package/x265/0002-arm-asm-primitives.patch

11          p.ssim_4x4x2_core = PFX(ssim_4x4x2_core_neon);
14 -         p.pu[LUMA_4x4].addAvg   = PFX(addAvg_4x4_neon);
15 -         p.pu[LUMA_4x8].addAvg   = PFX(addAvg_4x8_neon);
16 -         p.pu[LUMA_4x16].addAvg  = PFX(addAvg_4x16_neon);
17 -         p.pu[LUMA_8x4].addAvg   = PFX(addAvg_8x4_neon);
18 -         p.pu[LUMA_8x8].addAvg   = PFX(addAvg_8x8_neon);
19 -         p.pu[LUMA_8x16].addAvg  = PFX(addAvg_8x16_neon);
20 -         p.pu[LUMA_8x32].addAvg  = PFX(addAvg_8x32_neon);
21 -         p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_neon);
22 -         p.pu[LUMA_16x4].addAvg  = PFX(addAvg_16x4_neon);
23 -         p.pu[LUMA_16x8].addAvg  = PFX(addAvg_16x8_neon);
24 -         p.pu[LUMA_16x12].addAvg = PFX(addAvg_16x12_neon);
25 -         p.pu[LUMA_16x16].addAvg = PFX(addAvg_16x16_neon);
26 -         p.pu[LUMA_16x32].addAvg = PFX(addAvg_16x32_neon);
27 -         p.pu[LUMA_16x64].addAvg = PFX(addAvg_16x64_neon);
28 -         p.pu[LUMA_24x32].addAvg = PFX(addAvg_24x32_neon);
29 -         p.pu[LUMA_32x8].addAvg  = PFX(addAvg_32x8_neon);
30 -         p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_neon);
31 -         p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_neon);
32 -         p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_neon);
33 -         p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_neon);
34 -         p.pu[LUMA_48x64].addAvg = PFX(addAvg_48x64_neon);
35 -         p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_neon);
36 -         p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_neon);
37 -         p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_neon);
38 -         p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_neon);
39 +         p.pu[LUMA_4x4].addAvg[ALIGNED]   = PFX(addAvg_4x4_neon);
40 +         p.pu[LUMA_4x8].addAvg[ALIGNED]   = PFX(addAvg_4x8_neon);
41 +         p.pu[LUMA_4x16].addAvg[ALIGNED]  = PFX(addAvg_4x16_neon);
42 +         p.pu[LUMA_8x4].addAvg[ALIGNED]   = PFX(addAvg_8x4_neon);
43 +         p.pu[LUMA_8x8].addAvg[ALIGNED]   = PFX(addAvg_8x8_neon);
44 +         p.pu[LUMA_8x16].addAvg[ALIGNED]  = PFX(addAvg_8x16_neon);
45 +         p.pu[LUMA_8x32].addAvg[ALIGNED]  = PFX(addAvg_8x32_neon);
46 +         p.pu[LUMA_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon);
47 +         p.pu[LUMA_16x4].addAvg[ALIGNED]  = PFX(addAvg_16x4_neon);
48 +         p.pu[LUMA_16x8].addAvg[ALIGNED]  = PFX(addAvg_16x8_neon);
49 +         p.pu[LUMA_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon);
50 +         p.pu[LUMA_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
51 +         p.pu[LUMA_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
52 +         p.pu[LUMA_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon);
53 +         p.pu[LUMA_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon);
54 +         p.pu[LUMA_32x8].addAvg[ALIGNED]  = PFX(addAvg_32x8_neon);
55 +         p.pu[LUMA_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
56 +         p.pu[LUMA_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon);
57 +         p.pu[LUMA_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
58 +         p.pu[LUMA_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon);
59 +         p.pu[LUMA_48x64].addAvg[ALIGNED] = PFX(addAvg_48x64_neon);
60 +         p.pu[LUMA_64x16].addAvg[ALIGNED] = PFX(addAvg_64x16_neon);
61 +         p.pu[LUMA_64x32].addAvg[ALIGNED] = PFX(addAvg_64x32_neon);
62 +         p.pu[LUMA_64x48].addAvg[ALIGNED] = PFX(addAvg_64x48_neon);
63 +         p.pu[LUMA_64x64].addAvg[ALIGNED] = PFX(addAvg_64x64_neon);
66 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg   = PFX(addAvg_4x2_neon);
67 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg   = PFX(addAvg_4x4_neon);
68 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg   = PFX(addAvg_4x8_neon);
69 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg  = PFX(addAvg_4x16_neon);
