radeonsi: Rename R600->SI in some remaining defines.

I had previously considered that unsafe.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename radeonsi->si remaining identifiers in si_uvd.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifiers in si_state_draw.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifiers in si_resource.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifiers in si_query.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifiers in si_pipe.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifier in si_hw_context.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename radeonsi->si remaining identifiers in si_compute.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si remaining identifiers in si_blit.c.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for functions in si_pipe.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for functions in si.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for functions in si_resource.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for structs in si_resource.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for structs in si.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Rename r600->si for structs in si_pipe.h.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: Apply si_* file naming scheme.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

Use AC_PATH_TOOL instead of AC_PATH_PROG for llvm-config.

This should help with cross-compiling and multilib when $CHOST-specific
llvm-config is expected rather than build host default one.

It will help us a bit in Gentoo where we've started using
i686-pc-linux-gnu-llvm-config for 32-bit multilib LLVM.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Signed-off-by: Michał Górny <mgorny@gentoo.org>
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=73100
CC: "10.0" <mesa-stable@lists.freedesktop.org>

configure: Disable xvmc by default

The xvmc unit tests are failing on r300g and r600g.

Reviewed-by: Vinson Lee <vlee@freedesktop.org>

glsl: Remove exec_list iterators now that nothing uses them.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Replace iterators in ir_reader.cpp with ad-hoc list walking.

These can't use foreach_list since they want to skip over the first few
list elements.  Just doing the ad-hoc list walking isn't too bad.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Use a new foreach_two_lists macro for walking two lists at once.

When handling function calls, we often want to walk through the list of
formal parameters and list of actual parameters at the same time.
(Both are guaranteed to be the same length.)

Previously, we used a pattern of:

   exec_list_iterator 1st_iter = <1st list>.iterator();
   foreach_iter(exec_list_iterator, 2nd_iter, <2nd list>) {
      ...
      1st_iter.next();
   }

This was awkward, since you had to manually iterate through one of
the two lists.

This patch introduces a foreach_two_lists macro which safely walks
through two lists at the same time, so you can simply do:

   foreach_two_lists(1st_node, <1st list>, 2nd_node, <2nd list>) {
      ...
   }

v2: Rename macro from foreach_list2 to foreach_two_lists, as suggested
    by Ian Romanick.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Statically cast parameter exec_node to ir_variable.

Formal function parameters are always ir_variable objects, not an
arbitrary ir_instruction.  So there's no need to dynamically cast here.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Cast ir_call parameters to ir_rvalue, not ir_instruction.

A function call's parameters are always rvalues.  ir_rvalue may not
always be a subclass of ir_instruction in the future, so we should use
the right one.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Replace foreach_iter and iter.remove() with foreach_list_safe.

foreach_list_safe allows you to safely remove the current node.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Convert piles of foreach_iter to foreach_list_safe.

In these cases, we edit the list (or at least might be), so we use the
foreach_list_safe variant.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

glsl: Convert piles of foreach_iter to the newer foreach_list macro.

foreach_iter and exec_list_iterators have been deprecated for some time now;
we just hadn't ever bothered to convert code to the newer foreach_list
and foreach_list_safe macros.

In these cases, we aren't editing the list, so we can use foreach_list
rather than foreach_list_safe.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Ensure that all necessary state is re-emitted if we run out of aperture.

Prior to this patch, if we ran out of aperture space during
brw_try_draw_prims(), we would rewind the batch buffer pointer
(potentially throwing some state that may have been emitted by
brw_upload_state()), flush the batch, and then try again.  However, we
wouldn't reset the dirty bits to the state they had before the call to
brw_upload_state().  As a result, when we tried again, there was a
danger that we wouldn't re-emit all the necessary state.  (Note: prior
to the introduction of hardware contexts, this wasn't a problem
because flushing the batch forced all state to be re-emitted).

This patch fixes the problem by leaving the dirty bits set at the end
of brw_upload_state(); we only clear them after we have determined
that we don't need to rewind the batch buffer.

