mesa.git
8 years agoi965/vec4: Take into account misalignment in regs_written() and regs_read().
Francisco Jerez [Sat, 3 Sep 2016 00:57:34 +0000 (17:57 -0700)]
i965/vec4: Take into account misalignment in regs_written() and regs_read().

Unlike the FS counterpart of this commit this was likely not (yet) a
bug, but let's fix it already in preparation for implementing support
for sub-GRF offsets in the VEC4 back-end.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Take into account misalignment in regs_written() and regs_read().
Francisco Jerez [Wed, 7 Sep 2016 21:36:32 +0000 (14:36 -0700)]
i965/fs: Take into account misalignment in regs_written() and regs_read().

There was a workaround for this in fs_inst::size_read() for the
SHADER_OPCODE_MOV_INDIRECT instruction and FIXED_GRF register file
*only*.  We should take this possibility into account for the sources
and destinations of all instructions on all optimization passes that
need to quantize dataflow in 32B increments by adding the amount of
misalignment to the size read or written from the regs_read() and
regs_written() helpers respectively.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Take into account trailing padding in regs_written() and regs_read().
Francisco Jerez [Wed, 7 Sep 2016 21:33:55 +0000 (14:33 -0700)]
i965/fs: Take into account trailing padding in regs_written() and regs_read().

This fixes regs_written() and regs_read() to return a more accurate
value when the padding left between components due to a stride value
greater than one causes the region bounds given by size_written or
size_read to overflow into the next register.  This could become a
problem in optimization passes that keep track of dataflow using
fixed-size arrays with register granularity, because the overflow
register (not actually accessed by the region) may not have been
allocated at all which could lead to undefined memory access.

An alternative to this would be to subtract the trailing padding
already during the calculation of fs_inst::size_read and
::size_written, but that would break code that currently assumes that
::size_read and _written are whole multiples of the component size,
and would be hard to maintain looking forward because size_written is
assigned from a bunch of different places.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Handle fixed HW GRF subnr in reg_offset().
Francisco Jerez [Wed, 7 Sep 2016 20:41:08 +0000 (13:41 -0700)]
i965/fs: Handle fixed HW GRF subnr in reg_offset().

This will be useful later on when we start using reg_offset() on fixed
hardware registers.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Handle arbitrary offsets in brw_reg_from_fs_reg for MRF/VGRF registers.
Francisco Jerez [Fri, 2 Sep 2016 04:25:18 +0000 (21:25 -0700)]
i965/fs: Handle arbitrary offsets in brw_reg_from_fs_reg for MRF/VGRF registers.

This restriction seemed rather artificial...  Removing it actually
simplifies things slightly.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Return more accurate read size for LINTERP from fs_inst::size_read.
Francisco Jerez [Wed, 7 Sep 2016 20:02:55 +0000 (13:02 -0700)]
i965/fs: Return more accurate read size for LINTERP from fs_inst::size_read.

The LINTERP virtual instruction only reads three scalar components
from the first 16B of the second source, we can now teach size_read()
about it since its return value is represented with byte granularity.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Return more accurate read size from fs_inst::size_read for IMM and UNIFORM...
Francisco Jerez [Fri, 2 Sep 2016 23:23:44 +0000 (16:23 -0700)]
i965/fs: Return more accurate read size from fs_inst::size_read for IMM and UNIFORM files.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/vec4: Replace vec4_instruction::regs_read with ::size_read using byte units.
Francisco Jerez [Thu, 8 Sep 2016 00:00:30 +0000 (17:00 -0700)]
i965/vec4: Replace vec4_instruction::regs_read with ::size_read using byte units.

The previous regs_read value can be recovered by rewriting each
reference of regs_read() like 'x = i.regs_read(j)' to 'x =
DIV_ROUND_UP(i.size_read(j), reg_unit)'.

For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible.  I'll come
back later to clean up any ugliness introduced here.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Replace fs_inst::regs_read with ::size_read using byte units.
Francisco Jerez [Thu, 8 Sep 2016 00:00:07 +0000 (17:00 -0700)]
i965/fs: Replace fs_inst::regs_read with ::size_read using byte units.

The previous regs_read value can be recovered by rewriting each
reference of regs_read() like 'x = i.regs_read(j)' to 'x =
DIV_ROUND_UP(i.size_read(j), reg_unit)'.

For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible.  I'll come
back later to clean up any ugliness introduced here.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/ir: Drop backend_instruction::regs_written field.
Francisco Jerez [Sun, 4 Sep 2016 01:19:59 +0000 (18:19 -0700)]
i965/ir: Drop backend_instruction::regs_written field.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/vec4: Replace vec4_instruction::regs_written with ::size_written field in bytes.
Francisco Jerez [Sat, 3 Sep 2016 01:00:21 +0000 (18:00 -0700)]
i965/vec4: Replace vec4_instruction::regs_written with ::size_written field in bytes.

The previous regs_written field can be recovered by rewriting each
rvalue reference of regs_written like 'x = i.regs_written' to 'x =
DIV_ROUND_UP(i.size_written, reg_unit)', and each lvalue reference
like 'i.regs_written = x' to 'i.size_written = x * reg_unit'.

