mirror of
https://github.com/Rockbox/rockbox.git
synced 2026-10-10 08:03:04 -04:00
02d5520c5f
The decoder gave the DSP samples with two fractional bits (sample depth 17), far fewer than any other codec. The DSP's filters are only as precise as the samples they are given: with an equalizer band at a low frequency the output was noise, or held at full scale. Scale the last stage of the synthesis filter up by 11 bits and set the sample depth to 28. On ARMv4 the bits are taken from the 64-bit sums of the dewindowing, which had them. Elsewhere the input of that stage is scaled, in its matrixing step; that is not done on ARMv4 because its multiplier takes longer for the larger operands (51.97 against 48.85 MHz below). The Coldfire path uses the C matrixing and has not been run. Seven files decoded under perfsim for the Sansa e200v1 and Clip+: shifted down again, the e200v1 output is within one step of the old output, and the Clip+ output no further from the e200v1's than before. Through a ten band equalizer with bands at 32 and 64 Hz (the filter fix of the DSP included), the noise of atrac3_lp2_132.oma falls from -64 dB to -101 dB relative to the signal. Estimated with perfsim (not measured on a device), atrac3_lp2_132.oma: e200v1 48.81 -> 48.85 MHz, Clip+ 26.91 -> 26.96 MHz. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
226 lines
9.6 KiB
ArmAsm
226 lines
9.6 KiB
ArmAsm
/***************************************************************************
|
|
* __________ __ ___.
|
|
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
|
|
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
|
|
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
|
|
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
|
|
* \/ \/ \/ \/ \/
|
|
* $Id:
|
|
*
|
|
* Copyright (C) 2009 by Andree Buschmann
|
|
*
|
|
* This program is free software; you can redistribute it and/or
|
|
* modify it under the terms of the GNU General Public License
|
|
* as published by the Free Software Foundation; either version 2
|
|
* of the License, or (at your option) any later version.
|
|
*
|
|
* This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
|
|
* KIND, either express or implied.
|
|
*
|
|
****************************************************************************/
|
|
|
|
#include "config.h"
|
|
|
|
.syntax unified
|
|
.section .text, "ax", %progbits
|
|
|
|
/****************************************************************************
|
|
* void atrac3_iqmf_matrixing(int32_t *dest,
|
|
* int32_t *inlo,
|
|
* int32_t *inhi,
|
|
* unsigned int count);
|
|
*
|
|
* Matrixing step within iqmf of atrac3 synthesis. Reference implementation:
|
|
*
|
|
* for(i=0; i<counter; i+=2){
|
|
* dest[2*i+0] = inlo[i ] + inhi[i ];
|
|
* dest[2*i+1] = inlo[i ] - inhi[i ];
|
|
* dest[2*i+2] = inlo[i+1] + inhi[i+1];
|
|
* dest[2*i+3] = inlo[i+1] - inhi[i+1];
|
|
* }
|
|
* Note: r12 is a scratch register and can be used without restorage.
|
|
****************************************************************************/
|
|
.align 2
|
|
.global atrac3_iqmf_matrixing
|
|
.type atrac3_iqmf_matrixing, %function
|
|
|
|
atrac3_iqmf_matrixing:
|
|
/* r0 = dest */
|
|
/* r1 = inlo */
|
|
/* r2 = inhi */
|
|
/* r3 = counter */
|
|
stmfd sp!, {r4-r9, lr} /* save non-scratch registers */
|
|
|
|
.iqmf_matrixing_loop:
|
|
ldmia r1!, { r4, r6, r8, r12} /* load inlo[0...3] */
|
|
ldmia r2!, { r5, r7, r9, lr } /* load inhi[0...3] */
|
|
add r4, r4, r5 /* r4 = inlo[0] + inhi[0] */
|
|
sub r5, r4, r5, asl #1 /* r5 = inlo[0] - inhi[0] */
|
|
add r6, r6, r7 /* r6 = inlo[1] + inhi[1] */
|
|
sub r7, r6, r7, asl #1 /* r7 = inlo[1] - inhi[1] */
|
|
add r8, r8, r9 /* r8 = inlo[2] + inhi[2] */
|
|
sub r9, r8, r9, asl #1 /* r9 = inlo[2] - inhi[2] */
|
|
add r12, r12, lr /* r12 = inlo[3] + inhi[3] */
|
|
sub lr , r12, lr, asl #1 /* lr = inlo[3] - inhi[3] */
|
|
stmia r0!, {r4-r9, r12, lr} /* store results to dest */
|
|
subs r3, r3, #4 /* counter -= 4 */
|
|
bgt .iqmf_matrixing_loop
|
|
|
|
ldmpc regs=r4-r9 /* restore registers */
|
|
|
|
.atrac3_iqmf_matrixing_end:
|
|
.size atrac3_iqmf_matrixing,.atrac3_iqmf_matrixing_end-atrac3_iqmf_matrixing
|
|
|
|
|
|
/* ATRAC3_OUT_SHIFT of atrac3.h */
|
|
#define OUT_SHIFT 11
|
|
|
|
/****************************************************************************
|
|
* void atrac3_iqmf_matrixing_out(int32_t *dest,
|
|
* int32_t *inlo,
|
|
* int32_t *inhi,
|
|
* unsigned int count);
|
|
*
|
|
* The matrixing step of the last iqmf stage: as atrac3_iqmf_matrixing, with
|
|
* the results scaled up by OUT_SHIFT bits, which then is the scale of the
|
|
* decoder's output. Not used on ARMv4, where the multiplications of the
|
|
* dewindowing would take longer for it: see atrac3_iqmf_dewindowing_out.