Cc: 10.0 9.2 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

r600g: fix glClearBuffer by handling PIPE_CLEAR_COLORi flags correctly

also restructure the code

r600g: handle NULL colorbuffers correctly on R600-R700

r600g: handle NULL colorbuffers correctly on Evergreen

radeonsi: handle NULL colorbuffers correctly

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

gallium/util: easy fixes for NULL colorbuffers

Reviewed-by: Brian Paul <brianp@vmware.com>

st/mesa: bind NULL colorbuffers as specified by glDrawBuffers

An example why it is required:

    Let's say there's a fragment shader writing to gl_FragData[0..1].
    The user calls: glDrawBuffers(2, {GL_NONE, GL_COLOR_ATTACHMENT0});

    That means gl_FragData[0] is unused and gl_FragData[1] is written
    to GL_COLOR_ATTACHMENT0.

st/mesa was skipping the GL_NONE draw buffer, therefore gl_FragData[0]
was written to GL_COLOR_ATTACHMENT0, which was wrong.

This commit fixes it, but drivers must also be fixed not to crash when
binding NULL colorbuffers. There is also a new set of piglit tests for this.

The MSAA state also had to be fixed not to crash when reading fb->cbufs[0].

Reviewed-by: Brian Paul <brianp@vmware.com>

mesa: handle GL_NONE draw buffers correctly in glClear

Reviewed-by: Brian Paul <brianp@vmware.com>

st/mesa: use sRGB formats for MSAA resolving if destination is sRGB

Copied from the i965 driver, including the big comment.

Cc: 9.2 10.0 <mesa-stable@lists.freedesktop.org>

st/mesa: check depth and stencil writemask before clearing

st/mesa: always prefer pipe->clear over clear_with_quad (v2)

v2: clear depth and stencil together

st/egl: Flush resources before presentation

Fixes wayland regression on r600g due to fast clear introduced by commit
edbbfac6.

Signed-off-by: Marek Olšák <marek.olsak@amd.com>

dri: set yInverted default to GL_TRUE

yInverted is used by EGL_NOK_texture_from_pixmap to indicate that
window system rendering is y-inverted compared to OpenGL texture
representation. This extension is only known to be used with X11
window system where sane default is GL_TRUE.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=73371

Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>

egl_dri2: call dri2_add_configs_for_visuals after extensions set

dri2_add_config makes decisions based on NOK_texture_from_pixmap so
it needs to be enabled before calling dri2_add_configs_for_visuals.

Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>

mesa: Set the correct error in _mesa_BeginConditionalRender

Piglit was recently changed to expect the correct error code (piglit
commit 271b998), so it started failing on Mesa.  This corrects that
failing and adds some spec quotations to justify the errrors set.

The code was rearranged a little bit to match the order listed in the
spec.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

i965: Delete duplicate write_timestamp function.

brw_queryobj.c needs a version of write_timestamp that works on all
generations for the QueryCounter() driver hook.  So there's no point in
duplicating it in gen6_queryobj.c.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Eric Anholt <eric@anholt.net>

i965: Fix clears of layered framebuffers with mismatched layer counts.

Previously, Mesa enforced the following rule (from
ARB_geometry_shader4's list of criteria for framebuffer completeness):

  * If any framebuffer attachment is layered, all attachments must have
    the same layer count.  For three-dimensional textures, the layer count
    is the depth of the attached volume.  For cube map textures, the layer
    count is always six.  For one- and two-dimensional array textures, the
    layer count is simply the number of layers in the array texture.
    { FRAMEBUFFER_INCOMPLETE_LAYER_COUNT_ARB }

However, when ARB_geometry_shader4 was adopted into GL 3.2, this rule
was dropped; GL 3.2 permits different attachments to have different
layer counts.  This patch brings Mesa in line with GL 3.2.

In order to ensure that layered clears properly clear all layers, we
now have to keep track of the maximum number of layers in a layered
framebuffer.

Fixes the following piglit tests in spec/!OpenGL 3.2/layered-rendering:
- clear-color-all-types 1d_array mipmapped
- clear-color-all-types 1d_array single_level
- clear-color-mismatched-layer-count
- framebuffer-layer-count-mismatch

Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>

main: check texture target when validating layered framebuffers.

From section 4.4.4 (Framebuffer Completeness) of the GL 3.2 spec:

    If any framebuffer attachment is layered, all populated
    attachments must be layered. Additionally, all populated color
    attachments must be from textures of the same target.

We weren't checking that the attachments were from textures of the
same target.

Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>

i965/gen6/blorp: Remove redundant HiZ workaround

Commit 1a92881 added extra flushes to fix a HiZ hang in
WebGL Google Maps. With the extra flushes emitted by the previous two
patches, the flushes added by 1a92881 are redundant.

Tested with the same criteria as in 1a92881: by zooming in and out
continuously for 2 hours on Sandybridge Chrome OS (codename
Stumpy) without a hang.

