mesa.git
7 years agoradv: Use a shader for occlusion CmdCopyQueryPoolResults.
Bas Nieuwenhuizen [Sun, 9 Apr 2017 20:35:32 +0000 (22:35 +0200)]
radv: Use a shader for occlusion CmdCopyQueryPoolResults.

Use the new occlusion query copy shader.

We don't use the shader for the waiting as a polling loop ineracts badly
with having caching enabled. I noticed on my GPU (Tonga) that the values
are written out in order, so I just use a WAIT_REG_MEM on the last value.

If it turns out other chips don't do that we may need to look a bit more
into this. Having 8 WAIT_REG_MEM packets per query doesn't sound ideal.

This also restricts the availability word in the pool to timestamp queries
only, as occlusion queries don't use it, and pipeline statistic queries
likely won't either.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
7 years agoradv: Add occlusion query shader.
Bas Nieuwenhuizen [Sun, 26 Feb 2017 17:21:01 +0000 (18:21 +0100)]
radv: Add occlusion query shader.

Adds a shader for writing occlusion query results to a buffer, as the
CP packet isn't support on SI or secondary buffers, and doesn't handle
the availability bit (or partial results) nor truncation to 32-bit.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
7 years agoi965: Fix wonky indentation left by brw_bo_alloc_tiled rename.
Kenneth Graunke [Tue, 11 Apr 2017 06:23:13 +0000 (23:23 -0700)]
i965: Fix wonky indentation left by brw_bo_alloc_tiled rename.

7 years agonouveau: when mapping a persistent buffer, synchronize on former xfers
Ilia Mirkin [Sat, 8 Apr 2017 22:31:35 +0000 (18:31 -0400)]
nouveau: when mapping a persistent buffer, synchronize on former xfers

If the buffer is being used, we should wait for those uses to be
complete before returning the map.

Fixes: GL45-CTS.direct_state_access.buffers_functional
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
7 years agonvc0: increase texture buffer object alignment to 256 for pre-GM107
Ilia Mirkin [Sat, 8 Apr 2017 18:56:16 +0000 (14:56 -0400)]
nvc0: increase texture buffer object alignment to 256 for pre-GM107

We currently don't pass the low byte of the address via the surface
info, so in order to work with images, these have to implicitly be
aligned to 256. The proprietary driver also doesn't go out of its way to
provide lower alignment.

Fixes GL45-CTS.texture_buffer.texture_buffer_texture_buffer_range

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
7 years agomesa: fix typo and add assert() to _mesa_attach_renderbuffer_without_ref()
Timothy Arceri [Mon, 10 Apr 2017 23:57:45 +0000 (09:57 +1000)]
mesa: fix typo and add assert() to _mesa_attach_renderbuffer_without_ref()

This function should only be used with a "freshly created" renderbuffer
so assert RefCount is 1.

7 years agoi965/drm: Add stall warnings when mapping or waiting on BOs.
Kenneth Graunke [Mon, 10 Apr 2017 06:14:56 +0000 (23:14 -0700)]
i965/drm: Add stall warnings when mapping or waiting on BOs.

This restores the performance warnings removed in:

    i965: Drop brw_bo_map[_gtt] wrappers which issue perf warnings.

but adds them for nearly all BO mapping, and also for wait_rendering.

Because we add this to the core bufmgr, we automatically get stall
warnings in all callers, unlike before where only a few callsites used
the wrappers that gave stall warnings.

We also do it a bit differently: we simply measure how long set_domain
takes (the part that stalls), and complain if it's more than 0.01 ms.
We don't bother calling brw_bo_busy(), and we don't measure the mmap
time (which doesn't stall).  This should be more accurate.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
7 years agoi965/drm: Make a set_domain() helper function.
Kenneth Graunke [Mon, 10 Apr 2017 05:58:57 +0000 (22:58 -0700)]
i965/drm: Make a set_domain() helper function.

Less boilerplate.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
7 years agoi965/batch: Ensure we use a consistent offset in relocs
Daniel Vetter [Thu, 6 Apr 2017 07:13:47 +0000 (09:13 +0200)]
i965/batch: Ensure we use a consistent offset in relocs

In theory gcc is free to re-load them, and if a concurrent
execbuf races and updates bo->offset64 then we have a problem:
execbuffer api requires that the ->presumed_offset and the one
we used for the reloc matches. It does not require that the value
is sensible, which means no locks needed, just a consistent load.

Ken said his next series will nuke this, so just hand-roll the
kernel's READ_ONCE idea inline.

FIXME: Most callers of brw_emit_reloc recompute the relocation
themselves, which means this doesn't really fix the race. But the long
term plan is to move to per-context relocation handling, which will
fix this all properly. So leave this for now as just a reminder.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/bufmgr: Garbage-collect vma cache/pruning
Daniel Vetter [Thu, 6 Apr 2017 06:48:08 +0000 (08:48 +0200)]
i965/bufmgr: Garbage-collect vma cache/pruning

This was done because the kernel has 1 global address space, shared
with all render clients, for gtt mmap offsets, and that address space
was only 32bit on 32bit kernels.

This was fixed  in

commit 440fd5283a87345cdd4237bdf45fb01130ea0056
Author: Thierry Reding <treding@nvidia.com>
Date:   Fri Jan 23 09:05:06 2015 +0100

    drm/mm: Support 4 GiB and larger ranges

which shipped in 4.0. Of course you still want to limit the bo cache
to a reasonable size on 32bit apps to avoid ENOMEM, but that's better
solved by tuning the cache a bit. On 64bit, this was never an issue.

