Commit 2debd5b5 authored by Yunqing Wang's avatar Yunqing Wang

Improve vp8_sad16x16_sse3 function

In real-time mode, vp8_sad16x16 function is called heavily in
motion search part. Improvement of this function gives 1.2%
encoding performance gain (real-time mode, tulip clip).

Change-Id: I23c401fc40c061f732a9767e8d383737a179bd58
parent 404e998e
......@@ -586,52 +586,45 @@ sym(vp8_sad16x16_sse3):
lea end_ptr, [src_ptr+src_stride*8]
lea end_ptr, [end_ptr+src_stride*8]
pxor mm7, mm7
mov end_ptr, 4
pxor xmm7, xmm7
movdqa xmm0, XMMWORD PTR [src_ptr]
movdqu xmm1, XMMWORD PTR [ref_ptr]
movdqa xmm2, XMMWORD PTR [src_ptr+src_stride]
movdqu xmm3, XMMWORD PTR [ref_ptr+ref_stride]
movq ret_var, mm7
cmp ret_var, max_err
jg .vp8_sad16x16_early_exit
movq mm0, QWORD PTR [src_ptr]
movq mm2, QWORD PTR [src_ptr+8]
movq mm1, QWORD PTR [ref_ptr]
movq mm3, QWORD PTR [ref_ptr+8]
lea src_ptr, [src_ptr+src_stride*2]
lea ref_ptr, [ref_ptr+ref_stride*2]
movq mm4, QWORD PTR [src_ptr+src_stride]
movq mm5, QWORD PTR [ref_ptr+ref_stride]
movdqa xmm4, XMMWORD PTR [src_ptr]
movdqu xmm5, XMMWORD PTR [ref_ptr]
movdqa xmm6, XMMWORD PTR [src_ptr+src_stride]
psadbw mm0, mm1
psadbw mm2, mm3
psadbw xmm0, xmm1
movq mm1, QWORD PTR [src_ptr+src_stride+8]
movq mm3, QWORD PTR [ref_ptr+ref_stride+8]
movdqu xmm1, XMMWORD PTR [ref_ptr+ref_stride]
psadbw mm4, mm5
psadbw mm1, mm3
psadbw xmm2, xmm3
psadbw xmm4, xmm5
psadbw xmm6, xmm1
lea src_ptr, [src_ptr+src_stride*2]
lea ref_ptr, [ref_ptr+ref_stride*2]
paddw mm0, mm2
paddw mm4, mm1
paddw mm7, mm0
paddw mm7, mm4
paddw xmm7, xmm0
paddw xmm7, xmm2
paddw xmm7, xmm4
paddw xmm7, xmm6
cmp src_ptr, end_ptr
sub end_ptr, 1
jne .vp8_sad16x16_sse3_loop
movq ret_var, mm7
mov rax, ret_var
movq xmm0, xmm7
psrldq xmm7, 8
paddw xmm0, xmm7
movq rax, xmm0
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment