Speed up rounding to byte test

Speed up rounding to byte test

R=harryjin@google.com
BUG=libyuv:492
Committed: https://chromium.googlesource.com/libyuv/libyuv/+/f4c1ac10f02bb7346737bfcfd592ddfe866f2cf1

[harryjin, 2015-10-02 18:28:32]

lgtm

[fbarchard1, 2015-10-02 18:39:40]

Profiler shows the actual unittest is half the bottleneck of this test,
generating floats to test against.

3.88 │ 30:   cmp    $0x100,%eax
  0.28 │       cmovge %r12d,%edx
  5.60 │       or     0xc(%rsp),%edx
  3.60 │       add    $0x1,%ebx
       │       cmp    %ebp,%ebx
  0.03 │       mov    %edx,0xc(%rsp)
  3.85 │     ↓ je     8b
  0.05 │ 48:   mov    %ebx,%eax
  2.17 │       sar    %eax
  1.58 │       test   $0x1,%bl
       │     ↓ je     53
  1.76 │       xor    $0xb8,%al
  1.70 │ 53:   cvtsi2 %eax,%xmm0
  3.16 │       mulsd  0x1790c9(%rip),%xmm0        # 5b5518 <.LC44>
  4.62 │       subsd  0x1790c9(%rip),%xmm0        # 5b5520 <.LC45>
 16.74 │       unpckl %xmm0,%xmm0
  6.66 │       cvtpd2 %xmm0,%xmm0
 29.52 │     → callq  lrintf@plt
  0.01 │       test   %eax,%eax
  7.36 │       mov    %eax,%edx
       │     ↑ jns    30

[brucedawson, 2015-10-02 18:48:07]

brucedawson@chromium.org changed reviewers:
+ brucedawson@chromium.org

[brucedawson, 2015-10-02 18:48:07]

https://codereview.chromium.org/1367403007/diff/1/unit_test/color_test.cc
File unit_test/color_test.cc (right):

https://codereview.chromium.org/1367403007/diff/1/unit_test/color_test.cc#new...
unit_test/color_test.cc:232: #define ROUND(f) lrintf(f)
But, lrintf is ~376 times slower on VC++
(https://connect.microsoft.com/VisualStudio/feedback/details/1179496/cq-intrin...)

[fbarchard1, 2015-10-02 21:21:20]

Using float instead of double, and branchless, inner loop of test is faster
78:   movaps %xmm1,%xmm0
  0.04 │       movss  %xmm1,0xc(%rsp)
  0.06 │     → callq  lrintf@plt
 13.75 │       cmp    $0xff,%eax
  0.03 │       movss  0xc(%rsp),%xmm1
  0.06 │       cmovg  %ebp,%eax
 14.98 │       xor    %edx,%edx
       │       addss  0x17916a(%rip),%xmm1        # 5b55b8 <.LC46>
 28.45 │       test   %eax,%eax
  0.01 │       cmovns %eax,%edx
 14.65 │       add    $0x1,%r15d
  0.07 │       or     %edx,0x1c(%rsp)
 27.90 │       cmp    %ebx,%r15d
       │     ↑ jne    78

[fbarchard1, 2015-10-02 21:25:49]

linux TestRoundToByte (2244 ms)

[fbarchard, 2015-10-02 21:35:21]

On 2015/10/02 21:25:49, fbarchard1 wrote:
> linux TestRoundToByte (2244 ms)

On Windows
Was TestRoundToByte (52299 ms)
Now TestRoundToByte (209490 ms)

with multiple changes.  rintf is slower, but test is faster.

[fbarchard1, 2015-10-02 21:41:59]

On 2015/10/02 21:35:21, fbarchard wrote:
> On 2015/10/02 21:25:49, fbarchard1 wrote:
> > linux TestRoundToByte (2244 ms)
> 
> On Windows
> Was TestRoundToByte (52299 ms)
> Now TestRoundToByte (209490 ms)
> 
> with multiple changes.  rintf is slower, but test is faster.

TestRoundToByte (1137 ms)

In 64 bit gcc, casting uses cvttss instruction so full test loop inlines
90:   movaps %xmm0,%xmm1
 22.36 │       addss  %xmm2,%xmm0
       │       addss  %xmm3,%xmm1
  0.02 │       cvttss %xmm1,%eax
 21.68 │       cmp    $0xff,%eax
       │       cmovg  %r8d,%eax
  1.08 │       test   %eax,%eax
 21.81 │       cmovs  %edi,%eax
 10.14 │       add    $0x1,%edx
       │       or     %ecx,%eax
 22.87 │       cmp    %esi,%edx
  0.02 │       mov    %eax,%ecx
  0.02 │     ↑ jne    90

[fbarchard, 2015-10-02 21:44:53]

fbarchard@chromium.org changed reviewers:
+ fbarchard@chromium.org

[fbarchard, 2015-10-02 21:44:53]

switched back to cast but add 0.5f instead of 0.5 to avoid extra double to float
conversion

Windows
TestRoundToByte (1557 ms)

https://codereview.chromium.org/1367403007/diff/1/unit_test/color_test.cc
File unit_test/color_test.cc (right):

https://codereview.chromium.org/1367403007/diff/1/unit_test/color_test.cc#new...
unit_test/color_test.cc:232: #define ROUND(f) lrintf(f)
On 2015/10/02 18:48:07, brucedawson wrote:
> But, lrintf is ~376 times slower on VC++
>
(https://connect.microsoft.com/VisualStudio/feedback/details/1179496/cq-intrin...)

Done.

[fbarchard, 2015-10-02 21:54:37]

Visual C 32 bit generated this test loop

 0F 28 EC           movaps      xmm5,xmm4
  0000008D: F3 0F 58 EA        addss       xmm5,xmm2
  00000091: F3 0F 2C C5        cvttss2si   eax,xmm5
  00000095: 3D FF 00 00 00     cmp         eax,0FFh
  0000009A: 89 C2              mov         edx,eax
  0000009C: 0F 4F D7           cmovg       edx,edi
  0000009F: 85 C0              test        eax,eax
  000000A1: 0F 48 D6           cmovs       edx,esi
  000000A4: F3 0F 58 E3        addss       xmm4,xmm3
  000000A8: 09 D3              or          ebx,edx
  000000AA: 49                 dec         ecx
  000000AB: 75 DD              jne         0000008A

[fbarchard, 2015-10-02 22:20:16]

Visual C 32 bit
TestRoundToByte (1179 ms)

Visual C 64 bit
TestRoundToByte (1419 ms)

clangcl 32 bit
TestRoundToByte (1571 ms)

clangcl 64 bit
TestRoundToByte (1546 ms)

[fbarchard1, 2015-10-02 22:27:16]

Committed patchset #4 (id:60001) manually as
f4c1ac10f02bb7346737bfcfd592ddfe866f2cf1 (presubmit successful).