support for 32 bit mixing, and multiple sample formats

10 files changed, 237 insertions, 18 deletions

diff --git a/include/adlmidi.h b/include/adlmidi.h
index b21df2c..8ba8939 100644
--- a/include/adlmidi.h
+++ b/include/adlmidi.h
@@ -65,6 +65,22 @@ enum ADLMIDI_VolumeModels
     ADLMIDI_VolumeModel_9X
 };
 
+enum ADLMIDI_SampleType
+{
+    ADLMIDI_SampleType_S16 = 0,  /* signed PCM 16-bit */
+    ADLMIDI_SampleType_S8,       /* signed PCM 8-bit */
+    ADLMIDI_SampleType_F32,      /* float 32-bit */
+    ADLMIDI_SampleType_F64,      /* float 64-bit */
+    ADLMIDI_SampleType_Count,
+};
+
+struct ADLMIDI_AudioFormat
+{
+    enum ADLMIDI_SampleType type;  /* type of sample */
+    unsigned containerSize;        /* size in bytes of the storage type */
+    unsigned sampleOffset;         /* distance in bytes between consecutive samples */
+};
+
 struct ADL_MIDIPlayer
 {
     void *adl_midiPlayer;
@@ -237,8 +253,14 @@ extern struct Adl_MarkerEntry adl_metaMarker(struct ADL_MIDIPlayer *device, size
 /*Take a sample buffer and iterate MIDI timers */
 extern int  adl_play(struct ADL_MIDIPlayer *device, int sampleCount, short out[]);
 
+/*Take a sample buffer and iterate MIDI timers */
+extern int  adl_playFormat(struct ADL_MIDIPlayer *device, int sampleCount, ADL_UInt8 left[], ADL_UInt8 right[], const struct ADLMIDI_AudioFormat *format);
+
+/*Generate audio output from chip emulators without iteration of MIDI timers.*/
+extern int  adl_generate(struct ADL_MIDIPlayer *device, int sampleCount, short out[]);
+
 /*Generate audio output from chip emulators without iteration of MIDI timers.*/
-extern int  adl_generate(struct ADL_MIDIPlayer *device, int sampleCount, short *out);
+extern int  adl_generateFormat(struct ADL_MIDIPlayer *device, int sampleCount, ADL_UInt8 left[], ADL_UInt8 right[], const struct ADLMIDI_AudioFormat *format);
 
 /**
  * @brief Periodic tick handler.
diff --git a/src/adlmidi.cpp b/src/adlmidi.cpp
index 7e4d457..39bb5b6 100644
--- a/src/adlmidi.cpp
+++ b/src/adlmidi.cpp
@@ -37,6 +37,13 @@ static ADL_Version adl_version = {
     ADLMIDI_VERSION_PATCHLEVEL
 };
 
+static const ADLMIDI_AudioFormat adl_DefaultAudioFormat =
+{
+    ADLMIDI_SampleType_S16,
+    sizeof(short),
+    2 * sizeof(short),
+};
+
 /*---------------------------EXPORTS---------------------------*/
 
 ADLMIDI_EXPORT struct ADL_MIDIPlayer *adl_init(long sample_rate)
@@ -608,30 +615,107 @@ ADLMIDI_EXPORT void adl_setDebugMessageHook(struct ADL_MIDIPlayer *device, ADL_D
 }
 
 
+template <class Dst>
+static void CopySamplesRaw(ADL_UInt8 *dstLeft, ADL_UInt8 *dstRight, const int32_t *src,
+                           size_t frameCount, unsigned sampleOffset)
+{
+    for(size_t i = 0; i < frameCount; ++i) {
+        *(Dst *)(dstLeft + i * sampleOffset) = src[2 * i];
+        *(Dst *)(dstRight + i * sampleOffset) = src[2 * i + 1];
+    }
+}
 
