aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorLoren Merritt <lorenm@u.washington.edu>2008-08-13 23:33:48 +0000
committerLoren Merritt <lorenm@u.washington.edu>2008-08-13 23:33:48 +0000
commitac2e556456bf8be035bdee386e8ec2a63ebb870d (patch)
treed76c0fdd1bb6a9df9f061c6aab6db1cc2cbcff71
parent45d9d61889874bf7d2f0babddb6778e71d71630e (diff)
downloadffmpeg-ac2e556456bf8be035bdee386e8ec2a63ebb870d.tar.gz
simd downmix
13% faster ac3 if downmixing Originally committed as revision 14742 to svn://svn.ffmpeg.org/ffmpeg/trunk
-rw-r--r--libavcodec/ac3dec.c29
-rw-r--r--libavcodec/dsputil.c6
-rw-r--r--libavcodec/dsputil.h1
-rw-r--r--libavcodec/i386/dsputil_mmx.c100
4 files changed, 120 insertions, 16 deletions
diff --git a/libavcodec/ac3dec.c b/libavcodec/ac3dec.c
index 4b288df84a..44233b9468 100644
--- a/libavcodec/ac3dec.c
+++ b/libavcodec/ac3dec.c
@@ -632,27 +632,25 @@ static inline void do_imdct(AC3DecodeContext *s, int channels)
/**
* Downmix the output to mono or stereo.
*/
-static av_noinline void ac3_downmix(AC3DecodeContext *s,
- float samples[AC3_MAX_CHANNELS][256])
+void ff_ac3_downmix_c(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)
{
int i, j;
float v0, v1;
-
- if(s->output_mode == AC3_CHMODE_STEREO) {
- for(i=0; i<256; i++) {
+ if(out_ch == 2) {
+ for(i=0; i<len; i++) {
v0 = v1 = 0.0f;
- for(j=0; j<s->fbw_channels; j++) {
- v0 += samples[j][i] * s->downmix_coeffs[j][0];
- v1 += samples[j][i] * s->downmix_coeffs[j][1];
+ for(j=0; j<in_ch; j++) {
+ v0 += samples[j][i] * matrix[j][0];
+ v1 += samples[j][i] * matrix[j][1];
}
samples[0][i] = v0;
samples[1][i] = v1;
}
- } else if(s->output_mode == AC3_CHMODE_MONO) {
- for(i=0; i<256; i++) {
+ } else if(out_ch == 1) {
+ for(i=0; i<len; i++) {
v0 = 0.0f;
- for(j=0; j<s->fbw_channels; j++)
- v0 += samples[j][i] * s->downmix_coeffs[j][0];
+ for(j=0; j<in_ch; j++)
+ v0 += samples[j][i] * matrix[j][0];
samples[0][i] = v0;
}
}
@@ -1018,17 +1016,16 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
do_imdct(s, s->channels);
if(downmix_output) {
- ac3_downmix(s, s->output);
+ s->dsp.ac3_downmix(s->output, s->downmix_coeffs, s->out_channels, s->fbw_channels, 256);
}
} else {
if(downmix_output) {
- ac3_downmix(s, s->transform_coeffs+1);
+ s->dsp.ac3_downmix(s->transform_coeffs+1, s->downmix_coeffs, s->out_channels, s->fbw_channels, 256);
}
if(downmix_output && !s->downmixed) {
s->downmixed = 1;
- // FIXME delay[] is half the size of the other downmixes
- ac3_downmix(s, s->delay);
+ s->dsp.ac3_downmix(s->delay, s->downmix_coeffs, s->out_channels, s->fbw_channels, 128);
}
do_imdct(s, s->out_channels);
diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index a9661eb2b5..b6df158b3e 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -41,6 +41,9 @@ void ff_spatial_dwt(int *buffer, int width, int height, int stride, int type, in
/* vorbis.c */
void vorbis_inverse_coupling(float *mag, float *ang, int blocksize);
+/* ac3dec.c */
+void ff_ac3_downmix_c(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
+
/* flacenc.c */
void ff_flac_compute_autocorr(const int32_t *data, int len, int lag, double *autoc);
@@ -4476,6 +4479,9 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
#ifdef CONFIG_VORBIS_DECODER
c->vorbis_inverse_coupling = vorbis_inverse_coupling;
#endif
+#ifdef CONFIG_AC3_DECODER
+ c->ac3_downmix = ff_ac3_downmix_c;
+#endif
#ifdef CONFIG_FLAC_ENCODER
c->flac_compute_autocorr = ff_flac_compute_autocorr;
#endif
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 600f4fbcce..83102db501 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -360,6 +360,7 @@ typedef struct DSPContext {
