Merge commit '43717469f9daa402f6acb48997255827a56034e9'

* commit '43717469f9daa402f6acb48997255827a56034e9':
  ac3dsp: Reverse matrix in/out order in downmix()

Merged-by: Clément Bœsch <u@pkh.me>

6 files changed, 64 insertions, 48 deletions

diff --git a/libavcodec/ac3dec.c b/libavcodec/ac3dec.c
index 003ce17..f9bab94 100644
--- a/libavcodec/ac3dec.c
+++ b/libavcodec/ac3dec.c
@@ -362,52 +362,62 @@ static int parse_frame_header(AC3DecodeContext *s)
  * Set stereo downmixing coefficients based on frame header info.
  * reference: Section 7.8.2 Downmixing Into Two Channels
  */
-static void set_downmix_coeffs(AC3DecodeContext *s)
+static int set_downmix_coeffs(AC3DecodeContext *s)
 {
     int i;
     float cmix = gain_levels[s->  center_mix_level];
     float smix = gain_levels[s->surround_mix_level];
     float norm0, norm1;
-    float downmix_coeffs[AC3_MAX_CHANNELS][2];
+    float downmix_coeffs[2][AC3_MAX_CHANNELS];
+
+    if (!s->downmix_coeffs[0]) {
+        s->downmix_coeffs[0] = av_malloc_array(2 * AC3_MAX_CHANNELS,
+                                               sizeof(**s->downmix_coeffs));
+        if (!s->downmix_coeffs[0])
+            return AVERROR(ENOMEM);
+        s->downmix_coeffs[1] = s->downmix_coeffs[0] + AC3_MAX_CHANNELS;
+    }
 
     for (i = 0; i < s->fbw_channels; i++) {
-        downmix_coeffs[i][0] = gain_levels[ac3_default_coeffs[s->channel_mode][i][0]];
-        downmix_coeffs[i][1] = gain_levels[ac3_default_coeffs[s->channel_mode][i][1]];
+        downmix_coeffs[0][i] = gain_levels[ac3_default_coeffs[s->channel_mode][i][0]];
+        downmix_coeffs[1][i] = gain_levels[ac3_default_coeffs[s->channel_mode][i][1]];
     }
     if (s->channel_mode > 1 && s->channel_mode & 1) {
-        downmix_coeffs[1][0] = downmix_coeffs[1][1] = cmix;
+        downmix_coeffs[0][1] = downmix_coeffs[1][1] = cmix;
     }
     if (s->channel_mode == AC3_CHMODE_2F1R || s->channel_mode == AC3_CHMODE_3F1R) {
         int nf = s->channel_mode - 2;
-        downmix_coeffs[nf][0] = downmix_coeffs[nf][1] = smix * LEVEL_MINUS_3DB;
+        downmix_coeffs[0][nf] = downmix_coeffs[1][nf] = smix * LEVEL_MINUS_3DB;
     }
     if (s->channel_mode == AC3_CHMODE_2F2R || s->channel_mode == AC3_CHMODE_3F2R) {
         int nf = s->channel_mode - 4;
-        downmix_coeffs[nf][0] = downmix_coeffs[nf+1][1] = smix;
+        downmix_coeffs[0][nf] = downmix_coeffs[1][nf+1] = smix;
     }
 
     /* renormalize */
     norm0 = norm1 = 0.0;
     for (i = 0; i < s->fbw_channels; i++) {
-        norm0 += downmix_coeffs[i][0];
-        norm1 += downmix_coeffs[i][1];
+        norm0 += downmix_coeffs[0][i];
+        norm1 += downmix_coeffs[1][i];
     }
     norm0 = 1.0f / norm0;
     norm1 = 1.0f / norm1;
     for (i = 0; i < s->fbw_channels; i++) {
-        downmix_coeffs[i][0] *= norm0;
-        downmix_coeffs[i][1] *= norm1;
+        downmix_coeffs[0][i] *= norm0;
+        downmix_coeffs[1][i] *= norm1;
     }
 
