mesa.git
8 years agoi965/vec4: Spaces around operators.
Matt Turner [Tue, 19 Jan 2016 20:12:38 +0000 (12:12 -0800)]
i965/vec4: Spaces around operators.

8 years agoi965: Inform compiler of variable range to silence warning.
Matt Turner [Fri, 15 Jan 2016 21:38:46 +0000 (13:38 -0800)]
i965: Inform compiler of variable range to silence warning.

Extends commit 6531ccb70 to silence the warning in release builds as
well.

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agoglsl: Restore Mesa-style to shader_enums.c/h.
Matt Turner [Fri, 15 Jan 2016 21:31:34 +0000 (13:31 -0800)]
glsl: Restore Mesa-style to shader_enums.c/h.

8 years agost/va: fix motion adaptive deinterlacing
Christian König [Mon, 18 Jan 2016 19:56:06 +0000 (20:56 +0100)]
st/va: fix motion adaptive deinterlacing

Signed-off-by: Christian König <christian.koenig@amd.com>
8 years agoutil/u_pstipple.c: copy immediates during transformation
Nicolai Hähnle [Fri, 15 Jan 2016 21:56:15 +0000 (16:56 -0500)]
util/u_pstipple.c: copy immediates during transformation

Apparently, nobody has combined stippling with a fragment shader
containing immediates in almost five years...

Fixes a bug in Kodi with radeonsi reported by Christian König.

Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Tested-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agomesa: Move sanity check of BindVertexBuffer for OpenGL ES 3.1
Marta Lofstedt [Fri, 8 Jan 2016 13:55:55 +0000 (14:55 +0100)]
mesa: Move sanity check of BindVertexBuffer for OpenGL ES 3.1

Sanity check of BindVertexBuffer for OpenGL ES in
_mesa_handle_bind_buffer_gen breaks OpenGL ES 2 conformance.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93426
Signed-off-by: Marta Lofstedt <marta.lofstedt@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
8 years agoglsl: fix interface block error message
Timothy Arceri [Tue, 19 Jan 2016 03:35:50 +0000 (14:35 +1100)]
glsl: fix interface block error message

Print the stream value not the pointer to the expression,
also use the unsigned format specifier.

Cc: 11.1 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agonv50/ir: swap the least-ref'd source into src1 when both const/imm
Ilia Mirkin [Mon, 18 Jan 2016 03:28:19 +0000 (22:28 -0500)]
nv50/ir: swap the least-ref'd source into src1 when both const/imm

The whole point of inlining sources is to reduce loads. We can end up in
a situation where one value is used a lot of times, and one value is
used only once per instruction. The once-per-instruction one is the one
that should get inlined, but with the previous algorithm, it was given
no preference.

This flips things around to preferring putting less-referenced values
into src1 which increases the likelihood of them being inlined.

While we're at it, adjust the heuristic to not treat 0 as an immediate,
as well as (effectively) check for situations where LIMMs can't be
loaded. All this yields improvements on nvc0:

total instructions in shared programs : 6261157 -> 6255985 (-0.08%)
total gprs used in shared programs    : 945082 -> 943417 (-0.18%)
total local used in shared programs   : 30372 -> 30288 (-0.28%)
total bytes used in shared programs   : 50089256 -> 50047880 (-0.08%)

                local        gpr       inst      bytes
    helped          21         822        3332        3332
      hurt           0         278         565         565

And more importantly avoids generating really bad code with SSBOs, where
we end up checking a lot of different values (usually immediates) against
the length.

On nv50 we get comparable results, and even improve packing (bytes went
down more than instructions):

total instructions in shared programs : 6346564 -> 6341277 (-0.08%)
total gprs used in shared programs    : 728719 -> 725131 (-0.49%)
total local used in shared programs   : 3552 -> 3552 (0.00%)
total bytes used in shared programs   : 43995688 -> 43932928 (-0.14%)

                local        gpr       inst      bytes
    helped           0        1380        3252        3774
      hurt           0         287        1710        1365

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agost/mesa: restore the stObj's size if it was cleared out
Ilia Mirkin [Sun, 17 Jan 2016 21:25:00 +0000 (16:25 -0500)]
st/mesa: restore the stObj's size if it was cleared out

An issue could still occur if the base level is set, but fixing that
would require a lot more logic.

This fixes the recently-failing texelFetch 3D tests because the mipmaps
were no longer being generated, which in turn caused the copying logic
to be hit, which in turn didn't work because of the broken
width/height/depth.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agofreedreno/a4xx: use smaller threadsize for more registers
Rob Clark [Mon, 18 Jan 2016 20:30:53 +0000 (15:30 -0500)]
freedreno/a4xx: use smaller threadsize for more registers

Once we go past half of the "GPR" register file, it seems like we need
to run frag shader with smaller threadsize.  (The vertex shader already
runs at TWO_QUADS, which is the minimum.)

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno: per-generation OUT_IB packet
Rob Clark [Mon, 18 Jan 2016 20:22:27 +0000 (15:22 -0500)]
freedreno: per-generation OUT_IB packet

Some a4xx firmware doesn't implement the "PFD" (prefetch-disabled)
version of the CP_INDIRECT_BUFFER packet.  So allow for PFD vs PFE per
generation.  Switch a3xx and a4xx over to using prefetch-enabled version
(which is also what blob does.. it seems only on a2xx we cannot use
PFE).