70 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg   = PFX(addAvg_6x8_neon);
71 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg   = PFX(addAvg_8x2_neon);
72 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg   = PFX(addAvg_8x4_neon);
73 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg   = PFX(addAvg_8x6_neon);
74 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg   = PFX(addAvg_8x8_neon);
75 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg  = PFX(addAvg_8x16_neon);
76 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg  = PFX(addAvg_8x32_neon);
77 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_neon);
78 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg  = PFX(addAvg_16x4_neon);
79 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg  = PFX(addAvg_16x8_neon);
80 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg = PFX(addAvg_16x12_neon);
81 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg = PFX(addAvg_16x16_neon);
82 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg = PFX(addAvg_16x32_neon);
83 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg = PFX(addAvg_24x32_neon);
84 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg  = PFX(addAvg_32x8_neon);
85 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_neon);
86 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg = PFX(addAvg_32x24_neon);
87 -        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg = PFX(addAvg_32x32_neon);
89 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg   = PFX(addAvg_4x8_neon);
90 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg  = PFX(addAvg_4x16_neon);
91 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg  = PFX(addAvg_4x32_neon);
92 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg  = PFX(addAvg_6x16_neon);
93 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg   = PFX(addAvg_8x4_neon);
94 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg   = PFX(addAvg_8x8_neon);
95 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg  = PFX(addAvg_8x12_neon);
96 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg  = PFX(addAvg_8x16_neon);
97 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg  = PFX(addAvg_8x32_neon);
98 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg  = PFX(addAvg_8x64_neon);
99 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_neon);
100 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg  = PFX(addAvg_16x8_neon);
101 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg = PFX(addAvg_16x16_neon);
102 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg = PFX(addAvg_16x24_neon);
103 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg = PFX(addAvg_16x32_neon);
104 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg = PFX(addAvg_16x64_neon);
105 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg = PFX(addAvg_24x64_neon);
106 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_neon);
107 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_neon);
108 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_neon);
109 -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_neon);
110 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg[ALIGNED]   = PFX(addAvg_4x2_neon);
111 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg[ALIGNED]   = PFX(addAvg_4x4_neon);
112 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg[ALIGNED]   = PFX(addAvg_4x8_neon);
113 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg[ALIGNED]  = PFX(addAvg_4x16_neon);
114 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg[ALIGNED]   = PFX(addAvg_6x8_neon);
115 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg[ALIGNED]   = PFX(addAvg_8x2_neon);
116 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg[ALIGNED]   = PFX(addAvg_8x4_neon);
117 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg[ALIGNED]   = PFX(addAvg_8x6_neon);
118 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg[ALIGNED]   = PFX(addAvg_8x8_neon);
