Define some additional convenience operators, clean up the
implementation slightly, and rename it to 'intrusive_ptr' for reasons
that will be obvious in the next commit.
Tested-by: Tom Stellard <thomas.stellard@amd.com>
Fixes a build break in state_tracker/st_program.c
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=75278
Reviewed-by: Dave Airlie <airlied@redhat.com>
I'd neglected to port these to Broadwell. Most of this code is copy
and pasted from Gen7, but instead of using F32TO16/F16TO32, we just
use MOV with HF register types.
Fixes fs-packHalf2x16 and fs-unpackHalf2x16 tests (both the ARB
extension and ES 3.0 variants).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Broadwell removed the F32TO16 and F16TO32 instructions. However, it has
actual support for HF values, so they're actually just MOV.
Fixes vs-packHalf2x16 and vs-unpackHalf2x16 tests (both the ARB
extension and ES 3.0 variants).
v2: Emulate F32TO16's align16 zeroing bug, since Chad's front end code
relies on it happening. We can probably refactor this code to be
better later.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
brw_init_state() calls brw_upload_initial_gpu_state(). If hardware
contexts are enabled (brw->hw_ctx != NULL), this will upload some
initial invariant state for the GPU. Without hardware contexts, we
rely on this state being uploaded via atoms that subscribe to the
BRW_NEW_CONTEXT bit.
Commit 46d3c2bf4d accidentally moved
the call to brw_init_state() before creating a hardware context.
This meant brw_upload_initial_gpu_state would always early return.
Except on Gen6+, we stopped uploading the initial GPU state via
state atoms, so it never happened.
Fixes a regression since 46d3c2bf4d.
Cc: "10.0 10.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
To make sure that both the Gen4 and Gen7 style messages work, I
initially disabled the SHADER_OPCODE_GEN7_SCRATCH_READ optimization,
ran Piglit, re-enabled it, and ran Piglit again. Both worked fine.
Fixes 40 Piglit tests (most of the varying-packing category).
v2: Move num_regs assertion from gen8_fs_generator to
gen8_set_dp_scratch_message() (suggested by Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The new accessors will make it easy to do Gen7-style scratch messages.
v2: Move num_regs assertion from gen8_fs_generator into
gen8_set_dp_scratch_message() (suggested by Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
In the past, 3DSTATE_PS took an absolute number of threads. Conversely,
on Broadwell you always program 64, and it implicitly scales based on
the GT-level with no special programming. So, I stored 64 in
brw_device_info::max_wm_threads.
However, I didn't realize that we also use max_wm_threads to compute the
size of the scratch space buffer. In that case, we really need the
absolute number of threads.
This patch hardcodes 3DSTATE_PS to use the value it expects, and changes
max_wm_threads back to a (completely fake) absolute thread count (once
again copied from Haswell).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
On Broadwell, g0.5 contains the "Scratch Space Pointer"; using OR
puts some bits of that into "ignored" sections of our message header.
While this doesn't hurt, it's also not terribly /useful/. Using MOV
is sufficient to set the only interesting bits in this part of the
message header.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
According to the latest documentation, any PIPE_CONTROL with the
"Command Streamer Stall" bit set must also have another bit set,
with five different options:
- Render Target Cache Flush
- Depth Cache Flush
- Stall at Pixel Scoreboard
- Post-Sync Operation
- Depth Stall
I chose "Stall at Pixel Scoreboard" since we've used it effectively
in the past, but the choice is fairly arbitrary.
Implementing this in the PIPE_CONTROL emit helpers ensures that the
workaround will always take effect when it ought to.
