1 |
/*************************************************************************** |
2 |
* * |
3 |
* LinuxSampler - modular, streaming capable sampler * |
4 |
* * |
5 |
* Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck * |
6 |
* Copyright (C) 2005 Christian Schoenebeck * |
7 |
* * |
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 |
#include "../../common/global.h" |
28 |
#include "../../common/RTMath.h" |
29 |
#include "../common/Resampler.h" |
30 |
#include "../common/BiquadFilter.h" |
31 |
#include "Filter.h" |
32 |
#include "Voice.h" |
33 |
|
34 |
#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 |
#define SYNTHESIS_MODE_SET_PROFILING(iMode,bVal) if (bVal) iMode |= 0x40; else iMode &= ~0x40 /* (un)set mode bit 6 */ |
41 |
|
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 |
typedef void SynthesizeFragment_Fn(VOICE&, uint, sample_t*, uint); |
59 |
|
60 |
void* GetSynthesisFunction(const int SynthesisMode); |
61 |
void RunSynthesisFunction(const int SynthesisMode, VOICE& voice, uint Samples, sample_t* pSrc, uint Skip); |
62 |
|
63 |
enum channels_t { |
64 |
MONO, |
65 |
STEREO |
66 |
}; |
67 |
|
68 |
/** @brief Main Synthesis algorithms for the gig::Engine |
69 |
* |
70 |
* Implementation of the main synthesis algorithms of the Gigasampler |
71 |
* format capable sampler engine. This means resampling / interpolation |
72 |
* for pitching the audio signal, looping, filter and amplification. |
73 |
*/ |
74 |
template<implementation_t IMPLEMENTATION, channels_t CHANNELS, bool USEFILTER, bool INTERPOLATE, bool DOLOOP, bool CONSTPITCH> |
75 |
class Synthesizer : public __RTMath<IMPLEMENTATION>, public LinuxSampler::Resampler<INTERPOLATE> { |
76 |
|
77 |
// declarations of derived functions (see "Name lookup, |
78 |
// templates, and accessing members of base classes" in |
79 |
// the gcc manual for an explanation of why this is |
80 |
// needed). |
81 |
using __RTMath<IMPLEMENTATION>::Mul; |
82 |
using __RTMath<IMPLEMENTATION>::Float; |
83 |
using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleMonoCPP; |
84 |
using LinuxSampler::Resampler<INTERPOLATE>::GetNextSampleStereoCPP; |
85 |
#if CONFIG_ASM && ARCH_X86 |
86 |
using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesMonoMMXSSE; |
87 |
using LinuxSampler::Resampler<INTERPOLATE>::GetNext4SamplesStereoMMXSSE; |
88 |
#endif |
89 |
|
90 |
public: |
91 |
/** |
92 |
* Render audio for the current fragment for the given voice. |
93 |
* This is the toplevel method of this class. |
94 |
*/ |
95 |
template<typename VOICE_T> |
96 |
inline static void SynthesizeFragment(VOICE_T& Voice, uint Samples, sample_t* pSrc, uint i) { |
97 |
const float panLeft = Mul(Voice.PanLeft, Voice.pEngineChannel->GlobalPanLeft); |
98 |
const float panRight = Mul(Voice.PanRight, Voice.pEngineChannel->GlobalPanRight); |
99 |
if (IMPLEMENTATION == ASM_X86_MMX_SSE) { |
100 |
float fPos = (float) Voice.Pos; |
101 |
SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount, |