On top, mesa never set this, so it's all dead code. Collect an trash it.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/bufmgr: Remove some reuse functions
Daniel Vetter [Thu, 6 Apr 2017 06:28:51 +0000 (08:28 +0200)]
i965/bufmgr: Remove some reuse functions

is_reusable was needed by uxa because it couldn't keep track of its
scanout buffers and used this as a proxy. Disabling reuse is a silly
idea, we set this once at start. Remove both.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/bufmgr: remove start_gtt_access
Daniel Vetter [Thu, 6 Apr 2017 06:27:47 +0000 (08:27 +0200)]
i965/bufmgr: remove start_gtt_access

Iirc this was used by uxa for persistent mmpas of the frontbuffer. For
mesa all the set_domain stuff needed before a synchronized mmap is handled
within the bufmgr, so no reason ever to call this.

Inline the implementation into its only internal user.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/bufmgr: Delete set_tiling
Daniel Vetter [Thu, 6 Apr 2017 06:25:50 +0000 (08:25 +0200)]
i965/bufmgr: Delete set_tiling

Entirely unused, and really shouldn't be used. The alloc functions already
take care of this. And even in a future where we're not going to
h/v-align tiled buffers in the bufmgr, but only in isl, I think we
still want to adjust the tiling mode in the bufmgr, since that ties in
closely to mmaps and stuff like that.

get_tiling is still needed for the import paths (until we have modifiers
everywhere).

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/bufmgr: Delete alloc_for_render
Daniel Vetter [Thu, 6 Apr 2017 06:23:28 +0000 (08:23 +0200)]
i965/bufmgr: Delete alloc_for_render

Entirely unused, mesa instead used the BO_ALLOC_FOR_RENDER flag.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
7 years agoi965/drm: Use list_for_each_entry_safe in a couple of cases.
Kenneth Graunke [Wed, 5 Apr 2017 21:10:36 +0000 (14:10 -0700)]
i965/drm: Use list_for_each_entry_safe in a couple of cases.

Suggested by Chris Wilson.  A tiny bit simpler.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
7 years agoi965/drm: Rename intel_bufmgr_gem.c to brw_bufmgr.c.
Kenneth Graunke [Tue, 4 Apr 2017 04:01:51 +0000 (21:01 -0700)]
i965/drm: Rename intel_bufmgr_gem.c to brw_bufmgr.c.

Matches the class name and the header file name.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Reindent intel_bufmgr_gem.c and brw_bufmgr.h.
Kenneth Graunke [Tue, 4 Apr 2017 03:54:16 +0000 (20:54 -0700)]
i965/drm: Reindent intel_bufmgr_gem.c and brw_bufmgr.h.

indent -i3 -nut -br -brs -npcs -ce --no-tabs -Tuint32_t -Tuint64_t
plus some manual fixes because those aren't quite the right settings.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Rename drm_bacon_bo to brw_bo.
Kenneth Graunke [Tue, 4 Apr 2017 03:13:08 +0000 (20:13 -0700)]
i965/drm: Rename drm_bacon_bo to brw_bo.

The bacon is all gone.

This renames both the class and the related functions.  We're about to
run indent on the bufmgr code, so no need to worry about fixing bad
indentation.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Drop brw_bo_map[_gtt] wrappers which issue perf warnings.
Kenneth Graunke [Tue, 4 Apr 2017 03:22:59 +0000 (20:22 -0700)]
i965: Drop brw_bo_map[_gtt] wrappers which issue perf warnings.

The stupid reason for eliminating these functions is that I'm about
to rename drm_bacon_bo_map() to brw_bo_map(), which makes the real
function have the short name, rather than the wrapper.

I'm also planning on reworking our mapping code soon, so we use WC
mappings and proper unsynchronized mappings on non-LLC platforms.
It will be easier to do that without thinking about the stall
warnings and wrappers.

My eventual hope is to put the performance warnings in the BO map
function itself, so all callers gain the warning.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Rename drm_bacon_reg_read() to brw_reg_read().
Kenneth Graunke [Tue, 4 Apr 2017 03:14:11 +0000 (20:14 -0700)]
i965/drm: Rename drm_bacon_reg_read() to brw_reg_read().

Less bacon.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Rename drm_bacon_bufmgr to struct brw_bufmgr.
Kenneth Graunke [Tue, 4 Apr 2017 01:10:23 +0000 (18:10 -0700)]
i965/drm: Rename drm_bacon_bufmgr to struct brw_bufmgr.

Also stop using typedefs, per Mesa coding style.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Just use a uint32_t context handle rather than a malloc'd wrapper.
Kenneth Graunke [Tue, 4 Apr 2017 00:32:19 +0000 (17:32 -0700)]
i965: Just use a uint32_t context handle rather than a malloc'd wrapper.

drm_bacon_context is a malloc'd struct containing a uint32_t context ID
and a pointer back to the bufmgr.  The bufmgr pointer is pretty useless,
as everybody already has brw->bufmgr.  At that point...we may as well
just use the ctx_id handle directly.  A number of places already had to
call drm_bacon_gem_context_get_id() to extract the ID anyway.  Now they
just have it.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Fold drm_bacon_gem_reset_stats into the callers.
Kenneth Graunke [Tue, 4 Apr 2017 00:44:57 +0000 (17:44 -0700)]
i965/drm: Fold drm_bacon_gem_reset_stats into the callers.

We're going to get rid of drm_bacon_context shortly, so we'd have to
change the interface slightly.  It's basically just an ioctl wrapper
that isn't terribly bufmgr-related, so We may as well just combine it
with the code in brw_reset.c that actually uses it.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Rename drm_bacon_gem_bo_bucket to bo_cache_bucket.
Kenneth Graunke [Tue, 4 Apr 2017 00:17:36 +0000 (17:17 -0700)]
i965/drm: Rename drm_bacon_gem_bo_bucket to bo_cache_bucket.

