simple_idct_arm.S 21.2 KB
Newer Older
1
/*
2
 * Copyright (C) 2002 Frederic 'dilb' Boulay
3 4 5
 *
 * Author: Frederic Boulay <dilb@handhelds.org>
 *
6 7 8
 * The function defined in this file is derived from the simple_idct function
 * from the libavcodec library part of the FFmpeg project.
 *
9 10 11
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
12 13
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
14
 * version 2.1 of the License, or (at your option) any later version.
15
 *
16
 * FFmpeg is distributed in the hope that it will be useful,
17
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 19
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
20
 *
21
 * You should have received a copy of the GNU Lesser General Public
22
 * License along with FFmpeg; if not, write to the Free Software
23
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 25
 */

26
#include "libavutil/arm/asm.S"
27

28
/* useful constants for the algorithm */
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
#define W1  22725
#define W2  21407
#define W3  19266
#define W4  16383
#define W5  12873
#define W6  8867
#define W7  4520
#define MASK_MSHW 0xFFFF0000

#define ROW_SHIFT 11
#define ROW_SHIFT2MSHW (16-11)
#define COL_SHIFT 20
#define ROW_SHIFTED_1 1024 /* 1<< (ROW_SHIFT-1) */
#define COL_SHIFTED_1 524288 /* 1<< (COL_SHIFT-1) */


45 46
function ff_simple_idct_arm, export=1
        @@ void simple_idct_arm(int16_t *block)
47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
        @@ save stack for reg needed (take all of them),
        @@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block
        @@ so it must not be overwritten, if it is not saved!!
        @@ R12 is another scratch register, so it should not be saved too
        @@ save all registers
        stmfd sp!, {r4-r11, r14} @ R14 is also called LR
        @@ at this point, R0=block, other registers are free.
        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.
        @@ add 2 temporary variables in the stack: R0 and R14
        sub sp, sp, #8          @ allow 2 local variables
        str r0, [sp, #0]        @ save block in sp[0]
        @@ stack status
        @@ sp+4   free
        @@ sp+0   R0  (block)


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


__row_loop:
67
        @@ 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 :)
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
        ldr r1, [r14, #0]        @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer)
        ldr r2, [r14, #4]        @ R2=(int32)(R12)[1]=ROWr32[1]
        ldr r3, [r14, #8]        @ R3=ROWr32[2]
        ldr r4, [r14, #12]       @ R4=ROWr32[3]
        @@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop),
        @@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row)
        @@ else follow the complete algorithm.
        @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
        @@                R3=ROWr32[2], R4=ROWr32[3], R5-R11 free
        orr r5, r4, r3           @ R5=R4 | R3
        orr r5, r5, r2           @ R5=R4 | R3 | R2
        orrs r6, r5, r1          @ Test R5 | R1 (the aim is to check if everything is null)
        beq __end_row_loop
        mov r7, r1, asr #16      @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later)
        ldrsh r6, [r14, #0]      @ R6=ROWr16[0]
        orrs r5, r5, r7          @ R5=R4 | R3 | R2 | R7
        beq __almost_empty_row

86
@@ __b_evaluation:
87 88 89 90 91 92 93 94 95 96 97 98 99
        @@ at this point, R0=block (temp),  R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3],
        @@     R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free,
        @@     R12=__const_ptr_, R14=&block[n]
        @@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3

        @@ MUL16(b0, W1, row[1]);
        @@ MUL16(b1, W3, row[1]);
        @@ MUL16(b2, W5, row[1]);
        @@ MUL16(b3, W7, row[1]);
        @@ MAC16(b0, W3, row[3]);
        @@ MAC16(b1, -W7, row[3]);
        @@ MAC16(b2, -W1, row[3]);
        @@ MAC16(b3, -W5, row[3]);
100
        ldr r8, =W1              @ R8=W1
101 102
        mov r2, r2, asr #16      @ R2=ROWr16[3]
        mul r0, r8, r7           @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
103 104
        ldr r9, =W3              @ R9=W3
        ldr r10, =W5             @ R10=W5
105
        mul r1, r9, r7           @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
106
        ldr r11, =W7             @ R11=W7
107 108
        mul r5, r10, r7          @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
        mul r7, r11, r7          @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
109 110 111
        teq r2, #0               @ if null avoid muls
        itttt ne
        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)
112 113 114
        rsbne r2, r2, #0         @ R2=-ROWr16[3]
        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)
        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)
115
        it    ne
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
        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)

