Merge remote-tracking branch 'mans/ac3'

* mans/ac3:
  ac3enc: slightly faster quantize_mantissas_blk_ch()
  ac3enc: NEON optimised sum_square_butterfly_float
  ac3enc: neon optimised sum_square_butterfly_int32
  ac3enc: move inner loop of compute_rematrixing_strategy to ac3dsp

Conflicts:
	libavcodec/ac3enc_template.c

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

8 files changed, 128 insertions, 19 deletions

diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
index 96bd123e6f..a414db4107 100644
--- a/libavcodec/ac3dsp.c
+++ b/libavcodec/ac3dsp.c
@@ -23,6 +23,7 @@
 #include "avcodec.h"
 #include "ac3.h"
 #include "ac3dsp.h"
+#include "mathops.h"
 
 static void ac3_exponent_min_c(uint8_t *exp, int num_reuse_blocks, int nb_coefs)
 {
@@ -169,6 +170,48 @@ static void ac3_extract_exponents_c(uint8_t *exp, int32_t *coef, int nb_coefs)
     }
 }
 
+static void ac3_sum_square_butterfly_int32_c(int64_t sum[4],
+                                             const int32_t *coef0,
+                                             const int32_t *coef1,
+                                             int len)
+{
+    int i;
+
+    sum[0] = sum[1] = sum[2] = sum[3] = 0;
+
+    for (i = 0; i < len; i++) {
+        int lt = coef0[i];
+        int rt = coef1[i];
+        int md = lt + rt;
+        int sd = lt - rt;
+        MAC64(sum[0], lt, lt);
+        MAC64(sum[1], rt, rt);
+        MAC64(sum[2], md, md);
+        MAC64(sum[3], sd, sd);
+    }
+}
+
+static void ac3_sum_square_butterfly_float_c(float sum[4],
+                                             const float *coef0,
+                                             const float *coef1,
+                                             int len)
+{
+    int i;
+
+    sum[0] = sum[1] = sum[2] = sum[3] = 0;
+
+    for (i = 0; i < len; i++) {
+        float lt = coef0[i];
+        float rt = coef1[i];
+        float md = lt + rt;
+        float sd = lt - rt;
+        sum[0] += lt * lt;
+        sum[1] += rt * rt;
+        sum[2] += md * md;
+        sum[3] += sd * sd;
+    }
+}
+
 av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)
 {
     c->ac3_exponent_min = ac3_exponent_min_c;
@@ -180,6 +223,8 @@ av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)
     c->update_bap_counts = ac3_update_bap_counts_c;
     c->compute_mantissa_size = ac3_compute_mantissa_size_c;
     c->extract_exponents = ac3_extract_exponents_c;
+    c->sum_square_butterfly_int32 = ac3_sum_square_butterfly_int32_c;
+    c->sum_square_butterfly_float = ac3_sum_square_butterfly_float_c;
 
     if (ARCH_ARM)
         ff_ac3dsp_init_arm(c, bit_exact);
diff --git a/libavcodec/ac3dsp.h b/libavcodec/ac3dsp.h
index d7131d2c03..f0a6999ecb 100644
--- a/libavcodec/ac3dsp.h
+++ b/libavcodec/ac3dsp.h
@@ -125,6 +125,12 @@ typedef struct AC3DSPContext {
     int (*compute_mantissa_size)(uint16_t mant_cnt[6][16]);
 
     void (*extract_exponents)(uint8_t *exp, int32_t *coef, int nb_coefs);
+
+    void (*sum_square_butterfly_int32)(int64_t sum[4], const int32_t *coef0,
+                                       const int32_t *coef1, int len);
+
+    void (*sum_square_butterfly_float)(float sum[4], const float *coef0,
+                                       const float *coef1, int len);
 } AC3DSPContext;
 
 void ff_ac3dsp_init    (AC3DSPContext *c, int bit_exact);
diff --git a/libavcodec/ac3enc.c b/libavcodec/ac3enc.c
index 44c16073a2..48cafd2e81 100644
--- a/libavcodec/ac3enc.c
+++ b/libavcodec/ac3enc.c
@@ -1163,14 +1163,11 @@ static void quantize_mantissas_blk_ch(AC3Mant *s, int32_t *fixed_coef,
     int i;
 
