/* Decode a frame */
internal static int silk_Decode( /* O Returns error code */
SilkDecoder psDec, /* I/O State */
DecControlState decControl, /* I/O Control Structure */
int lostFlag, /* I 0: no loss, 1 loss, 2 decode fec */
int newPacketFlag, /* I Indicates first decoder call for this packet */
EntropyCoder psRangeDec, /* I/O Compressor data structure */
short[] samplesOut, /* O Decoded output speech vector */
int samplesOut_ptr,
out int nSamplesOut /* O Number of samples decoded */
)
{
int i, n, decode_only_middle = 0, ret = SilkError.SILK_NO_ERROR;
int LBRR_symbol;
BoxedValueInt nSamplesOutDec = new BoxedValueInt();
short[] samplesOut_tmp;
int[] samplesOut_tmp_ptrs = new int[2];
short[] samplesOut1_tmp_storage1;
short[] samplesOut1_tmp_storage2;
short[] samplesOut2_tmp;
int[] MS_pred_Q13 = new int[] { 0, 0 };
short[] resample_out;
int resample_out_ptr;
SilkChannelDecoder[] channel_state = psDec.channel_state;
int has_side;
int stereo_to_mono;
int delay_stack_alloc;
nSamplesOut = 0;
Inlines.OpusAssert(decControl.nChannelsInternal == 1 || decControl.nChannelsInternal == 2);
/**********************************/
/* Test if first frame in payload */
/**********************************/
if (newPacketFlag != 0)
{
for (n = 0; n < decControl.nChannelsInternal; n++)
{
channel_state[n].nFramesDecoded = 0; /* Used to count frames in packet */
}
}
/* If Mono . Stereo transition in bitstream: init state of second channel */
if (decControl.nChannelsInternal > psDec.nChannelsInternal)
{
ret += channel_state[1].silk_init_decoder();
}
stereo_to_mono = (decControl.nChannelsInternal == 1 && psDec.nChannelsInternal == 2 &&
(decControl.internalSampleRate == 1000 * channel_state[0].fs_kHz)) ? 1 : 0;
if (channel_state[0].nFramesDecoded == 0)
{
for (n = 0; n < decControl.nChannelsInternal; n++)
{
int fs_kHz_dec;
if (decControl.payloadSize_ms == 0)
{
/* Assuming packet loss, use 10 ms */
channel_state[n].nFramesPerPacket = 1;
channel_state[n].nb_subfr = 2;
}
else if (decControl.payloadSize_ms == 10)
{
channel_state[n].nFramesPerPacket = 1;
channel_state[n].nb_subfr = 2;
}
else if (decControl.payloadSize_ms == 20)
{
channel_state[n].nFramesPerPacket = 1;
channel_state[n].nb_subfr = 4;
}
else if (decControl.payloadSize_ms == 40)
{
channel_state[n].nFramesPerPacket = 2;
channel_state[n].nb_subfr = 4;
}
else if (decControl.payloadSize_ms == 60)
{
channel_state[n].nFramesPerPacket = 3;
channel_state[n].nb_subfr = 4;
}
else
{
Inlines.OpusAssert(false);
return(SilkError.SILK_DEC_INVALID_FRAME_SIZE);
}
fs_kHz_dec = (decControl.internalSampleRate >> 10) + 1;
if (fs_kHz_dec != 8 && fs_kHz_dec != 12 && fs_kHz_dec != 16)
{
Inlines.OpusAssert(false);
return(SilkError.SILK_DEC_INVALID_SAMPLING_FREQUENCY);
}
ret += channel_state[n].silk_decoder_set_fs(fs_kHz_dec, decControl.API_sampleRate);
}
}
if (decControl.nChannelsAPI == 2 && decControl.nChannelsInternal == 2 && (psDec.nChannelsAPI == 1 || psDec.nChannelsInternal == 1))
{
Arrays.MemSetShort(psDec.sStereo.pred_prev_Q13, 0, 2);
Arrays.MemSetShort(psDec.sStereo.sSide, 0, 2);
channel_state[1].resampler_state.Assign(channel_state[0].resampler_state);
}
psDec.nChannelsAPI = decControl.nChannelsAPI;
psDec.nChannelsInternal = decControl.nChannelsInternal;
if (decControl.API_sampleRate > (int)SilkConstants.MAX_API_FS_KHZ * 1000 || decControl.API_sampleRate < 8000)
{
ret = SilkError.SILK_DEC_INVALID_SAMPLING_FREQUENCY;
return(ret);
}
if (lostFlag != DecoderAPIFlag.FLAG_PACKET_LOST && channel_state[0].nFramesDecoded == 0)
{
/* First decoder call for this payload */
/* Decode VAD flags and LBRR flag */
for (n = 0; n < decControl.nChannelsInternal; n++)
{
for (i = 0; i < channel_state[n].nFramesPerPacket; i++)
{
channel_state[n].VAD_flags[i] = psRangeDec.dec_bit_logp(1);
}
channel_state[n].LBRR_flag = psRangeDec.dec_bit_logp(1);
}
/* Decode LBRR flags */
for (n = 0; n < decControl.nChannelsInternal; n++)
{
Arrays.MemSetInt(channel_state[n].LBRR_flags, 0, SilkConstants.MAX_FRAMES_PER_PACKET);
if (channel_state[n].LBRR_flag != 0)
{
if (channel_state[n].nFramesPerPacket == 1)
{
channel_state[n].LBRR_flags[0] = 1;
}
else
{
LBRR_symbol = psRangeDec.dec_icdf(Tables.silk_LBRR_flags_iCDF_ptr[channel_state[n].nFramesPerPacket - 2], 8) + 1;
for (i = 0; i < channel_state[n].nFramesPerPacket; i++)
{
channel_state[n].LBRR_flags[i] = Inlines.silk_RSHIFT(LBRR_symbol, i) & 1;
}
}
}
}
if (lostFlag == DecoderAPIFlag.FLAG_DECODE_NORMAL)
{
/* Regular decoding: skip all LBRR data */
for (i = 0; i < channel_state[0].nFramesPerPacket; i++)
{
for (n = 0; n < decControl.nChannelsInternal; n++)
{
if (channel_state[n].LBRR_flags[i] != 0)
{
short[] pulses = new short[SilkConstants.MAX_FRAME_LENGTH];
int condCoding;
if (decControl.nChannelsInternal == 2 && n == 0)
{
Stereo.silk_stereo_decode_pred(psRangeDec, MS_pred_Q13);
if (channel_state[1].LBRR_flags[i] == 0)
{
BoxedValueInt decodeOnlyMiddleBoxed = new BoxedValueInt(decode_only_middle);
Stereo.silk_stereo_decode_mid_only(psRangeDec, decodeOnlyMiddleBoxed);
decode_only_middle = decodeOnlyMiddleBoxed.Val;
}
}
/* Use conditional coding if previous frame available */
if (i > 0 && (channel_state[n].LBRR_flags[i - 1] != 0))
{
condCoding = SilkConstants.CODE_CONDITIONALLY;
}
else
{
condCoding = SilkConstants.CODE_INDEPENDENTLY;
}
DecodeIndices.silk_decode_indices(channel_state[n], psRangeDec, i, 1, condCoding);
DecodePulses.silk_decode_pulses(psRangeDec, pulses, channel_state[n].indices.signalType,
channel_state[n].indices.quantOffsetType, channel_state[n].frame_length);
}
}
}
}
}
/* Get MS predictor index */
if (decControl.nChannelsInternal == 2)
{
if (lostFlag == DecoderAPIFlag.FLAG_DECODE_NORMAL ||
(lostFlag == DecoderAPIFlag.FLAG_DECODE_LBRR && channel_state[0].LBRR_flags[channel_state[0].nFramesDecoded] == 1))
{
Stereo.silk_stereo_decode_pred(psRangeDec, MS_pred_Q13);
/* For LBRR data, decode mid-only flag only if side-channel's LBRR flag is false */
if ((lostFlag == DecoderAPIFlag.FLAG_DECODE_NORMAL && channel_state[1].VAD_flags[channel_state[0].nFramesDecoded] == 0) ||
(lostFlag == DecoderAPIFlag.FLAG_DECODE_LBRR && channel_state[1].LBRR_flags[channel_state[0].nFramesDecoded] == 0))
{
BoxedValueInt decodeOnlyMiddleBoxed = new BoxedValueInt(decode_only_middle);
