Add 8088flex TMV video decoder.
[ffmpeg.git] / libavcodec / wavpack.c
index 4a7a9e5..10254f0 100644 (file)
 #define WV_HYBRID_BITRATE 0x00000200
 #define WV_HYBRID_BALANCE 0x00000400
 
+#define WV_FLT_SHIFT_ONES 0x01
+#define WV_FLT_SHIFT_SAME 0x02
+#define WV_FLT_SHIFT_SENT 0x04
+#define WV_FLT_ZERO_SENT  0x08
+#define WV_FLT_ZERO_SIGN  0x10
+
 enum WP_ID_Flags{
     WP_IDF_MASK   = 0x1F,
     WP_IDF_IGNORE = 0x20,
@@ -57,7 +63,7 @@ enum WP_ID{
     WP_ID_INT32INFO,
     WP_ID_DATA,
     WP_ID_CORR,
-    WP_ID_FLT,
+    WP_ID_EXTRABITS,
     WP_ID_CHANINFO
 };
 
@@ -85,14 +91,21 @@ typedef struct WavpackContext {
     int joint;
     uint32_t CRC;
     GetBitContext gb;
+    int got_extra_bits;
+    uint32_t crc_extra_bits;
+    GetBitContext gb_extra_bits;
     int data_size; // in bits
     int samples;
     int terms;
     Decorr decorr[MAX_TERMS];
     int zero, one, zeroes;
+    int extra_bits;
     int and, or, shift;
     int post_shift;
     int hybrid, hybrid_bitrate;
+    int float_flag;
+    int float_shift;
+    int float_max_exp;
     WvChannel ch[2];
 } WavpackContext;
 
@@ -337,139 +350,106 @@ static int wv_get_value(WavpackContext *ctx, GetBitContext *gb, int channel, int
     return sign ? ~ret : ret;
 }
 
-static int wv_unpack_stereo(WavpackContext *s, GetBitContext *gb, int16_t *dst)
+static inline int wv_get_value_integer(WavpackContext *s, uint32_t *crc, int S)
 {
-    int i, j, count = 0;
-    int last, t;
-    int A, B, L, L2, R, R2, bit;
-    int pos = 0;
-    uint32_t crc = 0xFFFFFFFF;
-
-    s->one = s->zero = s->zeroes = 0;
-    do{
-        L = wv_get_value(s, gb, 0, &last);
-        if(last) break;
-        R = wv_get_value(s, gb, 1, &last);
-        if(last) break;
-        for(i = 0; i < s->terms; i++){
-            t = s->decorr[i].value;
-            j = 0;
-            if(t > 0){
-                if(t > 8){
-                    if(t & 1){
-                        A = 2 * s->decorr[i].samplesA[0] - s->decorr[i].samplesA[1];
-                        B = 2 * s->decorr[i].samplesB[0] - s->decorr[i].samplesB[1];
-                    }else{
-                        A = (3 * s->decorr[i].samplesA[0] - s->decorr[i].samplesA[1]) >> 1;
-                        B = (3 * s->decorr[i].samplesB[0] - s->decorr[i].samplesB[1]) >> 1;
-                    }
-                    s->decorr[i].samplesA[1] = s->decorr[i].samplesA[0];
-                    s->decorr[i].samplesB[1] = s->decorr[i].samplesB[0];
-                    j = 0;
-                }else{
-                    A = s->decorr[i].samplesA[pos];
-                    B = s->decorr[i].samplesB[pos];
-                    j = (pos + t) & 7;
-                }
-                L2 = L + ((s->decorr[i].weightA * A + 512) >> 10);
-                R2 = R + ((s->decorr[i].weightB * B + 512) >> 10);
-                if(A && L) s->decorr[i].weightA -= ((((L ^ A) >> 30) & 2) - 1) * s->decorr[i].delta;
-                if(B && R) s->decorr[i].weightB -= ((((R ^ B) >> 30) & 2) - 1) * s->decorr[i].delta;
-                s->decorr[i].samplesA[j] = L = L2;
-                s->decorr[i].samplesB[j] = R = R2;
-            }else if(t == -1){
-                L2 = L + ((s->decorr[i].weightA * s->decorr[i].samplesA[0] + 512) >> 10);
-                UPDATE_WEIGHT_CLIP(s->decorr[i].weightA, s->decorr[i].delta, s->decorr[i].samplesA[0], L);
-                L = L2;
-                R2 = R + ((s->decorr[i].weightB * L2 + 512) >> 10);
-                UPDATE_WEIGHT_CLIP(s->decorr[i].weightB, s->decorr[i].delta, L2, R);
-                R = R2;
-                s->decorr[i].samplesA[0] = R;
-            }else{
-                R2 = R + ((s->decorr[i].weightB * s->decorr[i].samplesB[0] + 512) >> 10);
-                UPDATE_WEIGHT_CLIP(s->decorr[i].weightB, s->decorr[i].delta, s->decorr[i].samplesB[0], R);
-                R = R2;
+    int bit;
 
