AudioResamplerSinc.cpp revision 54c3b66444ebfb9f2265ee70ac3b76ccefa0506a 
1/*
2 * Copyright (C) 2007 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include <string.h>
18#include "AudioResamplerSinc.h"
19
20namespace android {
21// ----------------------------------------------------------------------------
22
23
24/*
25 * These coeficients are computed with the "fir" utility found in
26 * tools/resampler_tools
27 * TODO: A good optimization would be to transpose this matrix, to take
28 * better advantage of the data-cache.
29 */
30const int32_t AudioResamplerSinc::mFirCoefsUp[] = {
31        0x7fffffff, 0x7f15d078, 0x7c5e0da6, 0x77ecd867, 0x71e2e251, 0x6a6c304a, 0x61be7269, 0x58170412, 0x4db8ab05, 0x42e92ea6, 0x37eee214, 0x2d0e3bb1, 0x22879366, 0x18951e95, 0x0f693d0d, 0x072d2621,
32        0x00000000, 0xf9f66655, 0xf51a5fd7, 0xf16bbd84, 0xeee0d9ac, 0xed67a922, 0xece70de6, 0xed405897, 0xee50e505, 0xeff3be30, 0xf203370f, 0xf45a6741, 0xf6d67d53, 0xf957db66, 0xfbc2f647, 0xfe00f2b9,
33        0x00000000, 0x01b37218, 0x0313a0c6, 0x041d930d, 0x04d28057, 0x053731b0, 0x05534dff, 0x05309bfd, 0x04da440d, 0x045c1aee, 0x03c1fcdd, 0x03173ef5, 0x02663ae8, 0x01b7f736, 0x0113ec79, 0x007fe6a9,
34        0x00000000, 0xff96b229, 0xff44f99f, 0xff0a86be, 0xfee5f803, 0xfed518fd, 0xfed521fd, 0xfee2f4fd, 0xfefb54f8, 0xff1b159b, 0xff3f4203, 0xff6539e0, 0xff8ac502, 0xffae1ddd, 0xffcdf3f9, 0xffe96798,
35        0x00000000, 0x00119de6, 0x001e6b7e, 0x0026cb7a, 0x002b4830, 0x002c83d6, 0x002b2a82, 0x0027e67a, 0x002356f9, 0x001e098e, 0x001875e4, 0x0012fbbe, 0x000de2d1, 0x00095c10, 0x00058414, 0x00026636,
36        0x00000000, 0xfffe44a9, 0xfffd206d, 0xfffc7b7f, 0xfffc3c8f, 0xfffc4ac2, 0xfffc8f2b, 0xfffcf5c4, 0xfffd6df3, 0xfffdeab2, 0xfffe6275, 0xfffececf, 0xffff2c07, 0xffff788c, 0xffffb471, 0xffffe0f2,
37        0x00000000, 0x000013e6, 0x00001f03, 0x00002396, 0x00002399, 0x000020b6, 0x00001c3c, 0x00001722, 0x00001216, 0x00000d81, 0x0000099c, 0x0000067c, 0x00000419, 0x0000025f, 0x00000131, 0x00000070,
38        0x00000000, 0xffffffc7, 0xffffffb3, 0xffffffb3, 0xffffffbe, 0xffffffcd, 0xffffffdb, 0xffffffe7, 0xfffffff0, 0xfffffff7, 0xfffffffb, 0xfffffffe, 0xffffffff, 0x00000000, 0x00000000, 0x00000000,
39        0x00000000 // this one is needed for lerping the last coefficient
40};
41
42/*
43 * These coefficients are optimized for 48KHz -> 44.1KHz (stop-band at 22.050KHz)
44 * It's possible to use the above coefficient for any down-sampling
45 * at the expense of a slower processing loop (we can interpolate
46 * these coefficient from the above by "Stretching" them in time).
