mesa.git
12 years agomesa: move/fix MAX_WIDTH/HEIGHT-related assertions
Brian Paul [Mon, 20 Feb 2012 03:08:52 +0000 (20:08 -0700)]
mesa: move/fix MAX_WIDTH/HEIGHT-related assertions

Max texture and viewport size is only limited by MAX_WIDTH/HEIGHT for swrast.

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agost/mesa: stop using MAX_WIDTH in st_cb_texture.c
Brian Paul [Mon, 20 Feb 2012 03:08:51 +0000 (20:08 -0700)]
st/mesa: stop using MAX_WIDTH in st_cb_texture.c

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agost/mesa: stop using MAX_WIDTH in st_cb_drawpixels.c
Brian Paul [Mon, 20 Feb 2012 03:08:51 +0000 (20:08 -0700)]
st/mesa: stop using MAX_WIDTH in st_cb_drawpixels.c

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agomesa: stop using MAX_WIDTH in glReadPixels code
Brian Paul [Mon, 20 Feb 2012 03:08:51 +0000 (20:08 -0700)]
mesa: stop using MAX_WIDTH in glReadPixels code

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agomesa: stop using MAX_WIDTH in texstore code
Brian Paul [Mon, 20 Feb 2012 03:08:51 +0000 (20:08 -0700)]
mesa: stop using MAX_WIDTH in texstore code

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agou_blitter: decrease minimum buffer range alignment to 4 for copy_buffer
Marek Olšák [Fri, 24 Feb 2012 01:25:25 +0000 (02:25 +0100)]
u_blitter: decrease minimum buffer range alignment to 4 for copy_buffer

12 years agou_blitter: check for invalid values in copy_buffer
Marek Olšák [Fri, 24 Feb 2012 01:24:22 +0000 (02:24 +0100)]
u_blitter: check for invalid values in copy_buffer

12 years agor600g: rework queries
Marek Olšák [Tue, 21 Feb 2012 23:25:55 +0000 (00:25 +0100)]
r600g: rework queries

We always mapped the query buffer in begin_query, causing stalls
if the buffer was busy.

This commit reworks it such that the query buffer is only mapped
in get_query_result as it's supposed to be.

The query buffer is no longer treated as a ring buffer. Instead, the results
are just appended and when the buffer is full, we create a new one. One query
can have more than one query buffer, though that's a very rare case.
Begin_query releases all query buffers.

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: cleanup magic numbers in set_xx_sampler
Marek Olšák [Tue, 21 Feb 2012 19:24:34 +0000 (20:24 +0100)]
r600g: cleanup magic numbers in set_xx_sampler

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: remove duplicated evergreen_context_ps_partial_flush
Marek Olšák [Tue, 21 Feb 2012 18:58:09 +0000 (19:58 +0100)]
r600g: remove duplicated evergreen_context_ps_partial_flush

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: remove duplicated evergreen_context_pipe_state_set_sampler
Marek Olšák [Tue, 21 Feb 2012 18:54:17 +0000 (19:54 +0100)]
r600g: remove duplicated evergreen_context_pipe_state_set_sampler

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: remove duplicated set_xx_resource functions
Marek Olšák [Tue, 21 Feb 2012 18:47:07 +0000 (19:47 +0100)]
r600g: remove duplicated set_xx_resource functions

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: remove duplicated function r600_state_sampler_init
Marek Olšák [Tue, 21 Feb 2012 18:40:22 +0000 (19:40 +0100)]
r600g: remove duplicated function r600_state_sampler_init

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: fixup name of evergreen-specific function
Marek Olšák [Tue, 21 Feb 2012 18:38:42 +0000 (19:38 +0100)]
r600g: fixup name of evergreen-specific function

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: simplify fail paths in create_context
Marek Olšák [Tue, 21 Feb 2012 18:03:14 +0000 (19:03 +0100)]
r600g: simplify fail paths in create_context

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agogallium/u_slab: fix possible crash in util_slab_destroy
Marek Olšák [Tue, 21 Feb 2012 17:40:11 +0000 (18:40 +0100)]
gallium/u_slab: fix possible crash in util_slab_destroy

It may happen if util_slab_create has not been called.