For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible.  I'll come
back later to clean up any ugliness introduced here.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Replace fs_inst::regs_written with ::size_written field in bytes.
Francisco Jerez [Wed, 7 Sep 2016 20:38:20 +0000 (13:38 -0700)]
i965/fs: Replace fs_inst::regs_written with ::size_written field in bytes.

The previous regs_written field can be recovered by rewriting each
rvalue reference of regs_written like 'x = i.regs_written' to 'x =
DIV_ROUND_UP(i.size_written, reg_unit)', and each lvalue reference
like 'i.regs_written = x' to 'i.size_written = x * reg_unit'.

For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible.  I'll come
back later to clean up any ugliness introduced here.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/vec4: Add wrapper functions for vec4_instruction::regs_read and ::regs_written.
Francisco Jerez [Thu, 1 Sep 2016 23:55:46 +0000 (16:55 -0700)]
i965/vec4: Add wrapper functions for vec4_instruction::regs_read and ::regs_written.

This is in preparation for dropping vec4_instruction::regs_read and
::regs_written in favor of more accurate alternatives expressed in
byte units.  The main reason these wrappers are useful is that a
number of optimization passes implement dataflow analysis with
register granularity, so these helpers will come in handy once we've
switched register offsets and sizes to the byte representation.  The
wrapper functions will also make sure that GRF misalignment (currently
neglected by most of the back-end) is taken into account correctly in
the calculation of regs_read and regs_written.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Add wrapper functions for fs_inst::regs_read and ::regs_written.
Francisco Jerez [Wed, 7 Sep 2016 23:59:35 +0000 (16:59 -0700)]
i965/fs: Add wrapper functions for fs_inst::regs_read and ::regs_written.

This is in preparation for dropping fs_inst::regs_read and
::regs_written in favor of more accurate alternatives expressed in
byte units.  The main reason these wrappers are useful is that a
number of optimization passes implement dataflow analysis with
register granularity, so these helpers will come in handy once we've
switched register offsets and sizes to the byte representation.  The
wrapper functions will also make sure that GRF misalignment (currently
neglected by most of the back-end) is taken into account correctly in
the calculation of regs_read and regs_written.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Replace fs_reg::subreg_offset with fs_reg::offset expressed in bytes.
Francisco Jerez [Thu, 1 Sep 2016 22:11:21 +0000 (15:11 -0700)]
i965/fs: Replace fs_reg::subreg_offset with fs_reg::offset expressed in bytes.

The fs_reg::subreg_offset and ::offset fields are now redundant, the
sub-GRF offset can just be added to the single ::offset field
expressed in byte units.  The current subreg_offset value can be
recovered by applying the following rule: Replace each rvalue
reference of subreg_offset like 'x = r.subreg_offset' with 'x =
r.offset % reg_unit', and each lvalue reference like 'r.subreg_offset
= x' with 'r.offset = ROUND_DOWN_TO(r.offset, reg_unit) + x'.

For the same reason as in the previous patches, this doesn't attempt
to be particularly clever about simplifying the result in the interest
of keeping the rather lengthy patch as obvious as possible.  I'll come
back later to clean up any ugliness introduced here.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/ir: Remove backend_reg::reg_offset.
Francisco Jerez [Thu, 1 Sep 2016 21:19:27 +0000 (14:19 -0700)]
i965/ir: Remove backend_reg::reg_offset.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/vec4: Replace dst/src_reg::reg_offset with dst/src_reg::offset expressed in...
Francisco Jerez [Thu, 1 Sep 2016 20:10:36 +0000 (13:10 -0700)]
i965/vec4: Replace dst/src_reg::reg_offset with dst/src_reg::offset expressed in bytes.

The dst/src_reg::offset field in byte units introduced in the previous
patch is a more straightforward alternative to an offset
representation split between ::reg_offset and ::subreg_offset fields.
The split representation makes it too easy to forget about one of the
offsets while dealing with the other, which has led to multiple FS
back-end bugs in the past.  To make the matter worse the unit
reg_offset was expressed in was rather inconsistent, for uniforms it
would be expressed in either 4B or 16B units depending on the
back-end, and for most other things it would be expressed in 32B
units.

This encodes reg_offset as a new offset field expressed consistently
in byte units.  Each rvalue reference of reg_offset in existing code
like 'x = r.reg_offset' is rewritten to 'x = r.offset / reg_unit', and
each lvalue reference like 'r.reg_offset = x' is rewritten to
'r.offset = r.offset % reg_unit + x * reg_unit'.

Because the change affects a lot of places and is rather non-trivial
to verify due to the inconsistent value of reg_unit, I've tried to
avoid making any additional changes other than applying the rewrite
rule above in order to keep the patch as simple as possible, sometimes
at the cost of introducing obvious stupidity (e.g. algebraic
expressions that could be simplified given some knowledge of the
context) -- I'll clean those up later on in a second pass.

v2: Fix division by the wrong reg_unit in the UNIFORM case of
    convert_to_hw_regs(). (Iago)

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoi965/fs: Replace fs_reg::reg_offset with fs_reg::offset expressed in bytes.
Francisco Jerez [Thu, 1 Sep 2016 19:42:20 +0000 (12:42 -0700)]
i965/fs: Replace fs_reg::reg_offset with fs_reg::offset expressed in bytes.

