2015-06-24 23:05 GMT+02:00 Davin McCall :
> Hi - I'm new here.
>
> I've recently started poking the Mesa codebase for little reason other than
> personal interest. In the "help wanted" section of the website it mentions
> aliasing violations as a target for newcomers to fix, so with that in mind
>
So its time for me to post the current state of my loop analysis pass.
So where can you find this? You can find it in a branch on github[1],
if you want a more reading-friendly experience.
(Oh, and don't bother with looking at the commit history.)
I've also pasted to code as-is below, as I've yet
2015-07-12 1:17 GMT+02:00 Timothy Arceri :
> Currently only one ir assignment is removed for each var in a single
> dead code optimisation pass. This means if a var has more than one
> assignment, then it requires all the glsl optimisations to be run again
> for each additional assignment to be rem
2015-07-14 9:46 GMT+02:00 Iago Toral Quiroga :
> From: Samuel Iglesias Gonsalvez
>
> Signed-off-by: Samuel Iglesias Gonsalvez
> ---
> src/glsl/lower_ubo_reference.cpp | 179
> +++
^ This is 80 characters wide in this case(for reference), so we're blowing
the
Hi Matt,
Just a small nitpick I noticed;
s/promition/promotion in the commit message.
-Thomas
2015-07-14 1:22 GMT+02:00 Matt Turner :
> In the process of looking at the assembly generated from the OUT_BATCH
> series, I noticed a bunch of float <-> double round trips. Mostly this
> arises because
them soak
on the list for some comments while I go head-to-head with
the loop unrolling. Happy reviewing =)
Thomas Helland (5):
nir/worklist: Split out the type-independent parts
nir: Add a worklist for SSA definitions
nir: Move nir_instr_get_dest_ssa_def to nir.c / nir.h
nir: Add a helper
Now, nir_block_worklist is just a thin wraper around nir_worklist.
We can now use nir_worklist also for SSA definitions.
Signed-off-by: Connor Abbott
V2: Correct comments in more places
Add file to makefile
Fix EOF blank line
idx -> index
Signed-off-by: Thomas Helland
---
We were already doing this internally for iterating over a function
implementation, so just expose it directly.
Signed-off-by: Connor Abbott
---
src/glsl/nir/nir.c | 7 +++
src/glsl/nir/nir.h | 3 +++
2 files changed, 10 insertions(+)
diff --git a/src/glsl/nir/nir.c b/src/glsl/nir/nir.c
ind
This was originally located in nir_opt_cse.
Move it out for reuse.
Signed-off-by: Thomas Helland
---
This probably is not that smart. We can ofcourse roll our
own in value range pass. It was just so conventient at the time.
src/glsl/nir/nir.c | 21 +
src/glsl/nir
Signed-off-by: Connor Abbott
V2: Added entry in makefile
---
src/glsl/Makefile.sources | 1 +
src/glsl/nir/nir_ssa_def_worklist.h | 100
2 files changed, 101 insertions(+)
create mode 100644 src/glsl/nir/nir_ssa_def_worklist.h
diff --git a/src/g
ken. Weird stuff.
2015-07-15 0:08 GMT+02:00 Thomas Helland :
> The deadline for GSoC is aproaching a bit faster than I'm
> comfortable with, so I need to get this out on the list
> for some review-action. It's not yet complete, but I think
> it's getting there. At least it s
Signed-off-by: Thomas Helland
---
src/glsl/Makefile.sources |1 +
src/glsl/nir/nir.h |2 +
src/glsl/nir/nir_opt_value_range.c | 1330
3 files changed, 1333 insertions(+)
create mode 100644 src/glsl/nir/nir_opt_value_range.c
Hi Matt,
I've commented on some of your feedback down below.
The rest is taken note of and I'll be fixing it up later.
2015-07-15 19:18 GMT+02:00 Matt Turner :
> On Tue, Jul 14, 2015 at 4:29 PM, Thomas Helland
> wrote:
>> Signed-off-by: Thomas Helland
>> ---
&
oid platform
> +wayland, surfaceless, android,
> +and haiku. The android platform
> can only be built as a system component, part of AOSP, while the
> -haiku and gdi platforms can only be built with
> SCons.
> +haiku platforms can only be built with SCons.
^ s/platforms/plat
This looks a lot better.