No need for a prefix as this struct is local to the .c file.

Less bacon.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop drm_bacon_* from static functions.
Kenneth Graunke [Mon, 3 Apr 2017 23:57:44 +0000 (16:57 -0700)]
i965/drm: Drop drm_bacon_* from static functions.

Mesa style is to not use lengthy prefixes for static functions.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop drm_bacon_gem_bo_madvise_internal().
Kenneth Graunke [Mon, 3 Apr 2017 23:55:06 +0000 (16:55 -0700)]
i965/drm: Drop drm_bacon_gem_bo_madvise_internal().

The only difference is that it takes an explicit bufmgr rather than
using bo->bufmgr, but there is only one bufmgr per screen so they
should be identical anyway.

Chris says this was added primarly to avoid bo/bo_gem casting,
which was inconvenient.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Merge drm_bacon_bo_gem into drm_bacon_bo.
Kenneth Graunke [Mon, 3 Apr 2017 22:55:16 +0000 (15:55 -0700)]
i965/drm: Merge drm_bacon_bo_gem into drm_bacon_bo.

The separate class gives us a bit of extra encapsulation, but I don't
know that it's really worth the boilerplate.  I think we can reasonably
expect the rest of the driver to be responsible.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Merge bo->handle and bo_gem->gem_handle.
Kenneth Graunke [Mon, 3 Apr 2017 22:39:09 +0000 (15:39 -0700)]
i965/drm: Merge bo->handle and bo_gem->gem_handle.

These fields are the same value.  In the bad old days, bo->handle could
have been an identifier from the pre-GEM fake bufmgr, but that's long
gone.  Keep the "gem_handle" name for clarity.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Rewrite relocation handling.
Kenneth Graunke [Wed, 29 Mar 2017 03:20:00 +0000 (20:20 -0700)]
i965/drm: Rewrite relocation handling.

The execbuf2 kernel API requires us to construct two kinds of lists.
First is a "validation list" (struct drm_i915_gem_exec_object2[])
containing each BO referenced by the batch.  (The batch buffer itself
must be the last entry in this list.)  Each validation list entry
contains a pointer to the second kind of list: a relocation list.
The relocation list contains information about pointers to BOs that
the kernel may need to patch up if it relocates objects within the VMA.

This is a very general mechanism, allowing every BO to contain pointers
to other BOs.  libdrm_intel models this by giving each drm_intel_bo a
list of relocations to other BOs.  Together, these form "reloc trees".

Processing relocations involves a depth-first-search of the relocation
trees, starting from the batch buffer.  Care has to be taken not to
double-visit buffers.  Creating the validation list has to be deferred
until the last minute, after all relocations are emitted, so we have the
full tree present.  Calculating the amount of aperture space required to
pin those BOs also involves tree walking, which is expensive, so libdrm
has hacks to try and perform less expensive estimates.

For some reason, it also stored the validation list in the global
(per-screen) bufmgr structure, rather than as an local variable in the
execbuffer function, requiring locking for no good reason.

It also assumed that the batch would probably contain a relocation
every 2 DWords - which is absurdly high - and simply aborted if there
were more relocations than the max.  This meant the first relocation
from a BO would allocate 180kB of data structures!

This is way too complicated for our needs.  i965 only emits relocations
from the batchbuffer - all GPU commands and state such as SURFACE_STATE
live in the batch BO.  No other buffer uses relocations.  This means we
can have a single relocation list for the batchbuffer.  We can add a BO
to the validation list (set) the first time we emit a relocation to it.
We can easily keep a running tally of the aperture space required for
that list by adding the BO size when we add it to the validation list.

This patch overhauls the relocation system to do exactly that.  There
are many nice benefits:

- We have a flat relocation list instead of trees.
- We can produce the validation list up front.
- We can allocate smaller arrays and dynamically grow them.
- Aperture space checks are now (a + b <= c) instead of a tree walk.
- brw_batch_references() is a trivial validation list walk.
  It should be straightforward to make it O(1) in the future.
- We don't need to bloat each drm_bacon_bo with 32B of reloc data.
- We don't need to lock in execbuffer, as the data structures are
  context-local, and not per-screen.
- Significantly less code and a better match for what we're doing.
- The simpler system should make it easier to take advantage of
  I915_EXEC_NO_RELOC in a future patch.

Improves performance in Synmark 7.0's OglBatch7:

    - Skylake GT4e: 12.1499% +/- 2.29531%  (n=130)
    - Apollolake:   3.89245% +/- 0.598945% (n=35)

Improves performance in GFXBench4's gl_driver2 test:

    - Skylake GT4e: 3.18616% +/- 0.867791% (n=229)
    - Apollolake:   4.1776%  +/- 0.240847% (n=120)

v2: Feedback from Chris Wilson:
    - Omit explicit zero initializers for garbage execbuf fields.
    - Use .rsvd1 = ctx_id rather than i915_execbuffer2_set_context_id
    - Drop unnecessary fencing assertions.
    - Only use _WR variant of execbuf ioctl when necessary.
    - Shrink the arrays to be smaller by default.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Make register write check handle execbuffer directly.
Kenneth Graunke [Tue, 28 Mar 2017 23:13:41 +0000 (16:13 -0700)]
i965/drm: Make register write check handle execbuffer directly.

