ARM: NEON optimised apply_window_int16()

Signed-off-by: Mans Rullgard <mans@mansr.com>

libavcodec/arm/dsputil_init_neon.c

@@ -168,6 +168,9 @@ 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, int n);
+
 void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
 {
     if (!avctx->lowres) {
@@ -319,4 +322,6 @@ 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;
 }

libavcodec/arm/dsputil_neon.S

@@ -789,3 +789,26 @@ NOVFP   ldr             r2,  [sp]
         vst1.f32        {q9},[r0,:128]!
         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