## ffmpeg / libavcodec / arm / simple_idct_arm.S @ 2912e87a

History | View | Annotate | Download (21.5 KB)

1 |
/* |
---|---|

2 |
* simple_idct_arm.S |

3 |
* Copyright (C) 2002 Frederic 'dilb' Boulay |

4 |
* |

5 |
* Author: Frederic Boulay <dilb@handhelds.org> |

6 |
* |

7 |
* The function defined in this file is derived from the simple_idct function |

8 |
* from the libavcodec library part of the Libav project. |

9 |
* |

10 |
* This file is part of Libav. |

11 |
* |

12 |
* Libav is free software; you can redistribute it and/or |

13 |
* modify it under the terms of the GNU Lesser General Public |

14 |
* License as published by the Free Software Foundation; either |

15 |
* version 2.1 of the License, or (at your option) any later version. |

16 |
* |

17 |
* Libav is distributed in the hope that it will be useful, |

18 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |

19 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |

20 |
* Lesser General Public License for more details. |

21 |
* |

22 |
* You should have received a copy of the GNU Lesser General Public |

23 |
* License along with Libav; if not, write to the Free Software |

24 |
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |

25 |
*/ |

26 | |

27 |
#include "asm.S" |

28 | |

29 |
/* useful constants for the algorithm, they are save in __constant_ptr__ at */ |

30 |
/* the end of the source code.*/ |

31 |
#define W1 22725 |

32 |
#define W2 21407 |

33 |
#define W3 19266 |

34 |
#define W4 16383 |

35 |
#define W5 12873 |

36 |
#define W6 8867 |

37 |
#define W7 4520 |

38 |
#define MASK_MSHW 0xFFFF0000 |

39 | |

40 |
/* offsets of the constants in the vector */ |

41 |
#define offW1 0 |

42 |
#define offW2 4 |

43 |
#define offW3 8 |

44 |
#define offW4 12 |

45 |
#define offW5 16 |

46 |
#define offW6 20 |

47 |
#define offW7 24 |

48 |
#define offMASK_MSHW 28 |

49 | |

50 |
#define ROW_SHIFT 11 |

51 |
#define ROW_SHIFT2MSHW (16-11) |

52 |
#define COL_SHIFT 20 |

53 |
#define ROW_SHIFTED_1 1024 /* 1<< (ROW_SHIFT-1) */ |

54 |
#define COL_SHIFTED_1 524288 /* 1<< (COL_SHIFT-1) */ |

55 | |

56 | |

57 |
.text |

58 | |

59 |
function ff_simple_idct_arm, export=1 |

60 |
@@ void simple_idct_arm(int16_t *block) |

61 |
@@ save stack for reg needed (take all of them), |

62 |
@@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block |

63 |
@@ so it must not be overwritten, if it is not saved!! |

64 |
@@ R12 is another scratch register, so it should not be saved too |

65 |
@@ save all registers |

66 |
stmfd sp!, {r4-r11, r14} @ R14 is also called LR |

67 |
@@ at this point, R0=block, other registers are free. |

68 |
add r14, r0, #112 @ R14=&block[8*7], better start from the last row, and decrease the value until row=0, i.e. R12=block. |

69 |
adr r12, __constant_ptr__ @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it |

70 |
@@ add 2 temporary variables in the stack: R0 and R14 |

71 |
sub sp, sp, #8 @ allow 2 local variables |

72 |
str r0, [sp, #0] @ save block in sp[0] |

73 |
@@ stack status |

74 |
@@ sp+4 free |

75 |
@@ sp+0 R0 (block) |

76 | |

77 | |

78 |
@@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free |

79 | |

80 | |

81 |
__row_loop: |

82 |
@@ read the row and check if it is null, almost null, or not, according to strongarm specs, it is not necessary to optimize ldr accesses (i.e. split 32bits in 2 16bits words), at least it gives more usable registers :) |

