lib/misc/shuffle2.cl

53 #define _CLC_SHUFFLE_SET_2_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
54     ret_val.s0 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s0); \
55     ret_val.s1 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s1);
57 #define _CLC_SHUFFLE_SET_4_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
58     _CLC_SHUFFLE_SET_2_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
59     ret_val.s2 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s2); \
60     ret_val.s3 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s3);
62 #define _CLC_SHUFFLE_SET_8_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
63     _CLC_SHUFFLE_SET_4_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
64     ret_val.s4 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s4); \
65     ret_val.s5 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s5); \
66     ret_val.s6 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s6); \
67     ret_val.s7 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s7);
69 #define _CLC_SHUFFLE_SET_16_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
70     _CLC_SHUFFLE_SET_8_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
71     ret_val.s8 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s8); \
72     ret_val.s9 = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.s9); \
73     ret_val.sA = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sA); \
74     ret_val.sB = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sB); \
75     ret_val.sC = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sC); \
76     ret_val.sD = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sD); \
77     ret_val.sE = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sE); \
78     ret_val.sF = __clc_get_el_##ARGTYPE##ARGSIZE##_##MASKTYPE(x, y, mask.sF); \
80 #define _CLC_SHUFFLE_DEFINE2(ARGTYPE, ARGSIZE, MASKTYPE) \
81 _CLC_DEF _CLC_OVERLOAD ARGTYPE##2 shuffle2(ARGTYPE##ARGSIZE x, ARGTYPE##ARGSIZE y, MASKTYPE##2 mask…
83     mask &= (MASKTYPE##2)(ARGSIZE * 2 - 1); \
84     _CLC_SHUFFLE_SET_2_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
88 #define _CLC_SHUFFLE_DEFINE4(ARGTYPE, ARGSIZE, MASKTYPE) \
89 _CLC_DEF _CLC_OVERLOAD ARGTYPE##4 shuffle2(ARGTYPE##ARGSIZE x, ARGTYPE##ARGSIZE y, MASKTYPE##4 mask…
91     mask &= (MASKTYPE##4)(ARGSIZE * 2 - 1); \
92     _CLC_SHUFFLE_SET_4_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
96 #define _CLC_SHUFFLE_DEFINE8(ARGTYPE, ARGSIZE, MASKTYPE) \
97 _CLC_DEF _CLC_OVERLOAD ARGTYPE##8 shuffle2(ARGTYPE##ARGSIZE x, ARGTYPE##ARGSIZE y, MASKTYPE##8 mask…
99     mask &= (MASKTYPE##8)(ARGSIZE * 2 - 1); \
100     _CLC_SHUFFLE_SET_8_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
104 #define _CLC_SHUFFLE_DEFINE16(ARGTYPE, ARGSIZE, MASKTYPE) \
105 _CLC_DEF _CLC_OVERLOAD ARGTYPE##16 shuffle2(ARGTYPE##ARGSIZE x, ARGTYPE##ARGSIZE y, MASKTYPE##16 ma…
107     mask &= (MASKTYPE##16)(ARGSIZE * 2 - 1); \
108     _CLC_SHUFFLE_SET_16_ELEMENTS(ARGTYPE, ARGSIZE, MASKTYPE) \
112 #define _CLC_VECTOR_SHUFFLE_MASKSIZE(INTYPE, ARGSIZE, MASKTYPE) \
113   _CLC_GET_ELEMENT_DEFINE(INTYPE, ARGSIZE, MASKTYPE) \
114   _CLC_SHUFFLE_DEFINE2(INTYPE, ARGSIZE, MASKTYPE) \
115   _CLC_SHUFFLE_DEFINE4(INTYPE, ARGSIZE, MASKTYPE) \
116   _CLC_SHUFFLE_DEFINE8(INTYPE, ARGSIZE, MASKTYPE) \
117   _CLC_SHUFFLE_DEFINE16(INTYPE, ARGSIZE, MASKTYPE) \
119 #define _CLC_VECTOR_SHUFFLE_INSIZE(TYPE, MASKTYPE) \
120   _CLC_VECTOR_SHUFFLE_MASKSIZE(TYPE, 2, MASKTYPE) \
121   _CLC_VECTOR_SHUFFLE_MASKSIZE(TYPE, 4, MASKTYPE) \
122   _CLC_VECTOR_SHUFFLE_MASKSIZE(TYPE, 8, MASKTYPE) \
123   _CLC_VECTOR_SHUFFLE_MASKSIZE(TYPE, 16, MASKTYPE) \