lavc/j2kenc: Support AV_PIX_FMT_PAL8.
[ffmpeg.git] / libavcodec / celp_filters.c
index ef1a578..fafedd9 100644 (file)
 
 #include "avcodec.h"
 #include "celp_filters.h"
+#include "libavutil/avassert.h"
+#include "libavutil/common.h"
 
-void ff_celp_convolve_circ(
-        int16_t* fc_out,
-        const int16_t* fc_in,
-        const int16_t* filter,
-        int len)
+void ff_celp_convolve_circ(int16_t* fc_out, const int16_t* fc_in,
+                           const int16_t* filter, int len)
 {
     int i, k;
 
@@ -37,69 +36,184 @@ void ff_celp_convolve_circ(
 
     /* Since there are few pulses over an entire subframe (i.e. almost
        all fc_in[i] are zero) it is faster to loop over fc_in first. */
-    for(i=0; i<len; i++)
-    {
-        if(fc_in[i])
-        {
-            for(k=0; k<i; k++)
+    for (i = 0; i < len; i++) {
+        if (fc_in[i]) {
+            for (k = 0; k < i; k++)
                 fc_out[k] += (fc_in[i] * filter[len + k - i]) >> 15;
 
-            for(k=i; k<len; k++)
+            for (k = i; k < len; k++)
                 fc_out[k] += (fc_in[i] * filter[      k - i]) >> 15;
         }
     }
 }
 
-int ff_celp_lp_synthesis_filter(
-        int16_t *out,
-        const int16_t* filter_coeffs,
-        const int16_t* in,
-        int buffer_length,
-        int filter_length,
-        int stop_on_overflow,
-        int rounder)
+void ff_celp_circ_addf(float *out, const float *in,
+                       const float *lagged, int lag, float fac, int n)
+{
+    int k;
+    for (k = 0; k < lag; k++)
+        out[k] = in[k] + fac * lagged[n + k - lag];
+    for (; k < n; k++)
+        out[k] = in[k] + fac * lagged[    k - lag];
+}
+
+int ff_celp_lp_synthesis_filter(int16_t *out, const int16_t *filter_coeffs,
+                                const int16_t *in, int buffer_length,
+                                int filter_length, int stop_on_overflow,
+                                int shift, int rounder)
 {
     int i,n;
 
-    // This line is to avoid a +1 subtraction in the main loop.
-    filter_length++;
+    for (n = 0; n < buffer_length; n++) {
+        int sum = -rounder, sum1;
+        for (i = 1; i <= filter_length; i++)
+            sum += (unsigned)(filter_coeffs[i-1] * out[n-i]);
 
-    for(n=0; n<buffer_length; n++)
-    {
-        int sum = rounder;
-        for(i=1; i<filter_length; i++)
-            sum -= filter_coeffs[i-1] * out[n-i];
+        sum1 = ((-sum >> 12) + in[n]) >> shift;
+        sum  = av_clip_int16(sum1);
 
-        sum = (sum >> 12) + in[n];
+        if (stop_on_overflow && sum != sum1)
+            return 1;
 
-        if(sum + 0x8000 > 0xFFFFU)
-        {
-            if(stop_on_overflow)
-                return 1;
-            sum = (sum >> 31) ^ 32767;
-        }
         out[n] = sum;
     }
 
     return 0;
 }
 
-void ff_celp_lp_synthesis_filterf(
-        float *out,
-        const float* filter_coeffs,
-        const float* in,
-        int buffer_length,
-        int filter_length)
+void ff_celp_lp_synthesis_filterf(float *out, const float *filter_coeffs,
+                                  const float* in, int buffer_length,
+                                  int filter_length)
 {
     int i,n;
 