The patch is:
Reviewed-by: Thomas Helland
2015-07-10 19:49 GMT+02:00 Emil Velikov :
> Move the code around rather than having it scattered. No functional
> change.
>
> Signed-off-by: Emil Velikov
> ---
> src/egl/drivers/dri
This patch is:
Reviewed-by: Thomas Helland
2015-07-10 19:49 GMT+02:00 Emil Velikov :
> Signed-off-by: Emil Velikov
> ---
> src/egl/drivers/dri2/platform_wayland.c | 26 +-
> 1 file changed, 13 insertions(+), 13 deletions(-)
>
> diff --git a/sr
Hi Emil,
I looked at the whole series, but didn't really feel qualified
to review the whole bunch. I've sent reviews on the patches
I felt I had a good take on. If you don't get anyone else to
review the remainder then shout out and I'll give it a shot.
- Thomas
2015-07-17 22:57 GMT+02:00 Emil V
2015-07-21 14:12 GMT+02:00 Emil Velikov :
> On 20/07/15 21:25, Chad Versace wrote:
>> On Fri 17 Jul 2015, Emil Velikov wrote:
>>> On 17 July 2015 at 19:11, Ilia Mirkin wrote:
On Fri, Jul 17, 2015 at 2:11 PM, Eric Anholt wrote:
> Matt Turner writes:
>
>> On Fri, Jul 17, 2015 at 1
2015-10-25 13:58 GMT+01:00 Rob Clark :
> On Sun, Oct 25, 2015 at 8:05 AM, Erik Faye-Lund wrote:
>> On Sat, Oct 24, 2015 at 7:08 PM, Rob Clark wrote:
>>> From: Rob Clark
>>>
>>> +#define NIR_PASS_PROGRESS(pass, nir, ...) ({ \
>>> + assert(nir_shader_is_mutable(nir));
Hi all,
Just thought I'd inform you of my experiments and findings this weekend.
This summer I discussed a possible optimization pass with Connor that
I've simply named "pull-through" for the lack of a better name.
This tries to reorder operations of the type
min(sqrt(a), sqrt(b)) into sqrt(min(
2015-11-09 12:55 GMT+01:00 Iago Toral :
> Hi,
>
> Currently, NIR defines vecN operations as unsigned (integer). The fp64
> patches from Connor change this to float (I guess because we need to
> know the case where we are packing vectors of 64-bit floats). However,
> this makes it so that nir_lower
sert: 4.13 % --> 2.13 %
hash_table_search: 4.12 % --> 2.29 %
To be done:
Do the same for hash set.
Change some hash table uses to the one in util.
Similar treatment for the other tables.
Test on something that is actually CPU bound.
-Thomas Helland
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
On Feb 26, 2015 4:32 PM, "Emil Velikov" wrote:
>
> On 26/02/15 01:30, Brian Paul wrote:
> > Another clean-up series. Mostly trimming down on unneeded #includes.
> >
> > I tested DRI, non-DRI and MSVC builds of this series but it's possible
> > I've missed some configuration and broke something.
asher
[2]
http://stackoverflow.com/questions/17645167/implementing-sse-4-2s-crc32c-in-software
Thomas Helland (8):
util: Change hash_table to use quadratic probing.
util: Change table size to be power of two
glsl: Change ir_const_expr to use util/hash_table
util: Change util/set to use quad
Should decrease collisions and therefore improve performance.
Result from oprofile on a run of shader-db shows:
mesa_hash_data 3.23 ---> 3.25
hash_table_insert 2.71 ---> 2.52
hash_table_search 2.87 ---> 2.64
set_add 1.69 ---> 1.65
set_search 2.11 -
Less code, and gives a small speedup due to
avoiding the extra modulo operation.
Also less memory footprint since rehash gets dropped.
Results from oprofile on a shader-db run:
set_add 2.80 ---> 2.70
set_search 2.86 ---> 2.74
runtime 162 ---> 165
---
Should be more efficient as we can use a bitmask
operation instead of an expensive modulo.
Results from oprofile on a shader-db run:
set_add 2.70 ---> 1.69
set_search 2.74 ---> 2.11
runtime 165 ---> 157
---
src/util/set.c | 82
This gives better performance as we can do bitmasking
instead of modulo operations.