Stereo.silk_stereo_decode_mid_only(psRangeDec, decodeOnlyMiddleBoxed);
decode_only_middle = decodeOnlyMiddleBoxed.Val;
}
else
{
decode_only_middle = 0;
}
}
else
{
for (n = 0; n < 2; n++)
{
MS_pred_Q13[n] = psDec.sStereo.pred_prev_Q13[n];
}
}
}
/* Reset side channel decoder prediction memory for first frame with side coding */
if (decControl.nChannelsInternal == 2 && decode_only_middle == 0 && psDec.prev_decode_only_middle == 1)
{
Arrays.MemSetShort(psDec.channel_state[1].outBuf, 0, SilkConstants.MAX_FRAME_LENGTH + 2 * SilkConstants.MAX_SUB_FRAME_LENGTH);
Arrays.MemSetInt(psDec.channel_state[1].sLPC_Q14_buf, 0, SilkConstants.MAX_LPC_ORDER);
psDec.channel_state[1].lagPrev = 100;
psDec.channel_state[1].LastGainIndex = 10;
psDec.channel_state[1].prevSignalType = SilkConstants.TYPE_NO_VOICE_ACTIVITY;
psDec.channel_state[1].first_frame_after_reset = 1;
}
/* Check if the temp buffer fits into the output PCM buffer. If it fits,
* we can delay allocating the temp buffer until after the SILK peak stack
* usage. We need to use a < and not a <= because of the two extra samples. */
delay_stack_alloc = (decControl.internalSampleRate * decControl.nChannelsInternal
< decControl.API_sampleRate * decControl.nChannelsAPI) ? 1 : 0;
if (delay_stack_alloc != 0)
{
samplesOut_tmp = samplesOut;
samplesOut_tmp_ptrs[0] = samplesOut_ptr;
samplesOut_tmp_ptrs[1] = samplesOut_ptr + channel_state[0].frame_length + 2;
}
else
{
samplesOut1_tmp_storage1 = new short[decControl.nChannelsInternal * (channel_state[0].frame_length + 2)];
samplesOut_tmp = samplesOut1_tmp_storage1;
samplesOut_tmp_ptrs[0] = 0;
samplesOut_tmp_ptrs[1] = channel_state[0].frame_length + 2;
}
if (lostFlag == DecoderAPIFlag.FLAG_DECODE_NORMAL)
{
has_side = (decode_only_middle == 0) ? 1 : 0;
}
else
{
has_side = (psDec.prev_decode_only_middle == 0 ||
(decControl.nChannelsInternal == 2 &&
lostFlag == DecoderAPIFlag.FLAG_DECODE_LBRR &&
channel_state[1].LBRR_flags[channel_state[1].nFramesDecoded] == 1)) ? 1 : 0;
}
/* Call decoder for one frame */
for (n = 0; n < decControl.nChannelsInternal; n++)
{
if (n == 0 || (has_side != 0))
{
int FrameIndex;
int condCoding;
FrameIndex = channel_state[0].nFramesDecoded - n;
/* Use independent coding if no previous frame available */
if (FrameIndex <= 0)
{
condCoding = SilkConstants.CODE_INDEPENDENTLY;
}
else if (lostFlag == DecoderAPIFlag.FLAG_DECODE_LBRR)
{
condCoding = (channel_state[n].LBRR_flags[FrameIndex - 1] != 0) ? SilkConstants.CODE_CONDITIONALLY : SilkConstants.CODE_INDEPENDENTLY;
}
else if (n > 0 && (psDec.prev_decode_only_middle != 0))
{
/* If we skipped a side frame in this packet, we don't
* need LTP scaling; the LTP state is well-defined. */
condCoding = SilkConstants.CODE_INDEPENDENTLY_NO_LTP_SCALING;
}
else
{
condCoding = SilkConstants.CODE_CONDITIONALLY;
}
ret += channel_state[n].silk_decode_frame(psRangeDec, samplesOut_tmp, samplesOut_tmp_ptrs[n] + 2, nSamplesOutDec, lostFlag, condCoding);
}
else
{
Arrays.MemSetWithOffset <short>(samplesOut_tmp, 0, samplesOut_tmp_ptrs[n] + 2, nSamplesOutDec.Val);
}
channel_state[n].nFramesDecoded++;
}
if (decControl.nChannelsAPI == 2 && decControl.nChannelsInternal == 2)
{
/* Convert Mid/Side to Left/Right */
Stereo.silk_stereo_MS_to_LR(psDec.sStereo, samplesOut_tmp, samplesOut_tmp_ptrs[0], samplesOut_tmp, samplesOut_tmp_ptrs[1], MS_pred_Q13, channel_state[0].fs_kHz, nSamplesOutDec.Val);
}
else
{
/* Buffering */
Array.Copy(psDec.sStereo.sMid, 0, samplesOut_tmp, samplesOut_tmp_ptrs[0], 2);
Array.Copy(samplesOut_tmp, samplesOut_tmp_ptrs[0] + nSamplesOutDec.Val, psDec.sStereo.sMid, 0, 2);
}
/* Number of output samples */
nSamplesOut = Inlines.silk_DIV32(nSamplesOutDec.Val * decControl.API_sampleRate, Inlines.silk_SMULBB(channel_state[0].fs_kHz, 1000));
/* Set up pointers to temp buffers */
if (decControl.nChannelsAPI == 2)
{
samplesOut2_tmp = new short[nSamplesOut];
resample_out = samplesOut2_tmp;
resample_out_ptr = 0;
}
else
{
resample_out = samplesOut;
resample_out_ptr = samplesOut_ptr;
}
if (delay_stack_alloc != 0)
{
samplesOut1_tmp_storage2 = new short[decControl.nChannelsInternal * (channel_state[0].frame_length + 2)];
Array.Copy(samplesOut, samplesOut_ptr, samplesOut1_tmp_storage2, 0, decControl.nChannelsInternal * (channel_state[0].frame_length + 2));
samplesOut_tmp = samplesOut1_tmp_storage2;
samplesOut_tmp_ptrs[0] = 0;
samplesOut_tmp_ptrs[1] = channel_state[0].frame_length + 2;
}
for (n = 0; n < Inlines.silk_min(decControl.nChannelsAPI, decControl.nChannelsInternal); n++)
{
/* Resample decoded signal to API_sampleRate */
ret += Resampler.silk_resampler(channel_state[n].resampler_state, resample_out, resample_out_ptr, samplesOut_tmp, samplesOut_tmp_ptrs[n] + 1, nSamplesOutDec.Val);
/* Interleave if stereo output and stereo stream */
if (decControl.nChannelsAPI == 2)
{
int nptr = samplesOut_ptr + n;
for (i = 0; i < nSamplesOut; i++)
{
samplesOut[nptr + 2 * i] = resample_out[resample_out_ptr + i];
}
}
}
/* Create two channel output from mono stream */
if (decControl.nChannelsAPI == 2 && decControl.nChannelsInternal == 1)
{
if (stereo_to_mono != 0)
{
/* Resample right channel for newly collapsed stereo just in case
* we weren't doing collapsing when switching to mono */
ret += Resampler.silk_resampler(channel_state[1].resampler_state, resample_out, resample_out_ptr, samplesOut_tmp, samplesOut_tmp_ptrs[0] + 1, nSamplesOutDec.Val);
for (i = 0; i < nSamplesOut; i++)
{
samplesOut[samplesOut_ptr + 1 + 2 * i] = resample_out[resample_out_ptr + i];
}
}
else
{
for (i = 0; i < nSamplesOut; i++)
{
samplesOut[samplesOut_ptr + 1 + 2 * i] = samplesOut[samplesOut_ptr + 2 * i];
}
}
}
/* Export pitch lag, measured at 48 kHz sampling rate */
if (channel_state[0].prevSignalType == SilkConstants.TYPE_VOICED)
{
int[] mult_tab = { 6, 4, 3 };
decControl.prevPitchLag = channel_state[0].lagPrev * mult_tab[(channel_state[0].fs_kHz - 8) >> 2];
}
else
{
decControl.prevPitchLag = 0;
}
if (lostFlag == DecoderAPIFlag.FLAG_PACKET_LOST)
{
/* On packet loss, remove the gain clamping to prevent having the energy "bounce back"
* if we lose packets when the energy is going down */
for (i = 0; i < psDec.nChannelsInternal; i++)
{
psDec.channel_state[i].LastGainIndex = 10;
}
}
else
{
psDec.prev_decode_only_middle = decode_only_middle;
}
