Merge remote-tracking branch 'qatar/master'

* qatar/master:
  lavr: add x86-optimized functions for mixing 1-to-2 s16p with flt coeffs
  lavr: add x86-optimized functions for mixing 1-to-2 fltp with flt coeffs
  Add Dolby/DPLII downmix support to libavresample
  vorbisdec: replace div/mod in loop with a counter
  fate: vorbis: add 5.1 surround test
  rtpenc: Allow requesting H264 RTP packetization mode 0
  configure: Sort the library listings in the help text alphabetically
  dwt: remove variable-length arrays
  RTMPT protocol support
  http: Properly handle chunked transfer-encoding for replies to post data
  http: Fail reading if the connection has gone away
  amr: Mark an array const
  amr: More space cleanup
  rtpenc: Fix memory leaks in the muxer open function

Conflicts:
	Changelog
	configure
	doc/APIchanges
	libavformat/version.h

Merged-by: Michael Niedermayer <michaelni@gmx.at>

8 files changed, 165 insertions, 11 deletions

diff --git a/libavresample/audio_mix.c b/libavresample/audio_mix.c
index 7ab11b0d4d..93192221cd 100644
--- a/libavresample/audio_mix.c
+++ b/libavresample/audio_mix.c
@@ -320,7 +320,8 @@ int ff_audio_mix_init(AVAudioResampleContext *avr)
                                       avr->center_mix_level,
                                       avr->surround_mix_level,
                                       avr->lfe_mix_level, 1, matrix_dbl,
-                                      avr->in_channels);
+                                      avr->in_channels,
+                                      avr->matrix_encoding);
         if (ret < 0) {
             av_free(matrix_dbl);
             return ret;
diff --git a/libavresample/audio_mix_matrix.c b/libavresample/audio_mix_matrix.c
index 6135b02422..f7121c846d 100644
--- a/libavresample/audio_mix_matrix.c
+++ b/libavresample/audio_mix_matrix.c
@@ -54,6 +54,8 @@
 #define SURROUND_DIRECT_LEFT   33
 #define SURROUND_DIRECT_RIGHT  34
 
+#define SQRT3_2      1.22474487139158904909  /* sqrt(3/2) */
+
 static av_always_inline int even(uint64_t layout)
 {
     return (!layout || (layout & (layout - 1)));
@@ -83,14 +85,21 @@ static int sane_layout(uint64_t layout)
 int avresample_build_matrix(uint64_t in_layout, uint64_t out_layout,
                             double center_mix_level, double surround_mix_level,
                             double lfe_mix_level, int normalize,
-                            double *matrix_out, int stride)
+                            double *matrix_out, int stride,
+                            enum AVMatrixEncoding matrix_encoding)
 {
     int i, j, out_i, out_j;
     double matrix[64][64] = {{0}};
-    int64_t unaccounted = in_layout & ~out_layout;
+    int64_t unaccounted;
     double maxcoef = 0;
     int in_channels, out_channels;
 
+    if ((out_layout & AV_CH_LAYOUT_STEREO_DOWNMIX) == AV_CH_LAYOUT_STEREO_DOWNMIX) {
+        out_layout = AV_CH_LAYOUT_STEREO;
+    }
+
+    unaccounted = in_layout & ~out_layout;
+
     in_channels  = av_get_channel_layout_nb_channels( in_layout);
     out_channels = av_get_channel_layout_nb_channels(out_layout);
 
