mesa.git
11 years agoi965/vs: Teach copy propagation about sends from GRFs.
Eric Anholt [Tue, 18 Dec 2012 01:03:02 +0000 (17:03 -0800)]
i965/vs: Teach copy propagation about sends from GRFs.

This incidentally also teaches it a bit about gen6 math -- we now allow
unswizzled, unmodified GRF temps as the sources for math.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965/vs: Prepare split_virtual_grfs() for the presence of SENDs from GRFs.
Eric Anholt [Tue, 18 Dec 2012 00:48:20 +0000 (16:48 -0800)]
i965/vs: Prepare split_virtual_grfs() for the presence of SENDs from GRFs.

v2: Fix silly bool handling, and don't add new tabs.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965/fs: Include everything but the final FB write in shader_time.
Eric Anholt [Tue, 19 Mar 2013 22:14:20 +0000 (15:14 -0700)]
i965/fs: Include everything but the final FB write in shader_time.

Previously, if you just wrote a constant color to the render target, no
time got noted at all.  This is convenient for doing single-instruction
timings, but not so much for actual program analysis.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965/fs: Switch shader_time writes to using GRFs.
Eric Anholt [Tue, 19 Mar 2013 22:28:11 +0000 (15:28 -0700)]
i965/fs: Switch shader_time writes to using GRFs.

This avoids conflicts between shader_time and FB writes, so we can include
more of the program under our profiling.  This does mean hiding more of
the message setup from the optimizer, which doesn't have a way to handle
multi-reg sends from GRFs.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Provide more detailed information to match shader_time to programs.
Eric Anholt [Tue, 19 Mar 2013 21:28:29 +0000 (14:28 -0700)]
i965: Provide more detailed information to match shader_time to programs.

Ken asked me the other day what -1 vs 0 vs 3 vs other meant in our shader
names, and I realized that it was really unclear.  I'd like to do even
better, like noting which one is the clear shader, but that would require
exposing the metaops struct to the driver.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Track ARB program state along with GLSL state for shader_time.
Eric Anholt [Tue, 19 Mar 2013 21:27:42 +0000 (14:27 -0700)]
i965: Track ARB program state along with GLSL state for shader_time.

This will let us do much better printouts for non-GLSL programs.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agost/dri: fix crash with HUD and single buffering
Marek Olšák [Wed, 27 Mar 2013 00:56:25 +0000 (01:56 +0100)]
st/dri: fix crash with HUD and single buffering

11 years agost/mesa: remove leftover printfs from ReadPixels
Marek Olšák [Thu, 28 Mar 2013 13:51:28 +0000 (14:51 +0100)]
st/mesa: remove leftover printfs from ReadPixels

Oops, I thought I had removed all debugging code.

11 years agoi965/fs: Improve performance of copy propagation dataflow using bitsets.
Eric Anholt [Fri, 22 Mar 2013 21:11:25 +0000 (14:11 -0700)]
i965/fs: Improve performance of copy propagation dataflow using bitsets.

Reduces compile time of l4d2's slowest shader by 17.8% +/- 1.3% (n=10).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agollvmpipe/draw: Fix texture sampling in geometry shaders
Zack Rusin [Wed, 27 Mar 2013 00:53:27 +0000 (17:53 -0700)]
llvmpipe/draw: Fix texture sampling in geometry shaders

We weren't correctly propagating the samplers and sampler views
when they were related to geometry shaders.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agodraw/llvm: Cleanup the store debugging code
Zack Rusin [Tue, 26 Mar 2013 19:35:45 +0000 (12:35 -0700)]
draw/llvm: Cleanup the store debugging code

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agodraw: Allocate the output buffer for output primitives
Zack Rusin [Tue, 26 Mar 2013 19:32:30 +0000 (12:32 -0700)]
draw: Allocate the output buffer for output primitives

We were allocating the output buffer but using the input
primitives. We need to allocate that buffer using the
maximum number of output, not input, primitives.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agogallivm: Implement the breakc instruction
Zack Rusin [Wed, 27 Mar 2013 09:38:32 +0000 (02:38 -0700)]
gallivm: Implement the breakc instruction

Required by more modern examples. Like BRK but with a condition.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agogallivm: implement implicit primitive flushing
Zack Rusin [Wed, 27 Mar 2013 09:30:38 +0000 (02:30 -0700)]
gallivm: implement implicit primitive flushing

TGSI semantics currently require an implicit endprim at the end
of GS if an ending primitive hasn't been emitted.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agogallium/llvm: implement geometry shaders in the llvm paths
Zack Rusin [Mon, 18 Feb 2013 12:00:19 +0000 (04:00 -0800)]
gallium/llvm: implement geometry shaders in the llvm paths

This commits implements code generation of the geometry shaders in
the SOA paths. All the code is there but bugs are likely present.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agodraw/gs: Fetch more than one primitive per invocation
Zack Rusin [Thu, 14 Mar 2013 08:07:28 +0000 (01:07 -0700)]
draw/gs: Fetch more than one primitive per invocation

