ffmpeg / libavcodec / armv4l / simple_idct_arm.S @ 5509bffa
History | View | Annotate | Download (21.5 KB)
| 1 |
/* |
|---|---|
| 2 |
* simple_idct_arm.S |
| 3 |
* Copyright (C) 2002 Frederic 'dilb' Boulay. |
| 4 |
* All Rights Reserved. |
| 5 |
* |
| 6 |
* Author: Frederic Boulay <dilb@handhelds.org> |
| 7 |
* |
| 8 |
* You can redistribute this file and/or modify |
| 9 |
* it under the terms of the GNU General Public License (version 2) |
| 10 |
* as published by the Free Software Foundation. |
| 11 |
* |
| 12 |
* This file is distributed in the hope that it will be useful, |
| 13 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 15 |
* GNU General Public License for more details. |
| 16 |
* |
| 17 |
* You should have received a copy of the GNU General Public License |
| 18 |
* along with this library; if not, write to the Free Software |
| 19 |
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 20 |
* |
| 21 |
* |
| 22 |
* The function defined in this file, is derived from the simple_idct function |
| 23 |
* from the libavcodec library part of the ffmpeg project. |
| 24 |
*/ |
| 25 |
|
| 26 |
/* useful constants for the algorithm, they are save in __constant_ptr__ at */ |
| 27 |
/* the end of the source code.*/ |
| 28 |
#define W1 22725 |
| 29 |
#define W2 21407 |
| 30 |
#define W3 19266 |
| 31 |
#define W4 16383 |
| 32 |
#define W5 12873 |
| 33 |
#define W6 8867 |
| 34 |
#define W7 4520 |
| 35 |
#define MASK_MSHW 0xFFFF0000 |
| 36 |
|
| 37 |
/* offsets of the constants in the vector */ |
| 38 |
#define offW1 0 |
| 39 |
#define offW2 4 |
| 40 |
#define offW3 8 |
| 41 |
#define offW4 12 |
| 42 |
#define offW5 16 |
| 43 |
#define offW6 20 |
| 44 |
#define offW7 24 |
| 45 |
#define offMASK_MSHW 28 |
| 46 |
|
| 47 |
#define ROW_SHIFT 11 |
| 48 |
#define ROW_SHIFT2MSHW (16-11) |
| 49 |
#define COL_SHIFT 20 |
| 50 |
#define ROW_SHIFTED_1 1024 /* 1<< (ROW_SHIFT-1) */ |
| 51 |
#define COL_SHIFTED_1 524288 /* 1<< (COL_SHIFT-1) */ |
| 52 |
|
| 53 |
|
| 54 |
.text |
| 55 |
.align |
| 56 |
.global simple_idct_ARM |
| 57 |
|
| 58 |
simple_idct_ARM: |
| 59 |
@@ void simple_idct_ARM(int16_t *block) |
| 60 |
@@ save stack for reg needed (take all of them), |
| 61 |
@@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block |
| 62 |
@@ so it must not be overwritten, if it is not saved!! |
| 63 |
@@ R12 is another scratch register, so it should not be saved too |
| 64 |
@@ save all registers |
| 65 |
stmfd sp!, {r4-r11, r14} @ R14 is also called LR
|
| 66 |
@@ at this point, R0=block, other registers are free. |
| 67 |
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. |
| 68 |
add r12, pc, #(__constant_ptr__-.-8) @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it |
| 69 |
@@ add 2 temporary variables in the stack: R0 and R14 |
| 70 |
sub sp, sp, #8 @ allow 2 local variables |
| 71 |
str r0, [sp, #0] @ save block in sp[0] |
| 72 |
@@ stack status |
| 73 |
@@ sp+4 free |
| 74 |
@@ sp+0 R0 (block) |
| 75 |
|
| 76 |
|
| 77 |
@@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free |
| 78 |
|
| 79 |
|
| 80 |
__row_loop: |
| 81 |
@@ read the row and check if it is null, almost null, or not, according to strongarm specs, it is not necessary to optimise ldr accesses (i.e. split 32bits in 2 16bits words), at least it gives more usable registers :) |
| 82 |
ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer) |
| 83 |
ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1] |
| 84 |
ldr r3, [r14, #8] @ R3=ROWr32[2] |
| 85 |
ldr r4, [r14, #12] @ R4=ROWr32[3] |
| 86 |