        @@ at this point, R0=b0,  R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
        @@     R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
        @@     R12=__const_ptr_, R14=&block[n]
        @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
        @@ if (temp != 0) {}
        orrs r2, r3, r4          @ R2=ROWr32[2] | ROWr32[3]
        beq __end_b_evaluation

        @@ at this point, R0=b0,  R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
        @@     R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
        @@     R12=__const_ptr_, R14=&block[n]
        @@ MAC16(b0, W5, row[5]);
        @@ MAC16(b2, W7, row[5]);
        @@ MAC16(b3, W3, row[5]);
        @@ MAC16(b1, -W1, row[5]);
        @@ MAC16(b0, W7, row[7]);
        @@ MAC16(b2, W3, row[7]);
        @@ MAC16(b3, -W1, row[7]);
        @@ MAC16(b1, -W5, row[7]);
        mov r3, r3, asr #16      @ R3=ROWr16[5]
138 139
        teq r3, #0               @ if null avoid muls
        it    ne
140 141
        mlane r0, r10, r3, r0    @ R0+=W5*ROWr16[5]=b0
        mov r4, r4, asr #16      @ R4=ROWr16[7]
142
        itttt ne
143 144 145 146 147
        mlane r5, r11, r3, r5    @ R5+=W7*ROWr16[5]=b2
        mlane r7, r9, r3, r7     @ R7+=W3*ROWr16[5]=b3
        rsbne r3, r3, #0         @ R3=-ROWr16[5]
        mlane r1, r8, r3, r1     @ R7-=W1*ROWr16[5]=b1
        @@ R3 is free now
148 149
        teq r4, #0               @ if null avoid muls
        itttt ne
150 151 152 153
        mlane r0, r11, r4, r0    @ R0+=W7*ROWr16[7]=b0
        mlane r5, r9, r4, r5     @ R5+=W3*ROWr16[7]=b2
        rsbne r4, r4, #0         @ R4=-ROWr16[7]
        mlane r7, r8, r4, r7     @ R7-=W1*ROWr16[7]=b3
154
        it    ne
155 156 157 158 159 160 161
        mlane r1, r10, r4, r1    @ R1-=W5*ROWr16[7]=b1
        @@ R4 is free now
__end_b_evaluation:
        @@ at this point, R0=b0,  R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free),
        @@     R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]

162
@@ __a_evaluation:
163 164 165 166 167
        @@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
        @@ a1 = a0 + W6 * row[2];
        @@ a2 = a0 - W6 * row[2];
        @@ a3 = a0 - W2 * row[2];
        @@ a0 = a0 + W2 * row[2];
168
        ldr r9, =W4              @ R9=W4
169
        mul r6, r9, r6           @ R6=W4*ROWr16[0]
170
        ldr r10, =W6             @ R10=W6
171 172 173 174
        ldrsh r4, [r14, #4]      @ R4=ROWr16[2] (a3 not defined yet)
        add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0)

        mul r11, r10, r4         @ R11=W6*ROWr16[2]
175
        ldr r8, =W2              @ R8=W2
176 177 178 179 180 181
        sub r3, r6, r11          @ R3=a0-W6*ROWr16[2] (a2)
        @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
        @@ if (temp != 0) {}
        teq r2, #0
        beq __end_bef_a_evaluation

182
        add r2, r6, r11          @ R2=a0+W6*ROWr16[2] (a1)
183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
        mul r11, r8, r4          @ R11=W2*ROWr16[2]
        sub r4, r6, r11          @ R4=a0-W2*ROWr16[2] (a3)
        add r6, r6, r11          @ R6=a0+W2*ROWr16[2] (a0)


        @@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
        @@     R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]


        @@ a0 += W4*row[4]
        @@ a1 -= W4*row[4]
        @@ a2 -= W4*row[4]
        @@ a3 += W4*row[4]
        ldrsh r11, [r14, #8]     @ R11=ROWr16[4]
198 199
        teq r11, #0              @ if null avoid muls
        it    ne
200 201 202
        mulne r11, r9, r11       @ R11=W4*ROWr16[4]
        @@ R9 is free now
        ldrsh r9, [r14, #12]     @ R9=ROWr16[6]
203
        itttt ne
204 205 206 207 208
        addne r6, r6, r11        @ R6+=W4*ROWr16[4] (a0)
        subne r2, r2, r11        @ R2-=W4*ROWr16[4] (a1)
        subne r3, r3, r11        @ R3-=W4*ROWr16[4] (a2)
        addne r4, r4, r11        @ R4+=W4*ROWr16[4] (a3)
        @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
209 210
        teq r9, #0               @ if null avoid muls
        itttt ne
211 212 213 214 215 216 217 218
        mulne r11, r10, r9       @ R11=W6*ROWr16[6]
        addne r6, r6, r11        @ R6+=W6*ROWr16[6] (a0)
        mulne r10, r8, r9        @ R10=W2*ROWr16[6]
        @@ a0 += W6*row[6];
        @@ a3 -= W6*row[6];
        @@ a1 -= W2*row[6];
        @@ a2 += W2*row[6];
        subne r4, r4, r11        @ R4-=W6*ROWr16[6] (a3)
219
        itt   ne
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
        subne r2, r2, r10        @ R2-=W2*ROWr16[6] (a1)
        addne r3, r3, r10        @ R3+=W2*ROWr16[6] (a2)