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: fix possible crashes in destroy_context when failing in create_context
Marek Olšák [Tue, 21 Feb 2012 17:30:18 +0000 (18:30 +0100)]
r600g: fix possible crashes in destroy_context when failing in create_context

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: consolidate common context init code
Marek Olšák [Tue, 21 Feb 2012 17:08:32 +0000 (18:08 +0100)]
r600g: consolidate common context init code

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agor600g: define GROUP_FORCE_NEW_BLOCK in common header
Marek Olšák [Tue, 21 Feb 2012 16:49:29 +0000 (17:49 +0100)]
r600g: define GROUP_FORCE_NEW_BLOCK in common header

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
12 years agost/dri: Fix crash in st/dri from dri format commit
Jakob Bornecrantz [Thu, 23 Feb 2012 17:17:37 +0000 (18:17 +0100)]
st/dri: Fix crash in st/dri from dri format commit

Introduced with 8de5c355fa2bf0f30df2c7cf39aee01e793284bf, it would probably just crash.

Signed-off-by: Jakob Bornecrantz <jakob@vmware.com>
Reviewed-by: Jesse Barnes <jbarnes@virtousgeek.org>
12 years agoglapi: Fix incorrect enum value.
Paul Berry [Wed, 15 Feb 2012 18:35:10 +0000 (10:35 -0800)]
glapi: Fix incorrect enum value.

From http://www.opengl.org/registry/specs/ARB/seamless_cube_map.txt:

    Accepted by the <cap> parameter of Enable, Disable and IsEnabled,
    and by the <pname> parameter of GetBooleanv, GetIntegerv, GetFloatv
    and GetDoublev:

    TEXTURE_CUBE_MAP_SEAMLESS                   0x884F

This caused a change in enums.c, which is manually built from the .xml
files.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
12 years agoutil: add mutex lock in u_debug_memory.c code
Brian Paul [Thu, 16 Feb 2012 19:25:22 +0000 (12:25 -0700)]
util: add mutex lock in u_debug_memory.c code

The linked list of memory allocations was not protected by a mutex.
This lead to sporadic failures with multi-threaded apps.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agosvga: no-op sampler view referencing if no change
Brian Paul [Fri, 17 Feb 2012 17:18:55 +0000 (10:18 -0700)]
svga: no-op sampler view referencing if no change

Just to be safe with ref counting and avoid atomic operations.

12 years agosvga: use pipe_sampler_view_release() to avoid segfault
Brian Paul [Thu, 16 Feb 2012 19:23:16 +0000 (12:23 -0700)]
svga: use pipe_sampler_view_release() to avoid segfault

This fixes another case of faulting when freeing a pipe_sampler_view
that belongs to a previously destroyed context.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agost/wgl: check for hPbuffer=0 in wgl pbuffer functions
Brian Paul [Thu, 16 Feb 2012 19:20:57 +0000 (12:20 -0700)]
st/wgl: check for hPbuffer=0 in wgl pbuffer functions

Per the GL_WGL_pbuffer spec, generate ERROR_INVALID_HANDLE and
return FALSE (and don't segfault).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agost/wgl: Fix argument of stw_pixelformat_get_info().
José Fonseca [Wed, 8 Feb 2012 15:46:28 +0000 (15:46 +0000)]
st/wgl: Fix argument of stw_pixelformat_get_info().

stw_pixelformat_get_info takes zero based index, not a 1 based pixel
format number.

12 years agost/mesa: use pipe_sampler_view_release()
Brian Paul [Fri, 3 Feb 2012 18:10:40 +0000 (11:10 -0700)]
st/mesa: use pipe_sampler_view_release()

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agogallium: added pipe_sampler_view_release() function
Brian Paul [Fri, 3 Feb 2012 18:09:39 +0000 (11:09 -0700)]
gallium: added pipe_sampler_view_release() function

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agost/mesa: add floating point formats in st_new_renderbuffer_fb()
Brian Paul [Fri, 3 Feb 2012 18:08:50 +0000 (11:08 -0700)]
st/mesa: add floating point formats in st_new_renderbuffer_fb()

To allow creating floating point buffers / pbuffers.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agosvga: re-enable deleted surface caching
Brian Paul [Thu, 8 Dec 2011 00:11:00 +0000 (17:11 -0700)]
svga: re-enable deleted surface caching

Basically, instead of immediately freeing deleted surfaces, hang onto
them in a cache to do quick re-allocation.  This helps when surfaces
are frequently destroyed and then reallocated a bit later.

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agosvga: implement surface cache size limit
Brian Paul [Wed, 7 Dec 2011 23:57:11 +0000 (16:57 -0700)]
svga: implement surface cache size limit

There was a SVGA_HOST_SURFACE_CACHE_BYTES symbol, but it was never
used.

Now when we go to add a newly deleted surface to the cache we check
if the cache size would be exceeded.  If so, try to free the least
recently "unused" surfaces until the cache is smaller.  If we can't
do that, simply don't cache the newly deleted surface.  The alternative
involves flushing and waiting and we don't want to do that.