83 |
ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer) |

84 |
ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1] |

85 |
ldr r3, [r14, #8] @ R3=ROWr32[2] |

86 |
ldr r4, [r14, #12] @ R4=ROWr32[3] |

87 |
@@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop), |

88 |
@@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row) |

89 |
@@ else follow the complete algorithm. |

90 |
@@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1], |

91 |
@@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free |

92 |
orr r5, r4, r3 @ R5=R4 | R3 |

93 |
orr r5, r5, r2 @ R5=R4 | R3 | R2 |

94 |
orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null) |

95 |
beq __end_row_loop |

96 |
mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later) |

97 |
ldrsh r6, [r14, #0] @ R6=ROWr16[0] |

98 |
orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7 |

99 |
beq __almost_empty_row |

100 | |

101 |
__b_evaluation: |

102 |
@@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3], |

103 |
@@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free, |

104 |
@@ R12=__const_ptr_, R14=&block[n] |

105 |
@@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3 |

106 | |

107 |
@@ MUL16(b0, W1, row[1]); |

108 |
@@ MUL16(b1, W3, row[1]); |

109 |
@@ MUL16(b2, W5, row[1]); |

110 |
@@ MUL16(b3, W7, row[1]); |

111 |
@@ MAC16(b0, W3, row[3]); |

112 |
@@ MAC16(b1, -W7, row[3]); |

113 |
@@ MAC16(b2, -W1, row[3]); |

114 |
@@ MAC16(b3, -W5, row[3]); |

115 |
ldr r8, [r12, #offW1] @ R8=W1 |

116 |
mov r2, r2, asr #16 @ R2=ROWr16[3] |

117 |
mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

118 |
ldr r9, [r12, #offW3] @ R9=W3 |

119 |
ldr r10, [r12, #offW5] @ R10=W5 |

120 |
mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

121 |
ldr r11, [r12, #offW7] @ R11=W7 |

122 |
mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

123 |
mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

124 |
teq r2, #0 @ if null avoid muls |

125 |
mlane r0, r9, r2, r0 @ R0+=W3*ROWr16[3]=b0 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

126 |
rsbne r2, r2, #0 @ R2=-ROWr16[3] |

127 |
mlane r1, r11, r2, r1 @ R1-=W7*ROWr16[3]=b1 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

128 |
mlane r5, r8, r2, r5 @ R5-=W1*ROWr16[3]=b2 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

129 |
mlane r7, r10, r2, r7 @ R7-=W5*ROWr16[3]=b3 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