This reduces oprofile hits on a shader-db run accordingly:
hash_table_insert 4.28 ---> 2.57
hash_table_search 4.59 ---> 2.67
runtime 175 ---> 170
Since the last patch hits assertion fa
This should give better cache locality, less memory consumption,
less code, and should also be faster since we avoid a modulo operation.
This is not the quadratic probing function you see most places.
They do not accumulate, so you try hash +1, +4, +9, etc.
My code accumulates; so it becomes hash
Copy-pasta from the wikipedia article.
Results from oprofile on a shader-db run:
mesa_hash_data 3.25 ---> 3.11
hash_table_insert 2.52 ---> 2.52
hash_table_search 2.64 ---> 2.64
set_add 1.65 ---> 1.74
set_search 2.07 ---> 2.08
runtime
This is faster than program_hash_table,
and we don't need the extra capabilities.
Results from oprofile on a shader-db run:
mesa_hash_data 3.09 ---> 3.05
hash_table_insert 2.57 ---> 2.58
hash_table_search 2.67 ---> 2.72
runtime 170 ---> 162
---
src/glsl/ir_co
There are probably better ways to do this.
Results from oprofile on a shader-db run:
mesa_hash_data 3.11 ---> 3.12
hash_table_insert 2.52 ---> 2.50
hash_table_search 2.64 ---> 2.59
set_add 1.74 ---> 1.72
set_search 2.08 ---> 2.09
runtime
Forgot to run piglit on this before sending out.
And well, it looks bad, really bad.
I'm looking into it now.
Will report back when I get some more data.
Regards,
Thomas
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.or
This is the commit that regresses piglit.
With this reverted there are no regressions.
Regards,
Thomas
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
8
GAINED:0
LOST: 0
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir_opt_algebraic.py | 2 ++
1 file changed, 2 insertions(+)
diff --git a/src/glsl/nir/nir_opt_algebraic.py
b/src/glsl/nir/nir_opt_algebraic.py
index 7bf6431..688ceff 100644
--- a/
0
LOST: 0
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir_opt_algebraic.py | 2 ++
1 file changed, 2 insertions(+)
diff --git a/src/glsl/nir/nir_opt_algebraic.py
b/src/glsl/nir/nir_opt_algebraic.py
index 688ceff..ef855aa 100644
--- a/
On Feb 28, 2015 8:39 PM, "Jason Ekstrand" wrote:
>
> Both patches are
>
> Reviewed-by: Jason Ekstrand
Could you commit them?
I don't have commit access.
Regards,
Thomas
>
> On Sat, Feb 28, 2015 at 11:32 AM, Thomas Helland <
thomashellan...@gmail.com> w
On Feb 28, 2015 5:05 PM, "Jason Ekstrand" wrote:
>
>
> On Feb 28, 2015 4:55 AM, "Thomas Helland"
wrote:
> >
> > So here comes my hash-table series mentioned earlier.
> >
> > So, first of all, there's some issues.
> > I've been
2015-02-28 17:05 GMT+01:00 Jason Ekstrand :
>
> On Feb 28, 2015 4:55 AM, "Thomas Helland" wrote:
>>
>> So here comes my hash-table series mentioned earlier.
>>
>> So, first of all, there's some issues.
>> I've been strugling with hitting ass
Humble ping?
2015-02-28 21:19 GMT+01:00 Matt Turner :
> On Sat, Feb 28, 2015 at 11:47 AM, Thomas Helland
> wrote:
>> On Feb 28, 2015 8:39 PM, "Jason Ekstrand" wrote:
>>>
>>> Both patches are
>>>
>>> Reviewed-by: Jason Ekstrand
>>
---
src/compiler/glsl/opt_constant_variable.cpp | 6 --
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/src/compiler/glsl/opt_constant_variable.cpp
b/src/compiler/glsl/opt_constant_variable.cpp
index 1c06ffe675..a07d64067b 100644
--- a/src/compiler/glsl/opt_constant_variable.cpp
This is not a hot path, but when we have the functionality we
should take advantage of it.
---
src/util/string_to_uint_map.h | 9 ++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/src/util/string_to_uint_map.h b/src/util/string_to_uint_map.h
index e0533ec6ea..3248696e88 10064
---
src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/src/compiler/glsl/opt_copy_propagation_elements.cpp
b/src/compiler/glsl/opt_copy_propagation_elements.cpp
index 9f79fa9202..d318417f6c 100644
--- a/src/compiler/glsl/
This should prove benefitial in the common case of inserting
and not rewriting anything.