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agosvga: add svga_format_size() function
Brian Paul [Thu, 8 Dec 2011 00:05:48 +0000 (17:05 -0700)]
svga: add svga_format_size() function

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agosvga: use dummy shaders when translation fails
Brian Paul [Thu, 17 Nov 2011 23:40:02 +0000 (16:40 -0700)]
svga: use dummy shaders when translation fails

Before, if shader translation failed for any reason we'd keep trying
to translate the shader over and over again during state validation.
The dummy fragment shader emits solid red so that might be visual
clue that translation is failing.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agosvga: move temp register index assertions
Brian Paul [Mon, 21 Nov 2011 21:53:46 +0000 (14:53 -0700)]
svga: move temp register index assertions

The assertion recently added in dst_register() was invalid because that
function is also (suprisingly) used to declare constant registers.

Move the assertion to the callers where we're really creating temp
registers and add some code to prevent emitting invalid temp register
indexes for release builds.

Also, update the comment for get_temp().  It didn't return -1 if it
ran out of registers and none of the callers checked for that.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agosvga: check that we don't exceed temp register limit
Brian Paul [Thu, 17 Nov 2011 23:36:26 +0000 (16:36 -0700)]
svga: check that we don't exceed temp register limit

And assert on the register index in dst_register().  The dest can
only be an output or temp reg and there's more of the later.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agoscons: Update windows llvm libraries for version 3.0.
José Fonseca [Thu, 23 Feb 2012 11:50:42 +0000 (11:50 +0000)]
scons: Update windows llvm libraries for version 3.0.

12 years agoAdd a test for the compare opcode.
Alex Corscadden [Wed, 8 Sep 2010 17:42:41 +0000 (10:42 -0700)]
Add a test for the compare opcode.

This is a test for the compare opcode (CMP).  This should draw a green triangle
in the fs-test.

12 years agor600g: Fix build when libdrm is installed to non-standard dir.
Satyajit Sarangi [Mon, 20 Feb 2012 06:17:56 +0000 (01:17 -0500)]
r600g: Fix build when libdrm is installed to non-standard dir.

Signed-off-by: Satyajit Sarangi<kernel.saty@gmail.com>
Signed-off-by: Tom Stellard <tstellar@gmail.com>
12 years agoi915: Initialize swrast_texture_image structure fields.
Paul Berry [Fri, 17 Feb 2012 00:00:45 +0000 (16:00 -0800)]
i915: Initialize swrast_texture_image structure fields.

Commit 980f6f1 (mesa: move gl_texture_image::Width/Height/DepthScale
fields to swrast) moved the initialization of the Width, Height, and
DepthScale fields to _swrast_alloc_texture_image_buffer().  However,
i915 doesn't call this function because it performs its own buffer
allocation.  As a result, the Width, Height, and DepthScale fields
weren't getting initialized properly, and some operations requiring
swrast would fail.

This patch ensures that Width, Height, and DepthScale are properly
initialized by separating the code that sets them into a new function,
_swrast_init_texture_image(), which is called by
intel_alloc_texture_image_buffer() as well as
_swrast_alloc_texture_image_buffer().  It also moves the
initialization of _IsPowerOfTwo into this function.

Fixes piglit test fbo/fbo-cubemap on i915.

Partially fixes https://bugs.freedesktop.org/show_bug.cgi?id=41216

This is a candidate for the 8.0 branch.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
12 years agointel: bump DRI_IMAGE extension version to 3
Jesse Barnes [Wed, 22 Feb 2012 17:52:09 +0000 (09:52 -0800)]
intel: bump DRI_IMAGE extension version to 3

To indicate support for the format query.

Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
12 years agogbm: track buffer format through DRI drivers
Jesse Barnes [Tue, 21 Feb 2012 20:53:09 +0000 (12:53 -0800)]
gbm: track buffer format through DRI drivers

GBM needs the buffer format in order to communicate with DRM and clients
for things like scanout.

So track the DRI format requested in the various back ends and use it to
return the DRI format back to GBM when requested.  GBM will then map
this into the GBM surface type (which is in turn based on the DRM fb
format list).

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
12 years agotests/graw: Add a bunch of tests.
José Fonseca [Wed, 22 Feb 2012 15:22:31 +0000 (15:22 +0000)]
tests/graw: Add a bunch of tests.

These were rotting in an internal branch, but contain nothing confidential,
and would be much more useful if kept up-to-date with latest gallium
interface changes.

Several authors including Keith Whitwell, Zack Rusin, and Brian Paul.