     if (s->output_mode == AC3_CHMODE_MONO) {
         for (i = 0; i < s->fbw_channels; i++)
-            downmix_coeffs[i][0] = (downmix_coeffs[i][0] +
-                                    downmix_coeffs[i][1]) * LEVEL_MINUS_3DB;
+            downmix_coeffs[0][i] = (downmix_coeffs[0][i] +
+                                    downmix_coeffs[1][i]) * LEVEL_MINUS_3DB;
     }
     for (i = 0; i < s->fbw_channels; i++) {
-        s->downmix_coeffs[i][0] = FIXR12(downmix_coeffs[i][0]);
-        s->downmix_coeffs[i][1] = FIXR12(downmix_coeffs[i][1]);
+        s->downmix_coeffs[0][i] = FIXR12(downmix_coeffs[0][i]);
+        s->downmix_coeffs[1][i] = FIXR12(downmix_coeffs[1][i]);
     }
+
+    return 0;
 }
 
 /**
@@ -1562,7 +1572,10 @@ static int ac3_decode_frame(AVCodecContext * avctx, void *data,
         /* set downmixing coefficients if needed */
         if (s->channels != s->out_channels && !((s->output_mode & AC3_OUTPUT_LFEON) &&
                 s->fbw_channels == s->out_channels)) {
-            set_downmix_coeffs(s);
+            if ((ret = set_downmix_coeffs(s)) < 0) {
+                av_log(avctx, AV_LOG_ERROR, "error setting downmix coeffs\n");
+                return ret;
+            }
         }
     } else if (!s->channels) {
         av_log(avctx, AV_LOG_ERROR, "unable to determine channel mode\n");
@@ -1685,6 +1698,7 @@ static av_cold int ac3_decode_end(AVCodecContext *avctx)
     ff_mdct_end(&s->imdct_512);
     ff_mdct_end(&s->imdct_256);
     av_freep(&s->fdsp);
+    av_freep(&s->downmix_coeffs[0]);
 
     return 0;
 }
diff --git a/libavcodec/ac3dec.h b/libavcodec/ac3dec.h
index 495e9a6..bac661c 100644
--- a/libavcodec/ac3dec.h
+++ b/libavcodec/ac3dec.h
@@ -159,7 +159,7 @@ typedef struct AC3DecodeContext {
     int fbw_channels;                           ///< number of full-bandwidth channels
     int channels;                               ///< number of total channels
     int lfe_ch;                                 ///< index of LFE channel
-    SHORTFLOAT downmix_coeffs[AC3_MAX_CHANNELS][2];  ///< stereo downmix coefficients
+    SHORTFLOAT *downmix_coeffs[2];              ///< stereo downmix coefficients
     int downmixed;                              ///< indicates if coeffs are currently downmixed
     int output_mode;                            ///< output channel configuration
     int out_channels;                           ///< number of output channels
diff --git a/libavcodec/ac3dec_fixed.c b/libavcodec/ac3dec_fixed.c
index 1f79ade..682fe93 100644
--- a/libavcodec/ac3dec_fixed.c
+++ b/libavcodec/ac3dec_fixed.c
@@ -139,7 +139,7 @@ static void scale_coefs (
  * Downmix samples from original signal to stereo or mono (this is for 16-bit samples
  * and fixed point decoder - original (for 32-bit samples) is in ac3dsp.c).
  */
-static void ac3_downmix_c_fixed16(int16_t **samples, int16_t (*matrix)[2],
+static void ac3_downmix_c_fixed16(int16_t **samples, int16_t **matrix,
                                   int out_ch, int in_ch, int len)
 {
     int i, j;
@@ -148,8 +148,8 @@ static void ac3_downmix_c_fixed16(int16_t **samples, int16_t (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = v1 = 0;
             for (j = 0; j < in_ch; j++) {
-                v0 += samples[j][i] * matrix[j][0];
-                v1 += samples[j][i] * matrix[j][1];
+                v0 += samples[j][i] * matrix[0][j];
+                v1 += samples[j][i] * matrix[1][j];
             }
             samples[0][i] = (v0+2048)>>12;
             samples[1][i] = (v1+2048)>>12;
@@ -158,7 +158,7 @@ static void ac3_downmix_c_fixed16(int16_t **samples, int16_t (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = 0;
             for (j = 0; j < in_ch; j++)
-                v0 += samples[j][i] * matrix[j][0];
+                v0 += samples[j][i] * matrix[0][j];
             samples[0][i] = (v0+2048)>>12;
         }
     }
diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
index 9902f90..23abc56 100644
--- a/libavcodec/ac3dsp.c
+++ b/libavcodec/ac3dsp.c
@@ -213,7 +213,7 @@ static void ac3_sum_square_butterfly_float_c(float sum[4],
     }
 }
 