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agogallium: bundle the compat header u_pwr8.h in the tarball
Emil Velikov [Mon, 18 Jan 2016 11:34:14 +0000 (13:34 +0200)]
gallium: bundle the compat header u_pwr8.h in the tarball

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agomapi: include gl.xml in the tarball
Emil Velikov [Mon, 18 Jan 2016 11:06:28 +0000 (13:06 +0200)]
mapi: include gl.xml in the tarball

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoi965: adding missing headers to the dist tarball
Emil Velikov [Thu, 14 Jan 2016 07:28:21 +0000 (09:28 +0200)]
i965: adding missing headers to the dist tarball

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
8 years agost/va: add motion adaptive deinterlacing v2
Christian König [Sun, 13 Dec 2015 10:44:13 +0000 (11:44 +0100)]
st/va: add motion adaptive deinterlacing v2

v2: minor cleanup

Signed-off-by: Christian König <christian.koenig@amd.com>
8 years agogallium/radeon: Rename do_invalidate_resource to invalidate_buffer
Michel Dänzer [Fri, 15 Jan 2016 07:02:22 +0000 (16:02 +0900)]
gallium/radeon: Rename do_invalidate_resource to invalidate_buffer

And only call it from r600_invalidate_resource for buffer resources.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agost/dri: Don't call invalidate_resource for NULL depth/stencil buffers
Michel Dänzer [Fri, 15 Jan 2016 06:46:31 +0000 (15:46 +0900)]
st/dri: Don't call invalidate_resource for NULL depth/stencil buffers

Fixes crash in 4 EGL piglit tests with radeonsi.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoradeonsi: Avoid warning about LLVM generating R_0286D0_SPI_PS_INPUT_ADDR
Michel Dänzer [Fri, 15 Jan 2016 03:18:29 +0000 (12:18 +0900)]
radeonsi: Avoid warning about LLVM generating R_0286D0_SPI_PS_INPUT_ADDR

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
8 years agoradeonsi: Print "LLVM emitted unknown config register" warning only once
Michel Dänzer [Fri, 15 Jan 2016 03:13:15 +0000 (12:13 +0900)]
radeonsi: Print "LLVM emitted unknown config register" warning only once

Say "LLVM" instead of "Compiler" for clarity.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agollvmpipe: use vpkswss when dst is signed
Oded Gabbay [Sun, 17 Jan 2016 20:15:40 +0000 (22:15 +0200)]
llvmpipe: use vpkswss when dst is signed

This patch fixes a bug when building a pack instruction.

For POWER (altivec), in case the destination is signed and the
src width is 32, we need to use vpkswss. The original code used vpkuwus,
which emits an unsigned result.

This fixes the following piglit tests on ppc64le:
- spec@arb_color_buffer_float@gl_rgba8-drawpixels
- shaders@glsl-fs-fogscale

I've also corrected some coding style issues in the function.

v2: Returned else statements to vmware style

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoglsl: fix subroutine lowering reusing actual parmaters
Dave Airlie [Sun, 17 Jan 2016 04:23:35 +0000 (14:23 +1000)]
glsl: fix subroutine lowering reusing actual parmaters

One of the oglconform tests was crashing here, and it was
due to not cloning the actual parameters before creating the
new call. This makes a call clone function that does the right
things to make sure we clone all the needed info, and points
the callee at it. (It differs from ->clone due to this).

this may fix https://bugs.freedesktop.org/show_bug.cgi?id=93722, I had this
patch in my cts fixes tree, but hadn't had time to make sure I liked it.

Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
8 years agoglsl: remove special case for detecting stream duplicates
Timothy Arceri [Fri, 15 Jan 2016 02:45:49 +0000 (13:45 +1100)]
glsl: remove special case for detecting stream duplicates

Any duplicates in a single declaration will already fail the
generic duplicates test due to the explicit_stream flag being set.

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
8 years agoglsl: add missing explicit_stream flag to has_layout()
Timothy Arceri [Fri, 15 Jan 2016 02:45:48 +0000 (13:45 +1100)]
glsl: add missing explicit_stream flag to has_layout()

This will allow the ARB_shading_language_420pack rules in
glsl_parser.yy for catching duplicate layout qualifiers to be
triggered for the stream identifier rather than relying on the
code meant to catch duplicates within a single layout(...)

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
8 years agomesa: fix segfault in glUniformSubroutinesuiv()
Timothy Arceri [Sun, 17 Jan 2016 05:09:08 +0000 (16:09 +1100)]
mesa: fix segfault in glUniformSubroutinesuiv()

From Section 7.9 (SUBROUTINE UNIFORM VARIABLES) of the OpenGL
4.5 Core spec:

   "The command

       void UniformSubroutinesuiv(enum shadertype, sizei count,
                                  const uint *indices);

   will load all active subroutine uniforms for shader stage
   shadertype with subroutine indices from indices, storing
   indices[i] into the uniform at location i. The indices for
   any locations between zero and the value of
   ACTIVE_SUBROUTINE_UNIFORM_LOCATIONS minus one which are not
   used will be ignored."

V2: simplify NULL check suggested by Jason.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Cc: "11.0 11.1" mesa-stable@lists.freedesktop.org
https://bugs.freedesktop.org/show_bug.cgi?id=93731

8 years agoglsl: fix segfault linking subroutine uniform with explicit location
Timothy Arceri [Mon, 18 Jan 2016 00:13:27 +0000 (11:13 +1100)]
glsl: fix segfault linking subroutine uniform with explicit location

Reviewed-by: Dave Airlie <airlied@redhat.com>
Cc: "11.0 11.1" mesa-stable@lists.freedesktop.org
8 years agogm107/ir: don't do indirect frag shader inputs on GM107
Ilia Mirkin [Sun, 17 Jan 2016 21:24:02 +0000 (16:24 -0500)]
gm107/ir: don't do indirect frag shader inputs on GM107

Apparently the IPA op decided to stop working with offsets. Need to
figure out if we need to do an AL2P situation or something similar. For
now just turn it back off.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agotgsi: initialize Atomic field in tgsi_default_declaration
Ilia Mirkin [Sun, 17 Jan 2016 08:44:22 +0000 (03:44 -0500)]
tgsi: initialize Atomic field in tgsi_default_declaration

Spotted by Coverity.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agonvc0: bsp_bo can't be null
Ilia Mirkin [Sun, 17 Jan 2016 08:41:55 +0000 (03:41 -0500)]
nvc0: bsp_bo can't be null

We already deref it earlier. And these are all allocated on load.
Spotted by Coverity.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agollvmpipe: fix arguments order given to vec_andc
Oded Gabbay [Sun, 17 Jan 2016 12:25:32 +0000 (14:25 +0200)]
llvmpipe: fix arguments order given to vec_andc

This patch fixes a classic "confuse the enemy" bug.

