minor comments
[platform/upstream/flac.git] / src / libFLAC / lpc.c
index 06d6808..6ec5a91 100644 (file)
@@ -1,5 +1,5 @@
 /* libFLAC - Free Lossless Audio Codec library
- * Copyright (C) 2000,2001  Josh Coalson
+ * Copyright (C) 2000,2001,2002,2003  Josh Coalson
  *
  * This library is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Library General Public
  */
 
 #include <math.h>
-#include <stdio.h>
 #include "FLAC/assert.h"
 #include "FLAC/format.h"
+#include "private/bitmath.h"
 #include "private/lpc.h"
+#if defined DEBUG || defined FLAC__OVERFLOW_DETECT || defined FLAC__OVERFLOW_DETECT_VERBOSE
+#include <stdio.h>
+#endif
 
 #ifndef M_LN2
 /* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
 #define M_LN2 0.69314718055994530942
 #endif
 
-#define LOCAL_FABS(x) ((x)<0.0? -(x):(x))
-
-void FLAC__lpc_compute_autocorrelation(const real data[], unsigned data_len, unsigned lag, real autoc[])
+void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[])
 {
        /* a readable, but slower, version */
 #if 0
-       real d;
+       FLAC__real d;
        unsigned i;
 
        FLAC__ASSERT(lag > 0);
@@ -51,7 +52,7 @@ void FLAC__lpc_compute_autocorrelation(const real data[], unsigned data_len, uns
         * this version tends to run faster because of better data locality
         * ('data_len' is usually much larger than 'lag')
         */
-       real d;
+       FLAC__real d;
        unsigned sample, coeff;
        const unsigned limit = data_len - lag;
 
@@ -72,10 +73,10 @@ void FLAC__lpc_compute_autocorrelation(const real data[], unsigned data_len, uns
        }
 }
 
-void FLAC__lpc_compute_lp_coefficients(const real autoc[], unsigned max_order, real lp_coeff[][FLAC__MAX_LPC_ORDER], real error[])
+void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned max_order, FLAC__real lp_coeff[][FLAC__MAX_LPC_ORDER], FLAC__real error[])
 {
        unsigned i, j;
-       real r, err, ref[FLAC__MAX_LPC_ORDER], lpc[FLAC__MAX_LPC_ORDER];
+       double r, err, ref[FLAC__MAX_LPC_ORDER], lpc[FLAC__MAX_LPC_ORDER];
 
        FLAC__ASSERT(0 < max_order);
        FLAC__ASSERT(max_order <= FLAC__MAX_LPC_ORDER);
@@ -93,7 +94,7 @@ void FLAC__lpc_compute_lp_coefficients(const real autoc[], unsigned max_order, r
                /* Update LPC coefficients and total error. */
                lpc[i]=r;
                for(j = 0; j < (i>>1); j++) {
-                       real tmp = lpc[j];
+                       double tmp = lpc[j];
                        lpc[j] += r * lpc[i-1-j];
                        lpc[i-1-j] += r * tmp;
                }
@@ -104,66 +105,102 @@ void FLAC__lpc_compute_lp_coefficients(const real autoc[], unsigned max_order, r
 
                /* save this order */
                for(j = 0; j <= i; j++)
-                       lp_coeff[i][j] = -lpc[j]; /* negate FIR filter coeff to get predictor coeff */
-               error[i] = err;
+                       lp_coeff[i][j] = (FLAC__real)(-lpc[j]); /* negate FIR filter coeff to get predictor coeff */
+               error[i] = (FLAC__real)err;
        }
 }
 
-int FLAC__lpc_quantize_coefficients(const real lp_coeff[], unsigned order, unsigned precision, unsigned bits_per_sample, int32 qlp_coeff[], int *shift)
+int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], unsigned order, unsigned precision, FLAC__int32 qlp_coeff[], int *shift)
 {
        unsigned i;
-       real d, cmax = -1e32;
+       double d, cmax = -1e32;
+       FLAC__int32 qmax, qmin;
+       const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
+       const int min_shiftlimit = -max_shiftlimit - 1;
 