The fs_reg::offset field in byte units introduced in this patch is a
more straightforward alternative to the current register offset
representation split between fs_reg::reg_offset and ::subreg_offset.
The split representation makes it too easy to forget about one of the
offsets while dealing with the other, which has led to multiple
back-end bugs in the past.  To make the matter worse the unit
reg_offset was expressed in was rather inconsistent, for uniforms it
would be expressed in either 4B or 16B units depending on the
back-end, and for most other things it would be expressed in 32B
units.

This encodes reg_offset as a new offset field expressed consistently
in byte units.  Each rvalue reference of reg_offset in existing code
like 'x = r.reg_offset' is rewritten to 'x = r.offset / reg_unit', and
each lvalue reference like 'r.reg_offset = x' is rewritten to
'r.offset = r.offset % reg_unit + x * reg_unit'.

Because the change affects a lot of places and is rather non-trivial
to verify due to the inconsistent value of reg_unit, I've tried to
avoid making any additional changes other than applying the rewrite
rule above in order to keep the patch as simple as possible, sometimes
at the cost of introducing obvious stupidity (e.g. algebraic
expressions that could be simplified given some knowledge of the
context) -- I'll clean those up later on in a second pass.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
8 years agoglsl: grammar fix
Eero Tamminen [Wed, 14 Sep 2016 14:28:28 +0000 (15:28 +0100)]
glsl: grammar fix

Signed-off-by: Eero Tamminen <eero.t.tamminen@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agodocs: Mention AEP in release notes
Kenneth Graunke [Wed, 14 Sep 2016 19:43:16 +0000 (12:43 -0700)]
docs: Mention AEP in release notes

8 years agoi965: Enable ANDROID_extension_pack_es31a on Gen9+.
Kenneth Graunke [Wed, 14 Sep 2016 00:07:36 +0000 (17:07 -0700)]
i965: Enable ANDROID_extension_pack_es31a on Gen9+.

AEP requires ASTC, which is currently only enabled on Skylake and later.
(It may be possible to extend this to Cherryview/Braswell in the future,
but earlier hardware doesn't have ASTC support.)

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
8 years agonir: Report progress from nir_lower_phis_to_scalar.
Kenneth Graunke [Tue, 13 Sep 2016 22:14:28 +0000 (15:14 -0700)]
nir: Report progress from nir_lower_phis_to_scalar.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
8 years agonir: Report progress from nir_lower_alu_to_scalar.
Kenneth Graunke [Tue, 13 Sep 2016 22:14:28 +0000 (15:14 -0700)]
nir: Report progress from nir_lower_alu_to_scalar.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
8 years agonir: Call nir_metadata_preserve from nir_lower_alu_to_scalar().
Kenneth Graunke [Tue, 13 Sep 2016 22:17:29 +0000 (15:17 -0700)]
nir: Call nir_metadata_preserve from nir_lower_alu_to_scalar().

This is mandatory.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
8 years agonir/lower_tex: fix typo with sample_dim
Rob Clark [Thu, 8 Sep 2016 19:49:49 +0000 (15:49 -0400)]
nir/lower_tex: fix typo with sample_dim

Numeric 2 is actually GLSL_SAMPLER_DIM_3D, which I don't think is what
was intended.

Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agonir: move tex_instr_remove_src
Rob Clark [Thu, 8 Sep 2016 18:07:06 +0000 (14:07 -0400)]
nir: move tex_instr_remove_src

I want to re-use this in a different pass, so move to nir.h

Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agonir/lower_tex: remove tex_instr_find_src()
Rob Clark [Thu, 8 Sep 2016 18:05:39 +0000 (14:05 -0400)]
nir/lower_tex: remove tex_instr_find_src()

Turns out it already exists.. so don't duplicate it.

Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoegl: Add storage for EGL_KHR_debug's state to EGL objects
Kyle Brenneman [Mon, 12 Sep 2016 20:15:10 +0000 (16:15 -0400)]
egl: Add storage for EGL_KHR_debug's state to EGL objects

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Factor out _eglGetSyncAttribCommon
Kyle Brenneman [Mon, 12 Sep 2016 21:46:04 +0000 (17:46 -0400)]
egl: Factor out _eglGetSyncAttribCommon

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Factor out _eglWaitSyncCommon
Kyle Brenneman [Mon, 12 Sep 2016 21:44:36 +0000 (17:44 -0400)]
egl: Factor out _eglWaitSyncCommon

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Lock the display in _eglCreateSync's callers
Kyle Brenneman [Mon, 12 Sep 2016 21:40:29 +0000 (17:40 -0400)]
egl: Lock the display in _eglCreateSync's callers

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Factor out _eglCreateImageCommon (v2)
Kyle Brenneman [Mon, 12 Sep 2016 21:38:13 +0000 (17:38 -0400)]
egl: Factor out _eglCreateImageCommon (v2)

v2:
- Pass disp to RETURN_EGL_ERROR so we unlock the display

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Factor out _eglWaitClientCommon
Kyle Brenneman [Mon, 12 Sep 2016 21:35:22 +0000 (17:35 -0400)]
egl: Factor out _eglWaitClientCommon

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Use _eglCreatePixmapSurfaceCommon consistently
Kyle Brenneman [Mon, 12 Sep 2016 21:25:56 +0000 (17:25 -0400)]
egl: Use _eglCreatePixmapSurfaceCommon consistently

