[FFmpeg-cvslog] lavr: x86: optimized 6-channel s16p to s16 conversion
Justin Ruggles
git at videolan.org
Fri Aug 24 14:44:29 CEST 2012
ffmpeg | branch: master | Justin Ruggles <justin.ruggles at gmail.com> | Wed May 2 15:36:43 2012 -0400| [13df7d2d40237152a00a98a40c8e844a026ee5fa] | committer: Justin Ruggles
lavr: x86: optimized 6-channel s16p to s16 conversion
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=13df7d2d40237152a00a98a40c8e844a026ee5fa
---
libavresample/x86/audio_convert.asm | 123 ++++++++++++++++++++++++++++++++
libavresample/x86/audio_convert_init.c | 14 ++++
2 files changed, 137 insertions(+)
diff --git a/libavresample/x86/audio_convert.asm b/libavresample/x86/audio_convert.asm
index 4a92952..ee05efc 100644
--- a/libavresample/x86/audio_convert.asm
+++ b/libavresample/x86/audio_convert.asm
@@ -260,6 +260,129 @@ INIT_XMM avx
CONV_S16P_TO_S16_2CH
%endif
+;------------------------------------------------------------------------------
+; void ff_conv_s16p_to_s16_6ch(int16_t *dst, int16_t *const *src, int len,
+; int channels);
+;------------------------------------------------------------------------------
+
+;------------------------------------------------------------------------------
+; NOTE: In the 6-channel functions, len could be used as an index on x86-64
+; instead of just a counter, which would avoid incrementing the
+; pointers, but the extra complexity and amount of code is not worth
+; the small gain. On x86-32 there are not enough registers to use len
+; as an index without keeping two of the pointers on the stack and
+; loading them in each iteration.
+;------------------------------------------------------------------------------
+
+%macro CONV_S16P_TO_S16_6CH 0
+%if ARCH_X86_64
+cglobal conv_s16p_to_s16_6ch, 3,8,7, dst, src0, len, src1, src2, src3, src4, src5
+%else
+cglobal conv_s16p_to_s16_6ch, 2,7,7, dst, src0, src1, src2, src3, src4, src5
+%define lend dword r2m
+%endif
+ mov src1q, [src0q+1*gprsize]
+ mov src2q, [src0q+2*gprsize]
+ mov src3q, [src0q+3*gprsize]
+ mov src4q, [src0q+4*gprsize]
+ mov src5q, [src0q+5*gprsize]
+ mov src0q, [src0q]
+ sub src1q, src0q
+ sub src2q, src0q
+ sub src3q, src0q
+ sub src4q, src0q
+ sub src5q, src0q
+.loop:
+%if cpuflag(sse2slow)
+ movq m0, [src0q ] ; m0 = 0, 6, 12, 18, x, x, x, x
+ movq m1, [src0q+src1q] ; m1 = 1, 7, 13, 19, x, x, x, x
+ movq m2, [src0q+src2q] ; m2 = 2, 8, 14, 20, x, x, x, x
+ movq m3, [src0q+src3q] ; m3 = 3, 9, 15, 21, x, x, x, x
+ movq m4, [src0q+src4q] ; m4 = 4, 10, 16, 22, x, x, x, x
+ movq m5, [src0q+src5q] ; m5 = 5, 11, 17, 23, x, x, x, x