I'm about to rewrite how relocation handling works, at which point
drm_bacon_bo_emit_reloc() and drm_bacon_bo_mrb_exec() won't exist
anymore.  This code is already largely not using the batchbuffer
infrastructure, so just go all the way and handle relocations, the
validation list, and execbuffer ourselves.  That way, we don't have
to think the weird case where we only have a screen, and no context,
when redesigning the relocation handling.

v2: Write reloc.presumed_offset + reloc.delta into the batch, rather
    than duplicating the comment, so it's obvious that they match
    (suggested by Chris).  Also add a comment about why we don't do
    any error checking.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Make a screen::aperture_threshold field.
Kenneth Graunke [Fri, 31 Mar 2017 05:27:42 +0000 (22:27 -0700)]
i965: Make a screen::aperture_threshold field.

This is the threshold after which drm_intel_bufmgr_check_aperture_space
returns -ENOSPC, signalling that it thinks an execbuf is likely to fail
and we need to roll back and flush the batch.

We'll need this when we rewrite aperture space checking, shortly.
In the meantime, we can also use it in GLX_MESA_query_renderer.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Make/use a brw_batch_references() wrapper.
Kenneth Graunke [Tue, 28 Mar 2017 23:49:35 +0000 (16:49 -0700)]
i965: Make/use a brw_batch_references() wrapper.

We'll want to change the implementation of this shortly.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Use brw_emit_reloc() instead of drm_bacon_bo_emit_reloc().
Kenneth Graunke [Tue, 28 Mar 2017 21:41:39 +0000 (14:41 -0700)]
i965: Use brw_emit_reloc() instead of drm_bacon_bo_emit_reloc().

I'm about to make brw_emit_reloc do actual work, so everybody needs
to start using it and not the raw drm_bacon function.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Change intel_batchbuffer_reloc() into brw_emit_reloc().
Kenneth Graunke [Tue, 28 Mar 2017 22:03:49 +0000 (15:03 -0700)]
i965: Change intel_batchbuffer_reloc() into brw_emit_reloc().

This renames intel_batchbuffer_reloc to brw_emit_reloc and changes the
parameter naming and ordering to match drm_intel_bo_emit_reloc().

For now, it's a trivial wrapper that accesses batch->bo.  When we
rework relocations, it will start doing actual work.

target_offset should be expanded to a uint64_t to match the kernel,
but for now we leave it as its original 32-bit type.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop GEM_SW_FINISH stuff.
Kenneth Graunke [Mon, 3 Apr 2017 06:35:27 +0000 (23:35 -0700)]
i965/drm: Drop GEM_SW_FINISH stuff.

This is only useful when doing an incoherent CPU mapping of the current
scanout buffer.  That's a terrible plan, so we never do it.  We always
use an uncached GTT map.

So, this is useless.  Drop the code.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop code to search for an existing bufmgr.
Kenneth Graunke [Fri, 31 Mar 2017 16:44:54 +0000 (09:44 -0700)]
i965/drm: Drop code to search for an existing bufmgr.

This functionality was added by libdrm commit
743af59669386cb6e063fa4bd85f0a0b2da86295 (intel: make bufmgr_gem
shareable from different API) in an attempt to solve libva/mesa buffer
sharing problems.  Specifically, this was working around an issue hit
by Chromium, which used the same drm_fd for multiple APIs, and shared
buffers between them.

This code attempted to work around that issue by using the same bufmgr
for both libva and Mesa.  It worked because libdrm_intel was loaded by
both libraries.  However, now that Mesa has forked, we don't have a
common library, and this code cannot work.

The correct solution is to have each API open its own file descriptor
(and get a corresponding buffer manager), and then use PRIME export
and import to share BOs across those APIs.  Then the kernel can manage
those shared resources.  According to Chris, the kernel will pass back
the same handle for a prime FD if the lookup is from the same device FD.

We believe Chromium has since moved to this model.

In Mesa, there is already only one screen per FD, and so there will
only be one bufmgr per FD.  We don't need any of this code.

v2: Add a big warning comment written by Chris Wilson.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Unwrap the unnecessary drm_bacon_reloc_target_info struct.
Kenneth Graunke [Wed, 29 Mar 2017 00:25:27 +0000 (17:25 -0700)]
i965/drm: Unwrap the unnecessary drm_bacon_reloc_target_info struct.

This used to have another field, but now it's just a BO pointer.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Switch from uthash to Mesa's hash table.
Kenneth Graunke [Thu, 23 Mar 2017 00:29:39 +0000 (17:29 -0700)]
i965/drm: Switch from uthash to Mesa's hash table.

No performance data has been gathered about this choice.  I just don't
want that many hash tables.  Chris points out that this is not
performance critical - we should not be recreating that many handles
from scratch.  In the past we used a linear list, which became
unreasonable in stress tests that used hundreds of thousands of BOs.
In real usage, it shouldn't matter that much.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop bo_gem::kflags.
Kenneth Graunke [Mon, 3 Apr 2017 07:42:30 +0000 (00:42 -0700)]
i965/drm: Drop bo_gem::kflags.

It's always zero now.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop has_exec_async related API.
Kenneth Graunke [Mon, 3 Apr 2017 06:17:54 +0000 (23:17 -0700)]
i965/drm: Drop has_exec_async related API.

Mesa doesn't use this yet.  We'll almost certainly want to, but we can
add the functionality back after we clean up the messy drm code.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop softpin support for now.
Kenneth Graunke [Wed, 22 Mar 2017 19:46:08 +0000 (12:46 -0700)]
i965/drm: Drop softpin support for now.

We may want this eventually, but simplify for now.  We can add it back
later when we actually intend to use it.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop userptr support for now.
Kenneth Graunke [Wed, 22 Mar 2017 19:42:43 +0000 (12:42 -0700)]
i965/drm: Drop userptr support for now.

