1 /***************************************************************************
3 * Open \______ \ ____ ____ | | _\_ |__ _______ ___
4 * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
5 * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
6 * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
10 * Copyright (C) 2006 by David Bryant
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version 2
15 * of the License, or (at your option) any later version.
17 * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
18 * KIND, either express or implied.
20 ****************************************************************************/
22 /* This is an assembly optimized version of the following WavPack function:
24 * void decorr_stereo_pass_cont_arm (struct decorr_pass *dpp,
25 * long *buffer, long sample_count);
27 * It performs a single pass of stereo decorrelation on the provided buffer.
28 * Note that this version of the function requires that the 8 previous stereo
29 * samples are visible and correct. In other words, it ignores the "samples_*"
30 * fields in the decorr_pass structure and gets the history data directly
31 * from the buffer. It does, however, return the appropriate history samples
32 * to the decorr_pass structure before returning.
34 * This is written to work on a ARM7TDMI processor. This version only uses the
35 * 32-bit multiply-accumulate instruction and so will overflow with 24-bit
43 .global decorr_stereo_pass_cont_arm
48 * r0 = struct decorr_pass *dpp
50 * r2 = long sample_count
53 decorr_stereo_pass_cont_arm:
55 stmfd sp!, {r4 - r8, r10, r11, lr}
57 mov r11, #512 @ r11 = 512 for rounding
58 ldrsh r6, [r0, #2] @ r6 = dpp->delta
59 ldrsh r4, [r0, #4] @ r4 = dpp->weight_A
60 ldrsh r0, [r0, #6] @ r0 = dpp->weight_B
61 cmp r2, #0 @ exit if no samples to process
64 add r7, r1, r2, asl #3 @ r7 = buffer ending position
65 ldrsh r2, [r5, #0] @ r2 = dpp->term
69 ldr lr, [r1, #-16] @ load 2 sample history from buffer
70 ldr r10, [r1, #-12] @ for terms 2, 17, and 18
79 b term_default_loop @ else handle default (1-8, except 2)
82 mov r10, #1024 @ r10 = -1024 for weight clipping
83 rsb r10, r10, #0 @ (only used for negative terms)
93 ******************************************************************************
94 * Loop to handle term = 17 condition
96 * r0 = dpp->weight_B r8 = previous left sample
98 * r2 = current sample r10 = second previous left sample
99 * r3 = previous right sample r11 = 512 (for rounding)
100 * r4 = dpp->weight_A ip = current decorrelation value
102 * r6 = dpp->delta lr = second previous right sample
104 *******************************************************************************
108 rsbs ip, lr, r8, asl #1 @ decorr value = (2 * prev) - 2nd prev
109 mov lr, r8 @ previous becomes 2nd previous
110 ldr r2, [r1], #4 @ get sample & update pointer
111 mla r8, ip, r4, r11 @ mult decorr value by weight, round,
112 add r8, r2, r8, asr #10 @ shift, and add to new sample
113 strne r8, [r1, #-4] @ if change possible, store sample back
116 teq ip, r2 @ update weight based on signs
120 .L325: rsbs ip, r10, r3, asl #1 @ do same thing for right channel
124 add r3, r2, r3, asr #10
132 .L329: cmp r7, r1 @ loop back if more samples to do
134 b store_1718 @ common exit for terms 17 & 18
137 ******************************************************************************
138 * Loop to handle term = 18 condition
140 * r0 = dpp->weight_B r8 = previous left sample
142 * r2 = current sample r10 = second previous left sample
143 * r3 = previous right sample r11 = 512 (for rounding)
144 * r4 = dpp->weight_A ip = decorrelation value
146 * r6 = dpp->delta lr = second previous right sample
148 *******************************************************************************
152 sub ip, r8, lr @ decorr value =
153 mov lr, r8 @ ((3 * prev) - 2nd prev) >> 1
154 adds ip, r8, ip, asr #1
155 ldr r2, [r1], #4 @ get sample & update pointer
156 mla r8, ip, r4, r11 @ mult decorr value by weight, round,
157 add r8, r2, r8, asr #10 @ shift, and add to new sample
158 strne r8, [r1, #-4] @ if change possible, store sample back
161 teq ip, r2 @ update weight based on signs
165 .L337: sub ip, r3, r10 @ do same thing for right channel
167 adds ip, r3, ip, asr #1
170 add r3, r2, r3, asr #10
178 .L341: cmp r7, r1 @ loop back if more samples to do
181 /* common exit for terms 17 & 18 */
184 str r3, [r5, #40] @ store sample history into struct
188 b common_exit @ and return
191 ******************************************************************************
192 * Loop to handle term = 2 condition
193 * (note that this case can be handled by the default term handler (1-8), but
194 * this special case is faster because it doesn't have to read memory twice)
196 * r0 = dpp->weight_B r8 = previous left sample
198 * r2 = current sample r10 = second previous left sample
199 * r3 = previous right sample r11 = 512 (for rounding)
200 * r4 = dpp->weight_A ip = decorrelation value