-inline static void SendStereoAudio(int      &samples_requested,
-                                   ssize_t  &in_size,
-                                   short    *_in,
-                                   ssize_t   out_pos,
-                                   short    *_out)
+template <class Dst, class Ret>
+static void CopySamplesTransformed(ADL_UInt8 *dstLeft, ADL_UInt8 *dstRight, const int32_t *src,
+                                   size_t frameCount, unsigned sampleOffset,
+                                   Ret(&transform)(int32_t))
+{
+    for(size_t i = 0; i < frameCount; ++i) {
+        *(Dst *)(dstLeft + i * sampleOffset) = transform(src[2 * i]);
+        *(Dst *)(dstRight + i * sampleOffset) = transform(src[2 * i + 1]);
+    }
+}
+
+static int SendStereoAudio(int        samples_requested,
+                           ssize_t    in_size,
+                           int32_t   *_in,
+                           ssize_t    out_pos,
+                           ADL_UInt8 *left,
+                           ADL_UInt8 *right,
+                           const ADLMIDI_AudioFormat *format)
 {
     if(!in_size)
-        return;
-    size_t offset       = static_cast<size_t>(out_pos);
+        return 0;
+    size_t outputOffset = static_cast<size_t>(out_pos);
     size_t inSamples    = static_cast<size_t>(in_size * 2);
-    size_t maxSamples   = static_cast<size_t>(samples_requested) - offset;
+    size_t maxSamples   = static_cast<size_t>(samples_requested) - outputOffset;
     size_t toCopy       = std::min(maxSamples, inSamples);
-    std::memcpy(_out + out_pos, _in, toCopy * sizeof(short));
+
+    ADLMIDI_SampleType sampleType = format->type;
+    const unsigned containerSize = format->containerSize;
+    const unsigned sampleOffset = format->sampleOffset;
+
+    switch(sampleType) {
+    case ADLMIDI_SampleType_S8:
+        switch(containerSize) {
+        case sizeof(int8_t):
+            CopySamplesTransformed<int8_t>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtS8);
+            break;
+        case sizeof(int16_t):
+            CopySamplesTransformed<int16_t>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtS8);
+            break;
+        case sizeof(int32_t):
+            CopySamplesTransformed<int32_t>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtS8);
+            break;
+        default:
+            return -1;
+        }
+        break;
+    case ADLMIDI_SampleType_S16:
+        switch(containerSize) {
+        case sizeof(int16_t):
+            CopySamplesTransformed<int16_t>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtS16);
+            break;
+        case sizeof(int32_t):
+            CopySamplesRaw<int32_t>(left, right, _in, toCopy / 2, sampleOffset);
+            break;
+        default:
+            return -1;
+        }
+        break;
+    case ADLMIDI_SampleType_F32:
+        if(containerSize != sizeof(float))
+            return -1;
+        CopySamplesTransformed<float>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtReal<float>);
+        break;
+    case ADLMIDI_SampleType_F64:
+        if(containerSize != sizeof(double))
+            return -1;
+        CopySamplesTransformed<double>(left, right, _in, toCopy / 2, sampleOffset, adl_cvtReal<double>);
+        break;
+    default:
+        return -1;
+    }
+
+    return 0;
 }
 
 
-ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
+ADLMIDI_EXPORT int adl_play(struct ADL_MIDIPlayer *device, int sampleCount, short *out)
+{
+    return adl_playFormat(device, sampleCount, (ADL_UInt8 *)out, (ADL_UInt8 *)(out + 1), &adl_DefaultAudioFormat);
+}
+
+ADLMIDI_EXPORT int adl_playFormat(ADL_MIDIPlayer *device, int sampleCount,
+                                  ADL_UInt8 out_left[], ADL_UInt8 out_right[],
+                                  const ADLMIDI_AudioFormat *format)
 {
     #ifndef ADLMIDI_DISABLE_MIDI_SEQUENCER
     #ifdef ADLMIDI_HW_OPL
     (void)device;
     (void)sampleCount;
     (void)out;
+    (void)format;
     return 0;
     #else
     sampleCount -= sampleCount % 2; //Avoid even sample requests
@@ -685,8 +769,8 @@ ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
                 ssize_t in_generatedPhys = in_generatedStereo * 2;
                 //! Unsigned total sample count
                 //fill buffer with zeros
-                int16_t *out_buf = player->outBuf;
-                std::memset(out_buf, 0, static_cast<size_t>(in_generatedPhys) * sizeof(int16_t));
+                int32_t *out_buf = player->outBuf;
+                std::memset(out_buf, 0, static_cast<size_t>(in_generatedPhys) * sizeof(out_buf[0]));
                 unsigned int chips = player->opl.NumCards;
                 if(chips == 1)
                 {
@@ -699,7 +783,8 @@ ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
                         player->opl.cardsOP2[card]->generateAndMix(out_buf, (size_t)in_generatedStereo);
                 }
                 /* Process it */
-                SendStereoAudio(sampleCount, in_generatedStereo, out_buf, gotten_len, out);
+                if(SendStereoAudio(sampleCount, in_generatedStereo, out_buf, gotten_len, out_left, out_right, format) == -1)
+                    return 0;
 