We'll want userptr support for GL_AMD_pinned_memory support someday,
and possibly some other upload optimizations.  Chris says "not in this
form" though.  Drop it and simplify for now - we can add it back later
when we're ready to hook it up fully.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Delete engine checks.
Kenneth Graunke [Wed, 22 Mar 2017 19:39:44 +0000 (12:39 -0700)]
i965/drm: Delete engine checks.

This is basically handholding to prevent a bogus caller from trying to
execbuffer on a bogus engine.  i965 already does this correctly.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop intel_chipset.h in favor of using gen_device_info.
Kenneth Graunke [Tue, 4 Apr 2017 18:45:08 +0000 (11:45 -0700)]
i965/drm: Drop intel_chipset.h in favor of using gen_device_info.

This moves the PCI ID detection to intel_screen.c and makes
drm_bacon_bufmgr_gem_init() take a devinfo pointer.

We also drop the HAS_LLC query stuff - devinfo has that info already,
without kernel queries, and it makes no sense to have two has_llc flags
set by different mechanisms.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop deprecated drm_bacon_bo::offset.
Kenneth Graunke [Wed, 22 Mar 2017 18:39:39 +0000 (11:39 -0700)]
i965/drm: Drop deprecated drm_bacon_bo::offset.

This field was the wrong size, so we replaced it with offset64.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop has_wait_timeout.
Kenneth Graunke [Wed, 22 Mar 2017 22:28:07 +0000 (15:28 -0700)]
i965/drm: Drop has_wait_timeout.

The wait-ioctl was introduced in kernel v3.6 (20120930) and that is our
current minimum requirement for screen creation.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Assume aperture size query will work.
Kenneth Graunke [Fri, 24 Mar 2017 04:34:23 +0000 (21:34 -0700)]
i965/drm: Assume aperture size query will work.

This query has been available since 2.6.28.  We require 3.6.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Combine drm_bacon_bufmgr_gem and drm_bacon_bufmgr classes.
Kenneth Graunke [Wed, 22 Mar 2017 07:50:21 +0000 (00:50 -0700)]
i965/drm: Combine drm_bacon_bufmgr_gem and drm_bacon_bufmgr classes.

The distinction was required when the bufmgr was virtualised, now there
is only one class, we no longer need the distraction of pretending it is
a subclass.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Move _drm_bacon_context to intel_bufmgr_gem.c.
Kenneth Graunke [Thu, 30 Mar 2017 21:59:23 +0000 (14:59 -0700)]
i965/drm: Move _drm_bacon_context to intel_bufmgr_gem.c.

This moves us one step closer to killing off intel_bufmgr_priv.h.

We might want to nuke it altogether, since it's basically just a
uint32_t handle, but for now, let's focus on removing files.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop cliprects and dr4 from execbuf variants.
Kenneth Graunke [Tue, 4 Apr 2017 21:51:52 +0000 (14:51 -0700)]
i965/drm: Drop cliprects and dr4 from execbuf variants.

Legacy DRI1 leftovers.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Devirtualize the bufmgr.
Kenneth Graunke [Wed, 22 Mar 2017 00:27:25 +0000 (17:27 -0700)]
i965/drm: Devirtualize the bufmgr.

libdrm_bacon used to have a GEM-based bufmgr and a legacy fake bufmgr,
but that's long since dead (and we never imported it to i965).  So,
drop the extra layer of function pointers.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Check INTEL_DEBUG & DEBUG_BUFMGR directly.
Kenneth Graunke [Wed, 22 Mar 2017 05:40:04 +0000 (22:40 -0700)]
i965/drm: Check INTEL_DEBUG & DEBUG_BUFMGR directly.

Eliminates some API around this, and more importantly, the last
field in one bufmgr class.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Use Mesa's macros.h instead of duplicating them.
Kenneth Graunke [Wed, 22 Mar 2017 05:42:43 +0000 (22:42 -0700)]
i965/drm: Use Mesa's macros.h instead of duplicating them.

Replace the duplicated macros imported from libdrm:

   ARRAY_SIZE, MAX2, ALIGN, STATIC_ASSERT

and remove unused ROUND_UP_TO and ROUND_UP_TO_MB.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Use ALIGN, not ROUND_UP_TO.
Kenneth Graunke [Wed, 22 Mar 2017 18:12:27 +0000 (11:12 -0700)]
i965/drm: Use ALIGN, not ROUND_UP_TO.

ROUND_UP_TO handles a NPOT alignment, but all the alignments we use
are power of two anyway, so there's no need.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Delete execbuf1 support.
Kenneth Graunke [Wed, 22 Mar 2017 00:31:24 +0000 (17:31 -0700)]
i965/drm: Delete execbuf1 support.

execbuf2 has been around since v2.6.33.  We require v3.6.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Remove Gen2-3 fence accounting.
Kenneth Graunke [Thu, 30 Mar 2017 18:46:05 +0000 (11:46 -0700)]
i965/drm: Remove Gen2-3 fence accounting.

Since gen4, we do not use fence registers for any GPU access and so
never have to account for the fence during batch construction. All the
related fence functions are unused.

Based on Kristian Høgsberg's patch; commit message by Chris Wilson.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Remove some unused functions and macros.
Kenneth Graunke [Fri, 24 Mar 2017 04:19:24 +0000 (21:19 -0700)]
i965/drm: Remove some unused functions and macros.

Mesa doesn't use these functions or macros, so we can delete them,
and save work refactoring and cleaning them up.  We'll delete a lot
more later, too.

Based on a patch by Kristian Høgsberg.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Switch to util/list.h instead of libdrm_lists.h.
Kenneth Graunke [Tue, 21 Mar 2017 23:10:08 +0000 (16:10 -0700)]
i965/drm: Switch to util/list.h instead of libdrm_lists.h.

