Optimization of AMR NB and WB decoders for MIPS

AMR NB and WB decoders are optimized for MIPS architecture.
Appropriate Makefiles are changed accordingly.

Cnfigure script is changed in order to support optimizations.
 Optimizations are enabled by default when compiling is done for
  mips architecture.
 Appropriate cflags are automatically set.
 Support for several mips CPUs is added in configure script.

New ffmpeg options are added for disabling optimizations.

The FFMPEG option --disable-mipsfpu disables MIPS floating point
 optimizations.
The FFMPEG option --disable-mips32r2 disables MIPS32R2
 optimizations.
The FFMPEG option --disable-mipsdspr1 disables MIPS DSP ASE R1
 optimizations.
The FFMPEG option --disable-mipsdspr2 disables MIPS DSP ASE R2
 optimizations.

Signed-off-by: Nedeljko Babic <nbabic@mips.com>
Reviewed-by: Vitor Sessak <vitor1001@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

Makefile

@@ -70,6 +70,7 @@ SUBDIR_VARS := CLEANFILES EXAMPLES FFLIBS HOSTPROGS TESTPROGS TOOLS      \
                ARCH_HEADERS BUILT_HEADERS SKIPHEADERS                    \
                ALTIVEC-OBJS ARMV5TE-OBJS ARMV6-OBJS ARMVFP-OBJS MMI-OBJS \
                MMX-OBJS NEON-OBJS VIS-OBJS YASM-OBJS                     \
+               MIPSFPU-OBJS MIPSDSPR2-OBJS MIPSDSPR1-OBJS MIPS32R2-OBJS  \
                OBJS TESTOBJS
 
 define RESET

arch.mak

@@ -4,6 +4,10 @@ OBJS-$(HAVE_ARMVFP)  += $(ARMVFP-OBJS)  $(ARMVFP-OBJS-yes)
 OBJS-$(HAVE_NEON)    += $(NEON-OBJS)    $(NEON-OBJS-yes)
 
 OBJS-$(HAVE_MMI)     += $(MMI-OBJS)     $(MMI-OBJS-yes)
+OBJS-$(HAVE_MIPSFPU)   += $(MIPSFPU-OBJS)    $(MIPSFPU-OBJS-yes)
+OBJS-$(HAVE_MIPS32R2)  += $(MIPS32R2-OBJS)   $(MIPS32R2-OBJS-yes)
+OBJS-$(HAVE_MIPSDSPR1) += $(MIPSDSPR1-OBJS)  $(MIPSDSPR1-OBJS-yes)
+OBJS-$(HAVE_MIPSDSPR2) += $(MIPSDSPR2-OBJS)  $(MIPSDSPR2-OBJS-yes)
 
 OBJS-$(HAVE_ALTIVEC) += $(ALTIVEC-OBJS) $(ALTIVEC-OBJS-yes)
 

configure

@@ -268,6 +268,10 @@ Optimization options (experts only):
   --disable-neon           disable NEON optimizations
   --disable-vis            disable VIS optimizations
   --disable-yasm           disable use of yasm assembler
+  --disable-mips32r2       disable MIPS32R2 optimizations
+  --disable-mipsdspr1      disable MIPS DSP ASE R1 optimizations
+  --disable-mipsdspr2      disable MIPS DSP ASE R2 optimizations
+  --disable-mipsfpu        disable floating point MIPS optimizations
   --postproc-version=V     build libpostproc version V.
                            Where V can be '$ALT_PP_VER_MAJOR.$ALT_PP_VER_MINOR.$ALT_PP_VER_MICRO' or 'current'. [$postproc_version_default]
 
@@ -1149,6 +1153,10 @@ ARCH_EXT_LIST='
     ssse3
     vfpv3
     vis
+    mipsfpu
+    mips32r2
+    mipsdspr1
+    mipsdspr2
 '
 
 HAVE_LIST_PUB='
@@ -1368,6 +1376,10 @@ armvfp_deps="arm"
 neon_deps="arm"
 vfpv3_deps="armvfp"
 