+ ; unpack words:
+ punpcklwd m0, m1 ; m0 = 0, 1, 6, 7, 12, 13, 18, 19
+ punpcklwd m2, m3 ; m2 = 4, 5, 10, 11, 16, 17, 22, 23
+ punpcklwd m4, m5 ; m4 = 2, 3, 8, 9, 14, 15, 20, 21
+ ; blend dwords
+ shufps m1, m0, m2, q2020 ; m1 = 0, 1, 12, 13, 2, 3, 14, 15
+ shufps m0, m4, q2031 ; m0 = 6, 7, 18, 19, 4, 5, 16, 17
+ shufps m2, m4, q3131 ; m2 = 8, 9, 20, 21, 10, 11, 22, 23
+ ; shuffle dwords
+ pshufd m0, m0, q1302 ; m0 = 4, 5, 6, 7, 16, 17, 18, 19
+ pshufd m1, m1, q3120 ; m1 = 0, 1, 2, 3, 12, 13, 14, 15
+ pshufd m2, m2, q3120 ; m2 = 8, 9, 10, 11, 20, 21, 22, 23
+ movq [dstq+0*mmsize/2], m1
+ movq [dstq+1*mmsize/2], m0
+ movq [dstq+2*mmsize/2], m2
+ movhps [dstq+3*mmsize/2], m1
+ movhps [dstq+4*mmsize/2], m0
+ movhps [dstq+5*mmsize/2], m2
+ add src0q, mmsize/2
+ add dstq, mmsize*3
+ sub lend, mmsize/4
+%else
+ mova m0, [src0q ] ; m0 = 0, 6, 12, 18, 24, 30, 36, 42
+ mova m1, [src0q+src1q] ; m1 = 1, 7, 13, 19, 25, 31, 37, 43
+ mova m2, [src0q+src2q] ; m2 = 2, 8, 14, 20, 26, 32, 38, 44
+ mova m3, [src0q+src3q] ; m3 = 3, 9, 15, 21, 27, 33, 39, 45
+ mova m4, [src0q+src4q] ; m4 = 4, 10, 16, 22, 28, 34, 40, 46
+ mova m5, [src0q+src5q] ; m5 = 5, 11, 17, 23, 29, 35, 41, 47
+ ; unpack words:
+ SBUTTERFLY2 wd, 0, 1, 6 ; m0 = 0, 1, 6, 7, 12, 13, 18, 19
+ ; m1 = 24, 25, 30, 31, 36, 37, 42, 43
+ SBUTTERFLY2 wd, 2, 3, 6 ; m2 = 2, 3, 8, 9, 14, 15, 20, 21
+ ; m3 = 26, 27, 32, 33, 38, 39, 44, 45
+ SBUTTERFLY2 wd, 4, 5, 6 ; m4 = 4, 5, 10, 11, 16, 17, 22, 23
+ ; m5 = 28, 29, 34, 35, 40, 41, 46, 47
+ ; blend dwords
+ shufps m6, m0, m2, q2020 ; m6 = 0, 1, 12, 13, 2, 3, 14, 15
+ shufps m0, m4, q2031 ; m0 = 6, 7, 18, 19, 4, 5, 16, 17
+ shufps m2, m4, q3131 ; m2 = 8, 9, 20, 21, 10, 11, 22, 23
+ SWAP 4,6 ; m4 = 0, 1, 12, 13, 2, 3, 14, 15
+ shufps m6, m1, m3, q2020 ; m6 = 24, 25, 36, 37, 26, 27, 38, 39
+ shufps m1, m5, q2031 ; m1 = 30, 31, 42, 43, 28, 29, 40, 41
+ shufps m3, m5, q3131 ; m3 = 32, 33, 44, 45, 34, 35, 46, 47
+ SWAP 5,6 ; m5 = 24, 25, 36, 37, 26, 27, 38, 39
+ ; shuffle dwords
+ pshufd m0, m0, q1302 ; m0 = 4, 5, 6, 7, 16, 17, 18, 19
+ pshufd m2, m2, q3120 ; m2 = 8, 9, 10, 11, 20, 21, 22, 23
+ pshufd m4, m4, q3120 ; m4 = 0, 1, 2, 3, 12, 13, 14, 15
+ pshufd m1, m1, q1302 ; m1 = 28, 29, 30, 31, 40, 41, 42, 43
+ pshufd m3, m3, q3120 ; m3 = 32, 33, 34, 35, 44, 45, 46, 47
+ pshufd m5, m5, q3120 ; m5 = 24, 25, 26, 27, 36, 37, 38, 39
+ ; shuffle qwords
+ punpcklqdq m6, m4, m0 ; m6 = 0, 1, 2, 3, 4, 5, 6, 7
+ punpckhqdq m0, m2 ; m0 = 16, 17, 18, 19, 20, 21, 22, 23
+ shufps m2, m4, q3210 ; m2 = 8, 9, 10, 11, 12, 13, 14, 15
+ SWAP 4,6 ; m4 = 0, 1, 2, 3, 4, 5, 6, 7
