[Speex-dev] Optimised qmf_synth and iir_mem16

Robin Watts Robin.Watts at wss.co.uk
Sun Dec 2 04:42:43 PST 2007


Hi all,

I've taken preglows ARM versions of qmf_synth and iir_mem16 from
rockboxes speex codec, and tweaked them a bit further for some more
speed.

I attach them here so you can review and take on any changes you
want.

Please let me know if you have questions etc.

Thanks,

Robin
-- 
Robin Watts,             Email: <mailto:Robin.Watts at wss.co.uk>
Warm Silence Software,   WWW:   <http://www.wss.co.uk/>
P.O.Box 28, Woodstock,   Tel:   01608 737172 (or Mobile: 07885 487642)
Oxfordshire, OX20 1XX    Fax:   01608 737172
-------------- next part --------------
@ Copyright (C) 2007 Thom Johansen
@file filters_arm4.S
@brief Various analysis/synthesis filters (ARMv4 version)

@   Redistribution and use in source and binary forms, with or without
@   modification, are permitted provided that the following conditions
@   are met:
@
@   - Redistributions of source code must retain the above copyright
@   notice, this list of conditions and the following disclaimer.
@
@   - Redistributions in binary form must reproduce the above copyright
@   notice, this list of conditions and the following disclaimer in the
@   documentation and/or other materials provided with the distribution.
@
@   - Neither the name of the Xiph.org Foundation nor the names of its
@   contributors may be used to endorse or promote products derived from
@   this software without specific prior written permission.
@
@   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
@   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
@   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
@   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
@   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
@   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
@   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
@   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
@   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
@   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
@   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

    .text

@ void iir_mem16(const spx_word16_t *x,
@                const spx_coef_t   *den,
@                      spx_word16_t *y,
@                      int           N,
@                      int           ord,
@                      spx_mem_t    *mem,
@                      char         *stack)
    .global iir_mem16
iir_mem16:
    stmdb   sp!, { r4-r11, lr }
    ldr     r5, [sp, #36]           @ r0 = x, r1 = den, r2 = y, r3 = N
    ldr     r4, [sp, #40]           @ r4 = mem, r5 = ord
    cmp     r5, #10
    beq     .order_10
    cmp     r5, #8
    beq     .order_8
    ldmia   sp!, { r4-r11, pc }     @ Mon-supported order, return

    @ TODO: try using direct form 1 filtering
.order_8:
    ldmia   r4, { r5-r12 }          @ r5-r12 = mem[0..7]
0:
    ldrsh   r14, [r0], #2
    add     r5, r5, #4096           @ Rounding constant
    str     r0, [sp,#-4]!           @ push r0
    add     r14, r14, r5, asr #13   @ (mem[0] + 4096) >> 13 + x[i]
    mov     r5, #0x7f00
    orr     r5, r5, #0xff           @ r5 = 32767
    cmp     r14, r5
    movgt   r14, r5                 @ Clip positive
    cmn     r14, r5
    rsblt   r14, r5, #0             @ Clip negative
    strh    r14, [r2], #2           @ Write result to y[i]

    ldrsh   r4, [r1]
    ldrsh   r0, [r1, #2]
    rsb     r14,r14,#0              @ r14 = -y[i]
    mla     r5, r4, r14,r6          @ mem[0] = mem[1] - den[0]*y[i]
    ldrsh   r4, [r1, #4]
    mla     r6, r0, r14,r7          @ mem[1] = mem[2] - den[1]*y[i]
    ldrsh   r0, [r1, #6]
    mla     r7, r4, r14,r8          @ mem[2] = mem[3] - den[2]*y[i]
    ldrsh   r4, [r1, #8]
    mla     r8, r0, r14,r9          @ mem[3] = mem[4] - den[3]*y[i]
    ldrsh   r0, [r1, #10]
    mla     r9, r4, r14,r10         @ mem[4] = mem[5] - den[4]*y[i]
    ldrsh   r4, [r1, #12]
    mla     r10, r0, r14,r11        @ mem[5] = mem[6] - den[5]*y[i]
    ldrsh   r0, [r1, #14]
    mla     r11, r4, r14,r12        @ mem[6] = mem[7] - den[6]*y[i]
    subs    r3, r3, #1
    mul     r12, r0, r14            @ mem[7] = -den[7]*y[i]
    ldr     r0, [sp], #4            @ restore r0
    bne     0b
    ldr     r4, [sp, #40]           @ r4 = mem
    stmia   r4, { r5-r12 }          @ Save back mem[]
    ldmia   sp!, { r4-r11, pc }     @ Exit