+mipsfpu_deps="mips"
+mips32r2_deps="mips"
+mipsdspr1_deps="mips"
+mipsdspr2_deps="mips"
 mmi_deps="mips"
 
 altivec_deps="ppc"
@@ -2567,6 +2579,28 @@ elif enabled mips; then
 
     cpuflags="-march=$cpu"
 
+    case $cpu in
+        24kc)
+            disable mipsfpu
+            disable mipsdspr1
+            disable mipsdspr2
+        ;;
+        24kf*)
+            disable mipsdspr1
+            disable mipsdspr2
+        ;;
+        24kec|34kc|1004kc)
+            disable mipsfpu
+            disable mipsdspr2
+        ;;
+        24kef*|34kf*|1004kf*)
+            disable mipsdspr2
+        ;;
+        74kc)
+            disable mipsfpu
+        ;;
+    esac
+
 elif enabled avr32; then
 
     case $cpu in
@@ -2942,6 +2976,15 @@ elif enabled mips; then
 
     check_asm loongson '"dmult.g $1, $2, $3"'
     enabled mmi     && check_asm mmi     '"lq $2, 0($2)"'
+    enabled mips32r2  && add_cflags "-mips32r2" &&
+     check_asm mips32r2  '"rotr $t0, $t1, 1"'
+    enabled mipsdspr1 && add_cflags "-mdsp" && add_asflags "-mdsp" &&
+     check_asm mipsdspr1 '"addu.qb $t0, $t1, $t2"'
+    enabled mipsdspr2 && add_cflags "-mdspr2" && add_asflags "-mdspr2" &&
+     check_asm mipsdspr2 '"absq_s.qb $t0, $t1"'
+    enabled mipsfpu   && add_cflags "-mhard-float" &&
+     check_asm mipsfpu   '"madd.d $f0, $f2, $f4, $f6"'
+
 
 elif enabled ppc; then
 
@@ -3541,6 +3584,10 @@ if enabled arm; then
 fi
 if enabled mips; then
     echo "MMI enabled               ${mmi-no}"
+    echo "MIPS FPU enabled          ${mipsfpu-no}"
+    echo "MIPS32R2 enabled          ${mips32r2-no}"
+    echo "MIPS DSP R1 enabled       ${mipsdspr1-no}"
+    echo "MIPS DSP R2 enabled       ${mipsdspr2-no}"
 fi
 if enabled ppc; then
     echo "AltiVec enabled           ${altivec-no}"

libavcodec/acelp_filters.c

@@ -142,3 +142,12 @@ void ff_tilt_compensation(float *mem, float tilt, float *samples, int size)
     samples[0] -= tilt * *mem;
     *mem = new_tilt_mem;
 }
+
+void ff_acelp_filter_init(ACELPFContext *c)
+{
+    c->acelp_interpolatef                      = ff_acelp_interpolatef;
+    c->acelp_apply_order_2_transfer_function   = ff_acelp_apply_order_2_transfer_function;
+
+    if(HAVE_MIPSFPU)
+        ff_acelp_filter_init_mips(c);
+}

libavcodec/acelp_filters.h

@@ -25,6 +25,39 @@
 
 #include <stdint.h>
 
+typedef struct ACELPFContext {
+    /**
+    * Floating point version of ff_acelp_interpolate()
+    */
+    void (*acelp_interpolatef)(float *out, const float *in,
+                            const float *filter_coeffs, int precision,
+                            int frac_pos, int filter_length, int length);
+
+    /**
+     * Apply an order 2 rational transfer function in-place.
+     *
+     * @param out output buffer for filtered speech samples
+     * @param in input buffer containing speech data (may be the same as out)
+     * @param zero_coeffs z^-1 and z^-2 coefficients of the numerator
+     * @param pole_coeffs z^-1 and z^-2 coefficients of the denominator
+     * @param gain scale factor for final output
+     * @param mem intermediate values used by filter (should be 0 initially)
+     * @param n number of samples (should be a multiple of eight)
+     */
+    void (*acelp_apply_order_2_transfer_function)(float *out, const float *in,
+                                                  const float zero_coeffs[2],
+                                                  const float pole_coeffs[2],
+                                                  float gain,
+                                                  float mem[2], int n);
+
+}ACELPFContext;
+
+/**
+ * Initialize ACELPFContext.
+ */
+void ff_acelp_filter_init(ACELPFContext *c);
+void ff_acelp_filter_init_mips(ACELPFContext *c);
+
 /**
  * low-pass Finite Impulse Response filter coefficients.
  *