102 |
Voice.pSample->LoopStart, |
103 |
Voice.pSample->LoopEnd, |
104 |
Voice.pSample->LoopSize, |
105 |
Voice.LoopCyclesLeft, |
106 |
(void *)&fPos, |
107 |
Voice.PitchBase, |
108 |
Voice.PitchBend, |
109 |
&panLeft, &panRight); |
110 |
#if CONFIG_ASM && ARCH_X86 |
111 |
if (INTERPOLATE) EMMS; |
112 |
#endif |
113 |
Voice.Pos = (double) fPos; |
114 |
} else { |
115 |
SynthesizeFragment(Voice, Samples, pSrc, i, Voice.pSample->LoopPlayCount, |
116 |
Voice.pSample->LoopStart, |
117 |
Voice.pSample->LoopEnd, |
118 |
Voice.pSample->LoopSize, |
119 |
Voice.LoopCyclesLeft, |
120 |
(void *)&Voice.Pos, |
121 |
Voice.PitchBase, |
122 |
Voice.PitchBend, |
123 |
&panLeft, &panRight); |
124 |
} |
125 |
} |
126 |
|
127 |
//protected: |
128 |
|
129 |
/** |
130 |
* Render audio for the current fragment for the given voice. |
131 |
* Will be called by the toplevel SynthesizeFragment() method. |
132 |
*/ |
133 |
template<typename VOICE_T> |
134 |
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) { |
135 |
const float loopEnd = Float(LoopEnd); |
136 |
const float PBbyPB = Mul(PitchBase, PitchBend); |
137 |
const float f_LoopStart = Float(LoopStart); |
138 |
const float f_LoopSize = Float(LoopSize); |
139 |
if (DOLOOP) { |
140 |
if (LoopPlayCount) { |
141 |
// render loop (loop count limited) |
142 |
while (i < Samples && LoopCyclesLeft) { |
143 |
if (CONSTPITCH) { |
144 |
const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd,Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up |
145 |
while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight); |
146 |
} |
147 |
else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight); |
148 |
if (WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos)) LoopCyclesLeft--; |
149 |
} |
150 |
// render on without loop |
151 |
while (i < Samples) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight); |
152 |
} |
153 |
else { // render loop (endless loop) |
154 |
while (i < Samples) { |
155 |
if (CONSTPITCH) { |
156 |
const uint processEnd = Min(Samples, i + DiffToLoopEnd(loopEnd, Pos, PBbyPB) + 1); //TODO: instead of +1 we could also round up |
157 |
while (i < processEnd) Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight); |
158 |
} |
159 |
else Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight); |
160 |
WrapLoop(f_LoopStart, f_LoopSize, loopEnd, Pos); |
161 |
} |
162 |
} |
163 |
} |
164 |
else { // no looping |
165 |
while (i < Samples) { Synthesize(Voice, Pos, pSrc, i, PanLeft, PanRight);} |
166 |
} |
167 |
} |
168 |
|
169 |
/** |
170 |
* Atomicly render a piece for the voice. For the C++ |
171 |
* implementation this means rendering exactly one sample |
172 |
* point, whereas for the MMX/SSE implementation this means |
173 |
* rendering 4 sample points. |
174 |
*/ |
175 |
template<typename VOICE_T> |
176 |
inline static void Synthesize(VOICE_T& Voice, void* Pos, sample_t* pSrc, uint& i, const float* PanLeft, const float* PanRight) { |
