There are currently 2 users of this functionality. I have 2 more users coming
up, and having a simple function makes the results much cleaner. The existing
interface semantics was proposed by Matt.
v2 (Ken): Rename to region_matches()/has_scalar_region().
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
GEN8 added the QWORD as a valid type for certain operations on the EU.
In order to calculate the number of registers used one must have the type
size as part of the equation. Quoting the formula in the code:
regs_written = (dst.width * dst.stride * type_sz(dst.type) + 31) / 32;
Adding this separately for bisection since there is no simple way to add
an assert in the type_sz function.
NOTE: As a side note, I was confused for a while because it's impossible
to calculate the region, ie. registers needed, without vstride. However,
at this point these are all part of the IR, and so no vstride must exist.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Instead of passing a pointer to the scratch buffer via user sgprs, we
now patch the shader with the buffer address using reloc information
from the LLVM generated ELF.
v2:
- Make sure not to break older LLVM.
Gen4 hardware appears to GPU hang frequently when using Chromium, and
also when running 'glmark2 -b ideas'. Most of the error states contain
3DPRIMITIVE commands in quick succession, with very few state packets
between them - usually VERTEX_BUFFERS/ELEMENTS and CONSTANT_BUFFER.
I trimmed an apitrace of the glmark2 hang down to two draw calls with a
glUniformMatrix4fv call between the two. Either draw by itself works
fine, but together, they hang the GPU. Removing the glUniform call
makes the hangs disappear. In the hardware state, this translates to
removing the CONSTANT_BUFFER packet between the two 3DPRIMITIVE packets.
Flushing before emitting CONSTANT_BUFFER packets also appears to make
the hangs disappear. I observed a slowdown in glxgears by doing it all
the time, so I've chosen to only do it when BRW_NEW_BATCH and
BRW_NEW_PSP are unset (i.e. we haven't done a CS_URB_STATE change or
already flushed the whole pipeline).
I'd much rather understand the problem, but at this point, I don't see
how we'd ever be able to track it down further. We have no real tools,
and the hardware people moved on years ago. I've analyzed 20+ error
states and read every scrap of documentation I could find.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=80568
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=85367
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Matt Turner <mattst88@gmail.com>
Cc: "10.4 10.3" <mesa-stable@lists.freedesktop.org>
With the previous commits in place, it just works.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
offset() properly handles reg_width, so it'll work for SIMD16.
While we're in the area, simplify a few cases, and use retype() to cut a
few more lines of code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
brw_fs_nir.cpp creates almost all of its registers via:
fs_reg reg = fs_reg(GRF, virtual_grf_alloc(num_components));
When we add SIMD16 support, we'll need to set reg->width = 16 and
double the VGRF size...on pretty much every VGRF it allocates.
This patch replaces that pattern with a new "vgrf" helper method:
fs_reg reg = vgrf(num_components);
The new function correctly takes reg_width into account. For now,
reg_width is always 1, so this should have no functional change.
v2: Just make vgrf() account for reg_width right away, rather than
changing the behavior in the next patch.
v3: Replace one last virtual_grf_alloc I missed. It's used in code
that only runs for dispatch_width == 8, so it doesn't matter,
but consistency is nice.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
I dislike how fs_reg has a constructor that knows about fs_visitor.
Apart from that, it stands alone, with no need to interact with the
rest of the compiler. Which is sensible - a class that represents
a register should do just that. Allocating virtual register numbers
should be left up to the compiler (fs_visitor).
This patch replaces the constructor with a new fs_visitor::vgrf method,
eliminating fs_reg's dependency on fs_visitor. It ends up being no
more code.
v2: Rebase from May 2014 -> January 2015.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The filename of sha1.h was conflicting with the system-provided
sha1.h, (and in some confiurations, our sha1.c was unsuccessfully
attemping to include "sha1.h" and <sha1.h> as two different files).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88523
Commit 8ec6534 changed texture upload path and the way how texture
format is being checked, this commit adds support for GL_RGB with
GL_UNSIGNED_INT_2_10_10_10_REV as specified by the extension
EXT_texture_type_2_10_10_10_REV specification.
This fixes regression in ES3 conformance test
ES3-CTS.gtf.GL3Tests.packed_pixels.packed_pixels
v2: add MESA_FORMAT_R10G10B10X2_UNORM format (Iago Toral)
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88385
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
We hit an assertion that the destination of the FB write should not be
an immediate. (I don't know what we were thinking.) Use ARF null.
Trying to substitute real shaders with the dummy shader would crash
when trying to upload non-existent uniforms. Say there are none.
It also wouldn't generate any code because we didn't compute the CFG,
and code generation now requires it. Compute it.
Gen4-5 also require a message header to be present.
On Gen6+, there were assertion failures in SF/SBE state because
urb_setup was memset to 0 instad of -1, causing it to think there were
attributes when nothing was set up right. Set to no attributes.
Finally, you have to ensure "Setup URB Entry Read Length" is non-zero
or you get GPU hangs, at least on Crestline.