Allows executing gs on up to 4 primitives at a time. Will also be
required by the llvm code because there we definitely don't want
to flush with just a single primitive.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agodraw/gs: Abstract the portions of GS that are tgsi specific
Zack Rusin [Thu, 14 Mar 2013 07:42:06 +0000 (00:42 -0700)]
draw/gs: Abstract the portions of GS that are tgsi specific

To be able to add llvm paths later on we need to have some common
interface for them.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agodraw/llvm: Remove unused gs_constants from jit_context
Zack Rusin [Thu, 14 Mar 2013 00:13:21 +0000 (17:13 -0700)]
draw/llvm: Remove unused gs_constants from jit_context

The member was never used and we'll need to handle it differently
because gs will also need samplers/textures setup.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agograw/gs: add missing max output vertices to all tests
Zack Rusin [Wed, 13 Mar 2013 23:46:24 +0000 (16:46 -0700)]
graw/gs: add missing max output vertices to all tests

A few tests were missing this crucial property.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
11 years agoradeonsi: add cs tracing v3
Jerome Glisse [Mon, 25 Mar 2013 15:46:38 +0000 (11:46 -0400)]
radeonsi: add cs tracing v3

Same as on r600, trace cs execution by writting cs offset after each
states, this allow to pin point lockup inside command stream and
narrow down the scope of lockup investigation.

v2: Use WRITE_DATA packet instead of WRITE_MEM
v3: Remove useless nop packet

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
11 years agomesa: only check sample count if we actually wanted multisampling
Chris Forbes [Mon, 25 Mar 2013 10:19:07 +0000 (23:19 +1300)]
mesa: only check sample count if we actually wanted multisampling

Fixes various test fallout from 90b5a2425a on Pineview, which claims to
support ARB_internalformat_query but doesn't actually provide the
driverfunc.

That driver is still broken [GetInternalformativ will still segfault!]
but it was silly to be going through the sample count logic in the
nonmultisampling case at all.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoradeon/llvm: document LLVM commit
Christian König [Tue, 26 Mar 2013 14:08:00 +0000 (15:08 +0100)]
radeon/llvm: document LLVM commit

We need at least that revision to work correctly now.

Signed-off-by: Christian König <christian.koenig@amd.com>
11 years agoradeonsi: add preloading for all samplers
Christian König [Sun, 17 Mar 2013 15:02:42 +0000 (16:02 +0100)]
radeonsi: add preloading for all samplers

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoradeonsi: add preloading of all constants
Christian König [Fri, 15 Mar 2013 14:53:25 +0000 (15:53 +0100)]
radeonsi: add preloading of all constants

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoradeonsi: mark most intrinsics as readnone/nounwind
Christian König [Wed, 20 Mar 2013 11:10:35 +0000 (12:10 +0100)]
radeonsi: mark most intrinsics as readnone/nounwind

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoradeonsi: mark all loads as constant
Christian König [Wed, 20 Mar 2013 13:37:21 +0000 (14:37 +0100)]
radeonsi: mark all loads as constant

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoradeonsi: remove wqm intrinsic
Christian König [Tue, 19 Mar 2013 12:36:26 +0000 (13:36 +0100)]
radeonsi: remove wqm intrinsic

Now the backend handles that itself.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoradeon/llvm: remove uneeded inclusion
Christian König [Tue, 26 Mar 2013 10:37:45 +0000 (11:37 +0100)]
radeon/llvm: remove uneeded inclusion

The include isn't needed and the file has moved with LLVM master.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agoglsl_to_tgsi: avoid creating arrays if driver doesn't support them
Christian König [Sun, 24 Mar 2013 15:24:52 +0000 (16:24 +0100)]
glsl_to_tgsi: avoid creating arrays if driver doesn't support them

Avoid creating arrays if we replace indirect addressing anyway.

Signed-off-by: Christian König <christian.koenig@amd.com>
11 years agoglsl_to_tgsi: make simplify_cmp work with arrays
Christian König [Mon, 25 Mar 2013 09:57:48 +0000 (10:57 +0100)]
glsl_to_tgsi: make simplify_cmp work with arrays

Even when we have arrays it is possible for simplify_cmp
to work on temps, just not on arrays.

Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=62696
Signed-off-by: Christian König <christian.koenig@amd.com>
11 years agogallium/docs: document get_driver_query_info
Marek Olšák [Tue, 26 Mar 2013 00:37:40 +0000 (01:37 +0100)]
gallium/docs: document get_driver_query_info

11 years agor600g: add a driver query returning the amount of requested VRAM and GTT memory
Marek Olšák [Fri, 22 Mar 2013 01:39:42 +0000 (02:39 +0100)]
r600g: add a driver query returning the amount of requested VRAM and GTT memory

11 years agor600g: add a driver query returning the number of draw_vbo calls
Marek Olšák [Thu, 21 Mar 2013 18:44:18 +0000 (19:44 +0100)]
r600g: add a driver query returning the number of draw_vbo calls

between begin_query and end_query

11 years agost/dri: integrate the HUD
Marek Olšák [Thu, 21 Mar 2013 18:51:30 +0000 (19:51 +0100)]
st/dri: integrate the HUD

Reviewed-by: Brian Paul <brianp@vmware.com>
11 years agogallium: implement a heads-up display module
Marek Olšák [Thu, 21 Mar 2013 18:47:06 +0000 (19:47 +0100)]
gallium: implement a heads-up display module

Reviewed-by: Brian Paul <brianp@vmware.com>
v2: lots of cosmetic changes

11 years agogallium: add interface for driver queries like performance counters, etc.
Marek Olšák [Thu, 21 Mar 2013 18:32:24 +0000 (19:32 +0100)]
gallium: add interface for driver queries like performance counters, etc.

