Revert "Update opus to 1.3.1 and opusfile to 0.11"

This reverts commit e00426c512a7905f5f925d382c443bab7a0ca693.

The way we handle platform-specific intrinsics is not good, so the
current state will not compile on armv8. This commit also requires
SSE4.1 support, which is likely not a good idea for portable binaries.

We'll have to redo this with more caution after 3.2 is released, or
we might simply drop opus as we're only using it as dependency for
theora right now.

Fixes #33606.

6 files changed, 71 insertions, 33 deletions

diff --git a/thirdparty/opus/silk/x86/NSQ_del_dec_sse4_1.c b/thirdparty/opus/silk/x86/NSQ_del_dec_sse.c
index 2c75ede2dd..21d4a8bc1e 100644
--- a/thirdparty/opus/silk/x86/NSQ_del_dec_sse4_1.c
+++ b/thirdparty/opus/silk/x86/NSQ_del_dec_sse.c
@@ -107,12 +107,12 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
     opus_int            predictLPCOrder,        /* I    Prediction filter order             */
     opus_int            warping_Q16,            /* I                                        */
     opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
-    opus_int            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
+    opus_int            *smpl_buf_idx,          /* I    Index to newest samples in buffers  */
     opus_int            decisionDelay           /* I                                        */
 );
 
 void silk_NSQ_del_dec_sse4_1(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -234,8 +234,7 @@ void silk_NSQ_del_dec_sse4_1(
                     psDD = &psDelDec[ Winner_ind ];
                     last_smple_idx = smpl_buf_idx + decisionDelay;
                     for( i = 0; i < decisionDelay; i++ ) {
-                        last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
-                        if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
+                        last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
                         pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
                         pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
                             silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) );
@@ -247,7 +246,7 @@ void silk_NSQ_del_dec_sse4_1(
 
                 /* Rewhiten with new A coefs */
                 start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
-                celt_assert( start_idx > 0 );
+                silk_assert( start_idx > 0 );
 
                 silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
                     A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
@@ -286,8 +285,7 @@ void silk_NSQ_del_dec_sse4_1(
     last_smple_idx = smpl_buf_idx + decisionDelay;
     Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
     for( i = 0; i < decisionDelay; i++ ) {
-        last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
-        if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
+        last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
         pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
         pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
             silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) );
@@ -301,6 +299,7 @@ void silk_NSQ_del_dec_sse4_1(
     NSQ->lagPrev        = pitchL[ psEncC->nb_subfr - 1 ];
 
     /* Save quantized speech signal */
+    /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */
     silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
     silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
     RESTORE_STACK;
@@ -334,7 +333,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
     opus_int            predictLPCOrder,        /* I    Prediction filter order             */
     opus_int            warping_Q16,            /* I                                        */
     opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
-    opus_int            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
+    opus_int            *smpl_buf_idx,          /* I    Index to newest samples in buffers  */
     opus_int            decisionDelay           /* I                                        */
 )
 {
@@ -353,7 +352,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
     __m128i b_Q12_0123, b_sr_Q12_0123;
     SAVE_STACK;
 
-    celt_assert( nStatesDelayedDecision > 0 );
+    silk_assert( nStatesDelayedDecision > 0 );
     ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair );
 
     shp_lag_ptr  = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
@@ -639,9 +638,8 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
                 psSS[ 1 ].xq_Q14       = xq_Q14;
             }
         }
-        *smpl_buf_idx  = ( *smpl_buf_idx - 1 ) % DECISION_DELAY;
-        if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY;
-        last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY;
+        *smpl_buf_idx  = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK;                   /* Index to newest samples              */
+        last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK;       /* Index to decisionDelay old samples   */
 
         /* Find winner */
         RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
diff --git a/thirdparty/opus/silk/x86/NSQ_sse4_1.c b/thirdparty/opus/silk/x86/NSQ_sse.c
index b0315e35fc..bb3c5f1955 100644
--- a/thirdparty/opus/silk/x86/NSQ_sse4_1.c
+++ b/thirdparty/opus/silk/x86/NSQ_sse.c
@@ -71,7 +71,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1(
 );
 
 void silk_NSQ_sse4_1(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -199,7 +199,7 @@ void silk_NSQ_sse4_1(
             if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
                 /* Rewhiten with new A coefs */
                 start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
-                celt_assert( start_idx > 0 );
+                silk_assert( start_idx > 0 );
 
                 silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
                     A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
@@ -233,6 +233,7 @@ void silk_NSQ_sse4_1(
     NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ];
 