_mm_andnot_si128 (SSE) and vec_andc (VMX) do the same operation, but the
arguments are opposite.

_mm_andnot_si128 performs "r = (~a) & b" while
vec_andc performs "r = a & (~b)"

To make sure this error won't return in another place, I added a wrapper
function, vec_andnot_si128, in u_pwr8.h, which makes the swap inside.

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agofreedreno/ir3: fix mad 3rd src delay calc
Rob Clark [Sun, 17 Jan 2016 17:21:45 +0000 (12:21 -0500)]
freedreno/ir3: fix mad 3rd src delay calc

In fad158a0 ("freedreno/ir3: array rework") the src # (n) shifted by
one, but missed updating delay-slot calc.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: better array register allocation
Rob Clark [Sat, 16 Jan 2016 00:45:51 +0000 (19:45 -0500)]
freedreno/ir3: better array register allocation

Detect arrays which don't conflict with each other and allow overlapping
register allocation.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: array offset can be negative
Rob Clark [Fri, 15 Jan 2016 23:22:40 +0000 (18:22 -0500)]
freedreno/ir3: array offset can be negative

It at least happens with some piglit tests, like
$piglit/bin/vp-address-01

  VERT
  DCL IN[0]
  DCL IN[1]
  DCL OUT[0], POSITION
  DCL OUT[1], COLOR
  DCL CONST[0..7]
  DCL ADDR[0]
    0: ARL ADDR[0].x, IN[1].xxxx
    1: MOV_SAT OUT[1], CONST[ADDR[0].x-1]
    2: DP4 OUT[0].x, CONST[4], IN[0]
    3: DP4 OUT[0].y, CONST[5], IN[0]
    4: DP4 OUT[0].z, CONST[6], IN[0]
    5: DP4 OUT[0].w, CONST[7], IN[0]
    6: END

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: workaround bug/feature
Rob Clark [Fri, 15 Jan 2016 22:07:02 +0000 (17:07 -0500)]
freedreno/ir3: workaround bug/feature

Seems like in certain cases, we cannot use c<a0.x+0> as the third src to
cat3 instructions.  This may be slightly conservative, we may only have
this restriction when the first src is also const.

This fixes, for example, +24/-0 of the variable-indexing piglit tests.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agottn: use writemask for store_var
Rob Clark [Tue, 12 Jan 2016 15:24:10 +0000 (10:24 -0500)]
ttn: use writemask for store_var

Only user is freedreno, and after array-rework it can cope.  Avoids
generating loads for a store.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: array rework
Rob Clark [Sun, 10 Jan 2016 19:10:08 +0000 (14:10 -0500)]
freedreno/ir3: array rework

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: refactor/simplify cp
Rob Clark [Sun, 10 Jan 2016 16:03:46 +0000 (11:03 -0500)]
freedreno/ir3: refactor/simplify cp

If we handle separately the special case of eliminating output mov
(which includes keeps and various other cases where we don't have a
consuming instruction's src register to collapse things into), we
can simplify the logic.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: fix incorrect decoding of mov instructions
Rob Clark [Wed, 13 Jan 2016 18:07:51 +0000 (13:07 -0500)]
freedreno/ir3: fix incorrect decoding of mov instructions

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: remove unused tgsi tokens ptr
Rob Clark [Sat, 9 Jan 2016 19:46:36 +0000 (14:46 -0500)]
freedreno/ir3: remove unused tgsi tokens ptr

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: bit of ra refactor
Rob Clark [Wed, 6 Jan 2016 19:08:34 +0000 (14:08 -0500)]
freedreno/ir3: bit of ra refactor

Shuffle things slightly, passing instr-data to ra_name() to reduce the
number of places where we need to add support for array names.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: cosmetic de-indent
Rob Clark [Wed, 6 Jan 2016 18:32:24 +0000 (13:32 -0500)]
freedreno/ir3: cosmetic de-indent

Collapse two nested if's into one to reduce indent level.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agottn: add missing writemask on store_output
Rob Clark [Wed, 13 Jan 2016 23:39:56 +0000 (18:39 -0500)]
ttn: add missing writemask on store_output

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonir/print: const_index is signed
Rob Clark [Fri, 15 Jan 2016 23:24:11 +0000 (18:24 -0500)]
nir/print: const_index is signed

Noticed this with $piglit/bin/vp-address-01

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agonir: few missing struct names
Rob Clark [Mon, 4 Jan 2016 18:24:08 +0000 (13:24 -0500)]
nir: few missing struct names

nir.h is a bit inconsistent about 'typedef struct {} nir_foo' vs
'typedef struct nir_foo {} nir_foo'.  But missing struct name tags is
inconvenient when you need a fwd declaration without pulling in all
of nir.