libavcodec/acelp_vectors.c

@@ -260,3 +260,11 @@ void ff_clear_fixed_vector(float *out, const AMRFixed *in, int size)
             } while (x < size && repeats);
     }
 }
+
+void ff_acelp_vectors_init(ACELPVContext *c)
+{
+    c->weighted_vector_sumf   = ff_weighted_vector_sumf;
+
+    if(HAVE_MIPSFPU)
+        ff_acelp_vectors_init_mips(c);
+}

libavcodec/acelp_vectors.h

@@ -25,6 +25,30 @@
 
 #include <stdint.h>
 
+typedef struct ACELPVContext {
+    /**
+     * float implementation of weighted sum of two vectors.
+     * @param[out] out result of addition
+     * @param in_a first vector
+     * @param in_b second vector
+     * @param weight_coeff_a first vector weight coefficient
+     * @param weight_coeff_a second vector weight coefficient
+     * @param length vectors length (should be a multiple of two)
+     *
+     * @note It is safe to pass the same buffer for out and in_a or in_b.
+     */
+    void (*weighted_vector_sumf)(float *out, const float *in_a, const float *in_b,
+                                 float weight_coeff_a, float weight_coeff_b,
+                                 int length);
+
+}ACELPVContext;
+
+/**
+ * Initialize ACELPVContext.
+ */
+void ff_acelp_vectors_init(ACELPVContext *c);
+void ff_acelp_vectors_init_mips(ACELPVContext *c);
+
 /** Sparse representation for the algebraic codebook (fixed) vector */
 typedef struct {
     int      n;

libavcodec/amrnbdec.c

@@ -136,6 +136,11 @@ typedef struct AMRContext {
 
     float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
 
+    ACELPFContext                     acelpf_ctx; ///< context for filters for ACELP-based codecs
+    ACELPVContext                     acelpv_ctx; ///< context for vector operations for ACELP-based codecs
+    CELPFContext                       celpf_ctx; ///< context for filters for CELP-based codecs
+    CELPMContext                       celpm_ctx; ///< context for fixed point math operations
+
 } AMRContext;
 
 /** Double version of ff_weighted_vector_sumf() */
@@ -171,6 +176,11 @@ static av_cold int amrnb_decode_init(AVCodecContext *avctx)
     avcodec_get_frame_defaults(&p->avframe);
     avctx->coded_frame = &p->avframe;
 
+    ff_acelp_filter_init(&p->acelpf_ctx);
+    ff_acelp_vectors_init(&p->acelpv_ctx);
+    ff_celp_filter_init(&p->celpf_ctx);
+    ff_celp_math_init(&p->celpm_ctx);
+
     return 0;
 }
 
@@ -214,15 +224,16 @@ static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
  * Interpolate the LSF vector (used for fixed gain smoothing).
  * The interpolation is done over all four subframes even in MODE_12k2.
  *
+ * @param[in]     ctx       The Context
  * @param[in,out] lsf_q     LSFs in [0,1] for each subframe
  * @param[in]     lsf_new   New LSFs in [0,1] for subframe 4
  */
-static void interpolate_lsf(float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
+static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
 {
     int i;
 
     for (i = 0; i < 4; i++)
-        ff_weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
+        ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
                                 0.25 * (3 - i), 0.25 * (i + 1),
                                 LP_FILTER_ORDER);
 }
@@ -266,7 +277,7 @@ static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
     ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
 
     if (update)
-        interpolate_lsf(p->lsf_q, lsf_q);
+        interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
 
     ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
 }
@@ -329,7 +340,7 @@ static void lsf2lsp_3(AMRContext *p)
     ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
 
     // store data for computing the next frame's LSFs
-    interpolate_lsf(p->lsf_q, lsf_q);
+    interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
     memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
 
     ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
@@ -395,7 +406,8 @@ static void decode_pitch_vector(AMRContext *p,
 
