Merge branch 'master' into seekability

# Conflicts:
#	libADLMIDI-test.pro
#	src/adlmidi.cpp
#	src/adlmidi_private.hpp

4 files changed, 99 insertions, 70 deletions

diff --git a/src/adlmidi.cpp b/src/adlmidi.cpp
index a7bab22..7aa9e8b 100644
--- a/src/adlmidi.cpp
+++ b/src/adlmidi.cpp
@@ -346,43 +346,6 @@ ADLMIDI_EXPORT void adl_setTempo(struct ADL_MIDIPlayer *device, double tempo)
 }
 
 
-#ifdef ADLMIDI_USE_DOSBOX_OPL
-
-#define ADLMIDI_CLAMP(V, MIN, MAX) std::max(std::min(V, (MAX)), (MIN))
-
-inline static void SendStereoAudio(MIDIplay::Setup &device,
-                                   int      &samples_requested,
-                                   ssize_t  &in_size,
-                                   int      *_in,
-                                   ssize_t  out_pos,
-                                   short    *_out)
-{
-    if(!in_size) return;
-
-    device.stored_samples = 0;
-    ssize_t out;
-    ssize_t offset;
-    ssize_t pos = static_cast<ssize_t>(out_pos);
-
-    for(ssize_t p = 0; p < in_size; ++p)
-    {
-        for(ssize_t w = 0; w < 2; ++w)
-        {
-            out    = _in[p * 2 + w];
-            offset = pos + p * 2 + w;
-
-            if(offset < samples_requested)
-                _out[offset] = static_cast<short>(ADLMIDI_CLAMP(out, static_cast<ssize_t>(INT16_MIN), static_cast<ssize_t>(INT16_MAX)));
-            else
-            {
-                device.backup_samples[device.backup_samples_size] = static_cast<short>(out);
-                device.backup_samples_size++;
-                device.stored_samples++;
-            }
-        }
-    }
-}
-#else
 inline static void SendStereoAudio(MIDIplay::Setup &device,
                                    int      &samples_requested,
                                    ssize_t  &in_size,
@@ -409,7 +372,6 @@ inline static void SendStereoAudio(MIDIplay::Setup &device,
         device.stored_samples += (ssize_t)appendSize;
     }
 }
-#endif
 
 
 ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
@@ -460,70 +422,51 @@ ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
             setup.delay -= eat_delay;
             setup.carry += setup.PCM_RATE * eat_delay;
             n_periodCountStereo = static_cast<ssize_t>(setup.carry);
-            //n_periodCountPhys = n_periodCountStereo * 2;
             setup.carry -= n_periodCountStereo;
 
             if(setup.SkipForward > 0)
                 setup.SkipForward -= 1;
             else
             {
-                #ifdef ADLMIDI_USE_DOSBOX_OPL
-                std::vector<int>     out_buf;
-                #else
-                std::vector<int16_t> out_buf;
-                #endif
-                out_buf.resize(1024, 0);
-
                 if((player->atEnd) && (setup.delay <= 0.0))
                     break;//Stop to fetch samples at reaching the song end with disabled loop
 
+                //! Count of stereo samples
                 ssize_t in_generatedStereo = (n_periodCountStereo > 512) ? 512 : n_periodCountStereo;
+                //! Total count of samples
                 ssize_t in_generatedPhys = in_generatedStereo * 2;
-
+                //! Unsigned total sample count
                 //fill buffer with zeros
-                size_t in_countStereoU = static_cast<size_t>(in_generatedStereo * 2);
+                int16_t *out_buf = player->outBuf;
+                std::memset(out_buf, 0, static_cast<size_t>(in_generatedPhys) * sizeof(int16_t));
 