@@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop), |
| 87 |
@@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row) |
| 88 |
@@ else follow the complete algorithm. |
| 89 |
@@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1], |
| 90 |
@@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free |
| 91 |
orr r5, r4, r3 @ R5=R4 | R3 |
| 92 |
orr r5, r5, r2 @ R5=R4 | R3 | R2 |
| 93 |
orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null) |
| 94 |
beq __end_row_loop |
| 95 |
mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later) |
| 96 |
ldrsh r6, [r14, #0] @ R6=ROWr16[0] |
| 97 |
orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7 |
| 98 |
beq __almost_empty_row |
| 99 |
|
| 100 |
__b_evaluation: |
| 101 |
@@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3], |
| 102 |
@@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free, |
| 103 |
@@ R12=__const_ptr_, R14=&block[n] |
| 104 |
@@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3 |
| 105 |
|
| 106 |
@@ MUL16(b0, W1, row[1]); |
| 107 |
@@ MUL16(b1, W3, row[1]); |
| 108 |
@@ MUL16(b2, W5, row[1]); |
| 109 |
@@ MUL16(b3, W7, row[1]); |
| 110 |
@@ MAC16(b0, W3, row[3]); |
| 111 |
@@ MAC16(b1, -W7, row[3]); |
| 112 |
@@ MAC16(b2, -W1, row[3]); |
| 113 |
@@ MAC16(b3, -W5, row[3]); |
| 114 |
ldr r8, [r12, #offW1] @ R8=W1 |
| 115 |
mov r2, r2, asr #16 @ R2=ROWr16[3] |
| 116 |
mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 117 |
ldr r9, [r12, #offW3] @ R9=W3 |
| 118 |
ldr r10, [r12, #offW5] @ R10=W5 |
| 119 |
mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 120 |
ldr r11, [r12, #offW7] @ R11=W7 |
| 121 |
mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 122 |
mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 123 |
teq r2, #0 @ if null avoid muls |
| 124 |
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) |
| 125 |
rsbne r2, r2, #0 @ R2=-ROWr16[3] |
| 126 |
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) |
| 127 |
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) |
| 128 |
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) |
| 129 |
|
| 130 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3], |
| 131 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |
| 132 |
@@ R12=__const_ptr_, R14=&block[n] |
| 133 |
@@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3]; |
| 134 |
@@ if (temp != 0) {}
|
| 135 |
orrs r2, r3, r4 @ R2=ROWr32[2] | ROWr32[3] |
| 136 |
beq __end_b_evaluation |
| 137 |
|
| 138 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3], |
| 139 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |
| 140 |
@@ R12=__const_ptr_, R14=&block[n] |
| 141 |
@@ MAC16(b0, W5, row[5]); |
| 142 |
@@ MAC16(b2, W7, row[5]); |
| 143 |
@@ MAC16(b3, W3, row[5]); |
| 144 |
@@ MAC16(b1, -W1, row[5]); |
| 145 |
@@ MAC16(b0, W7, row[7]); |
| 146 |
@@ MAC16(b2, W3, row[7]); |
| 147 |
@@ MAC16(b3, -W1, row[7]); |
| 148 |
@@ MAC16(b1, -W5, row[7]); |
| 149 |
mov r3, r3, asr #16 @ R3=ROWr16[5] |
| 150 |
teq r3, #0 @ if null avoid muls |
| 151 |
mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0 |
| 152 |
mov r4, r4, asr #16 @ R4=ROWr16[7] |
| 153 |
mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5]=b2 |
| 154 |
mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5]=b3 |
| 155 |
rsbne r3, r3, #0 @ R3=-ROWr16[5] |
| 156 |
mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5]=b1 |
| 157 |
@@ R3 is free now |
| 158 |
teq r4, #0 @ if null avoid muls |
| 159 |
mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0 |
| 160 |
mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2 |
| 161 |
rsbne r4, r4, #0 @ R4=-ROWr16[7] |
| 162 |
mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3 |
| 163 |
mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1 |
| 164 |
@@ R4 is free now |
| 165 |
__end_b_evaluation: |
| 166 |
@@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free), |
| 167 |
@@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |
| 168 |
@@ R12=__const_ptr_, R14=&block[n] |
| 169 |
|
| 170 |
__a_evaluation: |
| 171 |
@@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1)); |
| 172 |
@@ a1 = a0 + W6 * row[2]; |
| 173 |
@@ a2 = a0 - W6 * row[2]; |
| 174 |
@@ a3 = a0 - W2 * row[2]; |
| 175 |
@@ a0 = a0 + W2 * row[2]; |
| 176 |
ldr r9, [r12, #offW4] @ R9=W4 |
| 177 |
mul r6, r9, r6 @ R6=W4*ROWr16[0] |
| 178 |
ldr r10, [r12, #offW6] @ R10=W6 |
| 179 |
ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet) |
| 180 |
add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0) |
| 181 |
|
| 182 |
mul r11, r10, r4 @ R11=W6*ROWr16[2] |
| 183 |
ldr r8, [r12, #offW2] @ R8=W2 |
| 184 |
sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2) |
| 185 |
@@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3]; |
| 186 |
@@ if (temp != 0) {}
|
| 187 |
teq r2, #0 |
| 188 |
beq __end_bef_a_evaluation |
| 189 |
|
| 190 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |
| 191 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |
| 192 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |
| 193 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |
| 194 |
|
| 195 |
|
| 196 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |
| 197 |
@@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free), |
| 198 |
@@ R12=__const_ptr_, R14=&block[n] |
| 199 |
|
| 200 |
|
| 201 |
@@ a0 += W4*row[4] |
| 202 |
@@ a1 -= W4*row[4] |
| 203 |
@@ a2 -= W4*row[4] |
| 204 |
@@ a3 += W4*row[4] |
| 205 |
ldrsh r11, [r14, #8] @ R11=ROWr16[4] |
| 206 |
teq r11, #0 @ if null avoid muls |
| 207 |
mulne r11, r9, r11 @ R11=W4*ROWr16[4] |
| 208 |
@@ R9 is free now |
| 209 |
ldrsh r9, [r14, #12] @ R9=ROWr16[6] |
| 210 |
addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0) |
| 211 |
subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1) |
| 212 |
subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2) |
| 213 |
addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3) |
| 214 |
@@ W6 alone is no more useful, save W2*ROWr16[6] in it instead |
| 215 |
teq r9, #0 @ if null avoid muls |
| 216 |
mulne r11, r10, r9 @ R11=W6*ROWr16[6] |
| 217 |
addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0) |
| 218 |
mulne r10, r8, r9 @ R10=W2*ROWr16[6] |
| 219 |
@@ a0 += W6*row[6]; |
| 220 |
@@ a3 -= W6*row[6]; |
| 221 |
@@ a1 -= W2*row[6]; |
| 222 |
@@ a2 += W2*row[6]; |
| 223 |
subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3) |
| 224 |
subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1) |
| 225 |
addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2) |
| 226 |
|
| 227 |
__end_a_evaluation: |
| 228 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |
| 229 |
@@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |
| 230 |
@@ R12=__const_ptr_, R14=&block[n] |
| 231 |
@@ row[0] = (a0 + b0) >> ROW_SHIFT; |
| 232 |
@@ row[1] = (a1 + b1) >> ROW_SHIFT; |
| 233 |
@@ row[2] = (a2 + b2) >> ROW_SHIFT; |
| 234 |
@@ row[3] = (a3 + b3) >> ROW_SHIFT; |
| 235 |
@@ row[4] = (a3 - b3) >> ROW_SHIFT; |
| 236 |
@@ row[5] = (a2 - b2) >> ROW_SHIFT; |
| 237 |
@@ row[6] = (a1 - b1) >> ROW_SHIFT; |
| 238 |
@@ row[7] = (a0 - b0) >> ROW_SHIFT; |