     /* Calculate the pitch vector by interpolating the past excitation at the
        pitch lag using a b60 hamming windowed sinc function.   */
-    ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int,
+    p->acelpf_ctx.acelp_interpolatef(p->excitation,
+                          p->excitation + 1 - pitch_lag_int,
                           ff_b60_sinc, 6,
                           pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
                           10, AMR_SUBFRAME_SIZE);
@@ -780,12 +792,12 @@ static int synthesis(AMRContext *p, float *lpc,
         for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
             p->pitch_vector[i] *= 0.25;
 
-    ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
+    p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
                             p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
 
     // emphasize pitch vector contribution
     if (p->pitch_gain[4] > 0.5 && !overflow) {
-        float energy = ff_dot_productf(excitation, excitation,
+        float energy = p->celpm_ctx.dot_productf(excitation, excitation,
                                        AMR_SUBFRAME_SIZE);
         float pitch_factor =
             p->pitch_gain[4] *
@@ -800,7 +812,8 @@ static int synthesis(AMRContext *p, float *lpc,
                                                 AMR_SUBFRAME_SIZE);
     }
 
-    ff_celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE,
+    p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
+                                 AMR_SUBFRAME_SIZE,
                                  LP_FILTER_ORDER);
 
     // detect overflow
@@ -846,10 +859,11 @@ static void update_state(AMRContext *p)
 /**
  * Get the tilt factor of a formant filter from its transfer function
  *
+ * @param p     The Context
  * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
  * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
  */
-static float tilt_factor(float *lpc_n, float *lpc_d)
+static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d)
 {
     float rh0, rh1; // autocorrelation at lag 0 and 1
 
@@ -859,11 +873,12 @@ static float tilt_factor(float *lpc_n, float *lpc_d)
 
     hf[0] = 1.0;
     memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
-    ff_celp_lp_synthesis_filterf(hf, lpc_d, hf, AMR_TILT_RESPONSE,
+    p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf,
+                                 AMR_TILT_RESPONSE,
                                  LP_FILTER_ORDER);
 
-    rh0 = ff_dot_productf(hf, hf,     AMR_TILT_RESPONSE);
-    rh1 = ff_dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
+    rh0 = p->celpm_ctx.dot_productf(hf, hf,     AMR_TILT_RESPONSE);
+    rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
 
     // The spec only specifies this check for 12.2 and 10.2 kbit/s
     // modes. But in the ref source the tilt is always non-negative.
@@ -883,7 +898,7 @@ static void postfilter(AMRContext *p, float *lpc, float *buf_out)
     int i;
     float *samples          = p->samples_in + LP_FILTER_ORDER; // Start of input
 
-    float speech_gain       = ff_dot_productf(samples, samples,
+    float speech_gain       = p->celpm_ctx.dot_productf(samples, samples,
                                               AMR_SUBFRAME_SIZE);
 
     float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER];  // Output of pole filter
@@ -904,16 +919,16 @@ static void postfilter(AMRContext *p, float *lpc, float *buf_out)
     }
 
     memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
-    ff_celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
+    p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
                                  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
     memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
            sizeof(float) * LP_FILTER_ORDER);
 
-    ff_celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
+    p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
                                       pole_out + LP_FILTER_ORDER,
                                       AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
 
-    ff_tilt_compensation(&p->tilt_mem, tilt_factor(lpc_n, lpc_d), buf_out,
+    ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out,
                          AMR_SUBFRAME_SIZE);
 
     ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
@@ -990,7 +1005,7 @@ static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
 
         p->fixed_gain[4] =
             ff_amr_set_fixed_gain(fixed_gain_factor,
-                       ff_dot_productf(p->fixed_vector, p->fixed_vector,
+                       p->celpm_ctx.dot_productf(p->fixed_vector, p->fixed_vector,
                                        AMR_SUBFRAME_SIZE)/AMR_SUBFRAME_SIZE,
                        p->prediction_error,
                        energy_mean[p->cur_frame_mode], energy_pred_fac);
@@ -1034,7 +1049,8 @@ static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
         update_state(p);
     }
 