12 years agomesa: add special case in texstore.c for GL_LUMINANCE_ALPHA src image
Brian Paul [Thu, 17 Nov 2011 23:03:35 +0000 (16:03 -0700)]
mesa: add special case in texstore.c for GL_LUMINANCE_ALPHA src image

About a 10% improvement over the swizzle-copy path.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
12 years agogallium/cso: Put the comment about shader in the code for future reference.
José Fonseca [Wed, 22 Feb 2012 08:41:19 +0000 (08:41 +0000)]
gallium/cso: Put the comment about shader in the code for future reference.

12 years agoi965/gen6: Fix near-NULL deref in setting up GS binding table for non-XFB.
Eric Anholt [Tue, 21 Feb 2012 21:17:40 +0000 (13:17 -0800)]
i965/gen6: Fix near-NULL deref in setting up GS binding table for non-XFB.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Correct the size of the state batch space allocated for binding tables.
Eric Anholt [Tue, 21 Feb 2012 21:05:57 +0000 (13:05 -0800)]
i965: Correct the size of the state batch space allocated for binding tables.

In the gen6 GS case, we were under-counting and so other state would
get smashed.  In the VS case, we were over-counting, so everything was
fine.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Fix a bad comment in gen6 sol setup.
Eric Anholt [Tue, 21 Feb 2012 21:04:12 +0000 (13:04 -0800)]
i965: Fix a bad comment in gen6 sol setup.

This was copy and paste from the VS where I had similar code.  We're
only looking at things derived from BRW_NEW_VERTEX_PROGRAM in this
block.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965/gen6: Fix the size of the GS surface binding table.
Eric Anholt [Tue, 21 Feb 2012 20:56:29 +0000 (12:56 -0800)]
i965/gen6: Fix the size of the GS surface binding table.

I obviously didn't test on gen6 before pushing.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agor600g: move invariant register updates into start_cs for evergreen and cayman
Marek Olšák [Wed, 8 Feb 2012 22:25:58 +0000 (23:25 +0100)]
r600g: move invariant register updates into start_cs for evergreen and cayman

12 years agor600g: move invariant register updates into start_cs for r6xx-r7xx
Marek Olšák [Fri, 3 Feb 2012 04:05:31 +0000 (05:05 +0100)]
r600g: move invariant register updates into start_cs for r6xx-r7xx

12 years agor600g: add a depth misc state which depends on occlusion queries
Marek Olšák [Thu, 2 Feb 2012 13:01:12 +0000 (14:01 +0100)]
r600g: add a depth misc state which depends on occlusion queries

This is a state which is derived from other states and is actually the first
state which doesn't correspond to any gallium state.

There are two state flags:
  bool occlusion_query_enabled
  bool flush_depthstencil_enabled

Additional flags can be added later if needed, e.g. bool hiz_enabled.
The emit function will have to figure out the register values by itself.

It basically just emits the registers when the state changes.
This commit also adds a few helper functions for writing registers directly
into a command stream.

12 years agor600g: consolidate the main draw code
Marek Olšák [Thu, 2 Feb 2012 09:04:29 +0000 (10:04 +0100)]
r600g: consolidate the main draw code

The code was almost the same for r600 and eg. What can't be consolidated is
in the *_prepare functions.

12 years agor600g: move all invariant state from draw_vbo into start_cs
Marek Olšák [Thu, 2 Feb 2012 07:27:01 +0000 (08:27 +0100)]
r600g: move all invariant state from draw_vbo into start_cs

12 years agor600g: turn init_config into a command buffer for starting a CS
Marek Olšák [Tue, 31 Jan 2012 09:50:51 +0000 (10:50 +0100)]
r600g: turn init_config into a command buffer for starting a CS

This is the first pure command buffer. It contains CS initialization
packets and emits invariant state (i.e. the registers which never or rarely
change).

The affected registers are removed from *_hw_context.c, so that both ways
of emitting commands can co-exist.

v2: emit context_control in cayman's start_cs too

12 years agoi965: Only set Last Render Target Select on the last FB write.
Kenneth Graunke [Sun, 19 Feb 2012 05:29:29 +0000 (21:29 -0800)]
i965: Only set Last Render Target Select on the last FB write.

Fixes GPU hangs in OilRush, Trine, and Amnesia: The Dark Descent,
which all use MRT (multiple render targets).

NOTE: This is a candidate for release branches.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38720
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=40059
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=45216
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agogallium/cso: kill off non-functional shader caching
Marek Olšák [Fri, 17 Feb 2012 01:13:32 +0000 (02:13 +0100)]
gallium/cso: kill off non-functional shader caching

Suggested by José.

We don't provide shader caching in CSO. Most of the time the api provides
object semantics for shaders anyway, and the cases where it doesn't
(eg mesa's internall-generated texenv programs), it will be up to
the state tracker to implement their own specialized caching.