This moves the native pixmap fixup to a helper function so we don't
repeat ourselves.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Use _eglCreateWindowSurfaceCommon consistently
Kyle Brenneman [Mon, 12 Sep 2016 21:04:38 +0000 (17:04 -0400)]
egl: Use _eglCreateWindowSurfaceCommon consistently

This moves the native window fixup to a helper function so we don't
repeat ourselves.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Factor out _eglGetPlatformDisplayCommon
Kyle Brenneman [Mon, 12 Sep 2016 20:42:56 +0000 (16:42 -0400)]
egl: Factor out _eglGetPlatformDisplayCommon

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Fix typo
Kyle Brenneman [Mon, 12 Sep 2016 21:12:52 +0000 (17:12 -0400)]
egl: Fix typo

Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agoegl: Tear down images and syncs at eglTerminate
Adam Jackson [Fri, 9 Sep 2016 17:45:09 +0000 (13:45 -0400)]
egl: Tear down images and syncs at eglTerminate

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Signed-off-by: Adam Jackson <ajax@redhat.com>
8 years agoegl: Update eglext.h (v2)
Kyle Brenneman [Wed, 6 Jul 2016 16:33:42 +0000 (10:33 -0600)]
egl: Update eglext.h (v2)

Updated eglext.h to revision 33111 from the Khronos repository.

v2:
- Don't (re)move extension includes from eglext.h (Emil Velikov)
- Bump to revision 33111 (Adam Jackson)

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
8 years agoconfigure.ac: fix the name of the Wayland Scanner pc file
Brendan King [Tue, 13 Sep 2016 16:31:05 +0000 (17:31 +0100)]
configure.ac: fix the name of the Wayland Scanner pc file

The Wayland Scanner pkg-config file is called wayland-scanner.pc.

Fixes: 153539bd9d4445b50411 ("configure: rework wayland_scanner
       handling (fix make distcheck)")

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Tested-by: Eric Engestrom <eric.engestrom@imgtec.com>
Signed-off-by: Brendan King <Brendan.King@imgtec.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agogbm: remove left-over array
Eric Engestrom [Tue, 13 Sep 2016 16:32:39 +0000 (17:32 +0100)]
gbm: remove left-over array

e7c8c85785b3a8f29e3f ("gbm: Removed unused function.") forgot to remove
the global array used only by that function.

Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
8 years agogallium: fix return value check
Martina Kollarova [Thu, 8 Sep 2016 12:12:42 +0000 (15:12 +0300)]
gallium: fix return value check

A possible error (-1) was being lost because it was first converted to an
unsigned int and only then checked.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Martina Kollarova <martina.kollarova@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
8 years agoradeonsi: reload PS inputs with direct indexing at each use (v2)
Marek Olšák [Tue, 13 Sep 2016 15:33:23 +0000 (17:33 +0200)]
radeonsi: reload PS inputs with direct indexing at each use (v2)

The LLVM compiler can CSE interp intrinsics thanks to
LLVMReadNoneAttribute.

26011 shaders in 14651 tests
Totals:
SGPRS: 1146340 -> 1132676 (-1.19 %)
VGPRS: 727371 -> 711730 (-2.15 %)
Spilled SGPRs: 2218 -> 2078 (-6.31 %)
Spilled VGPRs: 369 -> 369 (0.00 %)
Scratch VGPRs: 1344 -> 1344 (0.00 %) dwords per thread
Code Size: 35841268 -> 36009732 (0.47 %) bytes
LDS: 767 -> 767 (0.00 %) blocks
Max Waves: 222559 -> 224779 (1.00 %)
Wait states: 0 -> 0 (0.00 %)

v2: don't call load_input for fragment shaders in emit_declaration

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: get rid of constant buffer preloading
Marek Olšák [Tue, 13 Sep 2016 12:30:50 +0000 (14:30 +0200)]
radeonsi: get rid of constant buffer preloading

26011 shaders in 14651 tests
Totals:
SGPRS: 1152636 -> 1146340 (-0.55 %)
VGPRS: 728198 -> 727371 (-0.11 %)
Spilled SGPRs: 3776 -> 2218 (-41.26 %)
Spilled VGPRs: 369 -> 369 (0.00 %)
Scratch VGPRs: 1344 -> 1344 (0.00 %) dwords per thread
Code Size: 35835152 -> 35841268 (0.02 %) bytes
LDS: 767 -> 767 (0.00 %) blocks
Max Waves: 222372 -> 222559 (0.08 %)
Wait states: 0 -> 0 (0.00 %)

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
8 years agoradeonsi: get rid of img/buf/sampler descriptor preloading (v2)
Marek Olšák [Tue, 13 Sep 2016 12:25:44 +0000 (14:25 +0200)]
radeonsi: get rid of img/buf/sampler descriptor preloading (v2)

26011 shaders in 14651 tests
Totals:
SGPRS: 1251920 -> 1152636 (-7.93 %)
VGPRS: 728421 -> 728198 (-0.03 %)
Spilled SGPRs: 16644 -> 3776 (-77.31 %)
Spilled VGPRs: 369 -> 369 (0.00 %)
Scratch VGPRs: 1344 -> 1344 (0.00 %) dwords per thread
Code Size: 36001064 -> 35835152 (-0.46 %) bytes
LDS: 767 -> 767 (0.00 %) blocks
Max Waves: 222221 -> 222372 (0.07 %)
Wait states: 0 -> 0 (0.00 %)

v2: merge codepaths where possible

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: rename get_sampler_desc -> load_sampler_desc
Marek Olšák [Tue, 13 Sep 2016 11:37:16 +0000 (13:37 +0200)]
radeonsi: rename get_sampler_desc -> load_sampler_desc

