mesa.git
9 years agoconfigure.ac: roll up a program for the sse4.1 check
Emil Velikov [Sat, 15 Nov 2014 18:37:22 +0000 (18:37 +0000)]
configure.ac: roll up a program for the sse4.1 check

So when checking/building sse code we have three possibilities:
 1 Old compiler, throws an error when using -msse*
 2 New compiler, user disables sse* (-mno-sse*)
 3 New compiler, user doesn't disable sse

The original code, added code for #1 but not #2. Later on we patched
around the lack of handling #2 by wrapping the code in __SSE4_1__.
Yet it lead to a missing/undefined symbol in case of #1 or #2, which
might cause an issue for #2 when using the i965 driver.

A bit later we "fixed" the undefined symbol by using #1, rather than
updating it to handle #2. With this commit we set things straight :)

To top it all up, conventions state that in case of conflicting
(-enable-foo -disable-foo) options, the latter one takes precedence.
Thus we need to make sure to prepend -msse4.1 to CFLAGS in our test.

v2: Clean the #includes. Suggested by Ilia, Matt & Siavash.

Cc: "10.3 10.4" <mesa-stable@lists.freedesktop.org>
Tested-by: David Heidelberg <david@ixit.cz>
Tested-by: Siavash Eliasi <siavashserver@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
9 years agonv50,nvc0: use clip_halfz setting when creating rasterizer state
Ilia Mirkin [Thu, 23 Oct 2014 04:43:45 +0000 (00:43 -0400)]
nv50,nvc0: use clip_halfz setting when creating rasterizer state

This enables the ARB_clip_control extension.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
9 years agofreedreno: add adreno 420 support
Rob Clark [Thu, 31 Jul 2014 19:42:55 +0000 (15:42 -0400)]
freedreno: add adreno 420 support

Very initial support.  Basic stuff working (es2gears, es2tri, and maybe
about half of glmark2).  Expect broken stuff.  Still missing: mem->gmem
(restore), queries, mipmaps (blob segfaults!), hw binning, etc.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agofreedreno: update generated headers
Rob Clark [Thu, 31 Jul 2014 19:09:52 +0000 (15:09 -0400)]
freedreno: update generated headers

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agoi965: Move fs_visitor ra pass to new fs_visitor::allocate_registers()
Kristian Høgsberg [Fri, 14 Nov 2014 00:28:19 +0000 (16:28 -0800)]
i965: Move fs_visitor ra pass to new fs_visitor::allocate_registers()

This will be reused for the scalar VS pass.

v2 (Ken): Rebase on master.

Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agoi965: Move fs_visitor optimization pass into new method fs_visitor::optimize()
Kristian Høgsberg [Fri, 14 Nov 2014 00:28:18 +0000 (16:28 -0800)]
i965: Move fs_visitor optimization pass into new method fs_visitor::optimize()

We'll reuse this toplevel optimization driver for the scalar VS.

Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agoi965: Move more code into codegen-branch of the fs_visitor::run() if statement
Kristian Høgsberg [Fri, 14 Nov 2014 00:28:17 +0000 (16:28 -0800)]
i965: Move more code into codegen-branch of the fs_visitor::run() if statement

These last few operations all only apply when we've actually generated
code, optimized and allocated registers.  The dummy and the repclear
shaders don't need the gen4 send workaround, and don't spill.  This
means we can move these lines into the else-branch, which will make
the following refactoring easier.

v2 (Ken): Rebase on master, which removed the uncompressed stack.

Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agoi965: Refactor fs_generator API
Kristian Høgsberg [Fri, 14 Nov 2014 00:28:08 +0000 (16:28 -0800)]
i965: Refactor fs_generator API

We split out SIMD8 and SIMD16 generation into seperate calls to
new method generate_code(), which returns the start offset for the
generated code.  A new get_assembly() method returns the generated code.

This avoids asserting MESA_SHADER_FRAGMENT and accessing wm_prog_data
in the generator.

Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agost/wgl: Implement WGL_EXT_create_context_es/es2_profile.
José Fonseca [Fri, 14 Nov 2014 17:20:21 +0000 (17:20 +0000)]
st/wgl: Implement WGL_EXT_create_context_es/es2_profile.

Derived from st/glx's GLX_EXT_create_context_es/es2_profile implementation.

Tested with an OpenGL ES 2.0 ApiTrace.

Reviewed-by: Brian Paul <brianp@vmware.com>
9 years agost/xlib: Generate errors as specified.
José Fonseca [Fri, 14 Nov 2014 17:04:26 +0000 (17:04 +0000)]
st/xlib: Generate errors as specified.

Tested with piglit glx tests.

Reviewed-by: Brian Paul <brianp@vmware.com>
9 years agofreedreno/ir3: move some helpers
Rob Clark [Wed, 5 Nov 2014 17:08:04 +0000 (12:08 -0500)]
freedreno/ir3: move some helpers

Split out a few helpers from fd3_program so we don't have to duplicate
for fd4_program.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agofreedreno: rename draw->draw_vbo
Rob Clark [Wed, 5 Nov 2014 15:42:13 +0000 (10:42 -0500)]
freedreno: rename draw->draw_vbo

Gets rid of a namespace conflict w/ a4xx which wants an fd4_draw()
version of fd_draw()..