12 years agogallium/util: remove u_simple_screen
Marek Olšák [Fri, 17 Feb 2012 00:54:37 +0000 (01:54 +0100)]
gallium/util: remove u_simple_screen

Deprecated and unused.

12 years agogallium: remove unused winsys pointers in pipe_screen and pipe_context
Marek Olšák [Fri, 17 Feb 2012 00:49:49 +0000 (01:49 +0100)]
gallium: remove unused winsys pointers in pipe_screen and pipe_context

A winsys is already a private object of a driver.

12 years agointel: Silence valgrind warning for getparam ioctl argument.
Eric Anholt [Thu, 16 Feb 2012 19:30:49 +0000 (11:30 -0800)]
intel: Silence valgrind warning for getparam ioctl argument.

It was concerned that the 4 pad bytes on LP64 were uninitialized.

12 years agoi965: Rename the original binding table to mention that it's the WM now.
Eric Anholt [Wed, 15 Feb 2012 22:24:37 +0000 (14:24 -0800)]
i965: Rename the original binding table to mention that it's the WM now.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Split the gen6 GS binding table to a separate table.
Eric Anholt [Wed, 15 Feb 2012 22:15:14 +0000 (14:15 -0800)]
i965: Split the gen6 GS binding table to a separate table.

Improves VS state change microbenchmark performance by 7.08729% +/-
1.22289% (n=10) on gen7, because we don't upload the 64 dwords of
unused binding table any more.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Split the VS binding table to a separate table.
Eric Anholt [Wed, 15 Feb 2012 21:33:07 +0000 (13:33 -0800)]
i965: Split the VS binding table to a separate table.

This is a step toward making the samplers/binding tables reflect
sampler uniform mappings instead of embedding those in the programs.
No significant performance difference on the microbenchmark (n=10).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965/gen6+: Avoid recomputing whether we use noperspective.
Eric Anholt [Wed, 15 Feb 2012 20:57:25 +0000 (12:57 -0800)]
i965/gen6+: Avoid recomputing whether we use noperspective.

Improves VS state change microbenchmark performance 2.38246% +/-
1.15046% (n=20).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965/gen7: Skip checking if we need a GS program for now.
Eric Anholt [Wed, 15 Feb 2012 18:55:02 +0000 (10:55 -0800)]
i965/gen7: Skip checking if we need a GS program for now.

We always say no.  Improves VS state change microbenchmark performance
7.68747% +/- 1.40826% (n=10).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Compute required barycentric interp modes once at FS compile time.
Eric Anholt [Wed, 15 Feb 2012 05:48:43 +0000 (21:48 -0800)]
i965: Compute required barycentric interp modes once at FS compile time.

Improves VS state change microbenchmark performance 1.78817% +/-
0.556878% (n=25).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Move VUE map computation to once at VS compile time.
Eric Anholt [Mon, 13 Feb 2012 18:08:23 +0000 (10:08 -0800)]
i965: Move VUE map computation to once at VS compile time.

With this and the previous patch, 640x480 nexuiz is running 0.169118%
+/- 0.0863696% faster (n=121).  On a VS state change microbenchmark,
performance is increased 8.28645% +/- 0.460478% (n=52).

v2: Fix CACHE_NEW_VS comment.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agoi965: Make the userclip flag for the VUE map come from VS prog data.
Eric Anholt [Mon, 13 Feb 2012 18:00:14 +0000 (10:00 -0800)]
i965: Make the userclip flag for the VUE map come from VS prog data.

This reduces recomputation of state based on non-clipping-related
transform changes, and is a step toward removing VUE map
recomputation.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
12 years agomesa: fix issues with texture border and array textures
Anuj Phogat [Wed, 15 Feb 2012 23:40:51 +0000 (15:40 -0800)]
mesa: fix issues with texture border and array textures

For a 1D texture array, the border only applies to the width.  For a 2D
texture array the border applies to the width and height but not the depth.
Sucha cases were  not handled correctly in _mesa_init_teximage_fields().

Note: This is a candidate for stable branches

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
12 years agoradeon/r600g: fix virtual address space allocation
Jerome Glisse [Tue, 21 Feb 2012 17:34:54 +0000 (12:34 -0500)]
radeon/r600g: fix virtual address space allocation

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
12 years agollvmpipe: Remove lp_test_sincos.
José Fonseca [Tue, 21 Feb 2012 16:04:54 +0000 (16:04 +0000)]
llvmpipe: Remove lp_test_sincos.

Completely replaced by lp_test_arit.

12 years agovl: add support for bob deinterlacing
Christian König [Fri, 10 Feb 2012 13:32:16 +0000 (14:32 +0100)]
vl: add support for bob deinterlacing

v2: return VDP_STATUS_INVALID_VIDEO_MIXER_PICTURE_STRUCTURE
    for unknown picture structure.

