engines/gig/Synthesizer.h

/***************************************************************************
 *                                                                         *
 *   LinuxSampler - modular, streaming capable sampler                     *
 *                                                                         *
 *   Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck   *
 *   Copyright (C) 2005 Christian Schoenebeck                              *
 *                                                                         *
 *   This program is free software; you can redistribute it and/or modify  *
 *   it under the terms of the GNU General Public License as published by  *
 *   the Free Software Foundation; either version 2 of the License, or     *
 *   (at your option) any later version.                                   *
 *                                                                         *
 *   This program is distributed in the hope that it will be useful,       *
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of        *
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the         *
 *   GNU General Public License for more details.                          *
 *                                                                         *
 *   You should have received a copy of the GNU General Public License     *
 *   along with this program; if not, write to the Free Software           *
 *   Foundation, Inc., 59 Temple Place, Suite 330, Boston,                 *
 *   MA  02111-1307  USA                                                   *
 ***************************************************************************/

#ifndef __LS_GIG_SYNTHESIZER_H__
#define __LS_GIG_SYNTHESIZER_H__

#include "../../common/global.h"
#include "../../common/RTMath.h"
#include "../common/Resampler.h"
#include "../common/BiquadFilter.h"
#include "Filter.h"
#include "Voice.h"

#define SYNTHESIS_MODE_SET_CONSTPITCH(iMode,bVal)       if (bVal) iMode |= 0x01; else iMode &= ~0x01   /* (un)set mode bit 0 */
#define SYNTHESIS_MODE_SET_LOOP(iMode,bVal)             if (bVal) iMode |= 0x02; else iMode &= ~0x02   /* (un)set mode bit 1 */
#define SYNTHESIS_MODE_SET_INTERPOLATE(iMode,bVal)      if (bVal) iMode |= 0x04; else iMode &= ~0x04   /* (un)set mode bit 2 */
#define SYNTHESIS_MODE_SET_FILTER(iMode,bVal)           if (bVal) iMode |= 0x08; else iMode &= ~0x08   /* (un)set mode bit 3 */
#define SYNTHESIS_MODE_SET_CHANNELS(iMode,bVal)         if (bVal) iMode |= 0x10; else iMode &= ~0x10   /* (un)set mode bit 4 */
#define SYNTHESIS_MODE_SET_IMPLEMENTATION(iMode,bVal)   if (bVal) iMode |= 0x20; else iMode &= ~0x20   /* (un)set mode bit 5 */
#define SYNTHESIS_MODE_SET_PROFILING(iMode,bVal)        if (bVal) iMode |= 0x40; else iMode &= ~0x40   /* (un)set mode bit 6 */

#define SYNTHESIS_MODE_GET_CONSTPITCH(iMode)            iMode & 0x01
#define SYNTHESIS_MODE_GET_LOOP(iMode)                  iMode & 0x02
#define SYNTHESIS_MODE_GET_INTERPOLATE(iMode)           iMode & 0x04
#define SYNTHESIS_MODE_GET_FILTER(iMode)                iMode & 0x08
#define SYNTHESIS_MODE_GET_CHANNELS(iMode)              iMode & 0x10
#define SYNTHESIS_MODE_GET_IMPLEMENTATION(iMode)        iMode & 0x20

// that's usually gig::Voice of course, but we make it a macro so we can
// include this code for our synthesis benchmark which uses fake data
// structures
#ifndef VOICE
# define VOICE Voice
#endif // VOICE

namespace LinuxSampler { namespace gig {

    typedef void SynthesizeFragment_Fn(VOICE&, uint, sample_t*, uint);

    void* GetSynthesisFunction(const int SynthesisMode);
    void RunSynthesisFunction(const int SynthesisMode, VOICE& voice, uint Samples, sample_t* pSrc, uint Skip);

    enum channels_t {
        MONO,
        STEREO
    };

    template<implementation_t IMPLEMENTATION, channels_t CHANNELS, bool USEFILTER, bool INTERPOLATE, bool DOLOOP, bool CONSTPITCH>
    class Synthesizer : public __RTMath<IMPLEMENTATION>, public LinuxSampler::Resampler<INTERPOLATE> {

            // declarations of derived functions (see "Name lookup,
            // templates, and accessing members of base classes" in
            // the gcc manual for an explanation of why this is
            // needed).
            using __RTMath<IMPLEMENTATION>::Mul;
            using __RTMath<IMPLEMENTATION>::Float;
            using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleMonoCPP;
            using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleStereoCPP;
#if ARCH_X86
            using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesMonoMMXSSE;
            using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesStereoMMXSSE;
#endif

        public:
            template<typename VOICE_T>
            inline static void SynthesizeFragment(VOICE_T& Voice, uint Samples, sample_t* pSrc, uint i) {
                const float panLeft  = Mul(Voice.PanLeft,  Voice.pEngineChannel->GlobalPanLeft);
                const float panRight = Mul(Voice.PanRight, Voice.pEngineChannel->GlobalPanRight);
                if (IMPLEMENTATION == ASM_X86_MMX_SSE) {
                    float fPos = (float) Voice.Pos;
                    SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount,
                                       Voice.pSample->LoopStart,
                                       Voice.pSample->LoopEnd,
                                       Voice.pSample->LoopSize,
                                       Voice.LoopCyclesLeft,
                                       (void *)&fPos,
                                       Voice.PitchBase,
                                       Voice.PitchBend,
                                       &panLeft, &panRight);
                    #if  ARCH_X86
                    if (INTERPOLATE) EMMS;
                    #endif
                    Voice.Pos = (double) fPos;
                } else {
                    SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount,
                                       Voice.pSample->LoopStart,
                                       Voice.pSample->LoopEnd,
                                       Voice.pSample->LoopSize,
                                       Voice.LoopCyclesLeft,
                                       (void *)&Voice.Pos,
                                       Voice.PitchBase,
                                       Voice.PitchBend,
                                       &panLeft, &panRight);
                }
            }

        //protected:

            template<typename VOICE_T>
            inline static void SynthesizeFragment(VOICE_T& Voice, uint Samples, sample_t* pSrc, uint& i, uint& LoopPlayCount, uint LoopStart, uint LoopEnd, uint LoopSize, uint& LoopCyclesLeft, void* Pos, float& PitchBase, float& PitchBend, const float* PanLeft, const float* PanRight) {
                const float loopEnd = Float(LoopEnd);
                const float PBbyPB = Mul(PitchBase, PitchBend);
                const float f_LoopStart = Float(LoopStart);
                const float f_LoopSize = Float(LoopSize);
                if (DOLOOP) {
                    if (LoopPlayCount) {
                        // render loop (loop count limited)
                        while (i < Samples && LoopCyclesLeft) {
                            if (CONSTPITCH) {
                                const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd,Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up
                                while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
                            }
                            else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
                            if (WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos)) LoopCyclesLeft--;
                        }
                        // render on without loop
                        while (i < Samples) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
                    }
                    else { // render loop (endless loop)
                        while (i < Samples) {
                            if (CONSTPITCH) {
                                const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd, Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up
                                while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
                            }
                            else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
                            WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos);
                        }
                    }
                }
                else { // no looping
                    while (i < Samples) { Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);}
                }
            }

            template<typename VOICE_T>
            inline static void Synthesize(VOICE_T& Voice, void* Pos, sample_t* pSrc, uint& i, const float* PanLeft, const float* PanRight) {
                Synthesize(pSrc, Pos,
                           Voice.pEngine->pSynthesisParameters[Event::destination_vco][i],
                           Voice.pEngineChannel->pOutputLeft,
                           Voice.pEngineChannel->pOutputRight,
                           i,
                           Voice.pEngine->pSynthesisParameters[Event::destination_vca],
                           PanLeft,
                           PanRight,
                           Voice.FilterLeft,
                           Voice.FilterRight,
                           Voice.pEngine->pBasicFilterParameters[i],
                           Voice.pEngine->pMainFilterParameters[i]);
            }

            inline static int DiffToLoopEnd(const float& LoopEnd, const void* Pos, const float& Pitch) {
                switch (IMPLEMENTATION) {
                    // pure C++ implementation (thus platform independent)
                    case CPP: {
                        return uint((LoopEnd - *((double *)Pos)) / Pitch);
                    }
                    #if ARCH_X86
                    case ASM_X86_MMX_SSE: {
                        int result;
                        __asm__ __volatile__ (
                            "movss    (%1), %%xmm0  #read loopend\n\t"
                            "subss    (%2), %%xmm0  #sub  pos\n\t"
                            "divss    (%3), %%xmm0  #div  by pitch\n\t"
                            "cvtss2si %%xmm0, %0    #convert to int\n\t"
                            : "=r" (result)   /* %0 */
                            : "r" (&LoopEnd), /* %1 */
                              "r" (Pos),      /* %2 */
                              "r" (&Pitch)    /* %3 */
                        );
                        return result;
                    }
                    #endif // ARCH_X86
                }
            }

            inline static int WrapLoop(const float& LoopStart, const float& LoopSize, const float& LoopEnd, void* vPos) {
                switch (IMPLEMENTATION) {
                    // pure C++ implementation (thus platform independent)
                    case CPP: {
                        double * Pos = (double *)vPos;
                        if (*Pos < LoopEnd) return 0;
                        *Pos = fmod(*Pos - LoopEnd, LoopSize) + LoopStart;
                        return 1;
                    }
                    #if ARCH_X86
                    case ASM_X86_MMX_SSE: {
                        int result = 0;
                        __asm__ __volatile__ (
                            "movss  (%2), %%xmm0          # load LoopEnd\n\t"
                            "movss  (%1), %%xmm1          # load Pos\n\t"
                            "comiss %%xmm0, %%xmm1      # LoopEnd <> Pos\n\t"
                            "jb     1f                  # jump if no work needs to be done\n\t"
                            "movss    (%3), %%xmm2        # load LoopSize\n\t"
                            "subss    %%xmm0, %%xmm1    # Pos - LoopEnd\n\t"
                            //now the fmodf
                            "movss    %%xmm1, %%xmm3    # xmm3 = (Pos - LoopEnd)\n\t"
                            "divss    %%xmm2, %%xmm1    # (Pos - LoopEnd) / LoopSize\n\t"
                            "cvttss2si %%xmm1, %%eax    # convert to int\n\t"
                            "cvtsi2ss  %%eax, %%xmm1    # convert back to float\n\t"
                            "movss    (%4), %%xmm0      # load LoopStart\n\t"
                            "mulss    %%xmm2, %%xmm1    # LoopSize * int((Pos-LoopEnd)/LoopSize)\n\t"
                            "subss    %%xmm1, %%xmm3    # xmm2 = fmodf(Pos - LoopEnd, LoopSize)\n\t"
                            //done with fmodf
                            "addss    %%xmm0, %%xmm3      # add LoopStart\n\t"
                            "movss    %%xmm3, (%1)        # update Pos\n\t"
                            "movl    $1, (%0)             # result = 1\n\t"
                            ".balign 16 \n\t"
                            "1:\n\t"
                            :: "r" (&result),   /* %0 */
                              "r"  (vPos),      /* %1 */
                              "r"  (&LoopEnd),  /* %2 */
                              "r"  (&LoopSize), /* %3 */
                              "r"  (&LoopStart) /* %4 */
                        );
                        return result;
                    }
                    #endif // ARCH_X86
                }
            }

            inline static void Synthesize(sample_t* pSrc, void* Pos, float& Pitch, float* pOutL, float* pOutR, uint& i, float* Volume, const float* PanL, const float* PanR, Filter& FilterL, Filter& FilterR, biquad_param_t& bqBase, biquad_param_t& bqMain) {
                switch (IMPLEMENTATION) {
                    // pure C++ implementation (thus platform independent)
                    case CPP: {
                        switch (CHANNELS) {
                            case MONO: {
                                float samplePoint = GetNextSampleMonoCPP(pSrc, (double *)Pos, Pitch);
                                if (USEFILTER) samplePoint = FilterL.Apply(&bqBase, &bqMain, samplePoint);
                                pOutL[i] += samplePoint * Volume[i] * *PanL;
                                pOutR[i] += samplePoint * Volume[i] * *PanR;
                                i++;
                                break;
                            }
                            case STEREO: {
                                stereo_sample_t samplePoint = GetNextSampleStereoCPP(pSrc, (double *)Pos, Pitch);
                                if (USEFILTER) {
                                    samplePoint.left  = FilterL.Apply(&bqBase, &bqMain, samplePoint.left);
                                    samplePoint.right = FilterR.Apply(&bqBase, &bqMain, samplePoint.right);
                                }
                                pOutL[i] += samplePoint.left  * Volume[i] * *PanL;
                                pOutR[i] += samplePoint.right * Volume[i] * *PanR;
                                i++;
                                break;
                            }
                        }
                        break;
                    }
                    #if ARCH_X86
                    // Assembly optimization using the MMX & SSE(1) instruction set (thus only for x86)
                    case ASM_X86_MMX_SSE: {
                        const int ii = i & 0xfffffffc;
                        i += 4;
                        switch (CHANNELS) {
                            case MONO: {
                                GetNext4SamplesMonoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2
                                if (USEFILTER) {
                                    /* prepare filter input */
                                    __asm__ __volatile__ (
                                        "movaps %xmm2,%xmm0"
                                    );
                                    FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
                                    __asm__ __volatile__ (
                                        "movaps %xmm7,%xmm2       # mono filter result -> xmm2"
                                    );
                                }
                                /* apply panorama and volume factors */
                                __asm__ __volatile__ (
                                    "movss    (%1),%%xmm0             # load pan left\n\t"
                                    "movss    (%2),%%xmm1             # load pan right\n\t"
                                    "movaps   (%0),%%xmm4             # load vca\n\t"
                                    "shufps   $0x00,%%xmm0,%%xmm0     # copy pan left to the other 3 cells\n\t"
                                    "shufps   $0x00,%%xmm1,%%xmm1     # copy pan right to the other 3 cells\n\t"
                                    "mulps    %%xmm2,%%xmm0           # left  = sample * pan_left\n\t"
                                    "mulps    %%xmm2,%%xmm1           # right = sample * pan_right\n\t"
                                    "mulps    %%xmm4,%%xmm0           # left  = vca * (sample * pan_left)\n\t"
                                    "mulps    %%xmm4,%%xmm1           # right = vca * (sample * pan_right)\n\t"
                                    : /* no output */
                                    : "r" (&Volume[ii]), /* %0 */
                                      "r" (PanL),   /* %1 */
                                      "r" (PanR)    /* %2 */
                                    : "xmm0", /* holds final left  sample (for the 4 samples) at the end */
                                      "xmm1"  /* holds final right sample (for the 4 samples) at the end */
                                );
                                break;
                            }
                            case STEREO: {
                                GetNext4SamplesStereoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2 (left channel) and xmm3 (right channel)
                                if (USEFILTER) {
                                    __asm__ __volatile__ (
                                        "movaps %xmm2,%xmm0     # prepare left channel for filter\n\t"
                                        "movaps %xmm3,%xmm1     # save right channel not to get overwritten by filter algorithms\n\t"
                                    );
                                    FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
                                    __asm__ __volatile__ (
                                        "movaps %xmm1,%xmm0     # prepare right channel for filter\n\t"
                                        "movaps %xmm7,%xmm1     # save filter output for left channel\n\t"
                                    );
                                    FilterR.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
                                    __asm__ __volatile__ (
                                        "movaps %xmm1,%xmm2     # result left channel -> xmm2\n\t"
                                        "movaps %xmm7,%xmm3     # result right channel -> xmm3\n\t"
                                    );
                                }
                                /* apply panorama and volume factors */
                                __asm__ __volatile__ (
                                    "movss    (%1),%%xmm0             # load pan left\n\t"
                                    "movss    (%2),%%xmm1             # load pan right\n\t"
                                    "movaps   (%0),%%xmm4             # load vca\n\t"
                                    "shufps   $0x00,%%xmm0,%%xmm0     # copy pan left to the other 3 cells\n\t"
                                    "shufps   $0x00,%%xmm1,%%xmm1     # copy pan right to the other 3 cells\n\t"
                                    "mulps    %%xmm2,%%xmm0           # left  = sample_left  * pan_left\n\t"
                                    "mulps    %%xmm3,%%xmm1           # right = sample_right * pan_right\n\t"
                                    "mulps    %%xmm4,%%xmm0           # left  = vca * (sample_left  * pan_left)\n\t"
                                    "mulps    %%xmm4,%%xmm1           # right = vca * (sample_right * pan_right)\n\t"
                                    : /* no output */
                                    : "r" (&Volume[ii]), /* %0 */
                                      "r" (PanL),   /* %1 */
                                      "r" (PanR)    /* %2 */
                                );
                                break;
                            }
                        }
                        /* mix the 4 samples to the output channels */
                        __asm__ __volatile__ (
                            "addps  (%0),%%xmm0       # mix calculated sample(s) to output left\n\t"
                            "movaps %%xmm0,(%0)       # output to left channel\n\t"
                            "addps  (%1),%%xmm1       # mix calculated sample(s) to output right\n\t"
                            "movaps %%xmm1,(%1)       # output to right channel\n\t"
                            : /* no output */
                            : "r" (&pOutL[ii]), /* %0 - must be 16 byte aligned ! */
                              "r" (&pOutR[ii])  /* %1 - must be 16 byte aligned ! */
                        );
                    }
                    #endif // ARCH_X86
                }
            }
    };

}} // namespace LinuxSampler::gig

#endif // __LS_GIG_SYNTHESIZER_H__
1	schoenebeck	320	/***************************************************************************
2			* *
3			* LinuxSampler - modular, streaming capable sampler *
4			* *
5			* Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck *
6	schoenebeck	411	* Copyright (C) 2005 Christian Schoenebeck *
7	schoenebeck	320	* *
8			* This program is free software; you can redistribute it and/or modify *
9			* it under the terms of the GNU General Public License as published by *
10			* the Free Software Foundation; either version 2 of the License, or *
11			* (at your option) any later version. *
12			* *
13			* This program is distributed in the hope that it will be useful, *
14			* but WITHOUT ANY WARRANTY; without even the implied warranty of *
15			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the *
16			* GNU General Public License for more details. *
17			* *
18			* You should have received a copy of the GNU General Public License *
19			* along with this program; if not, write to the Free Software *
20			* Foundation, Inc., 59 Temple Place, Suite 330, Boston, *
21			* MA 02111-1307 USA *
22			***************************************************************************/
23
24			#ifndef __LS_GIG_SYNTHESIZER_H__
25			#define __LS_GIG_SYNTHESIZER_H__
26
27	schoenebeck	328	#include "../../common/global.h"
28	schoenebeck	320	#include "../../common/RTMath.h"
29			#include "../common/Resampler.h"
30			#include "../common/BiquadFilter.h"
31			#include "Filter.h"
32			#include "Voice.h"
33
34	senkov	332	#define SYNTHESIS_MODE_SET_CONSTPITCH(iMode,bVal) if (bVal) iMode \|= 0x01; else iMode &= ~0x01 /* (un)set mode bit 0 */
35			#define SYNTHESIS_MODE_SET_LOOP(iMode,bVal) if (bVal) iMode \|= 0x02; else iMode &= ~0x02 /* (un)set mode bit 1 */
36			#define SYNTHESIS_MODE_SET_INTERPOLATE(iMode,bVal) if (bVal) iMode \|= 0x04; else iMode &= ~0x04 /* (un)set mode bit 2 */
37			#define SYNTHESIS_MODE_SET_FILTER(iMode,bVal) if (bVal) iMode \|= 0x08; else iMode &= ~0x08 /* (un)set mode bit 3 */
38			#define SYNTHESIS_MODE_SET_CHANNELS(iMode,bVal) if (bVal) iMode \|= 0x10; else iMode &= ~0x10 /* (un)set mode bit 4 */
39			#define SYNTHESIS_MODE_SET_IMPLEMENTATION(iMode,bVal) if (bVal) iMode \|= 0x20; else iMode &= ~0x20 /* (un)set mode bit 5 */
40	senkov	325	#define SYNTHESIS_MODE_SET_PROFILING(iMode,bVal) if (bVal) iMode \|= 0x40; else iMode &= ~0x40 /* (un)set mode bit 6 */
41	schoenebeck	320
42			#define SYNTHESIS_MODE_GET_CONSTPITCH(iMode) iMode & 0x01
43			#define SYNTHESIS_MODE_GET_LOOP(iMode) iMode & 0x02
44			#define SYNTHESIS_MODE_GET_INTERPOLATE(iMode) iMode & 0x04
45			#define SYNTHESIS_MODE_GET_FILTER(iMode) iMode & 0x08
46			#define SYNTHESIS_MODE_GET_CHANNELS(iMode) iMode & 0x10
47			#define SYNTHESIS_MODE_GET_IMPLEMENTATION(iMode) iMode & 0x20
48
49			// that's usually gig::Voice of course, but we make it a macro so we can
50			// include this code for our synthesis benchmark which uses fake data
51			// structures
52			#ifndef VOICE
53			# define VOICE Voice
54			#endif // VOICE
55
56			namespace LinuxSampler { namespace gig {
57
58	senkov	325	typedef void SynthesizeFragment_Fn(VOICE&, uint, sample_t*, uint);
59	schoenebeck	320
60			void* GetSynthesisFunction(const int SynthesisMode);
61	senkov	325	void RunSynthesisFunction(const int SynthesisMode, VOICE& voice, uint Samples, sample_t* pSrc, uint Skip);
62	schoenebeck	320
63			enum channels_t {
64			MONO,
65			STEREO
66			};
67
68			template<implementation_t IMPLEMENTATION, channels_t CHANNELS, bool USEFILTER, bool INTERPOLATE, bool DOLOOP, bool CONSTPITCH>
69			class Synthesizer : public __RTMath<IMPLEMENTATION>, public LinuxSampler::Resampler<INTERPOLATE> {
70	persson	497
71			// declarations of derived functions (see "Name lookup,
72			// templates, and accessing members of base classes" in
73			// the gcc manual for an explanation of why this is
74			// needed).
75			using __RTMath<IMPLEMENTATION>::Mul;
76			using __RTMath<IMPLEMENTATION>::Float;
77			using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleMonoCPP;
78			using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleStereoCPP;
79			#if ARCH_X86
80			using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesMonoMMXSSE;
81			using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesStereoMMXSSE;
82			#endif
83
84	schoenebeck	320	public:
85			template<typename VOICE_T>
86	senkov	325	inline static void SynthesizeFragment(VOICE_T& Voice, uint Samples, sample_t* pSrc, uint i) {
87	schoenebeck	424	const float panLeft = Mul(Voice.PanLeft, Voice.pEngineChannel->GlobalPanLeft);
88			const float panRight = Mul(Voice.PanRight, Voice.pEngineChannel->GlobalPanRight);
89	schoenebeck	320	if (IMPLEMENTATION == ASM_X86_MMX_SSE) {
90			float fPos = (float) Voice.Pos;
91			SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount,
92			Voice.pSample->LoopStart,
93			Voice.pSample->LoopEnd,
94			Voice.pSample->LoopSize,
95			Voice.LoopCyclesLeft,
96			(void *)&fPos,
97			Voice.PitchBase,
98	schoenebeck	424	Voice.PitchBend,
99			&panLeft, &panRight);
100	schoenebeck	361	#if ARCH_X86
101	schoenebeck	320	if (INTERPOLATE) EMMS;
102	schoenebeck	361	#endif
103	schoenebeck	320	Voice.Pos = (double) fPos;
104			} else {
105			SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount,
106			Voice.pSample->LoopStart,
107			Voice.pSample->LoopEnd,
108			Voice.pSample->LoopSize,
109			Voice.LoopCyclesLeft,
110			(void *)&Voice.Pos,
111			Voice.PitchBase,
112	schoenebeck	424	Voice.PitchBend,
113			&panLeft, &panRight);
114	schoenebeck	320	}
115			}
116
117			//protected:
118
119			template<typename VOICE_T>
120	schoenebeck	424	inline static void SynthesizeFragment(VOICE_T& Voice, uint Samples, sample_t* pSrc, uint& i, uint& LoopPlayCount, uint LoopStart, uint LoopEnd, uint LoopSize, uint& LoopCyclesLeft, void* Pos, float& PitchBase, float& PitchBend, const float* PanLeft, const float* PanRight) {
121	schoenebeck	320	const float loopEnd = Float(LoopEnd);
122			const float PBbyPB = Mul(PitchBase, PitchBend);
123			const float f_LoopStart = Float(LoopStart);
124			const float f_LoopSize = Float(LoopSize);
125			if (DOLOOP) {
126			if (LoopPlayCount) {
127			// render loop (loop count limited)
128			while (i < Samples && LoopCyclesLeft) {
129			if (CONSTPITCH) {
130	senkov	325	const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd,Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up
131	schoenebeck	424	while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
132	schoenebeck	320	}
133	schoenebeck	424	else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
134	schoenebeck	320	if (WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos)) LoopCyclesLeft--;
135			}
136			// render on without loop
137	schoenebeck	424	while (i < Samples) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
138	schoenebeck	320	}
139			else { // render loop (endless loop)
140			while (i < Samples) {
141			if (CONSTPITCH) {
142	senkov	325	const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd, Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up
143	schoenebeck	424	while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
144	schoenebeck	320	}
145	schoenebeck	424	else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);
146	schoenebeck	320	WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos);
147			}
148			}
149			}
150			else { // no looping
151	schoenebeck	424	while (i < Samples) { Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);}
152	schoenebeck	320	}
153			}
154
155			template<typename VOICE_T>
156	schoenebeck	424	inline static void Synthesize(VOICE_T& Voice, void* Pos, sample_t* pSrc, uint& i, const float* PanLeft, const float* PanRight) {
157	schoenebeck	320	Synthesize(pSrc, Pos,
158			Voice.pEngine->pSynthesisParameters[Event::destination_vco][i],
159	schoenebeck	411	Voice.pEngineChannel->pOutputLeft,
160			Voice.pEngineChannel->pOutputRight,
161	schoenebeck	320	i,
162			Voice.pEngine->pSynthesisParameters[Event::destination_vca],
163	schoenebeck	424	PanLeft,
164			PanRight,
165	schoenebeck	320	Voice.FilterLeft,
166			Voice.FilterRight,
167			Voice.pEngine->pBasicFilterParameters[i],
168			Voice.pEngine->pMainFilterParameters[i]);
169			}
170
171			inline static int DiffToLoopEnd(const float& LoopEnd, const void* Pos, const float& Pitch) {
172			switch (IMPLEMENTATION) {
173			// pure C++ implementation (thus platform independent)
174			case CPP: {
175	senkov	325	return uint((LoopEnd - ((double )Pos)) / Pitch);
176	schoenebeck	320	}
177	schoenebeck	328	#if ARCH_X86
178	schoenebeck	320	case ASM_X86_MMX_SSE: {
179			int result;
180			__asm__ __volatile__ (
181			"movss (%1), %%xmm0 #read loopend\n\t"
182			"subss (%2), %%xmm0 #sub pos\n\t"
183			"divss (%3), %%xmm0 #div by pitch\n\t"
184			"cvtss2si %%xmm0, %0 #convert to int\n\t"
185			: "=r" (result) /* %0 */
186			: "r" (&LoopEnd), /* %1 */
187			"r" (Pos), /* %2 */
188			"r" (&Pitch) /* %3 */
189			);
190			return result;
191			}
192	schoenebeck	328	#endif // ARCH_X86
193	schoenebeck	320	}
194			}
195
196			inline static int WrapLoop(const float& LoopStart, const float& LoopSize, const float& LoopEnd, void* vPos) {
197			switch (IMPLEMENTATION) {
198			// pure C++ implementation (thus platform independent)
199			case CPP: {
200			double * Pos = (double *)vPos;
201			if (*Pos < LoopEnd) return 0;
202			Pos = fmod(Pos - LoopEnd, LoopSize) + LoopStart;
203			return 1;
204			}
205	schoenebeck	328	#if ARCH_X86
206	schoenebeck	320	case ASM_X86_MMX_SSE: {
207	senkov	336	int result = 0;
208	schoenebeck	320	__asm__ __volatile__ (
209			"movss (%2), %%xmm0 # load LoopEnd\n\t"
210			"movss (%1), %%xmm1 # load Pos\n\t"
211			"comiss %%xmm0, %%xmm1 # LoopEnd <> Pos\n\t"
212			"jb 1f # jump if no work needs to be done\n\t"
213			"movss (%3), %%xmm2 # load LoopSize\n\t"
214			"subss %%xmm0, %%xmm1 # Pos - LoopEnd\n\t"
215			//now the fmodf
216			"movss %%xmm1, %%xmm3 # xmm3 = (Pos - LoopEnd)\n\t"
217			"divss %%xmm2, %%xmm1 # (Pos - LoopEnd) / LoopSize\n\t"
218			"cvttss2si %%xmm1, %%eax # convert to int\n\t"
219			"cvtsi2ss %%eax, %%xmm1 # convert back to float\n\t"
220			"movss (%4), %%xmm0 # load LoopStart\n\t"
221			"mulss %%xmm2, %%xmm1 # LoopSize * int((Pos-LoopEnd)/LoopSize)\n\t"
222			"subss %%xmm1, %%xmm3 # xmm2 = fmodf(Pos - LoopEnd, LoopSize)\n\t"
223			//done with fmodf
224			"addss %%xmm0, %%xmm3 # add LoopStart\n\t"
225			"movss %%xmm3, (%1) # update Pos\n\t"
226	senkov	336	"movl $1, (%0) # result = 1\n\t"
227	schoenebeck	320	".balign 16 \n\t"
228			"1:\n\t"
229	senkov	336	:: "r" (&result), /* %0 */
230			"r" (vPos), /* %1 */
231	schoenebeck	320	"r" (&LoopEnd), /* %2 */
232			"r" (&LoopSize), /* %3 */
233			"r" (&LoopStart) /* %4 */
234			);
235			return result;
236			}
237	schoenebeck	328	#endif // ARCH_X86
238	schoenebeck	320	}
239			}
240
241	schoenebeck	424	inline static void Synthesize(sample_t* pSrc, void* Pos, float& Pitch, float* pOutL, float* pOutR, uint& i, float* Volume, const float* PanL, const float* PanR, Filter& FilterL, Filter& FilterR, biquad_param_t& bqBase, biquad_param_t& bqMain) {
242	schoenebeck	320	switch (IMPLEMENTATION) {
243			// pure C++ implementation (thus platform independent)
244			case CPP: {
245			switch (CHANNELS) {
246			case MONO: {
247			float samplePoint = GetNextSampleMonoCPP(pSrc, (double *)Pos, Pitch);
248			if (USEFILTER) samplePoint = FilterL.Apply(&bqBase, &bqMain, samplePoint);
249			pOutL[i] += samplePoint * Volume[i] * *PanL;
250			pOutR[i] += samplePoint * Volume[i] * *PanR;
251			i++;
252			break;
253			}
254			case STEREO: {
255			stereo_sample_t samplePoint = GetNextSampleStereoCPP(pSrc, (double *)Pos, Pitch);
256			if (USEFILTER) {
257			samplePoint.left = FilterL.Apply(&bqBase, &bqMain, samplePoint.left);
258			samplePoint.right = FilterR.Apply(&bqBase, &bqMain, samplePoint.right);
259			}
260			pOutL[i] += samplePoint.left * Volume[i] * *PanL;
261			pOutR[i] += samplePoint.right * Volume[i] * *PanR;
262			i++;
263			break;
264			}
265			}
266			break;
267			}
268	schoenebeck	328	#if ARCH_X86
269	schoenebeck	320	// Assembly optimization using the MMX & SSE(1) instruction set (thus only for x86)
270			case ASM_X86_MMX_SSE: {
271			const int ii = i & 0xfffffffc;
272			i += 4;
273			switch (CHANNELS) {
274			case MONO: {
275			GetNext4SamplesMonoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2
276			if (USEFILTER) {
277			/* prepare filter input */
278			__asm__ __volatile__ (
279			"movaps %xmm2,%xmm0"
280			);
281			FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
282			__asm__ __volatile__ (
283			"movaps %xmm7,%xmm2 # mono filter result -> xmm2"
284			);
285			}
286			/* apply panorama and volume factors */
287			__asm__ __volatile__ (
288			"movss (%1),%%xmm0 # load pan left\n\t"
289			"movss (%2),%%xmm1 # load pan right\n\t"
290			"movaps (%0),%%xmm4 # load vca\n\t"
291			"shufps $0x00,%%xmm0,%%xmm0 # copy pan left to the other 3 cells\n\t"
292			"shufps $0x00,%%xmm1,%%xmm1 # copy pan right to the other 3 cells\n\t"
293			"mulps %%xmm2,%%xmm0 # left = sample * pan_left\n\t"
294			"mulps %%xmm2,%%xmm1 # right = sample * pan_right\n\t"
295			"mulps %%xmm4,%%xmm0 # left = vca * (sample * pan_left)\n\t"
296			"mulps %%xmm4,%%xmm1 # right = vca * (sample * pan_right)\n\t"
297			: /* no output */
298			: "r" (&Volume[ii]), /* %0 */
299			"r" (PanL), /* %1 */
300			"r" (PanR) /* %2 */
301			: "xmm0", /* holds final left sample (for the 4 samples) at the end */
302			"xmm1" /* holds final right sample (for the 4 samples) at the end */
303			);
304			break;
305			}
306			case STEREO: {
307			GetNext4SamplesStereoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2 (left channel) and xmm3 (right channel)
308			if (USEFILTER) {
309			__asm__ __volatile__ (
310			"movaps %xmm2,%xmm0 # prepare left channel for filter\n\t"
311			"movaps %xmm3,%xmm1 # save right channel not to get overwritten by filter algorithms\n\t"
312			);
313			FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
314			__asm__ __volatile__ (
315			"movaps %xmm1,%xmm0 # prepare right channel for filter\n\t"
316			"movaps %xmm7,%xmm1 # save filter output for left channel\n\t"
317			);
318			FilterR.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output
319			__asm__ __volatile__ (
320			"movaps %xmm1,%xmm2 # result left channel -> xmm2\n\t"
321			"movaps %xmm7,%xmm3 # result right channel -> xmm3\n\t"
322			);
323			}
324			/* apply panorama and volume factors */
325			__asm__ __volatile__ (
326			"movss (%1),%%xmm0 # load pan left\n\t"
327			"movss (%2),%%xmm1 # load pan right\n\t"
328			"movaps (%0),%%xmm4 # load vca\n\t"
329			"shufps $0x00,%%xmm0,%%xmm0 # copy pan left to the other 3 cells\n\t"
330			"shufps $0x00,%%xmm1,%%xmm1 # copy pan right to the other 3 cells\n\t"
331			"mulps %%xmm2,%%xmm0 # left = sample_left * pan_left\n\t"
332			"mulps %%xmm3,%%xmm1 # right = sample_right * pan_right\n\t"
333			"mulps %%xmm4,%%xmm0 # left = vca * (sample_left * pan_left)\n\t"
334			"mulps %%xmm4,%%xmm1 # right = vca * (sample_right * pan_right)\n\t"
335			: /* no output */
336			: "r" (&Volume[ii]), /* %0 */
337			"r" (PanL), /* %1 */
338			"r" (PanR) /* %2 */
339			);
340			break;
341			}
342			}
343			/* mix the 4 samples to the output channels */
344			__asm__ __volatile__ (
345			"addps (%0),%%xmm0 # mix calculated sample(s) to output left\n\t"
346			"movaps %%xmm0,(%0) # output to left channel\n\t"
347			"addps (%1),%%xmm1 # mix calculated sample(s) to output right\n\t"
348			"movaps %%xmm1,(%1) # output to right channel\n\t"
349			: /* no output */
350			: "r" (&pOutL[ii]), /* %0 - must be 16 byte aligned ! */
351			"r" (&pOutR[ii]) /* %1 - must be 16 byte aligned ! */
352			);
353			}
354	schoenebeck	328	#endif // ARCH_X86
355	schoenebeck	320	}
356			}
357			};
358
359			}} // namespace LinuxSampler::gig
360
361			#endif // __LS_GIG_SYNTHESIZER_H__