-       FLAC__ASSERT(bits_per_sample > 0);
-       FLAC__ASSERT(bits_per_sample <= sizeof(int32)*8);
        FLAC__ASSERT(precision > 0);
        FLAC__ASSERT(precision >= FLAC__MIN_QLP_COEFF_PRECISION);
-       FLAC__ASSERT(precision + bits_per_sample < sizeof(int32)*8);
-#ifdef NDEBUG
-       (void)bits_per_sample; /* silence compiler warning about unused parameter */
-#endif
 
        /* drop one bit for the sign; from here on out we consider only |lp_coeff[i]| */
        precision--;
+       qmax = 1 << precision;
+       qmin = -qmax;
+       qmax--;
 
        for(i = 0; i < order; i++) {
                if(lp_coeff[i] == 0.0)
                        continue;
-               d = LOCAL_FABS(lp_coeff[i]);
+               d = fabs(lp_coeff[i]);
                if(d > cmax)
                        cmax = d;
        }
-       if(cmax < 0.0) {
+redo_it:
+       if(cmax <= 0.0) {
                /* => coefficients are all 0, which means our constant-detect didn't work */
                return 2;
        }
        else {
-               const int maxshift = (int)precision - (int)floor(log(cmax) / M_LN2) - 1;
-               const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
-               const int min_shiftlimit = -max_shiftlimit - 1;
+               int log2cmax;
 
-               *shift = maxshift;
+               (void)frexp(cmax, &log2cmax);
+               log2cmax--;
+               *shift = (int)precision - log2cmax - 1;
 
                if(*shift < min_shiftlimit || *shift > max_shiftlimit) {
-                       return 1;
+#if 0
+                       /*@@@ this does not seem to help at all, but was not extensively tested either: */
+                       if(*shift > max_shiftlimit)
+                               *shift = max_shiftlimit;
+                       else
+#endif
+                               return 1;
                }
        }
 
-       if(*shift != 0) { /* just to avoid wasting time... */
-               for(i = 0; i < order; i++)
-                       qlp_coeff[i] = (int32)floor(lp_coeff[i] * (real)(1 << *shift));
+       if(*shift >= 0) {
+               for(i = 0; i < order; i++) {
+                       qlp_coeff[i] = (FLAC__int32)floor((double)lp_coeff[i] * (double)(1 << *shift));
+
+                       /* double-check the result */
+                       if(qlp_coeff[i] > qmax || qlp_coeff[i] < qmin) {
+#ifdef FLAC__OVERFLOW_DETECT
+                               fprintf(stderr,"FLAC__lpc_quantize_coefficients: compensating for overflow, qlp_coeff[%u]=%d, lp_coeff[%u]=%f, cmax=%f, precision=%u, shift=%d, q=%f, f(q)=%f\n", i, qlp_coeff[i], i, lp_coeff[i], cmax, precision, *shift, (double)lp_coeff[i] * (double)(1 << *shift), floor((double)lp_coeff[i] * (double)(1 << *shift)));
+#endif
+                               cmax *= 2.0;
+                               goto redo_it;
+                       }
+               }
+       }
+       else { /* (*shift < 0) */
+               const int nshift = -(*shift);
+#ifdef DEBUG
+               fprintf(stderr,"FLAC__lpc_quantize_coefficients: negative shift = %d\n", *shift);
+#endif
+               for(i = 0; i < order; i++) {
+                       qlp_coeff[i] = (FLAC__int32)floor((double)lp_coeff[i] / (double)(1 << nshift));
+
+                       /* double-check the result */
+                       if(qlp_coeff[i] > qmax || qlp_coeff[i] < qmin) {
+#ifdef FLAC__OVERFLOW_DETECT
+                               fprintf(stderr,"FLAC__lpc_quantize_coefficients: compensating for overflow, qlp_coeff[%u]=%d, lp_coeff[%u]=%f, cmax=%f, precision=%u, shift=%d, q=%f, f(q)=%f\n", i, qlp_coeff[i], i, lp_coeff[i], cmax, precision, *shift, (double)lp_coeff[i] / (double)(1 << nshift), floor((double)lp_coeff[i] / (double)(1 << nshift)));
+#endif
+                               cmax *= 2.0;
+                               goto redo_it;
+                       }
+               }
        }
+
        return 0;
 }
 