47 */
48const int32_t AudioResamplerSinc::mFirCoefsDown[] = {
49        0x7fffffff, 0x7f55e46d, 0x7d5b4c60, 0x7a1b4b98, 0x75a7fb14, 0x7019f0bd, 0x698f875a, 0x622bfd59, 0x5a167256, 0x5178cc54, 0x487e8e6c, 0x3f53aae8, 0x36235ad4, 0x2d17047b, 0x245539ab, 0x1c00d540,
50        0x14383e57, 0x0d14d5ca, 0x06aa910b, 0x0107c38b, 0xfc351654, 0xf835abae, 0xf5076b45, 0xf2a37202, 0xf0fe9faa, 0xf00a3bbd, 0xefb4aa81, 0xefea2b05, 0xf0959716, 0xf1a11e83, 0xf2f6f7a0, 0xf481fff4,
51        0xf62e48ce, 0xf7e98ca5, 0xf9a38b4c, 0xfb4e4bfa, 0xfcde456f, 0xfe4a6d30, 0xff8c2fdf, 0x009f5555, 0x0181d393, 0x0233940f, 0x02b62f06, 0x030ca07d, 0x033afa62, 0x03461725, 0x03334f83, 0x030835fa,
52        0x02ca59cc, 0x027f12d1, 0x022b570d, 0x01d39a49, 0x017bb78f, 0x0126e414, 0x00d7aaaf, 0x008feec7, 0x0050f584, 0x001b73e3, 0xffefa063, 0xffcd46ed, 0xffb3ddcd, 0xffa29aaa, 0xff988691, 0xff949066,
53        0xff959d24, 0xff9a959e, 0xffa27195, 0xffac4011, 0xffb72d2b, 0xffc28569, 0xffcdb706, 0xffd85171, 0xffe20364, 0xffea97e9, 0xfff1f2b2, 0xfff80c06, 0xfffcec92, 0x0000a955, 0x00035fd8, 0x000532cf,
54        0x00064735, 0x0006c1f9, 0x0006c62d, 0x000673ba, 0x0005e68f, 0x00053630, 0x000475a3, 0x0003b397, 0x0002fac1, 0x00025257, 0x0001be9e, 0x0001417a, 0x0000dafd, 0x000089eb, 0x00004c28, 0x00001f1d,
55        0x00000000, 0xffffec10, 0xffffe0be, 0xffffdbc5, 0xffffdb39, 0xffffdd8b, 0xffffe182, 0xffffe638, 0xffffeb0a, 0xffffef8f, 0xfffff38b, 0xfffff6e3, 0xfffff993, 0xfffffba6, 0xfffffd30, 0xfffffe4a,
56        0xffffff09, 0xffffff85, 0xffffffd1, 0xfffffffb, 0x0000000f, 0x00000016, 0x00000015, 0x00000012, 0x0000000d, 0x00000009, 0x00000006, 0x00000003, 0x00000002, 0x00000001, 0x00000000, 0x00000000,
57        0x00000000 // this one is needed for lerping the last coefficient
58};
59
60// ----------------------------------------------------------------------------
61
62static inline
63int32_t mulRL(int left, int32_t in, uint32_t vRL)
64{
65#if defined(__arm__) && !defined(__thumb__)
66    int32_t out;
67    if (left) {
68        asm( "smultb %[out], %[in], %[vRL] \n"
69             : [out]"=r"(out)
70             : [in]"%r"(in), [vRL]"r"(vRL)
71             : );
72    } else {
73        asm( "smultt %[out], %[in], %[vRL] \n"
74             : [out]"=r"(out)
75             : [in]"%r"(in), [vRL]"r"(vRL)
76             : );
77    }
78    return out;
79#else
80    if (left) {
81        return int16_t(in>>16) * int16_t(vRL&0xFFFF);
82    } else {
83        return int16_t(in>>16) * int16_t(vRL>>16);
84    }
85#endif
86}
87
88static inline
89int32_t mulAdd(int16_t in, int32_t v, int32_t a)
90{
91#if defined(__arm__) && !defined(__thumb__)
92    int32_t out;
93    asm( "smlawb %[out], %[v], %[in], %[a] \n"
94         : [out]"=r"(out)
95         : [in]"%r"(in), [v]"r"(v), [a]"r"(a)
96         : );
97    return out;
98#else
99    return a + in * (v>>16);
100    // improved precision
101    // return a + in * (v>>16) + ((in * (v & 0xffff)) >> 16);
102#endif
103}
104
105static inline
106int32_t mulAddRL(int left, uint32_t inRL, int32_t v, int32_t a)
107{
108#if defined(__arm__) && !defined(__thumb__)
109    int32_t out;
110    if (left) {
111        asm( "smlawb %[out], %[v], %[inRL], %[a] \n"
112             : [out]"=r"(out)
113             : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
114             : );
115    } else {
116        asm( "smlawt %[out], %[v], %[inRL], %[a] \n"
117             : [out]"=r"(out)
118             : [inRL]"%r"(inRL), [v]"r"(v), [a]"r"(a)
119             : );
120    }
121    return out;
122#else
123    if (left) {
124        return a + (int16_t(inRL&0xFFFF) * (v>>16));
125        //improved precision
126        // return a + (int16_t(inRL&0xFFFF) * (v>>16)) + ((int16_t(inRL&0xFFFF) * (v & 0xffff)) >> 16);
127    } else {
128        return a + (int16_t(inRL>>16) * (v>>16));
129    }
130#endif
131}
132
133// ----------------------------------------------------------------------------
134
135AudioResamplerSinc::AudioResamplerSinc(int bitDepth,
136        int inChannelCount, int32_t sampleRate)
137    : AudioResampler(bitDepth, inChannelCount, sampleRate),
138    mState(0)
139{
140    /*
141     * Layout of the state buffer for 32 tap:
142     *
143     * "present" sample            beginning of 2nd buffer
144     *                 v                v
145     *  0              01               2              23              3
146     *  0              F0               0              F0              F
147     * [pppppppppppppppInnnnnnnnnnnnnnnnpppppppppppppppInnnnnnnnnnnnnnnn]
148     *                 ^               ^ head
149     *
150     * p = past samples, convoluted with the (p)ositive side of sinc()
151     * n = future samples, convoluted with the (n)egative side of sinc()
152     * r = extra space for implementing the ring buffer
153     *
154     */
155
156    const size_t numCoefs = 2*halfNumCoefs;
157    const size_t stateSize = numCoefs * inChannelCount * 2;
158    mState = new int16_t[stateSize];
159    memset(mState, 0, sizeof(int16_t)*stateSize);
160    mImpulse = mState + (halfNumCoefs-1)*inChannelCount;
161    mRingFull = mImpulse + (numCoefs+1)*inChannelCount;
162}
163
164AudioResamplerSinc::~AudioResamplerSinc()
165{
166    delete [] mState;
167}
168
169void AudioResamplerSinc::init() {
170}
171
172void AudioResamplerSinc::resample(int32_t* out, size_t outFrameCount,
173            AudioBufferProvider* provider)
174{
175    mFirCoefs = (mInSampleRate <= mSampleRate) ? mFirCoefsUp : mFirCoefsDown;
176
177    // select the appropriate resampler
178    switch (mChannelCount) {
179    case 1:
180        resample<1>(out, outFrameCount, provider);
181        break;
182    case 2:
183        resample<2>(out, outFrameCount, provider);
184        break;
185    }
186}
187
188
189template<int CHANNELS>
190void AudioResamplerSinc::resample(int32_t* out, size_t outFrameCount,
191        AudioBufferProvider* provider)
192{
193    int16_t* impulse = mImpulse;
194    uint32_t vRL = mVolumeRL;
195    size_t inputIndex = mInputIndex;
196    uint32_t phaseFraction = mPhaseFraction;
197    uint32_t phaseIncrement = mPhaseIncrement;
198    size_t outputIndex = 0;
199    size_t outputSampleCount = outFrameCount * 2;
200    size_t inFrameCount = (outFrameCount*mInSampleRate)/mSampleRate;
201
202    AudioBufferProvider::Buffer& buffer(mBuffer);
203    while (outputIndex < outputSampleCount) {
204        // buffer is empty, fetch a new one
205        while (buffer.frameCount == 0) {
206            buffer.frameCount = inFrameCount;
207            provider->getNextBuffer(&buffer);
208            if (buffer.raw == NULL) {
209                goto resample_exit;
210            }
211            const uint32_t phaseIndex = phaseFraction >> kNumPhaseBits;
212            if (phaseIndex == 1) {
213                // read one frame
214                read<CHANNELS>(impulse, phaseFraction, buffer.i16, inputIndex);
215            } else if (phaseIndex == 2) {
216                // read 2 frames
217                read<CHANNELS>(impulse, phaseFraction, buffer.i16, inputIndex);
218                inputIndex++;
219                if (inputIndex >= mBuffer.frameCount) {
220                    inputIndex -= mBuffer.frameCount;
221                    provider->releaseBuffer(&buffer);
222                } else {
223                    read<CHANNELS>(impulse, phaseFraction, buffer.i16, inputIndex);
224                }
225           }
226        }
227        int16_t *in = buffer.i16;
228        const size_t frameCount = buffer.frameCount;
229
230        // Always read-in the first samples from the input buffer
231        int16_t* head = impulse + halfNumCoefs*CHANNELS;
232        head[0] = in[inputIndex*CHANNELS + 0];
233        if (CHANNELS == 2)
234            head[1] = in[inputIndex*CHANNELS + 1];
235
236        // handle boundary case
237        int32_t l, r;
238        while (outputIndex < outputSampleCount) {
239            filterCoefficient<CHANNELS>(l, r, phaseFraction, impulse);
240            out[outputIndex++] += 2 * mulRL(1, l, vRL);
241            out[outputIndex++] += 2 * mulRL(0, r, vRL);
242
243            phaseFraction += phaseIncrement;
244            const uint32_t phaseIndex = phaseFraction >> kNumPhaseBits;
245            if (phaseIndex == 1) {
246                inputIndex++;
247                if (inputIndex >= frameCount)
