-fPIC buildable version of yuv for 64 bit linux.

-fPIC buildable version of yuv for 64 bit linux.
BUG=25070
TEST=add -fPIC to gcc.  Old code failed to build.  New code builds with 64 bit.


Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=29329

[fbarchard, 2009-10-16 18:55:42]

64 bit buildable with -fPIC

[awong, 2009-10-16 18:58:22]

Adding Mark to verify this doesn't break any subtle assumptions.

[awong, 2009-10-16 19:27:10]

LGTM.

Misread the change description...I thought we were thinking of adding -fPIC into
our build which we aren't.  Mark, if you want to look over, you're welcome too,
but dont' think it's critical anymore.

http://codereview.chromium.org/269106/diff/3001/3002
File media/base/yuv_row_linux.cc (right):

http://codereview.chromium.org/269106/diff/3001/3002#newcode251
Line 251: const uint8* u_buf,   // rsi
indentation is off.

Add another comment explaining tha tyou aren't directly referring to this
register mapping, but it's still valid.

http://codereview.chromium.org/269106/diff/3001/3002#newcode306
Line 306: const uint8* v_buf,   // rdx
indentation. Same thing about having comments for the mapping that you are no
longer directly relying upon in code.

[Mark Mentovai, 2009-10-16 19:30:54]

LG otherwise

http://codereview.chromium.org/269106/diff/3001/3002
File media/base/yuv_row_linux.cc (right):

http://codereview.chromium.org/269106/diff/3001/3002#newcode251
Line 251: const uint8* u_buf,   // rsi
This has fallen out of alignment, the parameter types should line up.

http://codereview.chromium.org/269106/diff/3001/3002#newcode298
Line 298: : "r"(y_buf), "r"(u_buf), "r"(v_buf), "r"(rgb_buf), "r"(width),
I like to line these guys up with comments identifying the indices:

: "r"(y_buf),  // %0
  "r"(u_buf),  // %1

http://codereview.chromium.org/269106/diff/3001/3002#newcode300
Line 300: : "memory", "r10", "r11"
Can you put the SSE registers you use into the clobber list too?

http://codereview.chromium.org/269106/diff/3001/3002#newcode305
Line 305: const uint8* u_buf,   // rsi
Same as on 251.

http://codereview.chromium.org/269106/diff/3001/3002#newcode362
Line 362: : "r"(y_buf), "r"(u_buf), "r"(v_buf), "r"(rgb_buf), "r"(width),
Same as on 298.

http://codereview.chromium.org/269106/diff/3001/3002#newcode365
Line 365: : "memory", "r10", "r11", "rax"
Same as on 300.

[fbarchard, 2009-10-16 20:11:27]

fixed alignment/comments.
Changed code to use one table instead of 3.
The scale function uses more registers than there are 'temporaries' and had to
push some on the stack.
Using one table, like the Mac does, uses only one register.
Original performance was 302 ms
3 pointer version is 333 ms
1 pointer version is 326 ms
A 7.9% slowdown compared to the original, but still substantially faster than 32
bit, which is about 400ms or C code that is 1200 ms
The non-scaled version has more registers free, so no slowdown.