     for (i = start_freq; i < end_freq; i++) {
-        int v;
         int c = fixed_coef[i];
         int e = exp[i];
-        int b = bap[i];
-        switch (b) {
-        case 0:
-            v = 0;
-            break;
+        int v = bap[i];
+        if (v)
+        switch (v) {
         case 1:
             v = sym_quant(c, e, 3);
             switch (s->mant1_cnt) {
@@ -1239,7 +1236,7 @@ static void quantize_mantissas_blk_ch(AC3Mant *s, int32_t *fixed_coef,
             v = asym_quant(c, e, 16);
             break;
         default:
-            v = asym_quant(c, e, b - 1);
+            v = asym_quant(c, e, v - 1);
             break;
         }
         qmant[i] = v;
diff --git a/libavcodec/ac3enc_fixed.c b/libavcodec/ac3enc_fixed.c
index 0c7f81b7d3..46d971ddf7 100644
--- a/libavcodec/ac3enc_fixed.c
+++ b/libavcodec/ac3enc_fixed.c
@@ -103,6 +103,12 @@ static void scale_coefficients(AC3EncodeContext *s)
     }
 }
 
+static void sum_square_butterfly(AC3EncodeContext *s, int64_t sum[4],
+                                 const int32_t *coef0, const int32_t *coef1,
+                                 int len)
+{
+    s->ac3dsp.sum_square_butterfly_int32(sum, coef0, coef1, len);
+}
 
 /**
  * Clip MDCT coefficients to allowable range.
diff --git a/libavcodec/ac3enc_float.c b/libavcodec/ac3enc_float.c
index 3876a8007b..c87ae40e86 100644
--- a/libavcodec/ac3enc_float.c
+++ b/libavcodec/ac3enc_float.c
@@ -110,6 +110,12 @@ static void scale_coefficients(AC3EncodeContext *s)
                                chan_size * (s->channels + cpl));
 }
 
+static void sum_square_butterfly(AC3EncodeContext *s, float sum[4],
+                                 const float *coef0, const float *coef1,
+                                 int len)
+{
+    s->ac3dsp.sum_square_butterfly_float(sum, coef0, coef1, len);
+}
 
 /**
  * Clip MDCT coefficients to allowable range.
diff --git a/libavcodec/ac3enc_template.c b/libavcodec/ac3enc_template.c
index 7e2bf3a9ca..85513b1cdd 100644
--- a/libavcodec/ac3enc_template.c
+++ b/libavcodec/ac3enc_template.c
@@ -43,6 +43,9 @@ static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len)
 
 static CoefType calc_cpl_coord(CoefSumType energy_ch, CoefSumType energy_cpl);
 
+static void sum_square_butterfly(AC3EncodeContext *s, CoefSumType sum[4],
+                                 const CoefType *coef0, const CoefType *coef1,
+                                 int len);
 
 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
 {
@@ -334,7 +337,7 @@ static void apply_channel_coupling(AC3EncodeContext *s)
 static void compute_rematrixing_strategy(AC3EncodeContext *s)
 {
     int nb_coefs;
-    int blk, bnd, i;
+    int blk, bnd;
     AC3Block *block, *av_uninit(block0);
 
     if (s->channel_mode != AC3_CHMODE_STEREO)
@@ -362,17 +365,9 @@ static void compute_rematrixing_strategy(AC3EncodeContext *s)
             /* calculate calculate sum of squared coeffs for one band in one block */
             int start = ff_ac3_rematrix_band_tab[bnd];
             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
-            CoefSumType sum[4] = {0,};
-            for (i = start; i < end; i++) {
-                CoefType lt = block->mdct_coef[1][i];
-                CoefType rt = block->mdct_coef[2][i];
-                CoefType md = lt + rt;
-                CoefType sd = lt - rt;
-                MAC_COEF(sum[0], lt, lt);
-                MAC_COEF(sum[1], rt, rt);
-                MAC_COEF(sum[2], md, md);
-                MAC_COEF(sum[3], sd, sd);
-            }
+            CoefSumType sum[4];
+            sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
+                                 block->mdct_coef[2] + start, end - start);
 
