Move pitch vector interpolation code to acelp_filters
authorVladimir Voroshilov <voroshil@gmail.com>
Sat, 24 May 2008 17:18:42 +0000 (17:18 +0000)
committerVladimir Voroshilov <voroshil@gmail.com>
Sat, 24 May 2008 17:18:42 +0000 (17:18 +0000)
and convert it to a generic interpolation routine.

Originally committed as revision 13284 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/acelp_filters.c
libavcodec/acelp_filters.h

index 7e33a67..2aedbdd 100644 (file)
 #define FRAC_BITS 13
 #include "mathops.h"
 
 #define FRAC_BITS 13
 #include "mathops.h"
 
+const int16_t ff_acelp_interp_filter[61] =
+{ /* (0.15) */
+  29443, 28346, 25207, 20449, 14701,  8693,
+   3143, -1352, -4402, -5865, -5850, -4673,
+  -2783,  -672,  1211,  2536,  3130,  2991,
+   2259,  1170,     0, -1001, -1652, -1868,
+  -1666, -1147,  -464,   218,   756,  1060,
+   1099,   904,   550,   135,  -245,  -514,
+   -634,  -602,  -451,  -231,     0,   191,
+    308,   340,   296,   198,    78,   -36,
+   -120,  -163,  -165,  -132,   -79,   -19,
+     34,    73,    91,    89,    70,    38,
+      0,
+};
+
+void ff_acelp_interpolate(
+        int16_t* out,
+        const int16_t* in,
+        const int16_t* filter_coeffs,
+        int precision,
+        int pitch_delay_frac,
+        int filter_length,
+        int length)
+{
+    int n, i;
+
+    assert(pitch_delay_frac >= 0 && pitch_delay_frac < precision);
+
+    for(n=0; n<length; n++)
+    {
+        int idx = 0;
+        int v = 0x4000;
+
+        for(i=0; i<filter_length;)
+        {
+
+            /* The reference G.729 and AMR fixed point code performs clipping after
+               each of the two following accumulations.
+               Since clipping affects only the synthetic OVERFLOW test without
+               causing an int type overflow, it was moved outside the loop. */
+
+            /*  R(x):=ac_v[-k+x]
+                v += R(n-i)*ff_acelp_interp_filter(t+6i)
+                v += R(n+i+1)*ff_acelp_interp_filter(6-t+6i) */
+
+            v += in[n + i] * filter_coeffs[idx + pitch_delay_frac];
+            idx += precision;
+            i++;
+            v += in[n - i] * filter_coeffs[idx - pitch_delay_frac];
+        }
+        out[n] = av_clip_int16(v >> 15);
+    }
+}
+
 void ff_acelp_convolve_circ(
         int16_t* fc_out,
         const int16_t* fc_in,
 void ff_acelp_convolve_circ(
         int16_t* fc_out,
         const int16_t* fc_in,
index 9e497ca..167faf7 100644 (file)
 
 #include <stdint.h>
 
 
 #include <stdint.h>
 
+/**
+ * low-pass FIR (Finite Impulse Response) filter coefficients
+ *
+ *   A similar filter is named b30 in G.729.
+ *
+ *   G.729 specification says:
+ *     b30 is based on Hamming windowed sinc functions, truncated at +/-29 and
+ *     padded with zeros at +/-30 b30[30]=0.
+ *     The filter has a cut-off frequency (-3 dB) at 3600 Hz in the oversampled
+ *     domain.
+ *
+ *   After some analysis, I found this approximation:
+ *
+ *                                    PI * x
+ *   Hamm(x,N) = 0.53836-0.46164*cos(--------)
+ *                                      N-1
+ *                                      ---
+ *                                       2
+ *
+ *                                                             PI * x
+ *   Hamm'(x,k) = Hamm(x - k, 2*k+1) =  0.53836 + 0.46164*cos(--------)
+ *                                                                k
+ *
+ *             sin(PI * x)
+ *   Sinc(x) = ----------- (normalized sinc function)
+ *               PI * x
+ *
+ *   h(t,B) = 2 * B * Sinc(2 * B * t) (impulse response of sinc low-pass filter)
+ *
+ *   b(k,B, n) = Hamm'(n, k) * h(n, B)
+ *
+ *
+ *       3600
+ *   B = ----
+ *       8000
+ *
+ *   3600 - cut-off frequency
+ *   8000 - sampling rate
+ *   k    - filter order
+ *
+ *   ff_acelp_interp_filter[6*i+j] = b(10, 3600/8000, i+j/6)
+ *
+ * The filter assumes the following order of fractions (X - integer delay):
+ *
+ * 1/3 precision: X     1/3      2/3      X     1/3      2/3      X
+ * 1/6 precision: X 1/6 2/6 3/6  4/6  5/6 X 1/6 2/6 3/6  4/6  5/6 X
+ *
+ * The filter can be used for 1/3 precision, too, by
+ * passing 2*pitch_delay_frac as third parameter to the interpolation routine.
+ *
+ */
+extern const int16_t ff_acelp_interp_filter[61];
+
+/**
+ * \brief Generic interpolation routine
+ * \param out [out] buffer for interpolated data
+ * \param in input data
+ * \param filter_coeffs interpolation filter coefficients (0.15)
+ * \param precision filter is able to interpolate with 1/precision precision of pitch delay
+ * \param pitch_delay_frac pitch delay, fractional part [0..precision-1]
+ * \param filter_length filter length
+ * \param length length of speech data to process
+ *
+ * filter_coeffs contains coefficients of the positive half of the symmetric
+ * interpolation filter. filter_coeffs[0] should the central (unpaired) coefficient.
+ * See ff_acelp_interp_filter fot example.
+ *
+ */
+void ff_acelp_interpolate(
+        int16_t* out,
+        const int16_t* in,
+        const int16_t* filter_coeffs,
+        int precision,
+        int pitch_delay_frac,
+        int filter_length,
+        int length);
+
 /**
  * \brief Circularly convolve fixed vector with a phase dispersion impulse
  *        response filter (D.6.2 of G.729 and 6.1.5 of AMR).
 /**
  * \brief Circularly convolve fixed vector with a phase dispersion impulse
  *        response filter (D.6.2 of G.729 and 6.1.5 of AMR).