                 left -= (int)in_generatedPhys;
                 gotten_len += (in_generatedPhys) /* - setup.stored_samples*/;
@@ -726,10 +811,18 @@ ADLMIDI_EXPORT int adl_play(ADL_MIDIPlayer *device, int sampleCount, short *out)
 
 ADLMIDI_EXPORT int adl_generate(struct ADL_MIDIPlayer *device, int sampleCount, short *out)
 {
+    return adl_generateFormat(device, sampleCount, (ADL_UInt8 *)out, (ADL_UInt8 *)(out + 1), &adl_DefaultAudioFormat);
+}
+
+ADLMIDI_EXPORT int adl_generateFormat(struct ADL_MIDIPlayer *device, int sampleCount,
+                                      ADL_UInt8 out_left[], ADL_UInt8 out_right[],
+                                      const ADLMIDI_AudioFormat *format)
+{
     #ifdef ADLMIDI_HW_OPL
     (void)device;
     (void)sampleCount;
     (void)out;
+    (void)format;
     return 0;
     #else
     sampleCount -= sampleCount % 2; //Avoid even sample requests
@@ -766,8 +859,8 @@ ADLMIDI_EXPORT int adl_generate(struct ADL_MIDIPlayer *device, int sampleCount,
                 ssize_t in_generatedPhys = in_generatedStereo * 2;
                 //! Unsigned total sample count
                 //fill buffer with zeros
-                int16_t *out_buf = player->outBuf;
-                std::memset(out_buf, 0, static_cast<size_t>(in_generatedPhys) * sizeof(int16_t));
+                int32_t *out_buf = player->outBuf;
+                std::memset(out_buf, 0, static_cast<size_t>(in_generatedPhys) * sizeof(out_buf[0]));
                 unsigned int chips = player->opl.NumCards;
                 if(chips == 1)
                     player->opl.cardsOP2[0]->generate(out_buf, (size_t)in_generatedStereo);
@@ -778,7 +871,8 @@ ADLMIDI_EXPORT int adl_generate(struct ADL_MIDIPlayer *device, int sampleCount,
                         player->opl.cardsOP2[card]->generateAndMix(out_buf, (size_t)in_generatedStereo);
                 }
                 /* Process it */
-                SendStereoAudio(sampleCount, in_generatedStereo, out_buf, gotten_len, out);
+                if(SendStereoAudio(sampleCount, in_generatedStereo, out_buf, gotten_len, out_left, out_right, format) == -1)
+                    return 0;
 
                 left -= (int)in_generatedPhys;
                 gotten_len += (in_generatedPhys) /* - setup.stored_samples*/;
diff --git a/src/adlmidi_private.hpp b/src/adlmidi_private.hpp
index caaf4f9..63cadb3 100644
--- a/src/adlmidi_private.hpp
+++ b/src/adlmidi_private.hpp
@@ -132,6 +132,25 @@ typedef int32_t ssize_t;
 extern std::string ADLMIDI_ErrorString;
 
