(external) bl00mbox: more inlining, performance++

components/bl00mbox/plugins/lowpass.c

 #include "lowpass.h"
 
-//#define LOWPASS_DEBUG_PRINT
-
-#ifdef LOWPASS_DEBUG_PRINT
-#include <stdio.h>
-#endif
-
 radspa_t * lowpass_create(uint32_t init_var);
 radspa_descriptor_t lowpass_desc = {
     .name = "lowpass",
@@ -23,63 +17,37 @@ radspa_descriptor_t lowpass_desc = {
 #define LOWPASS_GAIN 4
 
 #define LOWPASS_INTERNAL_SHIFT 14
-#define LOWPASS_OUT_COEFF_SHIFT 11
-
-static void coeff_shift(uint64_t coeff, int32_t *shift, int32_t * shifted_coeff) {
-    int32_t ret = 22;
-    uint64_t ret_coeff = coeff >> (32-ret);
-    while(1){
-        if(ret > 31){
-            ret = 31;
-            ret_coeff = coeff >> (32-ret);
-            break;
-        }
-        if(ret < (LOWPASS_INTERNAL_SHIFT)){
-            ret = (LOWPASS_INTERNAL_SHIFT);
-            ret_coeff = coeff >> (32-ret);
-            break;
-        }
-        if(ret_coeff > (1UL<<14)){
-            ret -= 3;
-            ret_coeff = coeff >> (32-ret);
-        } else if (ret_coeff < (1UL<<11)){
-            ret += 3;
-            ret_coeff = coeff >> (32-ret);
-        } else {
-            break;
-        }
-    }
-    
-    * shift = ret;
-    * shifted_coeff = (int32_t) ret_coeff;
-}
+#define LOWPASS_OUT_COEFF_SHIFT 30
+#define LOWPASS_SUM_SHIFT 25
 
 static void set_lowpass_coeffs(lowpass_data_t * data, int32_t freq, int32_t mq){
     if(freq < 15.) freq = 15.;
     if(freq > 15000.) freq = 15000.;
-    if(mq < 130) mq = 130;
+    if(mq < 140) mq = 140;
     int32_t K = 15287; // 2*48000/6.28
     int32_t omega = freq;
+    int32_t mq_rec = (1UL<<31)/(mq+1);
     int32_t A[3];
     A[0] = K * K;
-    A[1] = K * 1000;
-    A[1] = (int64_t) A[1] * omega / (mq+1);
+    A[1] = K * omega * 8;
+    A[1] = ((int64_t) A[1] * mq_rec) >> 32;
+    A[1] *= 250;
     A[2] = omega*omega;
-    int32_t sum_a = A[0] + A[1] + A[2];
-    int64_t round = A[2];
+    // slow, figure out smth smarther:
+    int32_t sum_a_rec = (1ULL<<(32 + (LOWPASS_SUM_SHIFT))) / (A[0] + A[1] + A[2]);
 
-    round = (round << 32) / sum_a;
-    int32_t shift;
-    int32_t shifted_coeff;
-    coeff_shift(round, &shift, &shifted_coeff);
-    data->in_coeff_shift = shift - (LOWPASS_INTERNAL_SHIFT);
-    data->in_coeff = shifted_coeff;
+    int32_t tmp;
 
-    round = 2.*(A[2]-A[0]);
-    data->out_coeff[0] = round * (1LL<<(LOWPASS_OUT_COEFF_SHIFT)) / sum_a;
+    tmp = (int64_t) A[2] * sum_a_rec >> 32;
+    data->in_coeff = tmp << (32-LOWPASS_SUM_SHIFT);
 
-    round = (A[0]-A[1]+A[2]);
-    data->out_coeff[1] = round * (1LL<<(LOWPASS_OUT_COEFF_SHIFT)) / sum_a;
+    tmp = 2.*(A[2]-A[0]);
+    tmp = ((int64_t) tmp * sum_a_rec) >> 32;
+    data->out_coeff[0] = tmp<<((LOWPASS_OUT_COEFF_SHIFT) - (LOWPASS_SUM_SHIFT));
+
+    tmp = (A[0]-A[1]+A[2]);
+    tmp = ((int64_t) tmp * sum_a_rec) >> 32;
+    data->out_coeff[1] = tmp<<((LOWPASS_OUT_COEFF_SHIFT) - (LOWPASS_SUM_SHIFT));
 }
 
 
@@ -100,11 +68,6 @@ void lowpass_run(radspa_t * lowpass, uint16_t num_samples, uint32_t render_pass_
 
         if((freq != data->prev_freq) | (q != data->prev_q)){
             set_lowpass_coeffs(data, freq, q);
-#ifdef LOWPASS_DEBUG_PRINT
-            printf("\nfreq: %ld, q: %d\n", freq, q);
-            printf("in_coeff: %ld >> %ld, ", data->in_coeff, data->in_coeff_shift);
-            printf("out_coeffs: %ld, %ld\n", data->out_coeff[0], data->out_coeff[1]);
-#endif
             data->prev_freq = freq;
             data->prev_q = q;
         }
@@ -121,11 +84,13 @@ void lowpass_run(radspa_t * lowpass, uint16_t num_samples, uint32_t render_pass_
             int8_t pos = data->pos - i - 1;
             if(pos < 0) pos += 3;
             in_acc += (2-i)*data->in_history[pos];
-
-            ret -= (((int64_t) data->out_history[pos]) * data->out_coeff[i]) >> (LOWPASS_OUT_COEFF_SHIFT);
+            ret -= (((int64_t) data->out_history[pos]) * data->out_coeff[i]) >> 32;
         }
 
-        ret += (in_acc * data->in_coeff) >> data->in_coeff_shift;
+        ret = ret << (32 - LOWPASS_OUT_COEFF_SHIFT);
+        in_acc = in_acc << (LOWPASS_INTERNAL_SHIFT);
+        in_acc = ((int64_t) in_acc * data->in_coeff) >> 32;
+        ret += in_acc;
 
         data->out_history[data->pos] = ret;