CC: Kenneth Graunke <kenneth@whitecape.org>
CC: Stéphane Marchesin <marcheu@chromium.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/gen6/blorp: Set need_workaround_flush at top of blorp

Unconditionally set brw->need_workaround_flush at the top of gen6 blorp
state emission.

The art of emitting workaround flushes on Sandybridge is mysterious and
not fully understood. Ken and I believe that
intel_emit_post_sync_nonzero_flush() may be required when switching from
regular drawing to blorp.  This is an extra safety measure to prevent
undiscovered difficult-to-diagnose gpu hangs.

I verified that on ChromeOS, pre-patch, need_workaround_flush was not
set at the top of blorp, as Paul expected. To verify, I inserted the
following debug code at the top of gen6_blorp_exec(), restarted the ui,
and inspected the logs in /var/log/ui. The abort gets triggered so early
that the browser never appears on the display.

    static void
    gen6_blorp_exec(...)
    {
        if (!brw->need_workaround_flush) {
            fprintf(stderr, "chadv: %s:%d\n", __FILE__, __LINE__);
            abort();
        }
        ...
    }

CC: Kenneth Graunke <kenneth@whitecape.org>
CC: Stéphane Marchesin <marcheu@chromium.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

i965/gen6/blorp: Set need_workaround_flush immediately after primitive

This patch makes the workaround code in gen6 blorp follow the pattern
established in the regular draw path. It shouldn't result in any
behavioral change.

On gen6, there are two places where we emit 3D_CMD_PRIM: brw_emit_prim()
and gen6_blorp_emit_primitive().  brw_emit_prim() sets
need_workaround_flush immediately after emitting the primitive, but
blorp does not. Blorp sets need_workaround_flush at the bottom of
brw_blorp_exec().

This patch moves the need_workaround_flush from brw_blorp_exec() to
gen6_blorp_emit_primitive().  There is no need to set
need_workaround_flush in gen7_blorp_emit_primitive() because the
workaround applies only to gen6.

Reviewed-by: Paul Berry <stereotype441@gmail.com>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>

docs: Import 10.0.2 release notes, add news item.

mesa: add missing SNORM formats in _mesa_base_fbo_format()

We weren't handling the LUMINANCE_SNORM, LUMINANCE_ALPHA_SNORM and
INTENSITY_SNORM cases.  Note that adding these cases here does not
require a driver to support rendering to these surface types.  If
the driver can't do it we'll report an incomplete framebuffer.

NVIDIA doesn't support GL_EXT_texture_snorm but their driver
accepts these formats in glRenderBufferStorage().

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

mesa: remove dead geom shader code

I doubt the swrast-based drivers will ever support GS.

Reviewed-by: Matt Turner <mattst88@gmail.com>

docs: minor updates to VMware SVGA3D driver page

Signed-off-by: Brian Paul <brianp@vmware.com>

mesa: check bits per channel for GL_RGBA_SIGNED_COMPONENTS_EXT query

If a channel has zero bits it's not signed.

v2: also check for luminance and intensity format bits.  Bruce
Merry's proposed piglit test hits the luminance case.

Bugzilla: http://bugs.freedesktop.org/show_bug.cgi?id=73096
Cc: 10.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>

mesa: check for MESA_FORMAT_RGB9_E5_FLOAT in _mesa_is_format_signed()

This packed floating point format only stores positive values.

Bugzilla: http://bugs.freedesktop.org/show_bug.cgi?id=73096
Cc: 10.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

st/mesa: fix breakage from gl_constant::Program[] change

mesa: Use functions to convert gl_shader_stage to PROGRAM enum or pipe target.

Suggested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Improve assert message.

main: Change init_program_limits() to use gl_shader_stage.

This allows the caller to execute it in a loop rather than
hand-rolling a separate call for each stage.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: Index into ctx->Const.Program[] rather than using ad-hoc code.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Index into ctx->Const.Program[] rather than using ad-hoc code.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: replace ctx->Const.{Vertex,Fragment,Geomtery}Program with an array.

These are replaced with
ctx->Const.Program[MESA_SHADER_{VERTEX,FRAGMENT,GEOMETRY}].  In
patches to follow, this will allow us to replace a lot of ad-hoc logic
with a variable index into the array.