| 239 |
add r8, r6, r0 @ R8=a0+b0 |
| 240 |
add r9, r2, r1 @ R9=a1+b1 |
| 241 |
@@ put 2 16 bits half-words in a 32bits word |
| 242 |
@@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!) |
| 243 |
ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000 |
| 244 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5) |
| 245 |
mvn r11, r10 @ R11= NOT R10= 0x0000FFFF |
| 246 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11) |
| 247 |
orr r8, r8, r9 |
| 248 |
str r8, [r14, #0] |
| 249 |
|
| 250 |
add r8, r3, r5 @ R8=a2+b2 |
| 251 |
add r9, r4, r7 @ R9=a3+b3 |
| 252 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5) |
| 253 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11) |
| 254 |
orr r8, r8, r9 |
| 255 |
str r8, [r14, #4] |
| 256 |
|
| 257 |
sub r8, r4, r7 @ R8=a3-b3 |
| 258 |
sub r9, r3, r5 @ R9=a2-b2 |
| 259 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5) |
| 260 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11) |
| 261 |
orr r8, r8, r9 |
| 262 |
str r8, [r14, #8] |
| 263 |
|
| 264 |
sub r8, r2, r1 @ R8=a1-b1 |
| 265 |
sub r9, r6, r0 @ R9=a0-b0 |
| 266 |
and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5) |
| 267 |
and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11) |
| 268 |
orr r8, r8, r9 |
| 269 |
str r8, [r14, #12] |
| 270 |
|
| 271 |
bal __end_row_loop |
| 272 |
|
| 273 |
__almost_empty_row: |
| 274 |
@@ the row was empty, except ROWr16[0], now, management of this special case |
| 275 |
@@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1], |
| 276 |
@@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1], |
| 277 |
@@ R8=0xFFFF (temp), R9-R11 free |
| 278 |
mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run). |
| 279 |
sub r8, r8, #1 @ R8 is now ready. |
| 280 |
and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF |
| 281 |
orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16) |
| 282 |
str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5 |
| 283 |
str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5 |
| 284 |
str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5 |
| 285 |
str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5 |
| 286 |
|
| 287 |
__end_row_loop: |
| 288 |
@@ at this point, R0-R11 (free) |
| 289 |
@@ R12=__const_ptr_, R14=&block[n] |
| 290 |
ldr r0, [sp, #0] @ R0=block |
| 291 |
teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished. |
| 292 |
sub r14, r14, #16 |
| 293 |
bne __row_loop |
| 294 |
|
| 295 |
|
| 296 |
|
| 297 |
@@ at this point, R0=block, R1-R11 (free) |
| 298 |
@@ R12=__const_ptr_, R14=&block[n] |
| 299 |
add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block. |
| 300 |
__col_loop: |
| 301 |
|
| 302 |
__b_evaluation2: |
| 303 |
@@ at this point, R0=block (temp), R1-R11 (free) |
| 304 |
@@ R12=__const_ptr_, R14=&block[n] |
| 305 |
@@ proceed with b0-b3 first, followed by a0-a3 |
| 306 |
@@ MUL16(b0, W1, col[8x1]); |
| 307 |
@@ MUL16(b1, W3, col[8x1]); |
| 308 |
@@ MUL16(b2, W5, col[8x1]); |
| 309 |
@@ MUL16(b3, W7, col[8x1]); |
| 310 |
@@ MAC16(b0, W3, col[8x3]); |
| 311 |
@@ MAC16(b1, -W7, col[8x3]); |
| 312 |
@@ MAC16(b2, -W1, col[8x3]); |
| 313 |
@@ MAC16(b3, -W5, col[8x3]); |
| 314 |
ldr r8, [r12, #offW1] @ R8=W1 |
| 315 |
ldrsh r7, [r14, #16] |
| 316 |
mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 317 |
ldr r9, [r12, #offW3] @ R9=W3 |
| 318 |
ldr r10, [r12, #offW5] @ R10=W5 |
| 319 |
mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 320 |
ldr r11, [r12, #offW7] @ R11=W7 |
| 321 |
mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 322 |