Signed-off-by: Christian König <deathsimple@vodafone.de>
12 years agost/vdpau: Use transfer_inline_write to upload ycbcr data
Christian König [Fri, 10 Feb 2012 11:56:27 +0000 (12:56 +0100)]
st/vdpau: Use transfer_inline_write to upload ycbcr data

Uses less code and looks at least a bit cleaner than mapping manually.

Signed-off-by: Christian König <deathsimple@vodafone.de>
12 years agost/vdpau: remove unnecessary tracing and adjust tracing levels a bit
Christian König [Fri, 10 Feb 2012 11:34:22 +0000 (12:34 +0100)]
st/vdpau: remove unnecessary tracing and adjust tracing levels a bit

Tracing function entry/exits is a bit pointless
when VDPAU_TRACE=1 does the same thing.

v2: use WARN instead of ERR for application problems

Signed-off-by: Christian König <deathsimple@vodafone.de>
12 years agogallivm: Fix TGSI_OPCODE_ARR's translation.
José Fonseca [Mon, 20 Feb 2012 20:49:03 +0000 (20:49 +0000)]
gallivm: Fix TGSI_OPCODE_ARR's translation.

Like TGSI_OPCODE_ARL, destination should be an integer.

This fixes invalid LLVM IR on an internal state tracker (currently Mesa
never emits this opcode).

In the future consider making ADDR register also a integer-as-float array,
like all other register kinds, or simply replace ADDR & ARR/ARL with
integer temp and instructions.

Reviewed-by: Dave Airlie <airlied@redhat.com>
12 years agost/egl: Move drm_display_authenticate into HAVE_WAYLAND_BACKEND section.
Vinson Lee [Sat, 18 Feb 2012 19:16:18 +0000 (11:16 -0800)]
st/egl: Move drm_display_authenticate into HAVE_WAYLAND_BACKEND section.

Fixes this GCC warning.
native_drm.c:153:1: warning: ‘drm_display_authenticate’ defined but not
used [-Wunused-function]

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agollvmpipe: Don't assume vector is 4 wide in lp_build_sin()/lp_build_cos()
José Fonseca [Fri, 17 Feb 2012 16:22:18 +0000 (16:22 +0000)]
llvmpipe: Don't assume vector is 4 wide in lp_build_sin()/lp_build_cos()

Reviewed-by: Dave Airlie <airlied@redhat.com>
12 years agomesa: check for no state change in glPrimitiveRestartIndex()
Brian Paul [Mon, 20 Feb 2012 02:50:32 +0000 (19:50 -0700)]
mesa: check for no state change in glPrimitiveRestartIndex()

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agomesa: check for no state change in Enable/DisableVertexAttribArray()
Brian Paul [Mon, 20 Feb 2012 02:50:31 +0000 (19:50 -0700)]
mesa: check for no state change in Enable/DisableVertexAttribArray()

Avoid setting dirty state flags when enabling or disabling a vertex
attribute arrays when there's no change.

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agomesa: add missing return after _mesa_error() in update_array()
Brian Paul [Mon, 20 Feb 2012 02:50:31 +0000 (19:50 -0700)]
mesa: add missing return after _mesa_error() in update_array()

NOTE: This is a candidate for the 8.0 branch.

Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agor300g: Fix build when libdrm is installed to non-standard dir
Tom Stellard [Mon, 20 Feb 2012 00:37:30 +0000 (19:37 -0500)]
r300g: Fix build when libdrm is installed to non-standard dir

12 years agoi965: Make the dummy fragment shader work in SIMD16 mode.
Kenneth Graunke [Sat, 18 Feb 2012 21:08:29 +0000 (13:08 -0800)]
i965: Make the dummy fragment shader work in SIMD16 mode.

If you're resorting to the dummy shader, you've probably already turned
off SIMD16 mode.  But if you didn't, it would die in a fire.

We could either fail to compile in SIMD16 mode...or just fix it.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
12 years agoi965: Fix GPU hangs in the dummy fragment shader.
Kenneth Graunke [Sat, 18 Feb 2012 20:59:40 +0000 (12:59 -0800)]
i965: Fix GPU hangs in the dummy fragment shader.

The dummy FB write failed to specify EOT and a message length, causing
the GPU to hang.  Now we can enjoy "everyone's favorite color" again.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
12 years agost/vega: Remove unused variable.
Vinson Lee [Sat, 18 Feb 2012 08:32:53 +0000 (00:32 -0800)]
st/vega: Remove unused variable.