130 | |

131 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3], |

132 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |

133 |
@@ R12=__const_ptr_, R14=&block[n] |

134 |
@@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3]; |

135 |
@@ if (temp != 0) {} |

136 |
orrs r2, r3, r4 @ R2=ROWr32[2] | ROWr32[3] |

137 |
beq __end_b_evaluation |

138 | |

139 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3], |

140 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |

141 |
@@ R12=__const_ptr_, R14=&block[n] |

142 |
@@ MAC16(b0, W5, row[5]); |

143 |
@@ MAC16(b2, W7, row[5]); |

144 |
@@ MAC16(b3, W3, row[5]); |

145 |
@@ MAC16(b1, -W1, row[5]); |

146 |
@@ MAC16(b0, W7, row[7]); |

147 |
@@ MAC16(b2, W3, row[7]); |

148 |
@@ MAC16(b3, -W1, row[7]); |

149 |
@@ MAC16(b1, -W5, row[7]); |

150 |
mov r3, r3, asr #16 @ R3=ROWr16[5] |

151 |
teq r3, #0 @ if null avoid muls |

152 |
mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0 |

153 |
mov r4, r4, asr #16 @ R4=ROWr16[7] |

154 |
mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5]=b2 |

155 |
mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5]=b3 |

156 |
rsbne r3, r3, #0 @ R3=-ROWr16[5] |

157 |
mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5]=b1 |

158 |
@@ R3 is free now |

159 |
teq r4, #0 @ if null avoid muls |

160 |
mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0 |

161 |
mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2 |

162 |
rsbne r4, r4, #0 @ R4=-ROWr16[7] |

163 |
mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3 |

164 |
mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1 |

165 |
@@ R4 is free now |

166 |
__end_b_evaluation: |

167 |
@@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free), |

168 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |

169 |
@@ R12=__const_ptr_, R14=&block[n] |

170 | |

171 |
__a_evaluation: |

172 |
@@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1)); |

173 |
@@ a1 = a0 + W6 * row[2]; |

174 |
@@ a2 = a0 - W6 * row[2]; |

175 |
@@ a3 = a0 - W2 * row[2]; |

176 |
@@ a0 = a0 + W2 * row[2]; |

177 |
ldr r9, [r12, #offW4] @ R9=W4 |

178 |
mul r6, r9, r6 @ R6=W4*ROWr16[0] |

179 |
ldr r10, [r12, #offW6] @ R10=W6 |

180 |
ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet) |

181 |
add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0) |

182 | |

183 |
mul r11, r10, r4 @ R11=W6*ROWr16[2] |

184 |
ldr r8, [r12, #offW2] @ R8=W2 |

185 |
sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2) |

186 |
@@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3]; |

187 |
@@ if (temp != 0) {} |

188 |
teq r2, #0 |

189 |
beq __end_bef_a_evaluation |

190 | |

191 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |

192 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |

193 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |

194 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |

195 | |

196 | |

197 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |

198 |
@@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free), |

199 |
@@ R12=__const_ptr_, R14=&block[n] |

200 | |

201 | |

202 |
@@ a0 += W4*row[4] |

203 |
@@ a1 -= W4*row[4] |

204 |
@@ a2 -= W4*row[4] |

205 |
@@ a3 += W4*row[4] |

206 |
ldrsh r11, [r14, #8] @ R11=ROWr16[4] |

207 |
teq r11, #0 @ if null avoid muls |

208 |
mulne r11, r9, r11 @ R11=W4*ROWr16[4] |

209 |
@@ R9 is free now |

210 |
ldrsh r9, [r14, #12] @ R9=ROWr16[6] |

211 |
addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0) |

212 |
subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1) |

213 |
subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2) |

214 |
addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3) |

215 |
@@ W6 alone is no more useful, save W2*ROWr16[6] in it instead |

216 |
teq r9, #0 @ if null avoid muls |

217 |
mulne r11, r10, r9 @ R11=W6*ROWr16[6] |

218 |
addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0) |

219 |
mulne r10, r8, r9 @ R10=W2*ROWr16[6] |

220 |
@@ a0 += W6*row[6]; |

221 |
@@ a3 -= W6*row[6]; |

222 |
@@ a1 -= W2*row[6]; |

223 |
@@ a2 += W2*row[6]; |

224 |
subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3) |

225 |
subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1) |

226 |
addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2) |

227 | |

228 |