Apparently, this workaround may be necessary on older hardware as well;
for now I've only added it to Broadwell as it's absolutely necessary
there. Subsequent patches could add it to older platforms, provided
someone tests it there.
v2: Only flag "Stall at Pixel Scoreboard" when none of the other bits
are set (suggested by Ian Romanick).
v3: Prefix the function with "gen8" (requested by Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v2)
Reviewed-by: Eric Anholt <eric@anholt.net>
v2:
* Make gl_InvocationID a system value
v3:
* Properly shift from R0.1 into DST.4 by adding
GS_OPCODE_GET_INSTANCE_ID
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
v2:
* Make gl_InvocationID a system value
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Grab the parsed invocation count, check for consistency
during linking, and finally save the result in
gl_shader_program Geom.Invocations.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
_mesa_glsl_parse_state in_qualifier->invocations will store the
invocations count.
v3:
* Use in_qualifier to allow the primitive to be specied
separately from the invocations count (merge_qualifiers)
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
We introduce a new merge_in_qualifier ast_type_qualifier
which allows specialized handling of merging input layout
qualifiers.
By merging layout qualifiers into state->in_qualifier, we
allow multiple input qualifiers. For example, the primitive
type can be specified specified separately from the
invocations count (ARB_gpu_shader5).
state->gs_input_prim_type is moved into state->in_qualifier->prim_type
state->gs_input_prim_type_specified is still processed separately
so we can determine when the input primitive is specified. This
is important since certain scenerios are not supported until after
the primitive type has been specified in the shader code.
v4:
* Merge with compute shader input layout qualifiers
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Improves performance of a dolphin emulator trace I had laying around by
3.60131% +/- 0.995887% (n=128).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We generate steaming piles of these for the centroid workaround, and this
quickly cleans them up.
total instructions in shared programs: 1591228 -> 1590047 (-0.07%)
instructions in affected programs: 26111 -> 24930 (-4.52%)
GAINED: 0
LOST: 0
(Improved apps are l4d2, csgo, and dolphin)
Reviewed-by: Matt Turner <mattst88@gmail.com>
The previous code relied on cpu denorm support for converting small float
formats (such r11g11b10_float and r16_float) to floats, otherwise denorms
are flushed to zero. We worked around that in llvmpipe blend code by
reenabling denorms, but this did nothing for texture sampling. Now it would
be possible to reenable it there too but I'm not really a fan of messing
with fpu flags (and it seems we can't actually do it reliably with llvm in
any case looking at some bug reports). (Not to mention if you actually have
a lot of denorms in there, you can expect some order-of-magnitude slowdown
with x86 cpus.)
So instead use code which adjusts exponents etc. directly hence not relying
on cpu denorm support for the rescaling mul.
(We still need the fpu flag handling as we can't do float-to-smallfloat
without using cpu denorms at least for now - I actually wanted to keep
both the old and new code and using one or the other depending on from where
it's called but that didn't work out as the parameter would have to be passed
through too many layers than I'd like.)
Reviewed-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Si Chen <sichen@vmware.com>
We need to advertise 8x, 4x, and 2x multisamples. Previously, we only
claimed to support 0/1 samples.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
I can't find any documentation to explain what ought to be done here, so
I simply guessed based on the pattern I observed in the 4x/8x cases.
It appears to work, but it could be totally wrong.
I was able to find the Sandybridge PRM quote from the comments in the
latest documentation: Shared Functions > 3D Sampler > Multisampled
Surface Behavior. However, it only mentions 4x MSAA - not even 8x.
After a substantial amount more digging, I was able to find a second
page (incorrectly tagged) which confirmed the formulas in our code for
8x MSAA. However, that page didn't mention 2x MSAA at all.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
According to the "Point Multisample Rasterization" of the OpenGL
specification (3.0 or later), smooth points are supposed to be enabled
implicitly when multisampling, regardless of the GL_POINT_SMOOTH flag.
However, if GL_POINT_SPRITE is enabled, you get square points no matter
what. Core contexts always enable point sprites, so this effectively
makes smooth points go away, even in the case of multisampling.
Fixes Piglit's EXT_framebuffer_multisample/point-smooth tests.
(Yes, that's right folks, we actually have Piglit tests for this.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
The meaning and effects of this bit are surprisingly complicated.