The pipe query interface is reused. The list of available queries can be
obtained using pipe_screen::get_driver_query_info.

Reviewed-by: Brian Paul <brianp@vmware.com>
11 years agogallium/tgsi: fix valgrind warning
Marek Olšák [Fri, 22 Mar 2013 16:04:15 +0000 (17:04 +0100)]
gallium/tgsi: fix valgrind warning

"Conditional jump or move depends on uninitialised value(s)"

Reviewed-by: Brian Paul <brianp@vmware.com>
11 years agost/mesa: fix crash with blit-based GetTexImage
Marek Olšák [Thu, 21 Mar 2013 23:31:47 +0000 (00:31 +0100)]
st/mesa: fix crash with blit-based GetTexImage

https://bugs.freedesktop.org/show_bug.cgi?id=62573

Tested-by: Andreas Boll <andreas.boll.dev@gmail.com>
11 years agocso: add constant buffer save/restore feature for postprocessing
Marek Olšák [Mon, 18 Mar 2013 21:36:21 +0000 (22:36 +0100)]
cso: add constant buffer save/restore feature for postprocessing

Postprocessing is an internal meta op and should restore the states
it changes.

11 years agoradeonsi: fix crash while binding a NULL constant buffer
Marek Olšák [Thu, 21 Mar 2013 18:29:29 +0000 (19:29 +0100)]
radeonsi: fix crash while binding a NULL constant buffer

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
11 years agor600g: fix crash while binding a NULL constant buffer
Marek Olšák [Thu, 21 Mar 2013 18:29:29 +0000 (19:29 +0100)]
r600g: fix crash while binding a NULL constant buffer

11 years agor300g: fix crash while binding a NULL constant buffer
Marek Olšák [Thu, 21 Mar 2013 18:29:29 +0000 (19:29 +0100)]
r300g: fix crash while binding a NULL constant buffer

11 years agor600g: Use virtual address for PIPE_QUERY_SO* in r600_emit_query_end
Martin Andersson [Mon, 25 Mar 2013 22:11:34 +0000 (23:11 +0100)]
r600g: Use virtual address for PIPE_QUERY_SO* in r600_emit_query_end

Virtual address is used for PIPE_QUERY_SO* queries in
r600_emit_query_begin, but not in r600_emit_query_end.

This will trigger a GPU fault when one of those queries is
made and virtual address is enabled.

Note: this is a candidate for the 9.1 branch

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
11 years agofreedreno: use u_debug for debug env vars
Rob Clark [Mon, 25 Mar 2013 18:57:24 +0000 (14:57 -0400)]
freedreno: use u_debug for debug env vars

Signed-off-by: Rob Clark <robdclark@gmail.com>
11 years agoglsl ir: add as_dereference_record
Jordan Justen [Sun, 10 Mar 2013 01:12:09 +0000 (17:12 -0800)]
glsl ir: add as_dereference_record

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agogallium: undef PACKAGE_* macros to silence warnings
Brian Paul [Mon, 25 Mar 2013 16:24:01 +0000 (10:24 -0600)]
gallium: undef PACKAGE_* macros to silence warnings

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
11 years agogallivm: init vars to silence warnings
Brian Paul [Mon, 25 Mar 2013 16:23:42 +0000 (10:23 -0600)]
gallivm: init vars to silence warnings

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
11 years agoswrast: init vars to silence warnings
Brian Paul [Mon, 25 Mar 2013 16:23:27 +0000 (10:23 -0600)]
swrast: init vars to silence warnings

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
11 years agofreedreno: prefer sw upload for textures
Rob Clark [Mon, 25 Mar 2013 15:55:18 +0000 (11:55 -0400)]
freedreno: prefer sw upload for textures

Since we are UMA, in most cases the GPU blit doesn't make much sense for
texture upload.

Signed-off-by: Rob Clark <robdclark@gmail.com>
11 years agofreedreno: track maximal scissor bounds
Rob Clark [Wed, 6 Mar 2013 15:45:58 +0000 (10:45 -0500)]
freedreno: track maximal scissor bounds

Optimize out parts of the render target that are scissored out by taking
into account maximal scissor bounds in fd_gmem_render_tiles().

This is a big win on things like gnome-shell which frequently do partial
screen updates.