return(ret);
}
/// <summary>
/// Encode frame with Silk
/// Note: if prefillFlag is set, the input must contain 10 ms of audio, irrespective of what
/// encControl.payloadSize_ms is set to
/// </summary>
/// <param name="psEnc">I/O State</param>
/// <param name="encControl">I Control status</param>
/// <param name="samplesIn">I Speech sample input vector</param>
/// <param name="nSamplesIn">I Number of samples in input vector</param>
/// <param name="psRangeEnc">I/O Compressor data structure</param>
/// <param name="nBytesOut">I/O Number of bytes in payload (input: Max bytes)</param>
/// <param name="prefillFlag">I Flag to indicate prefilling buffers no coding</param>
/// <returns>error code</returns>
internal static int silk_Encode(
SilkEncoder psEnc,
EncControlState encControl,
short[] samplesIn,
int nSamplesIn,
EntropyCoder psRangeEnc,
BoxedValueInt nBytesOut,
int prefillFlag)
{
int ret = SilkError.SILK_NO_ERROR;
int n, i, nBits, flags, tmp_payloadSize_ms = 0, tmp_complexity = 0;
int nSamplesToBuffer, nSamplesToBufferMax, nBlocksOf10ms;
int nSamplesFromInput = 0, nSamplesFromInputMax;
int speech_act_thr_for_switch_Q8;
int TargetRate_bps, channelRate_bps, LBRR_symbol, sum;
int[] MStargetRates_bps = new int[2];
short[] buf;
int transition, curr_block, tot_blocks;
nBytesOut.Val = 0;
if (encControl.reducedDependency != 0)
{
psEnc.state_Fxx[0].first_frame_after_reset = 1;
psEnc.state_Fxx[1].first_frame_after_reset = 1;
}
psEnc.state_Fxx[0].nFramesEncoded = psEnc.state_Fxx[1].nFramesEncoded = 0;
/* Check values in encoder control structure */
ret += encControl.check_control_input();
if (ret != SilkError.SILK_NO_ERROR)
{
Inlines.OpusAssert(false);
return(ret);
}
encControl.switchReady = 0;
if (encControl.nChannelsInternal > psEnc.nChannelsInternal)
{
/* Mono . Stereo transition: init state of second channel and stereo state */
ret += SilkEncoder.silk_init_encoder(psEnc.state_Fxx[1]);
Arrays.MemSetShort(psEnc.sStereo.pred_prev_Q13, 0, 2);
Arrays.MemSetShort(psEnc.sStereo.sSide, 0, 2);
psEnc.sStereo.mid_side_amp_Q0[0] = 0;
psEnc.sStereo.mid_side_amp_Q0[1] = 1;
psEnc.sStereo.mid_side_amp_Q0[2] = 0;
psEnc.sStereo.mid_side_amp_Q0[3] = 1;
psEnc.sStereo.width_prev_Q14 = 0;
psEnc.sStereo.smth_width_Q14 = (short)(((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/);
if (psEnc.nChannelsAPI == 2)
{
psEnc.state_Fxx[1].resampler_state.Assign(psEnc.state_Fxx[0].resampler_state);
Array.Copy(psEnc.state_Fxx[0].In_HP_State, psEnc.state_Fxx[1].In_HP_State, 2);
}
}
transition = ((encControl.payloadSize_ms != psEnc.state_Fxx[0].PacketSize_ms) || (psEnc.nChannelsInternal != encControl.nChannelsInternal)) ? 1 : 0;
psEnc.nChannelsAPI = encControl.nChannelsAPI;
psEnc.nChannelsInternal = encControl.nChannelsInternal;
nBlocksOf10ms = Inlines.silk_DIV32(100 * nSamplesIn, encControl.API_sampleRate);
tot_blocks = (nBlocksOf10ms > 1) ? nBlocksOf10ms >> 1 : 1;
curr_block = 0;
if (prefillFlag != 0)
{
/* Only accept input length of 10 ms */
if (nBlocksOf10ms != 1)
{
Inlines.OpusAssert(false);
return(SilkError.SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES);
}
/* Reset Encoder */
for (n = 0; n < encControl.nChannelsInternal; n++)
{
ret += SilkEncoder.silk_init_encoder(psEnc.state_Fxx[n]);
Inlines.OpusAssert(ret == SilkError.SILK_NO_ERROR);
}
tmp_payloadSize_ms = encControl.payloadSize_ms;
encControl.payloadSize_ms = 10;
tmp_complexity = encControl.complexity;
encControl.complexity = 0;
for (n = 0; n < encControl.nChannelsInternal; n++)
{
psEnc.state_Fxx[n].controlled_since_last_payload = 0;
psEnc.state_Fxx[n].prefillFlag = 1;
}
}
else
{
/* Only accept input lengths that are a multiple of 10 ms */
if (nBlocksOf10ms * encControl.API_sampleRate != 100 * nSamplesIn || nSamplesIn < 0)
{
Inlines.OpusAssert(false);
return(SilkError.SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES);
}
/* Make sure no more than one packet can be produced */
if (1000 * (int)nSamplesIn > encControl.payloadSize_ms * encControl.API_sampleRate)
{
Inlines.OpusAssert(false);
return(SilkError.SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES);
}
}
TargetRate_bps = Inlines.silk_RSHIFT32(encControl.bitRate, encControl.nChannelsInternal - 1);
for (n = 0; n < encControl.nChannelsInternal; n++)
{
/* Force the side channel to the same rate as the mid */
int force_fs_kHz = (n == 1) ? psEnc.state_Fxx[0].fs_kHz : 0;
ret += psEnc.state_Fxx[n].silk_control_encoder(encControl, TargetRate_bps, psEnc.allowBandwidthSwitch, n, force_fs_kHz);
if (ret != SilkError.SILK_NO_ERROR)
{
Inlines.OpusAssert(false);
return(ret);
}
if (psEnc.state_Fxx[n].first_frame_after_reset != 0 || transition != 0)
{
for (i = 0; i < psEnc.state_Fxx[0].nFramesPerPacket; i++)
{
psEnc.state_Fxx[n].LBRR_flags[i] = 0;
}
}
psEnc.state_Fxx[n].inDTX = psEnc.state_Fxx[n].useDTX;
}
Inlines.OpusAssert(encControl.nChannelsInternal == 1 || psEnc.state_Fxx[0].fs_kHz == psEnc.state_Fxx[1].fs_kHz);
/* Input buffering/resampling and encoding */
nSamplesToBufferMax = 10 * nBlocksOf10ms * psEnc.state_Fxx[0].fs_kHz;
nSamplesFromInputMax =
Inlines.silk_DIV32_16(nSamplesToBufferMax *
psEnc.state_Fxx[0].API_fs_Hz,
(short)(psEnc.state_Fxx[0].fs_kHz * 1000));
buf = new short[nSamplesFromInputMax];
int samplesIn_ptr = 0;
while (true)
{
nSamplesToBuffer = psEnc.state_Fxx[0].frame_length - psEnc.state_Fxx[0].inputBufIx;
nSamplesToBuffer = Inlines.silk_min(nSamplesToBuffer, nSamplesToBufferMax);
nSamplesFromInput = Inlines.silk_DIV32_16(nSamplesToBuffer * psEnc.state_Fxx[0].API_fs_Hz, psEnc.state_Fxx[0].fs_kHz * 1000);
/* Resample and write to buffer */
if (encControl.nChannelsAPI == 2 && encControl.nChannelsInternal == 2)
{
int id = psEnc.state_Fxx[0].nFramesEncoded;
for (n = 0; n < nSamplesFromInput; n++)
{
buf[n] = samplesIn[samplesIn_ptr + (2 * n)];
}
/* Making sure to start both resamplers from the same state when switching from mono to stereo */
if (psEnc.nPrevChannelsInternal == 1 && id == 0)
{
//silk_memcpy(&psEnc.state_Fxx[1].resampler_state, &psEnc.state_Fxx[0].resampler_state, sizeof(psEnc.state_Fxx[1].resampler_state));
psEnc.state_Fxx[1].resampler_state.Assign(psEnc.state_Fxx[0].resampler_state);
}
ret += Resampler.silk_resampler(
psEnc.state_Fxx[0].resampler_state,
psEnc.state_Fxx[0].inputBuf,