 /*
+  Sample conversions to various formats
+*/
+template <class Real>
+inline Real adl_cvtReal(int32_t x)
+{
+    return x * ((Real)1 / INT16_MAX);
+}
+inline int32_t adl_cvtS16(int32_t x)
+{
+    x = (x < INT16_MIN) ? INT16_MIN : x;
+    x = (x > INT16_MAX) ? INT16_MAX : x;
+    return x;
+}
+inline int32_t adl_cvtS8(int32_t x)
+{
+    return adl_cvtS16(x) / 256;
+}
+
+/*
     Smart pointer for C heaps, created with malloc() call.
     FAQ: Why not std::shared_ptr? Because of Android NDK now doesn't supports it
 */
@@ -909,7 +928,7 @@ public:
 #endif
     OPL3 opl;
 
-    int16_t outBuf[1024];
+    int32_t outBuf[1024];
 
     Setup m_setup;
 
diff --git a/src/chips/dosbox_opl3.cpp b/src/chips/dosbox_opl3.cpp
index b696343..57b1412 100644
--- a/src/chips/dosbox_opl3.cpp
+++ b/src/chips/dosbox_opl3.cpp
@@ -78,6 +78,34 @@ int DosBoxOPL3::generateAndMix(int16_t *output, size_t frames)
     return (int)frames;
 }
 
+int DosBoxOPL3::generate(int32_t *output, size_t frames)
+{
+    DBOPL::Handler *chip_r = reinterpret_cast<DBOPL::Handler*>(m_chip);
+    ssize_t left = (ssize_t)frames;
+    while(left > 0)
+    {
+        ssize_t frames_i = left;
+        chip_r->GenerateArr(output, &frames_i);
+        output += (frames_i * 2);
+        left -= frames_i;
+    }
+    return (int)frames;
+}
+
+int DosBoxOPL3::generateAndMix(int32_t *output, size_t frames)
+{
+    DBOPL::Handler *chip_r = reinterpret_cast<DBOPL::Handler*>(m_chip);
+    ssize_t left = (ssize_t)frames;
+    while(left > 0)
+    {
+        ssize_t frames_i = left;
+        chip_r->GenerateArrMix(output, &frames_i);
+        output += (frames_i * 2);
+        left -= frames_i;
+    }
+    return (int)frames;
+}
+
 const char *DosBoxOPL3::emulatorName()
 {
     return "DosBox 0.74 OPL3";
diff --git a/src/chips/dosbox_opl3.h b/src/chips/dosbox_opl3.h
index 46c20fa..05682ae 100644
--- a/src/chips/dosbox_opl3.h
+++ b/src/chips/dosbox_opl3.h
@@ -17,6 +17,8 @@ public:
     virtual void writeReg(uint16_t addr, uint8_t data) override;
     virtual int generate(int16_t *output, size_t frames) override;
     virtual int generateAndMix(int16_t *output, size_t frames) override;
+    virtual int generate(int32_t *output, size_t frames) override;
+    virtual int generateAndMix(int32_t *output, size_t frames) override;
     virtual const char *emulatorName() override;
 };
 
diff --git a/src/chips/nuked_opl3.cpp b/src/chips/nuked_opl3.cpp
index 0898907..16cc73b 100644
--- a/src/chips/nuked_opl3.cpp
+++ b/src/chips/nuked_opl3.cpp
@@ -61,6 +61,30 @@ int NukedOPL3::generateAndMix(int16_t *output, size_t frames)
     return (int)frames;
 }
 