Signed-off-by: Rob Clark <robdclark@gmail.com>
11 years agoandroid: fix Android.mk bug in mesa/drivers/dri/common
Adrian Marius Negreanu [Fri, 22 Mar 2013 11:42:40 +0000 (13:42 +0200)]
android: fix Android.mk bug in mesa/drivers/dri/common

target-specific variables are undefined when used as pre-requisites.
instead, use secondary-expansion.

I noticed this when building the patch:
     i965: Add a driconf option to disable flush throttling

Signed-off-by: Adrian Marius Negreanu <adrian.m.negreanu@intel.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
11 years agomesa: Disable validate_ir_tree() on release builds.
Eric Anholt [Mon, 18 Mar 2013 15:42:19 +0000 (08:42 -0700)]
mesa: Disable validate_ir_tree() on release builds.

Since half of ir_validate uses asserts() (the other using printf() then
abort()), there's not much use to calling it in a release build.  Cuts
6.3% of the startup time of TF2.

NOTE: This is a candidate for the stable branches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agogallivm: move code for dealing with rgb9e5 and r11g11b10 formats to own file
Roland Scheidegger [Sun, 24 Mar 2013 01:08:01 +0000 (02:08 +0100)]
gallivm: move code for dealing with rgb9e5 and r11g11b10 formats to own file

This is really not generic conversion stuff and the code very particular to
these formats.

11 years agollvmpipe: Fix assertions with assignment instead of comparison.
Vinson Lee [Sat, 23 Mar 2013 07:24:52 +0000 (00:24 -0700)]
llvmpipe: Fix assertions with assignment instead of comparison.

Fixes assign instead of compare defects reported by Coverity.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
11 years agoi965: Shrink brw_vue_map struct.
Paul Berry [Fri, 22 Mar 2013 00:14:53 +0000 (17:14 -0700)]
i965: Shrink brw_vue_map struct.

This patch changes the arrays in brw_vue_map (which only ever contain
values from -1 to 58) from ints to signed chars.  This reduces the
size of the struct from 488 bytes to 136 bytes.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: fix STATIC_ASSERT to use 127 instead of 128.

Reviewed-by: Eric Anholt <eric@anholt.net>
11 years agoi965/fs: Rename vp_outputs_written to input_slots_valid.
Paul Berry [Wed, 20 Mar 2013 17:15:52 +0000 (10:15 -0700)]
i965/fs: Rename vp_outputs_written to input_slots_valid.

With the introduction of geometry shaders, fragment inputs will no
longer come exclusively from the vertex shader; sometimes they come
from the geometry shader.  So the name "vp_outputs_written" will
become a misnomer.  This patch renames vp_outputs_written to
input_slots_valid, to reflect the true meaning of the bitfield from
the fragment shader's point of view: it indicates which of the
possible input slots contain valid data that was written by the
previous shader stage.

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Use brw.vue_map_geom_out instead of VS output VUE map where appropriate.
Paul Berry [Sun, 17 Mar 2013 18:29:28 +0000 (11:29 -0700)]
i965: Use brw.vue_map_geom_out instead of VS output VUE map where appropriate.

This patch modifies post-GS pipeline stages (transform feedback, clip,
sf, fs) to refer to the VUE map through brw->vue_map_geom_out rather
than brw->vs.prog_data->vue_map.  This ensures that when geometry
shader support is added, these pipeline stages will consult the
geometry shader output VUE map when appropriate, rather than the
vertex shader output VUE map.

v2: Fixed some stale "CACHE_NEW_VS_PROG" comments.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Store the geometry output VUE map in brw_context.
Paul Berry [Mon, 18 Feb 2013 18:16:02 +0000 (10:16 -0800)]
i965: Store the geometry output VUE map in brw_context.

Currently, the GPU pipeline has one active VUE map in effect at any
given time--the one representing the layout of vertex data coming from
the vertex shader.  However, when geometry shaders are added, they
will have their own independent VUE map.  Later pipeline stages (clip,
sf, fs) will need to consult the geometry shader VUE map if a geometry
shader is in use, and the vertex shader VUE map otherwise.

This patch adds a new field to brw_context, vue_map_geom_out, which
contains the VUE map that should be used by later pipeline stages.  It
also adds a new state flag, BRW_NEW_VUE_MAP_GEOM_OUT, which is
signalled whenever the contents of the VUE map changes.

Since we don't support geometry shaders yet, vue_map_geom_out is
currently set only by the brw_vs_prog state atom.

v2: Don't set vue_map_geom_out in do_vs_prog--that's redundant and
possibly problematic for precompiles.  Only set it in
brw_upload_vs_prog.  Also, make a copy instead of using a
pointer--this makes it possible to detect when the VUE map hasn't
changed, so we can avoid redundant state uploads.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Move brw_vs_prog_data::outputs_written into VUE map.
Paul Berry [Sun, 17 Mar 2013 18:13:56 +0000 (11:13 -0700)]
i965: Move brw_vs_prog_data::outputs_written into VUE map.