-    ff_acelp_apply_order_2_transfer_function(buf_out, buf_out, highpass_zeros,
+    p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out,
+                                             buf_out, highpass_zeros,
                                              highpass_poles,
                                              highpass_gain * AMR_SAMPLE_SCALE,
                                              p->high_pass_mem, AMR_BLOCK_SIZE);
@@ -1045,7 +1061,7 @@ static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
      * for fixed_gain_smooth.
      * The specification has an incorrect formula: the reference decoder uses
      * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
-    ff_weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
+    p->acelpv_ctx.weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
                             0.84, 0.16, LP_FILTER_ORDER);
 
     *got_frame_ptr   = 1;

libavcodec/amrwbdec.c

@@ -38,6 +38,7 @@
 #include "amr.h"
 
 #include "amrwbdata.h"
+#include "mips/amrwbdec_mips.h"
 
 typedef struct {
     AVFrame                              avframe; ///< AVFrame for decoded samples
@@ -82,6 +83,11 @@ typedef struct {
 
     AVLFG                                   prng; ///< random number generator for white noise excitation
     uint8_t                          first_frame; ///< flag active during decoding of the first frame
+    ACELPFContext                     acelpf_ctx; ///< context for filters for ACELP-based codecs
+    ACELPVContext                     acelpv_ctx; ///< context for vector operations for ACELP-based codecs
+    CELPFContext                       celpf_ctx; ///< context for filters for CELP-based codecs
+    CELPMContext                       celpm_ctx; ///< context for fixed point math operations
+
 } AMRWBContext;
 
 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
@@ -105,6 +111,11 @@ static av_cold int amrwb_decode_init(AVCodecContext *avctx)
     avcodec_get_frame_defaults(&ctx->avframe);
     avctx->coded_frame = &ctx->avframe;
 
+    ff_acelp_filter_init(&ctx->acelpf_ctx);
+    ff_acelp_vectors_init(&ctx->acelpv_ctx);
+    ff_celp_filter_init(&ctx->celpf_ctx);
+    ff_celp_math_init(&ctx->celpm_ctx);
+
     return 0;
 }
 
@@ -319,7 +330,8 @@ static void decode_pitch_vector(AMRWBContext *ctx,
 
     /* Calculate the pitch vector by interpolating the past excitation at the
        pitch lag using a hamming windowed sinc function */
-    ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
+    ctx->acelpf_ctx.acelp_interpolatef(exc,
+                          exc + 1 - pitch_lag_int,
                           ac_inter, 4,
                           pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
                           LP_ORDER, AMRWB_SFR_SIZE + 1);
@@ -578,15 +590,17 @@ static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
  *
  * @param[in] p_vector, f_vector   Pitch and fixed excitation vectors
  * @param[in] p_gain, f_gain       Pitch and fixed gains
+ * @param[in] ctx                  The context
  */
 // XXX: There is something wrong with the precision here! The magnitudes
 // of the energies are not correct. Please check the reference code carefully
 static float voice_factor(float *p_vector, float p_gain,
-                          float *f_vector, float f_gain)
+                          float *f_vector, float f_gain,
+                          CELPMContext *ctx)
 {
-    double p_ener = (double) ff_dot_productf(p_vector, p_vector,
+    double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
                                              AMRWB_SFR_SIZE) * p_gain * p_gain;
-    double f_ener = (double) ff_dot_productf(f_vector, f_vector,
+    double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
                                              AMRWB_SFR_SIZE) * f_gain * f_gain;
 
     return (p_ener - f_ener) / (p_ener + f_ener);
@@ -749,13 +763,13 @@ static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
                       float fixed_gain, const float *fixed_vector,
                       float *samples)
 {
-    ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
+    ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
                             ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
 
     /* emphasize pitch vector contribution in low bitrate modes */
     if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
         int i;
-        float energy = ff_dot_productf(excitation, excitation,
+        float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
                                        AMRWB_SFR_SIZE);
 
         // XXX: Weird part in both ref code and spec. A unknown parameter
@@ -769,7 +783,7 @@ static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
                                                 energy, AMRWB_SFR_SIZE);
     }
 
-    ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
+    ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
                                  AMRWB_SFR_SIZE, LP_ORDER);
 }
 