.order_10:
    ldmia   r4, { r5-r9 }           @ r5-r9 = mem[0..4]
    add     r5, r5, #4096           @ Rounding constant
    ldrsh   r14, [r0], #2
    add     r14, r14, r5, asr #13   @ (mem[0] + 4096) >> 13 + x[i]
    mov     r5, #0x7f00
    orr     r5, r5, #0xff           @ r5 = 32767
    cmp     r14, r5
    movgt   r14, r5                 @ Clip positive
    cmn     r14, r5
    rsblt   r14, r5, #0             @ Clip negative
    strh    r14, [r2], #2           @ Write result to y[i]
    rsb     r14, r14, #0            @ r14 = -y[i]

    ldmia   r1!, { r10-r12 }        @ r10-r12 = den[0..5]
    mov     r5, r10, lsl #16
    mov     r5, r5, asr #16
    mla     r5, r14, r5, r6         @ mem[0] = mem[1] - den[0]*y[i]
    mov     r10, r10, asr #16
    mla     r6, r14, r10, r7        @ mem[1] = mem[2] - den[1]*y[i]
    mov     r10, r11, lsl #16
    mov     r10, r10, asr #16
    mla     r7, r14, r10, r8        @ mem[2] = mem[3] - den[2]*y[i]
    mov     r10, r11, asr #16
    mla     r8, r14, r10, r9        @ mem[3] = mem[4] - den[3]*y[i]
    stmia   r4!, { r5-r8 }          @ Write back mem[0..3], r4 = &mem[4]
    mov     r10, r12, lsl #16
    mov     r11, r10, asr #16
    ldmib   r4, { r6-r10 }          @ r6-r10 = mem[5..9]
    mla     r5, r14, r11, r6        @ mem[4] = mem[5] - den[4]*y[i]
    mov     r12, r12, asr #16
    mla     r6, r14, r12, r7        @ mem[5] = mem[6] - den[5]*y[i]
    ldmia   r1!, { r11-r12 }        @ r11-r12 = den[6..9]
    mov     r7, r11, lsl #16
    mov     r7, r7, asr #16
    mla     r7, r14, r7, r8         @ mem[6] = mem[7] - den[6]*y[i]
    mov     r11, r11, asr #16
    mla     r8, r14, r11, r9        @ mem[7] = mem[8] - den[7]*y[i]
    mov     r11, r12, lsl #16
    mov     r11, r11, asr #16
    mla     r9, r14, r11, r10       @ mem[8] = mem[9] - den[8]*y[i]
    mov     r12, r12, asr #16
    mul     r10, r14, r12           @ mem[9] = -den[9]*y[i]
    stmia   r4!, { r5-r10 }         @ Write back mem[4..9]
    sub     r4, r4, #10*4
    sub     r1, r1, #10*2
    subs    r3, r3, #1
    bne     .order_10
    ldmia   sp!, { r4-r11, pc }     @ Exit