-static void ac3_downmix_c(float **samples, float (*matrix)[2],
+static void ac3_downmix_c(float **samples, float **matrix,
                           int out_ch, int in_ch, int len)
 {
     int i, j;
@@ -222,8 +222,8 @@ static void ac3_downmix_c(float **samples, float (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = v1 = 0.0f;
             for (j = 0; j < in_ch; j++) {
-                v0 += samples[j][i] * matrix[j][0];
-                v1 += samples[j][i] * matrix[j][1];
+                v0 += samples[j][i] * matrix[0][j];
+                v1 += samples[j][i] * matrix[1][j];
             }
             samples[0][i] = v0;
             samples[1][i] = v1;
@@ -232,13 +232,13 @@ static void ac3_downmix_c(float **samples, float (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = 0.0f;
             for (j = 0; j < in_ch; j++)
-                v0 += samples[j][i] * matrix[j][0];
+                v0 += samples[j][i] * matrix[0][j];
             samples[0][i] = v0;
         }
     }
 }
 
-static void ac3_downmix_c_fixed(int32_t **samples, int16_t (*matrix)[2],
+static void ac3_downmix_c_fixed(int32_t **samples, int16_t **matrix,
                                 int out_ch, int in_ch, int len)
 {
     int i, j;
@@ -247,8 +247,8 @@ static void ac3_downmix_c_fixed(int32_t **samples, int16_t (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = v1 = 0;
             for (j = 0; j < in_ch; j++) {
-                v0 += (int64_t)samples[j][i] * matrix[j][0];
-                v1 += (int64_t)samples[j][i] * matrix[j][1];
+                v0 += (int64_t)samples[j][i] * matrix[0][j];
+                v1 += (int64_t)samples[j][i] * matrix[1][j];
             }
             samples[0][i] = (v0+2048)>>12;
             samples[1][i] = (v1+2048)>>12;
@@ -257,7 +257,7 @@ static void ac3_downmix_c_fixed(int32_t **samples, int16_t (*matrix)[2],
         for (i = 0; i < len; i++) {
             v0 = 0;
             for (j = 0; j < in_ch; j++)
-                v0 += (int64_t)samples[j][i] * matrix[j][0];
+                v0 += (int64_t)samples[j][i] * matrix[0][j];
             samples[0][i] = (v0+2048)>>12;
         }
     }
diff --git a/libavcodec/ac3dsp.h b/libavcodec/ac3dsp.h
index ed98c8c..b4de307 100644
--- a/libavcodec/ac3dsp.h
+++ b/libavcodec/ac3dsp.h
@@ -132,10 +132,10 @@ typedef struct AC3DSPContext {
     void (*sum_square_butterfly_float)(float sum[4], const float *coef0,
                                        const float *coef1, int len);
 
-    void (*downmix)(float **samples, float (*matrix)[2], int out_ch,
+    void (*downmix)(float **samples, float **matrix, int out_ch,
                     int in_ch, int len);
 