-void FLAC__lpc_compute_residual_from_qlp_coefficients(const int32 data[], unsigned data_len, const int32 qlp_coeff[], unsigned order, int lp_quantization, int32 residual[])
+void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 data[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
 {
 #ifdef FLAC__OVERFLOW_DETECT
-       int64 sumo;
+       FLAC__int64 sumo;
 #endif
        unsigned i, j;
-       int32 sum;
-       const int32 *history;
+       FLAC__int32 sum;
+       const FLAC__int32 *history;
 
 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
        fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
@@ -182,8 +219,13 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients(const int32 data[], unsign
                for(j = 0; j < order; j++) {
                        sum += qlp_coeff[j] * (*(--history));
 #ifdef FLAC__OVERFLOW_DETECT
-                       sumo += (int64)qlp_coeff[j] * (int64)(*history);
-                       if(sumo > 2147483647ll || sumo < -2147483648ll) {
+                       sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
+#if defined _MSC_VER
+                       if(sumo > 2147483647I64 || sumo < -2147483648I64)
+#else
+                       if(sumo > 2147483647ll || sumo < -2147483648ll)
+#endif
+                       {
                                fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
                        }
 #endif
@@ -201,14 +243,47 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients(const int32 data[], unsign
        */
 }
 
-void FLAC__lpc_restore_signal(const int32 residual[], unsigned data_len, const int32 qlp_coeff[], unsigned order, int lp_quantization, int32 data[])
+void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 data[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
 {
+       unsigned i, j;
+       FLAC__int64 sum;
+       const FLAC__int32 *history;
+
+#ifdef FLAC__OVERFLOW_DETECT_VERBOSE
+       fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
+       for(i=0;i<order;i++)
+               fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
+       fprintf(stderr,"\n");
+#endif
+       FLAC__ASSERT(order > 0);
+
+       for(i = 0; i < data_len; i++) {
+               sum = 0;
+               history = data;
+               for(j = 0; j < order; j++)
+                       sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
 #ifdef FLAC__OVERFLOW_DETECT
-       int64 sumo;
+               if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
+                       fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, sum=%lld\n", i, sum >> lp_quantization);
+                       break;
+               }
+               if(FLAC__bitmath_silog2_wide((FLAC__int64)(*data) - (sum >> lp_quantization)) > 32) {
+                       fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, data=%d, sum=%lld, residual=%lld\n", i, *data, sum >> lp_quantization, (FLAC__int64)(*data) - (sum >> lp_quantization));
+                       break;
+               }
+#endif
+               *(residual++) = *(data++) - (FLAC__int32)(sum >> lp_quantization);
+       }
+}
+
+void FLAC__lpc_restore_signal(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
+{
+#ifdef FLAC__OVERFLOW_DETECT
+       FLAC__int64 sumo;
 #endif
        unsigned i, j;
-       int32 sum;
-       const int32 *history;
+       FLAC__int32 sum;
+       const FLAC__int32 *history;
 
 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
        fprintf(stderr,"FLAC__lpc_restore_signal: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
@@ -227,8 +302,13 @@ void FLAC__lpc_restore_signal(const int32 residual[], unsigned data_len, const i
                for(j = 0; j < order; j++) {
                        sum += qlp_coeff[j] * (*(--history));
 #ifdef FLAC__OVERFLOW_DETECT
-                       sumo += (int64)qlp_coeff[j] * (int64)(*history);
-                       if(sumo > 2147483647ll || sumo < -2147483648ll) {
+                       sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
+#if defined _MSC_VER
+                       if(sumo > 2147483647I64 || sumo < -2147483648I64)
+#else
+                       if(sumo > 2147483647ll || sumo < -2147483648ll)
+#endif
+                       {
                                fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
                        }
 #endif
@@ -246,61 +326,83 @@ void FLAC__lpc_restore_signal(const int32 residual[], unsigned data_len, const i
        */
 }
 