119 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg[ALIGNED]  = PFX(addAvg_8x16_neon);
120 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg[ALIGNED]  = PFX(addAvg_8x32_neon);
121 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon);
122 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg[ALIGNED]  = PFX(addAvg_16x4_neon);
123 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg[ALIGNED]  = PFX(addAvg_16x8_neon);
124 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon);
125 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
126 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
127 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon);
128 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg[ALIGNED]  = PFX(addAvg_32x8_neon);
129 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
130 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon);
131 +        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
133 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg[ALIGNED]   = PFX(addAvg_4x8_neon);
134 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg[ALIGNED]  = PFX(addAvg_4x16_neon);
135 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg[ALIGNED]  = PFX(addAvg_4x32_neon);
136 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg[ALIGNED]  = PFX(addAvg_6x16_neon);
137 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg[ALIGNED]   = PFX(addAvg_8x4_neon);
138 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg[ALIGNED]   = PFX(addAvg_8x8_neon);
139 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg[ALIGNED]  = PFX(addAvg_8x12_neon);
140 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg[ALIGNED]  = PFX(addAvg_8x16_neon);
141 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg[ALIGNED]  = PFX(addAvg_8x32_neon);
142 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg[ALIGNED]  = PFX(addAvg_8x64_neon);
143 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg[ALIGNED] = PFX(addAvg_12x32_neon);
144 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg[ALIGNED]  = PFX(addAvg_16x8_neon);
145 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
146 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg[ALIGNED] = PFX(addAvg_16x24_neon);
147 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
148 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon);
149 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg[ALIGNED] = PFX(addAvg_24x64_neon);
150 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
151 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
152 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg[ALIGNED] = PFX(addAvg_32x48_neon);
153 +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon);
156           p.quant = PFX(quant_neon);
158          p.scale2D_64to32  = PFX(scale2D_64to32_neon);
161 -        p.scale1D_128to64 = PFX(scale1D_128to64_neon);
162 +        p.scale1D_128to64[ALIGNED] = PFX(scale1D_128to64_neon);
165          p.cu[BLOCK_4x4].copy_cnt     = PFX(copy_cnt_4_neon);
167          p.cu[BLOCK_32x32].copy_cnt   = PFX(copy_cnt_32_neon);
170 -        p.pu[LUMA_4x4].convert_p2s   = PFX(filterPixelToShort_4x4_neon);
171 -        p.pu[LUMA_4x8].convert_p2s   = PFX(filterPixelToShort_4x8_neon);
172 -        p.pu[LUMA_4x16].convert_p2s  = PFX(filterPixelToShort_4x16_neon);
173 -        p.pu[LUMA_8x4].convert_p2s   = PFX(filterPixelToShort_8x4_neon);
174 -        p.pu[LUMA_8x8].convert_p2s   = PFX(filterPixelToShort_8x8_neon);
175 -        p.pu[LUMA_8x16].convert_p2s  = PFX(filterPixelToShort_8x16_neon);
176 -        p.pu[LUMA_8x32].convert_p2s  = PFX(filterPixelToShort_8x32_neon);
177 -        p.pu[LUMA_12x16].convert_p2s = PFX(filterPixelToShort_12x16_neon);
178 -        p.pu[LUMA_16x4].convert_p2s  = PFX(filterPixelToShort_16x4_neon);
179 -        p.pu[LUMA_16x8].convert_p2s  = PFX(filterPixelToShort_16x8_neon);
180 -        p.pu[LUMA_16x12].convert_p2s = PFX(filterPixelToShort_16x12_neon);
181 -        p.pu[LUMA_16x16].convert_p2s = PFX(filterPixelToShort_16x16_neon);
