Reverts all changes past 2ab504b.

Individual optimizations will be cherry-picked and included in subsequent PRs.

Author: tmiw

src/codec2.c

@@ -2088,11 +2088,7 @@ void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[])
     for(i=0; i<n_samp; i++)
       c2->Sn[i+m_pitch-n_samp] = speech[i];
 
-#if defined(STM32F40_41xxx)
     dft_speech(&c2->c2const, c2->fft_fwd_cfg, Sw, c2->Sn, c2->w);
-#else
-    dft_speech(&c2->c2const, c2->fftr_fwd_cfg, Sw, c2->Sn, c2->w);
-#endif // STM32F40_41xxx
 
     /* Estimate pitch */
     nlp(c2->nlp, c2->Sn, n_samp, &pitch, Sw, c2->W, &c2->prev_f0_enc);

src/filter.c

@@ -20,10 +20,6 @@
 #include <math.h>
 #include <complex.h>
 
-#ifdef FDV_ARM_MATH
-#include "arm_math.h"
-#endif // FDV_ARM_MATH
-
 #include "filter.h"
 #include "filter_coef.h"
 #include "debug_alloc.h"
@@ -258,28 +254,6 @@ void quisk_cfTune(struct quisk_cfFilter * filter, float freq) {
 \*---------------------------------------------------------------------------*/
 
 void quisk_ccfFilter(complex float * inSamples, complex float * outSamples, int count, struct quisk_cfFilter * filter) {
-#if defined(FDV_ARM_MATH)
-    complex float ptSample[filter->nTaps + count - 1];
-    complex float *ptSamplePtr = &ptSample[0];
-    complex float ptCoeff[filter->nTaps];
-    memcpy(&ptSample, &filter->cSamples[1], (filter->nTaps - 1) * sizeof(complex float));
-    memcpy(&ptSample[filter->nTaps - 1], inSamples, count * sizeof(complex float));
-
-    // Reverse coefficients since we're going to go forwards over samples.
-    for (int i = 0, k = filter->nTaps - 1; i < filter->nTaps; i++, k--)
-    {
-        ptCoeff[i] = filter->cpxCoefs[k];
-    }
-
-    for (int i = 0; i < count; i++, ptSamplePtr++)
-    {
-        complex float accum = 0;
-        arm_cmplx_dot_prod_f32((float32_t*)ptSamplePtr, (float32_t*)&ptCoeff, filter->nTaps, (float32_t*)&accum, (float32_t*)&accum + 1);
-        outSamples[i] = accum;
-    }
-
-    memcpy((float32_t*)filter->cSamples, (float32_t*)&ptSample[count - 1 - filter->nTaps], filter->nTaps * sizeof(complex float));
-#else
     int i, k;
     complex float * ptSample;
     complex float * ptCoef;
@@ -303,6 +277,5 @@ void quisk_ccfFilter(complex float * inSamples, complex float * outSamples, int
         if (++filter->ptcSamp >= filter->cSamples + filter->nTaps)
             filter->ptcSamp = filter->cSamples;
     }
-#endif
 }
 

src/fmfsk.c

@@ -168,7 +168,7 @@ void fmfsk_mod(struct FMFSK *fmfsk, float fmfsk_out[],uint8_t bits_in[]){
  * float fsk_in[] - nin samples of modualted FMFSK from an FM radio
  */
 void fmfsk_demod(struct FMFSK *fmfsk, uint8_t rx_bits[],float fmfsk_in[]){
-    int i,j;
+    int i,j,k;
     int Ts          = fmfsk->Ts;
     int Fs          = fmfsk->Fs;
     int Rs          = fmfsk->Rs;
@@ -187,6 +187,9 @@ void fmfsk_demod(struct FMFSK *fmfsk, uint8_t rx_bits[],float fmfsk_in[]){
     int next_nin;
     float apeven,apodd;     /* Approx. prob of even or odd stream being correct */
     float currv,mdiff,lastv;
+    int neyesamp;
+    int neyeoffset;
+    float eye_max;
     uint8_t mbit;
     float var_signal = 0, var_noise = 0, lastFabsV;
 