So add missing struct name tag for nir_variable, and a couple other
spots where it would likely be useful.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
8 years agonv50/ir: add saturate support on ex2
Ilia Mirkin [Fri, 15 Jan 2016 22:12:27 +0000 (17:12 -0500)]
nv50/ir: add saturate support on ex2

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agogallivm: avoid crashing in mod by 0 with llvmpipe
Jeff Muizelaar [Sat, 16 Jan 2016 02:35:26 +0000 (03:35 +0100)]
gallivm: avoid crashing in mod by 0 with llvmpipe

This adds code that is basically the same as the code in umod, udiv and idiv.
However, unlike idiv we return -1.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoglsl: Allow implicit int -> uint conversions for bitwise operators (&, ^, |).
Kenneth Graunke [Fri, 15 Jan 2016 07:27:03 +0000 (23:27 -0800)]
glsl: Allow implicit int -> uint conversions for bitwise operators (&, ^, |).

The ARB has decided that implicit conversions should be performed for
bitwise operators in future language revisions.  Implementations of
current language revisions may or may not perform them.

This patch makes Mesa apply implicti conversions even on current
language versions.  Applications appear to expect this behavior,
and there's really no downside to doing so.

Fixes shader compilation in Shadow of Mordor.

Bugzilla: https://www.khronos.org/bugzilla/show_bug.cgi?id=1405
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Cc: mesa-stable@lists.freedesktop.org
8 years agoi965/fs: Always set channel 2 of texture headers in some stages
Jason Ekstrand [Fri, 15 Jan 2016 04:42:47 +0000 (20:42 -0800)]
i965/fs: Always set channel 2 of texture headers in some stages

In the vertex and fragment stages, the hardware is nice to us and leaves
g0.2 zerod out for us so we can use it for headers.  However, in compute,
geometry, and tessellation stages, the hardware is not so nice.  In
particular, for compute shaders on BDW, the hardware places some debug bits
in 23:15.  As it happens, bit 15 is interpreted by the sampler as the alpha
channel mask.  This means that if you use a texturing instruction with a
header in a compute shader, you may randomly get the alpha channel
disabled.  Since channel masks affect the return length of the sampler
message, this can lead the GPU to expect a different mlen to the one you
specified in the shader and this, in turn, hangs your GPU.

Cc: "11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
8 years agoi965/fs/generator: Take an actual shader stage rather than a string
Jason Ekstrand [Fri, 15 Jan 2016 04:27:51 +0000 (20:27 -0800)]
i965/fs/generator: Take an actual shader stage rather than a string

Cc: "11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoi965/vec4: Use UW type for multiply into accumulator on GEN8+
Jason Ekstrand [Thu, 14 Jan 2016 20:08:57 +0000 (12:08 -0800)]
i965/vec4: Use UW type for multiply into accumulator on GEN8+

BDW adds the following restriction: "When multiplying DW x DW, the dst
cannot be accumulator."

Cc: "11.1,11.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agollvmpipe: ditch additional ref counting for vertex/geometry sampler views
Roland Scheidegger [Fri, 15 Jan 2016 02:10:40 +0000 (03:10 +0100)]
llvmpipe: ditch additional ref counting for vertex/geometry sampler views

The cleaning up was quite a performance hog (making pipe_resource_reference
the number two in profilers on the vertex path, and 3rd overall, with its
cousin pipe_reference_described not far behind) if there were lots
of tiny draw calls (ipers). Now the reason was really that it was blindly
calling this for all potential shader views (so 32 each for vs and gs) even
though the app never touched a single one which could have been fixed,
however I can't come up with a good reason why we refcount these. We've got
references, of course, in the sampler views, which should be quite sufficient
as we do all vertex and geometry shader execution fully synchronous.
(Calling prepare_shader_sampling for all draw calls even if there were no
changes looks quite suboptimal too, but generally we don't really expect vs/gs
shader sampling to be used much with llvmpipe, and there's even an early exit
if there aren't any views to avoid the "null loop" albeit it's now no longer
always trying to loop through all 32 slots. Maybe improve another time...).
Of course, if we manage to make vertex loads run asynchronously some day,
we need references again, but adding that back would be the least of the
problems...
Also only set LP_NEW_SAMPLER_VIEW for fragment sampler views. Nothing on the
vertex side depends on it (I suppose we'd really wanted a separate flag in
any case).
(Good for a 3% improvement or so in ipers under the right conditions.)

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agollvmpipe: fix "leaking" textures
Roland Scheidegger [Fri, 15 Jan 2016 19:12:24 +0000 (20:12 +0100)]
llvmpipe: fix "leaking" textures

This was not really a leak per se, but we were referencing the textures for
longer than intended. If textures were set via llvmpipe_set_sampler_views()
(for fs) and then picked up by lp_setup_set_fragment_sampler_views(), they
were referenced in the setup state. However, the only way to unreference them
was by replacing them with another texture, and not when the texture slot
was replaced with a NULL sampler view. (They were then further also referenced
by the scene too which might have additional minor side effects as we limit
the memory size which is allowed to be referenced by a scene in a rather crude
way.) Only setup destruction (at context destruction time) then finally would
get rid of the references.
Fix this by noting the number of textures the last time, and unreference
things if the new view is NULL (avoiding having to unreference things
always up to PIPE_MAX_SHADER_SAMPLER_VIEWS which would also have worked).
Found by code inspection, no test...

v2: rename var

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agoglsl: restrict consumer stage condition to modify interpolation type
Samuel Iglesias Gonsálvez [Tue, 12 Jan 2016 14:36:56 +0000 (15:36 +0100)]
glsl: restrict consumer stage condition to modify interpolation type

Only modify interpolation type for integer-based varyings or when the
consumer is known and different than fragment shader.

If we are linking separate shader programs and the consumer is unknown,
the consumer could be added later and be a fragment shader. If we
modify the interpolation type in this case, we could read wrong
values in the fragment shader inputs, as shown in bug 93320.

