summaryrefslogtreecommitdiff
path: root/media/libopus/silk/x86/main_sse.h
diff options
context:
space:
mode:
Diffstat (limited to 'media/libopus/silk/x86/main_sse.h')
-rw-r--r--media/libopus/silk/x86/main_sse.h199
1 files changed, 84 insertions, 115 deletions
diff --git a/media/libopus/silk/x86/main_sse.h b/media/libopus/silk/x86/main_sse.h
index d8d61310ed..0a0391a2da 100644
--- a/media/libopus/silk/x86/main_sse.h
+++ b/media/libopus/silk/x86/main_sse.h
@@ -38,66 +38,68 @@
void silk_VQ_WMat_EC_sse4_1(
opus_int8 *ind, /* O index of best codebook vector */
- opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
+ opus_int32 *res_nrg_Q15, /* O best residual energy */
+ opus_int32 *rate_dist_Q8, /* O best total bitrate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
- const opus_int16 *in_Q14, /* I input vector to be quantized */
- const opus_int32 *W_Q18, /* I weighting matrix */
+ const opus_int32 *XX_Q17, /* I correlation matrix */
+ const opus_int32 *xX_Q17, /* I correlation vector */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
- const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
+ const opus_int subfr_len, /* I number of samples per subframe */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
- opus_int L /* I number of vectors in codebook */
+ const opus_int L /* I number of vectors in codebook */
);
#if defined OPUS_X86_PRESUME_SSE4_1
-#define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
- mu_Q9, max_gain_Q7, L, arch) \
- ((void)(arch),silk_VQ_WMat_EC_sse4_1(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
- mu_Q9, max_gain_Q7, L))
+#define silk_VQ_WMat_EC(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, \
+ subfr_len, max_gain_Q7, L, arch) \
+ ((void)(arch),silk_VQ_WMat_EC_sse4_1(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, \
+ subfr_len, max_gain_Q7, L))
#else
extern void (*const SILK_VQ_WMAT_EC_IMPL[OPUS_ARCHMASK + 1])(
opus_int8 *ind, /* O index of best codebook vector */
- opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
+ opus_int32 *res_nrg_Q15, /* O best residual energy */
+ opus_int32 *rate_dist_Q8, /* O best total bitrate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
- const opus_int16 *in_Q14, /* I input vector to be quantized */
- const opus_int32 *W_Q18, /* I weighting matrix */
+ const opus_int32 *XX_Q17, /* I correlation matrix */
+ const opus_int32 *xX_Q17, /* I correlation vector */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
- const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
+ const opus_int subfr_len, /* I number of samples per subframe */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
- opus_int L /* I number of vectors in codebook */
+ const opus_int L /* I number of vectors in codebook */
);
-# define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
- mu_Q9, max_gain_Q7, L, arch) \
- ((*SILK_VQ_WMAT_EC_IMPL[(arch) & OPUS_ARCHMASK])(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
- mu_Q9, max_gain_Q7, L))
+# define silk_VQ_WMat_EC(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, \
+ subfr_len, max_gain_Q7, L, arch) \
+ ((*SILK_VQ_WMAT_EC_IMPL[(arch) & OPUS_ARCHMASK])(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, \
+ subfr_len, max_gain_Q7, L))
#endif
# define OVERRIDE_silk_NSQ
void silk_NSQ_sse4_1(
- const silk_encoder_state *psEncC, /* I/O Encoder State */
- silk_nsq_state *NSQ, /* I/O NSQ state */
- SideInfoIndices *psIndices, /* I/O Quantization Indices */
- const opus_int32 x_Q3[], /* I Prefiltered input signal */
- opus_int8 pulses[], /* O Quantized pulse signal */
- const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
- const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
- const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
- const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
- const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
- const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
- const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
- const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
- const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
- const opus_int LTP_scale_Q14 /* I LTP state scaling */
+ const silk_encoder_state *psEncC, /* I Encoder State */
+ silk_nsq_state *NSQ, /* I/O NSQ state */
+ SideInfoIndices *psIndices, /* I/O Quantization Indices */
+ const opus_int16 x16[], /* I Input */
+ opus_int8 pulses[], /* O Quantized pulse signal */
+ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
+ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
+ const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
+ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
+ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
+ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
+ const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
+ const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
+ const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
+ const opus_int LTP_scale_Q14 /* I LTP state scaling */
);
#if defined OPUS_X86_PRESUME_SSE4_1
@@ -110,21 +112,21 @@ void silk_NSQ_sse4_1(
#else
extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])(
- const silk_encoder_state *psEncC, /* I/O Encoder State */
- silk_nsq_state *NSQ, /* I/O NSQ state */
- SideInfoIndices *psIndices, /* I/O Quantization Indices */
- const opus_int32 x_Q3[], /* I Prefiltered input signal */
- opus_int8 pulses[], /* O Quantized pulse signal */
- const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
- const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
- const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
- const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
- const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
- const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
- const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
- const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
- const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
- const opus_int LTP_scale_Q14 /* I LTP state scaling */
+ const silk_encoder_state *psEncC, /* I Encoder State */
+ silk_nsq_state *NSQ, /* I/O NSQ state */
+ SideInfoIndices *psIndices, /* I/O Quantization Indices */
+ const opus_int16 x16[], /* I Input */
+ opus_int8 pulses[], /* O Quantized pulse signal */
+ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
+ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
+ const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
+ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
+ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
+ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
+ const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
+ const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
+ const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
+ const opus_int LTP_scale_Q14 /* I LTP state scaling */
);
# define silk_NSQ(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
@@ -137,53 +139,53 @@ extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])(
# define OVERRIDE_silk_NSQ_del_dec
void silk_NSQ_del_dec_sse4_1(
- const silk_encoder_state *psEncC, /* I/O Encoder State */
- silk_nsq_state *NSQ, /* I/O NSQ state */
- SideInfoIndices *psIndices, /* I/O Quantization Indices */
- const opus_int32 x_Q3[], /* I Prefiltered input signal */
- opus_int8 pulses[], /* O Quantized pulse signal */
- const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
- const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
- const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
- const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
- const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
- const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
- const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
- const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
- const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
- const opus_int LTP_scale_Q14 /* I LTP state scaling */
+ const silk_encoder_state *psEncC, /* I Encoder State */
+ silk_nsq_state *NSQ, /* I/O NSQ state */
+ SideInfoIndices *psIndices, /* I/O Quantization Indices */
+ const opus_int16 x16[], /* I Input */
+ opus_int8 pulses[], /* O Quantized pulse signal */
+ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
+ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
+ const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
+ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
+ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
+ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
+ const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
+ const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
+ const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
+ const opus_int LTP_scale_Q14 /* I LTP state scaling */
);
#if defined OPUS_X86_PRESUME_SSE4_1
-#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \
HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
- ((void)(arch),silk_NSQ_del_dec_sse4_1(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+ ((void)(arch),silk_NSQ_del_dec_sse4_1(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \
HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
#else
extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
- const silk_encoder_state *psEncC, /* I/O Encoder State */
- silk_nsq_state *NSQ, /* I/O NSQ state */
- SideInfoIndices *psIndices, /* I/O Quantization Indices */
- const opus_int32 x_Q3[], /* I Prefiltered input signal */
- opus_int8 pulses[], /* O Quantized pulse signal */
- const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
- const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
- const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
- const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
- const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
- const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
- const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
- const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
- const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
- const opus_int LTP_scale_Q14 /* I LTP state scaling */
+ const silk_encoder_state *psEncC, /* I Encoder State */
+ silk_nsq_state *NSQ, /* I/O NSQ state */
+ SideInfoIndices *psIndices, /* I/O Quantization Indices */
+ const opus_int16 x16[], /* I Input */
+ opus_int8 pulses[], /* O Quantized pulse signal */
+ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
+ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
+ const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
+ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
+ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
+ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
+ const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
+ const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
+ const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
+ const opus_int LTP_scale_Q14 /* I LTP state scaling */
);
-# define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+# define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \
HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
- ((*SILK_NSQ_DEL_DEC_IMPL[(arch) & OPUS_ARCHMASK])(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+ ((*SILK_NSQ_DEL_DEC_IMPL[(arch) & OPUS_ARCHMASK])(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \
HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
#endif
@@ -238,39 +240,6 @@ extern opus_int (*const SILK_VAD_GETSA_Q8_IMPL[OPUS_ARCHMASK + 1])(
silk_encoder_state *psEnC,
const opus_int16 pIn[]);
-# define OVERRIDE_silk_warped_LPC_analysis_filter_FIX
-
-#endif
-
-void silk_warped_LPC_analysis_filter_FIX_sse4_1(
- opus_int32 state[], /* I/O State [order + 1] */
- opus_int32 res_Q2[], /* O Residual signal [length] */
- const opus_int16 coef_Q13[], /* I Coefficients [order] */
- const opus_int16 input[], /* I Input signal [length] */
- const opus_int16 lambda_Q16, /* I Warping factor */
- const opus_int length, /* I Length of input signal */
- const opus_int order /* I Filter order (even) */
-);
-
-#if defined(OPUS_X86_PRESUME_SSE4_1)
-#define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \
- ((void)(arch),silk_warped_LPC_analysis_filter_FIX_c(state, res_Q2, coef_Q13, input, lambda_Q16, length, order))
-
-#else
-
-extern void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[OPUS_ARCHMASK + 1])(
- opus_int32 state[], /* I/O State [order + 1] */
- opus_int32 res_Q2[], /* O Residual signal [length] */
- const opus_int16 coef_Q13[], /* I Coefficients [order] */
- const opus_int16 input[], /* I Input signal [length] */
- const opus_int16 lambda_Q16, /* I Warping factor */
- const opus_int length, /* I Length of input signal */
- const opus_int order /* I Filter order (even) */
-);
-
-# define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \
- ((*SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[(arch) & OPUS_ARCHMASK])(state, res_Q2, coef_Q13, input, lambda_Q16, length, order))
-
#endif
# endif
generated by cgit v1.2.3 (git 2.26.2) at 2024-11-09 17:05:17 +0000