With the exception of the changes to mtypes.h, this patch was
generated entirely by the command:

    find src -type f '(' -iname '*.c' -o -iname '*.cpp' -o -iname '*.py' \
    -o -iname '*.y' ')' -print0 | xargs -0 sed -i \
    -e 's/Const\.VertexProgram/Const.Program[MESA_SHADER_VERTEX]/g' \
    -e 's/Const\.GeometryProgram/Const.Program[MESA_SHADER_GEOMETRY]/g' \
    -e 's/Const\.FragmentProgram/Const.Program[MESA_SHADER_FRAGMENT]/g'

Suggested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

llvmpipe: Honour pipe_rasterizer::point_quad_rasterization.

Commit eda21d2a3010d9fc5a68b55a843c5e44b2abf8dd fixed the rasterization
of points for Direct3D but ended up breaking the rasterization of OpenGL
non-sprite points, in particular conform's pntrast.c test.

The only way to get both working is to properly honour
pipe_rasterizer::point_quad_rasterization, and follow the weird OpenGL
rule when it is false.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

i965: Don't do the temporary-and-blit-copy for INVALIDATE_RANGE maps.

We definitely want to fall through to the unsynchronized map case, instead
of wasting bandwidth on a copy.  Prevents a -43.2407% +/- 1.06113% (n=49)
performance regression on aa10perf when teaching glamor to provide the
GL_INVALIDATE_RANGE_BIT information.

This is a performance fix, which I usually wouldn't cherry-pick to stable.
But this was really was just a bug in the code, its presence would
discourage developers from giving us the best information they can, and I
think we've got fairly high confidence in the unsynchronized map path
already.

Cc: 10.0 9.2 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Fix handling of MESA_pack_invert in blit (PBO) readpixels.

Fixes piglit GL_MESA_pack_invert/readpixels and GPU hangs with glamor and
cairo-gl.

Cc: 10.0 9.2 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

i965: Fix incorrect bounds tracking for blit readpixels's GPU access.

While incorrect, it probably wouldn't affect anyone ever: You'd have to do
an appropriately-formatted readpixels into a PBO, then overwrite the tail
end of the updated area of the PBO with glBufferSubData(), and you
wouldn't get appropriate synchronization.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

i965: Use SET_FIELD to safety check our x/y offsets in blits.

The earlier assert made sure that our math didn't exceed our bounds, but
this makes sure that we don't overflow from the high bits X into the low
bits of Y.  We've already put checks in intel_miptree_blit(), but I've
wanted to expand the type in our protoype from short to uint32_t, and we
could get in trouble with intel_emit_linear_blit() if we did.

v2: Add Ken's comment about the funny language extension used.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com> (v1)

i965: Add an assert for when SET_FIELD's value exceeds the field size.

This was one of the things we always wanted to do to this, to make it more
useful than just (value << FIELD_MASK).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

i965: Add a safety check for emitting blits.

With all of the flipping and pitch twiddling and miptree layout involved
in our blits, there are lots of ways for us to scribble outside of a
buffer.  Put in a check that we're not about to do so.

This catches a bug that glamor was running into.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

i965: Don't call the blitter on addresses it can't handle.

Noticed by tex3d-maxsize on my next commit to check that our addresses
don't overflow.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

mesa: Namespace qualify fma to override ambiguity with fma from math.h

MSVC 2013 version of math.h includes an fma() function.

Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa: Work around internal compiler error

This small rearrangement avoids MSVC 2013 ICE. Also, this should be
a better memory access order.

Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Fix compile error with MSVC 2013

This fixes the following compile error:
src\glsl\ir_constant_expression.cpp(1405) : error C2666: 'copysign' : 3
overloads have similar conversions

Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa: Preliminary support for MSVC_VERSION=12.0

Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

freedreno: add basic query support

Add for now some simple/basic query support (ie. things not actually
requiring the GPU).  Might change around a bit when I actually add
GPU queries, but for now this enables some useful performance info
in the GALLIUM_HUD.  For example:

  GALLIUM_HUD=fps+batches+batches-sysmem+batches-gmem+restores,draw-calls

The driver specific specific queries are:

  + draw-calls
  + batches - number of batches per second, sum of batches-sysmem
    plus batches-gmem
  + batches-gmem - render a set of tiles in GMEM, for each tile
    (optionally) system mem -> gmem (restore), plus N draws,
    plus gmem -> system mem (resolve) per second
  + batches-sysmem - N draws to system memory (GMEM bypass) per
    second
  + restores - number of GMEM batches that required restore per
    second