@ void qmf_synth(const spx_word16_t *x1,
@                const spx_word16_t *x2,
@                const spx_word16_t *a,
@                      spx_word16_t *y,
@                      int           N,
@                      int           M,
@                      spx_word32_t *mem1,
@                      spx_word32_t *mem2,
@                      char         *stack)
    .global qmf_synth
qmf_synth:
    stmdb   sp!, { r4-r11, lr }
    add     r7, sp, #36             @ r0 = x1, r1 = x2, r2 = a, r3 = y
    ldmia   r7, { r4-r7 }           @ r4 = N, r5 = M, r6 = mem1, r7 = mem2

    add     r8, r4, r5
    sub     r9, sp, r8              @ r9 = sp - (N + M >> 1) = xx2
    sub     r8, r9, r8              @ r8 = r9 - (N + M >> 1) = xx1
    str     sp, [r8, #-4]           @ Stack old sp
    sub     sp, r8, #4              @ Update sp

    add     r0, r0, r4              @ x1 += N >> 1
    add     r1, r1, r4              @ x2 += N >> 1
    mov     r14, r4                 @ Loop counter is N
0:
    @ Backwards copy x1 and x2 arrays to xx1 and xx2, assume N2 is power of two
    @ N should always be a multiple of four, so this should be OK
    ldmdb   r0!, { r10-r11 }
    subs    r14, r14, #8
    mov     r12, r10, ror #16
    mov     r11, r11, ror #16
    stmia   r8!, { r11-r12 }
    ldmdb   r1!, { r10-r11 }
    mov     r12, r10, ror #16
    mov     r11, r11, ror #16
    stmia   r9!, { r11-r12 }
    bne     0b

    @ Copy alternate members of mem1 and mem2 to last part of xx1 and xx2
    mov     r14, r5                 @ Loop counter is M
    add     r6, r6, #2
    add     r7, r7, #2
    stmdb   sp!, { r6-r7 }          @ Stack &mem1[1], &mem2[1]
0:
    ldrh    r10, [r6], #4
    ldrh    r11, [r6], #4
    ldrh    r12, [r7], #4
    @ 1 cycle stall on Xscale
    orr     r10, r10, r11, lsl #16
    ldrh    r11, [r7], #4
    str     r10, [r8], #4
    subs    r14, r14, #4
    orr     r11, r12, r11, lsl #16
    str     r11, [r9], #4
    bne     0b

    sub     r0, r8, r5              @ r0 = &xx1[N2]
    sub     r1, r9, r5              @ r1 = &xx2[N2]
    str     r4, [sp, #-4]           @ Stack N
    mov     r4, r5
    str     r4, [sp, #-8]           @ Stack M
    @ sp doesn't point to the end of the stack frame from here on, but we're not
    @ calling anything so it shouldn't matter
    @ Main loop, register usage:
    @ r0 = xx1, r1 = xx2, r2 = a, r3 = y, r4 = M, r5 = x10, r6 = x11, r7 = x20
    @ r8 = x21, r9 = [a1, a0], r10 = acc0, r11 = acc1, r12 = acc2, r14 = acc3
0:  @ Outerloop
    mov     r10, #16384             @ Init acccumulators to rounding const
    mov     r11, #16384
    mov     r12, #16384
    mov     r14, #16384

    ldrsh   r5, [r0, #-4]!          @ r5 = x10, r0 = &xx1[N2 - 2]
    ldrsh   r7, [r1, #-4]!          @ r7 = x20, r1 = &xx2[N2 - 2]
    str     r3, [sp, #-12]          @ save r3 onto stack
    sub     r5, r5, r7              @ r5 = x10 - x20
    add     r7, r5, r7, asl #1      @ r7 = x10 + x20
1:  @ Innerloop
    ldrsh   r9, [r2], #2            @ r9 = a0
    ldrsh   r6, [r0, #2]!           @ r6 = x11
    ldrsh   r8, [r1, #2]!           @ r8 = x21
    mla     r12, r9, r5, r12        @ acc2 += a0*(x10 - x20)
    ldrsh   r3, [r2], #2            @ r9 = a1
    sub     r6, r6, r8              @ r6 = x11 - x21
    add     r8, r6, r8, asl #1      @ r8 = x11 + x21
    mla     r10, r9, r6, r10        @ acc0 += a0*(x11 - x21)
    mla     r14, r3, r7, r14        @ acc3 += a1*(x10 + x20)
    mla     r11, r3, r8, r11        @ acc1 += a1*(x11 + x21)

