Add AVX FFT implementation.

Signed-off-by: Reinhard Tartler <siretart@tauware.de>

Changelog

@@ -5,7 +5,7 @@ releases are sorted from youngest to oldest.
 version <next>:
 
 - Lots of deprecated API cruft removed
-
+- fft and imdct optimizations for AVX (Sandy Bridge) processors
 
 version 0.7_beta1:
 

libavcodec/aac.h

@@ -223,9 +223,9 @@ typedef struct {
     float sf[120];                                  ///< scalefactors
     int sf_idx[128];                                ///< scalefactor indices (used by encoder)
     uint8_t zeroes[128];                            ///< band is not coded (used by encoder)
-    DECLARE_ALIGNED(16, float,   coeffs)[1024];     ///< coefficients for IMDCT
-    DECLARE_ALIGNED(16, float,   saved)[1024];      ///< overlap
-    DECLARE_ALIGNED(16, float,   ret)[2048];        ///< PCM output
+    DECLARE_ALIGNED(32, float,   coeffs)[1024];     ///< coefficients for IMDCT
+    DECLARE_ALIGNED(32, float,   saved)[1024];      ///< overlap
+    DECLARE_ALIGNED(32, float,   ret)[2048];        ///< PCM output
     DECLARE_ALIGNED(16, int16_t, ltp_state)[3072];  ///< time signal for LTP
     PredictorState predictor_state[MAX_PREDICTORS];
 } SingleChannelElement;
@@ -272,7 +272,7 @@ typedef struct {
      * @defgroup temporary aligned temporary buffers (We do not want to have these on the stack.)
      * @{
      */
-    DECLARE_ALIGNED(16, float, buf_mdct)[1024];
+    DECLARE_ALIGNED(32, float, buf_mdct)[1024];
     /** @} */
 
     /**
@@ -296,7 +296,7 @@ typedef struct {
     int sf_offset;                                    ///< offset into pow2sf_tab as appropriate for dsp.float_to_int16
     /** @} */
 
-    DECLARE_ALIGNED(16, float, temp)[128];
+    DECLARE_ALIGNED(32, float, temp)[128];
 
     enum OCStatus output_configured;
 } AACContext;

libavcodec/aacenc.h

@@ -64,7 +64,7 @@ typedef struct AACEncContext {
     int last_frame;
     float lambda;
     DECLARE_ALIGNED(16, int,   qcoefs)[96];      ///< quantized coefficients
-    DECLARE_ALIGNED(16, float, scoefs)[1024];    ///< scaled coefficients
+    DECLARE_ALIGNED(32, float, scoefs)[1024];    ///< scaled coefficients
 } AACEncContext;
 
 #endif /* AVCODEC_AACENC_H */

libavcodec/ac3dec.h

@@ -200,11 +200,11 @@ typedef struct {
 
 ///@defgroup arrays aligned arrays
     DECLARE_ALIGNED(16, int,   fixed_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS];       ///> fixed-point transform coefficients
-    DECLARE_ALIGNED(16, float, transform_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS];   ///< transform coefficients
-    DECLARE_ALIGNED(16, float, delay)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE];             ///< delay - added to the next block
-    DECLARE_ALIGNED(16, float, window)[AC3_BLOCK_SIZE];                              ///< window coefficients
-    DECLARE_ALIGNED(16, float, tmp_output)[AC3_BLOCK_SIZE];                          ///< temporary storage for output before windowing
-    DECLARE_ALIGNED(16, float, output)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE];            ///< output after imdct transform and windowing
+    DECLARE_ALIGNED(32, float, transform_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS];   ///< transform coefficients
+    DECLARE_ALIGNED(32, float, delay)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE];             ///< delay - added to the next block
+    DECLARE_ALIGNED(32, float, window)[AC3_BLOCK_SIZE];                              ///< window coefficients
+    DECLARE_ALIGNED(32, float, tmp_output)[AC3_BLOCK_SIZE];                          ///< temporary storage for output before windowing
+    DECLARE_ALIGNED(32, float, output)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE];            ///< output after imdct transform and windowing
 ///@}
 } AC3DecodeContext;
 

libavcodec/ac3enc.c

@@ -201,7 +201,7 @@ typedef struct AC3EncodeContext {
 
     uint8_t exp_strategy[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS]; ///< exponent strategies
 