@@ -340,11 +343,6 @@ void fmfsk_demod(struct FMFSK *fmfsk, uint8_t rx_bits[],float fmfsk_in[]){
     fmfsk->stats->snr_est = fmfsk->snr_mean;
 
 #ifndef __EMBEDDED__
-    int neyesamp;
-    int neyeoffset;
-    float eye_max;
-    int k;
-
     /* Collect an eye diagram */
     /* Take a sample for the eye diagrams */
     neyesamp = fmfsk->stats->neyesamp = Ts*4;

src/fsk.c

@@ -640,7 +640,8 @@ void fsk_demod_core(struct FSK *fsk, uint8_t rx_bits[], float rx_filt[], COMP fs
     float rx_timing,norm_rx_timing,old_norm_rx_timing,d_norm_rx_timing,appm;
 
     float fc_avg,fc_tx;
-    float meanebno,stdebno;
+    float meanebno,stdebno,eye_max;
+    int neyesamp,neyeoffset;
 
     #ifdef MODEMPROBE_ENABLE
     #define NMP_NAME 26
@@ -878,8 +879,6 @@ void fsk_demod_core(struct FSK *fsk, uint8_t rx_bits[], float rx_filt[], COMP fs
        trace.  So lets output a decimated version.  We use 2P
        as we want two symbols worth of samples in trace  */
 #ifndef __EMBEDDED__
-    float eye_max;
-    int neyesamp,neyeoffset;
     int neyesamp_dec = ceil(((float)P*2)/MODEM_STATS_EYE_IND_MAX);
     neyesamp = (P*2)/neyesamp_dec;
     assert(neyesamp <= MODEM_STATS_EYE_IND_MAX);
@@ -958,18 +957,18 @@ void fsk_demod_sd(struct FSK *fsk, float rx_filt[], COMP fsk_in[]){
 
 /* make sure stats have known values in case monitoring process reads stats before they are set */
 static void stats_init(struct FSK *fsk) {
+    /* Take a sample for the eye diagrams */
+    int i,j,m;
+    int P = fsk->P;
+    int M = fsk->mode;
+
     /* due to oversample rate P, we have too many samples for eye
        trace.  So lets output a decimated version */
 
     /* asserts below as we found some problems over-running eye matrix */
 
     /* TODO: refactor eye tracing code here and in fsk_demod */
 #ifndef __EMBEDDED__    
-    /* Take a sample for the eye diagrams */
-    int i,j,m;
-    int P = fsk->P;
-    int M = fsk->mode;
-
     int neyesamp_dec = ceil(((float)P*2)/MODEM_STATS_EYE_IND_MAX);
     int neyesamp = (P*2)/neyesamp_dec;
     assert(neyesamp <= MODEM_STATS_EYE_IND_MAX);

src/mbest.c

@@ -75,16 +75,18 @@ void mbest_destroy(struct MBEST *mbest) {
 \*---------------------------------------------------------------------------*/
 
 void mbest_insert(struct MBEST *mbest, int index[], float error) {
-    int                i, found;
+    int                i, j, found;
     struct MBEST_LIST *list    = mbest->list;
     int                entries = mbest->entries;
 
     found = 0;
     for(i=0; i<entries && !found; i++)
 	if (error < list[i].error) {
 	    found = 1;
-            memmove(&list[i+1], &list[i], sizeof(struct MBEST_LIST) * (entries - i - 1));
-            memcpy(&list[i].index[0], &index[0], sizeof(int) * MBEST_STAGES);
+	    for(j=entries-1; j>i; j--)
+		list[j] = list[j-1];
+	    for(j=0; j<MBEST_STAGES; j++)
+		list[i].index[j] = index[j];
 	    list[i].error = error;
 	}
 }
@@ -122,57 +124,17 @@ void mbest_search(
 )
 {
    float   e;
-   
-   for(int j = 0; j < m; j++) {
-        int i;
-
-        e = 0.0;
-        for(int i = 0; i < k; i++) {
-            float diff = (*cb++) - vec[i];
-            float diff2 = diff * diff;
-            float w2 = w[i] * w[i];
-            e += diff2 * w2;
-        }
-
-        index[0] = j;
-        if (e < mbest->list[mbest->entries - 1].error)
-            mbest_insert(mbest, index, e);
-   }
-}
+   int     i,j;
+   float   diff;
 
