@@ -239,6 +239,19 @@ static void ac3_downmix_c(float **samples, float (*matrix)[2],

     }

 }

+static void apply_window_int16_c(int16_t *output, const int16_t *input,

+                                 const int16_t *window, unsigned int len)

+{

+    int i;

+    int len2 = len >> 1;

+

+    for (i = 0; i < len2; i++) {

+        int16_t w       = window[i];

+        output[i]       = (MUL16(input[i],       w) + (1 << 14)) >> 15;

+        output[len-i-1] = (MUL16(input[len-i-1], w) + (1 << 14)) >> 15;

+    }

+}

+

 av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)

 {

     c->ac3_exponent_min = ac3_exponent_min_c;

@@ -253,6 +266,7 @@ av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)

     c->sum_square_butterfly_int32 = ac3_sum_square_butterfly_int32_c;

     c->sum_square_butterfly_float = ac3_sum_square_butterfly_float_c;

     c->downmix = ac3_downmix_c;

+    c->apply_window_int16 = apply_window_int16_c;

     if (ARCH_ARM)

         ff_ac3dsp_init_arm(c, bit_exact);

@@ -134,6 +134,20 @@ typedef struct AC3DSPContext {

     void (*downmix)(float **samples, float (*matrix)[2], int out_ch,

                     int in_ch, int len);

+

+    /**

+     * Apply symmetric window in 16-bit fixed-point.

+     * @param output destination array

+     *               constraints: 16-byte aligned

+     * @param input  source array

+     *               constraints: 16-byte aligned

+     * @param window window array

+     *               constraints: 16-byte aligned, at least len/2 elements

+     * @param len    full window length

+     *               constraints: multiple of ? greater than zero

+     */

+    void (*apply_window_int16)(int16_t *output, const int16_t *input,

+                               const int16_t *window, unsigned int len);

 } AC3DSPContext;

 void ff_ac3dsp_init    (AC3DSPContext *c, int bit_exact);

@@ -71,17 +71,6 @@ av_cold int AC3_NAME(mdct_init)(AC3EncodeContext *s)

 /*

- * Apply KBD window to input samples prior to MDCT.

- */

-static void apply_window(void *dsp, int16_t *output, const int16_t *input,

-                         const int16_t *window, unsigned int len)

-{

-    DSPContext *dsp0 = dsp;

-    dsp0->apply_window_int16(output, input, window, len);

-}

-

-

-/*

  * Normalize the input samples to use the maximum available precision.

  * This assumes signed 16-bit input samples.

  */

@@ -88,18 +88,6 @@ av_cold int ff_ac3_float_mdct_init(AC3EncodeContext *s)

 /*

- * Apply KBD window to input samples prior to MDCT.

- */

-static void apply_window(void *dsp, float *output,

-                         const float *input, const float *window,

-                         unsigned int len)

-{

-    AVFloatDSPContext *fdsp = dsp;

-    fdsp->vector_fmul(output, input, window, len);

-}

-

-

-/*

  * Normalize the input samples.

  * Not needed for the floating-point encoder.

  */

@@ -34,10 +34,6 @@

 static void scale_coefficients(AC3EncodeContext *s);

-static void apply_window(void *dsp, SampleType *output,

-                         const SampleType *input, const SampleType *window,

-                         unsigned int len);

-

 static int normalize_samples(AC3EncodeContext *s);

 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);

@@ -105,11 +101,11 @@ static void apply_mdct(AC3EncodeContext *s)

             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];

 #if CONFIG_AC3ENC_FLOAT

-            apply_window(&s->fdsp, s->windowed_samples, input_samples,

-                         s->mdct_window, AC3_WINDOW_SIZE);

+            s->fdsp.vector_fmul(s->windowed_samples, input_samples,

+                                s->mdct_window, AC3_WINDOW_SIZE);

 #else

-            apply_window(&s->dsp, s->windowed_samples, input_samples,

-                         s->mdct_window, AC3_WINDOW_SIZE);

+            s->ac3dsp.apply_window_int16(s->windowed_samples, input_samples,

+                                         s->mdct_window, AC3_WINDOW_SIZE);

 #endif

             if (s->fixed_point)

@@ -31,6 +31,8 @@ void ff_ac3_lshift_int16_neon(int16_t *src, unsigned len, unsigned shift);

 void ff_ac3_rshift_int32_neon(int32_t *src, unsigned len, unsigned shift);

 void ff_float_to_fixed24_neon(int32_t *dst, const float *src, unsigned int len);

 void ff_ac3_extract_exponents_neon(uint8_t *exp, int32_t *coef, int nb_coefs);

+void ff_apply_window_int16_neon(int16_t *dst, const int16_t *src,

+                                const int16_t *window, unsigned n);

 void ff_ac3_sum_square_butterfly_int32_neon(int64_t sum[4],

                                             const int32_t *coef0,

                                             const int32_t *coef1,

@@ -64,6 +66,7 @@ av_cold void ff_ac3dsp_init_arm(AC3DSPContext *c, int bit_exact)

         c->ac3_rshift_int32      = ff_ac3_rshift_int32_neon;

         c->float_to_fixed24      = ff_float_to_fixed24_neon;

         c->extract_exponents     = ff_ac3_extract_exponents_neon;