Fixes the following CTS test:
   ES31-CTS.vertex_attrib_binding.advanced-bindingUpdate

Fixes the following dEQP tests:

dEQP-GLES31.functional.separate_shader.random.102
dEQP-GLES31.functional.separate_shader.random.111
dEQP-GLES31.functional.separate_shader.random.115
dEQP-GLES31.functional.separate_shader.random.17
dEQP-GLES31.functional.separate_shader.random.22
dEQP-GLES31.functional.separate_shader.random.23
dEQP-GLES31.functional.separate_shader.random.3
dEQP-GLES31.functional.separate_shader.random.32
dEQP-GLES31.functional.separate_shader.random.39
dEQP-GLES31.functional.separate_shader.random.64
dEQP-GLES31.functional.separate_shader.random.73
dEQP-GLES31.functional.separate_shader.random.91

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93320
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
8 years agoi965: Apply add_const_offset_to_base for vec4 VS inputs too.
Kenneth Graunke [Wed, 13 Jan 2016 23:07:18 +0000 (15:07 -0800)]
i965: Apply add_const_offset_to_base for vec4 VS inputs too.

This shouldn't hurt anything, and I'm about to introduce a pass that
will want it.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoi965: Make add_const_offset_to_base() work at the shader level.
Kenneth Graunke [Wed, 13 Jan 2016 23:04:39 +0000 (15:04 -0800)]
i965: Make add_const_offset_to_base() work at the shader level.

This makes it a pass, hiding the parameter structs and block callbacks
so it's simpler to work with.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoi965: Make an is_scalar boolean in brw_compile_vs().
Kenneth Graunke [Wed, 13 Jan 2016 23:23:48 +0000 (15:23 -0800)]
i965: Make an is_scalar boolean in brw_compile_vs().

Shorter than compiler->scalar_stage[MESA_SHADER_VERTEX], which can
help with line-wrapping.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agonir/builder: Add a nir_build_ivec4() convenience helper.
Kenneth Graunke [Wed, 13 Jan 2016 21:32:44 +0000 (13:32 -0800)]
nir/builder: Add a nir_build_ivec4() convenience helper.

nir_build_ivec4 is more readable and succinct than using nir_build_imm
directly, even if you have C99.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
8 years agoglsl: mark explicit uniforms as explicit in other stages too
Tapani Pälli [Thu, 14 Jan 2016 12:10:59 +0000 (14:10 +0200)]
glsl: mark explicit uniforms as explicit in other stages too

If shader declares uniform explicit location in one stage but
implicit in another, explicit location should be used. Patch marks
implicit uniforms as explicit if they were explicit in previous stage.
This makes sure that we don't treat them implicit later when assigning
locations.

Fixes following CTS test:
   ES31-CTS.explicit_uniform_location.uniform-loc-implicit-in-some-stages3

v2: move check to cross_validate_globals (Timothy)

Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
8 years agoi965/gen7.5+: Disable resource streamer during GPGPU workloads.
Francisco Jerez [Sun, 3 Jan 2016 18:06:52 +0000 (10:06 -0800)]
i965/gen7.5+: Disable resource streamer during GPGPU workloads.

The RS and hardware binding tables are only supported on the 3D
pipeline and can lead to corruption if left enabled during a GPGPU
workload.  Disable it when switching to the GPGPU (or media) pipeline
and re-enable it when switching back to the 3D pipeline.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
8 years agoi965/gen7: Emit stall and dummy primitive draw after switching to the 3D pipeline.
Francisco Jerez [Sun, 3 Jan 2016 03:06:48 +0000 (19:06 -0800)]
i965/gen7: Emit stall and dummy primitive draw after switching to the 3D pipeline.

This hardware bug can supposedly lead to a hang on IVB and VLV.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/gen4-5: Emit MI_FLUSH as required prior to switching pipelines.
Francisco Jerez [Sun, 3 Jan 2016 03:05:48 +0000 (19:05 -0800)]
i965/gen4-5: Emit MI_FLUSH as required prior to switching pipelines.

AFAIK brw_emit_select_pipeline() is only called once during context
init on Gen4-5, at which point the pipeline is likely to be already
idle so it may just happen to work by luck regardless of the MI_FLUSH.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/gen6-7: Implement stall and flushes required prior to switching pipelines.
Francisco Jerez [Sun, 3 Jan 2016 03:02:09 +0000 (19:02 -0800)]
i965/gen6-7: Implement stall and flushes required prior to switching pipelines.

Switching the current pipeline while it's not completely idle or the
read and write caches aren't flushed can lead to corruption.  Fixes
misrendering of at least the following Khronos CTS test:

 ES31-CTS.shader_image_load_store.basic-allTargets-store-fs

The stall and flushes are no longer required on Gen8+.

v2: Emit PIPE_CONTROL with non-zero post-sync op before the write
    cache flush on SNB due to hardware bug. (Ken)

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93323
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/gen8+: Invalidate color calc state when switching to the GPGPU pipeline.
Francisco Jerez [Sun, 3 Jan 2016 02:35:28 +0000 (18:35 -0800)]
i965/gen8+: Invalidate color calc state when switching to the GPGPU pipeline.

This hardware bug can cause a hang on context restore while the
current pipeline is set to GPGPU (BDWGFX HSD 1909593).  In addition to
clearing the valid bit, mark the CC state as dirty to make sure that
the CC indirect state pointer is re-emitted when we switch back to the
3D pipeline.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965: Add state bit to trigger re-emission of color calculator state.
Francisco Jerez [Wed, 23 Dec 2015 16:44:59 +0000 (18:44 +0200)]
i965: Add state bit to trigger re-emission of color calculator state.