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
8 years agoradeonsi: cosmetic changes in si_shader.c
Marek Olšák [Tue, 13 Sep 2016 11:28:09 +0000 (13:28 +0200)]
radeonsi: cosmetic changes in si_shader.c

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
8 years agoradeonsi: load streamout buffer descriptors before use (v2)
Marek Olšák [Tue, 13 Sep 2016 11:12:33 +0000 (13:12 +0200)]
radeonsi: load streamout buffer descriptors before use (v2)

v2: inline the code and remove the conditional that's a no-op now

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agovc4: Implement job shuffling
Eric Anholt [Thu, 8 Sep 2016 19:56:11 +0000 (12:56 -0700)]
vc4: Implement job shuffling

Track rendering to each FBO independently and flush rendering only when
necessary.  This lets us avoid the overhead of storing and loading the
frame when an application momentarily switches to rendering to some other
texture in order to continue rendering the main scene.

Improves glmark -b desktop:effect=shadow:windows=4 by 27%
Improves glmark -b
    desktop:blur-radius=5:effect=blur:passes=1:separable=true:windows=4
    by 17%

While I haven't tested other apps, this should help X rendering a lot, and
I've heard GLBenchmark needed it too.

8 years agovc4: Handle resolve skipping at job submit time.
Eric Anholt [Thu, 8 Sep 2016 20:02:22 +0000 (13:02 -0700)]
vc4: Handle resolve skipping at job submit time.

This is done in vc4_flush currently, but I'm going to make the job always
track the surfaces it might be rendering to instead of putting in the
destinations at flush time.

8 years agovc4: Move the render job state into a separate structure.
Eric Anholt [Mon, 6 Apr 2015 20:17:58 +0000 (13:17 -0700)]
vc4: Move the render job state into a separate structure.

This is a preparation step for having multiple jobs being queued up at the
same time.

8 years agovc4: Always unref the current job surfaces at job reset time.
Eric Anholt [Wed, 7 Sep 2016 19:40:39 +0000 (12:40 -0700)]
vc4: Always unref the current job surfaces at job reset time.

Drops some tricky logic in vc4_flush() trying to update the pointers, and
fixes a broken lack of unref for MSAA surfaces at context destroy time.

8 years agovc4: Move job-submit skip cases to vc4_job_submit().
Eric Anholt [Thu, 8 Sep 2016 21:03:29 +0000 (14:03 -0700)]
vc4: Move job-submit skip cases to vc4_job_submit().

For calling job_submit() directly, I need the skipping here.

8 years agovc4: Move bin CL trailer to job_submit() time.
Eric Anholt [Thu, 8 Sep 2016 21:01:15 +0000 (14:01 -0700)]
vc4: Move bin CL trailer to job_submit() time.

To implement job shuffling, I want to be able to call submit() on specific
jobs, turning vc4_flush() into the context's flush-all-jobs hook.

8 years agovc4: Simplify the DISCARD_RANGE handling
Eric Anholt [Fri, 9 Sep 2016 05:40:44 +0000 (22:40 -0700)]
vc4: Simplify the DISCARD_RANGE handling

It's really just an upgrade to attempting WHOLE_RESOURCE.  Pulling the
logic out caught two bugs in it: We would try to do so on cubemaps (even
though we're only mapping 1 of the 6 slices), and we would break
persistent coherent mappings by trying to reallocate when we shouldn't.

8 years agovc4: Fix incorrect clearing of Z/stencil when cleared separately.
Eric Anholt [Fri, 9 Sep 2016 23:26:02 +0000 (16:26 -0700)]
vc4: Fix incorrect clearing of Z/stencil when cleared separately.

The clear of Z or stencil will end up clearing the other as well, instead
of masking.  There's no way around this that I know of, so if we are
clearing just one then we need to draw a quad.

Fixes a regression in the job-shuffling code, where the clear values move
to the job and don't just have the last clear's value laying around when
you do glClear(DEPTH) and then glClear(STENCIL) separately
(ext_framebuffer_multisample-clear 4 depth)).

This causes regressions in ext_framebuffer_multisample/multisample-blit
depth and ext_framebuffer_multisample/no-color depth, but these were
formerly false positives due to the reference image also being black.  Now
the reference and test images are both being drawn, and it looks like
there's an incorrect resolve of depth during blitting to an MSAA FBO.

8 years agoglsl: add core plumbing for GL_ANDROID_extension_pack_es31a
Ilia Mirkin [Sun, 28 Aug 2016 19:40:19 +0000 (15:40 -0400)]
glsl: add core plumbing for GL_ANDROID_extension_pack_es31a

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agomesa: introduce glPrimitiveBoundingBoxARB entrypoint
Ilia Mirkin [Sun, 28 Aug 2016 23:53:02 +0000 (19:53 -0400)]
mesa: introduce glPrimitiveBoundingBoxARB entrypoint

This requires a bit of rejiggering, since normally ES entrypoints alias
core ones, not vice-versa.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agomesa: add a GLES3.2 enums section, and expose new MS line width params
Ilia Mirkin [Sun, 28 Aug 2016 23:51:45 +0000 (19:51 -0400)]
mesa: add a GLES3.2 enums section, and expose new MS line width params

This also exposes them for ARB_ES3_2_compatibility.