-                if(t == -3){
-                    R2 = s->decorr[i].samplesA[0];
-                    s->decorr[i].samplesA[0] = R;
-                }
+    if(s->extra_bits){
+        S <<= s->extra_bits;
 
-                L2 = L + ((s->decorr[i].weightA * R2 + 512) >> 10);
-                UPDATE_WEIGHT_CLIP(s->decorr[i].weightA, s->decorr[i].delta, R2, L);
-                L = L2;
-                s->decorr[i].samplesB[0] = L;
-            }
+        if(s->got_extra_bits){
+            S |= get_bits(&s->gb_extra_bits, s->extra_bits);
+            *crc = *crc * 9 + (S&0xffff) * 3 + ((unsigned)S>>16);
         }
-        pos = (pos + 1) & 7;
-        if(s->joint)
-            L += (R -= (L >> 1));
-        crc = (crc * 3 + L) * 3 + R;
-        bit = (L & s->and) | s->or;
-        *dst++ = (((L + bit) << s->shift) - bit) << s->post_shift;
-        bit = (R & s->and) | s->or;
-        *dst++ = (((R + bit) << s->shift) - bit) << s->post_shift;
-        count++;
-    }while(!last && count < s->samples);
-
-    if(crc != s->CRC){
-        av_log(s->avctx, AV_LOG_ERROR, "CRC error\n");
-        return -1;
     }
-    return count * 2;
+    bit = (S & s->and) | s->or;
+    return (((S + bit) << s->shift) - bit) << s->post_shift;
 }
 
-static int wv_unpack_mono(WavpackContext *s, GetBitContext *gb, int16_t *dst)
+static float wv_get_value_float(WavpackContext *s, uint32_t *crc, int S)
 {
-    int i, j, count = 0;
-    int last, t;
-    int A, S, T, bit;
-    int pos = 0;
-    uint32_t crc = 0xFFFFFFFF;
+    union {
+        float    f;
+        uint32_t u;
+    } value;
 