This will be used on Gen8+ to make sure that the color calculator
state pointers are re-emitted when switching back to the 3D pipeline
after some GPGPU workload due to a hardware workaround.  There are
other state bits already defined that could be used to achieve the
same effect but they all cause a ton of unrelated state to be
re-emitted (e.g. BRW_NEW_STATE_BASE_ADDRESS), so just define a new
one, state bits are cheap.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonv50/ir: rebase indirect temp arrays to 0, so that we use less lmem space
Ilia Mirkin [Thu, 14 Jan 2016 06:09:25 +0000 (01:09 -0500)]
nv50/ir: rebase indirect temp arrays to 0, so that we use less lmem space

Reduces local memory usage in a lot of Metro 2033 Redux and a few KSP
shaders:

total local used in shared programs   : 54116 -> 30372 (-43.88%)

Probably modest advantage to execution, but it's an imporant
prerequisite to dropping some of the TGSI optimizations done by the
state tracker.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agonv50/ir: only use FILE_LOCAL_MEMORY for temp arrays that use indirection
Ilia Mirkin [Tue, 12 Jan 2016 19:41:52 +0000 (14:41 -0500)]
nv50/ir: only use FILE_LOCAL_MEMORY for temp arrays that use indirection

Previously we were treating any indirect temp array usage to mean that
everything should end up in lmem. The MemoryOpt pass would clean a lot
of that up later, but in the meanwhile we would lose a lot of
opportunity for optimization.

This helps a lot of Metro 2033 Redux and a handful of KSP shaders:

total instructions in shared programs : 6288373 -> 6261517 (-0.43%)
total gprs used in shared programs    : 944051 -> 945131 (0.11%)
total local used in shared programs   : 54116 -> 54116 (0.00%)

A typical case is for register usage to double and for instructions to
halve. A future commit can also optimize local memory usage size to be
reduced with better packing.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agonvc0/ir: be careful about propagating very large offsets into const load
Ilia Mirkin [Mon, 11 Jan 2016 21:41:18 +0000 (16:41 -0500)]
nvc0/ir: be careful about propagating very large offsets into const load

Indirect constbuf indexing works by using very large offsets. However if
an indirect constbuf index load is const-propagated, it becomes a very
large const offset. Take that into account when legalizing the SSA by
moving the high parts of that offset into the file index. Also disallow
very large (or small) indices on most other instructions.

This fixes regressions in ubo_array_indexing/*-two-arrays piglit tests.

Fixes: abd326e81b (nv50/ir: propagate indirect loads into instructions)
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agonvc0: allow fragment shader inputs to use indirect indexing
Ilia Mirkin [Mon, 11 Jan 2016 21:39:15 +0000 (16:39 -0500)]
nvc0: allow fragment shader inputs to use indirect indexing

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agost/mesa: use surface format to generate mipmaps when available
Ilia Mirkin [Thu, 14 Jan 2016 18:44:54 +0000 (13:44 -0500)]
st/mesa: use surface format to generate mipmaps when available

This fixes the recently posted mipmap + texture views piglit test.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
8 years agoradeonsi: don't miss changes to SPI_TMPRING_SIZE
Marek Olšák [Wed, 13 Jan 2016 17:42:02 +0000 (18:42 +0100)]
radeonsi: don't miss changes to SPI_TMPRING_SIZE

I'm not sure about the consequences of this bug, but it's definitely
dangerous.

This applies to SI, CIK, VI.

Cc: 11.0 11.1 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agosvga: add DXGenMips command support
Charmaine Lee [Tue, 22 Dec 2015 19:20:41 +0000 (11:20 -0800)]
svga: add DXGenMips command support

For those formats that support hw mipmap generation, use the
DXGenMips command. Otherwise fallback to the mipmap generation utility.

Tested with piglit, OpenGL apps (Heaven, Turbine, Cinebench)

v2: make sure the texture surface was created with the render target bind flag
    set relocation flag to SVGA_RELOC_WRITE for the texture surface

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agosvga: add num-generate-mipmap HUD query
Charmaine Lee [Mon, 21 Dec 2015 19:07:08 +0000 (11:07 -0800)]
svga: add num-generate-mipmap HUD query

The actual increment of the num-generate-mipmap counter will be done
in a subsequent patch when hw generate mipmap is supported.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agogallium/st: add pipe_context::generate_mipmap()
Charmaine Lee [Thu, 14 Jan 2016 17:22:17 +0000 (10:22 -0700)]
gallium/st: add pipe_context::generate_mipmap()

This patch adds a new interface to support hardware mipmap generation.
PIPE_CAP_GENERATE_MIPMAP is added to allow a driver to specify
if this new interface is supported; if not supported, the state tracker will
fallback to mipmap generation by rendering/texturing.

v2: add PIPE_CAP_GENERATE_MIPMAP to the disabled section for all drivers
v3: add format to the generate_mipmap interface to allow mipmap generation
    using a format other than the resource format
v4: fix return type of trace_context_generate_mipmap()

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
8 years agost/mesa: declare struct pipe_screen in st_cb_bufferobjects.h
Brian Paul [Thu, 14 Jan 2016 17:38:17 +0000 (10:38 -0700)]
st/mesa: declare struct pipe_screen in st_cb_bufferobjects.h

To silence a compiler warning.  Trivial.

8 years agonir: Lower bitfield_extract.
Matt Turner [Wed, 13 Jan 2016 19:09:11 +0000 (11:09 -0800)]
nir: Lower bitfield_extract.

The OpenGL specifications for bitfieldExtract() says:

   The result will be undefined if <offset> or <bits> is negative, or if
   the sum of <offset> and <bits> is greater than the number of bits
   used to store the operand.

Therefore passing bits=32, offset=0 is legal and defined in GLSL.

But the earlier SM5 ubfe/ibfe opcodes are specified to accept a bitfield width
ranging from 0-31. As such, Intel and AMD instructions read only the low 5 bits
of the width operand, making them not able to implement the GLSL-specified
behavior directly.