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agofreedreno/a3xx: missing u_upload_destroy
Rob Clark [Wed, 5 Nov 2014 15:43:18 +0000 (10:43 -0500)]
freedreno/a3xx: missing u_upload_destroy

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agofreedreno: fix borked check for a320.0
Rob Clark [Thu, 6 Nov 2014 20:56:43 +0000 (15:56 -0500)]
freedreno: fix borked check for a320.0

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agofreedreno/ir3: half vs full reg in standalone compiler output
Rob Clark [Sat, 1 Nov 2014 13:50:18 +0000 (09:50 -0400)]
freedreno/ir3: half vs full reg in standalone compiler output

Handle hrN.c in printing outputs/inputs.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
9 years agost/dri: Support EGL_CONTEXT_OPENGL_DEBUG_BIT_KHR/GLX_CONTEXT_DEBUG_BIT_ARB on ES...
José Fonseca [Wed, 12 Nov 2014 12:13:42 +0000 (12:13 +0000)]
st/dri: Support EGL_CONTEXT_OPENGL_DEBUG_BIT_KHR/GLX_CONTEXT_DEBUG_BIT_ARB on ES contexts.

The latest version of the specs explicitly allow it, and given that Mesa
universally supports KHR_debug we should definitely support it.

Totally untested.  (Just happened to noticed this while implementing
GLX_EXT_create_context_es2_profile for st/xlib.)

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
9 years agoegl: remove egl_gallium from the loader
Marek Olšák [Sat, 1 Nov 2014 13:31:09 +0000 (14:31 +0100)]
egl: remove egl_gallium from the loader

Acked-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Informally acked-by: Jose Fonseca

9 years agoconfigure.ac: remove enable flags for EGL and GBM Gallium state trackers
Marek Olšák [Sat, 1 Nov 2014 13:05:16 +0000 (14:05 +0100)]
configure.ac: remove enable flags for EGL and GBM Gallium state trackers

Acked-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Informally acked-by: Jose Fonseca

9 years agoi965/vec4: Combine all the math emitters.
Kenneth Graunke [Tue, 11 Nov 2014 22:40:08 +0000 (14:40 -0800)]
i965/vec4: Combine all the math emitters.

17 insertions(+), 102 deletions(-).  Works just as well.

v2: Make emit_math take const references (suggested by Matt),
    drop redundant WRITEMASK_XYZW setting (Matt and Curro).

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
9 years agoi965/vec4: Use const references in emit() functions.
Kenneth Graunke [Tue, 11 Nov 2014 01:20:37 +0000 (17:20 -0800)]
i965/vec4: Use const references in emit() functions.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
9 years agoi965: Use macros to create prototypes for emitter helpers.
Kenneth Graunke [Tue, 11 Nov 2014 01:14:48 +0000 (17:14 -0800)]
i965: Use macros to create prototypes for emitter helpers.

We do this almost everywhere else; this should make it easier to modify.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
9 years agoi965: Always enable VF statistics
Ben Widawsky [Thu, 13 Nov 2014 02:49:16 +0000 (18:49 -0800)]
i965: Always enable VF statistics

Every other unit in the geometry pipeline automatically enables
statistics gathering. This part of the pipe has been controlled by the
DEBUG_STATS variable, but this is asymmetric. This dates back to the
original implementation, and I am not sure if there is a reason for it.

I need access to these stats to implement ARB_pipeline_statistics_query.

Eric wrote it, and Ken touched it last. Do you have any opposition?

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86145
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
9 years agoi915: Emit 3DSTATE_SCISSOR_RECTANGLE_0 before 3DSTATE_SCISSOR_ENABLE
Ville Syrjälä [Sun, 6 Jul 2014 12:02:15 +0000 (15:02 +0300)]
i915: Emit 3DSTATE_SCISSOR_RECTANGLE_0 before 3DSTATE_SCISSOR_ENABLE

According to gen2 BSpec the pipeline must be flushed at least up to the
windower before changing the scissor rect enable field. Emitting the
3DSTATE_SCISSOR_RECTANGLE_0 before 3DSTATE_SCISSOR_ENABLE is sufficient
to do that.

gen3 BSpec no longer has that piece of text, but let's make the same
change there too for symmetry. The spec does still say that the scissor
rectangle must be defined before enabling it, so the new order does seem
more in line with the spec.

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Don't call _mesa_meta_glsl_Clear() on gen2
Ville Syrjälä [Tue, 8 Jul 2014 13:13:58 +0000 (16:13 +0300)]
i915: Don't call _mesa_meta_glsl_Clear() on gen2

Gen2 doesn't have fragment shaders so we shouldn't be calling
_mesa_meta_glsl_Clear() on gen2. Restore the appropriate
ARB_fragment_shader check to the clear path which was lost in:

 commit 94f22fbe787214580a1a13a774114d2650c166cb
 Author: Tapani Pälli <tapani.palli@intel.com>
 Date:   Wed Aug 8 20:46:45 2012 +0300

    intel: use _mesa_meta_Clear with OpenGL ES 1.1 v2

v2: Fix spelling in commit message

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Protect macro argument for TEXTURE_SET()
Ville Syrjälä [Mon, 30 Jun 2014 15:42:31 +0000 (18:42 +0300)]
i915: Protect macro argument for TEXTURE_SET()

TEXTURE_SET() is the only register macro that forgets to wrap the
argument evaluation in parens. Only simple integers are passed to this
macro so there's no bug but sitll it seems prudent to add the
parens.

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Kill intel_context::hw_stencil
Ville Syrjälä [Sun, 6 Jul 2014 12:10:44 +0000 (15:10 +0300)]
i915: Kill intel_context::hw_stencil

ctx.hw_stencil is not used anywhere so kill it.