-    s->one = s->zero = s->zeroes = 0;
-    do{
-        T = wv_get_value(s, gb, 0, &last);
-        S = 0;
-        if(last) break;
-        for(i = 0; i < s->terms; i++){
-            t = s->decorr[i].value;
-            if(t > 8){
-                if(t & 1)
-                    A = 2 * s->decorr[i].samplesA[0] - s->decorr[i].samplesA[1];
-                else
-                    A = (3 * s->decorr[i].samplesA[0] - s->decorr[i].samplesA[1]) >> 1;
-                s->decorr[i].samplesA[1] = s->decorr[i].samplesA[0];
-                j = 0;
+    int sign;
+    int exp = s->float_max_exp;
+
+    if(s->got_extra_bits){
+        const int max_bits = 1 + 23 + 8 + 1;
+        const int left_bits = s->gb_extra_bits.size_in_bits - get_bits_count(&s->gb_extra_bits);
+
+        if(left_bits + 8 * FF_INPUT_BUFFER_PADDING_SIZE < max_bits)
+            return 0.0;
+    }
+
+    if(S){
+        S <<= s->float_shift;
+        sign = S < 0;
+        if(sign)
+            S = -S;
+        if(S >= 0x1000000){
+            if(s->got_extra_bits && get_bits1(&s->gb_extra_bits)){
+                S = get_bits(&s->gb_extra_bits, 23);
             }else{
-                A = s->decorr[i].samplesA[pos];
-                j = (pos + t) & 7;
+                S = 0;
+            }
+            exp = 255;
+        }else if(exp){
+            int shift = 23 - av_log2(S);
+            exp = s->float_max_exp;
+            if(exp <= shift){
+                shift = --exp;
+            }
+            exp -= shift;
+
+            if(shift){
+                S <<= shift;
+                if((s->float_flag & WV_FLT_SHIFT_ONES) ||
+                   (s->got_extra_bits && (s->float_flag & WV_FLT_SHIFT_SAME) && get_bits1(&s->gb_extra_bits)) ){
+                    S |= (1 << shift) - 1;
+                } else if(s->got_extra_bits && (s->float_flag & WV_FLT_SHIFT_SENT)){
+                    S |= get_bits(&s->gb_extra_bits, shift);
+                }
+            }
+        }else{
+            exp = s->float_max_exp;
+        }
+        S &= 0x7fffff;
+    }else{
+        sign = 0;
+        exp = 0;
+        if(s->got_extra_bits && (s->float_flag & WV_FLT_ZERO_SENT)){
+            if(get_bits1(&s->gb_extra_bits)){
+                S = get_bits(&s->gb_extra_bits, 23);
+                if(s->float_max_exp >= 25)
+                    exp = get_bits(&s->gb_extra_bits, 8);
+                sign = get_bits1(&s->gb_extra_bits);
+            }else{
+                if(s->float_flag & WV_FLT_ZERO_SIGN)
+                    sign = get_bits1(&s->gb_extra_bits);
             }
-            S = T + ((s->decorr[i].weightA * A + 512) >> 10);
-            if(A && T) s->decorr[i].weightA -= ((((T ^ A) >> 30) & 2) - 1) * s->decorr[i].delta;
-            s->decorr[i].samplesA[j] = T = S;
         }
-        pos = (pos + 1) & 7;
-        crc = crc * 3 + S;
-        bit = (S & s->and) | s->or;
-        *dst++ = (((S + bit) << s->shift) - bit) << s->post_shift;
-        count++;
-    }while(!last && count < s->samples);
-
-    if(crc != s->CRC){
-        av_log(s->avctx, AV_LOG_ERROR, "CRC error\n");
-        return -1;
     }
-    return count;
+
+    *crc = *crc * 27 + S * 9 + exp * 3 + sign;
+
+    value.u = (sign << 31) | (exp << 23) | S;
+    return value.f;
 }
 
-static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t *dst)
+static inline int wv_unpack_stereo(WavpackContext *s, GetBitContext *gb, void *dst, const int type)
 {
     int i, j, count = 0;
     int last, t;
-    int A, B, L, L2, R, R2, bit;
+    int A, B, L, L2, R, R2;
     int pos = 0;
     uint32_t crc = 0xFFFFFFFF;
+    uint32_t crc_extra_bits = 0xFFFFFFFF;
+    int16_t *dst16 = dst;
+    int32_t *dst32 = dst;
+    float   *dstfl = dst;
 