It now works on at least Crestline and Haswell.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
glibc 2.19 introduced _DEFUAULT_SOURCE as a replacement for _BSD_SOURCE,
and deprecates _BSD_SOURCE with an annoying warning. Defining both is
how you're supposed to transition so let's do that. It gets rid of the
warning and we can figure out when/if we can drop _BSD_SOURCE later.
Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Fix build error.
CC libmesautil_la-sha1.lo
sha1.c: In function '_mesa_sha1_final':
sha1.c:210:22: error: 'grcy_md_hd_t' undeclared (first use in this function)
gcry_md_hd_t h = (grcy_md_hd_t) ctx;
^
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88519
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Fix build error on Mac OS X.
CC nir_to_ssa.lo
nir_to_ssa.c:29:10: fatal error: 'malloc.h' file not found
^
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88478
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
The driver name is no longer const, it's always allocated dynamically
one way or another. Drop const from dri_screen_create_dri2
driver_name argument to avoid warning.
Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
We don't actually have the code for the shader cache just yet, but
this configure machinery puts everything in place so that the shader
cache can be optionally compiled in.
Specifically, if the user passes no option (neither
--disable-shader-cache, nor --enable-shader-cache), then this feature
will be automatically detected based on the presence of a usable SHA-1
library. If no suitable library can be found, then the shader cache
will be automatically disabled, (and reported in the final output from
configure).
The user can force the shader-cache feature to not be compiled, (even
if a SHA-1 library is detected), by passing
--disable-shader-cache. This will prevent the compiled Mesa libraries
from depending on any library for SHA-1 implementation.
Finally, the user can also force the shader cache on with
--enable-shader-cache. This will cause configure to trigger a fatal
error if no sutiable SHA-1 implementation can be found for the
shader-cache feature.
Bug fix by José Fonseca <jfonseca@vmware.com>: Fix to put conditional
assignment in Makefile.am, not Makefile.sources to avoid breaking
scons build.
Note: As recommended by José, with this commit the scons build will
not compile any of the SHA-1-using code. This is waiting for someone
to write SConstruct detection of the available SHA-1 libraries, (and
set the appropriate HAVE_SHA1_* variables).
Reviewed-by: Matt Turner <mattst88@gmail.com>
The upcoming shader cache uses the SHA-1 algorithm for cryptographic
naming. These new mesa_sha1 functions are implemented with any one of
several differeny cryptographics libraries.
This code was copied from the xserver repository, (where it has
apparently been functioning well on a variety of operating systems),
and comes licensed with a license identical to that of Mesa.
Bug fixes by José Fonseca <jfonseca@vmware.com>: Fix to put
conditional assignment in Makefile.am, not Makefile.sources to avoid
breaking scons build. Fix include file for CryptoAPI section. Fix
missing cast in openssl section.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Prior to copying in code from the xserver configure.ac file, it makes
sense to have the license of this file clearly marked, (to show that
it's licensed identically to the configure.ac file from the xserver
repository).
And since the text of the license refers to "the above copyright
notice" it also makes sense to have an actual copyright attribution in
place.
I generated this list of names by looking at the output of:
git shortlog -n --format=%aD -- configure.ac
(and arbitrarily stopping for contributors with fewer than 15
commits). Then for each name, I looked for existing Copyright
attributions in the mesa source tree with the same name, (and using
"Intel Corporation" as the copyright holder where I knew that was
appropriate).
In addition to exercising all of the functions in blob.h, this
includes a stress test that forces some reallocing, and also tests to
verify the alignment and overrun-detection code in blob.c.
These functions are useful when serializing an unknown number of items
to a blob. The caller can first save the current offset, write a
placeholder uint32, write out (and count) the items, then use
blob_overwrite_uint32 with the saved offset to replace the placeholder
value.
Then, when deserializing, the reader will first read the count and
know how many subsequent items to expect.
(I wrote this code after reading a very similar patch written by
Tapani when he wrote serialization code for IR. Since I re-used the
idea of his code so directly, I've credited him as the author of this
code. --Carl)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This new interface allows for writing a series of objects to a chunk
of memory (a "blob").. The allocated memory is maintained within the
blob itself, (and re-allocated by doubling when necessary).
There are also functions for reading objects from a blob as well. If
code attempts to read beyond the available memory, the read functions
return 0 values (or its moral equivalent) without reading past the
allocated memory. Once the caller is done with the reads, it can check
blob->overrun to ensure whether any invalid values were previously
returned due to attempts to read too far.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The upcoming shader cache needs this to be able to cache hash data
from the gl_shader_program structure.
Edited-by: Carl Worth <cworth@cworth.org>:
There is an internal implementation detail that the hash table
underlying the struct string_to_uint_map stores each value internally
as (value+1). The user needn't be very concerned with this (other than
knowing that a value of UINT_MAX cannot be stored) since put() adds 1
and get() subtracts 1.
So in this commit, rather than call the user's function directly with
hash_table_call_foreach, we call through a wrapper that fixes up the
off-by-one values before the caller's callback sees them.
And with this wrapper in place, we also give a better signature to the
callback function being passed to iterate(), so that this callback
function can actually expect a char* and an unsigned argument, (rather
than a couple of void* ).