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Accept GL_DEPTH_STENCIL GL_DEPTH_COMPONENT formats for renderbuffers
Ville Syrjälä [Thu, 3 Jul 2014 12:38:07 +0000 (15:38 +0300)]
i915: Accept GL_DEPTH_STENCIL GL_DEPTH_COMPONENT formats for renderbuffers

Gen2 doesn't support depth/stencil textures, and since

 commit c1d4d4999303f9167b20f4e0674b9436e6295cf7
 Author: Ville Syrjälä <ville.syrjala@linux.intel.com>
 Date:   Thu Apr 24 14:11:43 2014 +0300

    i915: Don't advertise Z formats in TextureFormatSupported on gen2

depth/stencil formats are no longer accepted as texture formats.
However we still want depth/stencil renderbuffers, so add explicit
format checks to intel_alloc_renderbuffer_storage() to allow such
things.

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Override mip filter to nearest with aniso
Ville Syrjälä [Tue, 8 Jul 2014 12:34:27 +0000 (15:34 +0300)]
i915: Override mip filter to nearest with aniso

gen2 doesn't supporte linear mip filter with anisotropic min/mag
filtering. The hardware would automagically downgrade the min/mag
filters to linear in such cases, which IMO looks worse than forcing
the mip filter to nearest.

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Use L8A8 instead of I8 to simulate A8 on gen2
Ville Syrjälä [Thu, 7 Aug 2014 08:14:35 +0000 (11:14 +0300)]
i915: Use L8A8 instead of I8 to simulate A8 on gen2

Gen2 doesn't support the A8 texture format. Currently the driver
substitutes it with I8, but that results in incorrect RGB values.
Use A8L8 instead. We end up wasting a bit of memory, but at least
we should get the correct results.

v2: Handle the fallback in _mesa_choose_tex_format() and also
    do it for all alpha formats that currently accept A8

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=72819
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80050
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38873
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agoi915: Fix GL_DOT3_RGBA a bit
Ville Syrjälä [Thu, 3 Jul 2014 12:55:38 +0000 (15:55 +0300)]
i915: Fix GL_DOT3_RGBA a bit

The spec says using DOT4 for alpha is undefined unless DOT4 is also used
for color. It seems to do the right thing anyway, but better safe than sorry.

Also override numAlphaArgs to 2 for DOT4 since that's what it wants.
This migth fix something in case the specified alpha mode has only one
argument. Also avoids emitting a needless 3DSTATE_MAP_BLEND_ARG if
the specified alpha mode has three arguments.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
9 years agolinker: Add a missing space in an error message
Neil Roberts [Thu, 13 Nov 2014 15:31:44 +0000 (15:31 +0000)]
linker: Add a missing space in an error message

Reviewed-by: Brian Paul <brianp@vmware.com>
9 years agollvmpipe: Call pipe_thread_wait() on Linux.
José Fonseca [Thu, 13 Nov 2014 15:01:19 +0000 (15:01 +0000)]
llvmpipe: Call pipe_thread_wait() on Linux.

To address http://lists.freedesktop.org/archives/mesa-dev/2014-November/070569.html

In short, revert 706ad3b649e6a75fdac9dc9acc3caa9e6067b853 for non-Windows
OSes.

9 years agoi915g: we also have more than 0 viewports!
Kenneth Graunke [Wed, 12 Nov 2014 01:08:22 +0000 (17:08 -0800)]
i915g: we also have more than 0 viewports!

See 546d6c8d for the corresponding fix in freedreno.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Stephane Marchesin <stephane.marchesin@gmail.com>
Cc: "10.3" <mesa-stable@lists.freedesktop.org>
9 years agovc4: Avoid reusing a pointer from c->outputs[] after add_output().
Eric Anholt [Wed, 12 Nov 2014 22:14:32 +0000 (14:14 -0800)]
vc4: Avoid reusing a pointer from c->outputs[] after add_output().

add_output() can resize the qreg array, so we might use a stale pointer.

9 years agovc4: Fix assumption of TGSI OUT[0] being POSITION in the VS.
Eric Anholt [Fri, 7 Nov 2014 22:26:32 +0000 (14:26 -0800)]
vc4: Fix assumption of TGSI OUT[0] being POSITION in the VS.

All the shaders we've received so far had this be the case, but with
nir-to-tgsi that changed.

I might decide to make nir-to-tgsi keep the outputs in the same order, for
debugging sanity, but I'm not sure.

9 years agonvc0: remove unused mm_VRAM_fe0
Ilia Mirkin [Thu, 16 Oct 2014 06:58:28 +0000 (02:58 -0400)]
nvc0: remove unused mm_VRAM_fe0

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
9 years agost/glx: Implement GLX_EXT_create_context_es2_profile.
José Fonseca [Wed, 12 Nov 2014 12:18:41 +0000 (12:18 +0000)]
st/glx: Implement GLX_EXT_create_context_es2_profile.

apitrace now supports it, and it makes it much easier to test
tracing/replaying on OpenGL ES contexts since
GLX_EXT_create_context_{es2,es}_profile are widely available.

Reviewed-by: Brian Paul <brianp@vmware.com>
9 years agoRevert "clover: Fix build after llvm r221375"
Tom Stellard [Wed, 12 Nov 2014 17:28:34 +0000 (12:28 -0500)]
Revert "clover: Fix build after llvm r221375"

This reverts commit cd93d82ba9ec8cd8e4f54bbee16d7b47c542de71.

llvm r221375 was reverted, so this commit needs to be too.

9 years agogallivm: Fix build with LLVM 3.6 (r221751).
José Fonseca [Wed, 12 Nov 2014 10:32:38 +0000 (10:32 +0000)]
gallivm: Fix build with LLVM 3.6 (r221751).

Tested with LLVM 3.3, 3.4, 3.5, and 3.6.

Trivial.

