v2: new signature style for buffer intrinsics (offsets)
v3: new signature style for llvm.amdgcn.buffer.load.format (overloaded return)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v2)
Whether DCC is disabled depends on the access flags with which the image
is bound: image_load supports DCC, but store and atomic don't.
v2: remove an unnecessary masking of images->desc.enabled_mask
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Re-order flags in the order in which they appear in the OpenGL spec in the
description of MemoryBarrier().
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The svga winsys modules can use this to send debug messages to the
state tracker and Mesa.
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
The index buffer handle saved in the hw_state structure could
be invalid after the index buffer is destroyed. Instead of
rebinding the index buffer using the saved index buffer handle,
we will reset the index buffer handle in the hw_state structure
to force resending of the index buffer.
Fixes bug 1593320
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
To ensure stream output target surfaces are available for the draw commands,
we need to rebind the current stream output targets at the first draw in the
command buffer.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Similar to other resources, current index buffer needs to be
rebound at the first draw of the current command buffer to make
sure the buffer is available for the draw command.
Fixes bug 1587263.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
When a constant buffer slot is allocated in the upload buffer,
the allocated slot size is always in multiple of 256. But the actual buffer
size might not be in multiple of 256. This causes a gap between
the ending offset of a slot and the starting offset of the next slot.
The gap will prevent the two slots to be updated in a single update command.
In order to maximize the chance of merging the contiguous dirty ranges,
when a slot is to be allocated in the constant upload buffer,
specify a buffer size in multiple of 256.
There is about 10% performance improvement with Lightsmark2008 and
30% with Cinebench R11.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
This patch adds the following HUD queries:
.num-resource-updates -- number of resource update. Commands include
UPDATE_SUBRESOURCE, UPDATE_GB_IMAGE.
.num-buffer-uploads -- number of buffer uploads.
.num-const-buf-updates -- number of set constant buffer.
.num-const-updates -- number of set shader constant.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
To find out how many image readback command is issued.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Previously, we looked at the bound textures (via the pipe_sampler_views)
to determine texture dimensions (1D/2D/3D/etc) and datatype (float vs.
int). But this could fail in out of memory conditions. If we failed to
allocate a texture and didn't create a pipe_sampler_view, we'd default
to using 0 (PIPE_BUFFER) as the texture type. This led to device errors
because of inconsistent shader code.
This change relies on all TGSI shaders having an SVIEW declaration for
each SAMP declaration. The previous patch series does that.
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
In June 2015, Rob Clark started updating the tgsi utility code to emit
SVIEW declarations in various shaders (for polygon stipple, blitting,
etc). These patches do the same for the Mesa state tracker.
The VMware driver will use this.
v2: support both TGSI_TEXTURE_2D and _RECT
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Depending on the driver's support for NPOT textures, we might use
a RECT texture instead of 2D texture. We should propogate that info
to the fragment shader's TEX instruction.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Instead of hard-coded 2D tex target in tgsi_transform_tex_2d_inst()
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
These magic file-index defines where only ever used in the nouveau code
and that no longer uses them.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v2)
handeLOAD / handleSTORE / handleATOM can only handle TGSI_FILE_BUFFER
and TGSI_FILE_MEMORY. Make things fail explictly when another
register-file is used in these functions.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
Commit c3083c7082 ("nv50/ir: add support for BUFFER accesses") disabled /
commented out some of the old resource handling code, but not all of it.
Effectively all of it is dead already, if we ever enter the old code
paths in handeLOAD / handleSTORE / handleATOM we will get an exception
due to trying to access the now always zero-sized resources vector.
Disable all the dead code.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
Make the store offset handling in CodeEmitterGK110::emitSTORE identical
to the one in CodeEmitterGK110::emitLOAD handling.
This is just a cleanup, it does not cause any functional changes.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Use the dst temp variable which was used in the TGSI_FILE_OUTPUT
case everywhere. This makes the code somewhat easier to reads
and helps avoiding going over 80 chars with upcoming changes.
This also brings the dst handling more in line with the src
handling.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Extend the MEMORY file support to differentiate between global, private
and shared memory, as well as "input" memory.
"MEMORY[x], INPUT" is intended to access OpenCL kernel parameters, a
special memory type is added for this, since the actual storage of these
(e.g. UBO-s) may differ per implementation. The uploading of kernel
parameters is handled by launch_grid, "MEMORY[x], INPUT" allows drivers
to use an access mechanism for parameter reads which matches with the
upload method.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu> (v1)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
When support for decl.Atomic and .Shared was added, tgsi_build_declaration
was not updated to propagate these properly.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu> (v1)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
tgsi_default_instruction_memory / tgsi_build_instruction_memory were
returning uninitialized memory for tgsi_instruction_memory.Texture and
tgsi_instruction_memory.Format. Note 0 means not set, and thus is a
correct default initializer for these.
Fixes: 3243b6fc97 ("tgsi: add Texture and Format to tgsi_instruction_memory")
Cc: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
When we have native integers, these have full precision. Whether they're
low/medium/high isn't piped through the TGSI yet, but eventually those
might have differing precisions. For now they're just 32-bit ints.
Fixes the following dEQP tests:
dEQP-GLES3.functional.state_query.shader.precision_vertex_highp_int
dEQP-GLES3.functional.state_query.shader.precision_fragment_highp_int
which expected highp ints to have full 32-bit precision, not the default
23-bit float precision.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
If a layer parameter is provided, we want to flip it to position 0 (and
combine it with any indirect params). However if the target is not an
array, there is no layer, so we have to shift all of the arguments down
by one to make room for it.
This fixes situations where there were non-coordinate parameters, such
as bias, lod, depth compare, explicit derivatives. Instead of adding a
new parameter at the front for the indirect reference, we would swap one
of those in its place.
Fixes dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.uniform.compute.*shadow
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reported-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
We make sure that that image depth matches the level's depth before
copying it into place. However we should only be minifying the first
level's depth for 3d textures - array textures have the same depth for
all levels.
This fixes tests such as
dEQP-GLES3.functional.texture.specification.texsubimage3d_depth.* and I
suspect account for a number of other odd situations I've run into where
level > 0 of array textures was messed up.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
In "manual" derivative mode (always used on nv50 and sometimes on nvc0
but always for cube), the idea is that using the quadop instruction, we
set up the "other" quads to have values such that the derivatives work
out, and then run the texture instruction as if nothing were strange. It
pulls values from the other lanes, and does its magic.
However cube coordinates have to be normalized - one of the 3 coords has
to be 1, to determine which is the major axis, to say which face is
being sampled. We were normalizing the coordinates first, and then
adding the derivatives. This is wrong for two reasons:
- the coordinates got normalized by a scaling factor but the
derivatives didn't
- the result of the addition didn't end up normalized
To resolve this, we flip the logic around to normalize *after* the
per-lane coordinates are set up.
This fixes a bunch of textureGrad cube dEQP tests.
NOTE: nv50 cube arrays with explicit derivatives are still broken, to be
resolved at a later date.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