Future patches will allow for there to be separate VUE maps when both
a geometry shader and a vertex shader are in use.  When this happens,
we will want to have correspondingly separate outputs_written
bitfields.  Moving outputs_written into the VUE map will make this
easy.

For consistency with the terminology used in the VUE map, the bitfield
is renamed to "slots_valid" in the process.

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965/gen7: Use WE_all mode when enabling channel masks for URB write.
Paul Berry [Sat, 23 Mar 2013 15:23:03 +0000 (08:23 -0700)]
i965/gen7: Use WE_all mode when enabling channel masks for URB write.

Gen7 adds mask bits to the message header for a URB write which allow
the write to apply only to certain channels.  We don't use this
functionality, so to ensure that the entire write always occurs, we
emit an OR instruction to set the mask bits.

With the advent of geometry shaders, URB writes won't just happen at
the end of a thread; they will happen in mid-thread too.  Thus, we can
no longer rely on channel 0 being enabled, so we need to emit the OR
instruction in WE_all mode to ensure that it is executed.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Rename BRW_VARYING_SLOT_MAX -> BRW_VARYING_SLOT_COUNT.
Paul Berry [Sat, 23 Mar 2013 22:53:33 +0000 (15:53 -0700)]
i965: Rename BRW_VARYING_SLOT_MAX -> BRW_VARYING_SLOT_COUNT.

The new name clarifies that it represents *one more* than the maximum
possible brw_varying_slot value.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Clarify nomenclature: vert_result -> varying
Paul Berry [Fri, 22 Mar 2013 16:39:11 +0000 (09:39 -0700)]
i965: Clarify nomenclature: vert_result -> varying

This patch removes the terminology "vert_result" from the i965 driver,
replacing it with "varying".  The old terminology, "vert_result", was
confusing because (a) it referred to the enum gl_vert_result, which no
longer exists (it was replaced with gl_varying_slot), and (b) it
implied a vertex output, but with the advent of geometry shaders, it
could be either a vertex or a geometry output, depending what shaders
are in use.  The generic term "varying" is less confusing.

No functional change.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Whitespace fixes.

11 years agoi965: bump MAX_DEPTH_TEXTURE_SAMPLES to 4/8
Chris Forbes [Sun, 24 Mar 2013 03:21:01 +0000 (16:21 +1300)]
i965: bump MAX_DEPTH_TEXTURE_SAMPLES to 4/8

Bump MAX_DEPTH_TEXTURE_SAMPLES to match what GetInternalformativ is
claiming. Since that limit is what is actually enforced now, this
doesn't actually change anything except the queried value.

There's still no piglits verifying that multisample depth textures work,
but this works in the Unigine demos.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agomesa: use _mesa_check_sample_count() for multisample textures
Chris Forbes [Sun, 3 Mar 2013 08:46:12 +0000 (21:46 +1300)]
mesa: use _mesa_check_sample_count() for multisample textures

Extends _mesa_check_sample_count() to properly support the
TEXTURE_2D_MULTISAMPLE and TEXTURE_2D_MULTISAMPLE_ARRAY targets, which
have subtly different limits than renderbuffers.

This resolves the remaining TODO in the implementation of
TexImage*DMultisample.

V2: - Don't introduce spurious block.
    - Do this in multisample.c instead.
    - Fix typo in error message.
    - Inline spec quotes

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agomesa: helper for checking renderbuffer sample count
Chris Forbes [Wed, 6 Feb 2013 07:42:53 +0000 (20:42 +1300)]
mesa: helper for checking renderbuffer sample count

Pulls the checking of the sample count into a helper function, and
extends the existing logic to include the interactions with both
ARB_texture_multisample and ARB_internalformat_query.

_mesa_check_sample_count() checks a desired sample count against a
a combination of target/internalformat, and returns the error enum
to be produced, if any. Unfortunately the conditions are messy and the
errors vary.

V2: - Tidy up spurious block.
    - Move _mesa_check_sample_count() to multisample.c instead; It
      doesn't really belong in fbobject.c or teximage.c.
    - Inlined spec quotes

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agomesa: allow internalformat_query with multisample texture targets
Chris Forbes [Sat, 16 Feb 2013 07:47:11 +0000 (20:47 +1300)]
mesa: allow internalformat_query with multisample texture targets

Now that we support ARB_texture_multisample, there are multiple targets
accepted for this query, and they may have target-dependent limits, so
pass the target to the driverfunc.

For example, the sampling hardware may not be able to do general
texelFetch() for some format/sample count combination, but the driver
may still be able to implement a reasonable resolve operation, so it can
be supported for renderbuffers.

V2: - Don't break Gallium compile.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoclover: add dynamic_cast results checking down in clSetKernelArgument() code path.
Dmitry Cherkassov [Sat, 23 Mar 2013 19:51:22 +0000 (23:51 +0400)]
clover: add dynamic_cast results checking down in clSetKernelArgument() code path.