@@ -140,8 +149,19 @@ int avresample_build_matrix(uint64_t in_layout, uint64_t out_layout,
             matrix[SIDE_LEFT ][BACK_CENTER] += M_SQRT1_2;
             matrix[SIDE_RIGHT][BACK_CENTER] += M_SQRT1_2;
         } else if (out_layout & AV_CH_FRONT_LEFT) {
-            matrix[FRONT_LEFT ][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
-            matrix[FRONT_RIGHT][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
+            if (matrix_encoding == AV_MATRIX_ENCODING_DOLBY ||
+                matrix_encoding == AV_MATRIX_ENCODING_DPLII) {
+                if (unaccounted & (AV_CH_BACK_LEFT | AV_CH_SIDE_LEFT)) {
+                    matrix[FRONT_LEFT ][BACK_CENTER] -= surround_mix_level * M_SQRT1_2;
+                    matrix[FRONT_RIGHT][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
+                } else {
+                    matrix[FRONT_LEFT ][BACK_CENTER] -= surround_mix_level;
+                    matrix[FRONT_RIGHT][BACK_CENTER] += surround_mix_level;
+                }
+            } else {
+                matrix[FRONT_LEFT ][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
+            }
         } else if (out_layout & AV_CH_FRONT_CENTER) {
             matrix[FRONT_CENTER][BACK_CENTER] += surround_mix_level * M_SQRT1_2;
         } else
@@ -163,8 +183,20 @@ int avresample_build_matrix(uint64_t in_layout, uint64_t out_layout,
                 matrix[SIDE_RIGHT][BACK_RIGHT] += 1.0;
             }
         } else if (out_layout & AV_CH_FRONT_LEFT) {
-            matrix[FRONT_LEFT ][BACK_LEFT ] += surround_mix_level;
-            matrix[FRONT_RIGHT][BACK_RIGHT] += surround_mix_level;
+            if (matrix_encoding == AV_MATRIX_ENCODING_DOLBY) {
+                matrix[FRONT_LEFT ][BACK_LEFT ] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_LEFT ][BACK_RIGHT] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][BACK_LEFT ] += surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][BACK_RIGHT] += surround_mix_level * M_SQRT1_2;
+            } else if (matrix_encoding == AV_MATRIX_ENCODING_DPLII) {
+                matrix[FRONT_LEFT ][BACK_LEFT ] -= surround_mix_level * SQRT3_2;
+                matrix[FRONT_LEFT ][BACK_RIGHT] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][BACK_LEFT ] += surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][BACK_RIGHT] += surround_mix_level * SQRT3_2;
+            } else {
+                matrix[FRONT_LEFT ][BACK_LEFT ] += surround_mix_level;
+                matrix[FRONT_RIGHT][BACK_RIGHT] += surround_mix_level;
+            }
         } else if (out_layout & AV_CH_FRONT_CENTER) {
             matrix[FRONT_CENTER][BACK_LEFT ] += surround_mix_level * M_SQRT1_2;
             matrix[FRONT_CENTER][BACK_RIGHT] += surround_mix_level * M_SQRT1_2;
@@ -187,8 +219,20 @@ int avresample_build_matrix(uint64_t in_layout, uint64_t out_layout,
             matrix[BACK_CENTER][SIDE_LEFT ] += M_SQRT1_2;
             matrix[BACK_CENTER][SIDE_RIGHT] += M_SQRT1_2;
         } else if (out_layout & AV_CH_FRONT_LEFT) {
-            matrix[FRONT_LEFT ][SIDE_LEFT ] += surround_mix_level;
-            matrix[FRONT_RIGHT][SIDE_RIGHT] += surround_mix_level;
+            if (matrix_encoding == AV_MATRIX_ENCODING_DOLBY) {
+                matrix[FRONT_LEFT ][SIDE_LEFT ] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_LEFT ][SIDE_RIGHT] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][SIDE_LEFT ] += surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][SIDE_RIGHT] += surround_mix_level * M_SQRT1_2;
+            } else if (matrix_encoding == AV_MATRIX_ENCODING_DPLII) {
+                matrix[FRONT_LEFT ][SIDE_LEFT ] -= surround_mix_level * SQRT3_2;
+                matrix[FRONT_LEFT ][SIDE_RIGHT] -= surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][SIDE_LEFT ] += surround_mix_level * M_SQRT1_2;
+                matrix[FRONT_RIGHT][SIDE_RIGHT] += surround_mix_level * SQRT3_2;
+            } else {
+                matrix[FRONT_LEFT ][SIDE_LEFT ] += surround_mix_level;
+                matrix[FRONT_RIGHT][SIDE_RIGHT] += surround_mix_level;
+            }
         } else if (out_layout & AV_CH_FRONT_CENTER) {
             matrix[FRONT_CENTER][SIDE_LEFT ] += surround_mix_level * M_SQRT1_2;
             matrix[FRONT_CENTER][SIDE_RIGHT] += surround_mix_level * M_SQRT1_2;
diff --git a/libavresample/avresample.h b/libavresample/avresample.h
index 65d4d2d6e2..002bec21fb 100644
--- a/libavresample/avresample.h
+++ b/libavresample/avresample.h
@@ -131,12 +131,13 @@ void avresample_free(AVAudioResampleContext **avr);
  *                            the weight of input channel i in output channel o.
  * @param stride              distance between adjacent input channels in the
  *                            matrix array
+ * @param matrix_encoding     matrixed stereo downmix mode (e.g. dplii)
  * @return                    0 on success, negative AVERROR code on failure
  */
 int avresample_build_matrix(uint64_t in_layout, uint64_t out_layout,
                             double center_mix_level, double surround_mix_level,
                             double lfe_mix_level, int normalize, double *matrix,
-                            int stride);
+                            int stride, enum AVMatrixEncoding matrix_encoding);
 