+int NukedOPL3::generate(int32_t *output, size_t frames)
+{
+    opl3_chip *chip_r = reinterpret_cast<opl3_chip*>(m_chip);
+    for(size_t i = 0; i < frames; ++i) {
+        int16_t frame[2];
+        OPL3_GenerateResampled(chip_r, frame);
+        output[2 * i] = frame[0];
+        output[2 * i + 1] = frame[1];
+    }
+    return (int)frames;
+}
+
+int NukedOPL3::generateAndMix(int32_t *output, size_t frames)
+{
+    opl3_chip *chip_r = reinterpret_cast<opl3_chip*>(m_chip);
+    for(size_t i = 0; i < frames; ++i) {
+        int16_t frame[2];
+        OPL3_GenerateResampled(chip_r, frame);
+        output[2 * i] += frame[0];
+        output[2 * i + 1] += frame[1];
+    }
+    return (int)frames;
+}
+
 const char *NukedOPL3::emulatorName()
 {
     return "Nuked OPL3 (v 1.8)";
diff --git a/src/chips/nuked_opl3.h b/src/chips/nuked_opl3.h
index db665fe..144fb23 100644
--- a/src/chips/nuked_opl3.h
+++ b/src/chips/nuked_opl3.h
@@ -17,6 +17,8 @@ public:
     virtual void writeReg(uint16_t addr, uint8_t data) override;
     virtual int generate(int16_t *output, size_t frames) override;
     virtual int generateAndMix(int16_t *output, size_t frames) override;
+    virtual int generate(int32_t *output, size_t frames) override;
+    virtual int generateAndMix(int32_t *output, size_t frames) override;
     virtual const char *emulatorName() override;
 };
 
diff --git a/src/chips/nuked_opl3_v174.cpp b/src/chips/nuked_opl3_v174.cpp
index 7188e83..59f7953 100644
--- a/src/chips/nuked_opl3_v174.cpp
+++ b/src/chips/nuked_opl3_v174.cpp
@@ -61,6 +61,30 @@ int NukedOPL3v174::generateAndMix(int16_t *output, size_t frames)
     return (int)frames;
 }
 
+int NukedOPL3v174::generate(int32_t *output, size_t frames)
+{
+    opl3_chip *chip_r = reinterpret_cast<opl3_chip*>(m_chip);
+    for(size_t i = 0; i < frames; ++i) {
+        int16_t frame[2];
+        OPL3v17_GenerateResampled(chip_r, frame);
+        output[2 * i] = frame[0];
+        output[2 * i + 1] = frame[1];
+    }
+    return (int)frames;
+}
+
+int NukedOPL3v174::generateAndMix(int32_t *output, size_t frames)
+{
+    opl3_chip *chip_r = reinterpret_cast<opl3_chip*>(m_chip);
+    for(size_t i = 0; i < frames; ++i) {
+        int16_t frame[2];
+        OPL3v17_GenerateResampled(chip_r, frame);
+        output[2 * i] += frame[0];
+        output[2 * i + 1] += frame[1];
+    }
+    return (int)frames;
+}
+
 const char *NukedOPL3v174::emulatorName()
 {
     return "Nuked OPL3 (v 1.7.4)";
diff --git a/src/chips/nuked_opl3_v174.h b/src/chips/nuked_opl3_v174.h
index cf23187..b9a0466 100644
--- a/src/chips/nuked_opl3_v174.h
+++ b/src/chips/nuked_opl3_v174.h
@@ -17,6 +17,8 @@ public:
     virtual void writeReg(uint16_t addr, uint8_t data) override;
     virtual int generate(int16_t *output, size_t frames) override;
     virtual int generateAndMix(int16_t *output, size_t frames) override;
+    virtual int generate(int32_t *output, size_t frames) override;
+    virtual int generateAndMix(int32_t *output, size_t frames) override;
     virtual const char *emulatorName() override;
 };
 
diff --git a/src/chips/opl_chip_base.h b/src/chips/opl_chip_base.h
index aa6934b..1aaddf8 100644
--- a/src/chips/opl_chip_base.h
+++ b/src/chips/opl_chip_base.h
@@ -19,6 +19,8 @@ public:
     virtual void writeReg(uint16_t addr, uint8_t data) = 0;
     virtual int generate(int16_t *output, size_t frames) = 0;
     virtual int generateAndMix(int16_t *output, size_t frames) = 0;
+    virtual int generate(int32_t *output, size_t frames) = 0;
+    virtual int generateAndMix(int32_t *output, size_t frames) = 0;
     virtual const char* emulatorName() = 0;
 };