     /* Save quantized speech and noise shaping signals */
+    /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */
     silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
     silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
     RESTORE_STACK;
diff --git a/thirdparty/opus/silk/x86/VAD_sse4_1.c b/thirdparty/opus/silk/x86/VAD_sse.c
index d02ddf4ad0..4e90f4410d 100644
--- a/thirdparty/opus/silk/x86/VAD_sse4_1.c
+++ b/thirdparty/opus/silk/x86/VAD_sse.c
@@ -65,9 +65,9 @@ opus_int silk_VAD_GetSA_Q8_sse4_1(                  /* O    Return value, 0 if s
 
     /* Safety checks */
     silk_assert( VAD_N_BANDS == 4 );
-    celt_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
-    celt_assert( psEncC->frame_length <= 512 );
-    celt_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
+    silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
+    silk_assert( psEncC->frame_length <= 512 );
+    silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
 
     /***********************/
     /* Filter and Decimate */
diff --git a/thirdparty/opus/silk/x86/VQ_WMat_EC_sse4_1.c b/thirdparty/opus/silk/x86/VQ_WMat_EC_sse.c
index 74d6c6d0ec..74d6c6d0ec 100644
--- a/thirdparty/opus/silk/x86/VQ_WMat_EC_sse4_1.c
+++ b/thirdparty/opus/silk/x86/VQ_WMat_EC_sse.c
diff --git a/thirdparty/opus/silk/x86/main_sse.h b/thirdparty/opus/silk/x86/main_sse.h
index 2f15d44869..d8d61310ed 100644
--- a/thirdparty/opus/silk/x86/main_sse.h
+++ b/thirdparty/opus/silk/x86/main_sse.h
@@ -34,7 +34,6 @@
 
 # if defined(OPUS_X86_MAY_HAVE_SSE4_1)
 
-#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */
 #  define OVERRIDE_silk_VQ_WMat_EC
 
 void silk_VQ_WMat_EC_sse4_1(
@@ -80,13 +79,11 @@ extern void (*const SILK_VQ_WMAT_EC_IMPL[OPUS_ARCHMASK + 1])(
                           mu_Q9, max_gain_Q7, L))
 
 #endif
-#endif
 
-#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */
 #  define OVERRIDE_silk_NSQ
 
 void silk_NSQ_sse4_1(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -113,7 +110,7 @@ void silk_NSQ_sse4_1(
 #else
 
 extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -140,7 +137,7 @@ 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    Encoder State                   */
+    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        */
@@ -167,7 +164,7 @@ void silk_NSQ_del_dec_sse4_1(
 #else
 
 extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -190,7 +187,6 @@ extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
                            HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
 
 #endif
-#endif
 
 void silk_noise_shape_quantizer(
     silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
@@ -242,6 +238,39 @@ 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
diff --git a/thirdparty/opus/silk/x86/x86_silk_map.c b/thirdparty/opus/silk/x86/x86_silk_map.c
index 32dcc3cab7..818841f2c1 100644
--- a/thirdparty/opus/silk/x86/x86_silk_map.c
+++ b/thirdparty/opus/silk/x86/x86_silk_map.c
@@ -66,9 +66,8 @@ opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )(
   MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 )  /* avx */
 };
 
-#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */
 void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -90,9 +89,7 @@ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
   MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */
   MAY_HAVE_SSE4_1( silk_NSQ )  /* avx */
 };
-#endif
 
-#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */
 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       */
@@ -112,11 +109,9 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
   MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */
   MAY_HAVE_SSE4_1( silk_VQ_WMat_EC )  /* avx */
 };
-#endif
 
-#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */
 void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
-    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    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        */
@@ -138,10 +133,25 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
   MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */
   MAY_HAVE_SSE4_1( silk_NSQ_del_dec )  /* avx */
 };
-#endif
 
 #if defined(FIXED_POINT)
 
+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)                 */
+) = {
+  silk_warped_LPC_analysis_filter_FIX_c,                  /* non-sse */
+  silk_warped_LPC_analysis_filter_FIX_c,
+  silk_warped_LPC_analysis_filter_FIX_c,
+  MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX )  /* avx */
+};
+
 void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
     opus_int32                  *res_nrg,           /* O    Residual energy                                             */
     opus_int                    *res_nrg_Q,         /* O    Residual energy Q value                                     */