                 if(player->m_setup.NumCards == 1)
                 {
                     #ifdef ADLMIDI_USE_DOSBOX_OPL
-                    player->opl.cards[0].GenerateArr(out_buf.data(), &in_generatedStereo);
+                    player->opl.cards[0].GenerateArr(out_buf, &in_generatedStereo);
                     in_generatedPhys = in_generatedStereo * 2;
                     #else
-                    OPL3_GenerateStream(&player->opl.cards[0], out_buf.data(), static_cast<Bit32u>(in_generatedStereo));
+                    OPL3_GenerateStream(&player->opl.cards[0], out_buf, static_cast<Bit32u>(in_generatedStereo));
                     #endif
-                    /* Process it */
-                    SendStereoAudio(setup, sampleCount, in_generatedStereo, out_buf.data(), gotten_len, out);
                 }
                 else if(n_periodCountStereo > 0)
                 {
-                    #ifdef ADLMIDI_USE_DOSBOX_OPL
-                    std::vector<int32_t> in_mixBuffer;
-                    in_mixBuffer.resize(1024); //n_samples * 2
-                    ssize_t in_generatedStereo = n_periodCountStereo;
-                    #endif
-                    std::memset(out_buf.data(), 0, in_countStereoU * sizeof(short));
-
                     /* Generate data from every chip and mix result */
                     for(unsigned card = 0; card < player->m_setup.NumCards; ++card)
                     {
                         #ifdef ADLMIDI_USE_DOSBOX_OPL
-                        player->opl.cards[card].GenerateArr(in_mixBuffer.data(), &in_generatedStereo);
+                        player->opl.cards[card].GenerateArrMix(out_buf, &in_generatedStereo);
                         in_generatedPhys = in_generatedStereo * 2;
-                        size_t in_samplesCount = static_cast<size_t>(in_generatedPhys);
-                        for(size_t a = 0; a < in_samplesCount; ++a)
-                            out_buf[a] += in_mixBuffer[a];
                         #else
-                        OPL3_GenerateStreamMix(&player->opl.cards[card], out_buf.data(), static_cast<Bit32u>(in_generatedStereo));
+                        OPL3_GenerateStreamMix(&player->opl.cards[card], out_buf, static_cast<Bit32u>(in_generatedStereo));
                         #endif
                     }
-
-                    /* Process it */
-                    SendStereoAudio(setup, sampleCount, in_generatedStereo, out_buf.data(), gotten_len, out);
                 }
+                /* Process it */
+                SendStereoAudio(setup, sampleCount, in_generatedStereo, out_buf, gotten_len, out);
 
                 left -= (int)in_generatedPhys;
                 gotten_len += (in_generatedPhys) - setup.stored_samples;
-                out_buf.clear();
             }
 
             setup.delay = player->Tick(eat_delay, setup.mindelay);
diff --git a/src/adlmidi_private.hpp b/src/adlmidi_private.hpp
index a2ec8c1..622e9f2 100644
--- a/src/adlmidi_private.hpp
+++ b/src/adlmidi_private.hpp
@@ -479,6 +479,8 @@ public:
     char ____padding2[2];
     OPL3 opl;
 
+    int16_t outBuf[1024];
+
     struct Setup
     {
         unsigned int AdlBank;
diff --git a/src/dbopl.cpp b/src/dbopl.cpp
index 8bb3eb4..fb28e20 100644
--- a/src/dbopl.cpp
+++ b/src/dbopl.cpp
@@ -45,6 +45,10 @@ typedef struct vswprintf {} swprintf;
 #include <string.h>
 #include "dbopl.h"
 
+#define DB_MAX(x, y) ((x) > (y) ? (x) : (y))
+#define DB_MIN(x, y) ((x) < (y) ? (x) : (y))
+
+#define DBOPL_CLAMP(V, MIN, MAX) DB_MAX(DB_MIN(V, (MAX)), (MIN))
 
 #ifndef PI
 #define PI 3.14159265358979323846
@@ -1482,7 +1486,7 @@ namespace DBOPL
     {
         while(total > 0)
         {
-            Bit32u samples = ForwardLFO(total);
+            Bit32u samples = ForwardLFO((Bit32u)total);
             memset(output, 0, sizeof(Bit32s) * samples * 2);
             int count = 0;
 