See Rasterization > Windower > Multisampling > Multisample ModesState.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
This restriction carries forward from earlier platforms. The code is
ported straight from gen7_wm_state.c.
v2: Actually do it right.
v3: Add missing _NEW_MULTISAMPLE bit (caught by Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
v2: Also set the "oMask Present to Render Target" bit, which is required
for shaders that write oMask. Otherwise the hardware won't expect
the extra data.
v3: Add missing _NEW_MULTISAMPLE (caught by Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
I made a few changes which I think simplify the code a bit compared to
the Gen7 implementation, but which are largely pointless.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Largely cut and paste from Gen7; it works the same way.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
v2: Add a perf_debug() message to remind us to come back to this.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
We already set the number of samples, but were missing the MSAA layout
mode. Reusing gen7_surface_msaa_bits makes it easy to set both.
This also lets us drop the Gen8 surface_num_multisamples function.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
The enumerations are just log2(num_samples) shifted by 3, which we can
easily compute via ffs().
This also makes it reusable for Broadwell, which has 2x MSAA.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
These enumerations are simply log2 of the number of multisamples shifted
by a bit, so we can calculate them using ffs() in a lot less code.
Suggested by Eric Anholt.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
The const in
const unsigned foo(void);
is meaningless. Removing it silences this warning:
src/glsl/ast_to_hir.cpp:1802:56: warning: type qualifiers ignored on function return type [-Wignored-qualifiers]
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
From page 14 (page 20 of the PDF) of the GLSL 1.10 spec:
"In addition, all identifiers containing two consecutive underscores
(__) are reserved as possible future keywords."
The intention is that names containing __ are reserved for internal use
by the implementation, and names prefixed with GL_ are reserved for use
by Khronos. Names simply containing __ are dangerous to use, but should
be allowed.
Per the Khronos bug mentioned below, a future version of the GLSL
specification will clarify this.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: "9.2 10.0 10.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Tested-by: Darius Spitznagel <d.spitznagel@goodbytez.de>
Cc: Tapani Pälli <lemody@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=71870
Bugzilla: Khronos #11702
Section 3.3 (Preprocessor) of the GLSL 1.30 spec (and later) and the
GLSL ES spec (all versions) say:
"All macro names containing two consecutive underscores ( __ ) are
reserved for future use as predefined macro names. All macro names
prefixed with "GL_" ("GL" followed by a single underscore) are also
reserved."
The intention is that names containing __ are reserved for internal use
by the implementation, and names prefixed with GL_ are reserved for use
by Khronos. Since every extension adds a name prefixed with GL_ (i.e.,
the name of the extension), that should be an error. Names simply
containing __ are dangerous to use, but should be allowed. In similar
cases, the C++ preprocessor specification says, "no diagnostic is
required."
Per the Khronos bug mentioned below, a future version of the GLSL
specification will clarify this.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: "9.2 10.0 10.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Tested-by: Darius Spitznagel <d.spitznagel@goodbytez.de>
Cc: Tapani Pälli <lemody@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=71870
Bugzilla: Khronos #11702
Linking with LLVM static libraries is easily broken by changes to
the llvm-config program or when LLVM adds, removes, or changes library
components. Keeping up with these changes requires a lot of maintanence
effort to keep the build working on the master and stable branches.
Also, because of issues in the past LLVM static libraries, the release
manager is currently configuring with --with-llvm-shared-libs when
checking the build before release. Enabling shared libraries by
default would allow the release manager to run ./configure with
no arguments, and be reasonably confident that the build would succeed.
Acked-by: Emil Velikov <emil.l.velikov@gmail.com>
Useful because the total number of uniform components might exceed
MAX_UNIFORMS * 4 in some cases because of the image metadata we'll be
passing as push constants.
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Like the VEC4 back-end does. It will make dynamic allocation of the
param_size array easier in a future commit.
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Since we are now consuming two ringbuffers at a time, we probably want a
pool larger than 4.. but we don't need each individual ringbuffer to be
so large, so offset the pool size increase by reducing rb size.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
It seems the write-after-read hazard that applies to texture fetch
instructions, also applies to sfu instructions.
Also, cat5/cat6 instructions do not have a (ss) bit, so in these
cases we need to insert a dummy nop instruction with (ss) bit set.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Francisco Jerez
Jordan Justen
Kenneth Graunke
Eric Anholt
Roland Scheidegger
Leo Liu
Christian König
Ian Romanick
Tom Stellard
Rob Clark