Speed up GrBinHashKey computation

Replace Adler32 hash with One-at-a-Time hash, do 32b-wide operations instead of 8b. Approximately 6x speedup for this function in CPU profiles of FishIETank on Linux.

[reed1, 2011-06-01 19:20:31]

LGTM

http://codereview.appspot.com/4536100/diff/1/gpu/src/GrBinHashKey.h
File gpu/src/GrBinHashKey.h (left):

http://codereview.appspot.com/4536100/diff/1/gpu/src/GrBinHashKey.h#oldcode142
gpu/src/GrBinHashKey.h:142: GrAssert(fPass);
trivial. how about

SkAlign4(len) == len

so we can't accidentally forget one of the '='

http://codereview.appspot.com/4536100/diff/1/gpu/src/GrBinHashKey.h
File gpu/src/GrBinHashKey.h (right):

http://codereview.appspot.com/4536100/diff/1/gpu/src/GrBinHashKey.h#newcode159
gpu/src/GrBinHashKey.h:159: register uint32_t a (fA);
Don't think we use the register keyword anywhere in skia (except for inline
assem), and we don't initialize ints like that.

uint32_t a = fA;

http://codereview.appspot.com/4536100/diff/1/gpu/src/GrBinHashKey.h#newcode160
gpu/src/GrBinHashKey.h:160: while (len >= 4) {
a+= *dataToAdd++;

That pattern is probably more common in skia.

[TomH, 2011-06-01 19:28:00]

Committed in revision 1472.

[junov1, 2011-06-01 19:35:48]

http://codereview.appspot.com/4536100/diff/5001/gpu/src/GrBinHashKey.h
File gpu/src/GrBinHashKey.h (right):

http://codereview.appspot.com/4536100/diff/5001/gpu/src/GrBinHashKey.h#newcod...
gpu/src/GrBinHashKey.h:160: while (len >= 4) {
Maybe I am missing something, but it looks to me like the last len%4 bytes will
be ignored by this hash function.  I am pretty sure this also fails the unit
test.

[TomH, 2011-06-01 19:39:54]

You're right, that should fail the unit test.

In practice, this seems to only be called with 32b-aligned 32b-sized data, so
len is divisible by 4 (and there's an assert to verify the size).

[TomH, 2011-06-01 20:18:27]

Calls GrIsALIGN4 instead of SkAlign4, fixing the debug compile problem.
Partial patch for the unit tests, but on Linux they appear to be (1) not easily
runnable and (2) previously otherwise broken.

[junov1, 2011-06-02 15:00:30]

On 2011/06/01 20:18:27, TomH wrote:
> Calls GrIsALIGN4 instead of SkAlign4, fixing the debug compile problem.
> Partial patch for the unit tests, but on Linux they appear to be (1) not
easily
> runnable and (2) previously otherwise broken.

LGTM