|
|
****************************************************************************/
|
|
#if ARM_ARCH >= 5
|
|
.align 2
|
|
.global atrac3_iqmf_matrixing_out
|
|
.type atrac3_iqmf_matrixing_out, %function
|
|
|
|
atrac3_iqmf_matrixing_out:
|
|
/* r0 = dest */
|
|
/* r1 = inlo */
|
|
/* r2 = inhi */
|
|
/* r3 = counter */
|
|
stmfd sp!, {r4-r9, lr} /* save non-scratch registers */
|
|
|
|
.iqmf_matrixing_out_loop:
|
|
ldmia r1!, { r4, r6, r8, r12} /* load inlo[0...3] */
|
|
ldmia r2!, { r5, r7, r9, lr } /* load inhi[0...3] */
|
|
add r4, r4, r5 /* r4 = inlo[0] + inhi[0] */
|
|
mov r4, r4, asl #OUT_SHIFT
|
|
sub r5, r4, r5, asl #OUT_SHIFT+1 /* r5 = inlo[0] - inhi[0] */
|
|
add r6, r6, r7 /* r6 = inlo[1] + inhi[1] */
|
|
mov r6, r6, asl #OUT_SHIFT
|
|
sub r7, r6, r7, asl #OUT_SHIFT+1 /* r7 = inlo[1] - inhi[1] */
|
|
add r8, r8, r9 /* r8 = inlo[2] + inhi[2] */
|
|
mov r8, r8, asl #OUT_SHIFT
|
|
sub r9, r8, r9, asl #OUT_SHIFT+1 /* r9 = inlo[2] - inhi[2] */
|
|
add r12, r12, lr /* r12 = inlo[3] + inhi[3] */
|
|
mov r12, r12, asl #OUT_SHIFT
|
|
sub lr , r12, lr, asl #OUT_SHIFT+1 /* lr = inlo[3] - inhi[3] */
|
|
stmia r0!, {r4-r9, r12, lr} /* store results to dest */
|
|
subs r3, r3, #4 /* counter -= 4 */
|
|
bgt .iqmf_matrixing_out_loop
|
|
|
|
ldmpc regs=r4-r9 /* restore registers */
|
|
|
|
.atrac3_iqmf_matrixing_out_end:
|
|
.size atrac3_iqmf_matrixing_out,.atrac3_iqmf_matrixing_out_end-atrac3_iqmf_matrixing_out
|
|
#endif /* ARM_ARCH >= 5 */
|
|
|
|
|
|
/****************************************************************************
|
|
* atrac3_iqmf_dewindowing(int32_t *out,
|
|
* int32_t *in,
|
|
* int32_t *win,
|
|
* unsigned int nIn);
|
|
*
|
|
* Dewindowing step within iqmf of atrac3 synthesis. Reference implementation:
|
|
*
|
|
* for (j = nIn; j != 0; j--) {
|
|
* s1 = fixmul32(in[0], win[0]);
|
|
* s2 = fixmul32(in[1], win[1]);
|
|
* for (i = 2; i < 48; i += 2) {
|
|
* s1 += fixmul32(in[i ], win[i ]);
|
|
* s2 += fixmul32(in[i+1], win[i+1]);
|
|
* }
|
|
* out[0] = s2 << 1;
|
|
* out[1] = s1 << 1;
|
|
* in += 2;
|
|
* out += 2;
|
|
* }
|
|
*
|
|
* The window is symmetric, win[i] = win[47-i], so only its first half is
|
|
* read: win[2k] and win[2k+1] serve in[2k] and in[2k+1] from the front of
|
|
* the 48 inputs and, swapped, in[46-2k] and in[47-2k] from the back. That
|
|
* is 24 coefficient loads for an output pair where there were 48, and
|
|
* they can be loaded four at a time. The sums are kept in 64 bits, so the
|
|
* order of the additions does not change the result.
|
|
*
|
|
* Note: r12 is a scratch register and can be used without restorage.