177 |
Synthesize(pSrc, Pos, |
178 |
Voice.pEngine->pSynthesisParameters[Event::destination_vco][i], |
179 |
Voice.pEngineChannel->pOutputLeft, |
180 |
Voice.pEngineChannel->pOutputRight, |
181 |
i, |
182 |
Voice.pEngine->pSynthesisParameters[Event::destination_vca], |
183 |
PanLeft, |
184 |
PanRight, |
185 |
Voice.FilterLeft, |
186 |
Voice.FilterRight, |
187 |
Voice.pEngine->pBasicFilterParameters[i], |
188 |
Voice.pEngine->pMainFilterParameters[i]); |
189 |
} |
190 |
|
191 |
/** |
192 |
* Returns the difference to the sample's loop end. |
193 |
*/ |
194 |
inline static int DiffToLoopEnd(const float& LoopEnd, const void* Pos, const float& Pitch) { |
195 |
switch (IMPLEMENTATION) { |
196 |
// pure C++ implementation (thus platform independent) |
197 |
case CPP: { |
198 |
return uint((LoopEnd - *((double *)Pos)) / Pitch); |
199 |
} |
200 |
#if CONFIG_ASM && ARCH_X86 |
201 |
case ASM_X86_MMX_SSE: { |
202 |
int result; |
203 |
__asm__ __volatile__ ( |
204 |
"movss (%1), %%xmm0 #read loopend\n\t" |
205 |
"subss (%2), %%xmm0 #sub pos\n\t" |
206 |
"divss (%3), %%xmm0 #div by pitch\n\t" |
207 |
"cvtss2si %%xmm0, %0 #convert to int\n\t" |
208 |
: "=r" (result) /* %0 */ |
209 |
: "r" (&LoopEnd), /* %1 */ |
210 |
"r" (Pos), /* %2 */ |
211 |
"r" (&Pitch) /* %3 */ |
212 |
); |
213 |
return result; |
214 |
} |
215 |
#endif // CONFIG_ASM && ARCH_X86 |
216 |
} |
217 |
} |
218 |
|
219 |
/** |
220 |
* This method handles looping of the RAM playback part of the |
221 |
* sample, thus repositioning the playback position once the |
222 |
* loop limit was reached. Note: looping of the disk streaming |
223 |
* part is handled by libgig (ReadAndLoop() method which will |
224 |
* be called by the DiskThread). |
225 |
*/ |
226 |
inline static int WrapLoop(const float& LoopStart, const float& LoopSize, const float& LoopEnd, void* vPos) { |
227 |
switch (IMPLEMENTATION) { |
228 |
// pure C++ implementation (thus platform independent) |
229 |
case CPP: { |
230 |
double * Pos = (double *)vPos; |
231 |
if (*Pos < LoopEnd) return 0; |
232 |
*Pos = fmod(*Pos - LoopEnd, LoopSize) + LoopStart; |
233 |
return 1; |
234 |
} |
235 |
#if CONFIG_ASM && ARCH_X86 |
236 |
case ASM_X86_MMX_SSE: { |
237 |
int result = 0; |
238 |
__asm__ __volatile__ ( |
239 |
"movss (%2), %%xmm0 # load LoopEnd\n\t" |
240 |
"movss (%1), %%xmm1 # load Pos\n\t" |
241 |
"comiss %%xmm0, %%xmm1 # LoopEnd <> Pos\n\t" |
242 |
"jb 1f # jump if no work needs to be done\n\t" |
243 |
"movss (%3), %%xmm2 # load LoopSize\n\t" |
244 |
"subss %%xmm0, %%xmm1 # Pos - LoopEnd\n\t" |
245 |
//now the fmodf |
246 |
"movss %%xmm1, %%xmm3 # xmm3 = (Pos - LoopEnd)\n\t" |
247 |
"divss %%xmm2, %%xmm1 # (Pos - LoopEnd) / LoopSize\n\t" |
248 |
"cvttss2si %%xmm1, %2 # convert to int\n\t" |
249 |
"cvtsi2ss %2, %%xmm1 # convert back to float\n\t" |
250 |
"movss (%4), %%xmm0 # load LoopStart\n\t" |
251 |
"mulss %%xmm2, %%xmm1 # LoopSize * int((Pos-LoopEnd)/LoopSize)\n\t" |
252 |
"subss %%xmm1, %%xmm3 # xmm2 = fmodf(Pos - LoopEnd, LoopSize)\n\t" |
253 |
//done with fmodf |
254 |
"addss %%xmm0, %%xmm3 # add LoopStart\n\t" |
255 |
"movss %%xmm3, (%1) # update Pos\n\t" |
256 |
"movl $1, (%0) # result = 1\n\t" |
257 |
".balign 16 \n\t" |
258 |
"1:\n\t" |
259 |
:: "r" (&result), /* %0 */ |
260 |
"r" (vPos), /* %1 */ |
261 |
"r" (&LoopEnd), /* %2 */ |
262 |
"r" (&LoopSize), /* %3 */ |
263 |
"r" (&LoopStart) /* %4 */ |
264 |
); |
265 |
return result; |
266 |
} |
267 |
#endif // CONFIG_ASM && ARCH_X86 |
268 |
} |
269 |
} |
270 |
|
271 |
/** |
272 |
* Atomicly render a piece for the voice. For the C++ |
273 |
* implementation this means rendering exactly one sample |
274 |
* point, whereas for the MMX/SSE implementation this means |
275 |
* rendering 4 sample points. |
276 |
*/ |
277 |
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) { |
278 |
switch (IMPLEMENTATION) { |
279 |
// pure C++ implementation (thus platform independent) |
280 |
case CPP: { |
281 |
switch (CHANNELS) { |
282 |
case MONO: { |
283 |
float samplePoint = GetNextSampleMonoCPP(pSrc, (double *)Pos, Pitch); |
284 |
if (USEFILTER) samplePoint = FilterL.Apply(&bqBase, &bqMain, samplePoint); |
285 |
pOutL[i] += samplePoint * Volume[i] * *PanL; |
286 |
pOutR[i] += samplePoint * Volume[i] * *PanR; |
287 |
i++; |
288 |
break; |
289 |
} |
290 |
case STEREO: { |
291 |
stereo_sample_t samplePoint = GetNextSampleStereoCPP(pSrc, (double *)Pos, Pitch); |
292 |
if (USEFILTER) { |
293 |
samplePoint.left = FilterL.Apply(&bqBase, &bqMain, samplePoint.left); |
294 |
samplePoint.right = FilterR.Apply(&bqBase, &bqMain, samplePoint.right); |
295 |
} |
296 |
pOutL[i] += samplePoint.left * Volume[i] * *PanL; |
297 |
pOutR[i] += samplePoint.right * Volume[i] * *PanR; |
298 |
i++; |
299 |
break; |
300 |
} |
301 |
} |
302 |
break; |
303 |
} |
304 |
#if CONFIG_ASM && ARCH_X86 |
305 |
// Assembly optimization using the MMX & SSE(1) instruction set (thus only for x86) |
306 |
case ASM_X86_MMX_SSE: { |
307 |
const int ii = i & 0xfffffffc; |
308 |
i += 4; |
309 |
switch (CHANNELS) { |
310 |
case MONO: { |
311 |
GetNext4SamplesMonoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2 |
312 |
if (USEFILTER) { |
313 |
/* prepare filter input */ |
314 |
__asm__ __volatile__ ( |
315 |
"movaps %xmm2,%xmm0" |
316 |
); |
317 |
FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output |
318 |
__asm__ __volatile__ ( |
319 |
"movaps %xmm7,%xmm2 # mono filter result -> xmm2" |
320 |
); |
321 |
} |
322 |
/* apply panorama and volume factors */ |
323 |
__asm__ __volatile__ ( |
324 |
"movss (%1),%%xmm0 # load pan left\n\t" |
325 |
"movss (%2),%%xmm1 # load pan right\n\t" |
326 |
"movaps (%0),%%xmm4 # load vca\n\t" |
327 |
"shufps $0x00,%%xmm0,%%xmm0 # copy pan left to the other 3 cells\n\t" |
328 |
"shufps $0x00,%%xmm1,%%xmm1 # copy pan right to the other 3 cells\n\t" |
329 |
"mulps %%xmm2,%%xmm0 # left = sample * pan_left\n\t" |
