opnassg: 3 -> 4 samples per frame to ease SSE data load

4 files changed, 20 insertions, 20 deletions

diff --git a/libopna/opnassg-sinc-c.c b/libopna/opnassg-sinc-c.c
index bf93039..5f9baee 100644
--- a/libopna/opnassg-sinc-c.c
+++ b/libopna/opnassg-sinc-c.c
@@ -6,7 +6,7 @@ void opna_ssg_sinc_calc_c(unsigned resampler_index, const int16_t *inbuf, int32_
     for (int j = 0; j < OPNA_SSG_SINCTABLELEN; j++) {
       unsigned sincindex = j;
       if (!(resampler_index&1)) sincindex += OPNA_SSG_SINCTABLELEN;
-      chsample += inbuf[(((resampler_index)>>1)+j)*3+c] * opna_ssg_sinctable[sincindex];
+      chsample += inbuf[(((resampler_index)>>1)+j)*4+c] * opna_ssg_sinctable[sincindex];
     }
     outbuf[c] = chsample;
   }
diff --git a/libopna/opnassg-sinc-neon.s b/libopna/opnassg-sinc-neon.s
index 19dc3c6..1396562 100644
--- a/libopna/opnassg-sinc-neon.s
+++ b/libopna/opnassg-sinc-neon.s
@@ -22,8 +22,7 @@ opna_ssg_sinc_calc_neon:
 
 @ add offset to ssg input buffer address
   bic r0, #1
-  add r0, r0, lsl #1
-  add r1, r0
+  add r1, r0, lsl #2
 
 @ initialize output register
   vmov.i64 q12, #0
@@ -39,12 +38,12 @@ opna_ssg_sinc_calc_neon:
   blo .end
 
 @ load SSG channel data
-  vld3.16 {d0-d2}, [r1]!
-  vld3.16 {d3-d5}, [r1]!
-  vld3.16 {d6-d8}, [r1]!
-  vld3.16 {d9-d11}, [r1]!
-  vld3.16 {d12-d14}, [r1]!
-  vld3.16 {d15-d17}, [r1]!
+  vld4.16 {d0-d3}, [r1]!
+  vld4.16 {d3-d6}, [r1]!
+  vld4.16 {d6-d9}, [r1]!
+  vld4.16 {d9-d12}, [r1]!
+  vld4.16 {d12-d15}, [r1]!
+  vld4.16 {d15-d18}, [r1]!
 
 @ load sinc data
   vld1.16 {d18-d21}, [r3]!
@@ -73,7 +72,7 @@ opna_ssg_sinc_calc_neon:
 
 .end:
 @ 8 samples left
-  vld3.16 {d0-d2}, [r1]!
+  vld4.16 {d0-d3}, [r1]!
   vld3.16 {d3-d5}, [r1]
   vld1.16 {d18-d19}, [r3]
 
diff --git a/libopna/opnassg.c b/libopna/opnassg.c
index 4a12f76..86d9c9a 100644
--- a/libopna/opnassg.c
+++ b/libopna/opnassg.c
@@ -208,6 +208,7 @@ int opna_ssg_channel_level(const struct opna_ssg *ssg, int ch) {
 #define COEFFSH 14
 
 // 3 samples per frame
+// output buf: 0 1 2 x 0 1 2 x ...
 void opna_ssg_generate_raw(struct opna_ssg *ssg, int16_t *buf, int samples) {
   for (int i = 0; i < samples; i++) {
     if (((++ssg->noise_counter) >> 1) >= opna_ssg_noise_period(ssg)) {
@@ -234,7 +235,7 @@ void opna_ssg_generate_raw(struct opna_ssg *ssg, int16_t *buf, int samples) {
 
     //int16_t out = 0;
     for (int ch = 0; ch < 3; ch++) {
-      buf[i*3+ch] = 0;
+      buf[i*4+ch] = 0;
       if (++ssg->ch[ch].tone_counter >= opna_ssg_tone_period(ssg, ch)) {
         ssg->ch[ch].tone_counter = 0;
         ssg->ch[ch].out = !ssg->ch[ch].out;
@@ -252,13 +253,13 @@ void opna_ssg_generate_raw(struct opna_ssg *ssg, int16_t *buf, int samples) {
       previntmp *= COEFF;
       ssg->prevout[ch] = previntmp - ssg->previn[ch] + ((((int64_t)COEFF)*ssg->prevout[ch]) >> COEFFSH);
       ssg->previn[ch] = previntmp;
-      buf[i*3+ch] = ssg->prevout[ch] >> COEFFSH;
-      //buf[i*3+ch] = voltable[level]/2;
+      buf[i*4+ch] = ssg->prevout[ch] >> COEFFSH;
+      //buf[i*4+ch] = voltable[level]/2;
 #else
       if (!opna_ssg_tone_silent(ssg, ch)) {
         int level = opna_ssg_channel_level(ssg, ch);
         //out += (opna_ssg_tone_out(ssg, ch) ? voltable[level] : -voltable[level]) / 2;
-        buf[i*3+ch] = (opna_ssg_tone_out(ssg, ch) ? voltable[level] : -voltable[level]) / 4;
+        buf[i*4+ch] = (opna_ssg_tone_out(ssg, ch) ? voltable[level] : -voltable[level]) / 4;
       }
 #endif
       
@@ -288,18 +289,18 @@ void opna_ssg_mix_55466(
   for (int i = 0; i < samples; i++) {
     {
       int ssg_samples = ((resampler->index + 9)>>1) - ((resampler->index)>>1);
-      int16_t ssgbuf[15];
+      int16_t ssgbuf[20];
       opna_ssg_generate_raw(ssg, ssgbuf, ssg_samples);
       for (int j = 0; j < ssg_samples; j++) {
-        resampler->buf[BUFINDEX(j)*3+0] = ssgbuf[j*3+0];
-        resampler->buf[BUFINDEX(j)*3+1] = ssgbuf[j*3+1];
-        resampler->buf[BUFINDEX(j)*3+2] = ssgbuf[j*3+2];
+        resampler->buf[BUFINDEX(j)*4+0] = ssgbuf[j*4+0];
+        resampler->buf[BUFINDEX(j)*4+1] = ssgbuf[j*4+1];
+        resampler->buf[BUFINDEX(j)*4+2] = ssgbuf[j*4+2];
       }
       resampler->index += 9;
     }
     int32_t sample = 0;
     resampler->index &= (1u<<(OPNA_SSG_SINCTABLEBIT+1))-1;
-    memcpy(resampler->buf + OPNA_SSG_SINCTABLELEN*3, resampler->buf, OPNA_SSG_SINCTABLELEN*3*sizeof(*resampler->buf));
+    memcpy(resampler->buf + OPNA_SSG_SINCTABLELEN*4, resampler->buf, OPNA_SSG_SINCTABLELEN*4*sizeof(*resampler->buf));
     int32_t outbuf[3];
     opna_ssg_sinc_calc_func(resampler->index, resampler->buf, outbuf);
     for (int ch = 0; ch < 3; ch++) {
diff --git a/libopna/opnassg.h b/libopna/opnassg.h
index 231db4d..223d542 100644
--- a/libopna/opnassg.h
+++ b/libopna/opnassg.h
@@ -33,7 +33,7 @@ struct opna_ssg {
 };
 
 struct opna_ssg_resampler {
-  int16_t buf[OPNA_SSG_SINCTABLELEN*3 * 2];
+  int16_t buf[OPNA_SSG_SINCTABLELEN*4 * 2];
   unsigned index;
 };