-    DECLARE_ALIGNED(16, SampleType, windowed_samples)[AC3_WINDOW_SIZE];
+    DECLARE_ALIGNED(32, SampleType, windowed_samples)[AC3_WINDOW_SIZE];
 } AC3EncodeContext;
 
 typedef struct AC3Mant {

libavcodec/atrac1.c

@@ -60,11 +60,11 @@ typedef struct {
     int                 log2_block_count[AT1_QMF_BANDS];    ///< log2 number of blocks in a band
     int                 num_bfus;                           ///< number of Block Floating Units
     float*              spectrum[2];
-    DECLARE_ALIGNED(16, float, spec1)[AT1_SU_SAMPLES];     ///< mdct buffer
-    DECLARE_ALIGNED(16, float, spec2)[AT1_SU_SAMPLES];     ///< mdct buffer
-    DECLARE_ALIGNED(16, float, fst_qmf_delay)[46];         ///< delay line for the 1st stacked QMF filter
-    DECLARE_ALIGNED(16, float, snd_qmf_delay)[46];         ///< delay line for the 2nd stacked QMF filter
-    DECLARE_ALIGNED(16, float, last_qmf_delay)[256+23];    ///< delay line for the last stacked QMF filter
+    DECLARE_ALIGNED(32, float, spec1)[AT1_SU_SAMPLES];     ///< mdct buffer
+    DECLARE_ALIGNED(32, float, spec2)[AT1_SU_SAMPLES];     ///< mdct buffer
+    DECLARE_ALIGNED(32, float, fst_qmf_delay)[46];         ///< delay line for the 1st stacked QMF filter
+    DECLARE_ALIGNED(32, float, snd_qmf_delay)[46];         ///< delay line for the 2nd stacked QMF filter
+    DECLARE_ALIGNED(32, float, last_qmf_delay)[256+23];    ///< delay line for the last stacked QMF filter
 } AT1SUCtx;
 
 /**
@@ -72,13 +72,13 @@ typedef struct {
  */
 typedef struct {
     AT1SUCtx            SUs[AT1_MAX_CHANNELS];              ///< channel sound unit
-    DECLARE_ALIGNED(16, float, spec)[AT1_SU_SAMPLES];      ///< the mdct spectrum buffer
+    DECLARE_ALIGNED(32, float, spec)[AT1_SU_SAMPLES];      ///< the mdct spectrum buffer
 
-    DECLARE_ALIGNED(16, float,  low)[256];
-    DECLARE_ALIGNED(16, float,  mid)[256];
-    DECLARE_ALIGNED(16, float, high)[512];
+    DECLARE_ALIGNED(32, float,  low)[256];
+    DECLARE_ALIGNED(32, float,  mid)[256];
+    DECLARE_ALIGNED(32, float, high)[512];
     float*              bands[3];
-    DECLARE_ALIGNED(16, float, out_samples)[AT1_MAX_CHANNELS][AT1_SU_SAMPLES];
+    DECLARE_ALIGNED(32, float, out_samples)[AT1_MAX_CHANNELS][AT1_SU_SAMPLES];
     FFTContext          mdct_ctx[3];
     int                 channels;
     DSPContext          dsp;

libavcodec/atrac3.c

@@ -74,8 +74,8 @@ typedef struct {
     int               gcBlkSwitch;
     gain_block        gainBlock[2];
 
-    DECLARE_ALIGNED(16, float, spectrum)[1024];
-    DECLARE_ALIGNED(16, float, IMDCT_buf)[1024];
+    DECLARE_ALIGNED(32, float, spectrum)[1024];
+    DECLARE_ALIGNED(32, float, IMDCT_buf)[1024];
 
     float             delayBuf1[46]; ///<qmf delay buffers
     float             delayBuf2[46];
@@ -122,7 +122,7 @@ typedef struct {
     FFTContext          mdct_ctx;
 } ATRAC3Context;
 
-static DECLARE_ALIGNED(16, float,mdct_window)[512];
+static DECLARE_ALIGNED(32, float, mdct_window)[512];
 static VLC              spectral_coeff_tab[7];
 static float            gain_tab1[16];
 static float            gain_tab2[31];

libavcodec/binkaudio.c

@@ -55,7 +55,7 @@ typedef struct {
     int num_bands;
     unsigned int *bands;
     float root;
-    DECLARE_ALIGNED(16, FFTSample, coeffs)[BINK_BLOCK_MAX_SIZE];
+    DECLARE_ALIGNED(32, FFTSample, coeffs)[BINK_BLOCK_MAX_SIZE];
     DECLARE_ALIGNED(16, short, previous)[BINK_BLOCK_MAX_SIZE / 16];  ///< coeffs from previous audio block
     float *coeffs_ptr[MAX_CHANNELS]; ///< pointers to the coeffs arrays for float_to_int16_interleave
     union {