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Previously, if __builtin_unreachable() was unavailable, the
unreachable macro was defined to do nothing. We do better here, by at
least still making it an assert.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This is similar to the existing functions get_instance,
get_array_instance, etc. for getting a type singleton. The new
get_sampler_instance() function will be used by the upcoming shader
cache.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, we generated this for FB writes in SIMD16 mode:
load_payload(16) vgrf5@8+0.0:F, vgrf1:F, vgrf2:F, vgrf3:F, vgrf4:F
fb_write(8) (null):UD, vgrf5@8+0.0:F 1sthalf
The LOAD_PAYLOAD's destination had its register width set to 8, and the
FB_WRITE had its execution size set to 8. This seems wrong, and while
it probably doesn't affect anything, we should fix it.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
In order to support calling lower_load_payload() inside a condition,
this patch makes OPT() a statement expression:
https://gcc.gnu.org/onlinedocs/gcc/Statement-Exprs.html
We recently did the equivalent change in the vec4 backend (commit
9b8bd67768).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
When converting to a format that has fewer bits the previous code was just
shifting off the bits. This doesn't provide very accurate results. For example
when converting from 8 bits to 5 bits it is equivalent to doing this:
x * 32 / 256
This works as if it's taking a value from a range where 256 represents 1.0 and
scaling it down to a range where 32 represents 1.0. However this is not
correct because it is actually 255 and 31 that represent 1.0.
We can do better with a formula like this:
(x * 31 + 127) / 255
The +127 is to make it round correctly.
The new code has a special case to use uint64_t when the result of the
multiplication would overflow an unsigned int. This function is inline and
only ever called with constant values so hopefully the if statements will be
folded.
The main incentive to do this is to make the CPU conversion path pick the same
values as the hardware would if it did the conversion. This fixes failures
with the ‘texsubimage pbo’ test when using the patches from here:
http://lists.freedesktop.org/archives/mesa-dev/2015-January/074312.html
v2: Use 64-bit arithmetic when src_bits+dst_bits > 32
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Rendering with a GS and then using transform feedback with a program that does
not have a GS can crash in gen6. The reason for this is that
brw_begin_transform_feedback checks brw->geometry_program to decide if there
is a GS program, but this is not correct: brw->geometry_program is updated when
issuing drawing commands, so after rendering with a GS it will be non-NULL
until we draw again with a program that does not have a GS. If the next
program uses TF, we will call glBegintransformFeedback before issuing
the drawing command and hence brw->geometry_program will be non-NULL if
the previous rendering used a GS. The right thing to do here is to check
ctx->_Shader->CurrentProgram[MESA_SHADER_GEOMETRY] instead. This is what the
gen7 code path does too.
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=87694
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
This is a rework of the liveness algorithm using a worklist as suggested by
Connor. Doing so reduces the number of times we walk over the instructions
because we don't have to do an entire pointless walk over the instructions
just to figure out it's time to stop. Also, the stuff after the last loop
in the funciton will only ever get visited once.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
A worklist is a common concept in optimizations. This adds a structure
that we can reuse for many different types of optimizations.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Changes the initial internal format of a render buffer
to GL_RGBA4 in GLES 3. This fixes a failure in the following
DrawElements test:
dEQP-GLES3.functional.state_query.rbo.renderbuffer_internal_format
Reviewed-by: Chad Versace <chad.versace@intel.com>
Previously, the set API required the user to do all of the hashing of keys
as it passed them in. Since the hashing function is intrinsically tied to
the comparison function, it makes sense for the hash set to know about
it. Also, it makes for a somewhat clumsy API as the user is constantly
calling hashing functions many of which have long names. This is
especially bad when the standard call looks something like
_mesa_set_add(ht, _mesa_pointer_hash(key), key);
In the above case, there is no reason why the hash set shouldn't do the
hashing for you. We leave the option for you to do your own hashing if
it's more efficient, but it's no longer needed. Also, if you do do your
own hashing, the hash set will assert that your hash matches what it
expects out of the hashing function. This should make it harder to mess up
your hashing.
This is analygous to 94303a0750 where we did this for hash_table
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We already have search_pre_hashed. This makes the APIs match better.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
When performing common subexpression elimination on instructions with
non-null destinations we emit a MOV to copy the result to a new
register that must have no other uses. In the case of:
cmp.g.f0.0(8) null:D, vgrf43:F, 0.500000f
...
cmp.g.f0.0(8) vgrf113:D, vgrf43:F, 0.500000f
we put the first instruction in the AEB and decided that we could reuse
its result when we found the second. Unfortunately, that meant that we'd
emit a MOV from the first's destination, which is null.
Don't do anything if the entry's destination is null and the
instruction's destination is non-null.
Tested-by: Tapani Pälli <tapani.palli@intel.com>
Ben Widawsky
Eric Anholt
Rob Clark
Tom Stellard
Kenneth Graunke
Marek Olšák
Carl Worth
Martin Peres
Tapani Pälli
Micah Fedke
Kristian Høgsberg
Vinson Lee
Neil Roberts
Iago Toral Quiroga
Jason Ekstrand
Brian Paul
Mike Mason
Matt Turner