This commit adds ubfe/ibfe operations from SM5 and a lowering pass for
bitfield_extract to to handle the trivial case of <bits> = 32 as

   bitfieldExtract:
      bits > 31 ? value : bfe(value, offset, bits)

Fixes:
   ES31-CTS.shader_bitfield_operation.bitfieldExtract.uvec3_0
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
8 years agonir: Handle <bits>=32 case in bitfield_insert lowering.
Matt Turner [Wed, 13 Jan 2016 19:08:37 +0000 (11:08 -0800)]
nir: Handle <bits>=32 case in bitfield_insert lowering.

The OpenGL specifications for bitfieldInsert() says:

   The result will be undefined if <offset> or <bits> is negative, or if
   the sum of <offset> and <bits> is greater than the number of bits
   used to store the operand.

Therefore passing bits=32, offset=0 is legal and defined in GLSL.

But the earlier SM5 bfi opcode is specified to accept a bitfield width
ranging from 0-31. As such, Intel and AMD instructions read only the low
5 bits of the width operand, making them not able to implement the
GLSL-specified behavior directly.

This commit fixes the lowering of bitfield_insert to handle the trivial
case of <bits> = 32 as

   bitfieldInsert:
      bits > 31 ? insert : bfi(bfm(bits, offset), insert, base)

Fixes:
   ES31-CTS.shader_bitfield_operation.bitfieldInsert.uint_2
   ES31-CTS.shader_bitfield_operation.bitfieldInsert.uvec4_3
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92595
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Tested-by: Marta Lofstedt <marta.lofstedt@intel.com>
8 years agost/mesa: add check for color logicop in blit_copy_pixels()
Brian Paul [Wed, 13 Jan 2016 23:20:09 +0000 (16:20 -0700)]
st/mesa: add check for color logicop in blit_copy_pixels()

We check that a bunch of raster operations are disabled in
blit_copy_pixels().  We also need to check that color logicop is
disabled.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agogallium/radeon: do not reallocate user memory buffers
Nicolai Hähnle [Tue, 12 Jan 2016 14:29:18 +0000 (09:29 -0500)]
gallium/radeon: do not reallocate user memory buffers

The whole point of AMD_pinned_memory is that applications don't have to map
buffers via OpenGL - but they're still allowed to, so make sure we don't break
the link between buffer object and user memory unless explicitly instructed
to.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agogallium/radeon: implement PIPE_CAP_INVALIDATE_BUFFER
Nicolai Hähnle [Thu, 14 Jan 2016 14:41:04 +0000 (09:41 -0500)]
gallium/radeon: implement PIPE_CAP_INVALIDATE_BUFFER

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agogallium/radeon: reset valid_buffer_range on PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
Nicolai Hähnle [Mon, 11 Jan 2016 00:54:44 +0000 (19:54 -0500)]
gallium/radeon: reset valid_buffer_range on PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE

This accomodates a streaming pattern where the discard flag is set when the
application wraps back to the beginning of the buffer.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agost/mesa: implement Driver.InvalidateBufferSubData
Nicolai Hähnle [Sat, 9 Jan 2016 23:05:58 +0000 (18:05 -0500)]
st/mesa: implement Driver.InvalidateBufferSubData

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agost/mesa: use pipe->invalidate_resource instead of buffer re-allocation
Nicolai Hähnle [Mon, 11 Jan 2016 22:44:45 +0000 (17:44 -0500)]
st/mesa: use pipe->invalidate_resource instead of buffer re-allocation

Drivers are expected to avoid unnecessary work when possible in this code
path.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agogallium: add PIPE_CAP_INVALIDATE_BUFFER
Nicolai Hähnle [Mon, 11 Jan 2016 22:38:08 +0000 (17:38 -0500)]
gallium: add PIPE_CAP_INVALIDATE_BUFFER

It makes sense to re-use pipe->invalidate_resource for the purpose of
glInvalidateBufferData, but this function is already implemented in vc4
where it doesn't have the expected behavior. So add a capability flag
to indicate that the driver supports the expected behavior.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agomesa: add Driver.InvalidateBufferSubData
Nicolai Hähnle [Sat, 9 Jan 2016 22:53:07 +0000 (17:53 -0500)]
mesa: add Driver.InvalidateBufferSubData

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agomesa: fix the checks in _mesa_InvalidateBuffer(Sub)Data
Nicolai Hähnle [Sat, 9 Jan 2016 22:51:39 +0000 (17:51 -0500)]
mesa: fix the checks in _mesa_InvalidateBuffer(Sub)Data

Change the check to be in line with what the quoted spec fragment says.

I have sent out a piglit test for this as well.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agowinsys/radeon: fix warnings about incompatible pointer types
Nicolai Hähnle [Tue, 12 Jan 2016 17:28:11 +0000 (12:28 -0500)]
winsys/radeon: fix warnings about incompatible pointer types

Some confusion between pb_buffer and radeon_bo as well as between
radeon_drm_winsys and radeon_winsys.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agotexobj: Check completeness with InternalFormat rather than Mesa format
Neil Roberts [Wed, 13 Jan 2016 19:28:45 +0000 (19:28 +0000)]
texobj: Check completeness with InternalFormat rather than Mesa format

The internal Mesa format used for a texture might not match the one
requested in the internalFormat when the texture was created, for
example if the driver is internally remapping RGB textures to RGBA.
Otherwise it can cause false positives for completeness if one mipmap
image is created as RGBA and the other as RGB because they would both
have an RGBA Mesa format. If we check the InternalFormat instead then
we are directly checking the API usage which I think better matches
the intention of the check.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93700
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoi965: Remove unused hw_must_use_separate_stencil
Ben Widawsky [Wed, 30 Dec 2015 17:47:17 +0000 (09:47 -0800)]
i965: Remove unused hw_must_use_separate_stencil

I spotted this while looking for what needs updating in future platforms.