9 years agoi965/cfg: Remove if_block/else_block.
Matt Turner [Tue, 11 Nov 2014 06:04:41 +0000 (22:04 -0800)]
i965/cfg: Remove if_block/else_block.

I used these in the SEL peephole, but they require extra tracking and
fix ups. The SEL peephole can pretty easily find the blocks it needs
without these.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
9 years agoi965/fs: Don't use if_block/else_block in SEL peephole.
Matt Turner [Tue, 11 Nov 2014 06:02:38 +0000 (22:02 -0800)]
i965/fs: Don't use if_block/else_block in SEL peephole.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
9 years agoilo: clean up gen6_3DSTATE_SF()
Chia-I Wu [Tue, 11 Nov 2014 07:15:21 +0000 (15:15 +0800)]
ilo: clean up gen6_3DSTATE_SF()

Make the helpers fill out valid Gen7 3DSTATE_SF and 3STATE_SBE.  This
prevents the helpers from having to do

  dw[0] = GEN7_SBE_DW1_x; // setting DW1 value to dw[0]!?

and simplifies gen7_3DSTATE_{SF,SBE}().

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: clean up gen7_3DSTATE_STREAMOUT()
Chia-I Wu [Tue, 11 Nov 2014 05:28:32 +0000 (13:28 +0800)]
ilo: clean up gen7_3DSTATE_STREAMOUT()

Render stream and render enable are independent from so enable.  Having a
single return point makes it easier to see that.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: rework gen7_3DSTATE_SO_DECL_LIST()
Chia-I Wu [Tue, 11 Nov 2014 03:53:57 +0000 (11:53 +0800)]
ilo: rework gen7_3DSTATE_SO_DECL_LIST()

Started to make pipe_stream_output_info mandatory, but ended up adding support
for stream id and making a workaround Gen7-specific.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add 3DSTATE_SO_BUFFER variants
Chia-I Wu [Tue, 11 Nov 2014 03:21:01 +0000 (11:21 +0800)]
ilo: add 3DSTATE_SO_BUFFER variants

Add gen7_disable_3DSTATE_SO_BUFFER() to disable SO buffers.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add gen6_3dstate_constant()
Chia-I Wu [Tue, 11 Nov 2014 03:06:12 +0000 (11:06 +0800)]
ilo: add gen6_3dstate_constant()

It replaces gen6_fill_3dstate_constant().  gen6_3DSTATE_CONSTANT_{VS,GS,PS}
are made wrappers of the new function.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_{HS,DS}
Chia-I Wu [Tue, 11 Nov 2014 02:38:28 +0000 (10:38 +0800)]
ilo: add variants of 3DSTATE_{HS,DS}

Rename them to gen7_disable_3DSTATE_{HS,DS}() to reflect the fact.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_GS
Chia-I Wu [Tue, 11 Nov 2014 02:35:24 +0000 (10:35 +0800)]
ilo: add variants of 3DSTATE_GS

Add gen6_so_3DSTATE_GS(), gen6_disable_3DSTATE_GS(), and
gen7_disable_3DSTATE_GS() to do SO on GEN6 or to disable GS.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_VS
Chia-I Wu [Tue, 11 Nov 2014 02:31:55 +0000 (10:31 +0800)]
ilo: add variants of 3DSTATE_VS

Add gen6_disable_3DSTATE_VS() to disable VS.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_PS
Chia-I Wu [Tue, 11 Nov 2014 02:50:31 +0000 (10:50 +0800)]
ilo: add variants of 3DSTATE_PS

Add gen7_disable_3DSTATE_PS() to disable PS.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_WM
Chia-I Wu [Tue, 11 Nov 2014 02:25:20 +0000 (10:25 +0800)]
ilo: add variants of 3DSTATE_WM

Add gen6_hiz_3DSTATE_WM() and gen7_hiz_3DSTATE_WM() for HiZ ops without
dispatching.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: add variants of 3DSTATE_CLIP
Chia-I Wu [Tue, 11 Nov 2014 02:21:03 +0000 (10:21 +0800)]
ilo: add variants of 3DSTATE_CLIP

Add gen6_disable_3DSTATE_CLIP to disable clipping.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: prefix 3DSTATE_VF with gen75
Chia-I Wu [Tue, 11 Nov 2014 01:11:37 +0000 (09:11 +0800)]
ilo: prefix 3DSTATE_VF with gen75

3DSTATE_VF is Gen7.5+ only.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agost/va: MPEG4 call vlVaDecoderFixMPEG4Startcode()
Michael Varga [Thu, 23 Oct 2014 15:45:20 +0000 (10:45 -0500)]
st/va: MPEG4 call vlVaDecoderFixMPEG4Startcode()

If the VOP and GOV headers were truncated they will be regenerated.

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agost/va: MPEG4 generate GOV and VOP header
Michael Varga [Thu, 23 Oct 2014 15:41:24 +0000 (10:41 -0500)]
st/va: MPEG4 generate GOV and VOP header

Also, Implemented a small locally used interface for writing bits to a buffer.

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agost/va: MPEG4 populate the SPS structure
Michael Varga [Thu, 23 Oct 2014 15:36:52 +0000 (10:36 -0500)]
st/va: MPEG4 populate the SPS structure

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agost/va: MPEG4 populate the iq matrix buffers
Michael Varga [Thu, 23 Oct 2014 15:18:09 +0000 (10:18 -0500)]
st/va: MPEG4 populate the iq matrix buffers

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agost/va: MPEG4 populate the PPS structure
Michael Varga [Thu, 23 Oct 2014 15:13:05 +0000 (10:13 -0500)]
st/va: MPEG4 populate the PPS structure

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agost/va: refactored handleVASliceDataBufferType
Michael Varga [Fri, 17 Oct 2014 16:05:59 +0000 (11:05 -0500)]
st/va: refactored handleVASliceDataBufferType

This patch cleans the function handleVASliceDataBufferType() for better
readability.