psEnc.state_Fxx[0].inputBufIx + 2,
buf,
0,
nSamplesFromInput);
psEnc.state_Fxx[0].inputBufIx += nSamplesToBuffer;
nSamplesToBuffer = psEnc.state_Fxx[1].frame_length - psEnc.state_Fxx[1].inputBufIx;
nSamplesToBuffer = Inlines.silk_min(nSamplesToBuffer, 10 * nBlocksOf10ms * psEnc.state_Fxx[1].fs_kHz);
for (n = 0; n < nSamplesFromInput; n++)
{
buf[n] = samplesIn[samplesIn_ptr + (2 * n) + 1];
}
ret += Resampler.silk_resampler(
psEnc.state_Fxx[1].resampler_state,
psEnc.state_Fxx[1].inputBuf,
psEnc.state_Fxx[1].inputBufIx + 2,
buf,
0,
nSamplesFromInput);
psEnc.state_Fxx[1].inputBufIx += nSamplesToBuffer;
}
else if (encControl.nChannelsAPI == 2 && encControl.nChannelsInternal == 1)
{
/* Combine left and right channels before resampling */
for (n = 0; n < nSamplesFromInput; n++)
{
sum = samplesIn[samplesIn_ptr + (2 * n)] + samplesIn[samplesIn_ptr + (2 * n) + 1];
buf[n] = (short)Inlines.silk_RSHIFT_ROUND(sum, 1);
}
ret += Resampler.silk_resampler(
psEnc.state_Fxx[0].resampler_state,
psEnc.state_Fxx[0].inputBuf,
psEnc.state_Fxx[0].inputBufIx + 2,
buf,
0,
nSamplesFromInput);
/* On the first mono frame, average the results for the two resampler states */
if (psEnc.nPrevChannelsInternal == 2 && psEnc.state_Fxx[0].nFramesEncoded == 0)
{
ret += Resampler.silk_resampler(
psEnc.state_Fxx[1].resampler_state,
psEnc.state_Fxx[1].inputBuf,
psEnc.state_Fxx[1].inputBufIx + 2,
buf,
0,
nSamplesFromInput);
for (n = 0; n < psEnc.state_Fxx[0].frame_length; n++)
{
psEnc.state_Fxx[0].inputBuf[psEnc.state_Fxx[0].inputBufIx + n + 2] =
(short)(Inlines.silk_RSHIFT(psEnc.state_Fxx[0].inputBuf[psEnc.state_Fxx[0].inputBufIx + n + 2]
+ psEnc.state_Fxx[1].inputBuf[psEnc.state_Fxx[1].inputBufIx + n + 2], 1));
}
}
psEnc.state_Fxx[0].inputBufIx += nSamplesToBuffer;
}
else
{
Inlines.OpusAssert(encControl.nChannelsAPI == 1 && encControl.nChannelsInternal == 1);
Array.Copy(samplesIn, samplesIn_ptr, buf, 0, nSamplesFromInput);
ret += Resampler.silk_resampler(
psEnc.state_Fxx[0].resampler_state,
psEnc.state_Fxx[0].inputBuf,
psEnc.state_Fxx[0].inputBufIx + 2,
buf,
0,
nSamplesFromInput);
psEnc.state_Fxx[0].inputBufIx += nSamplesToBuffer;
}
samplesIn_ptr += (nSamplesFromInput * encControl.nChannelsAPI);
nSamplesIn -= nSamplesFromInput;
/* Default */
psEnc.allowBandwidthSwitch = 0;
/* Silk encoder */
if (psEnc.state_Fxx[0].inputBufIx >= psEnc.state_Fxx[0].frame_length)
{
/* Enough data in input buffer, so encode */
Inlines.OpusAssert(psEnc.state_Fxx[0].inputBufIx == psEnc.state_Fxx[0].frame_length);
Inlines.OpusAssert(encControl.nChannelsInternal == 1 || psEnc.state_Fxx[1].inputBufIx == psEnc.state_Fxx[1].frame_length);
/* Deal with LBRR data */
if (psEnc.state_Fxx[0].nFramesEncoded == 0 && prefillFlag == 0)
{
/* Create space at start of payload for VAD and FEC flags */
byte[] iCDF = { 0, 0 };
iCDF[0] = (byte)(256 - Inlines.silk_RSHIFT(256, (psEnc.state_Fxx[0].nFramesPerPacket + 1) * encControl.nChannelsInternal));
psRangeEnc.enc_icdf(0, iCDF, 8);
/* Encode any LBRR data from previous packet */
/* Encode LBRR flags */
for (n = 0; n < encControl.nChannelsInternal; n++)
{
LBRR_symbol = 0;
for (i = 0; i < psEnc.state_Fxx[n].nFramesPerPacket; i++)
{
LBRR_symbol |= Inlines.silk_LSHIFT(psEnc.state_Fxx[n].LBRR_flags[i], i);
}
psEnc.state_Fxx[n].LBRR_flag = (sbyte)(LBRR_symbol > 0 ? 1 : 0);
if (LBRR_symbol != 0 && psEnc.state_Fxx[n].nFramesPerPacket > 1)
{
psRangeEnc.enc_icdf(LBRR_symbol - 1, Tables.silk_LBRR_flags_iCDF_ptr[psEnc.state_Fxx[n].nFramesPerPacket - 2], 8);
}
}
/* Code LBRR indices and excitation signals */
for (i = 0; i < psEnc.state_Fxx[0].nFramesPerPacket; i++)
{
for (n = 0; n < encControl.nChannelsInternal; n++)
{
if (psEnc.state_Fxx[n].LBRR_flags[i] != 0)
{
int condCoding;
if (encControl.nChannelsInternal == 2 && n == 0)
{
Stereo.silk_stereo_encode_pred(psRangeEnc, psEnc.sStereo.predIx[i]);
/* For LBRR data there's no need to code the mid-only flag if the side-channel LBRR flag is set */
if (psEnc.state_Fxx[1].LBRR_flags[i] == 0)
{
Stereo.silk_stereo_encode_mid_only(psRangeEnc, psEnc.sStereo.mid_only_flags[i]);
}
}
/* Use conditional coding if previous frame available */
if (i > 0 && psEnc.state_Fxx[n].LBRR_flags[i - 1] != 0)
{
condCoding = SilkConstants.CODE_CONDITIONALLY;
}
else
{
condCoding = SilkConstants.CODE_INDEPENDENTLY;
}
EncodeIndices.silk_encode_indices(psEnc.state_Fxx[n], psRangeEnc, i, 1, condCoding);
EncodePulses.silk_encode_pulses(psRangeEnc, psEnc.state_Fxx[n].indices_LBRR[i].signalType, psEnc.state_Fxx[n].indices_LBRR[i].quantOffsetType,
psEnc.state_Fxx[n].pulses_LBRR[i], psEnc.state_Fxx[n].frame_length);
}
}
}
/* Reset LBRR flags */
for (n = 0; n < encControl.nChannelsInternal; n++)
{
Arrays.MemSetInt(psEnc.state_Fxx[n].LBRR_flags, 0, SilkConstants.MAX_FRAMES_PER_PACKET);
}
psEnc.nBitsUsedLBRR = psRangeEnc.tell();
}
HPVariableCutoff.silk_HP_variable_cutoff(psEnc.state_Fxx);
/* Total target bits for packet */
nBits = Inlines.silk_DIV32_16(Inlines.silk_MUL(encControl.bitRate, encControl.payloadSize_ms), 1000);
/* Subtract bits used for LBRR */
if (prefillFlag == 0)
{
nBits -= psEnc.nBitsUsedLBRR;
}
/* Divide by number of uncoded frames left in packet */
nBits = Inlines.silk_DIV32_16(nBits, psEnc.state_Fxx[0].nFramesPerPacket);
/* Convert to bits/second */
if (encControl.payloadSize_ms == 10)
{
TargetRate_bps = Inlines.silk_SMULBB(nBits, 100);
}
else
{
TargetRate_bps = Inlines.silk_SMULBB(nBits, 50);
}
/* Subtract fraction of bits in excess of target in previous frames and packets */
TargetRate_bps -= Inlines.silk_DIV32_16(Inlines.silk_MUL(psEnc.nBitsExceeded, 1000), TuningParameters.BITRESERVOIR_DECAY_TIME_MS);
if (prefillFlag == 0 && psEnc.state_Fxx[0].nFramesEncoded > 0)
{
/* Compare actual vs target bits so far in this packet */
int bitsBalance = psRangeEnc.tell() - psEnc.nBitsUsedLBRR - nBits * psEnc.state_Fxx[0].nFramesEncoded;
TargetRate_bps -= Inlines.silk_DIV32_16(Inlines.silk_MUL(bitsBalance, 1000), TuningParameters.BITRESERVOIR_DECAY_TIME_MS);
}
/* Never exceed input bitrate */
TargetRate_bps = Inlines.silk_LIMIT(TargetRate_bps, encControl.bitRate, 5000);
/* Convert Left/Right to Mid/Side */
if (encControl.nChannelsInternal == 2)
{
BoxedValueSbyte midOnlyFlagBoxed = new BoxedValueSbyte(psEnc.sStereo.mid_only_flags[psEnc.state_Fxx[0].nFramesEncoded]);