Signed-off-by: Dmitry Cherkassov <dcherkassov@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
11 years agogallivm: Add code for rgb9e5 shared exponent format to float conversion
Roland Scheidegger [Sat, 23 Mar 2013 01:05:54 +0000 (02:05 +0100)]
gallivm: Add code for rgb9e5 shared exponent format to float conversion

And use this (and the code for r11g11b10 packed float to float conversion)
in the soa texturing code (the generated code looks quite good).
Should be an order of magnitude faster probably than using the fallback
(not measured).
Tested with piglit texwrap GL_EXT_packed_float and
GL_EXT_texture_shared_exponent respectively (didn't find much else using
it).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
11 years agogallium,st/mesa: don't use blit-based transfers with software rasterizers
Marek Olšák [Thu, 14 Mar 2013 16:18:43 +0000 (17:18 +0100)]
gallium,st/mesa: don't use blit-based transfers with software rasterizers

The blit-based paths for TexImage, GetTexImage, and ReadPixels aren't very
fast with software rasterizer. Now Gallium drivers have the ability to turn
them off.

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agost/mesa: implement blit-based ReadPixels
Marek Olšák [Thu, 14 Mar 2013 15:36:22 +0000 (16:36 +0100)]
st/mesa: implement blit-based ReadPixels

Initial version contributed by: Martin Andersson <g02maran@gmail.com>

This is only used if the memcpy path cannot be used and if no transfer ops
are needed. It's pretty similar to our TexImage and GetTexImage
implementations.

The motivation behind this is to be able to use ReadPixels every frame and
still have at least 20 fps (or 60 fps with a powerful GPU and CPU)
instead of 0.5 fps.

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agomesa: add common format-independent memcpy-based ReadPixels path
Marek Olšák [Thu, 14 Mar 2013 14:20:27 +0000 (15:20 +0100)]
mesa: add common format-independent memcpy-based ReadPixels path

I'll need the _mesa_readpixels_needs_slow_path function for the blit-based
version, but it's also useful to have this memcpy-based path in one place
and not scattered across several functions.

v2: add "const" to function parameters

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agomesa: add helper func for checking combined depthstencil buffers from st/mesa
Marek Olšák [Thu, 14 Mar 2013 13:22:56 +0000 (14:22 +0100)]
mesa: add helper func for checking combined depthstencil buffers from st/mesa

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agomesa: add a common function returning transfer ops for ReadPixels
Marek Olšák [Thu, 14 Mar 2013 12:15:54 +0000 (13:15 +0100)]
mesa: add a common function returning transfer ops for ReadPixels

I'll need both new functions for later. For now, it consolidates the code
for determining what the transfer ops should be and makes it a little bit
smarter.

v2: added "const"

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agomesa: handle HALF_FLOAT like FLOAT in get_tex_rgba
Marek Olšák [Wed, 13 Mar 2013 15:47:21 +0000 (16:47 +0100)]
mesa: handle HALF_FLOAT like FLOAT in get_tex_rgba

NOTE: This is a candidate for the stable branches.

Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
11 years agollvmpipe: add EXT_packed_float render target format support
Roland Scheidegger [Fri, 22 Mar 2013 19:09:18 +0000 (20:09 +0100)]
llvmpipe: add EXT_packed_float render target format support

New conversion code to handle conversion from/to r11g11b10 AoS to/from
SoA floats, and also add code for conversion from rgb9e5 AoS to float SoA
(which works pretty much the same as r11g11b10 except for the packing).
(This code should also be used for texture sampling instead of
relying on u_format conversion but it's not yet, so rgb9e5 is unused.)
Unfortunately a crazy amount of hacks is necessary to get the conversion
code running in llvmpipe's generate_unswizzled_blend, which isn't well
suited for formats where the storage representation has nothing to do
with what's needed for blending (moreover, the conversion will convert
from packed AoS values, which is the storage format, to float SoA values,
because this is much more natural for the conversion, and likewise from
SoA values to packed AoS values - but the "blend" (which includes
trivial things like partial mask) works on AoS values, so incoming fs
values will go SoA->AoS, values from destination will go packed
AoS->SoA->AoS, then do blend, then AoS->SoA->packed AoS which probably
isn't the most efficient way though the shuffles are probably bearable).

Passes piglit fbo-blending-formats (with GL_EXT_packed_float parameter),
still need to verify Inf/NaNs (where most of the complexity in the
conversion comes from actually).

v2: drop the (very bogus) rgb9e5 part, and do component extraction
in the helper code for r11g11b10 to float conversion, making the code
slightly more compact (suggested by Jose), now that there are no other
callers left this works quite well. (Could do the same for the
opposite way but it's less than ideal there, final part of packing
needs to be done in caller anyway and there'd be another conditional.)

v3: minor style and comment fixes. Also fix a potential issue with
negative zero being potentially returned by max(src, zero) as we
don't have well-defined min/max behavior (fortunately no additonal cost).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
11 years agor600g: Honour legacy debugging environment variables
Michel Dänzer [Thu, 21 Mar 2013 16:56:52 +0000 (17:56 +0100)]
r600g: Honour legacy debugging environment variables

This helps minimize confusion / effort when moving between branches or
helping others.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Marek Olšák <maraeo@gmail.com>
11 years agodocs: Mark ARB_ES3_compatibility as done.
Matt Turner [Thu, 21 Mar 2013 22:59:08 +0000 (15:59 -0700)]
docs: Mark ARB_ES3_compatibility as done.