---
src/compiler/nir/nir_instr_set.c | 8 ++--
1 file changed, 6 insertions(+), 2 deletions(-)
diff --git a/src/compiler/nir/nir_instr_set.c b/src/compiler/nir/nir_instr_set.c
index 9cb9ed43e8..4c45d051f8
---
src/compiler/glsl/opt_constant_propagation.cpp | 8 +---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/src/compiler/glsl/opt_constant_propagation.cpp
b/src/compiler/glsl/opt_constant_propagation.cpp
index 4039512097..553de9c803 100644
--- a/src/compiler/glsl/opt_constant_p
This is one of our largest users of hash_table_insert. This causes some
code duplication and a parameter to pass in, so might not be worth it.
---
src/mesa/program/symbol_table.c | 20 +---
1 file changed, 13 insertions(+), 7 deletions(-)
diff --git a/src/mesa/program/symbol_table
hmark this
one will need to take care of that.
Thomas Helland (7):
mesa: Avoid unnecessary string hashing in symbol table
util: Avoid computing hash twice in string_to_uint_map
glsl: Prehash to avoid computing the hash twice
glsl: Prehash to avoid hashing two times when inserting
glsl:
_mesa_hash_table_search is one of our hottest function according to perf.
Callgrind shows the refcounting as one of the major users of the
searching functions. We can reduce the pain by prehashing, so that we
avoid hash two times when inserting in the table.
On a short shader-db run (with validati
2017-02-02 21:57 GMT+01:00 Thomas Helland :
> ---
> src/compiler/glsl/opt_constant_variable.cpp | 6 --
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/src/compiler/glsl/opt_constant_variable.cpp
> b/src/compiler/glsl/opt_constant_variable.cpp
Definitely not worth the added complexity. I'll clean up the other
patches that I have laying and post those instead, as they can actually
have a certain impact on things.
2017-02-02 21:57 GMT+01:00 Thomas Helland :
> This series converts multiple users of the hash table to reuse the
>
Add a test to exercise a worst case collision scenario
that may cause us to not be able to find an empty
slot in the table even though it is not full.
V2: Feedback from Emil Velikov
-Don't include code in the assert
---
src/util/tests/hash_table/collision.c | 14 ++
1 file changed
out it's effect though, as it sacrifices insert
speed for lookup speed, but one never knows until one tests.
Let me know if this is something I should persue. If not I'll
mark this series as "junk" in my git repo, and get on with the cleaning.
Thomas Helland (4):
util/tests: Ex
This follows the same pattern as in the hash_table.
Reviewed-by: Jason Ekstrand
---
src/util/set.c | 8 ++--
1 file changed, 2 insertions(+), 6 deletions(-)
diff --git a/src/util/set.c b/src/util/set.c
index 99c04369c5..110f182244 100644
--- a/src/util/set.c
+++ b/src/util/set.c
@@ -196,12
The same rationale applies here as for the hash table.
Power of two size should give better performance,
and using the algorithm hash = sh + i/2 + i*i/2
should result in only distinct hash values when hitting collisions.
V4: Feedback from Jason Ekstrand
- Split cleanup of set_rehash into separa
This will allow us to remove the large static table and use a
power of two hash table size that we can compute on the fly.
We can use bitmasking instead of modulo to fit our hash in the table,
and it's less code.
By using the algorithm hash = sh + i/2 + i*i/2
we are guaranteed that all retries fro
2017-02-08 21:35 GMT+01:00 Thomas Helland :
> I was cleaning up my local git repo, and came across this series.
> Last time it was discussed was all the way back in April 2015.
> Things looked pretty good back then, but we where seeing a smaller
> regression in CPU-bound scenarios a
2017-02-08 22:07 GMT+01:00 Thomas Helland :
> 2017-02-08 21:35 GMT+01:00 Thomas Helland :
>> I was cleaning up my local git repo, and came across this series.
>> Last time it was discussed was all the way back in April 2015.