Signed-off-by: Michael Varga <Michael.Varga@amd.com>
9 years agomesa: Remove _mesa_max_buffer_index
Ian Romanick [Fri, 7 Nov 2014 06:56:03 +0000 (22:56 -0800)]
mesa: Remove _mesa_max_buffer_index

It appears to be completely unused since f9be8543 (February 2012).

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agomesa: Uniform logging is very, very unlikely
Ian Romanick [Tue, 21 Oct 2014 00:02:55 +0000 (17:02 -0700)]
mesa: Uniform logging is very, very unlikely

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agoglsl: Swap the order of glsl_type::name and ::length
Ian Romanick [Fri, 24 Oct 2014 20:08:06 +0000 (13:08 -0700)]
glsl: Swap the order of glsl_type::name and ::length

On x86-64 this saves 8 bytes of padding in the structure, and this
reduces the size of the structure to 32 bytes.

v2: Fix constructor so that GCC won't warn about the order of
initialization.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agoglsl: Store glsl_type::vector_elements and ::matrix_columns as uint8_t
Ian Romanick [Fri, 24 Oct 2014 20:04:24 +0000 (13:04 -0700)]
glsl: Store glsl_type::vector_elements and ::matrix_columns as uint8_t

Due to the total number of bits used in the bitfield, this does not
increase the size of the structure.

It does, however, reduce the number of instructions required each time
one of these fields is accessed.  To access ::matrix_columns with the
bitfield, three instructions were required:

    movzbl 0x9(%rdx),%eax
    shr    %al
    and    $0x7,%eax

As a uint8_t, only one instruction is required.

    movzbl 0xa(%rdx),%eax

These fields are accessed *a lot*.

Valgrind callgrind results for a trace of Tesseract:

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (64-bit):       48,103,497       16,556,096          676,447
After  (64-bit):       45,722,616       15,737,964          670,607

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (32-bit):       61,472,611       21,051,222          821,361
After  (32-bit):       57,987,421       19,872,226          811,609

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Don't check for API_OPENGLES in _mesa_uniform_matrix
Ian Romanick [Mon, 20 Oct 2014 23:50:53 +0000 (16:50 -0700)]
mesa: Don't check for API_OPENGLES in _mesa_uniform_matrix

There are no uniforms in OpenGL ES 1.x, so we can't even get to this
code in that API.

Also, reorder the checks.  First check that transpose is true, then
check whether or not that is legal in the current API.  transpose should
never be true in an ES2 context, so this gets one check (the more
expensive one) out of the main path.

Valgrind callgrind results for a trace of Tesseract:

                 _mesa_UniformMatrix4fv  _mesa_UniformMatrix3fv
Before (64-bit):             96,119,025              24,240,510
After  (64-bit):             90,726,569              22,926,662

                 _mesa_UniformMatrix4fv  _mesa_UniformMatrix3fv
Before (32-bit):            132,434,452              29,051,808
After  (32-bit):            126,658,112              27,989,316

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Rework array error checks in validate_uniform_parameters
Ian Romanick [Mon, 20 Oct 2014 23:40:50 +0000 (16:40 -0700)]
mesa: Rework array error checks in validate_uniform_parameters

Before ARB_explicit_uniform_location, Mesa's location encoding allowed
locations for non-array types that had non-zero array indices.
Basically, part of the location was the uniform and part was the array
index.  This meant that some checks had to occur for arrays and
non-arrays.  This is no longer possible, we the checks can be split up.

Valgrind callgrind results for a trace of Tesseract:

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (64-bit):       50,499,557      17,487,316           686,227
After  (64-bit):       50,023,791      17,274,432           684,293

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (32-bit):       62,968,039       21,732,380          828,147
After  (32-bit):       62,373,967       21,490,756          826,223

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Get some gl_shader_program::LinkStatus checking out of the main path
Ian Romanick [Mon, 20 Oct 2014 21:47:45 +0000 (14:47 -0700)]
mesa: Get some gl_shader_program::LinkStatus checking out of the main path

I really wanted to remove 'shProg != NULL' as well, but that would have
required adding a dummy program as the default program.  That seemed
like more churn than removing one test was worth.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Rework location == -1 error checking
Ian Romanick [Mon, 20 Oct 2014 23:21:54 +0000 (16:21 -0700)]
mesa: Rework location == -1 error checking

Only one caller wanted to generate an error when location == -1, so move
the error generation to that caller.  There will be more callers in the
future that do not want to generate errors.

Move the location == -1 check later in validate_uniform_parameters.  As
currently implemented, glUniform1iv(-1, -1, data) would not generate an
error, but it should due to count being < 0.

The location that I have moved it to will make more sense with the next
commit.