Stereo.silk_stereo_LR_to_MS(psEnc.sStereo,
psEnc.state_Fxx[0].inputBuf,
2,
psEnc.state_Fxx[1].inputBuf,
2,
psEnc.sStereo.predIx[psEnc.state_Fxx[0].nFramesEncoded],
midOnlyFlagBoxed,
MStargetRates_bps,
TargetRate_bps,
psEnc.state_Fxx[0].speech_activity_Q8,
encControl.toMono,
psEnc.state_Fxx[0].fs_kHz,
psEnc.state_Fxx[0].frame_length);
psEnc.sStereo.mid_only_flags[psEnc.state_Fxx[0].nFramesEncoded] = midOnlyFlagBoxed.Val;
if (midOnlyFlagBoxed.Val == 0)
{
/* Reset side channel encoder memory for first frame with side coding */
if (psEnc.prev_decode_only_middle == 1)
{
psEnc.state_Fxx[1].sShape.Reset();
psEnc.state_Fxx[1].sPrefilt.Reset();
psEnc.state_Fxx[1].sNSQ.Reset();
Arrays.MemSetShort(psEnc.state_Fxx[1].prev_NLSFq_Q15, 0, SilkConstants.MAX_LPC_ORDER);
Arrays.MemSetInt(psEnc.state_Fxx[1].sLP.In_LP_State, 0, 2);
psEnc.state_Fxx[1].prevLag = 100;
psEnc.state_Fxx[1].sNSQ.lagPrev = 100;
psEnc.state_Fxx[1].sShape.LastGainIndex = 10;
psEnc.state_Fxx[1].prevSignalType = SilkConstants.TYPE_NO_VOICE_ACTIVITY;
psEnc.state_Fxx[1].sNSQ.prev_gain_Q16 = 65536;
psEnc.state_Fxx[1].first_frame_after_reset = 1;
}
psEnc.state_Fxx[1].silk_encode_do_VAD();
}
else
{
psEnc.state_Fxx[1].VAD_flags[psEnc.state_Fxx[0].nFramesEncoded] = 0;
}
if (prefillFlag == 0)
{
Stereo.silk_stereo_encode_pred(psRangeEnc, psEnc.sStereo.predIx[psEnc.state_Fxx[0].nFramesEncoded]);
if (psEnc.state_Fxx[1].VAD_flags[psEnc.state_Fxx[0].nFramesEncoded] == 0)
{
Stereo.silk_stereo_encode_mid_only(psRangeEnc, psEnc.sStereo.mid_only_flags[psEnc.state_Fxx[0].nFramesEncoded]);
}
}
}
else
{
/* Buffering */
Array.Copy(psEnc.sStereo.sMid, psEnc.state_Fxx[0].inputBuf, 2);
Array.Copy(psEnc.state_Fxx[0].inputBuf, psEnc.state_Fxx[0].frame_length, psEnc.sStereo.sMid, 0, 2);
}
psEnc.state_Fxx[0].silk_encode_do_VAD();
/* Encode */
for (n = 0; n < encControl.nChannelsInternal; n++)
{
int maxBits, useCBR;
/* Handling rate constraints */
maxBits = encControl.maxBits;
if (tot_blocks == 2 && curr_block == 0)
{
maxBits = maxBits * 3 / 5;
}
else if (tot_blocks == 3)
{
if (curr_block == 0)
{
maxBits = maxBits * 2 / 5;
}
else if (curr_block == 1)
{
maxBits = maxBits * 3 / 4;
}
}
useCBR = (encControl.useCBR != 0 && curr_block == tot_blocks - 1) ? 1 : 0;
if (encControl.nChannelsInternal == 1)
{
channelRate_bps = TargetRate_bps;
}
else
{
channelRate_bps = MStargetRates_bps[n];
if (n == 0 && MStargetRates_bps[1] > 0)
{
useCBR = 0;
/* Give mid up to 1/2 of the max bits for that frame */
maxBits -= encControl.maxBits / (tot_blocks * 2);
}
}
if (channelRate_bps > 0)
{
int condCoding;
psEnc.state_Fxx[n].silk_control_SNR(channelRate_bps);
/* Use independent coding if no previous frame available */
if (psEnc.state_Fxx[0].nFramesEncoded - n <= 0)
{
condCoding = SilkConstants.CODE_INDEPENDENTLY;
}
else if (n > 0 && psEnc.prev_decode_only_middle != 0)
{
/* If we skipped a side frame in this packet, we don't
* need LTP scaling; the LTP state is well-defined. */
condCoding = SilkConstants.CODE_INDEPENDENTLY_NO_LTP_SCALING;
}
else
{
condCoding = SilkConstants.CODE_CONDITIONALLY;
}
ret += psEnc.state_Fxx[n].silk_encode_frame(nBytesOut, psRangeEnc, condCoding, maxBits, useCBR);
Inlines.OpusAssert(ret == SilkError.SILK_NO_ERROR);
}
psEnc.state_Fxx[n].controlled_since_last_payload = 0;
psEnc.state_Fxx[n].inputBufIx = 0;
psEnc.state_Fxx[n].nFramesEncoded++;
}
psEnc.prev_decode_only_middle = psEnc.sStereo.mid_only_flags[psEnc.state_Fxx[0].nFramesEncoded - 1];
/* Insert VAD and FEC flags at beginning of bitstream */
if (nBytesOut.Val > 0 && psEnc.state_Fxx[0].nFramesEncoded == psEnc.state_Fxx[0].nFramesPerPacket)
{
flags = 0;
for (n = 0; n < encControl.nChannelsInternal; n++)
{
for (i = 0; i < psEnc.state_Fxx[n].nFramesPerPacket; i++)
{
flags = Inlines.silk_LSHIFT(flags, 1);
flags |= (int)psEnc.state_Fxx[n].VAD_flags[i];
}
flags = Inlines.silk_LSHIFT(flags, 1);
flags |= (int)psEnc.state_Fxx[n].LBRR_flag;
}
if (prefillFlag == 0)
{
psRangeEnc.enc_patch_initial_bits((uint)flags, (uint)((psEnc.state_Fxx[0].nFramesPerPacket + 1) * encControl.nChannelsInternal));
}
/* Return zero bytes if all channels DTXed */
if (psEnc.state_Fxx[0].inDTX != 0 && (encControl.nChannelsInternal == 1 || psEnc.state_Fxx[1].inDTX != 0))
{
nBytesOut.Val = 0;
}
psEnc.nBitsExceeded += nBytesOut.Val * 8;
psEnc.nBitsExceeded -= Inlines.silk_DIV32_16(Inlines.silk_MUL(encControl.bitRate, encControl.payloadSize_ms), 1000);
psEnc.nBitsExceeded = Inlines.silk_LIMIT(psEnc.nBitsExceeded, 0, 10000);
/* Update flag indicating if bandwidth switching is allowed */
speech_act_thr_for_switch_Q8 = Inlines.silk_SMLAWB(((int)((TuningParameters.SPEECH_ACTIVITY_DTX_THRES) * ((long)1 << (8)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SPEECH_ACTIVITY_DTX_THRES, 8)*/,
((int)(((1 - TuningParameters.SPEECH_ACTIVITY_DTX_THRES) / TuningParameters.MAX_BANDWIDTH_SWITCH_DELAY_MS) * ((long)1 << (16 + 8)) + 0.5)) /*Inlines.SILK_CONST((1 - TuningParameters.SPEECH_ACTIVITY_DTX_THRES) / TuningParameters.MAX_BANDWIDTH_SWITCH_DELAY_MS, 16 + 8)*/,
psEnc.timeSinceSwitchAllowed_ms);
if (psEnc.state_Fxx[0].speech_activity_Q8 < speech_act_thr_for_switch_Q8)
{
psEnc.allowBandwidthSwitch = 1;
psEnc.timeSinceSwitchAllowed_ms = 0;
}
else
{
psEnc.allowBandwidthSwitch = 0;
psEnc.timeSinceSwitchAllowed_ms += encControl.payloadSize_ms;
}
}
if (nSamplesIn == 0)
{
break;
}
}
else
{
break;
}
curr_block++;
}
psEnc.nPrevChannelsInternal = encControl.nChannelsInternal;
encControl.allowBandwidthSwitch = psEnc.allowBandwidthSwitch;
encControl.inWBmodeWithoutVariableLP = (psEnc.state_Fxx[0].fs_kHz == 16 && psEnc.state_Fxx[0].sLP.mode == 0) ? 1 : 0;
encControl.internalSampleRate = Inlines.silk_SMULBB(psEnc.state_Fxx[0].fs_kHz, 1000);
encControl.stereoWidth_Q14 = encControl.toMono != 0 ? 0 : psEnc.sStereo.smth_width_Q14;
if (prefillFlag != 0)
{
encControl.payloadSize_ms = tmp_payloadSize_ms;
encControl.complexity = tmp_complexity;
for (n = 0; n < encControl.nChannelsInternal; n++)
{
psEnc.state_Fxx[n].controlled_since_last_payload = 0;
psEnc.state_Fxx[n].prefillFlag = 0;
}
}
return(ret);
}
/*************************************************************/
/* FIXED POINT CORE PITCH ANALYSIS FUNCTION */
/*************************************************************/
