Factorize some duplicated code from CAVS and H.264 into a common file.
[ffmpeg.git] / libavcodec / i386 / dsputil_mmx.h
index fc5d01e70bf79eeb36d4a8fcb4f6191a01ccf1e7..a1571fc1e3f851f1f4eb3f9454dea49831a2b217 100644 (file)
@@ -52,4 +52,61 @@ extern const uint64_t ff_pb_FC;
 extern const double ff_pd_1[2];
 extern const double ff_pd_2[2];
 
+/* in/out: mma=mma+mmb, mmb=mmb-mma */
+#define SUMSUB_BA( a, b ) \
+    "paddw "#b", "#a" \n\t"\
+    "paddw "#b", "#b" \n\t"\
+    "psubw "#a", "#b" \n\t"
+
+#define SBUTTERFLY(a,b,t,n,m)\
+    "mov" #m " " #a ", " #t "         \n\t" /* abcd */\
+    "punpckl" #n " " #b ", " #a "     \n\t" /* aebf */\
+    "punpckh" #n " " #b ", " #t "     \n\t" /* cgdh */\
+
+#define TRANSPOSE4(a,b,c,d,t)\
+    SBUTTERFLY(a,b,t,wd,q) /* a=aebf t=cgdh */\
+    SBUTTERFLY(c,d,b,wd,q) /* c=imjn b=kolp */\
+    SBUTTERFLY(a,c,d,dq,q) /* a=aeim d=bfjn */\
+    SBUTTERFLY(t,b,c,dq,q) /* t=cgko c=dhlp */
+
+#ifdef ARCH_X86_64
+// permutes 01234567 -> 05736421
+#define TRANSPOSE8(a,b,c,d,e,f,g,h,t)\
+    SBUTTERFLY(a,b,%%xmm8,wd,dqa)\
+    SBUTTERFLY(c,d,b,wd,dqa)\
+    SBUTTERFLY(e,f,d,wd,dqa)\
+    SBUTTERFLY(g,h,f,wd,dqa)\
+    SBUTTERFLY(a,c,h,dq,dqa)\
+    SBUTTERFLY(%%xmm8,b,c,dq,dqa)\
+    SBUTTERFLY(e,g,b,dq,dqa)\
+    SBUTTERFLY(d,f,g,dq,dqa)\
+    SBUTTERFLY(a,e,f,qdq,dqa)\
+    SBUTTERFLY(%%xmm8,d,e,qdq,dqa)\
+    SBUTTERFLY(h,b,d,qdq,dqa)\
+    SBUTTERFLY(c,g,b,qdq,dqa)\
+    "movdqa %%xmm8, "#g"              \n\t"
+#else
+#define TRANSPOSE8(a,b,c,d,e,f,g,h,t)\
+    "movdqa "#h", "#t"                \n\t"\
+    SBUTTERFLY(a,b,h,wd,dqa)\
+    "movdqa "#h", 16"#t"              \n\t"\
+    "movdqa "#t", "#h"                \n\t"\
+    SBUTTERFLY(c,d,b,wd,dqa)\
+    SBUTTERFLY(e,f,d,wd,dqa)\
+    SBUTTERFLY(g,h,f,wd,dqa)\
+    SBUTTERFLY(a,c,h,dq,dqa)\
+    "movdqa "#h", "#t"                \n\t"\
+    "movdqa 16"#t", "#h"              \n\t"\
+    SBUTTERFLY(h,b,c,dq,dqa)\
+    SBUTTERFLY(e,g,b,dq,dqa)\
+    SBUTTERFLY(d,f,g,dq,dqa)\
+    SBUTTERFLY(a,e,f,qdq,dqa)\
+    SBUTTERFLY(h,d,e,qdq,dqa)\
+    "movdqa "#h", 16"#t"              \n\t"\
+    "movdqa "#t", "#h"                \n\t"\
+    SBUTTERFLY(h,b,d,qdq,dqa)\
+    SBUTTERFLY(c,g,b,qdq,dqa)\
+    "movdqa 16"#t", "#g"              \n\t"
+#endif
+
 #endif /* FFMPEG_DSPUTIL_MMX_H */