182 -        p.pu[LUMA_16x32].convert_p2s = PFX(filterPixelToShort_16x32_neon);
183 -        p.pu[LUMA_16x64].convert_p2s = PFX(filterPixelToShort_16x64_neon);
184 -        p.pu[LUMA_24x32].convert_p2s = PFX(filterPixelToShort_24x32_neon);
185 -        p.pu[LUMA_32x8].convert_p2s  = PFX(filterPixelToShort_32x8_neon);
186 -        p.pu[LUMA_32x16].convert_p2s = PFX(filterPixelToShort_32x16_neon);
187 -        p.pu[LUMA_32x24].convert_p2s = PFX(filterPixelToShort_32x24_neon);
188 -        p.pu[LUMA_32x32].convert_p2s = PFX(filterPixelToShort_32x32_neon);
189 -        p.pu[LUMA_32x64].convert_p2s = PFX(filterPixelToShort_32x64_neon);
190 -        p.pu[LUMA_48x64].convert_p2s = PFX(filterPixelToShort_48x64_neon);
191 -        p.pu[LUMA_64x16].convert_p2s = PFX(filterPixelToShort_64x16_neon);
192 -        p.pu[LUMA_64x32].convert_p2s = PFX(filterPixelToShort_64x32_neon);
193 -        p.pu[LUMA_64x48].convert_p2s = PFX(filterPixelToShort_64x48_neon);
194 -        p.pu[LUMA_64x64].convert_p2s = PFX(filterPixelToShort_64x64_neon);
195 +        p.pu[LUMA_4x4].convert_p2s[ALIGNED]   = PFX(filterPixelToShort_4x4_neon);
196 +        p.pu[LUMA_4x8].convert_p2s[ALIGNED]   = PFX(filterPixelToShort_4x8_neon);
197 +        p.pu[LUMA_4x16].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_4x16_neon);
198 +        p.pu[LUMA_8x4].convert_p2s[ALIGNED]   = PFX(filterPixelToShort_8x4_neon);
199 +        p.pu[LUMA_8x8].convert_p2s[ALIGNED]   = PFX(filterPixelToShort_8x8_neon);
200 +        p.pu[LUMA_8x16].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_8x16_neon);
201 +        p.pu[LUMA_8x32].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_8x32_neon);
202 +        p.pu[LUMA_12x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_12x16_neon);
203 +        p.pu[LUMA_16x4].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_16x4_neon);
204 +        p.pu[LUMA_16x8].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_16x8_neon);
205 +        p.pu[LUMA_16x12].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x12_neon);
206 +        p.pu[LUMA_16x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x16_neon);
207 +        p.pu[LUMA_16x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x32_neon);
208 +        p.pu[LUMA_16x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x64_neon);
209 +        p.pu[LUMA_24x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_24x32_neon);
210 +        p.pu[LUMA_32x8].convert_p2s[ALIGNED]  = PFX(filterPixelToShort_32x8_neon);
211 +        p.pu[LUMA_32x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x16_neon);
212 +        p.pu[LUMA_32x24].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x24_neon);
213 +        p.pu[LUMA_32x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x32_neon);
214 +        p.pu[LUMA_32x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x64_neon);
215 +        p.pu[LUMA_48x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_48x64_neon);
216 +        p.pu[LUMA_64x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x16_neon);
217 +        p.pu[LUMA_64x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x32_neon);
218 +        p.pu[LUMA_64x48].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x48_neon);
219 +        p.pu[LUMA_64x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x64_neon);
222 -        p.cu[BLOCK_4x4].blockfill_s   = PFX(blockfill_s_4x4_neon);
223 -        p.cu[BLOCK_8x8].blockfill_s   = PFX(blockfill_s_8x8_neon);
224 -        p.cu[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_neon);
225 -        p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_neon);
226 +        p.cu[BLOCK_4x4].blockfill_s[ALIGNED]   = PFX(blockfill_s_4x4_neon);
227 +        p.cu[BLOCK_8x8].blockfill_s[ALIGNED]   = PFX(blockfill_s_8x8_neon);
228 +        p.cu[BLOCK_16x16].blockfill_s[ALIGNED] = PFX(blockfill_s_16x16_neon);
229 +        p.cu[BLOCK_32x32].blockfill_s[ALIGNED] = PFX(blockfill_s_32x32_neon);
