Previously we didn't emit MAD instructions since they cannot take
immediate arguments, but with the opt_combine_constants() pass we can
handle this properly.
total instructions in shared programs: 5920017 -> 5733278 (-3.15%)
instructions in affected programs: 3625153 -> 3438414 (-5.15%)
helped: 22017
HURT: 870
GAINED: 91
LOST: 49
Without constant pooling, this patch is a complete loss:
total instructions in shared programs: 5912589 -> 5987888 (1.27%)
instructions in affected programs: 3190050 -> 3265349 (2.36%)
helped: 1564
HURT: 17827
GAINED: 27
LOST: 101
And since the constant pooling patch by itself hurt a bunch of things,
from before constant pooling to this patch the results are:
total instructions in shared programs: 5895414 -> 5747946 (-2.50%)
instructions in affected programs: 3617993 -> 3470525 (-4.08%)
helped: 20478
HURT: 4469
GAINED: 54
LOST: 146
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
And then the opt_combine_constants() pass will pull them out into
registers. This will allow us to do some algebraic optimizations on MAD
and LRP.
total instructions in shared programs: 5946656 -> 5931320 (-0.26%)
instructions in affected programs: 778247 -> 762911 (-1.97%)
helped: 3780
HURT: 6
GAINED: 12
LOST: 12
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The fs_visitor's dump_instruction() implementation calls cfg_t()
indirectly through calculate_live_intervals, so if you have an infinite
loop in the CFG code, you can't call cfg::dump(fs_visitor *) to debug
it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
To insert an instruction at the end of a basic block, we typically do
something like
inst = block->last_non_control_flow_inst();
inst->insert_after(block, new_inst);
But blocks can consist of a single control flow instruction, so inst
will actually be the exec_list's head sentinel. We shouldn't use it as
if it were a regular instruction, but it is safe to insert something after
it.
This patch avoids assert-failing because an exec_list sentinel wasn't in
the basic block's instruction list.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The macro is defined to provide a trailing ; so this caused the expansion
to end in ";;" which made the Solaris Studio compilers issue warnings for
every line of:
"builtin_type_macros.h", line 113: Warning: extra ";" ignored.
for every file that included the header, filling build logs with thousands
of useless warnings.
Signed-off-by: Alan Coopersmith <alan.coopersmith@oracle.com>
Cc: "10.5" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
opt_copy_propagation and opt_copy_propagation_elements create new ACP
and Kill sets each time they enter a new control flow block. For if
blocks, they also copy the entire existing ACP set contents into the
new set.
When we exit the control flow block, we discard the new sets. However,
we weren't freeing them - so they lived on until the pass finished.
This can waste a lot of memory (57MB on one pessimal shader).
This patch makes the pass allocate ACP entries using this->acp as the
memory context, and Kill entries out of this->kill. It also steals
kill entries when moving them from the inner kill list to the parent.
It then frees the lists, including their contents.
v2: Move ralloc_free(this->acp) just before this->acp = orig_acp
(suggested by Eric Anholt).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: "10.5 10.4" <mesa-stable@lists.freedesktop.org>
When we schedule an instructions with undefined value, we
eventually will use 0, which is a constant, however sb wasn't
taking this into account and creating ops with illegal scalar
swizzles.
this replaces my fix for op3 in t slots.
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Matt Turner noticed that the hardware has always had a MIN
instruction, but the driver always used MAX+MOV for no
apparent reason.
This should cut an instruction, and a temporary, allowing
more programs to run in hardware.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt Turner noticed that the hardware has always had a MIN
instruction, but the driver always used MAX+MOV for no
apparent reason.
This should cut an instruction, and a temporary, allowing
more programs to run in hardware.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
I used this a while back when debugging GPU hangs, and it seems like it
could be useful, so I figured I'd add it so people can use it in the
debugger.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Sandybridge requires the post-sync non-zero workaround in a ton of
places, and if you ever miss one, the GPU usually hangs.
Currently, we try to track exactly when a workaround flush is
necessary (via the brw->batch.need_workaround_flush flag). This is
tricky to get right, and we've botched it several times in the past.
This patch unconditionally performs the post-sync non-zero flush at the
start of each primitive's state upload (including BLORP). We drop the
needs_workaround_flush flag, and drop all the other callers, as the
flush has already been performed.
We have no data to indicate that simply flushing all the time will
hurt performance, and it has the potential to help stability.
v2: Add post-sync workaround to initial GPU state upload to be extra
cautious (suggested by Chad Versace).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Previously array textures were not working with GetCompressedTextureImage,
leading to failures in the test
arb_direct_state_access/getcompressedtextureimage.c.