202 * r6 = dpp->delta lr = second previous right sample
204 *******************************************************************************
208 movs ip, lr @ get decorrelation value & test
209 mov lr, r8 @ previous becomes 2nd previous
210 ldr r2, [r1], #4 @ get sample & update pointer
211 mla r8, ip, r4, r11 @ mult decorr value by weight, round,
212 add r8, r2, r8, asr #10 @ shift, and add to new sample
213 strne r8, [r1, #-4] @ if change possible, store sample back
216 teq ip, r2 @ update weight based on signs
220 .L225: movs ip, r10 @ do same thing for right channel
224 add r3, r2, r3, asr #10
232 .L229: cmp r7, r1 @ loop back if more samples to do
234 b default_term_exit @ this exit updates all dpp->samples
237 ******************************************************************************
238 * Loop to handle default term condition
240 * r0 = dpp->weight_B r8 = result accumulator
242 * r2 = dpp->term r10 =
243 * r3 = decorrelation value r11 = 512 (for rounding)
244 * r4 = dpp->weight_A ip = current sample
246 * r6 = dpp->delta lr =
248 *******************************************************************************
252 ldr ip, [r1] @ get original sample
253 ldr r3, [r1, -r2, asl #3] @ get decorrelation value based on term
254 mla r8, r3, r4, r11 @ mult decorr value by weight, round,
255 add r8, ip, r8, asr #10 @ shift and add to new sample
256 str r8, [r1], #4 @ store update sample
260 teq ip, r3 @ update weight based on signs
264 .L350: ldr ip, [r1] @ do the same thing for right channel
265 ldr r3, [r1, -r2, asl #3]
267 add r8, ip, r8, asr #10
276 .L354: cmp r7, r1 @ loop back if more samples to do
277 bhi term_default_loop
280 * This exit is used by terms 1-8 to store the previous 8 samples into the decorr
281 * structure (even if they are not all used for the given term)
289 .L358: and r3, ip, #7
290 add r3, r5, r3, asl #2
302 ******************************************************************************
303 * Loop to handle term = -1 condition
305 * r0 = dpp->weight_B r8 =
307 * r2 = intermediate result r10 = -1024 (for clipping)
308 * r3 = previous right sample r11 = 512 (for rounding)
309 * r4 = dpp->weight_A ip = current sample
311 * r6 = dpp->delta lr = updated left sample
313 *******************************************************************************
320 ldr ip, [r1] @ for left channel the decorrelation value
321 mla r2, r3, r4, r11 @ is the previous right sample (in r3)
322 add lr, ip, r2, asr #10
327 teq ip, r3 @ update weight based on signs
335 .L361: ldr r2, [r1, #-4] @ for right channel the decorrelation value
336 mla r3, lr, r0, r11 @ is the just updated right sample (in lr)
337 add r3, r2, r3, asr #10
345 cmp r0, #1024 @ then clip weight to +/-1024
350 .L369: cmp r7, r1 @ loop back if more samples to do
351 bhi term_minus_1_loop
353 str r3, [r5, #8] @ else store right sample and exit
357 ******************************************************************************
358 * Loop to handle term = -2 condition
359 * (note that the channels are processed in the reverse order here)
361 * r0 = dpp->weight_B r8 =
363 * r2 = intermediate result r10 = -1024 (for clipping)
364 * r3 = previous left sample r11 = 512 (for rounding)
365 * r4 = dpp->weight_A ip = current sample
367 * r6 = dpp->delta lr = updated right sample
369 *******************************************************************************
376 ldr ip, [r1, #4] @ for right channel the decorrelation value
377 mla r2, r3, r0, r11 @ is the previous left sample (in r3)
378 add lr, ip, r2, asr #10
383 teq ip, r3 @ update weight based on signs
386 cmp r0, #1024 @ then clip weight to +/-1024
391 .L380: ldr r2, [r1, #0] @ for left channel the decorrelation value
392 mla r3, lr, r4, r11 @ is the just updated left sample (in lr)
393 add r3, r2, r3, asr #10
406 .L388: cmp r7, r1 @ loop back if more samples to do
407 bhi term_minus_2_loop
409 str r3, [r5, #40] @ else store left channel and exit
413 ******************************************************************************
414 * Loop to handle term = -3 condition
416 * r0 = dpp->weight_B r8 = previous left sample
418 * r2 = current left sample r10 = -1024 (for clipping)
419 * r3 = previous right sample r11 = 512 (for rounding)
420 * r4 = dpp->weight_A ip = intermediate result
422 * r6 = dpp->delta lr =
424 *******************************************************************************
428 ldr r3, [r1, #-4] @ load previous samples
434 add r2, ip, r2, asr #10
439 teq ip, r3 @ update weight based on signs
442 cmp r4, #1024 @ then clip weight to +/-1024
447 .L399: movs ip, r8 @ ip = previous left we use now
448 mov r8, r2 @ r8 = current left we use next time
451 add r3, r2, r3, asr #10
463 .L407: cmp r7, r1 @ loop back if more samples to do
464 bhi term_minus_3_loop
466 str r3, [r5, #8] @ else store previous samples & exit
470 * Before finally exiting we must store weights back for next time
476 ldmpc regs="r4-r8, r10-r11"