             /* compare sums to determine if rematrixing will be used for this band */
             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
diff --git a/libavcodec/arm/ac3dsp_init_arm.c b/libavcodec/arm/ac3dsp_init_arm.c
index 5a9fdf0071..8ab685496d 100644
--- a/libavcodec/arm/ac3dsp_init_arm.c
+++ b/libavcodec/arm/ac3dsp_init_arm.c
@@ -29,6 +29,14 @@ void ff_ac3_lshift_int16_neon(int16_t *src, unsigned len, unsigned shift);
 void ff_ac3_rshift_int32_neon(int32_t *src, unsigned len, unsigned shift);
 void ff_float_to_fixed24_neon(int32_t *dst, const float *src, unsigned int len);
 void ff_ac3_extract_exponents_neon(uint8_t *exp, int32_t *coef, int nb_coefs);
+void ff_ac3_sum_square_butterfly_int32_neon(int64_t sum[4],
+                                            const int32_t *coef0,
+                                            const int32_t *coef1,
+                                            int len);
+void ff_ac3_sum_square_butterfly_float_neon(float sum[4],
+                                            const float *coef0,
+                                            const float *coef1,
+                                            int len);
 
 void ff_ac3_bit_alloc_calc_bap_armv6(int16_t *mask, int16_t *psd,
                                      int start, int end,
@@ -52,5 +60,7 @@ av_cold void ff_ac3dsp_init_arm(AC3DSPContext *c, int bit_exact)
         c->ac3_rshift_int32      = ff_ac3_rshift_int32_neon;
         c->float_to_fixed24      = ff_float_to_fixed24_neon;
         c->extract_exponents     = ff_ac3_extract_exponents_neon;
+        c->sum_square_butterfly_int32 = ff_ac3_sum_square_butterfly_int32_neon;
+        c->sum_square_butterfly_float = ff_ac3_sum_square_butterfly_float_neon;
     }
 }
diff --git a/libavcodec/arm/ac3dsp_neon.S b/libavcodec/arm/ac3dsp_neon.S
index e97197c27a..152afbfd3c 100644
--- a/libavcodec/arm/ac3dsp_neon.S
+++ b/libavcodec/arm/ac3dsp_neon.S
@@ -108,3 +108,47 @@ function ff_ac3_extract_exponents_neon, export=1
         bgt             1b
         bx              lr
 endfunc
+
+function ff_ac3_sum_square_butterfly_int32_neon, export=1
+        vmov.i64        q0,  #0
+        vmov.i64        q1,  #0
+        vmov.i64        q2,  #0
+        vmov.i64        q3,  #0
+1:
+        vld1.32         {d16},    [r1]!
+        vld1.32         {d17},    [r2]!
+        vadd.s32        d18, d16, d17
+        vsub.s32        d19, d16, d17
+        vmlal.s32       q0,  d16, d16
+        vmlal.s32       q1,  d17, d17
+        vmlal.s32       q2,  d18, d18
+        vmlal.s32       q3,  d19, d19
+        subs            r3,  r3,  #2
+        bgt             1b
+        vadd.s64        d0,  d0,  d1
+        vadd.s64        d1,  d2,  d3
+        vadd.s64        d2,  d4,  d5
+        vadd.s64        d3,  d6,  d7
+        vst1.64         {q0-q1},  [r0]
+        bx              lr
+endfunc
+
+function ff_ac3_sum_square_butterfly_float_neon, export=1
+        vmov.f32        q0,  #0.0
+        vmov.f32        q1,  #0.0
+1:
+        vld1.32         {d16},    [r1]!
+        vld1.32         {d17},    [r2]!
+        vadd.f32        d18, d16, d17
+        vsub.f32        d19, d16, d17
+        vmla.f32        d0,  d16, d16
+        vmla.f32        d1,  d17, d17
+        vmla.f32        d2,  d18, d18
+        vmla.f32        d3,  d19, d19
+        subs            r3,  r3,  #2
+        bgt             1b
+        vpadd.f32       d0,  d0,  d1
+        vpadd.f32       d1,  d2,  d3
+        vst1.32         {q0},     [r0]
+        bx              lr
+endfunc