@@ -1497,6 +1501,39 @@ namespace DBOPL
         }
     }
 
+    void Chip::GenerateBlock2_Mix(Bitu total, Bit32s *output)
+    {
+        while(total > 0)
+        {
+            Bit32u samples = ForwardLFO((Bit32u)total);
+            int count = 0;
+            for(Channel *ch = chan; ch < chan + 9;)
+            {
+                count++;
+                ch = (ch->*(ch->synthHandler))(this, samples, output);
+            }
+
+            total -= samples;
+            output += samples;
+        }
+    }
+
+    void Chip::GenerateBlock3_Mix(Bitu total, Bit32s *output)
+    {
+        while(total > 0)
+        {
+            Bit32u samples = ForwardLFO(total);
+            int count = 0;
+            for(Channel *ch = chan; ch < chan + 18;)
+            {
+                count++;
+                ch = (ch->*(ch->synthHandler))(this, samples, output);
+            }
+            total -= samples;
+            output += samples * 2;
+        }
+    }
+
     void Chip::Setup(Bit32u rate)
     {
         double original = OPLRATE;
@@ -1955,10 +1992,51 @@ namespace DBOPL
             chip.GenerateBlock3(static_cast<Bitu>(*samples), out);
     }
 
+    void Handler::GenerateArr(Bit16s *out, ssize_t *samples)
+    {
+        Bit32s out32[1024];
+        if(GCC_UNLIKELY(*samples > 512))
+            *samples = 512;
+        memset(out32, 0, sizeof(Bit32s) * 1024);
+        if(!chip.opl3Active)
+            chip.GenerateBlock2(static_cast<Bitu>(*samples), out32);
+        else
+            chip.GenerateBlock3(static_cast<Bitu>(*samples), out32);
+        ssize_t sz = *samples * 2;
+        for(ssize_t i = 0; i < sz; i++)
+            out[i] = static_cast<Bit16s>(DBOPL_CLAMP(out32[i], static_cast<ssize_t>(INT16_MIN), static_cast<ssize_t>(INT16_MAX)));
+    }
+
+    void Handler::GenerateArrMix(Bit32s *out, ssize_t *samples)
+    {
+        if(GCC_UNLIKELY(*samples > 512))
+            *samples = 512;
+        if(!chip.opl3Active)
+            chip.GenerateBlock2_Mix(static_cast<Bitu>(*samples), out);
+        else
+            chip.GenerateBlock3_Mix(static_cast<Bitu>(*samples), out);
+    }
+
+    void Handler::GenerateArrMix(Bit16s *out, ssize_t *samples)
+    {
+        Bit32s out32[1024];
+        if(GCC_UNLIKELY(*samples > 512))
+            *samples = 512;
+        memset(out32, 0, sizeof(Bit32s) * 1024);
+        if(!chip.opl3Active)
+            chip.GenerateBlock2(static_cast<Bitu>(*samples), out32);
+        else
+            chip.GenerateBlock3(static_cast<Bitu>(*samples), out32);
+        ssize_t sz = *samples * 2;
+        for(ssize_t i = 0; i < sz; i++)
+            out[i] += static_cast<Bit16s>(DBOPL_CLAMP(out32[i], static_cast<ssize_t>(INT16_MIN), static_cast<ssize_t>(INT16_MAX)));
+    }
+
+
     void Handler::Init(Bitu rate)
     {
         InitTables();
-        chip.Setup(rate);
+        chip.Setup((Bit32u)rate);
     }
 
 
diff --git a/src/dbopl.h b/src/dbopl.h
index 54384da..826cb44 100644
--- a/src/dbopl.h
+++ b/src/dbopl.h
@@ -277,6 +277,9 @@ namespace DBOPL
         void GenerateBlock2(Bitu samples, Bit32s *output);
         void GenerateBlock3(Bitu samples, Bit32s *output);
 
+        void GenerateBlock2_Mix(Bitu samples, Bit32s *output);
+        void GenerateBlock3_Mix(Bitu samples, Bit32s *output);
+
         void Generate(Bit32u samples);
         void Setup(Bit32u r);
 
@@ -293,6 +296,9 @@ namespace DBOPL
                       Bitu samples);
         void GenerateArr(Bit32s *out, Bitu *samples);
         void GenerateArr(Bit32s *out, ssize_t *samples);
+        void GenerateArr(Bit16s *out, ssize_t *samples);
+        void GenerateArrMix(Bit32s *out, ssize_t *samples);
+        void GenerateArrMix(Bit16s *out, ssize_t *samples);
         void Init(Bitu rate);
     };