Fixes this GCC warning.
mask.c: In function ‘mask_layer_fill’:
mask.c:387:12: warning: variable ‘alpha_color’ set but not used
[-Wunused-but-set-variable]

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agost/glx: Remove unused variables.
Vinson Lee [Sat, 18 Feb 2012 08:18:52 +0000 (00:18 -0800)]
st/glx: Remove unused variables.

Fixes these GCC warnings.
glx_api.c: In function ‘choose_visual’:
glx_api.c:678:8: warning: variable ‘trans_value’ set but not used
[-Wunused-but-set-variable]
glx_api.c:677:8: warning: variable ‘trans_type’ set but not used
[-Wunused-but-set-variable]
glx_api.c:663:8: warning: variable ‘min_ci’ set but not used
[-Wunused-but-set-variable]

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
12 years agodraw: add missing streamout state setup for draw/llvm.
Dave Airlie [Sat, 18 Feb 2012 12:43:37 +0000 (12:43 +0000)]
draw: add missing streamout state setup for draw/llvm.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agovbo: Eliminate short-circuiting in invalid-start case.
Kenneth Graunke [Wed, 8 Feb 2012 13:08:24 +0000 (05:08 -0800)]
vbo: Eliminate short-circuiting in invalid-start case.

Now that we have a index_range_invalid flag, we can just use that rather
than calling vbo_validated_drawrangeelements directly and returning.

NOTE: This is a candidate for release branches.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
12 years agovbo: Rework checking of 'end' against _MaxElement.
Kenneth Graunke [Wed, 8 Feb 2012 13:08:23 +0000 (05:08 -0800)]
vbo: Rework checking of 'end' against _MaxElement.

This failed to take basevertex into account:

If basevertex < 0:
   (end + basevertex) might actually be in-bounds while 'end' is not.
   We would have clamped in this case when we probably shouldn't.
   This could break application drawing.

If basevertex > 0:
   'end' might be in-bounds while (end + basevertex) might not.
   We would have failed to clamp in this place.  There's a comment
   indicating the TNL module depends on max_index being in-bounds;
   if so, it would likely break horribly.

Rather than trying to clamp correctly in the face of basevertex, simply
delete the clamping code and indicate that we don't have a valid range.
This causes _tnl_vbo_draw_prims to use vbo_get_minmax_indices() to
compute the actual bounds, which is much safer.

NOTE: This is a candidate for release branches.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
12 years agovbo: Ignore invalid element ranges which are outside VBO bounds.
Kenneth Graunke [Wed, 8 Feb 2012 13:08:22 +0000 (05:08 -0800)]
vbo: Ignore invalid element ranges which are outside VBO bounds.

Some applications, such as Regnum Online, appear to pass invalid
start/end values to glDrawRangeElements.  In particular, the 'start'
index sometimes exceeds the maximum array element.  This is clearly
invalid behavior, and although the spec isn't clear, seems to result
in undefined, implementation-specific behavior.

This patch takes the conservative approach and simply ignores the range,
while issuing a warning indicating that the application is broken and
should be fixed.

NOTE: This is a candidate for release branches.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=45214
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=44701
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=41152
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=40361
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=28138
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com> [v1]
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
12 years agovbo: Remove pedantic warning about 'end' beind out of bounds.
Kenneth Graunke [Wed, 8 Feb 2012 13:08:21 +0000 (05:08 -0800)]
vbo: Remove pedantic warning about 'end' beind out of bounds.

The application supplied [start, end] range is merely a conservative
hint of the ranges of index values inside the index buffer.  There is no
requirement that all vertices in the range [start, end] be referenced.

Passing an 'end' value larger than the maximum legal index is perfectly
acceptible; applications can legally pass 0xffffffff when they don't
have a tighter bound readily available.

Thus, the warning doesn't indicate a correctness issue; it could only
indicate a performance issue.  However, it does not even do that.

glDrawRangeElements is designed to optimize non-VBO vertex data uploads
by providing an upper bound on the size of buffers a driver would need
to allocate.  With VBOs, the data is already in an uploaded buffer, so
the range doesn't help.

The clincher is: we only know _MaxElement for VBOs.  For user-space
arrays, we just set it to 2,000,000,000 (see mesa/main/varray.h:63.)
So we can only check this in the case where it is not useful.

Many applications, including the Unigine demos, currently trigger this
warning, which suggests the applications are buggy when they're actually
fine.  Eliminating the warning should confuse users less while not
actually losing any benefit to application developers.

NOTE: This is a candidate for release branches.

Suggested-by: Jose Fonseca <jfonseca@vmware.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
12 years agointel: Fix rendering from textures after RenderTexture().
Eric Anholt [Fri, 10 Feb 2012 20:54:25 +0000 (12:54 -0800)]
intel: Fix rendering from textures after RenderTexture().

