24 #ifndef AVUTIL_PPC_UTIL_ALTIVEC_H 25 #define AVUTIL_PPC_UTIL_ALTIVEC_H 41 #define WORD_0 0x00,0x01,0x02,0x03 42 #define WORD_1 0x04,0x05,0x06,0x07 43 #define WORD_2 0x08,0x09,0x0a,0x0b 44 #define WORD_3 0x0c,0x0d,0x0e,0x0f 45 #define WORD_s0 0x10,0x11,0x12,0x13 46 #define WORD_s1 0x14,0x15,0x16,0x17 47 #define WORD_s2 0x18,0x19,0x1a,0x1b 48 #define WORD_s3 0x1c,0x1d,0x1e,0x1f 49 #define vcprm(a,b,c,d) (const vec_u8){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d} 51 #define SWP_W2S0 0x02,0x03,0x00,0x01 52 #define SWP_W2S1 0x06,0x07,0x04,0x05 53 #define SWP_W2S2 0x0a,0x0b,0x08,0x09 54 #define SWP_W2S3 0x0e,0x0f,0x0c,0x0d 55 #define SWP_W2Ss0 0x12,0x13,0x10,0x11 56 #define SWP_W2Ss1 0x16,0x17,0x14,0x15 57 #define SWP_W2Ss2 0x1a,0x1b,0x18,0x19 58 #define SWP_W2Ss3 0x1e,0x1f,0x1c,0x1d 59 #define vcswapi2s(a,b,c,d) (const vector unsigned char){SWP_W2S ## a, SWP_W2S ## b, SWP_W2S ## c, SWP_W2S ## d} 62 (const vector unsigned char){0x0f,0x0e,0x0d,0x0c,0x0b,0x0a,0x09,0x08,0x07,0x06,0x05,0x04,0x03,0x02,0x01,0x00} 66 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \ 68 vec_s16 A1, B1, C1, D1, E1, F1, G1, H1; \ 69 vec_s16 A2, B2, C2, D2, E2, F2, G2, H2; \ 71 A1 = vec_mergeh (a, e); \ 72 B1 = vec_mergel (a, e); \ 73 C1 = vec_mergeh (b, f); \ 74 D1 = vec_mergel (b, f); \ 75 E1 = vec_mergeh (c, g); \ 76 F1 = vec_mergel (c, g); \ 77 G1 = vec_mergeh (d, h); \ 78 H1 = vec_mergel (d, h); \ 80 A2 = vec_mergeh (A1, E1); \ 81 B2 = vec_mergel (A1, E1); \ 82 C2 = vec_mergeh (B1, F1); \ 83 D2 = vec_mergel (B1, F1); \ 84 E2 = vec_mergeh (C1, G1); \ 85 F2 = vec_mergel (C1, G1); \ 86 G2 = vec_mergeh (D1, H1); \ 87 H2 = vec_mergel (D1, H1); \ 89 a = vec_mergeh (A2, E2); \ 90 b = vec_mergel (A2, E2); \ 91 c = vec_mergeh (B2, F2); \ 92 d = vec_mergel (B2, F2); \ 93 e = vec_mergeh (C2, G2); \ 94 f = vec_mergel (C2, G2); \ 95 g = vec_mergeh (D2, H2); \ 96 h = vec_mergel (D2, H2); \ 101 #define VEC_LD(offset,b) \ 102 vec_perm(vec_ld(offset, b), vec_ld((offset)+15, b), vec_lvsl(offset, b)) 104 #define VEC_LD(offset,b) \ 105 vec_vsx_ld(offset, b) 113 register vec_u8 first = vec_ld(offset, src);
114 register vec_u8 second = vec_ld(offset + 15, src);
116 return vec_perm(first, second, mask);
120 vec_u8 a = vec_ld(offset, src);
121 vec_u8 b = vec_ld(offset + 15, src);
122 return vec_perm(a, b, perm_vec);
125 #define unaligned_load(a,b) VEC_LD(a,b) 126 #define load_with_perm_vec(a,b,c) VEC_LD(a,b) 135 #define vec_unaligned_load(b) VEC_LD(0, b) 138 #define VEC_MERGEH(a, b) vec_mergeh(a, b) 139 #define VEC_MERGEL(a, b) vec_mergel(a, b) 141 #define VEC_MERGEH(a, b) vec_mergeh(b, a) 142 #define VEC_MERGEL(a, b) vec_mergel(b, a) 146 #define VEC_ST(a,b,c) vec_st(a,b,c) 148 #define VEC_ST(a,b,c) vec_vsx_st(a,b,c) 152 #define VEC_SPLAT16(a,b) vec_splat((vec_s16)(a), b) 154 #define VEC_SPLAT16(a,b) vec_splat((vec_s16)(vec_perm(a, a, vcswapi2s(0,1,2,3))), b) 158 #define VEC_SLD16(a,b,c) vec_sld(a, b, c) 160 #define VEC_SLD16(a,b,c) vec_sld(b, a, c) 167 #define vsx_ld_u8_s16(off, p) \ 168 ((vec_s16)vec_mergeh((vec_u8)vec_splat_u8(0), \ 169 (vec_u8)vec_vsx_ld((off), (p)))) 171 #define vsx_ld_u8_s16(off, p) \ 172 ((vec_s16)vec_mergeh((vec_u8)vec_vsx_ld((off), (p)), \ 173 (vec_u8)vec_splat_u8(0)))
static const uint16_t mask[17]
static const uint8_t offset[127][2]