330 |
"mulps %%xmm2,%%xmm1 # right = sample * pan_right\n\t" |
331 |
"mulps %%xmm4,%%xmm0 # left = vca * (sample * pan_left)\n\t" |
332 |
"mulps %%xmm4,%%xmm1 # right = vca * (sample * pan_right)\n\t" |
333 |
: /* no output */ |
334 |
: "r" (&Volume[ii]), /* %0 */ |
335 |
"r" (PanL), /* %1 */ |
336 |
"r" (PanR) /* %2 */ |
337 |
: "xmm0", /* holds final left sample (for the 4 samples) at the end */ |
338 |
"xmm1" /* holds final right sample (for the 4 samples) at the end */ |
339 |
); |
340 |
break; |
341 |
} |
342 |
case STEREO: { |
343 |
GetNext4SamplesStereoMMXSSE(pSrc, (float *)Pos, Pitch); // outputs samples in xmm2 (left channel) and xmm3 (right channel) |
344 |
if (USEFILTER) { |
345 |
__asm__ __volatile__ ( |
346 |
"movaps %xmm2,%xmm0 # prepare left channel for filter\n\t" |
347 |
"movaps %xmm3,%xmm1 # save right channel not to get overwritten by filter algorithms\n\t" |
348 |
); |
349 |
FilterL.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output |
350 |
__asm__ __volatile__ ( |
351 |
"movaps %xmm1,%xmm0 # prepare right channel for filter\n\t" |
352 |
"movaps %xmm7,%xmm1 # save filter output for left channel\n\t" |
353 |
); |
354 |
FilterR.Apply4StepsSSE(&bqBase, &bqMain); // xmm0 input, xmm7 output |
355 |
__asm__ __volatile__ ( |
356 |
"movaps %xmm1,%xmm2 # result left channel -> xmm2\n\t" |
357 |
"movaps %xmm7,%xmm3 # result right channel -> xmm3\n\t" |
358 |
); |
359 |
} |
360 |
/* apply panorama and volume factors */ |
361 |
__asm__ __volatile__ ( |
362 |
"movss (%1),%%xmm0 # load pan left\n\t" |
363 |
"movss (%2),%%xmm1 # load pan right\n\t" |
364 |
"movaps (%0),%%xmm4 # load vca\n\t" |
365 |
"shufps $0x00,%%xmm0,%%xmm0 # copy pan left to the other 3 cells\n\t" |
366 |
"shufps $0x00,%%xmm1,%%xmm1 # copy pan right to the other 3 cells\n\t" |
367 |
"mulps %%xmm2,%%xmm0 # left = sample_left * pan_left\n\t" |
368 |
"mulps %%xmm3,%%xmm1 # right = sample_right * pan_right\n\t" |
369 |
"mulps %%xmm4,%%xmm0 # left = vca * (sample_left * pan_left)\n\t" |
370 |
"mulps %%xmm4,%%xmm1 # right = vca * (sample_right * pan_right)\n\t" |
371 |
: /* no output */ |
372 |
: "r" (&Volume[ii]), /* %0 */ |
373 |
"r" (PanL), /* %1 */ |
374 |
"r" (PanR) /* %2 */ |
375 |
); |
376 |
break; |
377 |
} |
378 |
} |
379 |
/* mix the 4 samples to the output channels */ |
380 |
__asm__ __volatile__ ( |
381 |
"addps (%0),%%xmm0 # mix calculated sample(s) to output left\n\t" |
382 |
"movaps %%xmm0,(%0) # output to left channel\n\t" |
383 |
"addps (%1),%%xmm1 # mix calculated sample(s) to output right\n\t" |
384 |
"movaps %%xmm1,(%1) # output to right channel\n\t" |
385 |
: /* no output */ |
386 |
: "r" (&pOutL[ii]), /* %0 - must be 16 byte aligned ! */ |
387 |
"r" (&pOutR[ii]) /* %1 - must be 16 byte aligned ! */ |
388 |
); |
389 |
} |
390 |
#endif // CONFIG_ASM && ARCH_X86 |
391 |
} |
392 |
} |
393 |
}; |
394 |
|
395 |
}} // namespace LinuxSampler::gig |
396 |
|
397 |
#endif // __LS_GIG_SYNTHESIZER_H__ |