__end_a_evaluation: |

229 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |

230 |
@@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |

231 |
@@ R12=__const_ptr_, R14=&block[n] |

232 |
@@ row[0] = (a0 + b0) >> ROW_SHIFT; |

233 |
@@ row[1] = (a1 + b1) >> ROW_SHIFT; |

234 |
@@ row[2] = (a2 + b2) >> ROW_SHIFT; |

235 |
@@ row[3] = (a3 + b3) >> ROW_SHIFT; |

236 |
@@ row[4] = (a3 - b3) >> ROW_SHIFT; |

237 |
@@ row[5] = (a2 - b2) >> ROW_SHIFT; |

238 |
@@ row[6] = (a1 - b1) >> ROW_SHIFT; |

239 |
@@ row[7] = (a0 - b0) >> ROW_SHIFT; |

240 |
add r8, r6, r0 @ R8=a0+b0 |

241 |
add r9, r2, r1 @ R9=a1+b1 |

242 |
@@ put 2 16 bits half-words in a 32bits word |

243 |
@@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!) |

244 |
ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000 |

245 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5) |

246 |
mvn r11, r10 @ R11= NOT R10= 0x0000FFFF |

247 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11) |

248 |
orr r8, r8, r9 |

249 |
str r8, [r14, #0] |

250 | |

251 |
add r8, r3, r5 @ R8=a2+b2 |

252 |
add r9, r4, r7 @ R9=a3+b3 |

253 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5) |

254 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11) |

255 |
orr r8, r8, r9 |

256 |
str r8, [r14, #4] |

257 | |

258 |
sub r8, r4, r7 @ R8=a3-b3 |

259 |
sub r9, r3, r5 @ R9=a2-b2 |

260 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5) |

261 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11) |

262 |
orr r8, r8, r9 |

263 |
str r8, [r14, #8] |

264 | |

265 |
sub r8, r2, r1 @ R8=a1-b1 |

266 |
sub r9, r6, r0 @ R9=a0-b0 |

267 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5) |

268 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11) |

269 |
orr r8, r8, r9 |

270 |
str r8, [r14, #12] |

271 | |

272 |
bal __end_row_loop |

273 | |

274 |
__almost_empty_row: |

275 |
@@ the row was empty, except ROWr16[0], now, management of this special case |

276 |
@@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1], |

277 |
@@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1], |

278 |
@@ R8=0xFFFF (temp), R9-R11 free |

279 |
mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run). |

280 |
sub r8, r8, #1 @ R8 is now ready. |

281 |
and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF |

282 |
orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16) |

283 |
str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5 |

284 |
str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5 |

285 |
str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5 |

286 |
str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5 |

287 | |

288 |
__end_row_loop: |

289 |
@@ at this point, R0-R11 (free) |

290 |
@@ R12=__const_ptr_, R14=&block[n] |

291 |
ldr r0, [sp, #0] @ R0=block |

292 |
teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished. |

293 |
sub r14, r14, #16 |

294 |
bne __row_loop |

295 | |

296 | |

297 | |

298 |
@@ at this point, R0=block, R1-R11 (free) |

299 |
@@ R12=__const_ptr_, R14=&block[n] |

300 |
add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block. |

301 |
__col_loop: |

302 | |

303 |
__b_evaluation2: |

304 |
@@ at this point, R0=block (temp), R1-R11 (free) |

305 |
@@ R12=__const_ptr_, R14=&block[n] |

306 |
@@ proceed with b0-b3 first, followed by a0-a3 |

307 |
@@ MUL16(b0, W1, col[8x1]); |

308 |
@@ MUL16(b1, W3, col[8x1]); |

309 |
@@ MUL16(b2, W5, col[8x1]); |

310 |
@@ MUL16(b3, W7, col[8x1]); |

311 |
@@ MAC16(b0, W3, col[8x3]); |

312 |
@@ MAC16(b1, -W7, col[8x3]); |

313 |
@@ MAC16(b2, -W1, col[8x3]); |

314 |
@@ MAC16(b3, -W5, col[8x3]); |

315 |
ldr r8, [r12, #offW1] @ R8=W1 |

316 |
ldrsh r7, [r14, #16] |

317 |
mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

318 |
ldr r9, [r12, #offW3] @ R9=W3 |

319 |
ldr r10, [r12, #offW5] @ R10=W5 |

320 |
mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

321 |
ldr r11, [r12, #offW7] @ R11=W7 |

322 |
mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

323 |
ldrsh r2, [r14, #48] |

324 |
mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |

325 |
teq r2, #0 @ if 0, then avoid muls |

326 |
mlane r0, r9, r2, r0 @ R0+=W3*ROWr16[3]=b0 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

327 |
rsbne r2, r2, #0 @ R2=-ROWr16[3] |

328 |
mlane r1, r11, r2, r1 @ R1-=W7*ROWr16[3]=b1 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

