| 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 | tst r0, #8 |
| 27 | beq 0f |
| 28 | @ align the dst |
| 29 | vld1.16 {d0}, [r1]! |
| 30 | sub r2, r2, #8 |
| 31 | vshl.u16 d0, d0, #1 |
| 32 | vshl.u16 d1, d0, #10 |
| 33 | vsri.u16 d1, d0, #11 |
| 34 | vbit d1, d0, d30 |
| 35 | vst1.16 {d1}, [r0]! |
| 36 | 0: |
| 37 | subs r2, r2, #64 |
| 38 | blt btr16_end64 |
| 39 | 0: |
| 40 | pld [r1, #64*2] |
| 41 | @ Pulls 15-bit BGR color values (which are actually 16 bits) into q0-q3. |
| 42 | @ example: q0 = 0111 1110 0101 0011 |
| 43 | vld1.16 {d0-d3}, [r1]! |
| 44 | vld1.16 {d4-d7}, [r1]! |
| 45 | @ Shift BGR color 1 bit to the left, discarding MSB and preparing for vbit. |
| 46 | @ MSB is used for transparency (not needed here, and can mess with green). |
| 47 | @ example: q0 = 1111 1100 1010 0110 |
| 48 | vshl.u16 q0, q0, #1 |
| 49 | vshl.u16 q1, q1, #1 |
| 50 | vshl.u16 q2, q2, #1 |
| 51 | vshl.u16 q3, q3, #1 |
| 52 | @ Places red value in left most bits, clears bits to the right. |
| 53 | @ example: q8 = 1001 1000 0000 0000 |
| 54 | vshl.u16 q8, q0, #10 |
| 55 | vshl.u16 q9, q1, #10 |
| 56 | vshl.u16 q10, q2, #10 |
| 57 | vshl.u16 q11, q3, #10 |
| 58 | @ Places blue value in right most bits, leaving bits to the left unchanged. |
| 59 | @ example: q8 = 1001 1000 0001 1111 |
| 60 | vsri.u16 q8, q0, #11 |
| 61 | vsri.u16 q9, q1, #11 |
| 62 | vsri.u16 q10, q2, #11 |
| 63 | vsri.u16 q11, q3, #11 |
| 64 | @ Sets green value from shifted BGR color by apply a mask. |
| 65 | @ example: q15 = 0000 0111 1100 0000 |
| 66 | @ q8 = 1001 1100 1001 1111 |
| 67 | vbit q8, q0, q15 |
| 68 | vbit q9, q1, q15 |
| 69 | vbit q10, q2, q15 |
| 70 | vbit q11, q3, q15 |
| 71 | vstmia r0!, {q8-q11} |
| 72 | subs r2, r2, #64 |
| 73 | bge 0b |
| 74 | |
| 75 | btr16_end64: |
| 76 | adds r2, r2, #64 |
| 77 | bxeq lr |
| 78 | subs r2, r2, #16 |
| 79 | blt btr16_end16 |
| 80 | |
| 81 | @ handle the remainder (reasonably rare) |
| 82 | 0: |
| 83 | vld1.16 {q0}, [r1]! |
| 84 | vshl.u16 q0, q0, #1 |
| 85 | vshl.u16 q1, q0, #10 |
| 86 | vsri.u16 q1, q0, #11 |
| 87 | vbit q1, q0, q15 |
| 88 | subs r2, r2, #16 |
| 89 | vst1.16 {q1}, [r0]! |
| 90 | bge 0b |
| 91 | |
| 92 | btr16_end16: |
| 93 | adds r2, r2, #16 |
| 94 | bxeq lr |
| 95 | subs r2, r2, #8 |
| 96 | bxlt lr |