-/*---------------------------------------------------------------------------*\
-
-  mbest_search_equalweight
-
-  Searches vec[] to a codebbook of vectors, and maintains a list of the mbest
-  closest matches. Similar to mbest_search() but with w[] = 1.
-
-\*---------------------------------------------------------------------------*/
-
-void mbest_search_equalweight(
-		  const float  *cb,     /* VQ codebook to search         */
-		  float         vec[],  /* target vector                 */
-		  int           k,      /* dimension of vector           */
-		  int           m,      /* number on entries in codebook */
-		  struct MBEST *mbest,  /* list of closest matches       */
-		  int           index[] /* indexes that lead us here     */
-)
-{
-   float   e;
-   
-   for(int j = 0; j < m; j++) {
-        int i;
-
-        e = 0.0;
-        for(int i = 0; i < k; i++) {
-            float diff = (*cb++) - vec[i];
-            float diff2 = diff * diff;
-            e += diff2;
-        }
-
-        index[0] = j;
-        if (e < mbest->list[mbest->entries - 1].error)
-            mbest_insert(mbest, index, e);
+   for(j=0; j<m; j++) {
+	e = 0.0;
+	for(i=0; i<k; i++) {
+	    diff = cb[j*k+i]-vec[i];
+	    e += diff*w[i]*diff*w[i];
+	}
+	index[0] = j;
+	mbest_insert(mbest, index, e);
    }
 }
 

src/mbest.h

@@ -46,7 +46,6 @@ struct MBEST *mbest_create(int entries);
 void mbest_destroy(struct MBEST *mbest);
 void mbest_insert(struct MBEST *mbest, int index[], float error);
 void mbest_search(const float  *cb, float vec[], float w[], int k, int m, struct MBEST *mbest, int index[]);
-void mbest_search_equalweight(const float  *cb, float vec[], int k, int m, struct MBEST *mbest, int index[]);
 void mbest_search450(const float  *cb, float vec[], float w[], int k,int shorterK, int m, struct MBEST *mbest, int index[]);
 
 void mbest_print(char title[], struct MBEST *mbest);

src/modem_stats.c

@@ -33,14 +33,16 @@
 
 void modem_stats_open(struct MODEM_STATS *f)
 {
+    int i;
+
     /* zero out all the stats */
 
     memset(f, 0, sizeof(struct MODEM_STATS));
 
     /* init the FFT */
 
 #ifndef __EMBEDDED__
-    for(int i=0; i<2*MODEM_STATS_NSPEC; i++)
+    for(i=0; i<2*MODEM_STATS_NSPEC; i++)
 	f->fft_buf[i] = 0.0;
     f->fft_cfg = (void*)kiss_fft_alloc (2*MODEM_STATS_NSPEC, 0, NULL, NULL);
     assert(f->fft_cfg != NULL);

src/newamp1.c

@@ -166,14 +166,18 @@ float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest
   const float *codebook2 = newamp1vq_cb[1].cb;
   struct MBEST *mbest_stage1, *mbest_stage2;
   float target[ndim];
+  float w[ndim];
   int   index[MBEST_STAGES];
   float mse, tmp;
 
   /* codebook is compiled for a fixed K */
 
   assert(ndim == newamp1vq_cb[0].k);
 
-  /* note: using equal weights, could be argued mel freq axis gives freq dep weighting */
+  /* equal weights, could be argued mel freq axis gives freq dep weighting */
+
+  for(i=0; i<ndim; i++)
+      w[i] = 1.0;
 
   mbest_stage1 = mbest_create(mbest_entries);
   mbest_stage2 = mbest_create(mbest_entries);
@@ -182,15 +186,15 @@ float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest
 
   /* Stage 1 */
 
-  mbest_search_equalweight(codebook1, x, ndim, newamp1vq_cb[0].m, mbest_stage1, index);
+  mbest_search(codebook1, x, w, ndim, newamp1vq_cb[0].m, mbest_stage1, index);
 
