid it. Normally we'd just DCE it,
but no point in generating the instructions if they're not going to be
used.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Acked-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested with new piglit gl-3.2-adj-prims test.
v2: re-order trisadj and tristripadj code, per Roland.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The unfilled index translator/generator functions should only be
called when the primitive mode is one of the triangle types.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The original mode test was valid before we had GS support.
Regression tested with full piglit run. Though, I don't think we have
any piglit tests that exercise drawing unfilled adjacency primitives.
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Like constant buffers, samplers and textures are aliased on Fermi and
we need to invalidate the state when switching from 3D to CP and vice
versa.
This fixes rendering issues in the UE4 demos.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Right now libglsl.la depends on libnir.la so putting it in libnir.la
adds a dependency on libglsl.la that goes the wrong direction.
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
The R_028B50_VGT_TESS_DISTRIBUTION value is copied from
amdgpu-pro. Smaller values in the ACCUM fields seem to
decrease the performance advantage from this patch, higher
values don't seem to matter.
v2: Add distribution mode field enums.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Using more than 1 wave per threadgroup does increase performance
generally. Not using too many patches per threadgroup also
increases performance. Both catalyst and amdgpu-pro seem to
use 40 patches as their maximum, but I haven't really seen
any performance increase from limiting the number of patches
to 40 instead of 64.
Note that the trick where we overlap the input and output LDS
does not work anymore as the insertion of the tess factors
changes the patch stride.
v2: - Add comment about LDS assumptions.
- Add constant for buffer size.
- Fix code style.
v3: - Correct limits for not splitting patches between waves.
- Set max num_patches to 40 as in the proprietary driver.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The factors may be stored to LDs by another invocation than
the invocation for vertex 0.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This allows running the TES on different CU's than the
TCS which results in performance improvements.
v2: Only write the control word from one invocation.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
They are unused.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
We always try to use 4-component loads, as LLVM does not combine loads
and they bypass the L1 cache.
We can't use a similar strategy for stores and this is especially
notable with the tess factors, as they are often set with separate
MOV's per component in the TGSI.
We keep storing to LDS and the LDS space, so we can load the outputs
later, either due to the shader, of for wrting the tess factors.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
We need to copy the VS outputs to memory. I decided to do this
using a shader key, as the value depends on other shaders.
I also switch the fixed function TCS over to monolithic, as
otherwisze many of the user SGPR's need to be passed to the
epilog, which increases register pressure, or complexity to
avoid that. The main body of the fixed function TCS is not
that interesting to precompile anyway, since we do it on
demand and it is very small.
v2: Use u_bit_scan64.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Instead of creating a memory area per patch and per vertex, we put
the same attribute of every vertex & patch together. Most loads
and stores access the same attribute across all lanes, only for
different patches and vertices.
For the TCS this results in tightly packed data for 4-component
stores.
For the TES this is not the case as within a patch the loads
often also access the same vertex. However if there are < 4
vertices/patch, this still results in a reduction of the number
of cache lines. In the LDS situation we only do better than worst
case if the data per patch < 64 bytes, which due to the
tessellation factors is pretty much never.
We do not use hardware swizzling for this. It would slightly reduce
the number of executed VALU instructions, but I had issues with
increased wait times that I haven't been able to solve yet.
Furthermore, the tbuffer_store intrinsic does not support both
VGPR offset and an index, so we have a problem storing
indirectly indexed outputs. This can be solved by temporarily
storing arrays in LDS and then copying them, but I don't think
that is worth the effort. The difference in VALU cycles
hardware swizzling gives is about 0.2% of total busy cycles.
That is without handling the array case.
I chose for attributes instead of components as they are often
accessed together, and the software swizzling takes VALU cycles
for calculating offsets.
v2: - Rename functions to get_tcs_tes_buffer_address.
- multiply by 16 as late as possible.
- Use tgsi_full_src_register_from_dst.
- Remove some bad comments.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This happens to be in the right position, but that changes
when TCS/TES get new parameters.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
v2: - Use llvm.admgcn.buffer.load instrinsics for new LLVM.
- Code style fixes.
v3: - Code style fix.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The buffer is quite large, but should only be allocated if the
application uses tessellation. Most non-games don't.
v2: - Use the correct register for SI.
- Add define for block size.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
CID 1271532 (#1 of 1): Out-of-bounds read (OVERRUN)34. overrun-local:
Overrunning array of 2 16-byte elements at element index 2 (byte offset
32) by dereferencing pointer &inst.Dst[i].
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Not sure why coverity calls this an out-of-bounds read vs out-of-bounds
write.
CID 1358920 (#1 of 1): Out-of-bounds read (OVERRUN)9. overrun-local:
Overrunning array r of 3 16-byte elements at element index 3 (byte
offset 48) using index chan (which evaluates to 3).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
To read out MP perf counters we use a compute shader and need to upload
input data like a 64-bits addr used to store the values and a sequence
ID for synchronization. Currently, this input data is uploaded as user
uniforms which means that it's sticked to c0[], but if a compute shader
from a real application is used, monitoring those performance counters
will just overwrite some data and miserably crash.
Instead, sticking the 64-bits addr and the sequence into the driver
constant buffer seems like much better and will allow to monitor
counters with GL 4.3 apps.
Tested on GF119 and GK110, but should not hurt anything on GK104.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Example:
Gallium 0.4 on AMD TONGA (DRM 3.2.0 / 4.5.0, LLVM 3.9.0)
My kernel version is pretty long already (4.5.0-amd-01025-g32791c1)
and adding "kernel" into the string would make too it long for glxinfo
to display.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Ported from the initial amdgpu winsys from the private AMD branch.
The thread creates the buffer list, submits IBs, and cleans up
the submission context, which can also destroy buffers.
3-5% reduction in CPU overhead is expected for apps submitting a lot
of IBs per frame. This is most visible with DMA IBs.
v2: use a semaphore instead of a busy loop in amdgpu_ws_queue_cs
add another amdgpu_cs_sync_flush call into amdgpu_bo_map
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Not piping this all the way through yet, but no better place to note
this down. This will can be used with NV_viewport_array2.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
For fermi, this likely will require use of linked tsc mode. However on
bindless architectures, we can have as many as we want. As it stands,
the AUX_TEX_INFO has 32 teture handles reserved.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Indexed primitives were always using cut-aware primitive assembly,
whether primitive_restart was enabled or not. Correctly pass down
primitive_restart and select optimized PA when possible.
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