Ideally for GMEM rendering, you want batches-gmem to equal fps.  If
the app is doing something that triggers multiple passes (ie. requires
extra round trip gmem <-> system memory) then the # of batches per
second will go up relative to fps.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/a3xx: use cs patch instead of RFI+RMW

Since we now have the cmdstream patch mechanism needed for hw binning,
might as well also use it for RB_RENDER_CONTROL updates.  This avoids
the need to use RMW (and associated WFI) to update RB_RENDER_CONTROL.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno/a3xx: support for hw binning pass

The binning pass sorts vertices into which bins/tiles they apply to.
The visibility information generated during the binning pass can be
used to speed up the rendering pass by filtering out vertices which
do not apply to the current tile.  See:

 https://github.com/freedreno/freedreno/wiki/Adreno-tiling#optimized-approach

This brings a significant fps boost.  A rough assortment of tests
(supertuxkart, etracer, tremulous, glmark2 'build' test, etc) seems
to yield a ~35-45% fps improvement.

For now, to be conservative, the binning pass is not enabled yet by
default.  To enable it use:

  FD_MESA_DEBUG=binning

So far I haven't found anything that breaks with binning enabled,
but I'd like a bit more testing before I enable it as default.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno: be more clever about gmem usage

Only need to leave room for depth/stencil if it is actually used, etc.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno: resync generated headers

Signed-off-by: Rob Clark <robclark@freedesktop.org>

i965: fold offset into coord for textureOffset(gsampler2DRect)

The hardware is broken with nonzero texel offsets and unnormalized
coordinates; instead of doing correct offsetting, we get garbage.

This just extends the existing workaround for ir_txf and
ir_tg4+gsampler2DRect to also consider ir_tex+gsampler2DRect.

Fixes broken rendering in 'tesseract' when 'mesa_texrectoffset_bug' is
not enabled; also fixes the new piglit test
'tests/spec/glsl-1.30/execution/fs-textureOffset-Rect'.

Has been broken ~forever; suggesting including this in only 10.0 because
the lowering pass doesn't exist in 9.2 or earlier so would require quite
a different patch.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Cc: Lee Salzman <lsalzman@gmail.com>
Cc: "10.0" <mesa-stable@lists.freedesktop.org>

mesa: Remove _mesa_progshader_enum_to_string(), which is no longer used.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

glsl: Make more use of gl_shader_stage enum in ir_set_program_inouts.cpp.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

glsl: Make more use of gl_shader_stage enum in lower_clip_distance.cpp.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

glsl: Make more use of gl_shader_stage enum in link_varyings.cpp.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Also rename "shaderType" param of is_varying_var() to "stage".

Reviewed-by: Brian Paul <brianp@vmware.com>

glsl: Change _mesa_glsl_parse_state ctor to use gl_shader_stage enum.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Also rename "target" param to "stage".

Reviewed-by: Brian Paul <brianp@vmware.com>

mesa: Use gl_shader::Stage instead of gl_shader::Type where possible.

This reduces confusion since gl_shader::Type is sometimes
GL_SHADER_PROGRAM_MESA but is more frequently
GL_SHADER_{VERTEX,GEOMETRY,FRAGMENT}.  It also has the advantage that
when switching on gl_shader::Stage, the compiler will alert if one of
the possible enum types is unhandled.  Finally, many functions in
src/glsl (especially those dealing with linking) already use
gl_shader_stage to represent pipeline stages; using gl_shader::Stage
in those functions avoids the need for a conversion.

Note: in the process I changed _mesa_write_shader_to_file() so that if
it encounters an unexpected shader stage, it will use a file suffix of
"????" rather than "geom".

Reviewed-by: Brian Paul <brianp@vmware.com>
v2: Split from patch "mesa: Store gl_shader_stage enum in gl_shader objects."

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

mesa: Store gl_shader_stage enum in gl_shader objects.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Move declaration of gl_shader_stage earlier in mtypes.h.

Also move the related #define MESA_SHADER_STAGES.  This will allow
gl_shader_stage to be used in struct gl_shader.

Reviewed-by: Brian Paul <brianp@vmware.com>
v2: Split from patch "mesa: Store gl_shader_stage enum in gl_shader objects."

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

glsl: make _mesa_shader_stage_to_string() available to non-C++ code.

Reviewed-by: Brian Paul <brianp@vmware.com>
v2: Split from patch "mesa: Store gl_shader_stage enum in gl_shader objects."

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

mesa: Clean up nomenclature for pipeline stages.

