From d7497540e2bc9c4466dba93a7843c6c0e9b0a7b2 Mon Sep 17 00:00:00 2001
From: Zilog80 <rs1729@users.noreply.github.com>
Date: Thu, 22 Sep 2022 23:58:41 +0200
Subject: [PATCH] demod_mod: decimate/lowpass -Ofast speed-up

---
 demod/mod/demod_mod.c    | 18 +++++++++++++++++-
 demod/multi/demod_base.c | 18 +++++++++++++++++-
 imet/imet4iq.c           | 18 +++++++++++++++++-
 mk2a/mk2a1680mod.c       | 18 +++++++++++++++++-
 scan/dft_detect.c        | 18 +++++++++++++++++-
 5 files changed, 85 insertions(+), 5 deletions(-)

diff --git a/demod/mod/demod_mod.c b/demod/mod/demod_mod.c
index e62e41b..113a82d 100644
--- a/demod/mod/demod_mod.c
+++ b/demod/mod/demod_mod.c
@@ -625,7 +625,7 @@ static float complex lowpass0(float complex buffer[], ui32_t sample, ui32_t taps
     }
     return (float complex)w;
 }
-static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+static float complex lowpass1(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
     ui32_t s = sample % taps;
     double complex w = 0;
@@ -635,6 +635,22 @@ static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps,
     return (float complex)w;
 // symmetry: ws[n] == ws[taps-1-n]
 }
+static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+    float complex w = 0;     // -Ofast
+    int n;
+    int s = sample % taps; // lpIQ
+    int S1 = s+1;
+    int S1N = S1-taps;
+    int n0 = taps-1-s;
+    for (n = 0; n < n0; n++) {
+        w += buffer[S1+n]*ws[n];
+    }
+    for (n = n0; n < taps; n++) {
+        w += buffer[S1N+n]*ws[n];
+    }
+    return w;
+// symmetry: ws[n] == ws[taps-1-n]
+}
 
 static float re_lowpass0(float buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
diff --git a/demod/multi/demod_base.c b/demod/multi/demod_base.c
index 4103a94..b45fc98 100644
--- a/demod/multi/demod_base.c
+++ b/demod/multi/demod_base.c
@@ -699,7 +699,7 @@ static float complex lowpass0(float complex buffer[], ui32_t sample, ui32_t taps
     }
     return (float complex)w;
 }
-static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+static float complex lowpass1(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
     ui32_t s = sample % taps;
     double complex w = 0;
@@ -709,6 +709,22 @@ static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps,
     return (float complex)w;
 // symmetry: ws[n] == ws[taps-1-n]
 }
+static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+    float complex w = 0;     // -Ofast
+    int n;
+    int s = sample % taps; // lpIQ
+    int S1 = s+1;
+    int S1N = S1-taps;
+    int n0 = taps-1-s;
+    for (n = 0; n < n0; n++) {
+        w += buffer[S1+n]*ws[n];
+    }
+    for (n = n0; n < taps; n++) {
+        w += buffer[S1N+n]*ws[n];
+    }
+    return w;
+// symmetry: ws[n] == ws[taps-1-n]
+}
 
 static float re_lowpass0(float buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
diff --git a/imet/imet4iq.c b/imet/imet4iq.c
index 1afc74d..f678733 100644
--- a/imet/imet4iq.c
+++ b/imet/imet4iq.c
@@ -395,7 +395,7 @@ static int lowpass_init(float f, int taps, float **pws) {
 }
 
 
-static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+static float complex lowpass1(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
     ui32_t s = sample % taps;
     double complex w = 0;
@@ -405,6 +405,22 @@ static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps,
     return (float complex)w;
 // symmetry: ws[n] == ws[taps-1-n]
 }
+static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+    float complex w = 0;     // -Ofast
+    int n;
+    int s = sample % taps; // lpIQ
+    int S1 = s+1;
+    int S1N = S1-taps;
+    int n0 = taps-1-s;
+    for (n = 0; n < n0; n++) {
+        w += buffer[S1+n]*ws[n];
+    }
+    for (n = n0; n < taps; n++) {
+        w += buffer[S1N+n]*ws[n];
+    }
+    return w;
+// symmetry: ws[n] == ws[taps-1-n]
+}
 
 static float re_lowpass(float buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
diff --git a/mk2a/mk2a1680mod.c b/mk2a/mk2a1680mod.c
index cdf89cc..b020d25 100644
--- a/mk2a/mk2a1680mod.c
+++ b/mk2a/mk2a1680mod.c
@@ -733,7 +733,7 @@ static int lowpass_init(float f, int taps, float **pws) {
 }
 
 
-static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+static float complex lowpass1(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
     ui32_t s = sample % taps;
     double complex w = 0;
@@ -743,6 +743,22 @@ static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps,
     return (float complex)w;
 // symmetry: ws[n] == ws[taps-1-n]
 }
+static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+    float complex w = 0;     // -Ofast
+    int n;
+    int s = sample % taps; // lpIQ
+    int S1 = s+1;
+    int S1N = S1-taps;
+    int n0 = taps-1-s;
+    for (n = 0; n < n0; n++) {
+        w += buffer[S1+n]*ws[n];
+    }
+    for (n = n0; n < taps; n++) {
+        w += buffer[S1N+n]*ws[n];
+    }
+    return w;
+// symmetry: ws[n] == ws[taps-1-n]
+}
 
 static float re_lowpass(float buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
diff --git a/scan/dft_detect.c b/scan/dft_detect.c
index b01e33d..2cfd032 100644
--- a/scan/dft_detect.c
+++ b/scan/dft_detect.c
@@ -650,7 +650,7 @@ static float complex lowpass0(float complex buffer[], ui32_t sample, ui32_t taps
     }
     return (float complex)w;
 }
-static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+static float complex lowpass1(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
     ui32_t n;
     ui32_t s = sample % taps;
     double complex w = 0;
@@ -660,6 +660,22 @@ static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps,
     return (float complex)w;
 // symmetry: ws[n] == ws[taps-1-n]
 }
+static float complex lowpass(float complex buffer[], ui32_t sample, ui32_t taps, float *ws) {
+    float complex w = 0;     // -Ofast
+    int n;
+    int s = sample % taps; // lpIQ
+    int S1 = s+1;
+    int S1N = S1-taps;
+    int n0 = taps-1-s;
+    for (n = 0; n < n0; n++) {
+        w += buffer[S1+n]*ws[n];
+    }
+    for (n = n0; n < taps; n++) {
+        w += buffer[S1N+n]*ws[n];
+    }
+    return w;
+// symmetry: ws[n] == ws[taps-1-n]
+}
 
 
 static int f32buf_sample(FILE *fp, int inv) {