Valgrind callgrind results for a trace of Tesseract:

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (64-bit):       51,241,217      17,740,162           689,181
After  (64-bit):       50,499,557      17,487,316           686,227

                 _mesa_Uniform4fv  _mesa_Uniform4f  _mesa_Uniform1i
Before (32-bit):       63,940,605       21,987,918          831,065
After  (32-bit):       62,968,039       21,732,380          828,147

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Minor clean ups in _mesa_uniform
Ian Romanick [Fri, 10 Oct 2014 14:35:31 +0000 (16:35 +0200)]
mesa: Minor clean ups in _mesa_uniform

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa: Remove GLSL_TYPE_SAMPLER check
Ian Romanick [Fri, 10 Oct 2014 13:29:31 +0000 (15:29 +0200)]
mesa: Remove GLSL_TYPE_SAMPLER check

Noting the assertion just a few lines earlier, returnType cannot be
GLSL_TYPE_SAMPLER.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agomesa/main: Pass the data that _mesa_uniform actually wants
Ian Romanick [Thu, 9 Oct 2014 09:00:32 +0000 (11:00 +0200)]
mesa/main: Pass the data that _mesa_uniform actually wants

The GL_ enums were previously used because glsl_types.h couldn't be used
in C code.  That was fixed some time ago (and uniforms.c already
includes glsl_types.h), so this is no longer necessary.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
9 years agoilo: derive fb blending caps at bind time
Chia-I Wu [Mon, 10 Nov 2014 05:44:45 +0000 (13:44 +0800)]
ilo: derive fb blending caps at bind time

Derive whether a RT supports blending, logicop, and the like when
set_framebuffer_state() is called.  This enables us to simplify
gen6_BLEND_STATE().

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: remove inlined state functions
Chia-I Wu [Mon, 10 Nov 2014 05:11:40 +0000 (13:11 +0800)]
ilo: remove inlined state functions

We had some inlined state functions for dispatching.  They were not needed
with the new top/bottom split.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: use top/bottom split for state functions
Chia-I Wu [Mon, 10 Nov 2014 04:55:34 +0000 (12:55 +0800)]
ilo: use top/bottom split for state functions

Follow the builder and split state functions into top (vertex processing) and
bottom (pixel processing).

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoi965: Advertise a line width of 40.0 on Cherryview and Skylake.
Kenneth Graunke [Tue, 4 Nov 2014 01:56:38 +0000 (17:56 -0800)]
i965: Advertise a line width of 40.0 on Cherryview and Skylake.

According to the documentation, line widths higher than 40.0 may have
quality problems.  That's already 20 times larger than we've been
exposing, so it seems totally sufficient.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
9 years agoi965: Advertise larger line widths.
Kenneth Graunke [Tue, 4 Nov 2014 01:01:53 +0000 (17:01 -0800)]
i965: Advertise larger line widths.

We've artificially been limiting this to 5 for no particular reason.

On Gen4-5, the limit is [0, 7.5] with a granularity of 0.5 (U3.1).
On Gen6+, the limit is [0, 7.9921875].  Since it's a U3.7, the
granularity should be 0.125 (1/8).

This patch conservatively advertises one granularity smaller than the
hardware's maximum value, just in case there's a problem using the
largest possible value.  On Gen4-5, this is 7.5 - 0.5 = 7.0.  On Gen6+,
this is 8.0 - 0.125 = 7.875.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
9 years agoi965: Use ctx->Const.MaxLineWidth when clamping ctx->Line.Width.
Kenneth Graunke [Tue, 4 Nov 2014 00:26:48 +0000 (16:26 -0800)]
i965: Use ctx->Const.MaxLineWidth when clamping ctx->Line.Width.

Rather than hardcoding platform values in every code path, just use the
maximum value we set.

Currently, ctx->Const.LineWidth == 5, which is smaller than the hardware
limit.  But applications shouldn't be using a value larger than we
support anyway.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
9 years agoi965: Set Line Width correctly on Cherryview and Skylake.
Kenneth Graunke [Tue, 4 Nov 2014 00:10:55 +0000 (16:10 -0800)]
i965: Set Line Width correctly on Cherryview and Skylake.

Line Width moved to DW1 bits 29:12.  It's actually now a U11.7.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
9 years agodocs: add news item and link release notes for mesa 10.3.3
Emil Velikov [Sat, 8 Nov 2014 17:15:42 +0000 (17:15 +0000)]
docs: add news item and link release notes for mesa 10.3.3

Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
9 years agodocs: Add sha256 sums for the 10.3.3 release
Emil Velikov [Sat, 8 Nov 2014 17:07:34 +0000 (17:07 +0000)]
docs: Add sha256 sums for the 10.3.3 release

Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
(cherry picked from commit 9cc26056ee13f25c5785fef81b31487f1429baa4)

9 years agoAdd release notes for the 10.3.3 release
Emil Velikov [Sat, 8 Nov 2014 16:43:13 +0000 (16:43 +0000)]
Add release notes for the 10.3.3 release

Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
(cherry picked from commit 1a9cc5f50db5d27530a3449743b43aac389d781f)

9 years agoutil/format: Fix clamping to 32bit integers.
José Fonseca [Fri, 7 Nov 2014 21:20:25 +0000 (21:20 +0000)]
util/format: Fix clamping to 32bit integers.

Use clamping constants that guarantee no integer overflows.

As spotted by Chris Forbes.

This causes the code to change as:

-         value |= (uint32_t)CLAMP(src[0], 0.0f, 4294967295.0f);
+         value |= (uint32_t)CLAMP(src[0], 0.0f, 4294967040.0f);

-         value |= (uint32_t)((int32_t)CLAMP(src[0], -2147483648.0f, 2147483647.0f));
+         value |= (uint32_t)((int32_t)CLAMP(src[0], -2147483648.0f, 2147483520.0f));

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
9 years agoutil/format: Generate floating point constants for clamping.
José Fonseca [Fri, 7 Nov 2014 14:26:58 +0000 (14:26 +0000)]
util/format: Generate floating point constants for clamping.