Previously, we had an enum called gl_shader_type which represented
pipeline stages in the order they occur in the pipeline
(i.e. MESA_SHADER_VERTEX=0, MESA_SHADER_GEOMETRY=1, etc), and several
inconsistently named functions for converting between it and other
representations:

- _mesa_shader_type_to_string: gl_shader_type -> string
- _mesa_shader_type_to_index: GLenum (GL_*_SHADER) -> gl_shader_type
- _mesa_program_target_to_index: GLenum (GL_*_PROGRAM) -> gl_shader_type
- _mesa_shader_enum_to_string: GLenum (GL_*_{SHADER,PROGRAM}) -> string

This patch tries to clean things up so that we use more consistent
terminology: the enum is now called gl_shader_stage (to emphasize that
it is in the order of pipeline stages), and the conversion functions are:

- _mesa_shader_stage_to_string: gl_shader_stage -> string
- _mesa_shader_enum_to_shader_stage: GLenum (GL_*_SHADER) -> gl_shader_stage
- _mesa_program_enum_to_shader_stage: GLenum (GL_*_PROGRAM) -> gl_shader_stage
- _mesa_progshader_enum_to_string: GLenum (GL_*_{SHADER,PROGRAM}) -> string

In addition, MESA_SHADER_TYPES has been renamed to MESA_SHADER_STAGES,
for consistency with the new name for the enum.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v2: Also rename the "target" field of _mesa_glsl_parse_state and the
"target" parameter of _mesa_shader_stage_to_string to "stage".

Reviewed-by: Brian Paul <brianp@vmware.com>

llvmpipe: Fix the bottom_edge_rule adjustment for points.

The adjustment needs to be applied to the y coordinates and not the x
coordinates, just like the equivalent code for lines and triangles in
lp_setup_line.c and lp_setup_tri.c.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Zack Rusin <zackr@vmware.com>

llvmpipe: Respect bottom_edge_rule when computing the rasterization bounding boxes.

This was inadvertently forgotten when replacing gl_rasterization_rules
with lower_left_origin and half_pixel_center (commit
2737abb44efebfa10ac84b183c20fc5818d1514e).

This makes a difference when lower_left_origin != half_pixel_center, e.g,
D3D10.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Zack Rusin <zackr@vmware.com>

ilo: enable HiZ

The support is still early.  Fast depth buffer clear is not enabled yet.

HiZ can be forced off with ILO_DEBUG=nohiz.

ilo: resolve Z/HiZ correctly

When the depth buffer is to be read, perform a Depth Buffer Resolve if it has
been rendered.  When the depth buffer is to be rendered, perform a HiZ Buffer
Resolve when the depth buffer is modified externally.

ilo: add flags to texture slices

The flags are used to mark who (CPU, BLT, or RENDER) has accessed the resource
and how (READ or WRITE).

ilo: rename and add an accessor for texture slices

Rename ilo_texture::slice_offsets to ilo_texture::slices and add an accessor,
ilo_texture_get_slice().

ilo: add HiZ op support to the pipelines

Add blitter functions to perform Depth Buffer Clear, Depth Buffer Resolve, and
Hierarchical Depth Buffer Resolve.  Those functions set ilo_blitter up and
pass it to the pipelines to emit the commands.

ilo: add support for HiZ allocation

Add tex_create_hiz() to create HiZ bo.  It is not really called yet.

ilo: refactor separate stencil allocation

Move separate stencil allocation code to tex_create_separate_stencil to keep
tex_create sane.

ilo: assorted GPE fixes for HiZ

Allow HiZ op to be specified in 3DSTATE_WM.  Pass depth format directly in
gen7_emit_3DSTATE_SF.  Use tex->hiz.bo to determine if HiZ exists.  Fix
3DSTATE_SF for the case when there is no ilo_rasterizer_state.  Fix
3DSTATE_PS for the case when there is no ilo_shader_state.

ilo: no layer offsetting on GEN7+

Even though the Ivy Bridge PRM lists some restrictions that require layer
offsetting as the Sandy Bridge PRM does, it seems they are actually lifted.

ilo: offset to layers only when necessary

GEN6 has several requirements regarding the LOD/Depth/Width/Height of the
render targets and the depth buffer.  We used to offset to the layers in
question unconditionally to meet the requirements.  With this commit,
offseting is done only when the requirements are not met.

ilo: allow ilo_zs_surface to skip layer offsetting

Make offset to layer optional in ilo_gpe_init_zs_surface.