     s->one = s->zero = s->zeroes = 0;
     do{
@@ -479,7 +459,6 @@ static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t
         if(last) break;
         for(i = 0; i < s->terms; i++){
             t = s->decorr[i].value;
-            j = 0;
             if(t > 0){
                 if(t > 8){
                     if(t & 1){
@@ -497,22 +476,36 @@ static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t
                     B = s->decorr[i].samplesB[pos];
                     j = (pos + t) & 7;
                 }
-                L2 = L + ((s->decorr[i].weightA * (int64_t)A + 512) >> 10);
-                R2 = R + ((s->decorr[i].weightB * (int64_t)B + 512) >> 10);
+                if(type != SAMPLE_FMT_S16){
+                    L2 = L + ((s->decorr[i].weightA * (int64_t)A + 512) >> 10);
+                    R2 = R + ((s->decorr[i].weightB * (int64_t)B + 512) >> 10);
+                }else{
+                    L2 = L + ((s->decorr[i].weightA * A + 512) >> 10);
+                    R2 = R + ((s->decorr[i].weightB * B + 512) >> 10);
+                }
                 if(A && L) s->decorr[i].weightA -= ((((L ^ A) >> 30) & 2) - 1) * s->decorr[i].delta;
                 if(B && R) s->decorr[i].weightB -= ((((R ^ B) >> 30) & 2) - 1) * s->decorr[i].delta;
                 s->decorr[i].samplesA[j] = L = L2;
                 s->decorr[i].samplesB[j] = R = R2;
             }else if(t == -1){
-                L2 = L + ((s->decorr[i].weightA * (int64_t)s->decorr[i].samplesA[0] + 512) >> 10);
+                if(type != SAMPLE_FMT_S16)
+                    L2 = L + ((s->decorr[i].weightA * (int64_t)s->decorr[i].samplesA[0] + 512) >> 10);
+                else
+                    L2 = L + ((s->decorr[i].weightA * s->decorr[i].samplesA[0] + 512) >> 10);
                 UPDATE_WEIGHT_CLIP(s->decorr[i].weightA, s->decorr[i].delta, s->decorr[i].samplesA[0], L);
                 L = L2;
-                R2 = R + ((s->decorr[i].weightB * (int64_t)L2 + 512) >> 10);
+                if(type != SAMPLE_FMT_S16)
+                    R2 = R + ((s->decorr[i].weightB * (int64_t)L2 + 512) >> 10);
+                else
+                    R2 = R + ((s->decorr[i].weightB * L2 + 512) >> 10);
                 UPDATE_WEIGHT_CLIP(s->decorr[i].weightB, s->decorr[i].delta, L2, R);
                 R = R2;
                 s->decorr[i].samplesA[0] = R;
             }else{
-                R2 = R + ((s->decorr[i].weightB * (int64_t)s->decorr[i].samplesB[0] + 512) >> 10);
+                if(type != SAMPLE_FMT_S16)
+                    R2 = R + ((s->decorr[i].weightB * (int64_t)s->decorr[i].samplesB[0] + 512) >> 10);
+                else
+                    R2 = R + ((s->decorr[i].weightB * s->decorr[i].samplesB[0] + 512) >> 10);
                 UPDATE_WEIGHT_CLIP(s->decorr[i].weightB, s->decorr[i].delta, s->decorr[i].samplesB[0], R);
                 R = R2;
 
@@ -521,7 +514,10 @@ static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t
                     s->decorr[i].samplesA[0] = R;
                 }
 
-                L2 = L + ((s->decorr[i].weightA * (int64_t)R2 + 512) >> 10);
+                if(type != SAMPLE_FMT_S16)
+                    L2 = L + ((s->decorr[i].weightA * (int64_t)R2 + 512) >> 10);
+                else
+                    L2 = L + ((s->decorr[i].weightA * R2 + 512) >> 10);
                 UPDATE_WEIGHT_CLIP(s->decorr[i].weightA, s->decorr[i].delta, R2, L);
                 L = L2;
                 s->decorr[i].samplesB[0] = L;
@@ -531,10 +527,17 @@ static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t
         if(s->joint)
             L += (R -= (L >> 1));
         crc = (crc * 3 + L) * 3 + R;
-        bit = (L & s->and) | s->or;
-        *dst++ = (((L + bit) << s->shift) - bit) << s->post_shift;
-        bit = (R & s->and) | s->or;
-        *dst++ = (((R + bit) << s->shift) - bit) << s->post_shift;
+
+        if(type == SAMPLE_FMT_FLT){
+            *dstfl++ = wv_get_value_float(s, &crc_extra_bits, L);
+            *dstfl++ = wv_get_value_float(s, &crc_extra_bits, R);
+        } else if(type == SAMPLE_FMT_S32){
+            *dst32++ = wv_get_value_integer(s, &crc_extra_bits, L);
+            *dst32++ = wv_get_value_integer(s, &crc_extra_bits, R);
+        } else {
+            *dst16++ = wv_get_value_integer(s, &crc_extra_bits, L);
+            *dst16++ = wv_get_value_integer(s, &crc_extra_bits, R);
+        }
         count++;
     }while(!last && count < s->samples);
 