This commit causes the generated C code to change as

            union util_format_r32g32b32a32_sscaled pixel;
  -         pixel.chan.r = (int32_t)CLAMP(src[0], -21474836482147483647);
  -         pixel.chan.g = (int32_t)CLAMP(src[1], -21474836482147483647);
  -         pixel.chan.b = (int32_t)CLAMP(src[2], -21474836482147483647);
  -         pixel.chan.a = (int32_t)CLAMP(src[3], -21474836482147483647);
  +         pixel.chan.r = (int32_t)CLAMP(src[0], -2147483648.0f, 2147483647.0f);
  +         pixel.chan.g = (int32_t)CLAMP(src[1], -2147483648.0f, 2147483647.0f);
  +         pixel.chan.b = (int32_t)CLAMP(src[2], -2147483648.0f, 2147483647.0f);
  +         pixel.chan.a = (int32_t)CLAMP(src[3], -2147483648.0f, 2147483647.0f);
            memcpy(dst, &pixel, sizeof pixel);

which surprisingly makes a difference for MSVC.

Thanks to Juraj Svec for diagnosing this and drafting a fix.

Fixes https://bugs.freedesktop.org/show_bug.cgi?id=29661

9 years agoglsl/list: Revert unintentional file mode change in previous commit.
Vinson Lee [Sat, 8 Nov 2014 05:04:08 +0000 (21:04 -0800)]
glsl/list: Revert unintentional file mode change in previous commit.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
9 years agoglsl/list: Move declaration before code.
Vinson Lee [Fri, 7 Nov 2014 23:33:41 +0000 (15:33 -0800)]
glsl/list: Move declaration before code.

Fixes MSVC build error.

shaderapi.c
src\glsl\list.h(535) : error C2143: syntax error : missing ';' before 'type'
src\glsl\list.h(535) : error C2143: syntax error : missing ')' before 'type'
src\glsl\list.h(536) : error C2065: 'node' : undeclared identifier

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86025
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
9 years agoglsl/list: Add an exec_list_validate function
Jason Ekstrand [Wed, 5 Nov 2014 21:57:09 +0000 (13:57 -0800)]
glsl/list: Add an exec_list_validate function

This can be very useful for trying to debug list corruptions.

Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
9 years agollvmpipe: Avoid deadlock when unloading opengl32.dll
José Fonseca [Fri, 7 Nov 2014 16:15:43 +0000 (16:15 +0000)]
llvmpipe: Avoid deadlock when unloading opengl32.dll

On Windows, DllMain calls and thread creation/destruction are
serialized, so when llvmpipe is destroyed from DllMain waiting for the
rasterizer threads to finish will deadlock.

So, instead of waiting for rasterizer threads to have finished, simply wait for the
rasterizer threads to notify they are just about to finish.

Verified with this very simple program:

   #include <windows.h>
   int main() {
      HMODULE hModule = LoadLibraryA("opengl32.dll");
      FreeLibrary(hModule);
   }

Fixes https://bugs.freedesktop.org/show_bug.cgi?id=76252

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Cc: 10.2 10.3 <mesa-stable@lists.freedesktop.org>
9 years agodocs: Update minimum required LLVM version.
José Fonseca [Fri, 7 Nov 2014 14:39:00 +0000 (14:39 +0000)]
docs: Update minimum required LLVM version.

9 years agoi965: drop the custom gen8_instruction CFLAG
Emil Velikov [Fri, 7 Nov 2014 00:55:46 +0000 (00:55 +0000)]
i965: drop the custom gen8_instruction CFLAG

No longer needed as the file was removed with
commit 8c229d306b3f312adbdfbaf79967ee43fbfc839e
Author: Kenneth Graunke <kenneth@whitecape.org>
Date:   Mon Aug 11 10:07:07 2014 -0700

    i965: Delete the Gen8 code generators.

    We now use the brw_eu_emit.c code instead.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
9 years agogbm/dri: cleanup memory leak on teardown
Emil Velikov [Fri, 7 Nov 2014 03:45:07 +0000 (03:45 +0000)]
gbm/dri: cleanup memory leak on teardown

During teardown we free the driver_configs list pointer, but we forget
to deallocate each config in that list.

Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-and-tested-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agoegl_dri2: add a note about dri2_create_screen
Emil Velikov [Fri, 7 Nov 2014 03:42:15 +0000 (03:42 +0000)]
egl_dri2: add a note about dri2_create_screen

The function is not called by platform_drm. As such one needs to
pay special attention at teardown.

v2: Fix the comment block. Spotted by Ken.

Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-and-tested-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
9 years agoegl_dri2: fix double free on drm platforms
Emil Velikov [Fri, 7 Nov 2014 03:33:56 +0000 (03:33 +0000)]
egl_dri2: fix double free on drm platforms

Earlier commit failed to attribure that for drm platforms one does not
call dri2_create_screen, thus it does not create the screen and
driver_configs but inherits them from the "display" - gbm.

As such wrap cleanup in Platform != _EGL_PLATFORM_DRM to prevent
the issue and still cleanup correctly for non-drm platforms.

v2:
 - Drop the ifdef HAVE_DRM_PLATFORM, reindent the code and fix the
comment block. Suggested by Ken.

Reported-by: Kenneth Graunke <kenneth@whitecape.org>
Reported-by: Mark Janes <mark.a.janes@intel.com>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-and-tested-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
9 years agoilo: tidy up message descriptor decoding
Chia-I Wu [Fri, 7 Nov 2014 08:10:38 +0000 (16:10 +0800)]
ilo: tidy up message descriptor decoding

Move opcode to string mappings to functions of their own.  Have for consistent
outputs for similar opcodes.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoilo: decode INTERFACE_DESCRIPTOR_DATA
Chia-I Wu [Fri, 7 Nov 2014 07:45:25 +0000 (15:45 +0800)]
ilo: decode INTERFACE_DESCRIPTOR_DATA

This is at least much better than decoding as blobs.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
9 years agoi965/fs: Wire up control flow correctly in predicated break pass.
Matt Turner [Wed, 5 Nov 2014 23:23:05 +0000 (15:23 -0800)]
i965/fs: Wire up control flow correctly in predicated break pass.