libavcodec/cook.c

@@ -153,7 +153,7 @@ typedef struct cook {
     /* data buffers */
 
     uint8_t*            decoded_bytes_buffer;
-    DECLARE_ALIGNED(16, float,mono_mdct_output)[2048];
+    DECLARE_ALIGNED(32, float, mono_mdct_output)[2048];
     float               decode_buffer_1[1024];
     float               decode_buffer_2[1024];
     float               decode_buffer_0[1060]; /* static allocation for joint decode */

libavcodec/dca.c

@@ -321,16 +321,16 @@ typedef struct {
 
     /* Subband samples history (for ADPCM) */
     float subband_samples_hist[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][4];
-    DECLARE_ALIGNED(16, float, subband_fir_hist)[DCA_PRIM_CHANNELS_MAX][512];
-    DECLARE_ALIGNED(16, float, subband_fir_noidea)[DCA_PRIM_CHANNELS_MAX][32];
+    DECLARE_ALIGNED(32, float, subband_fir_hist)[DCA_PRIM_CHANNELS_MAX][512];
+    DECLARE_ALIGNED(32, float, subband_fir_noidea)[DCA_PRIM_CHANNELS_MAX][32];
     int hist_index[DCA_PRIM_CHANNELS_MAX];
-    DECLARE_ALIGNED(16, float, raXin)[32];
+    DECLARE_ALIGNED(32, float, raXin)[32];
 
     int output;                 ///< type of output
     float scale_bias;           ///< output scale
 
-    DECLARE_ALIGNED(16, float, subband_samples)[DCA_BLOCKS_MAX][DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][8];
-    DECLARE_ALIGNED(16, float, samples)[(DCA_PRIM_CHANNELS_MAX+1)*256];
+    DECLARE_ALIGNED(32, float, subband_samples)[DCA_BLOCKS_MAX][DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][8];
+    DECLARE_ALIGNED(32, float, samples)[(DCA_PRIM_CHANNELS_MAX+1)*256];
     const float *samples_chanptr[DCA_PRIM_CHANNELS_MAX+1];
 
     uint8_t dca_buffer[DCA_MAX_FRAME_SIZE + DCA_MAX_EXSS_HEADER_SIZE + DCA_BUFFER_PADDING_SIZE];

libavcodec/fft.c

@@ -93,6 +93,44 @@ av_cold void ff_init_ff_cos_tabs(int index)
 #endif
 }
 
+static const int avx_tab[] = {
+    0, 4, 1, 5, 8, 12, 9, 13, 2, 6, 3, 7, 10, 14, 11, 15
+};
+
+static int is_second_half_of_fft32(int i, int n)
+{
+    if (n <= 32)
+        return i >= 16;
+    else if (i < n/2)
+        return is_second_half_of_fft32(i, n/2);
+    else if (i < 3*n/4)
+        return is_second_half_of_fft32(i - n/2, n/4);
+    else
+        return is_second_half_of_fft32(i - 3*n/4, n/4);
+}
+
+static av_cold void fft_perm_avx(FFTContext *s)
+{
+    int i;
+    int n = 1 << s->nbits;
+
+    for (i = 0; i < n; i += 16) {
+        int k;
+        if (is_second_half_of_fft32(i, n)) {
+            for (k = 0; k < 16; k++)
+                s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] =
+                    i + avx_tab[k];
+
+        } else {
+            for (k = 0; k < 16; k++) {
+                int j = i + k;
+                j = (j & ~7) | ((j >> 1) & 3) | ((j << 2) & 4);
+                s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] = j;
+            }
+        }
+    }
+}
+
 av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
 {
     int i, j, n;
@@ -132,11 +170,16 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
     for(j=4; j<=nbits; j++) {
         ff_init_ff_cos_tabs(j);
     }
-    for(i=0; i<n; i++) {
-        int j = i;
-        if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
-            j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
-        s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
+
+    if (s->fft_permutation == FF_FFT_PERM_AVX) {
+        fft_perm_avx(s);
+    } else {
+        for(i=0; i<n; i++) {
+            int j = i;
+            if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
+                j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
+            s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
+        }
     }
 
     return 0;

libavcodec/fft.h

@@ -85,6 +85,7 @@ struct FFTContext {
     int fft_permutation;
 #define FF_FFT_PERM_DEFAULT   0
 #define FF_FFT_PERM_SWAP_LSBS 1
+#define FF_FFT_PERM_AVX       2
     int mdct_permutation;
 #define FF_MDCT_PERM_NONE       0
 #define FF_MDCT_PERM_INTERLEAVE 1
@@ -97,7 +98,7 @@ struct FFTContext {
 #endif
 