   /* Stage 2 */
 
   for (j=0; j<mbest_entries; j++) {
       index[1] = n1 = mbest_stage1->list[j].index[0];
       for(i=0; i<ndim; i++)
 	  target[i] = x[i] - codebook1[ndim*n1+i];
-      mbest_search_equalweight(codebook2, target, ndim, newamp1vq_cb[1].m, mbest_stage2, index);
+      mbest_search(codebook2, target, w, ndim, newamp1vq_cb[1].m, mbest_stage2, index);
   }
 
   n1 = mbest_stage2->list[0].index[1];

src/nlp.c

@@ -126,11 +126,7 @@ typedef struct {
     float         sq[PMAX_M];	     /* squared speech samples       */
     float         mem_x,mem_y;       /* memory for notch filter      */
     float         mem_fir[NLP_NTAP]; /* decimation FIR filter memory */
-#if defined(STM32F40_41xxx)
     codec2_fft_cfg  fft_cfg;         /* kiss FFT config              */
-#else
-    codec2_fftr_cfg fft_cfg;
-#endif // defined(STM32F40_41xxx)
     float        *Sn16k;	     /* Fs=16kHz input speech vector */
     FILE         *f;
 } NLP;
@@ -194,11 +190,7 @@ void *nlp_create(C2CONST *c2const)
     for(i=0; i<NLP_NTAP; i++)
 	nlp->mem_fir[i] = 0.0;
 
-#if defined(STM32F40_41xxx)
     nlp->fft_cfg = codec2_fft_alloc (PE_FFT_SIZE, 0, NULL, NULL);
-#else
-    nlp->fft_cfg = codec2_fftr_alloc (PE_FFT_SIZE, 0, NULL, NULL);
-#endif // defined(STM32F40_41xxx)
     assert(nlp->fft_cfg != NULL);
 
     return (void*)nlp;
@@ -218,11 +210,7 @@ void nlp_destroy(void *nlp_state)
     assert(nlp_state != NULL);
     nlp = (NLP*)nlp_state;
 
-#if defined(STM32F40_41xxx)
     codec2_fft_free(nlp->fft_cfg);
-#else
-    codec2_fftr_free(nlp->fft_cfg);
-#endif // defined(STM32F40_41xxx)
     if (nlp->Fs == 16000) {
         free(nlp->Sn16k);
     }
@@ -363,35 +351,15 @@ float nlp(
     #ifdef DUMP
     dump_dec(Fw);
     #endif
-    
-#if defined(STM32F40_41xxx)
+
     // FIXME: check if this can be converted to a real fft
     // since all imag inputs are 0
     codec2_fft_inplace(nlp->fft_cfg, Fw);
     PROFILE_SAMPLE_AND_LOG(fft, window, "      fft");
 
     for(i=0; i<PE_FFT_SIZE; i++)
-	    Fw[i].real = Fw[i].real*Fw[i].real + Fw[i].imag*Fw[i].imag;
-#else
-    float   Fw_input[PE_FFT_SIZE];
-    for (i = 0; i < PE_FFT_SIZE; i++)
-    {
-        Fw_input[i] = Fw[i].real;
-    }
-    codec2_fftr(nlp->fft_cfg, &Fw_input[0], &Fw[0]);
-    PROFILE_SAMPLE_AND_LOG(fft, window, "      fft");
-    
-    for(i=0; i<PE_FFT_SIZE/2; i++)
-    {
-	    Fw[i].real = Fw[i].real*Fw[i].real + Fw[i].imag*Fw[i].imag;
-        
-        // Set Fw[PE_FFT_SIZE - i - 1] to the same as Fw[i] due to the real FFT's symmetry
-        // (and to avoid having to change anything else further down).
-        Fw[PE_FFT_SIZE - i - 1].real = Fw[i].real;
-    }
-    
-#endif // defined(STM32F40_41xxx)
-    
+	Fw[i].real = Fw[i].real*Fw[i].real + Fw[i].imag*Fw[i].imag;
+
     PROFILE_SAMPLE_AND_LOG(magsq, fft, "      mag sq");
     #ifdef DUMP
     dump_sq(m, nlp->sq);