-    void (*downmix_fixed)(int32_t **samples, int16_t (*matrix)[2], int out_ch,
+    void (*downmix_fixed)(int32_t **samples, int16_t **matrix, int out_ch,
                           int in_ch, int len);
 
     /**
diff --git a/libavcodec/x86/ac3dsp_init.c b/libavcodec/x86/ac3dsp_init.c
index 9fd0aef..edb6c60 100644
--- a/libavcodec/x86/ac3dsp_init.c
+++ b/libavcodec/x86/ac3dsp_init.c
@@ -76,8 +76,8 @@ void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
 #define MIX5(mono, stereo)                                      \
     __asm__ volatile (                                          \
         "movss           0(%1), %%xmm5          \n"             \
-        "movss           8(%1), %%xmm6          \n"             \
-        "movss          24(%1), %%xmm7          \n"             \
+        "movss           4(%1), %%xmm6          \n"             \
+        "movss          12(%1), %%xmm7          \n"             \
         "shufps     $0, %%xmm5, %%xmm5          \n"             \
         "shufps     $0, %%xmm6, %%xmm6          \n"             \
         "shufps     $0, %%xmm7, %%xmm7          \n"             \
@@ -102,7 +102,7 @@ void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
         "add               $16, %0              \n"             \
         "jl                 1b                  \n"             \
         : "+&r"(i)                                              \
-        : "r"(matrix),                                          \
+        : "r"(matrix[0]),                                          \
           "r"(samples[0] + len),                                \
           "r"(samples[1] + len),                                \
           "r"(samples[2] + len),                                \
@@ -146,22 +146,22 @@ void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
         : "memory"                                              \
     );
 
-static void ac3_downmix_sse(float **samples, float (*matrix)[2],
+static void ac3_downmix_sse(float **samples, float **matrix,
                             int out_ch, int in_ch, int len)
 {
-    int (*matrix_cmp)[2] = (int(*)[2])matrix;
+    int **matrix_cmp = (int **)matrix;
     intptr_t i, j, k, m;
 
     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]))) {
+        !(matrix_cmp[1][0] | matrix_cmp[0][2]   |
+          matrix_cmp[1][3] | matrix_cmp[0][4]   |
+          (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
+          (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
         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]) {
+               matrix_cmp[0][0] == matrix_cmp[0][2] &&
+               matrix_cmp[0][3] == matrix_cmp[0][4]) {
         MIX5(IF1, IF0);
     } else {
         LOCAL_ALIGNED(16, float, matrix_simd, [AC3_MAX_CHANNELS], [2][4]);
@@ -171,18 +171,20 @@ static void ac3_downmix_sse(float **samples, float (*matrix)[2],
             samp[j] = samples[j] + len;
 
         j = 2 * in_ch * sizeof(float);
+        k =     in_ch * sizeof(float);
         __asm__ volatile (
             "1:                                 \n"
+            "sub             $4, %1             \n"
             "sub             $8, %0             \n"
-            "movss     (%2, %0), %%xmm4         \n"
-            "movss    4(%2, %0), %%xmm5         \n"
+            "movss     (%3, %1), %%xmm4         \n"
+            "movss     (%4, %1), %%xmm5         \n"
             "shufps          $0, %%xmm4, %%xmm4 \n"
             "shufps          $0, %%xmm5, %%xmm5 \n"
-            "movaps      %%xmm4,   (%1, %0, 4)  \n"
-            "movaps      %%xmm5, 16(%1, %0, 4)  \n"
+            "movaps      %%xmm4,   (%2, %0, 4)  \n"
+            "movaps      %%xmm5, 16(%2, %0, 4)  \n"
             "jg              1b                 \n"
-            : "+&r"(j)
-            : "r"(matrix_simd), "r"(matrix)
+            : "+&r"(j), "+&r"(k)
+            : "r"(matrix_simd), "r"(matrix[0]), "r"(matrix[1])
             : "memory"
         );
         if (out_ch == 2) {