xenia-project/FFmpeg
1
0
Fork 0
mirror of https://github.com/xenia-project/FFmpeg.git synced 2024-11-24 12:09:55 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

lavc/flacenc: partially unroll loop in flac_enc_lpc_16

Browse Source 
It now does 12 samples per iteration, up from 4.

From 1.8 to 3.2 times faster again.  3.6 to 5.7 times faster overall.
Runtime is reduced by a further 2 to 18%.  Overall runtime reduced by
4 to 50%.

Same conditions as before apply.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit is contained in:
James Darnley 2014-08-12 23:22:03 +02:00 committed by Michael Niedermayer
parent a8592db9bb
commit 54a51d3840
2 changed files with 22 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
libavcodec/flacenc.c
View File

@ -80,7 +80,7 @@ typedef struct FlacSubframe {
int shift;
RiceContext rc;
int32_t samples[FLAC_MAX_BLOCKSIZE];
int32_t residual[FLAC_MAX_BLOCKSIZE+3];
int32_t residual[FLAC_MAX_BLOCKSIZE+11];
} FlacSubframe;
typedef struct FlacFrame {

26
libavcodec/x86/flac_dsp_gpl.asm
View File

@ -26,13 +26,13 @@ SECTION_TEXT
INIT_XMM sse4
%if ARCH_X86_64
cglobal flac_enc_lpc_16, 5, 7, 4, 0, res, smp, len, order, coefs
cglobal flac_enc_lpc_16, 5, 7, 8, 0, res, smp, len, order, coefs
DECLARE_REG_TMP 5, 6
%define length r2d
movsxd orderq, orderd
%else
cglobal flac_enc_lpc_16, 5, 6, 4, 0, res, smp, len, order, coefs
cglobal flac_enc_lpc_16, 5, 6, 8, 0, res, smp, len, order, coefs
DECLARE_REG_TMP 2, 5
%define length r2mp
%endif
@ -59,6 +59,8 @@ neg orderq
.looplen:
pxor m0, m0
pxor m4, m4
pxor m6, m6
mov posj, orderq
xor negj, negj
@ -66,20 +68,34 @@ neg orderq
movd m2, [coefsq+posj*4] ; c = coefs[j]
SPLATD m2
movu m1, [smpq+negj*4-4] ; s = smp[i-j-1]
movu m5, [smpq+negj*4-4+mmsize]
movu m7, [smpq+negj*4-4+mmsize*2]
pmulld m1, m2
pmulld m5, m2
pmulld m7, m2
paddd m0, m1 ; p += c * s
paddd m4, m5
paddd m6, m7
dec negj
inc posj
jnz .looporder
psrad m0, m3 ; p >>= shift
psrad m4, m3
psrad m6, m3
movu m1, [smpq]
movu m5, [smpq+mmsize]
movu m7, [smpq+mmsize*2]
psubd m1, m0 ; smp[i] - p
psubd m5, m4
psubd m7, m6
movu [resq], m1 ; res[i] = smp[i] - (p >> shift)
movu [resq+mmsize], m5
movu [resq+mmsize*2], m7
add resq, mmsize
add smpq, mmsize
sub length, mmsize/4
add resq, 3*mmsize
add smpq, 3*mmsize
sub length, (3*mmsize)/4
jg .looplen
RET