>> Things looked pretty good back then, but we w
2017-02-09 1:57 GMT+01:00 Timothy Arceri :
> On Wed, 2017-02-08 at 21:35 +0100, Thomas Helland wrote:
>> I was cleaning up my local git repo, and came across this series.
>> Last time it was discussed was all the way back in April 2015.
>> Things looked pretty good back then,
The same rationale applies here as for the hash table.
Power of two size should give better performance,
and using the algorithm hash = sh + i/2 + i*i/2
should result in only distinct hash values when hitting collisions.
Collision performane is also a lot better.
V4: Feedback from Jason Ekstrand
st patch, that reworks our pointer hashing.
This patch has been tested on shader-db, but has not been tested on
other workloads, like the above mentioned games. That's on my todo-lsit.
Thomas Helland (7):
util/tests: Expand collision test for hash table
util: Change hash_table to use q
Add a test to exercise a worst case collision scenario
that may cause us to not be able to find an empty
slot in the table even though it is not full.
This hits the bug in my last revision of the series
converting the hash table to quadratic probing.
V2: Feedback from Emil Velikov
-Don't inclu
---
src/util/hash_table.c | 2 +-
src/util/set.c| 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/src/util/hash_table.c b/src/util/hash_table.c
index e1255a2484..8121c8e67a 100644
--- a/src/util/hash_table.c
+++ b/src/util/hash_table.c
@@ -85,7 +85,7 @@ _mesa_hash_tabl
This will allow us to remove the large static table and use a
power of two hash table size that we can compute on the fly.
We can use bitmasking instead of modulo to fit our hash in the table,
and it's less code.
By using the algorithm hash = sh + i/2 + i*i/2
we are guaranteed that all retries fro
This follows the same pattern as in the hash_table.
Reviewed-by: Jason Ekstrand
---
src/util/set.c | 8 ++--
1 file changed, 2 insertions(+), 6 deletions(-)
diff --git a/src/util/set.c b/src/util/set.c
index 99c04369c5..110f182244 100644
--- a/src/util/set.c
+++ b/src/util/set.c
@@ -196,12
---
src/util/hash_table.c | 2 +-
src/util/set.c| 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/src/util/hash_table.c b/src/util/hash_table.c
index a93326ec27..e1255a2484 100644
--- a/src/util/hash_table.c
+++ b/src/util/hash_table.c
@@ -83,7 +83,7 @@ _mesa_hash_tabl
Use our knowledge that pointers are at least 4 byte aligned to remove
the useless digits. Then shift by 6, 10, and 14 bits and add this to
the original pointer, effectively folding in the entropy of the higher
bits of the pointer into a 4-bit section. Stopping at 14 means we can
add the entropy fro
We were already doing this internally for iterating over a function
implementation, so just expose it directly.
Signed-off-by: Connor Abbott
---
src/glsl/nir/nir.c | 7 +++
src/glsl/nir/nir.h | 3 +++
2 files changed, 10 insertions(+)
diff --git a/src/glsl/nir/nir.c b/src/glsl/nir/nir.c
ind
e phi-nodes or other dead code that is left behind.
Signed-off-by: Thomas Helland
---
src/glsl/Makefile.sources | 1 +
src/glsl/nir/nir.h| 4 +
src/glsl/nir/nir_form_LCSSA.c | 227 ++
src/glsl/nir/nir_metadata.c | 2 +
4 files cha
I was hoping this should fix the problem with the pass
getting rerun forever, but apparently not. We still need
a solution like this though, so sending it out for some
feedback of whether or not this is the way to do it.
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir.h | 1
've solved the DFS here is they way I'm planning
on doing it for the loop analysis too, and in fact a lot of it will be
quite similar.
So if this looks OK-ish then I know I'm on the right track. I'll post the
loop-analysis patches later today when I've reworked them.
2015-0
or the other?
Or I can do a "foreach in cf node" and compare pointers.
It might be a bit hacky, but it will do the job.
TODO items:
I haven't fixed up the casting in the visitor-functions,
and I should probably rename some of them.
Signed-off-by: Thomas Helland
---
src/glsl/Make
ill be my next
big thing on the list, as I also need to do a job witht that for
the value range propagation pass I'm writing.
With that said, those are mainly mechanical changes. I don't
expect the aproach I've taken to change a lot, so I think
it is now ready to be put on the list fo
The series looks good. You could probably s/GLuint/unsigned
in patch three as it is not "in the API", but that's a nitpick.