ldrsh r2, [r14, #48] |
| 323 |
mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle) |
| 324 |
teq r2, #0 @ if 0, then avoid muls |
| 325 |
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) |
| 326 |
rsbne r2, r2, #0 @ R2=-ROWr16[3] |
| 327 |
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) |
| 328 |
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) |
| 329 |
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) |
| 330 |
|
| 331 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free), |
| 332 |
@@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7, |
| 333 |
@@ R12=__const_ptr_, R14=&block[n] |
| 334 |
@@ MAC16(b0, W5, col[5x8]); |
| 335 |
@@ MAC16(b2, W7, col[5x8]); |
| 336 |
@@ MAC16(b3, W3, col[5x8]); |
| 337 |
@@ MAC16(b1, -W1, col[5x8]); |
| 338 |
@@ MAC16(b0, W7, col[7x8]); |
| 339 |
@@ MAC16(b2, W3, col[7x8]); |
| 340 |
@@ MAC16(b3, -W1, col[7x8]); |
| 341 |
@@ MAC16(b1, -W5, col[7x8]); |
| 342 |
ldrsh r3, [r14, #80] @ R3=COLr16[5x8] |
| 343 |
teq r3, #0 @ if 0 then avoid muls |
| 344 |
mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0 |
| 345 |
mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2 |
| 346 |
mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3 |
| 347 |
rsbne r3, r3, #0 @ R3=-ROWr16[5x8] |
| 348 |
ldrsh r4, [r14, #112] @ R4=COLr16[7x8] |
| 349 |
mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1 |
| 350 |
@@ R3 is free now |
| 351 |
teq r4, #0 @ if 0 then avoid muls |
| 352 |
mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0 |
| 353 |
mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2 |
| 354 |
rsbne r4, r4, #0 @ R4=-ROWr16[7x8] |
| 355 |
mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3 |
| 356 |
mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1 |
| 357 |
@@ R4 is free now |
| 358 |
__end_b_evaluation2: |
| 359 |
@@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free), |
| 360 |
@@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |
| 361 |
@@ R12=__const_ptr_, R14=&block[n] |
| 362 |
|
| 363 |
__a_evaluation2: |
| 364 |
@@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1)); |
| 365 |
@@ a1 = a0 + W6 * row[2]; |
| 366 |
@@ a2 = a0 - W6 * row[2]; |
| 367 |
@@ a3 = a0 - W2 * row[2]; |
| 368 |
@@ a0 = a0 + W2 * row[2]; |
| 369 |
ldrsh r6, [r14, #0] |
| 370 |
ldr r9, [r12, #offW4] @ R9=W4 |
| 371 |
mul r6, r9, r6 @ R6=W4*ROWr16[0] |
| 372 |
ldr r10, [r12, #offW6] @ R10=W6 |
| 373 |
ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet) |
| 374 |
add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0) |
| 375 |
mul r11, r10, r4 @ R11=W6*ROWr16[2] |
| 376 |
ldr r8, [r12, #offW2] @ R8=W2 |
| 377 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |
| 378 |
sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2) |
| 379 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |
| 380 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |
| 381 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |
| 382 |
|
| 383 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |
| 384 |
@@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free), |
| 385 |
@@ R12=__const_ptr_, R14=&block[n] |
| 386 |
@@ a0 += W4*row[4] |
| 387 |
@@ a1 -= W4*row[4] |
| 388 |
@@ a2 -= W4*row[4] |
| 389 |
@@ a3 += W4*row[4] |
| 390 |
ldrsh r11, [r14, #64] @ R11=ROWr16[4] |
| 391 |
teq r11, #0 @ if null avoid muls |
| 392 |
mulne r11, r9, r11 @ R11=W4*ROWr16[4] |
| 393 |
@@ R9 is free now |
| 394 |
addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0) |
| 395 |
subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1) |
| 396 |
subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2) |
| 397 |
ldrsh r9, [r14, #96] @ R9=ROWr16[6] |
| 398 |
addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3) |
| 399 |
@@ W6 alone is no more useful, save W2*ROWr16[6] in it instead |
| 400 |
teq r9, #0 @ if null avoid muls |
| 401 |
mulne r11, r10, r9 @ R11=W6*ROWr16[6] |
| 402 |
addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0) |
| 403 |
mulne r10, r8, r9 @ R10=W2*ROWr16[6] |
| 404 |
@@ a0 += W6*row[6]; |
| 405 |
@@ a3 -= W6*row[6]; |
| 406 |
@@ a1 -= W2*row[6]; |
| 407 |
@@ a2 += W2*row[6]; |
| 408 |
subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3) |
| 409 |
subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1) |
| 410 |
addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2) |
| 411 |
__end_a_evaluation2: |
| 412 |
@@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3, |
| 413 |
@@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free), |
| 414 |
@@ R12=__const_ptr_, R14=&block[n] |
| 415 |
@@ col[0 ] = ((a0 + b0) >> COL_SHIFT); |
| 416 |
@@ col[8 ] = ((a1 + b1) >> COL_SHIFT); |
| 417 |
@@ col[16] = ((a2 + b2) >> COL_SHIFT); |
| 418 |
@@ col[24] = ((a3 + b3) >> COL_SHIFT); |
| 419 |
@@ col[32] = ((a3 - b3) >> COL_SHIFT); |
| 420 |
@@ col[40] = ((a2 - b2) >> COL_SHIFT); |
| 421 |
@@ col[48] = ((a1 - b1) >> COL_SHIFT); |
| 422 |
@@ col[56] = ((a0 - b0) >> COL_SHIFT); |
| 423 |
@@@@@ no optimisation here @@@@@ |
| 424 |
add r8, r6, r0 @ R8=a0+b0 |
| 425 |
add r9, r2, r1 @ R9=a1+b1 |
| 426 |
mov r8, r8, asr #COL_SHIFT |
| 427 |
mov r9, r9, asr #COL_SHIFT |
| 428 |
strh r8, [r14, #0] |
| 429 |
strh r9, [r14, #16] |
| 430 |
add r8, r3, r5 @ R8=a2+b2 |
| 431 |
add r9, r4, r7 @ R9=a3+b3 |
| 432 |
mov r8, r8, asr #COL_SHIFT |
| 433 |
mov r9, r9, asr #COL_SHIFT |
| 434 |
strh r8, [r14, #32] |
| 435 |
strh r9, [r14, #48] |
| 436 |
sub r8, r4, r7 @ R8=a3-b3 |
| 437 |
sub r9, r3, r5 @ R9=a2-b2 |
| 438 |
mov r8, r8, asr #COL_SHIFT |
| 439 |
mov r9, r9, asr #COL_SHIFT |
| 440 |
strh r8, [r14, #64] |
| 441 |
strh r9, [r14, #80] |
| 442 |
sub r8, r2, r1 @ R8=a1-b1 |
| 443 |
sub r9, r6, r0 @ R9=a0-b0 |
| 444 |
mov r8, r8, asr #COL_SHIFT |
| 445 |
mov r9, r9, asr #COL_SHIFT |
| 446 |
strh r8, [r14, #96] |
| 447 |
strh r9, [r14, #112] |
| 448 |
|
| 449 |
__end_col_loop: |
| 450 |
@@ at this point, R0-R11 (free) |
| 451 |
@@ R12=__const_ptr_, R14=&block[n] |
| 452 |
ldr r0, [sp, #0] @ R0=block |
| 453 |
teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished. |
| 454 |
sub r14, r14, #2 |
| 455 |
bne __col_loop |
| 456 |
|
| 457 |
|
| 458 |
|
| 459 |
|
| 460 |
__end_simple_idct_ARM: |
| 461 |
@@ restore registers to previous status! |
| 462 |
add sp, sp, #8 @@ the local variables! |
| 463 |
ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.
|
| 464 |
|
| 465 |
|
| 466 |
|
| 467 |
@@ kind of sub-function, here not to overload the common case. |
| 468 |
__end_bef_a_evaluation: |
| 469 |
add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1) |
| 470 |
mul r11, r8, r4 @ R11=W2*ROWr16[2] |
| 471 |
sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3) |
| 472 |
add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0) |
| 473 |
bal __end_a_evaluation |
| 474 |
|
| 475 |
|
| 476 |
__constant_ptr__: @@ see #defines at the beginning of the source code for values. |
| 477 |
.align |
| 478 |
.word W1 |
| 479 |
.word W2 |
| 480 |
.word W3 |
| 481 |
.word W4 |
| 482 |
.word W5 |
| 483 |
.word W6 |
| 484 |
.word W7 |
| 485 |
.word MASK_MSHW |