11 years agofreedreno: add pipe->blit
Rob Clark [Thu, 21 Mar 2013 21:32:00 +0000 (17:32 -0400)]
freedreno: add pipe->blit

Signed-off-by: Rob Clark <robdclark@gmail.com>
11 years agoi965: Add a driconf option to disable flush throttling.
Paul Berry [Tue, 19 Mar 2013 18:49:08 +0000 (11:49 -0700)]
i965: Add a driconf option to disable flush throttling.

Normally when submitting the first batch buffer after a flush, we
check whether the GPU has completed processing of the first batch
buffer of the previous frame.  If it hasn't, we wait for it to finish
before submitting any more batches.  This prevents GPU-heavy and
CPU-light applications from racing too far ahead of the current frame,
but at the expense of possibly lower frame rates.  Sometimes when
benchmarking we want to disable this mechanism.

This patch adds the driconf option "disable_throttling" to disable the
throttling mechanism.

Reviewed-by: Eric Anholt <eric@anholt.net>
11 years agomesa: Implement TEXTURE_IMMUTABLE_LEVELS for ES 3.0.
Matt Turner [Mon, 4 Mar 2013 19:03:58 +0000 (11:03 -0800)]
mesa: Implement TEXTURE_IMMUTABLE_LEVELS for ES 3.0.

NOTE: This is a candidate for the 9.1 branch.
Fixes piglit's texture-immutable-levels test.
Reported-by: Marek Olšák <maraeo@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
11 years agoglx: Build with VISIBILITY_CFLAGS in automake
Adam Jackson [Thu, 21 Mar 2013 17:21:18 +0000 (13:21 -0400)]
glx: Build with VISIBILITY_CFLAGS in automake

Note: This is a candidate for the stable branches.

Signed-off-by: Adam Jackson <ajax@redhat.com>
11 years agoscons: check for existance of 'MSVC_VERSION' in env
Brian Paul [Wed, 20 Mar 2013 15:41:52 +0000 (09:41 -0600)]
scons: check for existance of 'MSVC_VERSION' in env

Evidently, MSVC_VERSION isn't always defined so check for it before
checking the MSVC version.

Suggested by Jose.

11 years agosoftpipe: silence some asst. MSVC type warnings in sp_tex_sample.c
Brian Paul [Wed, 20 Mar 2013 17:05:13 +0000 (11:05 -0600)]
softpipe: silence some asst. MSVC type warnings in sp_tex_sample.c

11 years agosoftpipe: silence some MSVC signed/unsigned warnings
Brian Paul [Wed, 20 Mar 2013 16:56:47 +0000 (10:56 -0600)]
softpipe: silence some MSVC signed/unsigned warnings

11 years agosoftpipe: silence some MSVC float/double warnings
Brian Paul [Wed, 20 Mar 2013 16:56:24 +0000 (10:56 -0600)]
softpipe: silence some MSVC float/double warnings

11 years agorbug: silence some MSVC signed/unsigned warnings
Brian Paul [Wed, 20 Mar 2013 16:54:24 +0000 (10:54 -0600)]
rbug: silence some MSVC signed/unsigned warnings

11 years agopostprocess: silence some MSVC float/int warnings
Brian Paul [Wed, 20 Mar 2013 16:54:07 +0000 (10:54 -0600)]
postprocess: silence some MSVC float/int warnings

11 years agometa: fix incorrect slice, r coordinate computation
Brian Paul [Wed, 20 Mar 2013 15:58:18 +0000 (09:58 -0600)]
meta: fix incorrect slice, r coordinate computation

The arithmetic to convert a 3D texture slice to an R coordinate was
incorrect.  Found when MSVC warned of a divide by zero.

Note that we don't actually ever hit this path.  We don't decompress
slices of 3D textures and we don't support 3D mipmap generation yet.

11 years agovega: fix MSVC warning about missing return statement
Brian Paul [Wed, 20 Mar 2013 15:57:21 +0000 (09:57 -0600)]
vega: fix MSVC warning about missing return statement

11 years agometa: minor indentation fix
Brian Paul [Wed, 20 Mar 2013 15:27:14 +0000 (09:27 -0600)]
meta: minor indentation fix

11 years agoradeonsi: Emit pixel shader state even when only the vertex shader changed
Michel Dänzer [Tue, 19 Mar 2013 16:57:11 +0000 (17:57 +0100)]
radeonsi: Emit pixel shader state even when only the vertex shader changed

Fixes random failures with piglit glsl-max-varyings.

NOTE: This is a candidate for the 9.1 branch.