While both specs refer to the new MS line width parameters being
separate from the existing AA line widths, reality begs to differ. It's
the same on all hardware currently supported by mesa. Should hardware
come along that wants these to be different, they're easy enough to
separate out.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoaubinator: Remove bogus "end" parameter in gen_disasm_disassemble()
Sirisha Gandikota [Tue, 13 Sep 2016 23:19:31 +0000 (16:19 -0700)]
aubinator: Remove bogus "end" parameter in gen_disasm_disassemble()

Earlier, the loop pretends to loop over instructions from "start" to "end",
but the callers always pass 8192 for end, which is some huge bogus
value. The real loop termination condition is send-with-EOT or 0. (Ken)

Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoaubinator: Make gen_disasm_disassemble handle split sends
Sirisha Gandikota [Tue, 13 Sep 2016 23:19:30 +0000 (16:19 -0700)]
aubinator: Make gen_disasm_disassemble handle split sends

Skylake adds new SENDS and SENDSC opcodes, which should be
handled in the send-with-EOT check. Make an is_send() helper
that checks if the opcode is SEND/SENDC/SENDS/SENDSC (Ken)

v2: Make is_send() much more crispier, Mix declaration and
code to make the code compact (Ken)

Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoaubinator: Simplify print_dword_val() method
Sirisha Gandikota [Tue, 13 Sep 2016 23:19:29 +0000 (16:19 -0700)]
aubinator: Simplify print_dword_val() method

Remove the float/dword union and use the iter->p[f->start / 32]
directly as printf formatter %08x expects uint32_t (Ken)

v2: Make the cleanup much more crispier (Ken)

Signed-off-by: Sirisha Gandikota <Sirisha.Gandikota@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoanv/image: Set correct base_array_layer and array_len for storage images
Jason Ekstrand [Tue, 13 Sep 2016 21:02:32 +0000 (14:02 -0700)]
anv/image: Set correct base_array_layer and array_len for storage images

Since Vulkan doesn't allow single-slice 3D storage images, we need to just
set the base_array_layer and array_len to the full size of the 3-D LOD.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoRevert "intel/isl: Ignore base_array_layer and array_len for 3D storage..."
Jason Ekstrand [Tue, 13 Sep 2016 19:55:25 +0000 (12:55 -0700)]
Revert "intel/isl: Ignore base_array_layer and array_len for 3D storage..."

This reverts commit 3943888c94beca69e575b8d3d1ec7a6cbf474ee4.  It turns out
that commit was pretty-much bogus since it breaks binding a 3-D texture as a
2-D storage image.  The correct fix for the Vulkan CTS tests needs to be in
the Vulkan driver itself rather than ISL.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoanv: Use blorp for doing MSAA resolves
Jason Ekstrand [Wed, 31 Aug 2016 00:49:56 +0000 (17:49 -0700)]
anv: Use blorp for doing MSAA resolves

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoanv: Use blorp for ClearColorImage
Jason Ekstrand [Tue, 30 Aug 2016 23:56:25 +0000 (16:56 -0700)]
anv: Use blorp for ClearColorImage

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoanv: Delete meta_blit2d
Jason Ekstrand [Tue, 30 Aug 2016 23:23:59 +0000 (16:23 -0700)]
anv: Delete meta_blit2d

Everything that we were once using the blit2d framework for is now done
with blorp.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoanv/blorp: Add a gcd_pow2_u64 helper and use it for buffer alignments
Jason Ekstrand [Sat, 10 Sep 2016 02:21:18 +0000 (19:21 -0700)]
anv/blorp: Add a gcd_pow2_u64 helper and use it for buffer alignments

This is a lot cleaner and easier to read than the old piles of if
statements.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoanv: Use blorp for CopyBuffer and UpdateBuffer
Jason Ekstrand [Tue, 30 Aug 2016 22:43:46 +0000 (15:43 -0700)]
anv: Use blorp for CopyBuffer and UpdateBuffer

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoanv: Use blorp for CopyImage
Jason Ekstrand [Sun, 28 Aug 2016 04:05:13 +0000 (21:05 -0700)]
anv: Use blorp for CopyImage

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoanv: Use blorp for CopyBufferToImage
Jason Ekstrand [Sat, 27 Aug 2016 19:57:01 +0000 (12:57 -0700)]
anv: Use blorp for CopyBufferToImage

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoanv: Use blorp for CopyImageToBuffer
Jason Ekstrand [Wed, 24 Aug 2016 03:19:57 +0000 (20:19 -0700)]
anv: Use blorp for CopyImageToBuffer

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
8 years agoanv: Use blorp to implement VkBlitImage
Jason Ekstrand [Tue, 23 Aug 2016 05:33:06 +0000 (22:33 -0700)]
anv: Use blorp to implement VkBlitImage

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoanv: Make image_get_surface_for_aspect_mask const
Jason Ekstrand [Wed, 31 Aug 2016 00:49:32 +0000 (17:49 -0700)]
anv: Make image_get_surface_for_aspect_mask const

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
8 years agoanv: Add initial blorp support
Jason Ekstrand [Tue, 23 Aug 2016 04:37:28 +0000 (21:37 -0700)]
anv: Add initial blorp support