|
|
****************************************************************************/
|
|
|
|
/* Eight taps: coefficients win[2k..2k+3], two input pairs from the front
|
|
* (r1) and two from the back (r3). s1 is lr/r9 and s2 is r12/r8. The
|
|
* samples are the last operand of the multiplies, as they are the smaller
|
|
* one and that is what the multiplier's early termination goes by. */
|
|
#define DEWIN_8_SAMPLES_ASM(first1, first2) \
|
|
ldmia r2!, {r4-r7}; /* load win[2k..2k+3] */ \
|
|
ldmia r1!, {r10, r11}; /* load in[2k], in[2k+1] */ \
|
|
first1 lr , r9, r4, r10; /* s1 += win[2k ] * in[2k ] */ \
|
|
first2 r12, r8, r5, r11; /* s2 += win[2k+1] * in[2k+1] */ \
|
|
ldmdb r3!, {r10, r11}; /* load in[46-2k], in[47-2k] */ \
|
|
smlal lr , r9, r5, r10; /* s1 += win[2k+1] * in[46-2k] */ \
|
|
smlal r12, r8, r4, r11; /* s2 += win[2k ] * in[47-2k] */ \
|
|
ldmia r1!, {r10, r11}; /* load in[2k+2], in[2k+3] */ \
|
|
smlal lr , r9, r6, r10; /* s1 += win[2k+2] * in[2k+2] */ \
|
|
smlal r12, r8, r7, r11; /* s2 += win[2k+3] * in[2k+3] */ \
|
|
ldmdb r3!, {r10, r11}; /* load in[44-2k], in[45-2k] */ \
|
|
smlal lr , r9, r7, r10; /* s1 += win[2k+3] * in[44-2k] */ \
|
|
smlal r12, r8, r6, r11; /* s2 += win[2k+2] * in[45-2k] */
|
|
|
|
/* atrac3_iqmf_dewindowing is the function above. atrac3_iqmf_dewindowing_out
|
|
* is the same for the last iqmf stage, with the result scaled up by OUT_SHIFT
|
|
* bits, which then is the scale of the decoder's output: the bits are taken
|
|
* from the 64-bit sums, that have them. */
|
|
.macro DEWINDOWING name, shift
|
|
.align 2
|
|
.global \name
|
|
.type \name, %function
|
|
|
|
\name:
|
|
/* r0 = dest */
|
|
/* r1 = input samples */
|
|
/* r2 = window coefficients */
|
|
/* r3 = counter */
|
|
stmfd sp!, {r4-r11, lr} /* save non-scratch registers */
|
|
add r3, r0, r3, lsl #3 /* end of dest: two words a count */
|
|
stmfd sp!, {r3}
|
|
add r3, r1, #192 /* r3 = end of the 48 inputs */
|
|
|
|
1: /* outer loop 0...counter-1 */
|
|
|
|
DEWIN_8_SAMPLES_ASM(smull, smull) /* in[ 0.. 3], in[44..47] */
|
|
DEWIN_8_SAMPLES_ASM(smlal, smlal) /* in[ 4.. 7], in[40..43] */
|
|
DEWIN_8_SAMPLES_ASM(smlal, smlal) /* in[ 8..11], in[36..39] */
|
|
DEWIN_8_SAMPLES_ASM(smlal, smlal) /* in[12..15], in[32..35] */
|
|
DEWIN_8_SAMPLES_ASM(smlal, smlal) /* in[16..19], in[28..31] */
|
|
DEWIN_8_SAMPLES_ASM(smlal, smlal) /* in[20..23], in[24..27] */
|
|
|
|
mov lr , lr , lsr #31-\shift
|
|
orr r9, lr , r9, lsl #1+\shift /* s1 = low>>31 || hi<<1, << shift */
|
|
mov r12, r12, lsr #31-\shift
|
|
orr r8, r12, r8, lsl #1+\shift /* s2 = low>>31 || hi<<1, << shift */
|
|
|
|
stmia r0!, {r8, r9} /* store result out[0]=s2, out[1]=s1 */
|
|
ldr r4, [sp] /* end of dest */
|
|
sub r1, r1, #88 /* back 24 entries, on 2: in += 2 */
|
|
add r3, r3, #104 /* on 24 entries, and 2 more */
|
|
sub r2, r2, #96 /* roll back 24 entries = win[0] */
|
|
|
|
cmp r0, r4
|
|
bne 1b
|
|
|
|
add sp, sp, #4
|
|
ldmpc regs=r4-r11 /* restore registers */
|
|
|
|
.size \name, .-\name
|
|
.endm
|
|
|
|
DEWINDOWING atrac3_iqmf_dewindowing, 0
|
|
DEWINDOWING atrac3_iqmf_dewindowing_out, OUT_SHIFT
|