 #define COSTABLE(size) \
-    COSTABLE_CONST DECLARE_ALIGNED(16, FFTSample, FFT_NAME(ff_cos_##size))[size/2]
+    COSTABLE_CONST DECLARE_ALIGNED(32, FFTSample, FFT_NAME(ff_cos_##size))[size/2]
 
 extern COSTABLE(16);
 extern COSTABLE(32);

libavcodec/imc.c

@@ -88,7 +88,7 @@ typedef struct {
 
     DSPContext dsp;
     FFTContext fft;
-    DECLARE_ALIGNED(16, FFTComplex, samples)[COEFFS/2];
+    DECLARE_ALIGNED(32, FFTComplex, samples)[COEFFS/2];
     float *out_samples;
 } IMCContext;
 

libavcodec/nellymoserdec.c

@@ -47,7 +47,7 @@
 
 typedef struct NellyMoserDecodeContext {
     AVCodecContext* avctx;
-    DECLARE_ALIGNED(16, float,float_buf)[NELLY_SAMPLES];
+    DECLARE_ALIGNED(32, float, float_buf)[NELLY_SAMPLES];
     float           state[128];
     AVLFG           random_state;
     GetBitContext   gb;
@@ -55,7 +55,7 @@ typedef struct NellyMoserDecodeContext {
     DSPContext      dsp;
     FFTContext      imdct_ctx;
     FmtConvertContext fmt_conv;
-    DECLARE_ALIGNED(16, float,imdct_out)[NELLY_BUF_LEN * 2];
+    DECLARE_ALIGNED(32, float, imdct_out)[NELLY_BUF_LEN * 2];
 } NellyMoserDecodeContext;
 
 static void overlap_and_window(NellyMoserDecodeContext *s, float *state, float *audio, float *a_in)

libavcodec/nellymoserenc.c

@@ -55,9 +55,9 @@ typedef struct NellyMoserEncodeContext {
     int             have_saved;
     DSPContext      dsp;
     FFTContext      mdct_ctx;
-    DECLARE_ALIGNED(16, float, mdct_out)[NELLY_SAMPLES];
-    DECLARE_ALIGNED(16, float, in_buff)[NELLY_SAMPLES];
-    DECLARE_ALIGNED(16, float, buf)[2][3 * NELLY_BUF_LEN];     ///< sample buffer
+    DECLARE_ALIGNED(32, float, mdct_out)[NELLY_SAMPLES];
+    DECLARE_ALIGNED(32, float, in_buff)[NELLY_SAMPLES];
+    DECLARE_ALIGNED(32, float, buf)[2][3 * NELLY_BUF_LEN];     ///< sample buffer
     float           (*opt )[NELLY_BANDS];
     uint8_t         (*path)[NELLY_BANDS];
 } NellyMoserEncodeContext;

libavcodec/qdm2.c

@@ -120,7 +120,7 @@ typedef struct {
 } FFTCoefficient;
 
 typedef struct {
-    DECLARE_ALIGNED(16, QDM2Complex, complex)[MPA_MAX_CHANNELS][256];
+    DECLARE_ALIGNED(32, QDM2Complex, complex)[MPA_MAX_CHANNELS][256];
 } QDM2FFT;
 
 /**

libavcodec/wma.h

@@ -113,15 +113,15 @@ typedef struct WMACodecContext {
     uint8_t ms_stereo;                      ///< true if mid/side stereo mode
     uint8_t channel_coded[MAX_CHANNELS];    ///< true if channel is coded
     int exponents_bsize[MAX_CHANNELS];      ///< log2 ratio frame/exp. length
-    DECLARE_ALIGNED(16, float, exponents)[MAX_CHANNELS][BLOCK_MAX_SIZE];
+    DECLARE_ALIGNED(32, float, exponents)[MAX_CHANNELS][BLOCK_MAX_SIZE];
     float max_exponent[MAX_CHANNELS];
     WMACoef coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
-    DECLARE_ALIGNED(16, float, coefs)[MAX_CHANNELS][BLOCK_MAX_SIZE];
-    DECLARE_ALIGNED(16, FFTSample, output)[BLOCK_MAX_SIZE * 2];
+    DECLARE_ALIGNED(32, float, coefs)[MAX_CHANNELS][BLOCK_MAX_SIZE];
+    DECLARE_ALIGNED(32, FFTSample, output)[BLOCK_MAX_SIZE * 2];
     FFTContext mdct_ctx[BLOCK_NB_SIZES];
     float *windows[BLOCK_NB_SIZES];
     /* output buffer for one frame and the last for IMDCT windowing */
-    DECLARE_ALIGNED(16, float, frame_out)[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
+    DECLARE_ALIGNED(32, float, frame_out)[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
     /* last frame info */
     uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
     int last_bitoffset;

