| 1 | /* |
| 2 | * (C) GraÅžvydas "notaz" Ignotas, 2010 |
| 3 | * |
| 4 | * This work is licensed under the terms of any of these licenses |
| 5 | * (at your option): |
| 6 | * - GNU GPL, version 2 or later. |
| 7 | * - GNU LGPL, version 2.1 or later. |
| 8 | * See the COPYING file in the top-level directory. |
| 9 | */ |
| 10 | |
| 11 | #include "arm_features.h" |
| 12 | |
| 13 | /* sanity check */ |
| 14 | #ifndef __ARM_NEON__ |
| 15 | #error Compiling NEON code, but appropriate preprocessor flag is missing |
| 16 | #error This usually means -mfpu=neon or -mfloat-abi= is not correctly specified |
| 17 | #endif |
| 18 | |
| 19 | .text |
| 20 | .align 2 |
| 21 | |
| 22 | FUNCTION(bgr555_to_rgb565): @ dst, src, bytes |
| 23 | pld [r1] |
| 24 | mov r3, #0x07c0 |
| 25 | vdup.16 q15, r3 |
| 26 | subs r2, r2, #64 |
| 27 | blt btr16_end64 |
| 28 | 0: |
| 29 | pld [r1, #64*2] |
| 30 | vldmia r1!, {q0-q3} |
| 31 | vshl.u16 q8, q0, #11 |
| 32 | vshl.u16 q9, q1, #11 |
| 33 | vshl.u16 q10, q2, #11 |
| 34 | vshl.u16 q11, q3, #11 |
| 35 | vsri.u16 q8, q0, #10 |
| 36 | vsri.u16 q9, q1, #10 |
| 37 | vsri.u16 q10, q2, #10 |
| 38 | vsri.u16 q11, q3, #10 |
| 39 | vshl.u16 q0, q0, #1 |
| 40 | vshl.u16 q1, q1, #1 |
| 41 | vshl.u16 q2, q2, #1 |
| 42 | vshl.u16 q3, q3, #1 |
| 43 | vbit q8, q0, q15 |
| 44 | vbit q9, q1, q15 |
| 45 | vbit q10, q2, q15 |
| 46 | vbit q11, q3, q15 |
| 47 | vstmia r0!, {q8-q11} |
| 48 | subs r2, r2, #64 |
| 49 | bge 0b |
| 50 | |
| 51 | btr16_end64: |
| 52 | adds r2, r2, #64 |
| 53 | bxeq lr |
| 54 | subs r2, r2, #16 |
| 55 | blt btr16_end16 |
| 56 | |
| 57 | @ handle the remainder (reasonably rare) |
| 58 | 0: |
| 59 | vld1.16 {q0}, [r1]! |
| 60 | vshl.u16 q1, q0, #11 |
| 61 | vshl.u16 q2, q0, #1 |
| 62 | vsri.u16 q1, q0, #10 |
| 63 | vbit q1, q2, q15 |
| 64 | subs r2, r2, #16 |
| 65 | vst1.16 {q1}, [r0]! |
| 66 | bge 0b |
| 67 | |
| 68 | btr16_end16: |
| 69 | adds r2, r2, #16 |
| 70 | bxeq lr |
| 71 | subs r2, r2, #8 |
| 72 | bxlt lr |
| 73 | |
| 74 | @ very rare |
| 75 | vld1.16 {d0}, [r1]! |
| 76 | vshl.u16 d1, d0, #11 |
| 77 | vshl.u16 d2, d0, #1 |
| 78 | vsri.u16 d1, d0, #10 |
| 79 | vbit d1, d2, d30 |
| 80 | vst1.16 {d1}, [r0]! |
| 81 | bx lr |
| 82 | |
| 83 | |
| 84 | @ note: may overflow source |
| 85 | FUNCTION(bgr555_to_rgb565_b): @ dst, src, bytes, int brightness2k // 0-0x0800 |
| 86 | pld [r1] |
| 87 | vdup.16 q15, r3 |
| 88 | vpush {q4-q7} |
| 89 | mov r3, #0x1f |
| 90 | vdup.16 q14, r3 |
| 91 | 0: |
| 92 | pld [r1, #64*2] |
| 93 | vldmia r1!, {q0-q3} |
| 94 | vand.u16 q8, q0, q14 |
| 95 | vand.u16 q9, q1, q14 |
| 96 | vand.u16 q10, q2, q14 |
| 97 | vand.u16 q11, q3, q14 |
| 98 | vmul.u16 q4, q8, q15 |
| 99 | vmul.u16 q5, q9, q15 |
| 100 | vmul.u16 q6, q10, q15 |
| 101 | vmul.u16 q7, q11, q15 |
| 102 | |
| 103 | vshr.u16 q8, q0, #5 |
| 104 | vshr.u16 q9, q1, #5 |
| 105 | vshr.u16 q10, q2, #5 |
| 106 | vshr.u16 q11, q3, #5 |
| 107 | vand.u16 q8, q14 |
| 108 | vand.u16 q9, q14 |
| 109 | vand.u16 q10, q14 |
| 110 | vand.u16 q11, q14 |
| 111 | vmul.u16 q8, q15 |
| 112 | vmul.u16 q9, q15 |
| 113 | vmul.u16 q10, q15 |
| 114 | vmul.u16 q11, q15 |
| 115 | vsri.u16 q4, q8, #5 |
| 116 | vsri.u16 q5, q9, #5 |
| 117 | vsri.u16 q6, q10, #5 |
| 118 | vsri.u16 q7, q11, #5 |