+        c->apply_window_int16    = ff_apply_window_int16_neon;

         c->sum_square_butterfly_int32 = ff_ac3_sum_square_butterfly_int32_neon;

         c->sum_square_butterfly_float = ff_ac3_sum_square_butterfly_float_neon;

     }

@@ -109,6 +109,29 @@ function ff_ac3_extract_exponents_neon, export=1

         bx              lr

 endfunc

+function ff_apply_window_int16_neon, export=1

+        push            {r4,lr}

+        add             r4,  r1,  r3,  lsl #1

+        add             lr,  r0,  r3,  lsl #1

+        sub             r4,  r4,  #16

+        sub             lr,  lr,  #16

+        mov             r12, #-16

+1:

+        vld1.16         {q0},     [r1,:128]!

+        vld1.16         {q2},     [r2,:128]!

+        vld1.16         {q1},     [r4,:128], r12

+        vrev64.16       q3,  q2

+        vqrdmulh.s16    q0,  q0,  q2

+        vqrdmulh.s16    d2,  d2,  d7

+        vqrdmulh.s16    d3,  d3,  d6

+        vst1.16         {q0},     [r0,:128]!

+        vst1.16         {q1},     [lr,:128], r12

+        subs            r3,  r3,  #16

+        bgt             1b

+

+        pop             {r4,pc}

+endfunc

+

 function ff_ac3_sum_square_butterfly_int32_neon, export=1

         vmov.i64        q0,  #0

         vmov.i64        q1,  #0

@@ -45,9 +45,6 @@ int32_t ff_scalarproduct_int16_neon(const int16_t *v1, const int16_t *v2, int le

 int32_t ff_scalarproduct_and_madd_int16_neon(int16_t *v1, const int16_t *v2,

                                              const int16_t *v3, int len, int mul);

-void ff_apply_window_int16_neon(int16_t *dst, const int16_t *src,

-                                const int16_t *window, unsigned n);

-

 av_cold void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)

 {

     const int high_bit_depth = avctx->bits_per_raw_sample > 8;

@@ -76,6 +73,4 @@ av_cold void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)

     c->scalarproduct_int16 = ff_scalarproduct_int16_neon;

     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_neon;

-

-    c->apply_window_int16 = ff_apply_window_int16_neon;

 }

@@ -169,29 +169,6 @@ NOVFP   ldr             r2,  [sp]

         bx              lr

 endfunc

-function ff_apply_window_int16_neon, export=1

-        push            {r4,lr}

-        add             r4,  r1,  r3,  lsl #1

-        add             lr,  r0,  r3,  lsl #1

-        sub             r4,  r4,  #16

-        sub             lr,  lr,  #16

-        mov             r12, #-16

-1:

-        vld1.16         {q0},     [r1,:128]!

-        vld1.16         {q2},     [r2,:128]!

-        vld1.16         {q1},     [r4,:128], r12

-        vrev64.16       q3,  q2

-        vqrdmulh.s16    q0,  q0,  q2

-        vqrdmulh.s16    d2,  d2,  d7

-        vqrdmulh.s16    d3,  d3,  d6

-        vst1.16         {q0},     [r0,:128]!

-        vst1.16         {q1},     [lr,:128], r12

-        subs            r3,  r3,  #16

-        bgt             1b

-

-        pop             {r4,pc}

-endfunc

-

 function ff_vector_clip_int32_neon, export=1

         vdup.32         q0,  r2

         vdup.32         q1,  r3

@@ -2495,19 +2495,6 @@ static int32_t scalarproduct_and_madd_int16_c(int16_t *v1, const int16_t *v2, co

     return res;

 }

-static void apply_window_int16_c(int16_t *output, const int16_t *input,

-                                 const int16_t *window, unsigned int len)

-{

-    int i;

-    int len2 = len >> 1;

-

-    for (i = 0; i < len2; i++) {

-        int16_t w       = window[i];

-        output[i]       = (MUL16(input[i],       w) + (1 << 14)) >> 15;

-        output[len-i-1] = (MUL16(input[len-i-1], w) + (1 << 14)) >> 15;

-    }

-}

-

 static void vector_clip_int32_c(int32_t *dst, const int32_t *src, int32_t min,

                                 int32_t max, unsigned int len)

 {

@@ -2799,7 +2786,6 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)

     c->vector_clipf = vector_clipf_c;

     c->scalarproduct_int16 = scalarproduct_int16_c;

     c->scalarproduct_and_madd_int16 = scalarproduct_and_madd_int16_c;

-    c->apply_window_int16 = apply_window_int16_c;

     c->vector_clip_int32 = vector_clip_int32_c;

     c->shrink[0]= av_image_copy_plane;

@@ -275,20 +275,6 @@ typedef struct DSPContext {

     int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, const int16_t *v2, const int16_t *v3, int len, int mul);

     /**

-     * Apply symmetric window in 16-bit fixed-point.