libavcodec/wmaprodec.c

@@ -145,7 +145,7 @@ typedef struct {
     uint8_t  table_idx;                               ///< index in sf_offsets for the scale factor reference block
     float*   coeffs;                                  ///< pointer to the subframe decode buffer
     uint16_t num_vec_coeffs;                          ///< number of vector coded coefficients
-    DECLARE_ALIGNED(16, float, out)[WMAPRO_BLOCK_MAX_SIZE + WMAPRO_BLOCK_MAX_SIZE / 2]; ///< output buffer
+    DECLARE_ALIGNED(32, float, out)[WMAPRO_BLOCK_MAX_SIZE + WMAPRO_BLOCK_MAX_SIZE / 2]; ///< output buffer
 } WMAProChannelCtx;
 
 /**
@@ -170,7 +170,7 @@ typedef struct WMAProDecodeCtx {
                       FF_INPUT_BUFFER_PADDING_SIZE];///< compressed frame data
     PutBitContext    pb;                            ///< context for filling the frame_data buffer
     FFTContext       mdct_ctx[WMAPRO_BLOCK_SIZES];  ///< MDCT context per block size
-    DECLARE_ALIGNED(16, float, tmp)[WMAPRO_BLOCK_MAX_SIZE]; ///< IMDCT output buffer
+    DECLARE_ALIGNED(32, float, tmp)[WMAPRO_BLOCK_MAX_SIZE]; ///< IMDCT output buffer
     float*           windows[WMAPRO_BLOCK_SIZES];   ///< windows for the different block sizes
 
     /* frame size dependent frame information (set during initialization) */

libavcodec/wmavoice.c

@@ -275,11 +275,11 @@ typedef struct {
                                   ///< by postfilter
     float denoise_filter_cache[MAX_FRAMESIZE];
     int   denoise_filter_cache_size; ///< samples in #denoise_filter_cache
-    DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80];
+    DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
                                   ///< aligned buffer for LPC tilting
-    DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80];
+    DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
                                   ///< aligned buffer for denoise coefficients
-    DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
+    DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
                                   ///< aligned buffer for postfilter speech
                                   ///< synthesis
     /**

libavcodec/x86/fft.c

@@ -25,7 +25,14 @@ av_cold void ff_fft_init_mmx(FFTContext *s)
 {
 #if HAVE_YASM
     int has_vectors = av_get_cpu_flags();
-    if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE) {
+    if (has_vectors & AV_CPU_FLAG_AVX && HAVE_AVX && s->nbits >= 5) {
+        /* AVX for SB */
+        s->imdct_calc      = ff_imdct_calc_sse;
+        s->imdct_half      = ff_imdct_half_avx;
+        s->fft_permute     = ff_fft_permute_sse;
+        s->fft_calc        = ff_fft_calc_avx;
+        s->fft_permutation = FF_FFT_PERM_AVX;
+    } else if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE) {
         /* SSE for P3/P4/K8 */
         s->imdct_calc  = ff_imdct_calc_sse;
         s->imdct_half  = ff_imdct_half_sse;

libavcodec/x86/fft.h

@@ -22,6 +22,7 @@
 #include "libavcodec/fft.h"
 
 void ff_fft_permute_sse(FFTContext *s, FFTComplex *z);
+void ff_fft_calc_avx(FFTContext *s, FFTComplex *z);
 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z);
 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z);
@@ -32,6 +33,7 @@ void ff_imdct_calc_3dn2(FFTContext *s, FFTSample *output, const FFTSample *input
 void ff_imdct_half_3dn2(FFTContext *s, FFTSample *output, const FFTSample *input);
 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input);
 void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input);
+void ff_imdct_half_avx(FFTContext *s, FFTSample *output, const FFTSample *input);
 void ff_dct32_float_sse(FFTSample *out, const FFTSample *in);
 
 #endif

libavcodec/x86/fft_sse.c

@@ -28,6 +28,12 @@ DECLARE_ASM_CONST(16, int, ff_m1m1m1m1)[4] =
 
 void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
 void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
+void ff_fft_dispatch_interleave_avx(FFTComplex *z, int nbits);
+
+void ff_fft_calc_avx(FFTContext *s, FFTComplex *z)
+{
+    ff_fft_dispatch_interleave_avx(z, s->nbits);
+}
 
 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
 {
@@ -77,7 +83,7 @@ void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
     long n = s->mdct_size;
     long n4 = n >> 2;
 
-    ff_imdct_half_sse(s, output+n4, input);
+    s->imdct_half(s, output + n4, input);
 
     j = -n;
     k = n-16;