__end_a_evaluation:
        @@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
        @@     R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]
        @@ row[0] = (a0 + b0) >> ROW_SHIFT;
        @@ row[1] = (a1 + b1) >> ROW_SHIFT;
        @@ row[2] = (a2 + b2) >> ROW_SHIFT;
        @@ row[3] = (a3 + b3) >> ROW_SHIFT;
        @@ row[4] = (a3 - b3) >> ROW_SHIFT;
        @@ row[5] = (a2 - b2) >> ROW_SHIFT;
        @@ row[6] = (a1 - b1) >> ROW_SHIFT;
        @@ row[7] = (a0 - b0) >> ROW_SHIFT;
        add r8, r6, r0           @ R8=a0+b0
        add r9, r2, r1           @ R9=a1+b1
        @@ put 2 16 bits half-words in a 32bits word
        @@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!)
239
        ldr r10, =MASK_MSHW      @ R10=0xFFFF0000
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
        mvn r11, r10             @ R11= NOT R10= 0x0000FFFF
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
        orr r8, r8, r9
        str r8, [r14, #0]

        add r8, r3, r5           @ R8=a2+b2
        add r9, r4, r7           @ R9=a3+b3
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
        orr r8, r8, r9
        str r8, [r14, #4]

        sub r8, r4, r7           @ R8=a3-b3
        sub r9, r3, r5           @ R9=a2-b2
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
        orr r8, r8, r9
        str r8, [r14, #8]

        sub r8, r2, r1           @ R8=a1-b1
        sub r9, r6, r0           @ R9=a0-b0
        and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
        and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
        orr r8, r8, r9
        str r8, [r14, #12]

        bal __end_row_loop

__almost_empty_row:
        @@ the row was empty, except ROWr16[0], now, management of this special case
        @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
        @@                R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
        @@                R8=0xFFFF (temp), R9-R11 free
        mov r8, #0x10000         @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
        sub r8, r8, #1           @ R8 is now ready.
        and r5, r8, r6, lsl #3   @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
        orr r5, r5, r5, lsl #16  @ R5=R5 | (R5<<16)
        str r5, [r14, #0]        @ R14[0]=ROWr32[0]=R5
        str r5, [r14, #4]        @ R14[4]=ROWr32[1]=R5
        str r5, [r14, #8]        @ R14[8]=ROWr32[2]=R5
        str r5, [r14, #12]       @ R14[12]=ROWr32[3]=R5

__end_row_loop:
        @@ at this point, R0-R11 (free)
        @@     R12=__const_ptr_, R14=&block[n]
        ldr r0, [sp, #0]         @ R0=block
        teq r0, r14              @ compare current &block[8*n] to block, when block is reached, the loop is finished.
        sub r14, r14, #16
        bne __row_loop



293 294 295
        @@ at this point, R0=block, R1-R11 (free)
        @@     R12=__const_ptr_, R14=&block[n]
        add r14, r0, #14        @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
296 297
__col_loop:

298
@@ __b_evaluation2:
299 300 301 302 303 304 305 306 307 308 309
        @@ at this point, R0=block (temp),  R1-R11 (free)
        @@     R12=__const_ptr_, R14=&block[n]
        @@ proceed with b0-b3 first, followed by a0-a3
        @@ MUL16(b0, W1, col[8x1]);
        @@ MUL16(b1, W3, col[8x1]);
        @@ MUL16(b2, W5, col[8x1]);
        @@ MUL16(b3, W7, col[8x1]);
        @@ MAC16(b0, W3, col[8x3]);
        @@ MAC16(b1, -W7, col[8x3]);
        @@ MAC16(b2, -W1, col[8x3]);
        @@ MAC16(b3, -W5, col[8x3]);
310
        ldr r8, =W1              @ R8=W1
311 312
        ldrsh r7, [r14, #16]
        mul r0, r8, r7           @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
313 314
        ldr r9, =W3              @ R9=W3
        ldr r10, =W5             @ R10=W5
315
        mul r1, r9, r7           @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
316
        ldr r11, =W7             @ R11=W7
317 318 319 320
        mul r5, r10, r7          @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
        ldrsh r2, [r14, #48]
        mul r7, r11, r7          @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
        teq r2, #0               @ if 0, then avoid muls
321
        itttt ne
322 323 324 325
        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)
        rsbne r2, r2, #0         @ R2=-ROWr16[3]
        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)
        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)
326
        it    ne
327 328 329 330 331 332 333 334 335 336 337 338 339 340 341
        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)

        @@ at this point, R0=b0,  R1=b1, R2 (free), R3 (free), R4 (free),
        @@     R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
        @@     R12=__const_ptr_, R14=&block[n]
        @@ MAC16(b0, W5, col[5x8]);
        @@ MAC16(b2, W7, col[5x8]);
        @@ MAC16(b3, W3, col[5x8]);
        @@ MAC16(b1, -W1, col[5x8]);
        @@ MAC16(b0, W7, col[7x8]);
        @@ MAC16(b2, W3, col[7x8]);
        @@ MAC16(b3, -W1, col[7x8]);
        @@ MAC16(b1, -W5, col[7x8]);
        ldrsh r3, [r14, #80]     @ R3=COLr16[5x8]
        teq r3, #0               @ if 0 then avoid muls
342
        itttt ne
343 344 345 346 347
        mlane r0, r10, r3, r0    @ R0+=W5*ROWr16[5x8]=b0
        mlane r5, r11, r3, r5    @ R5+=W7*ROWr16[5x8]=b2
        mlane r7, r9, r3, r7     @ R7+=W3*ROWr16[5x8]=b3
        rsbne r3, r3, #0         @ R3=-ROWr16[5x8]
        ldrsh r4, [r14, #112]    @ R4=COLr16[7x8]
348
        it    ne
349 350 351
        mlane r1, r8, r3, r1     @ R7-=W1*ROWr16[5x8]=b1
        @@ R3 is free now
        teq r4, #0               @ if 0 then avoid muls
352
        itttt ne
353 354 355 356
        mlane r0, r11, r4, r0    @ R0+=W7*ROWr16[7x8]=b0
        mlane r5, r9, r4, r5     @ R5+=W3*ROWr16[7x8]=b2
        rsbne r4, r4, #0         @ R4=-ROWr16[7x8]
        mlane r7, r8, r4, r7     @ R7-=W1*ROWr16[7x8]=b3
357
        it    ne
358 359
        mlane r1, r10, r4, r1    @ R1-=W5*ROWr16[7x8]=b1
        @@ R4 is free now
360
@@ __end_b_evaluation2:
361 362 363
        @@ at this point, R0=b0,  R1=b1, R2 (free), R3 (free), R4 (free),
        @@     R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]
364

365
@@ __a_evaluation2:
366 367 368 369 370 371
        @@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
        @@ a1 = a0 + W6 * row[2];
        @@ a2 = a0 - W6 * row[2];
        @@ a3 = a0 - W2 * row[2];
        @@ a0 = a0 + W2 * row[2];
        ldrsh r6, [r14, #0]
372
        ldr r9, =W4              @ R9=W4
373
        mul r6, r9, r6           @ R6=W4*ROWr16[0]
374
        ldr r10, =W6             @ R10=W6
375 376 377
        ldrsh r4, [r14, #32]     @ R4=ROWr16[2] (a3 not defined yet)
        add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
        mul r11, r10, r4         @ R11=W6*ROWr16[2]
378
        ldr r8, =W2              @ R8=W2
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
        add r2, r6, r11          @ R2=a0+W6*ROWr16[2] (a1)
        sub r3, r6, r11          @ R3=a0-W6*ROWr16[2] (a2)
        mul r11, r8, r4          @ R11=W2*ROWr16[2]
        sub r4, r6, r11          @ R4=a0-W2*ROWr16[2] (a3)
        add r6, r6, r11          @ R6=a0+W2*ROWr16[2] (a0)