internal static int silk_pitch_analysis_core( /* O Voicing estimate: 0 voiced, 1 unvoiced */
short[] frame, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */
int[] pitch_out, /* O 4 pitch lag values */
BoxedValueShort lagIndex, /* O Lag Index */
BoxedValueSbyte contourIndex, /* O Pitch contour Index */
BoxedValueInt LTPCorr_Q15, /* I/O Normalized correlation; input: value from previous frame */
int prevLag, /* I Last lag of previous frame; set to zero is unvoiced */
int search_thres1_Q16, /* I First stage threshold for lag candidates 0 - 1 */
int search_thres2_Q13, /* I Final threshold for lag candidates 0 - 1 */
int Fs_kHz, /* I Sample frequency (kHz) */
int complexity, /* I Complexity setting, 0-2, where 2 is highest */
int nb_subfr /* I number of 5 ms subframes */
)
{
short[] frame_8kHz;
short[] frame_4kHz;
int[] filt_state = new int[6];
short[] input_frame_ptr;
int i, k, d, j;
short[] C;
int[] xcorr32;
short[] basis;
int basis_ptr;
short[] target;
int target_ptr;
int cross_corr, normalizer, energy, shift, energy_basis, energy_target;
int Cmax, length_d_srch, length_d_comp;
int[] d_srch = new int[SilkConstants.PE_D_SRCH_LENGTH];
short[] d_comp;
int sum, threshold, lag_counter;
int CBimax, CBimax_new, CBimax_old, lag, start_lag, end_lag, lag_new;
int CCmax, CCmax_b, CCmax_new_b, CCmax_new;
int[] CC = new int[SilkConstants.PE_NB_CBKS_STAGE2_EXT];
silk_pe_stage3_vals[] energies_st3;
silk_pe_stage3_vals[] cross_corr_st3;
int frame_length, frame_length_8kHz, frame_length_4kHz;
int sf_length;
int min_lag;
int max_lag;
int contour_bias_Q15, diff;
int nb_cbk_search;
int delta_lag_log2_sqr_Q7, lag_log2_Q7, prevLag_log2_Q7, prev_lag_bias_Q13;
sbyte[][] Lag_CB_ptr;
/* Check for valid sampling frequency */
Inlines.OpusAssert(Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16);
/* Check for valid complexity setting */
Inlines.OpusAssert(complexity >= SilkConstants.SILK_PE_MIN_COMPLEX);
Inlines.OpusAssert(complexity <= SilkConstants.SILK_PE_MAX_COMPLEX);
Inlines.OpusAssert(search_thres1_Q16 >= 0 && search_thres1_Q16 <= (1 << 16));
Inlines.OpusAssert(search_thres2_Q13 >= 0 && search_thres2_Q13 <= (1 << 13));
/* Set up frame lengths max / min lag for the sampling frequency */
frame_length = (SilkConstants.PE_LTP_MEM_LENGTH_MS + nb_subfr * SilkConstants.PE_SUBFR_LENGTH_MS) * Fs_kHz;
frame_length_4kHz = (SilkConstants.PE_LTP_MEM_LENGTH_MS + nb_subfr * SilkConstants.PE_SUBFR_LENGTH_MS) * 4;
frame_length_8kHz = (SilkConstants.PE_LTP_MEM_LENGTH_MS + nb_subfr * SilkConstants.PE_SUBFR_LENGTH_MS) * 8;
sf_length = SilkConstants.PE_SUBFR_LENGTH_MS * Fs_kHz;
min_lag = SilkConstants.PE_MIN_LAG_MS * Fs_kHz;
max_lag = SilkConstants.PE_MAX_LAG_MS * Fs_kHz - 1;
/* Resample from input sampled at Fs_kHz to 8 kHz */
frame_8kHz = new short[frame_length_8kHz];
if (Fs_kHz == 16)
{
Arrays.MemSetInt(filt_state, 0, 2);
Resampler.silk_resampler_down2(filt_state, frame_8kHz, frame, frame_length);
}
else if (Fs_kHz == 12)
{
Arrays.MemSetInt(filt_state, 0, 6);
Resampler.silk_resampler_down2_3(filt_state, frame_8kHz, frame, frame_length);
}
else
{
Inlines.OpusAssert(Fs_kHz == 8);
Array.Copy(frame, frame_8kHz, frame_length_8kHz);
}
/* Decimate again to 4 kHz */
Arrays.MemSetInt(filt_state, 0, 2); /* Set state to zero */
frame_4kHz = new short[frame_length_4kHz];
Resampler.silk_resampler_down2(filt_state, frame_4kHz, frame_8kHz, frame_length_8kHz);
/* Low-pass filter */
for (i = frame_length_4kHz - 1; i > 0; i--)
{
frame_4kHz[i] = Inlines.silk_ADD_SAT16(frame_4kHz[i], frame_4kHz[i - 1]);
}
/*******************************************************************************
** Scale 4 kHz signal down to prevent correlations measures from overflowing
** find scaling as max scaling for each 8kHz(?) subframe
*******************************************************************************/
/* Inner product is calculated with different lengths, so scale for the worst case */
SumSqrShift.silk_sum_sqr_shift(out energy, out shift, frame_4kHz, frame_length_4kHz);
if (shift > 0)
{
shift = Inlines.silk_RSHIFT(shift, 1);
for (i = 0; i < frame_length_4kHz; i++)
{
frame_4kHz[i] = Inlines.silk_RSHIFT16(frame_4kHz[i], shift);
}
}
/******************************************************************************
* FIRST STAGE, operating in 4 khz
******************************************************************************/
C = new short[nb_subfr * CSTRIDE_8KHZ];
xcorr32 = new int[MAX_LAG_4KHZ - MIN_LAG_4KHZ + 1];
Arrays.MemSetShort(C, 0, (nb_subfr >> 1) * CSTRIDE_4KHZ);
target = frame_4kHz;
target_ptr = Inlines.silk_LSHIFT(SF_LENGTH_4KHZ, 2);
for (k = 0; k < nb_subfr >> 1; k++)
{
basis = target;
basis_ptr = target_ptr - MIN_LAG_4KHZ;
CeltPitchXCorr.pitch_xcorr(target, target_ptr, target, target_ptr - MAX_LAG_4KHZ, xcorr32, SF_LENGTH_8KHZ, MAX_LAG_4KHZ - MIN_LAG_4KHZ + 1);
/* Calculate first vector products before loop */
cross_corr = xcorr32[MAX_LAG_4KHZ - MIN_LAG_4KHZ];
normalizer = Inlines.silk_inner_prod_self(target, target_ptr, SF_LENGTH_8KHZ);
normalizer = Inlines.silk_ADD32(normalizer, Inlines.silk_inner_prod_self(basis, basis_ptr, SF_LENGTH_8KHZ));
normalizer = Inlines.silk_ADD32(normalizer, Inlines.silk_SMULBB(SF_LENGTH_8KHZ, 4000));
Inlines.MatrixSet(C, k, 0, CSTRIDE_4KHZ,
(short)Inlines.silk_DIV32_varQ(cross_corr, normalizer, 13 + 1)); /* Q13 */
/* From now on normalizer is computed recursively */
for (d = MIN_LAG_4KHZ + 1; d <= MAX_LAG_4KHZ; d++)
{
basis_ptr--;
cross_corr = xcorr32[MAX_LAG_4KHZ - d];
/* Add contribution of new sample and remove contribution from oldest sample */
normalizer = Inlines.silk_ADD32(normalizer,
Inlines.silk_SMULBB(basis[basis_ptr], basis[basis_ptr]) -
Inlines.silk_SMULBB(basis[basis_ptr + SF_LENGTH_8KHZ], basis[basis_ptr + SF_LENGTH_8KHZ]));
Inlines.MatrixSet(C, k, d - MIN_LAG_4KHZ, CSTRIDE_4KHZ,
(short)Inlines.silk_DIV32_varQ(cross_corr, normalizer, 13 + 1)); /* Q13 */
}
/* Update target pointer */
target_ptr += SF_LENGTH_8KHZ;
}
/* Combine two subframes into single correlation measure and apply short-lag bias */
if (nb_subfr == SilkConstants.PE_MAX_NB_SUBFR)
{
for (i = MAX_LAG_4KHZ; i >= MIN_LAG_4KHZ; i--)
{
sum = (int)Inlines.MatrixGet(C, 0, i - MIN_LAG_4KHZ, CSTRIDE_4KHZ)
+ (int)Inlines.MatrixGet(C, 1, i - MIN_LAG_4KHZ, CSTRIDE_4KHZ); /* Q14 */