248                    break;  // need a new buffer
249                read<CHANNELS>(impulse, phaseFraction, in, inputIndex);
250            } else if(phaseIndex == 2) {    // maximum value
251                inputIndex++;
252                if (inputIndex >= frameCount)
253                    break;  // 0 frame available, 2 frames needed
254                // read first frame
255                read<CHANNELS>(impulse, phaseFraction, in, inputIndex);
256                inputIndex++;
257                if (inputIndex >= frameCount)
258                    break;  // 0 frame available, 1 frame needed
259                // read second frame
260                read<CHANNELS>(impulse, phaseFraction, in, inputIndex);
261            }
262        }
263
264        // if done with buffer, save samples
265        if (inputIndex >= frameCount) {
266            inputIndex -= frameCount;
267            provider->releaseBuffer(&buffer);
268        }
269    }
270
271resample_exit:
272    mImpulse = impulse;
273    mInputIndex = inputIndex;
274    mPhaseFraction = phaseFraction;
275}
276
277template<int CHANNELS>
278/***
279* read()
280*
281* This function reads only one frame from input buffer and writes it in
282* state buffer
283*
284**/
285void AudioResamplerSinc::read(
286        int16_t*& impulse, uint32_t& phaseFraction,
287        const int16_t* in, size_t inputIndex)
288{
289    const uint32_t phaseIndex = phaseFraction >> kNumPhaseBits;
290    impulse += CHANNELS;
291    phaseFraction -= 1LU<<kNumPhaseBits;
292    if (impulse >= mRingFull) {
293        const size_t stateSize = (halfNumCoefs*2)*CHANNELS;
294        memcpy(mState, mState+stateSize, sizeof(int16_t)*stateSize);
295        impulse -= stateSize;
296    }
297    int16_t* head = impulse + halfNumCoefs*CHANNELS;
298    head[0] = in[inputIndex*CHANNELS + 0];
299    if (CHANNELS == 2)
300        head[1] = in[inputIndex*CHANNELS + 1];
301}
302
303template<int CHANNELS>
304void AudioResamplerSinc::filterCoefficient(
305        int32_t& l, int32_t& r, uint32_t phase, const int16_t *samples)
306{
307    // compute the index of the coefficient on the positive side and
308    // negative side
309    uint32_t indexP = (phase & cMask) >> cShift;
310    uint16_t lerpP  = (phase & pMask) >> pShift;
311    uint32_t indexN = (-phase & cMask) >> cShift;
312    uint16_t lerpN  = (-phase & pMask) >> pShift;
313    if ((indexP == 0) && (lerpP == 0)) {
314        indexN = cMask >> cShift;
315        lerpN = pMask >> pShift;
316    }
317
318    l = 0;
319    r = 0;
320    const int32_t* coefs = mFirCoefs;
321    const int16_t *sP = samples;
322    const int16_t *sN = samples+CHANNELS;
323    for (unsigned int i=0 ; i<halfNumCoefs/4 ; i++) {
324        interpolate<CHANNELS>(l, r, coefs+indexP, lerpP, sP);
325        interpolate<CHANNELS>(l, r, coefs+indexN, lerpN, sN);
326        sP -= CHANNELS; sN += CHANNELS; coefs += 1<<coefsBits;
327        interpolate<CHANNELS>(l, r, coefs+indexP, lerpP, sP);
328        interpolate<CHANNELS>(l, r, coefs+indexN, lerpN, sN);
329        sP -= CHANNELS; sN += CHANNELS; coefs += 1<<coefsBits;
330        interpolate<CHANNELS>(l, r, coefs+indexP, lerpP, sP);
331        interpolate<CHANNELS>(l, r, coefs+indexN, lerpN, sN);
332        sP -= CHANNELS; sN += CHANNELS; coefs += 1<<coefsBits;
333        interpolate<CHANNELS>(l, r, coefs+indexP, lerpP, sP);
334        interpolate<CHANNELS>(l, r, coefs+indexN, lerpN, sN);
335        sP -= CHANNELS; sN += CHANNELS; coefs += 1<<coefsBits;
336    }
337}
338
339template<int CHANNELS>
340void AudioResamplerSinc::interpolate(
341        int32_t& l, int32_t& r,
342        const int32_t* coefs, int16_t lerp, const int16_t* samples)
343{
344    int32_t c0 = coefs[0];
345    int32_t c1 = coefs[1];
346    int32_t sinc = mulAdd(lerp, (c1-c0)<<1, c0);
347    if (CHANNELS == 2) {
348        uint32_t rl = *reinterpret_cast<const uint32_t*>(samples);
349        l = mulAddRL(1, rl, sinc, l);
350        r = mulAddRL(0, rl, sinc, r);
351    } else {
352        r = l = mulAdd(samples[0], sinc, l);
353    }
354}
355
356// ----------------------------------------------------------------------------
357}; // namespace android
358
359