| 119 | |
| 120 | vshr.u16 q8, q0, #10 |
| 121 | vshr.u16 q9, q1, #10 |
| 122 | vshr.u16 q10, q2, #10 |
| 123 | vshr.u16 q11, q3, #10 |
| 124 | vand.u16 q8, q14 |
| 125 | vand.u16 q9, q14 |
| 126 | vand.u16 q10, q14 |
| 127 | vand.u16 q11, q14 |
| 128 | vmul.u16 q8, q15 |
| 129 | vmul.u16 q9, q15 |
| 130 | vmul.u16 q10, q15 |
| 131 | vmul.u16 q11, q15 |
| 132 | vsri.u16 q4, q8, #11 |
| 133 | vsri.u16 q5, q9, #11 |
| 134 | vsri.u16 q6, q10, #11 |
| 135 | vsri.u16 q7, q11, #11 |
| 136 | |
| 137 | subs r2, r2, #64 |
| 138 | ble 1f |
| 139 | vstmia r0!, {q4-q7} |
| 140 | b 0b |
| 141 | |
| 142 | 1: |
| 143 | blt 0f |
| 144 | vstmia r0!, {q4-q7} |
| 145 | b btr16b_end |
| 146 | 0: |
| 147 | subs r2, r2, #8 |
| 148 | blt btr16b_end |
| 149 | vst1.16 {q4}, [r0]! |
| 150 | subs r2, r2, #8 |
| 151 | blt btr16b_end |
| 152 | vst1.16 {q5}, [r0]! |
| 153 | subs r2, r2, #8 |
| 154 | blt btr16b_end |
| 155 | vst1.16 {q6}, [r0]! |
| 156 | subs r2, r2, #8 |
| 157 | blt btr16b_end |
| 158 | vst1.16 {q7}, [r0]! |
| 159 | |
| 160 | btr16b_end: |
| 161 | vpop {q4-q7} |
| 162 | bx lr |
| 163 | |
| 164 | |
| 165 | FUNCTION(bgr888_to_rgb888): @ dst, src, bytes |
| 166 | pld [r1] |
| 167 | @ r2 /= 48 |
| 168 | mov r2, r2, lsr #4 |
| 169 | movw r3, #0x5556 |
| 170 | movt r3, #0x5555 |
| 171 | umull r12,r2, r3, r2 |
| 172 | 0: |
| 173 | pld [r1, #48*3] |
| 174 | vld3.8 {d0-d2}, [r1, :64]! |
| 175 | vld3.8 {d3-d5}, [r1, :64]! |
| 176 | vswp d0, d2 |
| 177 | vswp d3, d5 |
| 178 | vst3.8 {d0-d2}, [r0, :64]! |
| 179 | vst3.8 {d3-d5}, [r0, :64]! |
| 180 | subs r2, r2, #1 |
| 181 | bne 0b |
| 182 | |
| 183 | bx lr |
| 184 | |
| 185 | |
| 186 | FUNCTION(bgr888_to_rgb565): @ dst, src, bytes |
| 187 | pld [r1] |
| 188 | @ r2 /= 48 |
| 189 | mov r2, r2, lsr #4 |
| 190 | movw r3, #0x5556 |
| 191 | movt r3, #0x5555 |
| 192 | umull r12,r2, r3, r2 |
| 193 | |
| 194 | mov r3, #0x07e0 |
| 195 | vdup.16 q15, r3 |
| 196 | 0: |
| 197 | pld [r1, #48*3] |
| 198 | vld3.8 {d1-d3}, [r1, :64]! |
| 199 | vld3.8 {d5-d7}, [r1, :64]! |
| 200 | |
| 201 | vshll.u8 q8, d2, #3 @ g |
| 202 | vshll.u8 q9, d6, #3 |
| 203 | vshr.u8 d0, d3, #3 @ b |
| 204 | vshr.u8 d4, d7, #3 |
| 205 | vzip.8 d0, d1 @ rb |
| 206 | vzip.8 d4, d5 |
| 207 | vbit q0, q8, q15 |
| 208 | vbit q2, q9, q15 |
| 209 | |
| 210 | vstmia r0!, {d0,d1} |
| 211 | vstmia r0!, {d4,d5} |
| 212 | subs r2, r2, #1 |
| 213 | bne 0b |
| 214 | |
| 215 | bx lr |
| 216 | |
| 217 | |
| 218 | FUNCTION(rgb888_to_rgb565): @ dst, src, bytes |
| 219 | pld [r1] |
| 220 | @ r2 /= 48 |
| 221 | mov r2, r2, lsr #4 |
| 222 | movw r3, #0x5556 |
| 223 | movt r3, #0x5555 |
| 224 | umull r12,r2, r3, r2 |
| 225 | |
| 226 | mov r3, #0x07e0 |
| 227 | vdup.16 q15, r3 |
| 228 | 0: |
| 229 | pld [r1, #48*3] |
| 230 | vld3.8 {d1-d3}, [r1, :64]! |
| 231 | vld3.8 {d5-d7}, [r1, :64]! |
| 232 | |
| 233 | vshll.u8 q8, d2, #3 @ g |
| 234 | vshll.u8 q9, d6, #3 |
| 235 | vshr.u8 d2, d1, #3 @ b |
| 236 | vshr.u8 d6, d5, #3 |
| 237 | vzip.8 d2, d3 @ rb |
| 238 | vzip.8 d6, d7 |
| 239 | vbit q1, q8, q15 |
| 240 | vbit q3, q9, q15 |
| 241 | |
| 242 | vstmia r0!, {d2,d3} |
| 243 | vstmia r0!, {d6,d7} |
| 244 | subs r2, r2, #1 |
| 245 | bne 0b |
| 246 | |
| 247 | bx lr |
| 248 | |
| 249 | |
| 250 | @ vim:filetype=armasm |