@@ -542,16 +545,24 @@ static int wv_unpack_stereo_hires(WavpackContext *s, GetBitContext *gb, int32_t
         av_log(s->avctx, AV_LOG_ERROR, "CRC error\n");
         return -1;
     }
+    if(s->got_extra_bits && crc_extra_bits != s->crc_extra_bits){
+        av_log(s->avctx, AV_LOG_ERROR, "Extra bits CRC error\n");
+        return -1;
+    }
     return count * 2;
 }
 
-static int wv_unpack_mono_hires(WavpackContext *s, GetBitContext *gb, int32_t *dst)
+static inline int wv_unpack_mono(WavpackContext *s, GetBitContext *gb, void *dst, const int type)
 {
     int i, j, count = 0;
     int last, t;
-    int A, S, T, bit;
+    int A, S, T;
     int pos = 0;
     uint32_t crc = 0xFFFFFFFF;
+    uint32_t crc_extra_bits = 0xFFFFFFFF;
+    int16_t *dst16 = dst;
+    int32_t *dst32 = dst;
+    float   *dstfl = dst;
 
     s->one = s->zero = s->zeroes = 0;
     do{
@@ -571,14 +582,22 @@ static int wv_unpack_mono_hires(WavpackContext *s, GetBitContext *gb, int32_t *d
                 A = s->decorr[i].samplesA[pos];
                 j = (pos + t) & 7;
             }
-            S = T + ((s->decorr[i].weightA * (int64_t)A + 512) >> 10);
+            if(type != SAMPLE_FMT_S16)
+                S = T + ((s->decorr[i].weightA * (int64_t)A + 512) >> 10);
+            else
+                S = T + ((s->decorr[i].weightA * A + 512) >> 10);
             if(A && T) s->decorr[i].weightA -= ((((T ^ A) >> 30) & 2) - 1) * s->decorr[i].delta;
             s->decorr[i].samplesA[j] = T = S;
         }
         pos = (pos + 1) & 7;
         crc = crc * 3 + S;
-        bit = (S & s->and) | s->or;
-        *dst++ = (((S + bit) << s->shift) - bit) << s->post_shift;
+
+        if(type == SAMPLE_FMT_FLT)
+            *dstfl++ = wv_get_value_float(s, &crc_extra_bits, S);
+        else if(type == SAMPLE_FMT_S32)
+            *dst32++ = wv_get_value_integer(s, &crc_extra_bits, S);
+        else
+            *dst16++ = wv_get_value_integer(s, &crc_extra_bits, S);
         count++;
     }while(!last && count < s->samples);
 
@@ -586,6 +605,10 @@ static int wv_unpack_mono_hires(WavpackContext *s, GetBitContext *gb, int32_t *d
         av_log(s->avctx, AV_LOG_ERROR, "CRC error\n");
         return -1;
     }
+    if(s->got_extra_bits && crc_extra_bits != s->crc_extra_bits){
+        av_log(s->avctx, AV_LOG_ERROR, "Extra bits CRC error\n");
+        return -1;
+    }
     return count;
 }
 