| 97 | |
| 98 | @ very rare |
| 99 | vld1.16 {d0}, [r1]! |
| 100 | vshl.u16 d0, d0, #1 |
| 101 | vshl.u16 d1, d0, #10 |
| 102 | vsri.u16 d1, d0, #11 |
| 103 | vbit d1, d0, d30 |
| 104 | vst1.16 {d1}, [r0]! |
| 105 | bx lr |
| 106 | |
| 107 | |
| 108 | @ note: may overflow source |
| 109 | FUNCTION(bgr555_to_rgb565_b): @ dst, src, bytes, int brightness2k // 0-0x0800 |
| 110 | pld [r1] |
| 111 | vdup.16 q15, r3 |
| 112 | vpush {q4-q7} |
| 113 | mov r3, #0x1f |
| 114 | vdup.16 q14, r3 |
| 115 | 0: |
| 116 | pld [r1, #64*2] |
| 117 | vld1.16 {d0-d3}, [r1]! |
| 118 | vld1.16 {d4-d7}, [r1]! |
| 119 | vand.u16 q8, q0, q14 |
| 120 | vand.u16 q9, q1, q14 |
| 121 | vand.u16 q10, q2, q14 |
| 122 | vand.u16 q11, q3, q14 |
| 123 | vmul.u16 q4, q8, q15 |
| 124 | vmul.u16 q5, q9, q15 |
| 125 | vmul.u16 q6, q10, q15 |
| 126 | vmul.u16 q7, q11, q15 |
| 127 | |
| 128 | vshr.u16 q8, q0, #5 |
| 129 | vshr.u16 q9, q1, #5 |
| 130 | vshr.u16 q10, q2, #5 |
| 131 | vshr.u16 q11, q3, #5 |
| 132 | vand.u16 q8, q14 |
| 133 | vand.u16 q9, q14 |
| 134 | vand.u16 q10, q14 |
| 135 | vand.u16 q11, q14 |
| 136 | vmul.u16 q8, q15 |
| 137 | vmul.u16 q9, q15 |
| 138 | vmul.u16 q10, q15 |
| 139 | vmul.u16 q11, q15 |
| 140 | vsri.u16 q4, q8, #5 |
| 141 | vsri.u16 q5, q9, #5 |
| 142 | vsri.u16 q6, q10, #5 |
| 143 | vsri.u16 q7, q11, #5 |
| 144 | |
| 145 | vshr.u16 q8, q0, #10 |
| 146 | vshr.u16 q9, q1, #10 |
| 147 | vshr.u16 q10, q2, #10 |
| 148 | vshr.u16 q11, q3, #10 |
| 149 | vand.u16 q8, q14 |
| 150 | vand.u16 q9, q14 |
| 151 | vand.u16 q10, q14 |
| 152 | vand.u16 q11, q14 |
| 153 | vmul.u16 q8, q15 |
| 154 | vmul.u16 q9, q15 |
| 155 | vmul.u16 q10, q15 |
| 156 | vmul.u16 q11, q15 |
| 157 | vsri.u16 q4, q8, #11 |
| 158 | vsri.u16 q5, q9, #11 |
| 159 | vsri.u16 q6, q10, #11 |
| 160 | vsri.u16 q7, q11, #11 |
| 161 | |
| 162 | subs r2, r2, #64 |
| 163 | ble 1f |
| 164 | vstmia r0!, {q4-q7} |
| 165 | b 0b |
| 166 | |
| 167 | 1: |
| 168 | blt 0f |
| 169 | vstmia r0!, {q4-q7} |
| 170 | b btr16b_end |
| 171 | 0: |
| 172 | subs r2, r2, #8 |
| 173 | blt btr16b_end |
| 174 | vst1.16 {q4}, [r0]! |
| 175 | subs r2, r2, #8 |
| 176 | blt btr16b_end |
| 177 | vst1.16 {q5}, [r0]! |
| 178 | subs r2, r2, #8 |
| 179 | blt btr16b_end |
| 180 | vst1.16 {q6}, [r0]! |
| 181 | subs r2, r2, #8 |