@@ -801,8 +815,9 @@ static void de_emphasis(float *out, float *in, float m, float mem[1])
  * @param[out] out                 Buffer for interpolated signal
  * @param[in]  in                  Current signal data (length 0.8*o_size)
  * @param[in]  o_size              Output signal length
+ * @param[in] ctx                  The context
  */
-static void upsample_5_4(float *out, const float *in, int o_size)
+static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
 {
     const float *in0 = in - UPS_FIR_SIZE + 1;
     int i, j, k;
@@ -815,7 +830,8 @@ static void upsample_5_4(float *out, const float *in, int o_size)
         i++;
 
         for (k = 1; k < 5; k++) {
-            out[i] = ff_dot_productf(in0 + int_part, upsample_fir[4 - frac_part],
+            out[i] = ctx->dot_productf(in0 + int_part,
+                                     upsample_fir[4 - frac_part],
                                      UPS_MEM_SIZE);
             int_part++;
             frac_part--;
@@ -842,8 +858,8 @@ static float find_hb_gain(AMRWBContext *ctx, const float *synth,
     if (ctx->fr_cur_mode == MODE_23k85)
         return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
 
-    tilt = ff_dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
-           ff_dot_productf(synth, synth, AMRWB_SFR_SIZE);
+    tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
+           ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
 
     /* return gain bounded by [0.1, 1.0] */
     return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
@@ -862,7 +878,7 @@ static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
                                  const float *synth_exc, float hb_gain)
 {
     int i;
-    float energy = ff_dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
+    float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
 
     /* Generate a white-noise excitation */
     for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
@@ -993,7 +1009,7 @@ static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
         float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
         double e_isp[LP_ORDER_16k];
 
-        ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
+        ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
                                 1.0 - isfp_inter[subframe], LP_ORDER);
 
         extrapolate_isf(e_isf);
@@ -1007,7 +1023,7 @@ static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
         lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
     }
 
-    ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
+    ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
                                  (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
 }
 
@@ -1022,6 +1038,8 @@ static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
  *
  * @remark It is safe to pass the same array in in and out parameters
  */
+
+#ifndef hb_fir_filter
 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
                           float mem[HB_FIR_SIZE], const float *in)
 {
@@ -1039,6 +1057,7 @@ static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
 
     memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
 }
+#endif /* hb_fir_filter */
 
 /**
  * Update context state before the next subframe.
@@ -1155,14 +1174,15 @@ static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
 
         ctx->fixed_gain[0] =
             ff_amr_set_fixed_gain(fixed_gain_factor,
-                       ff_dot_productf(ctx->fixed_vector, ctx->fixed_vector,
+                       ctx->celpm_ctx.dot_productf(ctx->fixed_vector, ctx->fixed_vector,
                                        AMRWB_SFR_SIZE) / AMRWB_SFR_SIZE,
                        ctx->prediction_error,
                        ENERGY_MEAN, energy_pred_fac);
 
         /* Calculate voice factor and store tilt for next subframe */
         voice_fac      = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
-                                      ctx->fixed_vector, ctx->fixed_gain[0]);
+                                      ctx->fixed_vector, ctx->fixed_gain[0],
+                                      &ctx->celpm_ctx);
         ctx->tilt_coef = voice_fac * 0.25 + 0.25;
 
         /* Construct current excitation */
@@ -1188,15 +1208,15 @@ static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
         de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
                     &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
 
-        ff_acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
+        ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
             &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
             hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
 
         upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
-                     AMRWB_SFR_SIZE_16k);
+                     AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
 
         /* High frequency band (6.4 - 7.0 kHz) generation part */
-        ff_acelp_apply_order_2_transfer_function(hb_samples,
+        ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
             &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
             hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
 

libavcodec/celp_filters.c

@@ -205,3 +205,12 @@ void ff_celp_lp_zero_synthesis_filterf(float *out, const float *filter_coeffs,
             out[n] += filter_coeffs[i-1] * in[n-i];
     }
 }
+
+void ff_celp_filter_init(CELPFContext *c)
+{
+    c->celp_lp_synthesis_filterf        = ff_celp_lp_synthesis_filterf;
+    c->celp_lp_zero_synthesis_filterf   = ff_celp_lp_zero_synthesis_filterf;
+
+    if(HAVE_MIPSFPU)
+        ff_celp_filter_init_mips(c);
+}