There's a serious trap for drivers: RenderTexture() does not indicate
that the texture is currently bound to the draw buffer, despite
FinishRenderTexture() signaling that the texture is just now being
unbound from the draw buffer.

We were acting as if RenderTexture() *was* the start of rendering and
that we could make texturing incoherent with the current contents of
the renderbuffer.  This caused intel oglconform sRGB
Mipmap.1D_textures to fail, because we got a call to TexImage() and
thus RenderTexture() on a texture bound to a framebuffer that wasn't
the draw buffer, so we skipped validating the new image into the
texture object used for rendering.

We can't (easily) make RenderTexture() indicate the start of drawing,
because both our driver and gallium are using it as the moment to set
up the renderbuffer wrapper used for things like MapRenderbuffer().
Instead, postpone the setup of the workaround render target miptree
until update_renderbuffer time, so that we no longer need to skip
validation of miptrees used as render targets.  As a bonus, this
should make GL_NV_texture_barrier possible.

(This also fixes a regression in the gen4 small-mipmap rendering since
3b38b33c1648b07e75dc4d8340758171e109c598, which switched
set_draw_offset from image->mt to irb->mt but didn't move the irb->mt
replacement up before set_draw_offset).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=44961
NOTE: This is a candidate for the 8.0 branch.

12 years agointel: Improve the fallback debug for framebuffer status checks.
Eric Anholt [Fri, 10 Feb 2012 20:24:50 +0000 (12:24 -0800)]
intel: Improve the fallback debug for framebuffer status checks.

12 years agomesa: Only end render-to-texture at bind time for drawbuffers.
Eric Anholt [Fri, 10 Feb 2012 20:05:16 +0000 (12:05 -0800)]
mesa: Only end render-to-texture at bind time for drawbuffers.

If we're only starting for new draw buffers, why would we end for old
read buffers along with draw buffers?

12 years agogallivm: fetch immediates to correct type (v2)
Dave Airlie [Mon, 6 Feb 2012 18:48:17 +0000 (18:48 +0000)]
gallivm: fetch immediates to correct type (v2)

Fetch float/uint/int immediates.

v2: bitcast to uint/int to floats as per Jose's suggestions.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: enable stores of integer types. (v2) + fix ARL
Dave Airlie [Mon, 6 Feb 2012 15:37:56 +0000 (15:37 +0000)]
gallivm: enable stores of integer types. (v2) + fix ARL

Infer from the operand the type of value to store.
MOV is untyped but we use the float store path.

v2: make MOV use float store path.

I've had to squash merge the ARL fix to be stored
as an integer in here to avoid regressions in a number
of piglit tests.

From now on ARL stores to an integer just like HW does.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: enable fetch for integer opcodes. (v2)
Dave Airlie [Mon, 6 Feb 2012 15:35:05 +0000 (15:35 +0000)]
gallivm: enable fetch for integer opcodes. (v2)

The infers the type of data required using the opcode,
and casts the input to the appropriate type.

So far this only handles non-indirect constant and temporaries.

v2: as per Jose suggestion, fetch immediates via floats

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: add uint/int bld to the base builder. (v2)
Dave Airlie [Mon, 6 Feb 2012 15:28:45 +0000 (15:28 +0000)]
gallivm: add uint/int bld to the base builder. (v2)

These are used inside the action handlers for the integer opcodes.

v2: use uint_bld/int_bld, drop higher level uint_bld.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: fix build gather to take a bld context
Dave Airlie [Tue, 7 Feb 2012 09:52:00 +0000 (09:52 +0000)]
gallivm: fix build gather to take a bld context

Then pass the correct build context to it.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: pass build context to exec_mask_store.
Dave Airlie [Mon, 6 Feb 2012 13:35:46 +0000 (13:35 +0000)]
gallivm: pass build context to exec_mask_store.

For now just pass the current context, but when we want to
store int or unsigned we need to pass those later.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agotgsi: add source/destination type from opcodes. (v2)
Dave Airlie [Tue, 24 Jan 2012 11:55:14 +0000 (11:55 +0000)]
tgsi: add source/destination type from opcodes. (v2)

These two functions produce the src/dst types for an opcode.

MOV is special since it can be used to mov float->float and int->int,
so just return VOID.

v2: use a new enum for the opcode type as per Jose's suggestion.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agogallivm: Initialize x86 disassembler on x86_64 too.
José Fonseca [Fri, 17 Feb 2012 17:22:23 +0000 (17:22 +0000)]
gallivm: Initialize x86 disassembler on x86_64 too.