Speed up vertical filtering using v = a+(b-a)*x formula...

Speed up vertical filtering using v = a+(b-a)*x formula
BUG=42064
TEST=unittests should still pass

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=45265

[fbarchard, 2010-04-21 10:43:06]

Allow vertical and horizontal to be separately filtered.
Unravel MMX avoiding compiler bug on movd and allowing better scheduling for
Atom.
Using single multiple instead of 2 for better pipelining on Atom.

Disassembly of SSE2 code (simpliest)
Old code produced this
50: F3 0F 7E 02        movq        xmm0,mmword ptr [edx]
54: F3 0F 7E 26        movq        xmm4,mmword ptr [esi]
58: 66 0F 60 C1        punpcklbw   xmm0,xmm1
5C: 66 0F D5 C3        pmullw      xmm0,xmm3
60: 66 0F 60 E1        punpcklbw   xmm4,xmm1
64: 66 0F D5 E2        pmullw      xmm4,xmm2
68: 66 0F FD C4        paddw       xmm0,xmm4
6C: 66 0F 71 D0 08     psrlw       xmm0,8
71: 66 0F 67 C0        packuswb    xmm0,xmm0
75: 66 0F D6 00        movq        mmword ptr [eax],xmm0
79: 83 C0 08           add         eax,8
7C: 83 C2 08           add         edx,8
7F: 83 C6 08           add         esi,8
82: 3B C1              cmp         eax,ecx
84: 72 CA              jb          00000050
15 instructions

New code produces this:
34: F3 0F 7E 01        movq        xmm0,mmword ptr [ecx]
38: F3 0F 7E 1A        movq        xmm3,mmword ptr [edx]
3C: 66 0F 60 C1        punpcklbw   xmm0,xmm1
40: 66 0F 60 D9        punpcklbw   xmm3,xmm1
44: 66 0F F9 D8        psubw       xmm3,xmm0
48: 66 0F D5 DA        pmullw      xmm3,xmm2
4C: 66 0F 71 E3 08     psraw       xmm3,8
51: 66 0F FD C3        paddw       xmm0,xmm3
55: 66 0F 67 C0        packuswb    xmm0,xmm0
59: 66 0F D6 00        movq        mmword ptr [eax],xmm0
5D: 83 C0 08           add         eax,8
60: 83 C1 08           add         ecx,8
63: 83 C2 08           add         edx,8
66: 3B C6              cmp         eax,esi
68: 72 CA              jb          00000034
15 instructions

[Sergey Ulanov, 2010-04-21 18:03:06]

LGTM