sum = Inlines.silk_SMLAWB(sum, sum, Inlines.silk_LSHIFT(-i, 4)); /* Q14 */
C[i - MIN_LAG_4KHZ] = (short)sum; /* Q14 */
}
}
else
{
/* Only short-lag bias */
for (i = MAX_LAG_4KHZ; i >= MIN_LAG_4KHZ; i--)
{
sum = Inlines.silk_LSHIFT((int)C[i - MIN_LAG_4KHZ], 1); /* Q14 */
sum = Inlines.silk_SMLAWB(sum, sum, Inlines.silk_LSHIFT(-i, 4)); /* Q14 */
C[i - MIN_LAG_4KHZ] = (short)sum; /* Q14 */
}
}
/* Sort */
length_d_srch = Inlines.silk_ADD_LSHIFT32(4, complexity, 1);
Inlines.OpusAssert(3 * length_d_srch <= SilkConstants.PE_D_SRCH_LENGTH);
Sort.silk_insertion_sort_decreasing_int16(C, d_srch, CSTRIDE_4KHZ, length_d_srch);
/* Escape if correlation is very low already here */
Cmax = (int)C[0]; /* Q14 */
if (Cmax < ((int)((0.2f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(0.2f, 14)*/)
{
Arrays.MemSetInt(pitch_out, 0, nb_subfr);
LTPCorr_Q15.Val = 0;
lagIndex.Val = 0;
contourIndex.Val = 0;
return(1);
}
threshold = Inlines.silk_SMULWB(search_thres1_Q16, Cmax);
for (i = 0; i < length_d_srch; i++)
{
/* Convert to 8 kHz indices for the sorted correlation that exceeds the threshold */
if (C[i] > threshold)
{
d_srch[i] = Inlines.silk_LSHIFT(d_srch[i] + MIN_LAG_4KHZ, 1);
}
else
{
length_d_srch = i;
break;
}
}
Inlines.OpusAssert(length_d_srch > 0);
d_comp = new short[D_COMP_STRIDE];
for (i = D_COMP_MIN; i < D_COMP_MAX; i++)
{
d_comp[i - D_COMP_MIN] = 0;
}
for (i = 0; i < length_d_srch; i++)
{
d_comp[d_srch[i] - D_COMP_MIN] = 1;
}
/* Convolution */
for (i = D_COMP_MAX - 1; i >= MIN_LAG_8KHZ; i--)
{
d_comp[i - D_COMP_MIN] += (short)(d_comp[i - 1 - D_COMP_MIN] + d_comp[i - 2 - D_COMP_MIN]);
}
length_d_srch = 0;
for (i = MIN_LAG_8KHZ; i < MAX_LAG_8KHZ + 1; i++)
{
if (d_comp[i + 1 - D_COMP_MIN] > 0)
{
d_srch[length_d_srch] = i;
length_d_srch++;
}
}
/* Convolution */
for (i = D_COMP_MAX - 1; i >= MIN_LAG_8KHZ; i--)
{
d_comp[i - D_COMP_MIN] += (short)(d_comp[i - 1 - D_COMP_MIN] + d_comp[i - 2 - D_COMP_MIN] + d_comp[i - 3 - D_COMP_MIN]);
}
length_d_comp = 0;
for (i = MIN_LAG_8KHZ; i < D_COMP_MAX; i++)
{
if (d_comp[i - D_COMP_MIN] > 0)
{
d_comp[length_d_comp] = (short)(i - 2);
length_d_comp++;
}
}
/**********************************************************************************
** SECOND STAGE, operating at 8 kHz, on lag sections with high correlation
*************************************************************************************/
/******************************************************************************
** Scale signal down to avoid correlations measures from overflowing
*******************************************************************************/
/* find scaling as max scaling for each subframe */
SumSqrShift.silk_sum_sqr_shift(out energy, out shift, frame_8kHz, frame_length_8kHz);
if (shift > 0)
{
shift = Inlines.silk_RSHIFT(shift, 1);
for (i = 0; i < frame_length_8kHz; i++)
{
frame_8kHz[i] = Inlines.silk_RSHIFT16(frame_8kHz[i], shift);
}
}
/*********************************************************************************
* Find energy of each subframe projected onto its history, for a range of delays
*********************************************************************************/
Arrays.MemSetShort(C, 0, nb_subfr * CSTRIDE_8KHZ);
target = frame_8kHz;
target_ptr = SilkConstants.PE_LTP_MEM_LENGTH_MS * 8;
for (k = 0; k < nb_subfr; k++)
{
energy_target = Inlines.silk_ADD32(Inlines.silk_inner_prod(target, target_ptr, target, target_ptr, SF_LENGTH_8KHZ), 1);
for (j = 0; j < length_d_comp; j++)
{
d = d_comp[j];
basis = target;
basis_ptr = target_ptr - d;
cross_corr = Inlines.silk_inner_prod(target, target_ptr, basis, basis_ptr, SF_LENGTH_8KHZ);
if (cross_corr > 0)
{
energy_basis = Inlines.silk_inner_prod_self(basis, basis_ptr, SF_LENGTH_8KHZ);
Inlines.MatrixSet(C, k, d - (MIN_LAG_8KHZ - 2), CSTRIDE_8KHZ,
(short)Inlines.silk_DIV32_varQ(cross_corr,
Inlines.silk_ADD32(energy_target,
energy_basis),
13 + 1)); /* Q13 */
}
else
{
Inlines.MatrixSet <short>(C, k, d - (MIN_LAG_8KHZ - 2), CSTRIDE_8KHZ, 0);
}
}
target_ptr += SF_LENGTH_8KHZ;
}
/* search over lag range and lags codebook */
/* scale factor for lag codebook, as a function of center lag */
CCmax = int.MinValue;
CCmax_b = int.MinValue;
CBimax = 0; /* To avoid returning undefined lag values */
lag = -1; /* To check if lag with strong enough correlation has been found */
if (prevLag > 0)
{
if (Fs_kHz == 12)
{
prevLag = Inlines.silk_DIV32_16(Inlines.silk_LSHIFT(prevLag, 1), 3);
}
else if (Fs_kHz == 16)
{
prevLag = Inlines.silk_RSHIFT(prevLag, 1);
}
prevLag_log2_Q7 = Inlines.silk_lin2log((int)prevLag);
}
else
{
prevLag_log2_Q7 = 0;
}
Inlines.OpusAssert(search_thres2_Q13 == Inlines.silk_SAT16(search_thres2_Q13));
/* Set up stage 2 codebook based on number of subframes */
if (nb_subfr == SilkConstants.PE_MAX_NB_SUBFR)
{
Lag_CB_ptr = Tables.silk_CB_lags_stage2;
if (Fs_kHz == 8 && complexity > SilkConstants.SILK_PE_MIN_COMPLEX)
{
/* If input is 8 khz use a larger codebook here because it is last stage */
nb_cbk_search = SilkConstants.PE_NB_CBKS_STAGE2_EXT;
}
else
{
nb_cbk_search = SilkConstants.PE_NB_CBKS_STAGE2;
}
}
else
{
Lag_CB_ptr = Tables.silk_CB_lags_stage2_10_ms;
nb_cbk_search = SilkConstants.PE_NB_CBKS_STAGE2_10MS;
}
for (k = 0; k < length_d_srch; k++)
{
d = d_srch[k];
for (j = 0; j < nb_cbk_search; j++)
{
CC[j] = 0;
for (i = 0; i < nb_subfr; i++)
{
int d_subfr;
/* Try all codebooks */
d_subfr = d + Lag_CB_ptr[i][j];
CC[j] = CC[j]
+ (int)Inlines.MatrixGet(C, i,
d_subfr - (MIN_LAG_8KHZ - 2),
CSTRIDE_8KHZ);
}
}
/* Find best codebook */
CCmax_new = int.MinValue;
CBimax_new = 0;
for (i = 0; i < nb_cbk_search; i++)
{
if (CC[i] > CCmax_new)
{
CCmax_new = CC[i];
CBimax_new = i;
}
}
/* Bias towards shorter lags */
lag_log2_Q7 = Inlines.silk_lin2log(d); /* Q7 */
Inlines.OpusAssert(lag_log2_Q7 == Inlines.silk_SAT16(lag_log2_Q7));
Inlines.OpusAssert(nb_subfr * ((int)((SilkConstants.PE_SHORTLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_SHORTLAG_BIAS, 13)*/ == Inlines.silk_SAT16(nb_subfr * ((int)((SilkConstants.PE_SHORTLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_SHORTLAG_BIAS, 13)*/));
CCmax_new_b = CCmax_new - Inlines.silk_RSHIFT(Inlines.silk_SMULBB(nb_subfr * ((int)((SilkConstants.PE_SHORTLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_SHORTLAG_BIAS, 13)*/, lag_log2_Q7), 7); /* Q13 */