-    // This line is to avoid a +1 subtraction in the main loop
-    filter_length++;
+#if 0 // Unoptimized code path for improved readability
+    for (n = 0; n < buffer_length; n++) {
+        out[n] = in[n];
+        for (i = 1; i <= filter_length; i++)
+            out[n] -= filter_coeffs[i-1] * out[n-i];
+    }
+#else
+    float out0, out1, out2, out3;
+    float old_out0, old_out1, old_out2, old_out3;
+    float a,b,c;
+
+    a = filter_coeffs[0];
+    b = filter_coeffs[1];
+    c = filter_coeffs[2];
+    b -= filter_coeffs[0] * filter_coeffs[0];
+    c -= filter_coeffs[1] * filter_coeffs[0];
+    c -= filter_coeffs[0] * b;
+
+    av_assert2((filter_length&1)==0 && filter_length>=4);
+
+    old_out0 = out[-4];
+    old_out1 = out[-3];
+    old_out2 = out[-2];
+    old_out3 = out[-1];
+    for (n = 0; n <= buffer_length - 4; n+=4) {
+        float tmp0,tmp1,tmp2;
+        float val;
+
+        out0 = in[0];
+        out1 = in[1];
+        out2 = in[2];
+        out3 = in[3];
 
-    for(n=0; n<buffer_length; n++)
-    {
+        out0 -= filter_coeffs[2] * old_out1;
+        out1 -= filter_coeffs[2] * old_out2;
+        out2 -= filter_coeffs[2] * old_out3;
+
+        out0 -= filter_coeffs[1] * old_out2;
+        out1 -= filter_coeffs[1] * old_out3;
+
+        out0 -= filter_coeffs[0] * old_out3;
+
+        val = filter_coeffs[3];
+
+        out0 -= val * old_out0;
+        out1 -= val * old_out1;
+        out2 -= val * old_out2;
+        out3 -= val * old_out3;
+
+        for (i = 5; i < filter_length; i += 2) {
+            old_out3 = out[-i];
+            val = filter_coeffs[i-1];
+
+            out0 -= val * old_out3;
+            out1 -= val * old_out0;
+            out2 -= val * old_out1;
+            out3 -= val * old_out2;
+
+            old_out2 = out[-i-1];
+
+            val = filter_coeffs[i];
+
+            out0 -= val * old_out2;
+            out1 -= val * old_out3;
+            out2 -= val * old_out0;
+            out3 -= val * old_out1;
+
+            FFSWAP(float, old_out0, old_out2);
+            old_out1 = old_out3;
+        }
+
+        tmp0 = out0;
+        tmp1 = out1;
+        tmp2 = out2;
+
+        out3 -= a * tmp2;
+        out2 -= a * tmp1;
+        out1 -= a * tmp0;
+
+        out3 -= b * tmp1;
+        out2 -= b * tmp0;
+
+        out3 -= c * tmp0;
+
+
+        out[0] = out0;
+        out[1] = out1;
+        out[2] = out2;
+        out[3] = out3;
+
+        old_out0 = out0;
+        old_out1 = out1;
+        old_out2 = out2;
+        old_out3 = out3;
+
+        out += 4;
+        in  += 4;
+    }
+
+    out -= n;
+    in -= n;
+    for (; n < buffer_length; n++) {
         out[n] = in[n];
-        for(i=1; i<filter_length; i++)
+        for (i = 1; i <= filter_length; i++)
             out[n] -= filter_coeffs[i-1] * out[n-i];
     }
+#endif
+}
+
+void ff_celp_lp_zero_synthesis_filterf(float *out, const float *filter_coeffs,
+                                       const float *in, int buffer_length,
+                                       int filter_length)
+{
+    int i,n;
+
+    for (n = 0; n < buffer_length; n++) {
+        out[n] = in[n];
+        for (i = 1; i <= filter_length; i++)
+            out[n] += filter_coeffs[i-1] * in[n-i];
+    }
+}
+
+void ff_celp_filter_init(CELPFContext *c)
+{
+    c->celp_lp_synthesis_filterf        = ff_celp_lp_synthesis_filterf;
+    c->celp_lp_zero_synthesis_filterf   = ff_celp_lp_zero_synthesis_filterf;
+
+    if(HAVE_MIPSFPU)
+        ff_celp_filter_init_mips(c);
 }