(You're fixing warnings, and not making things worse)
Either way the series is:
Reviewed-by: Thomas Helland
Someone else needs to push these for you
und in Dota2 and Brutal Legend.
One shader is cut by 8%, from 323 -> 296 instructons in SIMD8
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir_opt_algebraic.py | 1 +
1 file changed, 1 insertion(+)
diff --git a/src/glsl/nir/nir_opt_algebraic.py
b/src/glsl/nir/nir_opt_algebraic.py
inde
These where found when going through the mysteriously
high gains I got from the VRP pass. My pass was doing
bad things, but upon inspection I found these.
First patch is benefitial, the rest shows no change
on a shader-db run on my system.
Thomas Helland (3):
nir: Simplify fne(fneg(a), a
The same rationale as in the other patches.
No shader-db changes.
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir_opt_algebraic.py | 2 ++
1 file changed, 2 insertions(+)
diff --git a/src/glsl/nir/nir_opt_algebraic.py
b/src/glsl/nir/nir_opt_algebraic.py
index 3e483fb..d005f81 100644
--- a
The positive and negative value of a float can only
be equal to each other if it is -0.0f and 0.0f.
This is safe for Nan and Inf, as -Nan != Nan, and -Inf != Inf
This gives no changes in my shader-db
Signed-off-by: Thomas Helland
---
src/glsl/nir/nir_opt_algebraic.py | 1 +
1 file changed, 1
t would not give
us any substantial benefit.
Phi-nodes might have some saying in if we are able to remove
complete nodes/blocks? If so, then we should probably revisit
this decision, but for now it seems the return-on-investment
and risk-reward ratio does not justify dealing with them.
Signed-of
2015-08-06 16:22 GMT+02:00 Boyan Ding :
> But, isn't -INT_MIN == INT_MIN?
>
> Regards,
> Boyan Ding
>
Yup. That's correct. Thanks for pointing that out.
That means this patch is wrong, so I NAK myself on this one.
Regards,
ThoHel
> 2015-08-06 19:36 GMT+08:00
2015-08-07 0:08 GMT+02:00 Matt Turner :
> (Don't top quote)
>
> On Thu, Aug 6, 2015 at 7:22 AM, Boyan Ding wrote:
>> But, isn't -INT_MIN == INT_MIN?
>
> In C (-INT_MIN) is undefined since it relies on behavior of signed
> integer overflow.
>
> In GLSL though, I think you're right. The spec says th
2015-08-06 19:22 GMT+02:00 Thomas Helland :
> It is now mostly working, and I've incorporated a bunch
> of the feedback I got on the last posting. I promised to
> have it out on the list by american office hours today,
> so here it goes.
>
> There is a massive issue w
"Members of structures cannot be declared as atomic
> counter"
> + *" types."
Unintentionally "split string" due to your editor? With the nit fixed:
Reviewed-by: Thomas Helland
> */
> YYLTYPE loc = decl_list-&
in a statistics-gathering I performed, and amount
to aproximately 90% of alu-ops on a scalar only run. If we
actually knew the input to these operations there is the
possibility that the effectiveness could be greatly improved.
A piglit-run shows no regression with or without INTEL_USE_NIR=1.
Sign
Signed-off-by: Thomas Helland
---
These are listed to the best of my knowledge, looking
at the nvidia driver dropped yesterday, and a glance at the
"what changed according to the ARB variant"
section of each extension spec. Hopefully it should not
be too far off. Feel free to leave you
Signed-off-by: Thomas Helland
---
This adds a section for the extensions nvidia has chosen to
call the "GL ARB 2015 Extensions" unveiled at SIGGRAPH.
Also some minor whitespace fixes for consistency.
We might also want to add other extensions that are not part
of the openGL core specs
2015-08-11 20:25 GMT+02:00 Eric Anholt :
> We may find a cause to do more undef optimization in the future, but for
> now this fixes up things after if flattening. vc4 was handling this
> internally most of the time, but a GLB2.7 shader that did a conditional
> discard and assign gl_FragColor in t
gt; + for (int i = 1; i <= 2; i++) {
> + if (!instr->src[i].src.is_ssa)
> + continue;
> +
> + nir_instr *parent = instr->src[i].src.ssa->parent_instr;
> + if (parent->type != nir_instr_type_ssa_undef)
> + continue;
> +
> +
wn reviewed patches. I'll then get someone to merge the
patches some time before the 11.0 branchpoint (21st August).
