-/* Description : Load vectors with 8 halfword elements with stride
- Arguments : Inputs - psrc (source pointer to load from)
- - stride
- Outputs - out0, out1
- Details : Loads 8 halfword elements in 'out0' from (psrc)
- Loads 8 halfword elements in 'out1' from (psrc + stride)
-*/
-#define LD_H2(RTYPE, psrc, stride, out0, out1) \
-{ \
- out0 = LD_H(RTYPE, (psrc)); \
- out1 = LD_H(RTYPE, (psrc) + (stride)); \
-}
-#define LD_UH2(...) LD_H2(v8u16, __VA_ARGS__)
-#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__)
-
-#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3) \
-{ \
- LD_H2(RTYPE, (psrc), stride, out0, out1); \
- LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \
-}
-#define LD_UH4(...) LD_H4(v8u16, __VA_ARGS__)
-#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
-
-#define LD_H6(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5) \
-{ \
- LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3); \
- LD_H2(RTYPE, (psrc) + 4 * stride, stride, out4, out5); \
-}
-#define LD_UH6(...) LD_H6(v8u16, __VA_ARGS__)
-#define LD_SH6(...) LD_H6(v8i16, __VA_ARGS__)
-
-#define LD_H8(RTYPE, psrc, stride, \