 /**
  * Get the current channel mixing matrix.
diff --git a/libavresample/internal.h b/libavresample/internal.h
index 49ea6a668e..fa9499a8ef 100644
--- a/libavresample/internal.h
+++ b/libavresample/internal.h
@@ -70,6 +70,7 @@ struct AVAudioResampleContext {
     AudioConvert *ac_out;       /**< output sample format conversion context */
     ResampleContext *resample;  /**< resampling context                      */
     AudioMix *am;               /**< channel mixing context                  */
+    enum AVMatrixEncoding matrix_encoding;      /**< matrixed stereo encoding */
 };
 
 #endif /* AVRESAMPLE_INTERNAL_H */
diff --git a/libavresample/options.c b/libavresample/options.c
index 5430c4ddf2..a1a0b0ca21 100644
--- a/libavresample/options.c
+++ b/libavresample/options.c
@@ -52,6 +52,10 @@ static const AVOption options[] = {
     { "phase_shift",            "Resampling Phase Shift",   OFFSET(phase_shift),            AV_OPT_TYPE_INT,    { 10                    }, 0,                    30, /* ??? */           PARAM },
     { "linear_interp",          "Use Linear Interpolation", OFFSET(linear_interp),          AV_OPT_TYPE_INT,    { 0                     }, 0,                    1,                      PARAM },
     { "cutoff",                 "Cutoff Frequency Ratio",   OFFSET(cutoff),                 AV_OPT_TYPE_DOUBLE, { 0.8                   }, 0.0,                  1.0,                    PARAM },
+    { "matrix_encoding",        "Matrixed Stereo Encoding", OFFSET(matrix_encoding),        AV_OPT_TYPE_INT,    { AV_MATRIX_ENCODING_NONE}, AV_MATRIX_ENCODING_NONE,     AV_MATRIX_ENCODING_NB-1, PARAM, "matrix_encoding" },
+        { "none",  "None",               0, AV_OPT_TYPE_CONST, { AV_MATRIX_ENCODING_NONE  }, INT_MIN, INT_MAX, PARAM, "matrix_encoding" },
+        { "dolby", "Dolby",              0, AV_OPT_TYPE_CONST, { AV_MATRIX_ENCODING_DOLBY }, INT_MIN, INT_MAX, PARAM, "matrix_encoding" },
+        { "dplii", "Dolby Pro Logic II", 0, AV_OPT_TYPE_CONST, { AV_MATRIX_ENCODING_DPLII }, INT_MIN, INT_MAX, PARAM, "matrix_encoding" },
     { NULL },
 };
 
diff --git a/libavresample/version.h b/libavresample/version.h
index 6211a56352..63f07f5e84 100644
--- a/libavresample/version.h
+++ b/libavresample/version.h
@@ -21,7 +21,7 @@
 
 #define LIBAVRESAMPLE_VERSION_MAJOR  0
 #define LIBAVRESAMPLE_VERSION_MINOR  0
-#define LIBAVRESAMPLE_VERSION_MICRO  2
+#define LIBAVRESAMPLE_VERSION_MICRO  3
 