Both are kernel style lists, so this is trivial.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Port to Mesa's atomic header.
Kenneth Graunke [Tue, 21 Mar 2017 21:54:07 +0000 (14:54 -0700)]
i965/drm: Port to Mesa's atomic header.

Drop xf86atomic.h in favor of Mesa's util/u_atomic.h.  We replace the
atomic_t wrapper struct with a bare integer, switch to the 'p_atomic'
naming conventions, and move over the one extra helper.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Use our internal libdrm (drm_bacon) rather than the real one.
Kenneth Graunke [Mon, 20 Mar 2017 23:42:55 +0000 (16:42 -0700)]
i965/drm: Use our internal libdrm (drm_bacon) rather than the real one.

Now we can actually test our changes.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: s/drm_intel/drm_bacon/g
Kenneth Graunke [Thu, 23 Mar 2017 01:51:42 +0000 (18:51 -0700)]
i965/drm: s/drm_intel/drm_bacon/g

Using drm_intel_* as a prefix is hazardous - we don't want to conflict
with the actual libdrm_intel symbols.  In particular, I think we could
get into trouble during the final megadrivers linking.

So, rename everything to an different yet arbitrary prefix.  bacon and
intel are the same number of characters, so we don't have to reindent
the world.  It's also an homage to Ian's "Bacon Trail" platform.

I was going to use "drm_relic" to poke fun at libdrm being ancient,
and so we could explain the name with a "historical reasons" pun,
but it sounds too much like ralloc.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Drop libpciaccess dependencies.
Kenneth Graunke [Thu, 23 Mar 2017 22:29:43 +0000 (15:29 -0700)]
i965/drm: Drop libpciaccess dependencies.

i965 doesn't use drm_intel_get_aperture_sizes(), so we can delete
support for it.  This avoids a build dependency on libpciaccess.

Chris also notes:

"There's a really old bug that hopefully has been closed already
 (although as far as I can tell, it has never been fixed) about
 how using libpciaccess from libdrm_intel breaks the world (since
 libpciaccess uses a singleton that is torn down at the first request
 rather than upon the last user)."

This bug should go away in two commits when we switch over to our
internal copy of libdrm_intel.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=84325
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: Make libdrm_lists.h compile by defining typeof.
Kenneth Graunke [Tue, 21 Mar 2017 22:55:29 +0000 (15:55 -0700)]
i965/drm: Make libdrm_lists.h compile by defining typeof.

typeof doesn't seem to exist, so this won't compile (but we don't yet
try).  Define it to __typeof__.  This code is going to die soon anyway.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965/drm: remove legacy defines, aub functions, and decoder prototypes
Kenneth Graunke [Fri, 31 Mar 2017 17:06:24 +0000 (10:06 -0700)]
i965/drm: remove legacy defines, aub functions, and decoder prototypes

We never imported any of this code, so drop the prototypes, unused
enums, and defines.

Based on patches by Emil Velikov.

Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Import libdrm_intel.
Kenneth Graunke [Mon, 20 Mar 2017 23:40:01 +0000 (16:40 -0700)]
i965: Import libdrm_intel.

This imports commit 19c4cfc54918d361f2535aec16650e9f0be667cd of
libdrm/intel/*.[ch], minus a few files that we're never going to use
(and would immediately delete), plus a few necessary dependencies.

We rename intel_bufmgr.h to brw_bufmgr.h to avoid #include conflicts.
We also fix UTF-8 symbol problems in intel_bufmgr_gem.c comments
because vim keeps trying to fix that every time I edit the file,
and we may as well fix it right away.

Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Make sure we don't use CPU maps for the scanout buffer.
Kenneth Graunke [Mon, 3 Apr 2017 07:03:01 +0000 (00:03 -0700)]
i965: Make sure we don't use CPU maps for the scanout buffer.

Using an incoherent CPU map on the active scanout buffer is really
sketchy - we may need extra flushing via GEM_SW_FINISH, or using
drmModeDirtyFB() and kernel commit a6a7cc4b7db6d (4.10+).

Chris suggests "never ever do that", which seems like a wise plan!

intel_miptree_map_raw() uses CPU maps on linear buffers.

Having a linear scanout buffer should be really rare, and mapping the
front buffer should be similarly rare.  Together, it should basically
never happen.  But, in case it does somehow...make sure that mapping
the scanout buffer always goes through an uncached GTT map.

v2: Add a giant comment written by Chris Wilson.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
7 years agoi965: Stop calling drm_intel_bufmgr_gem_enable_fenced_relocs().
Kenneth Graunke [Wed, 22 Mar 2017 00:08:20 +0000 (17:08 -0700)]
i965: Stop calling drm_intel_bufmgr_gem_enable_fenced_relocs().

This does nothing on Gen4+, which is the only hardware we support.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agoi965: Fix GLX_MESA_query_renderer video memory on 32-bit.
Kenneth Graunke [Thu, 30 Mar 2017 20:52:46 +0000 (13:52 -0700)]
i965: Fix GLX_MESA_query_renderer video memory on 32-bit.

On modern systems with 4GB apertures, the size in bytes is 4294967296,
or (1ull << 32).  The kernel gives us the aperture size as a __u64,
which works out great.

Unfortunately, libdrm "helpfully" returns the data as a size_t, which
on 32-bit systems means it truncates the aperture size to 0 bytes.
We've happily reported this value as 0 MB of video memory via
GLX_MESA_query_renderer since it was originally exposed.

This patch bypasses libdrm and calls the ioctl ourselves so we can
use a proper uint64_t, avoiding the 32-bit integer overflow.  We now
report a proper video memory size on 32-bit systems.