+ punpcklqdq m6, m5, m1 ; m6 = 24, 25, 26, 27, 28, 29, 30, 31
+ punpckhqdq m1, m3 ; m1 = 40, 41, 42, 43, 44, 45, 46, 47
+ shufps m3, m5, q3210 ; m3 = 32, 33, 34, 35, 36, 37, 38, 39
+ SWAP 5,6 ; m5 = 24, 25, 26, 27, 28, 29, 30, 31
+ mova [dstq+0*mmsize], m4
+ mova [dstq+1*mmsize], m2
+ mova [dstq+2*mmsize], m0
+ mova [dstq+3*mmsize], m5
+ mova [dstq+4*mmsize], m3
+ mova [dstq+5*mmsize], m1
+ add src0q, mmsize
+ add dstq, mmsize*6
+ sub lend, mmsize/2
+%endif
+ jg .loop
+ REP_RET
+%endmacro
+
+INIT_XMM sse2
+CONV_S16P_TO_S16_6CH
+INIT_XMM sse2slow
+CONV_S16P_TO_S16_6CH
+%if HAVE_AVX
+INIT_XMM avx
+CONV_S16P_TO_S16_6CH
+%endif
+
;-----------------------------------------------------------------------------
; void ff_conv_fltp_to_flt_6ch(float *dst, float *const *src, int len,
; int channels);
diff --git a/libavresample/x86/audio_convert_init.c b/libavresample/x86/audio_convert_init.c
index 9b7bcb1..d9d4714 100644
--- a/libavresample/x86/audio_convert_init.c
+++ b/libavresample/x86/audio_convert_init.c
@@ -47,6 +47,13 @@ extern void ff_conv_s16p_to_s16_2ch_sse2(int16_t *dst, int16_t *const *src,
extern void ff_conv_s16p_to_s16_2ch_avx (int16_t *dst, int16_t *const *src,
int len, int channels);
+extern void ff_conv_s16p_to_s16_6ch_sse2(int16_t *dst, int16_t *const *src,
+ int len, int channels);
+extern void ff_conv_s16p_to_s16_6ch_sse2slow(int16_t *dst, int16_t *const *src,
+ int len, int channels);
+extern void ff_conv_s16p_to_s16_6ch_avx (int16_t *dst, int16_t *const *src,
+ int len, int channels);
+
extern void ff_conv_fltp_to_flt_6ch_mmx (float *dst, float *const *src, int len,
int channels);
extern void ff_conv_fltp_to_flt_6ch_sse4(float *dst, float *const *src, int len,
@@ -69,6 +76,11 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S32,
0, 16, 16, "SSE2", ff_conv_s32_to_s16_sse2);
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
+ 6, 16, 8, "SSE2", ff_conv_s16p_to_s16_6ch_sse2);
+ } else {
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
+ 6, 1, 4, "SSE2SLOW", ff_conv_s16p_to_s16_6ch_sse2slow);
}
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE2", ff_conv_s16_to_s32_sse2);
@@ -96,6 +108,8 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
0, 32, 32, "AVX", ff_conv_flt_to_s32_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
2, 16, 16, "AVX", ff_conv_s16p_to_s16_2ch_avx);
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
+ 6, 16, 8, "AVX", ff_conv_s16p_to_s16_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
6, 16, 4, "AVX", ff_conv_fltp_to_flt_6ch_avx);
}
More information about the ffmpeg-cvslog
mailing list