232          p.cu[BLOCK_4x4].copy_ss   = PFX(blockcopy_ss_4x4_neon);
234          p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_neon);
237 -        p.cu[BLOCK_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
238 -        p.cu[BLOCK_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
239 -        p.cu[BLOCK_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
240 -        p.cu[BLOCK_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
241 -        p.cu[BLOCK_64x64].add_ps = PFX(pixel_add_ps_64x64_neon);
242 +        p.cu[BLOCK_4x4].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x4_neon);
243 +        p.cu[BLOCK_8x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_8x8_neon);
244 +        p.cu[BLOCK_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
245 +        p.cu[BLOCK_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
246 +        p.cu[BLOCK_64x64].add_ps[ALIGNED] = PFX(pixel_add_ps_64x64_neon);
249 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
250 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
251 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
252 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
253 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps   = PFX(pixel_add_ps_4x8_neon);
254 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps  = PFX(pixel_add_ps_8x16_neon);
255 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps = PFX(pixel_add_ps_16x32_neon);
256 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps = PFX(pixel_add_ps_32x64_neon);
257 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x4_neon);
258 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_8x8_neon);
259 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
260 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
261 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x8_neon);
262 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps[ALIGNED]  = PFX(pixel_add_ps_8x16_neon);
263 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps[ALIGNED] = PFX(pixel_add_ps_16x32_neon);
264 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps[ALIGNED] = PFX(pixel_add_ps_32x64_neon);
267          p.cu[BLOCK_4x4].cpy2Dto1D_shr   = PFX(cpy2Dto1D_shr_4x4_neon);
269          p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon);
272 -        p.cu[BLOCK_4x4].ssd_s   = PFX(pixel_ssd_s_4x4_neon);
273 -        p.cu[BLOCK_8x8].ssd_s   = PFX(pixel_ssd_s_8x8_neon);
274 -        p.cu[BLOCK_16x16].ssd_s = PFX(pixel_ssd_s_16x16_neon);
275 -        p.cu[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32x32_neon);
276 +        p.cu[BLOCK_4x4].ssd_s[ALIGNED]   = PFX(pixel_ssd_s_4x4_neon);
277 +        p.cu[BLOCK_8x8].ssd_s[ALIGNED]   = PFX(pixel_ssd_s_8x8_neon);
278 +        p.cu[BLOCK_16x16].ssd_s[ALIGNED] = PFX(pixel_ssd_s_16x16_neon);
279 +        p.cu[BLOCK_32x32].ssd_s[ALIGNED] = PFX(pixel_ssd_s_32x32_neon);
282          p.cu[BLOCK_4x4].sse_ss   = PFX(pixel_sse_ss_4x4_neon);
284          p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_neon);
287 -        p.cu[BLOCK_4x4].calcresidual   = PFX(getResidual4_neon);
288 -        p.cu[BLOCK_8x8].calcresidual   = PFX(getResidual8_neon);
289 -        p.cu[BLOCK_16x16].calcresidual = PFX(getResidual16_neon);
290 -        p.cu[BLOCK_32x32].calcresidual = PFX(getResidual32_neon);
291 +        p.cu[BLOCK_4x4].calcresidual[ALIGNED]   = PFX(getResidual4_neon);
292 +        p.cu[BLOCK_8x8].calcresidual[ALIGNED]   = PFX(getResidual8_neon);
293 +        p.cu[BLOCK_16x16].calcresidual[ALIGNED] = PFX(getResidual16_neon);
294 +        p.cu[BLOCK_32x32].calcresidual[ALIGNED] = PFX(getResidual32_neon);
297          p.cu[BLOCK_4x4].sse_pp   = PFX(pixel_sse_pp_4x4_neon);
299          p.pu[LUMA_64x64].sad_x4 = PFX(sad_x4_64x64_neon);
302 -        p.pu[LUMA_4x4].pixelavg_pp   = PFX(pixel_avg_pp_4x4_neon);
303 -        p.pu[LUMA_4x8].pixelavg_pp   = PFX(pixel_avg_pp_4x8_neon);
304 -        p.pu[LUMA_4x16].pixelavg_pp  = PFX(pixel_avg_pp_4x16_neon);
