added softvolume for all supported output formats; moved softvolume from output plugins to streamer

4 files changed, 53 insertions, 49 deletions

diff --git a/deadbeef.h b/deadbeef.h
index bd1655c7..02401f5e 100644
--- a/deadbeef.h
+++ b/deadbeef.h
@@ -749,6 +749,9 @@ typedef struct DB_output_s {
 
     // parameters of current output
     ddb_waveformat_t fmt;
+
+    // set to 1 if volume control is done internally by plugin
+    int has_volume;
 } DB_output_t;
 
 // dsp plugin
diff --git a/plugins/alsa/alsa.c b/plugins/alsa/alsa.c
index 755981fa..68981512 100644
--- a/plugins/alsa/alsa.c
+++ b/plugins/alsa/alsa.c
@@ -615,48 +615,7 @@ palsa_thread (void *context) {
 
 static int
 palsa_callback (char *stream, int len) {
-    int bytesread = deadbeef->streamer_read (stream, len);
-
-// FIXME: move volume control to streamer_read for copy optimization
-#if 0
-    int16_t vol[4];
-    vol[0] = volume_get_amp () * 255; // that will be extra 8 bits
-    // pack 4 times
-    vol[1] = vol[2] = vol[3] = vol[0];
-
-    // apply volume with mmx
-    __asm__ volatile(
-            "  mov %0, %%ecx\n\t"
-            "  shr $4, %%ecx\n\t"
-            "  mov %1, %%eax\n\t"
-            "  movq %2, %mm1\n\t"
-            "1:\n\t"
-            "  movq [%%eax], %mm0\n\t"
-            "  movq %mm0, %mm2\n\t"
-            "  movq %mm0, %mm3\n\t"
-            "  pmullw %mm1, %mm2\n\t"
-            "  pmulhw %mm1, %mm3\n\t"
-            "  psrlw $8, %mm2\n\t" // discard lowest 8 bits
-            "  psllw $8, %mm3\n\t" // shift left 8 lsbs of hiwords
-            "  por %mm3, %mm2\n\t" // OR them together
-            "  movq %mm3, [%%eax]\n\t" // load back to memory
-            "  add $8, %%eax\n\t"
-            "  dec %%ecx\n\t"
-            "  jnz 1b\n\t"
-            :
-            : "r"(len), "r"(stream), "r"(vol)
-            : "%ecx", "%eax"
-       );
-
-#else
-    if (plugin.fmt.bps == 16) {
-        int16_t ivolume = deadbeef->volume_get_amp () * 1000;
-        for (int i = 0; i < bytesread/2; i++) {
-            ((int16_t*)stream)[i] = (int16_t)(((int32_t)(((int16_t*)stream)[i])) * ivolume / 1000);
-        }
-    }
-#endif
-    return bytesread;
+    return deadbeef->streamer_read (stream, len);
 }
 
 static int
diff --git a/plugins/oss/oss.c b/plugins/oss/oss.c
index 0f837c43..b78a2290 100644
--- a/plugins/oss/oss.c
+++ b/plugins/oss/oss.c
@@ -278,13 +278,7 @@ oss_thread (void *context) {
 
 static int
 oss_callback (char *stream, int len) {
-    int bytesread = deadbeef->streamer_read (stream, len);
-    int16_t ivolume = deadbeef->volume_get_amp () * 1000;
-    for (int i = 0; i < bytesread/2; i++) {
-        ((int16_t*)stream)[i] = (int16_t)(((int32_t)(((int16_t*)stream)[i])) * ivolume / 1000);
-    }
-
-    return bytesread;
+    return deadbeef->streamer_read (stream, len);
 }
 
 static int
diff --git a/streamer.c b/streamer.c
index a6f0ff8c..861e39a2 100644
--- a/streamer.c
+++ b/streamer.c
@@ -1511,6 +1511,54 @@ streamer_read (char *bytes, int size) {
     int ms = (tm2.tv_sec*1000+tm2.tv_usec/1000) - (tm1.tv_sec*1000+tm1.tv_usec/1000);
     printf ("streamer_read took %d ms\n", ms);
 #endif
+
+    if (!output->has_volume) {
+        char *stream = bytes;
+        int bytesread = sz;
+        if (output->fmt.bps == 16) {
+            int16_t ivolume = volume_get_amp () * 1000;
+            for (int i = 0; i < bytesread/2; i++) {
+                int16_t sample = *((int16_t*)stream);
+                *((int16_t*)stream) = (int16_t)(((int32_t)sample) * ivolume / 1000);
+                stream += 2;
+            }
+        }
+        else if (output->fmt.bps == 8) {
+            int16_t ivolume = volume_get_amp () * 255;
+            for (int i = 0; i < bytesread; i++) {
+                *stream = (int8_t)(((int32_t)(*stream)) * ivolume / 1000);
+                stream++;
+            }
+        }
+        else if (output->fmt.bps == 24) {
+            int16_t ivolume = volume_get_amp () * 1000;
+            for (int i = 0; i < bytesread/3; i++) {
+                int32_t sample = ((unsigned char)stream[0]) | ((unsigned char)stream[1]<<8) | (stream[2]<<16);
+                int32_t newsample = (int64_t)sample * ivolume / 1000;
+                stream[0] = (newsample&0x0000ff);
+                stream[1] = (newsample&0x00ff00)>>8;
+                stream[2] = (newsample&0xff0000)>>16;
+                stream += 3;
+            }
+        }
+        else if (output->fmt.bps == 32 && !output->fmt.is_float) {
+            int16_t ivolume = volume_get_amp () * 1000;
+            for (int i = 0; i < bytesread/4; i++) {
+                int32_t sample = *((int32_t*)stream);
+                int32_t newsample = (int64_t)sample * ivolume / 1000;
+                *((int32_t*)stream) = newsample;
+                stream += 4;
+            }
+        }
+        else if (output->fmt.bps == 32 && output->fmt.is_float) {
+            float fvolume = volume_get_amp ();
+            for (int i = 0; i < bytesread/4; i++) {
+                *((float*)stream) = (*((float*)stream)) * fvolume;
+                stream += 4;
+            }
+        }
+    }
+
     return sz;
 }