I'm too lazy to go through the git logs, but it was probably missed by Jason
when all the brw refactoring happened.

Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965: Drop extra newline from shader compile messages.
Matt Turner [Thu, 14 Jan 2016 00:17:26 +0000 (16:17 -0800)]
i965: Drop extra newline from shader compile messages.

Ilia changed shader-db's run.c to not expect messages to contain a
newline in shader-db commit 51bbc8035.

8 years agonir: Change bfm's semantics to match Intel/AMD/SM5.
Matt Turner [Fri, 8 Jan 2016 00:16:35 +0000 (16:16 -0800)]
nir: Change bfm's semantics to match Intel/AMD/SM5.

Intel/AMD's hardware instructions do not handle arguments of 32.
Constant evaluation should not produce a result different from the
hardware instruction.

The s/1ull/1u/ change is intentional: previously we wanted defined
behavior for the "1 << 32" case, but we're making this case undefined so
we can make it 1u and save ourselves a 64-bit operation.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Fix undefined shifts.
Matt Turner [Wed, 30 Dec 2015 19:48:22 +0000 (14:48 -0500)]
glsl: Fix undefined shifts.

Shifting into the sign bit is undefined, as is shifting by 32.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Handle failure of Python codegen scripts.
Matt Turner [Mon, 11 Jan 2016 18:54:19 +0000 (10:54 -0800)]
glsl: Handle failure of Python codegen scripts.

If a Python codegen script failed, it would write a zero-byte file,
which on subsequent invocations of make would trick it into thinking the
file was appropriately generated.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl, nir: Make ir_triop_bitfield_extract a vectorized operation.
Kenneth Graunke [Fri, 8 Jan 2016 00:01:51 +0000 (16:01 -0800)]
glsl, nir: Make ir_triop_bitfield_extract a vectorized operation.

We would like to be able to combine

   result.x = bitfieldExtract(src0.x, src1.x, src2.x);
   result.y = bitfieldExtract(src0.y, src1.y, src2.y);
   result.z = bitfieldExtract(src0.z, src1.z, src2.z);
   result.w = bitfieldExtract(src0.w, src1.w, src2.w);

into a single ivec4 bitfieldInsert operation.  This should be possible
with most drivers.

This patch changes the offset and bits parameters from scalar ints
to ivecN or uvecN.  The type of all three operands will be the same,
for simplicity.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agoglsl, nir: Make ir_quadop_bitfield_insert a vectorized operation.
Kenneth Graunke [Tue, 5 Jan 2016 12:01:11 +0000 (04:01 -0800)]
glsl, nir: Make ir_quadop_bitfield_insert a vectorized operation.

We would like to be able to combine

   result.x = bitfieldInsert(src0.x, src1.x, src2.x, src3.x);
   result.y = bitfieldInsert(src0.y, src1.y, src2.y, src3.y);
   result.z = bitfieldInsert(src0.z, src1.z, src2.z, src3.z);
   result.w = bitfieldInsert(src0.w, src1.w, src2.w, src3.w);

into a single ivec4 bitfieldInsert operation.  This should be possible
with most drivers.

This patch changes the offset and bits parameters from scalar ints
to ivecN or uvecN.  The type of all four operands will be the same,
for simplicity.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agoglsl: Delete the ir_binop_bfm and ir_triop_bfi opcodes.
Kenneth Graunke [Thu, 7 Jan 2016 23:54:16 +0000 (15:54 -0800)]
glsl: Delete the ir_binop_bfm and ir_triop_bfi opcodes.

TGSI doesn't use these - it just translates ir_quadop_bitfield_insert
directly.  NIR can handle ir_quadop_bitfield_insert as well.

These opcodes were only used for i965, and with Jason's recent patches,
we can do this lowering in NIR (which also gains us SPIR-V handling).
So there's not much point to retaining this GLSL IR lowering code.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agonir: Fix constant evaluation of bfm.
Matt Turner [Mon, 11 Jan 2016 20:13:24 +0000 (12:13 -0800)]
nir: Fix constant evaluation of bfm.

NIR's bfm, like Intel/AMD's hardware instructions and GLSL IR's
ir_binop_bfm takes <bits> as src0 and <offset> as src1.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Skip assertion on NaN.
Matt Turner [Mon, 11 Jan 2016 17:34:50 +0000 (09:34 -0800)]
i965/fs: Skip assertion on NaN.

A shader in Unreal4 uses the result of divide by zero in its color
output, producing NaN and triggering this assertion since NaN is not
equal to itself.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93560
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Add debugging to constant combining pass.
Matt Turner [Fri, 27 Feb 2015 00:06:45 +0000 (16:06 -0800)]
i965/fs: Add debugging to constant combining pass.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agometa: remove const qualifier on _mesa_meta_fb_tex_blit_begin()
Brian Paul [Tue, 12 Jan 2016 15:46:40 +0000 (08:46 -0700)]
meta: remove const qualifier on _mesa_meta_fb_tex_blit_begin()

To silence a compiler warning about a const/non-const mismatch.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agost/mesa: fix incorrect buffer token passed to _mesa_BindFramebuffer()
Brian Paul [Tue, 12 Jan 2016 01:22:50 +0000 (18:22 -0700)]
st/mesa: fix incorrect buffer token passed to _mesa_BindFramebuffer()

I added this code right at the end, and got it wrong.
Only used by the WGL_ARB_render_texture code.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>
8 years agodocs: add news item and link release notes for 11.1.1
Emil Velikov [Wed, 13 Jan 2016 13:27:50 +0000 (15:27 +0200)]
docs: add news item and link release notes for 11.1.1

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>