ARM Skia NEON patches - 12 - S32_Blend

ARM Skia NEON patches - 12 - S32_Blend

Uncommenting the two prefetches can give a big speedup on A9-based
targets but has a measurable cost on A15-based ones.

Blitrow32: S32_Blend fix and little speed improvement

- the results are now exactly similar to the C code ones
- the speed has improved, especially for small values of count

+-------+-----------+------------+
| count | Cortex-A9 | Cortex-A15 |
+-------+-----------+------------+
| 1     | +30%      | +18%       |
+-------+-----------+------------+
| 2     | 0         | 0          |
+-------+-----------+------------+
| 4     | - <1%     | +14%       |
+-------+-----------+------------+
| > 4   | -0.5..+5% | -0.5..+4%  |
+-------+-----------+------------+

BUG=

[djsollen, 2013-07-15 12:52:22]

https://codereview.chromium.org/18173012/diff/1/src/opts/SkBlitRow_opts_arm_n...
File src/opts/SkBlitRow_opts_arm_neon.cpp (right):

https://codereview.chromium.org/18173012/diff/1/src/opts/SkBlitRow_opts_arm_n...
src/opts/SkBlitRow_opts_arm_neon.cpp:709: //__builtin_prefetch(src+32);
can you document in the code what you put in your CL description regarding these
prefetch commands.

[kevin.petit.not.used.account, 2013-07-15 13:55:59]

https://codereview.chromium.org/18173012/diff/1/src/opts/SkBlitRow_opts_arm_n...
File src/opts/SkBlitRow_opts_arm_neon.cpp (right):

https://codereview.chromium.org/18173012/diff/1/src/opts/SkBlitRow_opts_arm_n...
src/opts/SkBlitRow_opts_arm_neon.cpp:709: //__builtin_prefetch(src+32);
On 2013/07/15 12:52:22, djsollen wrote:
> can you document in the code what you put in your CL description regarding
these
> prefetch commands.

I'm not sure I understand. There's already a comment just under the prefetches.
Also, given that you were quite happy with the compromise in a previous patch,
you might want to try the code with the prefetches uncommented...

[kevin.petit.not.used.account, 2013-07-30 13:28:10]

ping

[djsollen, 2013-07-31 13:49:02]

I got similar perf numbers when running on my devices as well, so it looks good
from that perspective.  However, we do need to clean up the warnings generated
by the trybots before we submit it.

[kevin.petit.not.used.account, 2013-07-31 14:44:35]

On 2013/07/31 13:49:02, djsollen wrote:
> I got similar perf numbers when running on my devices as well, so it looks
good
> from that perspective.  However, we do need to clean up the warnings generated
> by the trybots before we submit it.

What kind of warnings are you observing?

[djsollen, 2013-07-31 15:15:20]

If you click on the trybot results you can see the compiler failures due to
-werror being turned on.

[11:31:06.276221] ../../../src/opts/SkBlitRow_opts_arm_neon.cpp: In function
'void S32_Blend_BlitRow32_neon(SkPMColor*, const SkPMColor*, int, U8CPU)':
[11:31:06.276266] ../../../src/opts/SkBlitRow_opts_arm_neon.cpp:737:88: error:
'vsrc' may be used uninitialized in this function [-Werror=uninitialized]
[11:31:06.276315] ../../../src/opts/SkBlitRow_opts_arm_neon.cpp:733:23: note:
'vsrc' was declared here
[11:31:06.276360] ../../../src/opts/SkBlitRow_opts_arm_neon.cpp:738:88: error:
'vdst' may be used uninitialized in this function [-Werror=uninitialized]
[11:31:06.276406] ../../../src/opts/SkBlitRow_opts_arm_neon.cpp:733:29: note:
'vdst' was declared here
[11:31:06.276451] cc1plus: all warnings being treated as errors

[kevin.petit.not.used.account, 2013-07-31 16:13:06]

I've uploaded a new patch to fix this and fixed a mistake in the description.

Have you tested with or without the prefetches?

[kevin.petit.not.used.account, 2013-08-07 10:30:56]

I've had a look at the trybot results. The mismatches are expected because the
existing code, contrary to the one in this patch, wasn't a perfect match for the
C code. However, there's nothing I can say about the GL crash on the Nexus10.

[kevin.petit.not.used.account, 2013-08-21 15:53:32]

ping

[kevin.petit.not.used.account, 2014-01-29 16:39:13]

I've added mismatching GMs to ignored-tests.txt. This change should now be able
to pass the trybots.

[kevin.petit.not.used.account, 2014-02-10 14:44:56]

On 2014/01/29 16:39:13, kevin.petit wrote:
> I've added mismatching GMs to ignored-tests.txt. This change should now be
able
> to pass the trybots.

Moved to https://codereview.chromium.org/158973002 following the change of my
email address.