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agointel/anv: Use #defines for all __gen_ helpers
Jason Ekstrand [Wed, 24 Aug 2016 00:23:29 +0000 (17:23 -0700)]
intel/anv: Use #defines for all __gen_ helpers

This allows us to #undef them later if we don't want them to persist

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoanv: Generalize emit_urb_setup
Jason Ekstrand [Tue, 23 Aug 2016 02:08:33 +0000 (19:08 -0700)]
anv: Generalize emit_urb_setup

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoanv/pipeline: Roll compute_urb_partition into emit_urb_setup
Jason Ekstrand [Sat, 20 Aug 2016 13:11:41 +0000 (06:11 -0700)]
anv/pipeline: Roll compute_urb_partition into emit_urb_setup

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agointel/blorp: Use #defines for all __gen_ helpers
Jason Ekstrand [Wed, 24 Aug 2016 00:20:34 +0000 (17:20 -0700)]
intel/blorp: Use #defines for all __gen_ helpers

This allows us to #undef them later if we don't want them to persist

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agointel/isl: Divide QPitch by 2 for 3-D stencil textures on SKL+
Jason Ekstrand [Thu, 8 Sep 2016 17:26:15 +0000 (10:26 -0700)]
intel/isl: Divide QPitch by 2 for 3-D stencil textures on SKL+

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
8 years agoisl/state: Don't set QPitch for GEN4_3D surfaces
Jason Ekstrand [Thu, 8 Sep 2016 17:23:47 +0000 (10:23 -0700)]
isl/state: Don't set QPitch for GEN4_3D surfaces

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
8 years agointel/blorp: Rework alloc_binding_table
Jason Ekstrand [Wed, 24 Aug 2016 03:51:26 +0000 (20:51 -0700)]
intel/blorp: Rework alloc_binding_table

The original blorp_alloc_binding_table helper was supposed to return the
binding table offset and map along with the surface state maps.  This isn't
quite what we want, however.  What we really want is the binding table
offsets, surface state offsets, and surface state maps.  In the GL driver,
the binding table map *is* an array of surface state offsets.  However, in
Vulkan, this isn't quite true as the entries in the binding table are
surface state offsets combined with another binding table block offset.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agotgsi/scan: don't set interp flags for inputs only used by INTERP instructions
Marek Olšák [Thu, 8 Sep 2016 17:12:00 +0000 (19:12 +0200)]
tgsi/scan: don't set interp flags for inputs only used by INTERP instructions

radeonsi depends on the interp flags a little bit too much.

This fixes 9 randomly failing tests:
  GL45-CTS.shader_multisample_interpolation.render.interpolate_at_centroid.*

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: fix FP64 UBO loads with indirect uniform block indexing
Marek Olšák [Sun, 11 Sep 2016 23:38:53 +0000 (01:38 +0200)]
radeonsi: fix FP64 UBO loads with indirect uniform block indexing

No known tests.

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agowinsys/amdgpu: don't assume GTT if the VRAM flag isn't set
Marek Olšák [Sun, 11 Sep 2016 19:49:24 +0000 (21:49 +0200)]
winsys/amdgpu: don't assume GTT if the VRAM flag isn't set

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: clean up CP DMA emit code
Marek Olšák [Sun, 11 Sep 2016 15:57:31 +0000 (17:57 +0200)]
radeonsi: clean up CP DMA emit code

Unify the clear and copy paths, clean up the definitions.
It looks more like a rework. It's a preparation for GDS support,
which might or might not come.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: print the IB and buffer list in VM fault reports
Marek Olšák [Mon, 12 Sep 2016 18:23:55 +0000 (20:23 +0200)]
radeonsi: print the IB and buffer list in VM fault reports

This is a fallout from reworking the debug flags.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: add sampler view BOs to the BO list last
Marek Olšák [Mon, 12 Sep 2016 10:49:34 +0000 (12:49 +0200)]
radeonsi: add sampler view BOs to the BO list last

If si_sampler_view_add_buffer ends up flushing, then the code
in begin_new_cs would previously have added the buffer(s) for
whatever was previously bound to that slot. Now it would add only
the new buffer.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeonsi: export SampleMask from pixel shaders at full rate
Marek Olšák [Fri, 9 Sep 2016 23:21:11 +0000 (01:21 +0200)]
radeonsi: export SampleMask from pixel shaders at full rate

Heaven and Valley write gl_SampleMask and not Z.
Use 16_ABGR instead of 32_ABGR if Z isn't written.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agogallium/radeon: set new r600_resource fields correctly in other places too
Marek Olšák [Mon, 12 Sep 2016 10:47:13 +0000 (12:47 +0200)]
gallium/radeon: set new r600_resource fields correctly in other places too

This was missed in:

    commit 0d2e43fcb1198a6e67c85feadb1ca8c360ddc284
    Author: Marek Olšák <marek.olsak@amd.com>
    Date:   Thu Aug 18 16:30:00 2016 +0200

        gallium/radeon: derive buffer placement and flags only at initialization

Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoddebug: dump shader buffers and images
Marek Olšák [Sat, 10 Sep 2016 08:16:54 +0000 (10:16 +0200)]
ddebug: dump shader buffers and images

this was unimplemented

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoddebug: fix a crash in resource_get_handle
Marek Olšák [Sat, 10 Sep 2016 07:50:42 +0000 (09:50 +0200)]
ddebug: fix a crash in resource_get_handle

broken recently

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
8 years agoradeon: Don't check DCC on pipe buffers
Jan Vesely [Tue, 13 Sep 2016 00:47:37 +0000 (20:47 -0400)]
radeon: Don't check DCC on pipe buffers