Chris points out that the aperture size (CPU mappable size limit)
isn't really the right thing to be checking.  But libdrm_intel uses
it to fail execbuffer, so it is an actual limit for now.  Once that's
fixed we can probably move to something else.  In the meantime, fix
the obvious typecasting bug.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agogallium/radeon: add HUD queries for GPU temperature and clocks
Samuel Pitoiset [Mon, 10 Apr 2017 09:49:05 +0000 (11:49 +0200)]
gallium/radeon: add HUD queries for GPU temperature and clocks

Only the Radeon kernel driver exposed the GPU temperature and
the shader/memory clocks, this implements the same functionality
for the AMDGPU kernel driver.

These queries will return 0 if the DRM version is less than 3.10,
I don't explicitely check the version here because the query
codepath is already a bit messy.

v2: - rebase on top of master

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoconfigure.ac: require libdrm_amdgpu 2.4.79
Samuel Pitoiset [Mon, 10 Apr 2017 09:49:04 +0000 (11:49 +0200)]
configure.ac: require libdrm_amdgpu 2.4.79

The sensor info requires amdgpu_query_sensor_info().

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoradeonsi: add new si_check_render_feedback_texture() helper
Samuel Pitoiset [Wed, 5 Apr 2017 22:07:35 +0000 (00:07 +0200)]
radeonsi: add new si_check_render_feedback_texture() helper

For bindless.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoradeonsi: add new si_decompress_color_texture() helper
Samuel Pitoiset [Wed, 5 Apr 2017 22:07:34 +0000 (00:07 +0200)]
radeonsi: add new si_decompress_color_texture() helper

For bindless.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoradeonsi: add new depth_needs_decompression() helper
Samuel Pitoiset [Wed, 5 Apr 2017 22:07:33 +0000 (00:07 +0200)]
radeonsi: add new depth_needs_decompression() helper

v2: - rename to depth_needs_decompression() instead

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoradeonsi: add a 'break' in si_check_render_feedback_*()
Samuel Pitoiset [Wed, 5 Apr 2017 22:07:32 +0000 (00:07 +0200)]
radeonsi: add a 'break' in si_check_render_feedback_*()

No need to check all color buffers.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoradeonsi: re-use 'desc' in si_set_shader_image()
Samuel Pitoiset [Wed, 5 Apr 2017 22:07:31 +0000 (00:07 +0200)]
radeonsi: re-use 'desc' in si_set_shader_image()

No need to compute the offset in the descriptor twice.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agoac: add unreachable() in ac_build_image_opcode()
Samuel Pitoiset [Fri, 7 Apr 2017 16:44:16 +0000 (18:44 +0200)]
ac: add unreachable() in ac_build_image_opcode()

To silent the following compiler warning:

common/ac_llvm_build.c: In function ‘ac_build_image_opcode’:
common/ac_llvm_build.c:1080:3: warning: ‘name’ may be used uninitialized in this function [-Wmaybe-uninitialized]
   snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    name,
    ~~~~~
    a->compare ? ".c" : "",
    ~~~~~~~~~~~~~~~~~~~~~~~
    a->bias ? ".b" :
    ~~~~~~~~~~~~~~~~
    a->lod ? ".l" :
    ~~~~~~~~~~~~~~~
    a->deriv ? ".d" :
    ~~~~~~~~~~~~~~~~~
    a->level_zero ? ".lz" : "",
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~
    a->offset ? ".o" : "",
    ~~~~~~~~~~~~~~~~~~~~~~
    type);
    ~~~~~

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agor600g: get rid of dummy pixel shader
Constantine Kharlamov [Mon, 10 Apr 2017 20:04:37 +0000 (23:04 +0300)]
r600g: get rid of dummy pixel shader

The idea is taken from radeonsi. The code mostly was already checking for null
pixel shader, so little checks had to be added.

Interestingly, acc. to testing with GTAⅣ, though binding of null shader happens
a lot at the start (then just stops), but draw_vbo() never actually sees null
ps.

v2: added a check I missed because of a macros using a prefix to choose
a shader.

Signed-off-by: Constantine Kharlamov <Hi-Angel@yandex.ru>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
7 years agor600g: add draw_vbo check for a NULL pixel shader
Constantine Kharlamov [Mon, 10 Apr 2017 20:04:36 +0000 (23:04 +0300)]
r600g: add draw_vbo check for a NULL pixel shader

Taken from radeonsi, required to remove dummy pixel shader in the next patch

Signed-off-by: Constantine Kharlamov <Hi-Angel@yandex.ru>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
7 years agor600g: skip repeating vs, gs, and tes shader binds
Constantine Kharlamov [Mon, 10 Apr 2017 20:04:35 +0000 (23:04 +0300)]
r600g: skip repeating vs, gs, and tes shader binds

The idea is taken from radeonsi. The code lacks some checks for null vs,
and I'm unsure about some changes against that, so I left it in place.

Some statistics for GTAⅣ:
Average tesselation bind skip per frame: ≈350
Average geometric shaders bind skip per frame: ≈260
Skip of binding vertex ones occurs rarely enough to not get into per-frame
counter at all, so I just gonna say: it happens.

v2: I've occasionally removed an empty line, don't do this.
v3: return a check for null tes and gs back, while I haven't figured out
the way to move stride assignment to r600_update_derived_state() (as it
is in radeonsi).