@@ -611,14 +634,13 @@ static int wavpack_decode_frame(AVCodecContext *avctx,
     const uint8_t *buf = avpkt->data;
     int buf_size = avpkt->size;
     WavpackContext *s = avctx->priv_data;
-    int16_t *samples = data;
-    int32_t *samples32 = data;
+    void *samples = data;
     int samplecount;
-    int got_terms = 0, got_weights = 0, got_samples = 0, got_entropy = 0, got_bs = 0;
+    int got_terms = 0, got_weights = 0, got_samples = 0, got_entropy = 0, got_bs = 0, got_float = 0;
     int got_hybrid = 0;
     const uint8_t* buf_end = buf + buf_size;
     int i, j, id, size, ssize, weights, t;
-    int bpp = avctx->bits_per_coded_sample <= 16 ? 2 : 4;
+    int bpp;
 
     if (buf_size == 0){
         *data_size = 0;
@@ -627,25 +649,39 @@ static int wavpack_decode_frame(AVCodecContext *avctx,
 
     memset(s->decorr, 0, MAX_TERMS * sizeof(Decorr));
     memset(s->ch, 0, sizeof(s->ch));
+    s->extra_bits = 0;
     s->and = s->or = s->shift = 0;
+    s->got_extra_bits = 0;
 
     s->samples = AV_RL32(buf); buf += 4;
     if(!s->samples){
         *data_size = 0;
         return buf_size;
     }
-    /* should not happen but who knows */
-    if(s->samples * bpp * avctx->channels > *data_size){
-        av_log(avctx, AV_LOG_ERROR, "Packet size is too big to be handled in lavc!\n");
-        return -1;
-    }
     s->frame_flags = AV_RL32(buf); buf += 4;
+    if(s->frame_flags&0x80){
+        bpp = sizeof(float);
+        avctx->sample_fmt = SAMPLE_FMT_FLT;
+    } else if((s->frame_flags&0x03) <= 1){
+        bpp = 2;
+        avctx->sample_fmt = SAMPLE_FMT_S16;
+    } else {
+        bpp = 4;
+        avctx->sample_fmt = SAMPLE_FMT_S32;
+    }
     s->stereo_in = (s->frame_flags & WV_FALSE_STEREO) ? 0 : s->stereo;
     s->joint = s->frame_flags & WV_JOINT_STEREO;
     s->hybrid = s->frame_flags & WV_HYBRID_MODE;
     s->hybrid_bitrate = s->frame_flags & WV_HYBRID_BITRATE;
     s->post_shift = 8 * (bpp-1-(s->frame_flags&0x03)) + ((s->frame_flags >> 13) & 0x1f);
     s->CRC = AV_RL32(buf); buf += 4;
+
+    /* should not happen but who knows */
+    if(s->samples * bpp * avctx->channels > *data_size){
+        av_log(avctx, AV_LOG_ERROR, "Packet size is too big to be handled in lavc!\n");
+        return -1;
+    }
+
     // parse metadata blocks
     while(buf < buf_end){
         id = *buf++;
@@ -780,12 +816,14 @@ static int wavpack_decode_frame(AVCodecContext *avctx,
             got_hybrid = 1;
             break;
         case WP_ID_INT32INFO:
-            if(size != 4 || *buf){
+            if(size != 4){
                 av_log(avctx, AV_LOG_ERROR, "Invalid INT32INFO, size = %i, sent_bits = %i\n", size, *buf);
                 buf += ssize;
                 continue;
             }
-            if(buf[1])
+            if(buf[0])
+                s->extra_bits = buf[0];
+            else if(buf[1])
                 s->shift = buf[1];
             else if(buf[2]){
                 s->and = s->or = 1;
@@ -796,12 +834,35 @@ static int wavpack_decode_frame(AVCodecContext *avctx,
             }
             buf += 4;
             break;
+        case WP_ID_FLOATINFO:
+            if(size != 4){
+                av_log(avctx, AV_LOG_ERROR, "Invalid FLOATINFO, size = %i\n", size);
+                buf += ssize;
+                continue;
+            }
+            s->float_flag = buf[0];
+            s->float_shift = buf[1];
+            s->float_max_exp = buf[2];
+            buf += 4;
+            got_float = 1;
+            break;
         case WP_ID_DATA:
             init_get_bits(&s->gb, buf, size * 8);
             s->data_size = size * 8;
             buf += size;
             got_bs = 1;
             break;
+        case WP_ID_EXTRABITS:
+            if(size <= 4){
+                av_log(avctx, AV_LOG_ERROR, "Invalid EXTRABITS, size = %i\n", size);
+                buf += size;
+                continue;
+            }
+            init_get_bits(&s->gb_extra_bits, buf, size * 8);
+            s->crc_extra_bits = get_bits_long(&s->gb_extra_bits, 32);
+            buf += size;
+            s->got_extra_bits = 1;
+            break;
         default:
             buf += size;
         }
@@ -831,29 +892,56 @@ static int wavpack_decode_frame(AVCodecContext *avctx,
         av_log(avctx, AV_LOG_ERROR, "Packed samples not found\n");
         return -1;
     }
+    if(!got_float && avctx->sample_fmt == SAMPLE_FMT_FLT){
+        av_log(avctx, AV_LOG_ERROR, "Float information not found\n");
+        return -1;
+    }
+    if(s->got_extra_bits && avctx->sample_fmt != SAMPLE_FMT_FLT){
+        const int size = s->gb_extra_bits.size_in_bits - get_bits_count(&s->gb_extra_bits);
+        const int wanted = s->samples * s->extra_bits << s->stereo_in;
+        if(size < wanted){
+            av_log(avctx, AV_LOG_ERROR, "Too small EXTRABITS\n");
+            s->got_extra_bits = 0;
+        }
+    }
 