    ldrsh   r9, [r2], #2            @ r9 = a1
    ldrsh   r5, [r0, #2]!           @ r5 = x10
    ldrsh   r7, [r1, #2]!           @ r7 = x20
    mla     r12, r9, r6, r12        @ acc2 += a0*(x11 - x21)
    ldrsh   r3, [r2], #2            @ r9 = a1
    sub     r5, r5, r7              @ r5 = x10 - x20
    add     r7, r5, r7, asl #1      @ r7 = x10 + x20
    mla     r10, r9, r5, r10        @ acc0 += a0*(x10 - x20)
    mla     r14, r3, r8, r14        @ acc3 += a1*(x11 + x21)
    mla     r11, r3, r7, r11        @ acc1 += a1*(x10 + x20)
    subs    r4, r4, #4
    bne     1b

    ldr     r3, [sp, #-12]          @ restore r3 from stack
    ldr     r4, [sp, #-8]           @ r4 = M
    mov     r10, r10, asr #15       @ Shift outputs down
    mov     r11, r11, asr #15
    mov     r12, r12, asr #15
    mov     r14, r14, asr #15
    sub     r2, r2, r4, lsl #1      @ r2 = &a[0]
    sub     r0, r0, r4              @ r0 = &xx1[N2 - 2 - i]
    sub     r1, r1, r4              @ r1 = &xx2[N2 - 2 - i]

    @ TODO: this can be optimized further
    mov     r9, #0x7f00             @ Clip all four outputs
    orr     r9, r9, #0xff           @ r9 = 32767
    cmp     r10, r9
    movgt   r10, r9
    cmn     r10, r9
    rsblt   r10, r9, #0
    cmp     r11, r9
    movgt   r11, r9
    cmn     r11, r9
    rsblt   r11, r9, #0
    cmp     r12, r9
    movgt   r12, r9
    cmn     r12, r9
    rsblt   r12, r9, #0
    cmp     r14, r9
    movgt   r14, r9
    cmn     r14, r9
    rsblt   r14, r9, #0

    ldr     r5, [sp, #-4]           @ Load N
    strh    r10, [r3], #2           @ Write outputs
    strh    r11, [r3], #2
    strh    r12, [r3], #2
    strh    r14, [r3], #2
    subs    r5, r5, #4              @ Are we done?
    strne   r5, [sp, #-4]
    bne     0b

    @ Copy start of xx1 and xx2 back to alternate mem1 and mem2 entries
    @ r0 and r1 are &xx1[0] and &xx2[0] at this point
    ldmia   sp, { r5-r6, sp }       @ Fetch &mem1[1], &mem2[1], restore sp
0:
    ldr     r7, [r0], #4
    ldr     r8, [r1], #4
    subs    r4, r4, #4
    strh    r7, [r5], #4
    strh    r8, [r6], #4
    mov     r7, r7, lsr #16
    mov     r8, r8, lsr #16
    strh    r7, [r5], #4
    strh    r8, [r6], #4
    bne     0b
    ldmia   sp!, { r4-r11, pc }     @ Exit


@ void signal_mul(const spx_sig_t    *x,
@                       spx_sig_t    *y,
@                       spx_word32_t  scale,
@                       int           len)
    .global signal_mul
signal_mul:
    stmdb   sp!, { r4-r8, lr }
0:
    ldmia   r0!, { r5-r8 }                  @ Load four input samples
    smull   r5, r12, r2, r5
    mov     r12, r12, lsl #18               @ Recombine upper and lower parts
    orr     r5, r12, r5, lsr #14
    smull   r6, r12, r2, r6
    mov     r12, r12, lsl #18
    orr     r6, r12, r6, lsr #14
    smull   r7, r12, r2, r7
    mov     r12, r12, lsl #18
    orr     r7, r12, r7, lsr #14
    smull   r8, r12, r2, r8
    mov     r12, r12, lsl #18
    orr     r8, r12, r8, lsr #14
    stmia   r1!, { r5-r8 }                  @ Store four output samples
    subs    r3, r3, #4                      @ Are we done?
    bne     0b

    ldmia   sp!, { r4-r8, pc }              @ Exit



More information about the Speex-dev mailing list