329 |
mlane r5, r8, r2, r5 @ R5-=W1*ROWr16[3]=b2 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

330 |
mlane r7, r10, r2, r7 @ R7-=W5*ROWr16[3]=b3 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle) |

331 | |

332 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free), |

333 |
@@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |

334 |
@@ R12=__const_ptr_, R14=&block[n] |

335 |
@@ MAC16(b0, W5, col[5x8]); |

336 |
@@ MAC16(b2, W7, col[5x8]); |

337 |
@@ MAC16(b3, W3, col[5x8]); |

338 |
@@ MAC16(b1, -W1, col[5x8]); |

339 |
@@ MAC16(b0, W7, col[7x8]); |

340 |
@@ MAC16(b2, W3, col[7x8]); |

341 |
@@ MAC16(b3, -W1, col[7x8]); |

342 |
@@ MAC16(b1, -W5, col[7x8]); |

343 |
ldrsh r3, [r14, #80] @ R3=COLr16[5x8] |

344 |
teq r3, #0 @ if 0 then avoid muls |

345 |
mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0 |

346 |
mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2 |

347 |
mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3 |

348 |
rsbne r3, r3, #0 @ R3=-ROWr16[5x8] |

349 |
ldrsh r4, [r14, #112] @ R4=COLr16[7x8] |

350 |
mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1 |

351 |
@@ R3 is free now |

352 |
teq r4, #0 @ if 0 then avoid muls |

353 |
mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0 |

354 |
mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2 |

355 |
rsbne r4, r4, #0 @ R4=-ROWr16[7x8] |

356 |
mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3 |

357 |
mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1 |

358 |
@@ R4 is free now |

359 |
__end_b_evaluation2: |

360 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free), |

361 |
@@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |

362 |
@@ R12=__const_ptr_, R14=&block[n] |

363 | |

364 |
__a_evaluation2: |

365 |
@@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1)); |

366 |
@@ a1 = a0 + W6 * row[2]; |

367 |
@@ a2 = a0 - W6 * row[2]; |

368 |
@@ a3 = a0 - W2 * row[2]; |

369 |
@@ a0 = a0 + W2 * row[2]; |

370 |
ldrsh r6, [r14, #0] |

371 |
ldr r9, [r12, #offW4] @ R9=W4 |

372 |
mul r6, r9, r6 @ R6=W4*ROWr16[0] |

373 |
ldr r10, [r12, #offW6] @ R10=W6 |

374 |
ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet) |

375 |
add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0) |

376 |
mul r11, r10, r4 @ R11=W6*ROWr16[2] |

377 |
ldr r8, [r12, #offW2] @ R8=W2 |

378 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |

379 |
sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2) |

380 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |

381 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |

382 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |

383 | |

384 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |

385 |
@@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free), |

386 |
@@ R12=__const_ptr_, R14=&block[n] |

387 |
@@ a0 += W4*row[4] |

388 |
@@ a1 -= W4*row[4] |

389 |
@@ a2 -= W4*row[4] |

390 |
@@ a3 += W4*row[4] |

391 |
ldrsh r11, [r14, #64] @ R11=ROWr16[4] |

392 |
teq r11, #0 @ if null avoid muls |

393 |
mulne r11, r9, r11 @ R11=W4*ROWr16[4] |

394 |
@@ R9 is free now |

395 |
addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0) |

396 |
subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1) |

397 |
subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2) |