-     * @param output destination array

-     *               constraints: 16-byte aligned

-     * @param input  source array

-     *               constraints: 16-byte aligned

-     * @param window window array

-     *               constraints: 16-byte aligned, at least len/2 elements

-     * @param len    full window length

-     *               constraints: multiple of ? greater than zero

-     */

-    void (*apply_window_int16)(int16_t *output, const int16_t *input,

-                               const int16_t *window, unsigned int len);

-

-    /**

      * Clip each element in an array of int32_t to a given minimum and maximum value.

      * @param dst  destination array

      *             constraints: 16-byte aligned

@@ -51,6 +51,19 @@ void ff_ac3_extract_exponents_3dnow(uint8_t *exp, int32_t *coef, int nb_coefs);

 void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int nb_coefs);

 void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int nb_coefs);

+void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,

+                                        const int16_t *window, unsigned int len);

+void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,

+                                      const int16_t *window, unsigned int len);

+void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,

+                                  const int16_t *window, unsigned int len);

+void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,

+                                const int16_t *window, unsigned int len);

+void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,

+                                 const int16_t *window, unsigned int len);

+void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,

+                                      const int16_t *window, unsigned int len);

+

 #if ARCH_X86_32 && defined(__INTEL_COMPILER)

 #       undef HAVE_7REGS

 #       define HAVE_7REGS 0

@@ -201,6 +214,11 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)

     if (EXTERNAL_MMXEXT(cpu_flags)) {

         c->ac3_exponent_min = ff_ac3_exponent_min_mmxext;

         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmxext;

+        if (bit_exact) {

+            c->apply_window_int16 = ff_apply_window_int16_mmxext;

+        } else {

+            c->apply_window_int16 = ff_apply_window_int16_round_mmxext;

+        }

     }

     if (EXTERNAL_SSE(cpu_flags)) {

         c->float_to_fixed24 = ff_float_to_fixed24_sse;

@@ -215,11 +233,19 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)

             c->ac3_lshift_int16 = ff_ac3_lshift_int16_sse2;

             c->ac3_rshift_int32 = ff_ac3_rshift_int32_sse2;

         }

+        if (bit_exact) {

+            c->apply_window_int16 = ff_apply_window_int16_sse2;

+        } else if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW)) {

+            c->apply_window_int16 = ff_apply_window_int16_round_sse2;

+        }

     }

     if (EXTERNAL_SSSE3(cpu_flags)) {

         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_ssse3;

-        if (!(cpu_flags & AV_CPU_FLAG_ATOM)) {

+        if (cpu_flags & AV_CPU_FLAG_ATOM) {

+            c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;

+        } else {

             c->extract_exponents = ff_ac3_extract_exponents_ssse3;

+            c->apply_window_int16 = ff_apply_window_int16_ssse3;

         }

     }

@@ -86,19 +86,6 @@ int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,

                                               const int16_t *v3,

                                               int order, int mul);

-void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,

-                                        const int16_t *window, unsigned int len);

-void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,

-                                      const int16_t *window, unsigned int len);

-void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,

-                                  const int16_t *window, unsigned int len);

-void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,

-                                const int16_t *window, unsigned int len);

-void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,

-                                 const int16_t *window, unsigned int len);

-void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,

-                                      const int16_t *window, unsigned int len);

-

 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);

 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);

@@ -586,12 +573,6 @@ static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,

     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;

     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;

-

-    if (avctx->flags & CODEC_FLAG_BITEXACT) {

-        c->apply_window_int16 = ff_apply_window_int16_mmxext;

-    } else {

-        c->apply_window_int16 = ff_apply_window_int16_round_mmxext;

-    }

 #endif /* HAVE_MMXEXT_EXTERNAL */

 }

@@ -647,11 +628,6 @@ static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,

     } else {

         c->vector_clip_int32 = ff_vector_clip_int32_sse2;

     }

-    if (avctx->flags & CODEC_FLAG_BITEXACT) {

-        c->apply_window_int16 = ff_apply_window_int16_sse2;

-    } else if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW)) {

-        c->apply_window_int16 = ff_apply_window_int16_round_sse2;

-    }

     c->bswap_buf = ff_bswap32_buf_sse2;

 #endif /* HAVE_SSE2_EXTERNAL */

 }

@@ -664,10 +640,6 @@ static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,

     if (cpu_flags & AV_CPU_FLAG_SSE4) // not really SSE4, just slow on Conroe

         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;

-    if (cpu_flags & AV_CPU_FLAG_ATOM)

-        c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;

-    else

-        c->apply_window_int16 = ff_apply_window_int16_ssse3;

     if (!(cpu_flags & (AV_CPU_FLAG_SSE42 | AV_CPU_FLAG_3DNOW))) // cachesplit

         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;

     c->bswap_buf = ff_bswap32_buf_ssse3;