remove meaningless matrix benches, add mapPts() and add new benches

remove meaningless matrix benches, add mapPts() and add new benches

mapPts definitely faster than mapPoints (identity and perspective same speed). Up to 3x for large values of N.

abandoned for https://codereview.chromium.org/1030653002

[reed2, 2015-03-22 02:56:44]

The CQ bit was checked by reed@chromium.org

[commit-bot: I haz the power, 2015-03-22 02:57:01]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1031443002/1

[commit-bot: I haz the power, 2015-03-22 02:57:01]

Note for Reviewers:
The CQ is waiting for an approval.
If you believe that the CL is not ready yet, or if you would like to L-G-T-M
with comments then please uncheck the CQ checkbox.

Waiting for LGTM from valid reviewer(s) till 2015-03-22 08:57 UTC

[reed2, 2015-03-22 02:58:43]

reed@chromium.org changed reviewers:
+ caryclark@google.com, mtklein@google.com, reed@google.com

[reed2, 2015-03-22 02:58:44]

Two ideas:

1. I think I need a swizzle of ABCD -> BADC to handle the affine case, unless
you guys see how to do it differently.

2. Seems like Sk8f might help if it actually existed (but only for large numbers
of points).

[reed2, 2015-03-22 03:02:55]

Another question (looking ahead to mapXY ...

Is there a faster way to convert int -> float (for 2 ints at a time in my case)?

I do a lot of the following inside shaders...

mapXY( SkIntToScalar(x) + 0.5f, SkIntToScalar(y) + 0.5f, ...); 

Inside mapXY, I will try using Sk2s, but all of the callers spend time
converting their 2 ints first...

[reed2, 2015-03-22 03:03:33]

reed@chromium.org changed reviewers:
+ msarett@google.com

[commit-bot: I haz the power, 2015-03-22 08:57:04]

The CQ bit was unchecked by commit-bot@chromium.org

[commit-bot: I haz the power, 2015-03-22 08:57:06]

No LGTM from a valid reviewer yet. Please ask for an LGTM from a
full Skia committer from
https://skia.googlesource.com/skia/+/master/CQ_COMMITTERS

[mtklein, 2015-03-23 14:57:08]

On 2015/03/22 03:02:55, reed2 wrote:
> Another question (looking ahead to mapXY ...
> 
> Is there a faster way to convert int -> float (for 2 ints at a time in my
case)?
> 
> I do a lot of the following inside shaders...
> 
> mapXY( SkIntToScalar(x) + 0.5f, SkIntToScalar(y) + 0.5f, ...); 
> 
> Inside mapXY, I will try using Sk2s, but all of the callers spend time
> converting their 2 ints first...

Yes, two ways: I add Sk2x::LoadInts(const int[2]) / Sk2x::storeInts(int[2]), or
I add Sk2i and Sk2x::cast().

Given that we've already tried the Sk4i approach, I'd be interested in seeing
how the more limited LoadInts/storeInts fares first.  A parallel int type is a
bit more of a maintenance burden.

[mtklein, 2015-03-23 15:00:17]

On 2015/03/22 02:58:44, reed2 wrote:
> Two ideas:
> 
> 1. I think I need a swizzle of ABCD -> BADC to handle the affine case, unless
> you guys see how to do it differently.


> 2. Seems like Sk8f might help if it actually existed (but only for large
numbers
> of points).

The tech to back Sk8f more efficiently than 2x Sk4f, AVX, is pretty implausible
to be available at compile time, but it's easy to write.
We can absolutely just back Sk8f with 2x Sk4f if it's semantically useful to
have an Sk8f.

[mtklein, 2015-03-23 15:17:49]

On 2015/03/22 02:58:44, reed2 wrote:
> Two ideas:
> 
> 1. I think I need a swizzle of ABCD -> BADC to handle the affine case, unless
> you guys see how to do it differently.

This is an interesting one.   I've been meaning to look into which swizzles we
can support efficiently on both platforms.

I think this is the list of swizzles that are fast on both SSE and NEON:
   zip on two registers: ABCD, XYZW -> AXBY and/or CZDW
   zip on one register:  ABCD -> AABB and/or CCDD 
  unzip on two registers: ABCD, XYZW -> ACXZ and/or BDYW
  unzip on one register: ABCD -> AACC and/or BBDD

On SSE, the "and" is more expensive than "or", but on NEON you generally get
both together.

We can do ABCD -> BADC by unzip(ABCD) -> AACC,BBDD , then zip(BBDD, AACC) ->
BADC.  I think that'd be 2 ops on NEON, 3 on SSE.

There is another "transpose" swizzle on NEON I need to figure out.  That may be
what we want in fewer steps.