Tested-by: Laura Ekstrand <laura@jlekstrand.net>
Reviewed-by: Brian Paul <brianp@vmware.com>
Cc: "10.4, 10.5" <mesa-stable@lists.freedesktop.org>
Just remove the _mesa_free_lighting_data function. The body has been
empty since the shine table was moved into the tnl module (commit
ba1d921).
main/light.c:1216:46: warning: unused parameter 'ctx' [-Wunused-parameter]
_mesa_free_lighting_data( struct gl_context *ctx )
^
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
delete_management.c:56:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (i = 0; i < size; i++) {
^
delete_management.c:69:27: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (i = size - 100; i < size; i++) {
^
delete_management.c:79:31: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
assert(key_value(entry->key) >= size - 100 &&
^
delete_management.c:79:70: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
assert(key_value(entry->key) >= size - 100 &&
^
insert_many.c:56:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (i = 0; i < size; i++) {
^
insert_many.c:62:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (i = 0; i < size; i++) {
^
insert_many.c:67:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
assert(ht->entries == size);
^
random_entry.c:62:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
for (i = 0; i < size; i++) {
^
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
glcpp/glcpp.c:124:1: warning: ‘static’ is not at beginning of declaration [-Wold-style-declaration]
const static struct option
^
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
+ minor indentation fixes
Discovered by Axel Davy.
This can't be reproduced with any app, because all state trackers set a DSA
state first.
Cc: 10.5 10.4 10.3 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Axel Davy <axel.davy@ens.fr>
It's not possible to query the current buffer binding, because the extension
doesn't define GL_..._BUFFER__BINDING_AMD.
Drivers should check the target parameter of Drivers.BufferData. If it's
equal to GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, the memory should be pinned.
That's all there is to it.
A piglit test is on the piglit mailing list.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
everytime I open this file in emacs with show trailing whitespace
or git add from it my screen flares with red.
Just do a general cleanup, makes working on fp64 support not as
jarring.
I'm not saying this is perfect, its just better than before.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The short version: we need to set bits in R0.7 which provide a mask to be used
for PS kill samples/pixels. Since the VS has no such concept, we just need to
set all 1.
The longer version...
Execution for SIMD8 atomics is defined as follows:
SIMD8: The low 8 bits of the execution mask are ANDed with 8 bits of the
Pixel/Sample Mask from the message header. For the typed messages, the Slot
Group in the message descriptor selects either the low or high 8 bits. For the
untyped messages, the low 8 bits are always selected. The resulting mask is used
to determine which slots are read into the destination GRF register (for read),
or which slots are written to the surface (for write). If the header is not
present, only the low 8 bits of the execution mask are used.
The message header for untyped messages is defined in R0.7 "This field contains
the 16-bit pixel/sample mask to be used for SIMD16 and SIMD8 messages. All 16
bits are used for SIMD16 messages. For typed SIMD8 messages, Slot Group selects
which 8 bits of this field are used. For untyped SIMD8 messages, the low 8 bits
of this field are used." Furthermore, "The message header for the untyped
messages only needs to be delivered for pixel shader threads, where the
execution mask may indicate pixels/samples that are enabled only due to
derivative (LOD) calculations, but the corresponding slot on the surface must
not be accessed." We're not using a pixel shader here, but AFAICT, this mask is
used for all stages.
This leaves two options, Remove the header, or make the VS code emit the correct
thing for the header. I believe one of the goals of using SIMD8 VS was to get as
much code reuse as possible, and so I chose the latter. Since the VS has no such
thing as kill instructions, the mask is derived simple as all 1's.
v2:
Add a comment to the code (stolen from Curro on the mailing list)
Change the control flow style (Curro + Jason)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=87258
Cc: Kristian Høgsberg <krh@bitplanet.net>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
When restoring the current state in _mesa_meta_end it was previously trying to
copy the on-going sample count of the current occlusion query into the new
query after restarting it so that the driver will continue adding to the
previous value. This wouldn't work for two reasons. Firstly, the query might
not be ready yet so the Result member will usually be zero. Secondly the saved
query is stored as a pointer to the query object, not a copy of the struct, so
it is actually restarting the exact same object. Copying the result value is
just copying between identical addresses with no effect. The call to
_mesa_BeginQuery will have always reset it back to zero.
This patch fixes it by making it actually wait for the query object to be
ready before grabbing the previous result. The downside of doing this is that
it could introduce a stall but I think this situation is unlikely so it might
not matter too much. A better solution might be to introduce a real
suspend/resume mechanism to the driver interface. This could be implemented in
the i965 driver by saving the depth count multiple times like it does in the
i945 driver.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88248
Reviewed-by: Carl Worth <cworth@cworth.org>
Cc: "10.5" <mesa-stable@lists.freedesktop.org>
This line was removed by accident in commit
16b9112574 causing a regression in the
ES3-CTS.gtf.GL3Tests.shadow.shadow_execution_vert Khronos conformance
test. It's necessary because the swizzle_result() code below expects
all four components of the vector to be valid.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89094
Tested-by: Lu Hua <huax.lu@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt Turner