When the earlier block ended with control flow, we'd mistakenly remove
some of its links to its children. The same happened with the later
block.

Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
9 years agoi965/cfg: Add functions to get first and last non-CF instructions.
Matt Turner [Mon, 27 Oct 2014 21:36:48 +0000 (14:36 -0700)]
i965/cfg: Add functions to get first and last non-CF instructions.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
9 years agoglsl: Skip loop-too-large heuristic if indexing arrays of a certain size
Kenneth Graunke [Thu, 30 Oct 2014 03:56:07 +0000 (20:56 -0700)]
glsl: Skip loop-too-large heuristic if indexing arrays of a certain size

A pattern in certain shaders is:

   uniform vec4 colors[NUM_LIGHTS];

   for (int i = 0; i < NUM_LIGHTS; i++) {
      ...use colors[i]...
   }

In this case, the application author expects the shader compiler to
unroll the loop.  By doing so, it replaces variable indexing of the
array with constant indexing, which is more efficient.

This patch extends the heuristic to see if arrays accessed within the
loop are indexed by an induction variable, and if the array size exactly
matches the number of loop iterations.  If so, the application author
probably intended us to unroll it.  If not, we rely on the existing
loop-too-large heuristic.

Improves performance in a phong shading microbenchmark by 2.88x, and a
shadow mapping microbenchmark by 1.63x.  Without variable indexing, we
can upload the small uniform arrays as push constants instead of pull
constants, avoiding shader memory access.  Affects several games, but
doesn't appear to impact their performance.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Kristian Høgsberg <krh@bitplanet.net>
9 years agoglsl: Lower constant arrays to uniform arrays.
Kenneth Graunke [Sat, 26 Apr 2014 07:18:54 +0000 (00:18 -0700)]
glsl: Lower constant arrays to uniform arrays.

Consider GLSL code such as:

   const ivec2 offsets[] =
      ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
              ivec2(0, -1),  ivec2(0, 0),  ivec2(0, 1),
              ivec2(1, -1),  ivec2(1, 0),  ivec2(1, 1));

   ivec2 offset = offsets[<non-constant expression>];

Both i965 and nv50 currently handle this very poorly.  On i965, this
becomes a pile of MOVs to load the immediate constants into registers,
a pile of scratch writes to move the whole array to memory, and one
scratch read to actually access the value - effectively the same as if
it were a non-constant array.

We'd much rather upload large blocks of constant data as uniform data,
so drivers can simply upload the data via constbufs, and not have to
populate it via shader instructions.

This is currently non-optional because both i965 and nouveau benefit
from it, and according to Marek radeonsi would benefit today as well.
(According to Tom, radeonsi may want to handle this itself in the long
term, but we can always add a flag when it becomes useful.)

Improves performance in a terrain rendering microbenchmark by about 2x,
and cuts the number of instructions in about half.  Helps a lot of
"Natural Selection 2" shaders, as well as one "HOARD" shader.

total instructions in shared programs: 5473459 -> 5471765 (-0.03%)
instructions in affected programs:     5880 -> 4186 (-28.81%)

v2: Use ir_var_hidden to avoid exposing the new uniform via the GL
    uniform introspection API.

v3: Alphabetize Makefile.sources properly.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77957
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
9 years agoglsl: Add infrastructure for "hidden" uniforms.
Kenneth Graunke [Mon, 27 Oct 2014 23:34:06 +0000 (16:34 -0700)]
glsl: Add infrastructure for "hidden" uniforms.

In the compiler, we'd like to generate implicit uniforms for internal
use.  These should not be visible via the GL uniform introspection API.

To support that, we add a new ir_variable::how_declared value of
ir_var_hidden, and plumb that through to gl_uniform_storage.

v2 (idr): Fix some memory management issues in
move_hidden_uniforms_to_end.  The comment block on the function has more
details.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
9 years agomesa: Add SSE 4.1 optimisation for glDrawElements.
Timothy Arceri [Wed, 29 Oct 2014 12:05:59 +0000 (23:05 +1100)]
mesa: Add SSE 4.1 optimisation for glDrawElements.

Makes use of SSE 4.1 to speed up compute of min and max elements.

Callgrind cpu usage results from pts benchmarks:

Openarena 0.8.8: 3.67% -> 1.03%
UrbanTerror: 2.36% -> 0.81%

V5:
- actually make use of the optimisation in android (Emil Velikov)
- set a better array size limit for using SSE and added TODO

V4:
- fixed bugs with incrementing pointer and updating counters

V3:
- Removed sse_minmax.c from Makefile.sources
- handle the first few values without SSE until the pointer is aligned
 and use _mm_load_si128 rather than _mm_loadu_si128
- guard the call to the SSE code better at build time

V2:
- removed GL* types
- use _mm_store_si128() rather than _mm_store_ps()
- add runtime check for SSE
- use aligned attribute for local mix/max
- bunch of tidyups

Reviewed-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Timothy Arceri <t_arceri@yahoo.com.au>
9 years agoi965: Remove non-existent vertical strides from array.
Matt Turner [Tue, 28 Oct 2014 16:04:41 +0000 (09:04 -0700)]
i965: Remove non-existent vertical strides from array.

These never existed, as far as I can tell.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>