Reviewed-by: Christian König <christian.koenig@amd.com>
11 years agoandroid: Define PACKAGE_VERSION/BUGREPORT in CFLAGS
Chad Versace [Mon, 18 Mar 2013 20:56:28 +0000 (13:56 -0700)]
android: Define PACKAGE_VERSION/BUGREPORT in CFLAGS

This fixes the Android build. Commit 439c3d4 broke it.

CC: Adrian M Negreanu <adrian.m.negreanu@intel.com>
CC: Matt Turner <mattst@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>
11 years agoi965/vs: Add IR dumping for immediates.
Kenneth Graunke [Mon, 11 Mar 2013 18:10:34 +0000 (11:10 -0700)]
i965/vs: Add IR dumping for immediates.

This makes dump_instructions more useful.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
11 years agoglsl: Add built-in functions for GLSL 1.50.
Kenneth Graunke [Tue, 19 Mar 2013 01:57:28 +0000 (18:57 -0700)]
glsl: Add built-in functions for GLSL 1.50.

This makes basic built-in functions work in GLSL 1.50.  It supports
everything except the new Geometry Shader functions.

The new 150.glsl file is 140.glsl plus ARB_texture_multisample.glsl;
150.frag is identical to 140.frag except for the #version bump.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
11 years agoglsl: Add sampler2DMS/sampler2DMSArray types to GLSL 1.50.
Kenneth Graunke [Tue, 19 Mar 2013 01:57:27 +0000 (18:57 -0700)]
glsl: Add sampler2DMS/sampler2DMSArray types to GLSL 1.50.

GLSL 1.50 includes support for the new sampler types introduced by
the ARB_texture_multisample extension.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
11 years agoglsl: Bump standalone compiler versions to 1.50.
Kenneth Graunke [Tue, 19 Mar 2013 01:57:26 +0000 (18:57 -0700)]
glsl: Bump standalone compiler versions to 1.50.

The version bumps are necessary in order to compile built-ins for 1.50.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
11 years agoi965: Don't use texture swizzling to force alpha to 1.0 if unnecessary.
Kenneth Graunke [Fri, 15 Mar 2013 21:48:24 +0000 (14:48 -0700)]
i965: Don't use texture swizzling to force alpha to 1.0 if unnecessary.

Commit 33599433c7 began setting the texture swizzle mode to XYZ1 for
RED, RG, and RGB textures in order to force alpha to 1.0 in case we
actually stored the texture as RGBA.

This had a unforseen performance implication: the shader precompile
assumes that the texture swizzle mode will be XYZW for non-shadow
sampler types.  By setting it to XYZ1, this means every shader used with
a RED, RG, or RGB texture has to be recompiled.  This is a very common
case.

Unfortunately, there's no way to improve the precompile, since RGBA
textures still need XYZW, and there's no way to know by looking at
the shader source what texture formats might be used.

However, we only need to smash alpha to 1.0 if the texture's memory
format actually has alpha bits.  If not, the sampler already returns 1.0
for us without any special swizzling.  XRGB8888, for example, is a very
common case where this occurs.

This partially fixes a performance regression since commit 33599433c7.
More work is required to fully fix it in all cases.  This at least helps
Warsow.

NOTE: This is a candidate for the 9.1 branch.

Reviewed-by: Carl Worth <cworth@cworth.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agoi965: Don't print a fatal-looking message if intelCreateContext fails.
Kenneth Graunke [Thu, 14 Mar 2013 18:48:36 +0000 (11:48 -0700)]
i965: Don't print a fatal-looking message if intelCreateContext fails.

With the old context creation mechanism, an application asked the GL to
give it a context.  Failing to produce a context was a fatal error.

Now, with GLX_ARB_create_context, the application can request a specific
version.  If it's higher than the maximum version we support, context
creation will fail.  But this is a normal error that applications
recover from.

In particular, the new glxinfo tries to create OpenGL 4.3, 4.2, 4.1,
4.0, 3.3, and 3.2 contexts before finally succeeding at creating a 3.1
context.  This led to it printing the following message 6 times:
"brwCreateContext: failed to init intel context"

There's no need to alarm users (and developers) with such a message.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
11 years agoi965/gen7: Align all depth miplevels to 8 in the X direction.
Eric Anholt [Mon, 18 Mar 2013 22:38:58 +0000 (15:38 -0700)]
i965/gen7: Align all depth miplevels to 8 in the X direction.

On an INTEL_DEBUG=perf piglit run on IVB, reduces the instances of "HW
workaround: blit" (the printouts from the misaligned-depth workaround
blits) from 725 to 675.

It doesn't totally eliminate the workaround blit, because we still have
problems with Y offsets that we can't fix (since texturing can only align
miplevels up to 2 or 4, not 8).

No regressions on piglit/es3conform on IVB.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
11 years agonvc0: fix max varying count, move CLIPVERTEX,FOG out of the way
Christoph Bumiller [Fri, 15 Mar 2013 22:39:01 +0000 (23:39 +0100)]
nvc0: fix max varying count, move CLIPVERTEX,FOG out of the way

The card spews an error if I use all 128 generic slots.
Apparently the real limit isn't just dictated by the address space
layout.