ARM: NEON optimised simple_idct

Originally committed as revision 16146 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/Makefile

@@ -450,6 +450,7 @@ OBJS-$(HAVE_IWMMXT)                    += armv4l/dsputil_iwmmxt.o       \
 
 OBJS-$(HAVE_NEON)                      += armv4l/dsputil_neon.o         \
                                           armv4l/dsputil_neon_s.o       \
+                                          armv4l/simple_idct_neon.o     \
 
 OBJS-$(ARCH_BFIN)                      += bfin/dsputil_bfin.o           \
                                           bfin/fdct_bfin.o              \

libavcodec/armv4l/dsputil_arm.c

@@ -39,6 +39,10 @@ void ff_simple_idct_armv6(DCTELEM *data);
 void ff_simple_idct_put_armv6(uint8_t *dest, int line_size, DCTELEM *data);
 void ff_simple_idct_add_armv6(uint8_t *dest, int line_size, DCTELEM *data);
 
+void ff_simple_idct_neon(DCTELEM *data);
+void ff_simple_idct_put_neon(uint8_t *dest, int line_size, DCTELEM *data);
+void ff_simple_idct_add_neon(uint8_t *dest, int line_size, DCTELEM *data);
+
 /* XXX: local hack */
 static void (*ff_put_pixels_clamped)(const DCTELEM *block, uint8_t *pixels, int line_size);
 static void (*ff_add_pixels_clamped)(const DCTELEM *block, uint8_t *pixels, int line_size);
@@ -128,6 +132,8 @@ void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx)
         if(idct_algo == FF_IDCT_AUTO){
 #if defined(HAVE_IPP)
             idct_algo = FF_IDCT_IPP;
+#elif defined(HAVE_NEON)
+            idct_algo = FF_IDCT_SIMPLENEON;
 #elif defined(HAVE_ARMV6)
             idct_algo = FF_IDCT_SIMPLEARMV6;
 #elif defined(HAVE_ARMV5TE)
@@ -167,6 +173,13 @@ void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx)
             c->idct_add= simple_idct_ipp_add;
             c->idct    = simple_idct_ipp;
             c->idct_permutation_type= FF_NO_IDCT_PERM;
+#endif
+#ifdef HAVE_NEON
+        } else if (idct_algo==FF_IDCT_SIMPLENEON){
+            c->idct_put= ff_simple_idct_put_neon;
+            c->idct_add= ff_simple_idct_add_neon;
+            c->idct    = ff_simple_idct_neon;
+            c->idct_permutation_type = FF_PARTTRANS_IDCT_PERM;
 #endif
         }
     }

libavcodec/avcodec.h

@@ -1390,6 +1390,7 @@ typedef struct AVCodecContext {
 #define FF_IDCT_WMV2          19
 #define FF_IDCT_FAAN          20
 #define FF_IDCT_EA            21
+#define FF_IDCT_SIMPLENEON    22
 
     /**
      * slice count

libavcodec/utils.c

@@ -564,6 +564,7 @@ static const AVOption options[]={
 {"simplearm", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_SIMPLEARM, INT_MIN, INT_MAX, V|E|D, "idct"},
 {"simplearmv5te", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_SIMPLEARMV5TE, INT_MIN, INT_MAX, V|E|D, "idct"},
 {"simplearmv6", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_SIMPLEARMV6, INT_MIN, INT_MAX, V|E|D, "idct"},
+{"simpleneon", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_SIMPLENEON, INT_MIN, INT_MAX, V|E|D, "idct"},
 {"h264", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_H264, INT_MIN, INT_MAX, V|E|D, "idct"},
 {"vp3", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_VP3, INT_MIN, INT_MAX, V|E|D, "idct"},
 {"ipp", NULL, 0, FF_OPT_TYPE_CONST, FF_IDCT_IPP, INT_MIN, INT_MAX, V|E|D, "idct"},