Fixes segfaults in EG compute since:
commit 21de3be8e62b2b093569a99550e6356ed2f106b4
radeonsi: fix texture format reinterpretation with DCC

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agovl/util: Fix YV12/I420 convert to NV12 U/V reversal
Andy Furniss [Tue, 13 Sep 2016 17:52:31 +0000 (13:52 -0400)]
vl/util: Fix YV12/I420 convert to NV12 U/V reversal

Fix VAAPI YV12/I420 convert to NV12 U/V reversal.
Input order is YVU when this is called.

Signed-off-by: Andy Furniss <adf.lists@gmail.com>
Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
8 years agoanv/allocator: Use VG_NOACCESS_WRITE in anv_bo_pool_free
Jason Ekstrand [Thu, 8 Sep 2016 04:33:48 +0000 (21:33 -0700)]
anv/allocator: Use VG_NOACCESS_WRITE in anv_bo_pool_free

Previously, we were relying on the fact that VALGRIND_MEMPOOL_FREE came
later on in the function to prevent "link->bo = bo" from causing an invalid
write.  However, in the case where the size requested by the user is very
small (less than sizeof(struct anv_bo)), this isn't sufficient.  Instead,
we should call VALGRIND_MEMPOOL_FREE early and then use VG_NOACCESS_WRITE.
We do, however, have to call VALGRIND_MEMPOOL_FREE after reading bo_in
because it may be stored in the bo itself.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agointel/isl: Ignore base_array_layer and array_len for 3D storage surfaces
Jason Ekstrand [Tue, 13 Sep 2016 05:31:00 +0000 (22:31 -0700)]
intel/isl: Ignore base_array_layer and array_len for 3D storage surfaces

The time we want to restrict the Z range of a 3-D surface is when rendering
to it.  For storage surfaces, we always want he full range.  However, we
still need to set MinimumArrayElement and RenderTargetViewExtent to
sensible values so we'll just set them to the reasonable defaults we used
before we started respecting the base_array_layer and array_len.

This fixes a bunch of Vulkan CTS regressions caused by 48f195d7c6483ed.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97790
Reviewed-by: Chad Versace <chadversary@chromium.org>
8 years agoappveyor: Update winflexbison download URL.
Jose Fonseca [Tue, 13 Sep 2016 16:53:30 +0000 (17:53 +0100)]
appveyor: Update winflexbison download URL.

This particular version got moved into a `old_versions` subdirectory.

8 years agoi965: Use blorp_copy for all copy_image operations on gen6+
Jason Ekstrand [Fri, 9 Sep 2016 04:34:16 +0000 (21:34 -0700)]
i965: Use blorp_copy for all copy_image operations on gen6+

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
8 years agoi965/blorp: Add a copy_miptrees helper
Jason Ekstrand [Fri, 9 Sep 2016 04:27:01 +0000 (21:27 -0700)]
i965/blorp: Add a copy_miptrees helper

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chad Versace <chadversary@chromium.org>
8 years agointel/isl: Add support for RGB formats in X and Y-tiled memory
Jason Ekstrand [Fri, 9 Sep 2016 05:18:47 +0000 (22:18 -0700)]
intel/isl: Add support for RGB formats in X and Y-tiled memory

Normally, using a non-linear tiling format helps improve cache locality by
ensuring that neighboring pixels are usually close-by in memory.  For RGB
formats, this still sort-of holds, but it can also lead to rather terrible
memory access patterns where a single RGB pixel value crosses a tile
boundary and gets split into two pieces in different 4K pages.  It also
makes for some rather awkward calculations because your tile size is no
longer an even multiple of surface element size.  For these reasons, we
chose to simply never create tiled RGB images in the Vulkan driver.

The GL driver, however, is not so kind so we need to support it somehow.  I
briefly toyed with a couple of different schemes but this is the best one I
could come up with.  The fundamental problem is that a tile no longer
contains an integer number of surface elements.  I briefly considered a
couple other options but found them wanting:

 1) Using floats for the logical tile size.  This leads to potential
    rounding error problems.

 2) When presented with a RGB format, just make the tile 3-times as wide.
    This isn't so nice because now our tiles are no longer power-of-two
    size.  Also, it can force the row_pitch to be larger than needed which,
    while not strictly a problem for ISL, causes incompatibility problems
    with the way the GL driver chooses surface pitches.

The chosen method requires that you pay attention and not just assume that
your tile_info is in the units you think it is.  However, it's nice because
it provides a nice "these are the units" declaration in isl_tile_info
itself.  Previously, the tile_info wasn't usable as a stand-alone structure
because you had to also know the format.  It also forces figuring out how
to deal with inconsistencies between tiling and format back to the caller
which is good because the two different consumers of isl_tile_info really
want to deal with it differently:  Computation of the surface size wants
the fewest number of horizontal tiles possible while get_intratile_offset
is far more concerned with things aligning nicely.

Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Chad Versace <chadversary@chromium.org>