     if(s->stereo_in){
-        if(bpp == 2)
-            samplecount = wv_unpack_stereo(s, &s->gb, samples);
+        if(avctx->sample_fmt == SAMPLE_FMT_S16)
+            samplecount = wv_unpack_stereo(s, &s->gb, samples, SAMPLE_FMT_S16);
+        else if(avctx->sample_fmt == SAMPLE_FMT_S32)
+            samplecount = wv_unpack_stereo(s, &s->gb, samples, SAMPLE_FMT_S32);
         else
-            samplecount = wv_unpack_stereo_hires(s, &s->gb, samples32);
+            samplecount = wv_unpack_stereo(s, &s->gb, samples, SAMPLE_FMT_FLT);
+
     }else{
-        if(bpp == 2)
-            samplecount = wv_unpack_mono(s, &s->gb, samples);
+        if(avctx->sample_fmt == SAMPLE_FMT_S16)
+            samplecount = wv_unpack_mono(s, &s->gb, samples, SAMPLE_FMT_S16);
+        else if(avctx->sample_fmt == SAMPLE_FMT_S32)
+            samplecount = wv_unpack_mono(s, &s->gb, samples, SAMPLE_FMT_S32);
         else
-            samplecount = wv_unpack_mono_hires(s, &s->gb, samples32);
-        if(s->stereo && bpp == 2){
-            int16_t *dst = samples + samplecount * 2;
-            int16_t *src = samples + samplecount;
+            samplecount = wv_unpack_mono(s, &s->gb, samples, SAMPLE_FMT_FLT);
+
+        if(s->stereo && avctx->sample_fmt == SAMPLE_FMT_S16){
+            int16_t *dst = (int16_t*)samples + samplecount * 2;
+            int16_t *src = (int16_t*)samples + samplecount;
+            int cnt = samplecount;
+            while(cnt--){
+                *--dst = *--src;
+                *--dst = *src;
+            }
+            samplecount *= 2;
+        }else if(s->stereo && avctx->sample_fmt == SAMPLE_FMT_S32){
+            int32_t *dst = (int32_t*)samples + samplecount * 2;
+            int32_t *src = (int32_t*)samples + samplecount;
             int cnt = samplecount;
             while(cnt--){
                 *--dst = *--src;
                 *--dst = *src;
             }
             samplecount *= 2;
-        }else if(s->stereo){ //32-bit output
-            int32_t *dst = samples32 + samplecount * 2;
-            int32_t *src = samples32 + samplecount;
+        }else if(s->stereo){
+            float *dst = (float*)samples + samplecount * 2;
+            float *src = (float*)samples + samplecount;
             int cnt = samplecount;
             while(cnt--){
                 *--dst = *--src;