305 -        p.pu[LUMA_8x4].pixelavg_pp   = PFX(pixel_avg_pp_8x4_neon);
306 -        p.pu[LUMA_8x8].pixelavg_pp   = PFX(pixel_avg_pp_8x8_neon);
307 -        p.pu[LUMA_8x16].pixelavg_pp  = PFX(pixel_avg_pp_8x16_neon);
308 -        p.pu[LUMA_8x32].pixelavg_pp  = PFX(pixel_avg_pp_8x32_neon);
309 -        p.pu[LUMA_12x16].pixelavg_pp = PFX(pixel_avg_pp_12x16_neon);
310 -        p.pu[LUMA_16x4].pixelavg_pp  = PFX(pixel_avg_pp_16x4_neon);
311 -        p.pu[LUMA_16x8].pixelavg_pp  = PFX(pixel_avg_pp_16x8_neon);
312 -        p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_pp_16x12_neon);
313 -        p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_pp_16x16_neon);
314 -        p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_pp_16x32_neon);
315 -        p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_pp_16x64_neon);
316 -        p.pu[LUMA_24x32].pixelavg_pp = PFX(pixel_avg_pp_24x32_neon);
317 -        p.pu[LUMA_32x8].pixelavg_pp  = PFX(pixel_avg_pp_32x8_neon);
318 -        p.pu[LUMA_32x16].pixelavg_pp = PFX(pixel_avg_pp_32x16_neon);
319 -        p.pu[LUMA_32x24].pixelavg_pp = PFX(pixel_avg_pp_32x24_neon);
320 -        p.pu[LUMA_32x32].pixelavg_pp = PFX(pixel_avg_pp_32x32_neon);
321 -        p.pu[LUMA_32x64].pixelavg_pp = PFX(pixel_avg_pp_32x64_neon);
322 -        p.pu[LUMA_48x64].pixelavg_pp = PFX(pixel_avg_pp_48x64_neon);
323 -        p.pu[LUMA_64x16].pixelavg_pp = PFX(pixel_avg_pp_64x16_neon);
324 -        p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_pp_64x32_neon);
325 -        p.pu[LUMA_64x48].pixelavg_pp = PFX(pixel_avg_pp_64x48_neon);
326 -        p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_pp_64x64_neon);
327 +        p.pu[LUMA_4x4].pixelavg_pp[ALIGNED]   = PFX(pixel_avg_pp_4x4_neon);
328 +        p.pu[LUMA_4x8].pixelavg_pp[ALIGNED]   = PFX(pixel_avg_pp_4x8_neon);
329 +        p.pu[LUMA_4x16].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_4x16_neon);
330 +        p.pu[LUMA_8x4].pixelavg_pp[ALIGNED]   = PFX(pixel_avg_pp_8x4_neon);
331 +        p.pu[LUMA_8x8].pixelavg_pp[ALIGNED]   = PFX(pixel_avg_pp_8x8_neon);
332 +        p.pu[LUMA_8x16].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_8x16_neon);
333 +        p.pu[LUMA_8x32].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_8x32_neon);
334 +        p.pu[LUMA_12x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_12x16_neon);
335 +        p.pu[LUMA_16x4].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_16x4_neon);
336 +        p.pu[LUMA_16x8].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_16x8_neon);
337 +        p.pu[LUMA_16x12].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x12_neon);
338 +        p.pu[LUMA_16x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x16_neon);
339 +        p.pu[LUMA_16x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x32_neon);
340 +        p.pu[LUMA_16x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x64_neon);
341 +        p.pu[LUMA_24x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_24x32_neon);
342 +        p.pu[LUMA_32x8].pixelavg_pp[ALIGNED]  = PFX(pixel_avg_pp_32x8_neon);
343 +        p.pu[LUMA_32x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x16_neon);
344 +        p.pu[LUMA_32x24].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x24_neon);
345 +        p.pu[LUMA_32x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x32_neon);
346 +        p.pu[LUMA_32x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x64_neon);
347 +        p.pu[LUMA_48x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_48x64_neon);
348 +        p.pu[LUMA_64x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x16_neon);
349 +        p.pu[LUMA_64x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x32_neon);
350 +        p.pu[LUMA_64x48].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x48_neon);
351 +        p.pu[LUMA_64x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x64_neon);
354          p.planecopy_cp = PFX(pixel_planecopy_cp_neon);