| 182 | blt btr16b_end |
| 183 | vst1.16 {q7}, [r0]! |
| 184 | |
| 185 | btr16b_end: |
| 186 | vpop {q4-q7} |
| 187 | bx lr |
| 188 | |
| 189 | |
| 190 | FUNCTION(bgr888_to_rgb888): @ dst, src, bytes |
| 191 | pld [r1] |
| 192 | @ r2 /= 48 |
| 193 | mov r2, r2, lsr #4 |
| 194 | movw r3, #0x5556 |
| 195 | movt r3, #0x5555 |
| 196 | umull r12,r2, r3, r2 |
| 197 | 0: |
| 198 | pld [r1, #48*3] |
| 199 | vld3.8 {d0-d2}, [r1]! |
| 200 | vld3.8 {d3-d5}, [r1]! |
| 201 | vswp d0, d2 |
| 202 | vswp d3, d5 |
| 203 | vst3.8 {d0-d2}, [r0, :64]! |
| 204 | vst3.8 {d3-d5}, [r0, :64]! |
| 205 | subs r2, r2, #1 |
| 206 | bne 0b |
| 207 | |
| 208 | bx lr |
| 209 | |
| 210 | |
| 211 | FUNCTION(bgr888_to_rgb565): @ dst, src, bytes |
| 212 | pld [r1] |
| 213 | @ r2 /= 48 |
| 214 | mov r2, r2, lsr #4 |
| 215 | movw r3, #0x5556 |
| 216 | movt r3, #0x5555 |
| 217 | umull r12,r2, r3, r2 |
| 218 | |
| 219 | mov r3, #0x07e0 |
| 220 | vdup.16 q15, r3 |
| 221 | 0: |
| 222 | pld [r1, #48*3] |
| 223 | vld3.8 {d1-d3}, [r1]! |
| 224 | vld3.8 {d5-d7}, [r1]! |
| 225 | |
| 226 | vshll.u8 q8, d2, #3 @ g |
| 227 | vshll.u8 q9, d6, #3 |
| 228 | vshr.u8 d0, d3, #3 @ b |
| 229 | vshr.u8 d4, d7, #3 |
| 230 | vzip.8 d0, d1 @ rb |
| 231 | vzip.8 d4, d5 |
| 232 | vbit q0, q8, q15 |
| 233 | vbit q2, q9, q15 |
| 234 | |
| 235 | vstmia r0!, {d0,d1} |
| 236 | vstmia r0!, {d4,d5} |
| 237 | subs r2, r2, #1 |
| 238 | bne 0b |
| 239 | |
| 240 | bx lr |
| 241 | |
| 242 | |
| 243 | FUNCTION(rgb888_to_rgb565): @ dst, src, bytes |
| 244 | pld [r1] |
| 245 | @ r2 /= 48 |
| 246 | mov r2, r2, lsr #4 |
| 247 | movw r3, #0x5556 |
| 248 | movt r3, #0x5555 |
| 249 | umull r12,r2, r3, r2 |
| 250 | |
| 251 | mov r3, #0x07e0 |
| 252 | vdup.16 q15, r3 |
| 253 | 0: |
| 254 | pld [r1, #48*3] |
| 255 | vld3.8 {d1-d3}, [r1, :64]! |
| 256 | vld3.8 {d5-d7}, [r1, :64]! |
| 257 | |
| 258 | vshll.u8 q8, d2, #3 @ g |
| 259 | vshll.u8 q9, d6, #3 |
| 260 | vshr.u8 d2, d1, #3 @ b |
| 261 | vshr.u8 d6, d5, #3 |
| 262 | vzip.8 d2, d3 @ rb |
| 263 | vzip.8 d6, d7 |
| 264 | vbit q1, q8, q15 |
| 265 | vbit q3, q9, q15 |
| 266 | |
| 267 | vstmia r0!, {d2,d3} |
| 268 | vstmia r0!, {d6,d7} |
| 269 | subs r2, r2, #1 |
| 270 | bne 0b |
| 271 | |
| 272 | bx lr |
| 273 | |
| 274 | |
| 275 | @ vim:filetype=armasm:expandtab |