BR,
Thomas
> On 6 August 2015 at 18:35, Thomas Helland wrote:
>>
>> The series looks good. You could probably s/GLuint/unsigned
>> in patch three as it is n
2015-08-12 7:45 GMT+02:00 Tapani Pälli :
>
>
> On 08/11/2015 08:48 PM, Thomas Helland wrote:
>>
>> Signed-off-by: Thomas Helland
>> ---
>> These are listed to the best of my knowledge, looking
>> at the nvidia driver dropped yesterday, and a glance at th
V2: -Change to "not started" for most entries
-Add status for multisample_2d_array
-Change shader_multisample_interpolation to "not_stared"
Signed-off-by: Thomas Helland
Reviewed-by: Tapani Pälli
---
docs/GL3.txt | 25 +
1 file changed, 25
>2015-08-12 15:59 GMT+02:00 Gustaw Smolarczyk :
> Hello. I've just spotted a simple typo.
>
> 2015-08-12 15:07 GMT+02:00 Thomas Helland :
>>
>> V2: -Change to "not started" for most entries
>> -Add status for multisample_2d_array
>> -Cha
2015-08-12 16:15 GMT+02:00 Ilia Mirkin :
> On Wed, Aug 12, 2015 at 9:07 AM, Thomas Helland
> wrote:
>> V2: -Change to "not started" for most entries
>> -Add status for multisample_2d_array
>> -Change shader_multisample_interpolation to "not_star
2015-08-12 17:48 GMT+02:00 Ilia Mirkin :
> On Tue, Aug 11, 2015 at 1:48 PM, Thomas Helland
> wrote:
>> Signed-off-by: Thomas Helland
>> ---
>> This adds a section for the extensions nvidia has chosen to
>> call the "GL ARB 2015 Extensions" unveiled a
2015-08-12 18:56 GMT+02:00 Kenneth Graunke :
> On Wednesday, August 12, 2015 06:32:50 PM Thomas Helland wrote:
>> 2015-08-12 17:48 GMT+02:00 Ilia Mirkin :
>> > On Tue, Aug 11, 2015 at 1:48 PM, Thomas Helland
>> > wrote:
>> >> Signed-off-by: Thomas Helland
... that are not mandated by the current openGL specifications
Signed-off-by: Thomas Helland
Reviewed-by: Edward O'Callaghan
Reviewed-by: Tapani Pälli
---
Tapani, Edward: Does your review still stand?
I decided to be kind and not rename the file. That way we don't
break thin
On Aug 13, 2015 12:33 PM, "Marek Olšák" wrote:
>
> On Wed, Aug 12, 2015 at 11:23 PM, Thomas Helland
> wrote:
> > 2015-08-12 18:56 GMT+02:00 Kenneth Graunke :
> >> On Wednesday, August 12, 2015 06:32:50 PM Thomas Helland wrote:
> >>> 2015-08-12 17:48
Hi all,
I'm now doing the last modifications to the loop analysis pass
I've been working on, and I stumbled upon something I found
a bit surprising. (Keep in mind I'm not very familiar with GCM)
Some background info: there are 550 loops in my shader-db.
The glsl unrolling unrolls 150 of those. My
me a lot, and so it might
not be 100% correctly implmented.
Writemasks? I haven't really delt with those.
I guess I should probably do that?
I'm unsure what more there is to share to make
review simpler, but just shout out.
[1]: https://parasol.tamu.edu/~rwerger/Courses/605/ssa_wt_
t; char **name,
> (*name)[name_length] = '\0';
> this->leave_record(t, *name, row_major);
>}
> - } else if (t->is_array() && (t->fields.array->is_record()
> -|| t->fields.array->is_interface())) {
t; - */
> void GLAPIENTRY
> _mesa_StencilOpSeparate(GLenum face, GLenum sfail, GLenum zfail, GLenum
> zpass)
> {
> --
> 2.1.4
>
> ___
> mesa-dev mailing list
> mesa-dev@lists.freedesktop.org
> http://lists.freedeskto
1 - 100 of 517 matches
Mail list logo