 #define LIBAVRESAMPLE_VERSION_INT  AV_VERSION_INT(LIBAVRESAMPLE_VERSION_MAJOR, \
                                                   LIBAVRESAMPLE_VERSION_MINOR, \
diff --git a/libavresample/x86/audio_mix.asm b/libavresample/x86/audio_mix.asm
index 8a4cf061cd..4b0434dd6d 100644
--- a/libavresample/x86/audio_mix.asm
+++ b/libavresample/x86/audio_mix.asm
@@ -150,3 +150,84 @@ cglobal mix_2_to_1_s16p_q8, 3,4,6, src, matrix, len, src1
     sub        lend, mmsize/2
     jg .loop
     REP_RET
+
+;-----------------------------------------------------------------------------
+; void ff_mix_1_to_2_fltp_flt(float **src, float **matrix, int len,
+;                             int out_ch, int in_ch);
+;-----------------------------------------------------------------------------
+
+%macro MIX_1_TO_2_FLTP_FLT 0
+cglobal mix_1_to_2_fltp_flt, 3,5,4, src0, matrix0, len, src1, matrix1
+    mov       src1q, [src0q+gprsize]
+    mov       src0q, [src0q]
+    sub       src1q, src0q
+    mov    matrix1q, [matrix0q+gprsize]
+    mov    matrix0q, [matrix0q]
+    VBROADCASTSS m2, [matrix0q]
+    VBROADCASTSS m3, [matrix1q]
+    ALIGN 16
+.loop:
+    mova         m0, [src0q]
+    mulps        m1, m0, m3
+    mulps        m0, m0, m2
+    mova  [src0q      ], m0
+    mova  [src0q+src1q], m1
+    add       src0q, mmsize
+    sub        lend, mmsize/4
+    jg .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse
+MIX_1_TO_2_FLTP_FLT
+%if HAVE_AVX
+INIT_YMM avx
+MIX_1_TO_2_FLTP_FLT
+%endif
+
+;-----------------------------------------------------------------------------
+; void ff_mix_1_to_2_s16p_flt(int16_t **src, float **matrix, int len,
+;                             int out_ch, int in_ch);
+;-----------------------------------------------------------------------------
+
+%macro MIX_1_TO_2_S16P_FLT 0
+cglobal mix_1_to_2_s16p_flt, 3,5,6, src0, matrix0, len, src1, matrix1
+    mov       src1q, [src0q+gprsize]
+    mov       src0q, [src0q]
+    sub       src1q, src0q
+    mov    matrix1q, [matrix0q+gprsize]
+    mov    matrix0q, [matrix0q]
+    VBROADCASTSS m4, [matrix0q]
+    VBROADCASTSS m5, [matrix1q]
+    ALIGN 16
+.loop:
+    mova         m0, [src0q]
+    S16_TO_S32_SX 0, 2
+    cvtdq2ps     m0, m0
+    cvtdq2ps     m2, m2
+    mulps        m1, m0, m5
+    mulps        m0, m0, m4
+    mulps        m3, m2, m5
+    mulps        m2, m2, m4
+    cvtps2dq     m0, m0
+    cvtps2dq     m1, m1
+    cvtps2dq     m2, m2
+    cvtps2dq     m3, m3
+    packssdw     m0, m2
+    packssdw     m1, m3
+    mova  [src0q      ], m0
+    mova  [src0q+src1q], m1
+    add       src0q, mmsize
+    sub        lend, mmsize/2
+    jg .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+MIX_1_TO_2_S16P_FLT
+INIT_XMM sse4
+MIX_1_TO_2_S16P_FLT
+%if HAVE_AVX
+INIT_XMM avx
+MIX_1_TO_2_S16P_FLT
+%endif
diff --git a/libavresample/x86/audio_mix_init.c b/libavresample/x86/audio_mix_init.c
index fa204d6d36..b8f3a90eef 100644
--- a/libavresample/x86/audio_mix_init.c
+++ b/libavresample/x86/audio_mix_init.c
@@ -35,6 +35,18 @@ extern void ff_mix_2_to_1_s16p_flt_sse4(int16_t **src, float **matrix, int len,
 extern void ff_mix_2_to_1_s16p_q8_sse2(int16_t **src, int16_t **matrix,
                                        int len, int out_ch, int in_ch);
 
+extern void ff_mix_1_to_2_fltp_flt_sse(float **src, float **matrix, int len,
+                                       int out_ch, int in_ch);
+extern void ff_mix_1_to_2_fltp_flt_avx(float **src, float **matrix, int len,
+                                       int out_ch, int in_ch);
+
+extern void ff_mix_1_to_2_s16p_flt_sse2(int16_t **src, float **matrix, int len,
+                                        int out_ch, int in_ch);
+extern void ff_mix_1_to_2_s16p_flt_sse4(int16_t **src, float **matrix, int len,
+                                        int out_ch, int in_ch);
+extern void ff_mix_1_to_2_s16p_flt_avx (int16_t **src, float **matrix, int len,
+                                        int out_ch, int in_ch);
+
 av_cold void ff_audio_mix_init_x86(AudioMix *am)
 {
 #if HAVE_YASM
@@ -43,20 +55,30 @@ av_cold void ff_audio_mix_init_x86(AudioMix *am)
     if (mm_flags & AV_CPU_FLAG_SSE && HAVE_SSE) {
         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
                               2, 1, 16, 8, "SSE", ff_mix_2_to_1_fltp_flt_sse);
+        ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
+                              1, 2, 16, 4, "SSE", ff_mix_1_to_2_fltp_flt_sse);
     }
     if (mm_flags & AV_CPU_FLAG_SSE2 && HAVE_SSE) {
         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
                               2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_flt_sse2);
         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_Q8,
                               2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_q8_sse2);
+        ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
+                              1, 2, 16, 8, "SSE2", ff_mix_1_to_2_s16p_flt_sse2);
     }
     if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {
         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
                               2, 1, 16, 8, "SSE4", ff_mix_2_to_1_s16p_flt_sse4);
+        ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
+                              1, 2, 16, 8, "SSE4", ff_mix_1_to_2_s16p_flt_sse4);
     }
     if (mm_flags & AV_CPU_FLAG_AVX && HAVE_AVX) {
         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
                               2, 1, 32, 16, "AVX", ff_mix_2_to_1_fltp_flt_avx);
+        ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
+                              1, 2, 32, 8, "AVX", ff_mix_1_to_2_fltp_flt_avx);
+        ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
+                              1, 2, 16, 8, "AVX", ff_mix_1_to_2_s16p_flt_avx);
     }
 #endif
 }