/* assume len is a multiple of 4, and arrays are 16-byte aligned */
void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
+ void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
/* no alignment needed */
void (*flac_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
/* assume len is a multiple of 8, and arrays are 16-byte aligned */
diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index b15d42a017..c7bcd3b783 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -1842,6 +1842,105 @@ static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
}
}
+#define IF1(x) x
+#define IF0(x)
+
+#define MIX5(mono,stereo)\
+ asm volatile(\
+ "movss 0(%2), %%xmm5 \n"\
+ "movss 8(%2), %%xmm6 \n"\
+ "movss 24(%2), %%xmm7 \n"\
+ "shufps $0, %%xmm5, %%xmm5 \n"\
+ "shufps $0, %%xmm6, %%xmm6 \n"\
+ "shufps $0, %%xmm7, %%xmm7 \n"\
+ "1: \n"\
+ "movaps (%0,%1), %%xmm0 \n"\
+ "movaps 0x400(%0,%1), %%xmm1 \n"\
+ "movaps 0x800(%0,%1), %%xmm2 \n"\
+ "movaps 0xc00(%0,%1), %%xmm3 \n"\
+ "movaps 0x1000(%0,%1), %%xmm4 \n"\
+ "mulps %%xmm5, %%xmm0 \n"\
+ "mulps %%xmm6, %%xmm1 \n"\
+ "mulps %%xmm5, %%xmm2 \n"\
+ "mulps %%xmm7, %%xmm3 \n"\
+ "mulps %%xmm7, %%xmm4 \n"\
+ stereo("addps %%xmm1, %%xmm0 \n")\
+ "addps %%xmm1, %%xmm2 \n"\
+ "addps %%xmm3, %%xmm0 \n"\
+ "addps %%xmm4, %%xmm2 \n"\
+ mono("addps %%xmm2, %%xmm0 \n")\
+ "movaps %%xmm0, (%0,%1) \n"\
+ stereo("movaps %%xmm2, 0x400(%0,%1) \n")\
+ "add $16, %0 \n"\
+ "jl 1b \n"\
+ :"+&r"(i)\
+ :"r"(samples[0]+len), "r"(matrix)\
+ :"memory"\
+ );
+
+#define MIX_MISC(stereo)\
+ asm volatile(\
+ "1: \n"\
+ "movaps (%3,%0), %%xmm0 \n"\
+ stereo("movaps %%xmm0, %%xmm1 \n")\
+ "mulps %%xmm6, %%xmm0 \n"\
+ stereo("mulps %%xmm7, %%xmm1 \n")\
+ "lea 1024(%3,%0), %1 \n"\
+ "mov %5, %2 \n"\
+ "2: \n"\
+ "movaps (%1), %%xmm2 \n"\
+ stereo("movaps %%xmm2, %%xmm3 \n")\
+ "mulps (%4,%2), %%xmm2 \n"\
+ stereo("mulps 16(%4,%2), %%xmm3 \n")\
+ "addps %%xmm2, %%xmm0 \n"\
+ stereo("addps %%xmm3, %%xmm1 \n")\
+ "add $1024, %1 \n"\
+ "add $32, %2 \n"\
+ "jl 2b \n"\
+ "movaps %%xmm0, (%3,%0) \n"\
+ stereo("movaps %%xmm1, 1024(%3,%0) \n")\
+ "add $16, %0 \n"\
+ "jl 1b \n"\
+ :"+&r"(i), "=&r"(j), "=&r"(k)\
+ :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\
+ :"memory"\
+ );
+
+static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len)
+{
+ int (*matrix_cmp)[2] = (int(*)[2])matrix;
+ intptr_t i,j,k;
+
+ i = -len*sizeof(float);
+ if(in_ch == 5 && out_ch == 2 && !(matrix_cmp[0][1]|matrix_cmp[2][0]|matrix_cmp[3][1]|matrix_cmp[4][0]|(matrix_cmp[1][0]^matrix_cmp[1][1])|(matrix_cmp[0][0]^matrix_cmp[2][1]))) {
+ MIX5(IF0,IF1);
+ } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) {
+ MIX5(IF1,IF0);
+ } else {
+ DECLARE_ALIGNED_16(float, matrix_simd[in_ch][2][4]);
+ j = 2*in_ch*sizeof(float);
+ asm volatile(
+ "1: \n"
+ "sub $8, %0 \n"
+ "movss (%2,%0), %%xmm6 \n"
+ "movss 4(%2,%0), %%xmm7 \n"
+ "shufps $0, %%xmm6, %%xmm6 \n"
+ "shufps $0, %%xmm7, %%xmm7 \n"
+ "movaps %%xmm6, (%1,%0,4) \n"
+ "movaps %%xmm7, 16(%1,%0,4) \n"
+ "jg 1b \n"
+ :"+&r"(j)
+ :"r"(matrix_simd), "r"(matrix)
+ :"memory"
+ );
+ if(out_ch == 2) {
+ MIX_MISC(IF1);
+ } else {
+ MIX_MISC(IF0);
+ }
+ }
+}
+
static void vector_fmul_3dnow(float *dst, const float *src, int len){
x86_reg i = (len-4)*4;
asm volatile(
@@ -2682,6 +2781,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
}
if(mm_flags & MM_SSE){
c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
+ c->ac3_downmix = ac3_downmix_sse;
c->vector_fmul = vector_fmul_sse;
c->vector_fmul_reverse = vector_fmul_reverse_sse;
c->vector_fmul_add_add = vector_fmul_add_add_sse;