-real FLAC__lpc_compute_expected_bits_per_residual_sample(real lpc_error, unsigned total_samples)
+void FLAC__lpc_restore_signal_wide(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
 {
-       real error_scale;
+       unsigned i, j;
+       FLAC__int64 sum;
+       const FLAC__int32 *history;
+
+#ifdef FLAC__OVERFLOW_DETECT_VERBOSE
+       fprintf(stderr,"FLAC__lpc_restore_signal_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
+       for(i=0;i<order;i++)
+               fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
+       fprintf(stderr,"\n");
+#endif
+       FLAC__ASSERT(order > 0);
+
+       for(i = 0; i < data_len; i++) {
+               sum = 0;
+               history = data;
+               for(j = 0; j < order; j++)
+                       sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
+#ifdef FLAC__OVERFLOW_DETECT
+               if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
+                       fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, sum=%lld\n", i, sum >> lp_quantization);
+                       break;
+               }
+               if(FLAC__bitmath_silog2_wide((FLAC__int64)(*residual) + (sum >> lp_quantization)) > 32) {
+                       fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, residual=%d, sum=%lld, data=%lld\n", i, *residual, sum >> lp_quantization, (FLAC__int64)(*residual) + (sum >> lp_quantization));
+                       break;
+               }
+#endif
+               *(data++) = *(residual++) + (FLAC__int32)(sum >> lp_quantization);
+       }
+}
+
+FLAC__real FLAC__lpc_compute_expected_bits_per_residual_sample(FLAC__real lpc_error, unsigned total_samples)
+{
+       double error_scale;
 
        FLAC__ASSERT(total_samples > 0);
 
-       error_scale = 0.5 * M_LN2 * M_LN2 / (real)total_samples;
+       error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__real)total_samples;
 
-       if(lpc_error > 0.0) {
-               real bps = 0.5 * log(error_scale * lpc_error) / M_LN2;
-               if(bps >= 0.0)
-                       return bps;
-               else
-                       return 0.0;
-       }
-       else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate float resolution */
-               return 1e32;
-       }
-       else {
-               return 0.0;
-       }
+       return FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error, error_scale);
 }
 
-real FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(real lpc_error, real error_scale)
+FLAC__real FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(FLAC__real lpc_error, double error_scale)
 {
        if(lpc_error > 0.0) {
-               real bps = 0.5 * log(error_scale * lpc_error) / M_LN2;
+               FLAC__real bps = (FLAC__real)((double)0.5 * log(error_scale * lpc_error) / M_LN2);
                if(bps >= 0.0)
                        return bps;
                else
                        return 0.0;
        }
        else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate float resolution */
-               return 1e32;
+               return (FLAC__real)1e32;
        }
        else {
                return 0.0;
        }
 }
 
-unsigned FLAC__lpc_compute_best_order(const real lpc_error[], unsigned max_order, unsigned total_samples, unsigned bits_per_signal_sample)
+unsigned FLAC__lpc_compute_best_order(const FLAC__real lpc_error[], unsigned max_order, unsigned total_samples, unsigned bits_per_signal_sample)
 {
        unsigned order, best_order;
-       real best_bits, tmp_bits, error_scale;
+       FLAC__real best_bits, tmp_bits;
+       double error_scale;
 
        FLAC__ASSERT(max_order > 0);
        FLAC__ASSERT(total_samples > 0);
 
-       error_scale = 0.5 * M_LN2 * M_LN2 / (real)total_samples;
+       error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__real)total_samples;
 
        best_order = 0;
-       best_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[0], error_scale) * (real)total_samples;
+       best_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[0], error_scale) * (FLAC__real)total_samples;
 
        for(order = 1; order < max_order; order++) {
-               tmp_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[order], error_scale) * (real)(total_samples - order) + (real)(order * bits_per_signal_sample);
+               tmp_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[order], error_scale) * (FLAC__real)(total_samples - order) + (FLAC__real)(order * bits_per_signal_sample);
                if(tmp_bits < best_bits) {
                        best_order = order;
                        best_bits = tmp_bits;