        @@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
        @@     R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]
        @@ a0 += W4*row[4]
        @@ a1 -= W4*row[4]
        @@ a2 -= W4*row[4]
        @@ a3 += W4*row[4]
        ldrsh r11, [r14, #64]    @ R11=ROWr16[4]
        teq r11, #0              @ if null avoid muls
394
        itttt ne
395 396 397 398 399 400
        mulne r11, r9, r11       @ R11=W4*ROWr16[4]
        @@ R9 is free now
        addne r6, r6, r11        @ R6+=W4*ROWr16[4] (a0)
        subne r2, r2, r11        @ R2-=W4*ROWr16[4] (a1)
        subne r3, r3, r11        @ R3-=W4*ROWr16[4] (a2)
        ldrsh r9, [r14, #96]     @ R9=ROWr16[6]
401
        it    ne
402 403 404
        addne r4, r4, r11        @ R4+=W4*ROWr16[4] (a3)
        @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
        teq r9, #0               @ if null avoid muls
405
        itttt ne
406 407 408 409 410 411 412 413
        mulne r11, r10, r9       @ R11=W6*ROWr16[6]
        addne r6, r6, r11        @ R6+=W6*ROWr16[6] (a0)
        mulne r10, r8, r9        @ R10=W2*ROWr16[6]
        @@ a0 += W6*row[6];
        @@ a3 -= W6*row[6];
        @@ a1 -= W2*row[6];
        @@ a2 += W2*row[6];
        subne r4, r4, r11        @ R4-=W6*ROWr16[6] (a3)
414
        itt   ne
415 416
        subne r2, r2, r10        @ R2-=W2*ROWr16[6] (a1)
        addne r3, r3, r10        @ R3+=W2*ROWr16[6] (a2)
417
@@ __end_a_evaluation2:
418 419 420 421 422 423 424 425 426 427 428
        @@ at this point, R0=b0,  R1=b1, R2=a1, R3=a2, R4=a3,
        @@     R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
        @@     R12=__const_ptr_, R14=&block[n]
        @@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
        @@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
        @@ col[16] = ((a2 + b2) >> COL_SHIFT);
        @@ col[24] = ((a3 + b3) >> COL_SHIFT);
        @@ col[32] = ((a3 - b3) >> COL_SHIFT);
        @@ col[40] = ((a2 - b2) >> COL_SHIFT);
        @@ col[48] = ((a1 - b1) >> COL_SHIFT);
        @@ col[56] = ((a0 - b0) >> COL_SHIFT);
429
        @@@@@ no optimization here @@@@@
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453
        add r8, r6, r0           @ R8=a0+b0
        add r9, r2, r1           @ R9=a1+b1
        mov r8, r8, asr #COL_SHIFT
        mov r9, r9, asr #COL_SHIFT
        strh r8, [r14, #0]
        strh r9, [r14, #16]
        add r8, r3, r5           @ R8=a2+b2
        add r9, r4, r7           @ R9=a3+b3
        mov r8, r8, asr #COL_SHIFT
        mov r9, r9, asr #COL_SHIFT
        strh r8, [r14, #32]
        strh r9, [r14, #48]
        sub r8, r4, r7           @ R8=a3-b3
        sub r9, r3, r5           @ R9=a2-b2
        mov r8, r8, asr #COL_SHIFT
        mov r9, r9, asr #COL_SHIFT
        strh r8, [r14, #64]
        strh r9, [r14, #80]
        sub r8, r2, r1           @ R8=a1-b1
        sub r9, r6, r0           @ R9=a0-b0
        mov r8, r8, asr #COL_SHIFT
        mov r9, r9, asr #COL_SHIFT
        strh r8, [r14, #96]
        strh r9, [r14, #112]
454

455
@@ __end_col_loop:
456 457 458 459 460 461
        @@ at this point, R0-R11 (free)
        @@     R12=__const_ptr_, R14=&block[n]
        ldr r0, [sp, #0]         @ R0=block
        teq r0, r14              @ compare current &block[n] to block, when block is reached, the loop is finished.
        sub r14, r14, #2
        bne __col_loop
462 463 464 465




466
@@ __end_simple_idct_arm:
467 468 469 470 471 472 473 474
        @@ restore registers to previous status!
        add sp, sp, #8 @@ the local variables!
        ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.



@@ kind of sub-function, here not to overload the common case.
__end_bef_a_evaluation:
475
        add r2, r6, r11          @ R2=a0+W6*ROWr16[2] (a1)
476 477 478
        mul r11, r8, r4          @ R11=W2*ROWr16[2]
        sub r4, r6, r11          @ R4=a0-W2*ROWr16[2] (a3)
        add r6, r6, r11          @ R6=a0+W2*ROWr16[2] (a0)
479
        bal __end_a_evaluation
480
endfunc