Signed-off-by: Constantine Kharlamov <Hi-Angel@yandex.ru>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
7 years agomesa: use single memcpy when strides match in glReadPixels, texstore code
Bartosz Tomczyk [Mon, 10 Apr 2017 18:31:00 +0000 (12:31 -0600)]
mesa: use single memcpy when strides match in glReadPixels, texstore code

v2: fix indentation

Reviewed-by: Brian Paul <brianp@vmware.com>
7 years agointel/blorp: Use ISL for emitting depth/stencil/hiz
Jason Ekstrand [Thu, 30 Mar 2017 06:00:16 +0000 (23:00 -0700)]
intel/blorp: Use ISL for emitting depth/stencil/hiz

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
7 years agointel/blorp: Emit 3DSTATE_STENCIL_BUFFER before HIER_DEPTH
Jason Ekstrand [Wed, 5 Apr 2017 23:59:06 +0000 (16:59 -0700)]
intel/blorp: Emit 3DSTATE_STENCIL_BUFFER before HIER_DEPTH

We're about to replace blorp's emit code with ISL and it emits them in
the other order.  This makes diffing the aubs easier.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
7 years agoanv: Use ISL for emitting depth/stencil/hiz
Jason Ekstrand [Thu, 30 Mar 2017 05:24:50 +0000 (22:24 -0700)]
anv: Use ISL for emitting depth/stencil/hiz

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
7 years agointel/isl: Add support for emitting depth/stencil/hiz
Jason Ekstrand [Thu, 30 Mar 2017 04:10:50 +0000 (21:10 -0700)]
intel/isl: Add support for emitting depth/stencil/hiz

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
7 years agoamd/addrlib: use correct variable name in header
Thomas Hindoe Paaboel Andersen [Sat, 8 Apr 2017 06:36:36 +0000 (08:36 +0200)]
amd/addrlib: use correct variable name in header

Since the inclusion in 7f160efcde41b52ad78e562316384373dab419e3
the header used x_biased, while the implementation used y_biased.
This changes the header to macth the implementation since the
uses of the function seems to expect y_biased.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
7 years agomesa/st: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:54:56 +0000 (10:54 +1000)]
mesa/st: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Tested-by: Bartosz Tomczyk <bartosz.tomczyk86@gmail.com>
7 years agox11: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:47:12 +0000 (10:47 +1000)]
x11: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agoosmesa: tidy up renderbuffer refCount initialisation
Timothy Arceri [Sat, 8 Apr 2017 00:42:57 +0000 (10:42 +1000)]
osmesa: tidy up renderbuffer refCount initialisation

32141e53d1520 changed _mesa_init_renderbuffer() to set it to 1 for
us.

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agoswrast: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:35:57 +0000 (10:35 +1000)]
swrast: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agoradeon: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:29:22 +0000 (10:29 +1000)]
radeon: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agonouveau: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:26:34 +0000 (10:26 +1000)]
nouveau: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agoi965: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:22:16 +0000 (10:22 +1000)]
i965: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agoi915: take ownership rather than adding reference for new renderbuffers
Timothy Arceri [Sat, 8 Apr 2017 00:13:24 +0000 (10:13 +1000)]
i915: take ownership rather than adding reference for new renderbuffers

This avoids locking in the reference calls and fixes a leak after the
RefCount initialisation was change from 0 to 1.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
7 years agomesa: create _mesa_attach_renderbuffer_without_ref() helper
Timothy Arceri [Sat, 8 Apr 2017 00:03:20 +0000 (10:03 +1000)]
mesa: create _mesa_attach_renderbuffer_without_ref() helper

This will be used to take ownership of freashly created renderbuffers,
avoiding the need to call the reference function which requires
locking.

V2: dereference any existing fb attachments and actually attach the
    new rb.

v3: split out validation and attachment type/complete setting into
    a shared static function.

Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Tested-by: Bartosz Tomczyk <bartosz.tomczyk86@gmail.com>
7 years agonv50/ir: remove unused swizzle field in ValueRef
Ilia Mirkin [Sun, 9 Apr 2017 18:56:59 +0000 (14:56 -0400)]
nv50/ir: remove unused swizzle field in ValueRef

The nv50 ir is scalar. Perhaps this was from some early attempts to
integrate the simd aspects of nv30. However at this point it's entirely
unused.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
7 years agonouveau: enable ARB_shader_clock on nv50 and nvc0
Boyan Ding [Tue, 4 Apr 2017 14:44:47 +0000 (22:44 +0800)]
nouveau: enable ARB_shader_clock on nv50 and nvc0

v2: Also enable support on nv50

Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
7 years agonv50/ir: Handle TGSI_OPCODE_CLOCK
Boyan Ding [Tue, 4 Apr 2017 14:44:46 +0000 (22:44 +0800)]
nv50/ir: Handle TGSI_OPCODE_CLOCK

Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
[imirkin: make zero mov non-fixed]
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
7 years agogm107/ir: Emit SV_CLOCK system value
Boyan Ding [Fri, 31 Mar 2017 02:33:05 +0000 (10:33 +0800)]
gm107/ir: Emit SV_CLOCK system value

Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
7 years agogbm: Assert modifiers and count are copacetic
Ben Widawsky [Wed, 5 Apr 2017 17:31:45 +0000 (10:31 -0700)]
gbm: Assert modifiers and count are copacetic

The API/entry point in mesa already checks the correct behavior,
however, it's possible to be handled by another implementation and those
implementations should not be able to abuse a weird combination of count
and pointer.

This fixes CID 1403193

Cc: Mark Janes <mark.a.janes@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
7 years agost/mesa: Use compressed fog mode for atifs.
Gustaw Smolarczyk [Thu, 30 Mar 2017 18:09:33 +0000 (20:09 +0200)]
st/mesa: Use compressed fog mode for atifs.

Signed-off-by: Gustaw Smolarczyk <wielkiegie@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>