/* Bias towards previous lag */
Inlines.OpusAssert(nb_subfr * ((int)((SilkConstants.PE_PREVLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_PREVLAG_BIAS, 13)*/ == Inlines.silk_SAT16(nb_subfr * ((int)((SilkConstants.PE_PREVLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_PREVLAG_BIAS, 13)*/));
if (prevLag > 0)
{
delta_lag_log2_sqr_Q7 = lag_log2_Q7 - prevLag_log2_Q7;
Inlines.OpusAssert(delta_lag_log2_sqr_Q7 == Inlines.silk_SAT16(delta_lag_log2_sqr_Q7));
delta_lag_log2_sqr_Q7 = Inlines.silk_RSHIFT(Inlines.silk_SMULBB(delta_lag_log2_sqr_Q7, delta_lag_log2_sqr_Q7), 7);
prev_lag_bias_Q13 = Inlines.silk_RSHIFT(Inlines.silk_SMULBB(nb_subfr * ((int)((SilkConstants.PE_PREVLAG_BIAS) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_PREVLAG_BIAS, 13)*/, LTPCorr_Q15.Val), 15); /* Q13 */
prev_lag_bias_Q13 = Inlines.silk_DIV32(Inlines.silk_MUL(prev_lag_bias_Q13, delta_lag_log2_sqr_Q7), delta_lag_log2_sqr_Q7 + ((int)((0.5f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(0.5f, 7)*/);
CCmax_new_b -= prev_lag_bias_Q13; /* Q13 */
}
if (CCmax_new_b > CCmax_b && /* Find maximum biased correlation */
CCmax_new > Inlines.silk_SMULBB(nb_subfr, search_thres2_Q13) && /* Correlation needs to be high enough to be voiced */
Tables.silk_CB_lags_stage2[0][CBimax_new] <= MIN_LAG_8KHZ /* Lag must be in range */
)
{
CCmax_b = CCmax_new_b;
CCmax = CCmax_new;
lag = d;
CBimax = CBimax_new;
}
}
if (lag == -1)
{
/* No suitable candidate found */
Arrays.MemSetInt(pitch_out, 0, nb_subfr);
LTPCorr_Q15.Val = 0;
lagIndex.Val = 0;
contourIndex.Val = 0;
return(1);
}
/* Output normalized correlation */
LTPCorr_Q15.Val = (int)Inlines.silk_LSHIFT(Inlines.silk_DIV32_16(CCmax, nb_subfr), 2);
Inlines.OpusAssert(LTPCorr_Q15.Val >= 0);
if (Fs_kHz > 8)
{
short[] scratch_mem;
/***************************************************************************/
/* Scale input signal down to avoid correlations measures from overflowing */
/***************************************************************************/
/* find scaling as max scaling for each subframe */
SumSqrShift.silk_sum_sqr_shift(out energy, out shift, frame, frame_length);
if (shift > 0)
{
scratch_mem = new short[frame_length];
/* Move signal to scratch mem because the input signal should be unchanged */
shift = Inlines.silk_RSHIFT(shift, 1);
for (i = 0; i < frame_length; i++)
{
scratch_mem[i] = Inlines.silk_RSHIFT16(frame[i], shift);
}
input_frame_ptr = scratch_mem;
}
else
{
input_frame_ptr = frame;
}
/* Search in original signal */
CBimax_old = CBimax;
/* Compensate for decimation */
Inlines.OpusAssert(lag == Inlines.silk_SAT16(lag));
if (Fs_kHz == 12)
{
lag = Inlines.silk_RSHIFT(Inlines.silk_SMULBB(lag, 3), 1);
}
else if (Fs_kHz == 16)
{
lag = Inlines.silk_LSHIFT(lag, 1);
}
else
{
lag = Inlines.silk_SMULBB(lag, 3);
}
lag = Inlines.silk_LIMIT_int(lag, min_lag, max_lag);
start_lag = Inlines.silk_max_int(lag - 2, min_lag);
end_lag = Inlines.silk_min_int(lag + 2, max_lag);
lag_new = lag; /* to avoid undefined lag */
CBimax = 0; /* to avoid undefined lag */
CCmax = int.MinValue;
/* pitch lags according to second stage */
for (k = 0; k < nb_subfr; k++)
{
pitch_out[k] = lag + 2 * Tables.silk_CB_lags_stage2[k][CBimax_old];
}
/* Set up codebook parameters according to complexity setting and frame length */
if (nb_subfr == SilkConstants.PE_MAX_NB_SUBFR)
{
nb_cbk_search = (int)Tables.silk_nb_cbk_searchs_stage3[complexity];
Lag_CB_ptr = Tables.silk_CB_lags_stage3;
}
else
{
nb_cbk_search = SilkConstants.PE_NB_CBKS_STAGE3_10MS;
Lag_CB_ptr = Tables.silk_CB_lags_stage3_10_ms;
}
/* Calculate the correlations and energies needed in stage 3 */
energies_st3 = new silk_pe_stage3_vals[nb_subfr * nb_cbk_search];
cross_corr_st3 = new silk_pe_stage3_vals[nb_subfr * nb_cbk_search];
for (int c = 0; c < nb_subfr * nb_cbk_search; c++)
{
energies_st3[c] = new silk_pe_stage3_vals(); // fixme: these can be replaced with a linearized array probably, or at least a struct
cross_corr_st3[c] = new silk_pe_stage3_vals();
}
silk_P_Ana_calc_corr_st3(cross_corr_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity);
silk_P_Ana_calc_energy_st3(energies_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity);
lag_counter = 0;
Inlines.OpusAssert(lag == Inlines.silk_SAT16(lag));
contour_bias_Q15 = Inlines.silk_DIV32_16(((int)((SilkConstants.PE_FLATCONTOUR_BIAS) * ((long)1 << (15)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.PE_FLATCONTOUR_BIAS, 15)*/, lag);
target = input_frame_ptr;
target_ptr = SilkConstants.PE_LTP_MEM_LENGTH_MS * Fs_kHz;
energy_target = Inlines.silk_ADD32(Inlines.silk_inner_prod_self(target, target_ptr, nb_subfr * sf_length), 1);
for (d = start_lag; d <= end_lag; d++)
{
for (j = 0; j < nb_cbk_search; j++)
{
cross_corr = 0;
energy = energy_target;
for (k = 0; k < nb_subfr; k++)
{
cross_corr = Inlines.silk_ADD32(cross_corr,
Inlines.MatrixGet(cross_corr_st3, k, j,
nb_cbk_search).Values[lag_counter]);
energy = Inlines.silk_ADD32(energy,
Inlines.MatrixGet(energies_st3, k, j,
nb_cbk_search).Values[lag_counter]);
Inlines.OpusAssert(energy >= 0);
}
if (cross_corr > 0)
{
CCmax_new = Inlines.silk_DIV32_varQ(cross_corr, energy, 13 + 1); /* Q13 */
/* Reduce depending on flatness of contour */
diff = short.MaxValue - Inlines.silk_MUL(contour_bias_Q15, j); /* Q15 */
Inlines.OpusAssert(diff == Inlines.silk_SAT16(diff));
CCmax_new = Inlines.silk_SMULWB(CCmax_new, diff); /* Q14 */
}
else
{
CCmax_new = 0;
}
if (CCmax_new > CCmax && (d + Tables.silk_CB_lags_stage3[0][j]) <= max_lag)
{
CCmax = CCmax_new;
lag_new = d;
CBimax = j;
}
}
lag_counter++;
}
for (k = 0; k < nb_subfr; k++)
{
pitch_out[k] = lag_new + Lag_CB_ptr[k][CBimax];
pitch_out[k] = Inlines.silk_LIMIT(pitch_out[k], min_lag, SilkConstants.PE_MAX_LAG_MS * Fs_kHz);
}
lagIndex.Val = (short)(lag_new - min_lag);
contourIndex.Val = (sbyte)CBimax;
}
else /* Fs_kHz == 8 */
/* Save Lags */
{
for (k = 0; k < nb_subfr; k++)
{
pitch_out[k] = lag + Lag_CB_ptr[k][CBimax];
pitch_out[k] = Inlines.silk_LIMIT(pitch_out[k], MIN_LAG_8KHZ, SilkConstants.PE_MAX_LAG_MS * 8);
}
lagIndex.Val = (short)(lag - MIN_LAG_8KHZ);
contourIndex.Val = (sbyte)CBimax;
}
Inlines.OpusAssert(lagIndex.Val >= 0);
/* return as voiced */
return(0);
}