398 |
ldrsh r9, [r14, #96] @ R9=ROWr16[6] |

399 |
addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3) |

400 |
@@ W6 alone is no more useful, save W2*ROWr16[6] in it instead |

401 |
teq r9, #0 @ if null avoid muls |

402 |
mulne r11, r10, r9 @ R11=W6*ROWr16[6] |

403 |
addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0) |

404 |
mulne r10, r8, r9 @ R10=W2*ROWr16[6] |

405 |
@@ a0 += W6*row[6]; |

406 |
@@ a3 -= W6*row[6]; |

407 |
@@ a1 -= W2*row[6]; |

408 |
@@ a2 += W2*row[6]; |

409 |
subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3) |

410 |
subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1) |

411 |
addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2) |

412 |
__end_a_evaluation2: |

413 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |

414 |
@@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |

415 |
@@ R12=__const_ptr_, R14=&block[n] |

416 |
@@ col[0 ] = ((a0 + b0) >> COL_SHIFT); |

417 |
@@ col[8 ] = ((a1 + b1) >> COL_SHIFT); |

418 |
@@ col[16] = ((a2 + b2) >> COL_SHIFT); |

419 |
@@ col[24] = ((a3 + b3) >> COL_SHIFT); |

420 |
@@ col[32] = ((a3 - b3) >> COL_SHIFT); |

421 |
@@ col[40] = ((a2 - b2) >> COL_SHIFT); |

422 |
@@ col[48] = ((a1 - b1) >> COL_SHIFT); |

423 |
@@ col[56] = ((a0 - b0) >> COL_SHIFT); |

424 |
@@@@@ no optimization here @@@@@ |

425 |
add r8, r6, r0 @ R8=a0+b0 |

426 |
add r9, r2, r1 @ R9=a1+b1 |

427 |
mov r8, r8, asr #COL_SHIFT |

428 |
mov r9, r9, asr #COL_SHIFT |

429 |
strh r8, [r14, #0] |

430 |
strh r9, [r14, #16] |

431 |
add r8, r3, r5 @ R8=a2+b2 |

432 |
add r9, r4, r7 @ R9=a3+b3 |

433 |
mov r8, r8, asr #COL_SHIFT |

434 |
mov r9, r9, asr #COL_SHIFT |

435 |
strh r8, [r14, #32] |

436 |
strh r9, [r14, #48] |

437 |
sub r8, r4, r7 @ R8=a3-b3 |

438 |
sub r9, r3, r5 @ R9=a2-b2 |

439 |
mov r8, r8, asr #COL_SHIFT |

440 |
mov r9, r9, asr #COL_SHIFT |

441 |
strh r8, [r14, #64] |

442 |
strh r9, [r14, #80] |

443 |
sub r8, r2, r1 @ R8=a1-b1 |

444 |
sub r9, r6, r0 @ R9=a0-b0 |

445 |
mov r8, r8, asr #COL_SHIFT |

446 |
mov r9, r9, asr #COL_SHIFT |

447 |
strh r8, [r14, #96] |

448 |
strh r9, [r14, #112] |

449 | |

450 |
__end_col_loop: |

451 |
@@ at this point, R0-R11 (free) |

452 |
@@ R12=__const_ptr_, R14=&block[n] |

453 |
ldr r0, [sp, #0] @ R0=block |

454 |
teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished. |

455 |
sub r14, r14, #2 |

456 |
bne __col_loop |

457 | |

458 | |

459 | |

460 | |

461 |
__end_simple_idct_arm: |

462 |
@@ restore registers to previous status! |

463 |
add sp, sp, #8 @@ the local variables! |

464 |
ldmfd sp!, {r4-r11, r15} @@ update PC with LR content. |

465 | |

466 | |

467 | |

468 |
@@ kind of sub-function, here not to overload the common case. |

469 |
__end_bef_a_evaluation: |

470 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |

471 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |

472 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |

473 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |

474 |
bal __end_a_evaluation |

475 | |

476 | |

477 |
__constant_ptr__: @@ see #defines at the beginning of the source code for values. |

478 |
.align |

479 |
.word W1 |

480 |
.word W2 |

481 |
.word W3 |

482 |
.word W4 |

483 |
.word W5 |

484 |
.word W6 |

485 |
.word W7 |

486 |
.word MASK_MSHW |