arm: Add VFP-accelerated version of qmf_32_subbands

               Before           After
               Mean    StdDev   Mean    StdDev  Change
This function   1323.0  98.0      746.2  60.6   +77.3%
Overall        15400.0 336.4    14147.5 288.4    +8.9%

Signed-off-by: Martin Storsjö <martin@martin.st>

1 files changed, 273 insertions, 0 deletions

diff --git a/libavcodec/arm/dcadsp_vfp.S b/libavcodec/arm/dcadsp_vfp.S
index 57e16196f7..6039e87dfc 100644
--- a/libavcodec/arm/dcadsp_vfp.S
+++ b/libavcodec/arm/dcadsp_vfp.S
@@ -218,3 +218,276 @@ endfunc
         .unreq  POST1
         .unreq  POST2
         .unreq  POST3
+
+
+IN      .req    a1
+SBACT   .req    a2
+OLDFPSCR .req   a3
+IMDCT   .req    a4
+WINDOW  .req    v1
+OUT     .req    v2
+BUF     .req    v3
+SCALEINT .req   v4 @ only used in softfp case
+COUNT   .req    v5
+
+SCALE   .req    s0
+
+/* Stack layout differs in softfp and hardfp cases:
+ *
+ * hardfp
+ *      fp -> 6 arg words saved by caller
+ *            a3,a4,v1-v3,v5,fp,lr on entry (a3 just to pad to 8 bytes)
+ *            s16-s23 on entry
+ *            align 16
+ *     buf -> 8*32*4 bytes buffer
+ *            s0 on entry
+ *      sp -> 3 arg words for callee
+ *
+ * softfp
+ *      fp -> 7 arg words saved by caller
+ *            a4,v1-v5,fp,lr on entry
+ *            s16-s23 on entry
+ *            align 16
+ *     buf -> 8*32*4 bytes buffer
+ *      sp -> 4 arg words for callee
+ */
+
+/* void ff_dca_qmf_32_subbands_vfp(float samples_in[32][8], int sb_act,
+ *                                 SynthFilterContext *synth, FFTContext *imdct,
+ *                                 float (*synth_buf_ptr)[512],
+ *                                 int *synth_buf_offset, float (*synth_buf2)[32],
+ *                                 const float (*window)[512], float *samples_out,
+ *                                 float (*raXin)[32], float scale);
+ */
+function ff_dca_qmf_32_subbands_vfp, export=1
+VFP     push    {a3-a4,v1-v3,v5,fp,lr}
+NOVFP   push    {a4,v1-v5,fp,lr}
+        add     fp, sp, #8*4
+        vpush   {s16-s23}
+        @ The buffer pointed at by raXin isn't big enough for us to do a
+        @ complete matrix transposition as we want to, so allocate an
+        @ alternative buffer from the stack. Align to 4 words for speed.
+        sub     BUF, sp, #8*32*4
+        bic     BUF, BUF, #15
+        mov     sp, BUF
+        ldr     lr, =0x03330000     @ RunFast mode, short vectors of length 4, stride 2
+        fmrx    OLDFPSCR, FPSCR
+        fmxr    FPSCR, lr
+        @ COUNT is used to count down 2 things at once:
+        @ bits 0-4 are the number of word pairs remaining in the output row
+        @ bits 5-31 are the number of words to copy (with possible negation)
+        @   from the source matrix before we start zeroing the remainder
+        mov     COUNT, #(-4 << 5) + 16
+        adds    COUNT, COUNT, SBACT, lsl #5
+        bmi     2f
+1:
+        vldr    s8,  [IN, #(0*8+0)*4]
+        vldr    s10, [IN, #(0*8+1)*4]
+        vldr    s12, [IN, #(0*8+2)*4]
+        vldr    s14, [IN, #(0*8+3)*4]
+        vldr    s16, [IN, #(0*8+4)*4]
+        vldr    s18, [IN, #(0*8+5)*4]
+        vldr    s20, [IN, #(0*8+6)*4]
+        vldr    s22, [IN, #(0*8+7)*4]
+        vneg.f  s8, s8
+        vldr    s9,  [IN, #(1*8+0)*4]
+        vldr    s11, [IN, #(1*8+1)*4]
+        vldr    s13, [IN, #(1*8+2)*4]
+        vldr    s15, [IN, #(1*8+3)*4]
+        vneg.f  s16, s16
+        vldr    s17, [IN, #(1*8+4)*4]
+        vldr    s19, [IN, #(1*8+5)*4]
+        vldr    s21, [IN, #(1*8+6)*4]
+        vldr    s23, [IN, #(1*8+7)*4]
+        vstr    d4,  [BUF, #(0*32+0)*4]
+        vstr    d5,  [BUF, #(1*32+0)*4]
+        vstr    d6,  [BUF, #(2*32+0)*4]
+        vstr    d7,  [BUF, #(3*32+0)*4]
+        vstr    d8,  [BUF, #(4*32+0)*4]
+        vstr    d9,  [BUF, #(5*32+0)*4]
+        vstr    d10, [BUF, #(6*32+0)*4]
+        vstr    d11, [BUF, #(7*32+0)*4]
+        vldr    s9,  [IN, #(3*8+0)*4]
+        vldr    s11, [IN, #(3*8+1)*4]
+        vldr    s13, [IN, #(3*8+2)*4]
+        vldr    s15, [IN, #(3*8+3)*4]
+        vldr    s17, [IN, #(3*8+4)*4]
+        vldr    s19, [IN, #(3*8+5)*4]
+        vldr    s21, [IN, #(3*8+6)*4]
+        vldr    s23, [IN, #(3*8+7)*4]
+        vneg.f  s9, s9
+        vldr    s8,  [IN, #(2*8+0)*4]
+        vldr    s10, [IN, #(2*8+1)*4]
+        vldr    s12, [IN, #(2*8+2)*4]
+        vldr    s14, [IN, #(2*8+3)*4]
+        vneg.f  s17, s17
+        vldr    s16, [IN, #(2*8+4)*4]
+        vldr    s18, [IN, #(2*8+5)*4]
+        vldr    s20, [IN, #(2*8+6)*4]
+        vldr    s22, [IN, #(2*8+7)*4]
+        vstr    d4,  [BUF, #(0*32+2)*4]
+        vstr    d5,  [BUF, #(1*32+2)*4]
+        vstr    d6,  [BUF, #(2*32+2)*4]
+        vstr    d7,  [BUF, #(3*32+2)*4]
+        vstr    d8,  [BUF, #(4*32+2)*4]
+        vstr    d9,  [BUF, #(5*32+2)*4]
+        vstr    d10, [BUF, #(6*32+2)*4]
+        vstr    d11, [BUF, #(7*32+2)*4]
+        add     IN, IN, #4*8*4
+        add     BUF, BUF, #4*4
+        subs    COUNT, COUNT, #(4 << 5) + 2
+        bpl     1b
+2:      @ Now deal with trailing < 4 samples
+        adds    COUNT, COUNT, #3 << 5
+        bmi     4f  @ sb_act was a multiple of 4
+        bics    lr, COUNT, #0x1F
+        bne     3f
+        @ sb_act was n*4+1
+        vldr    s8,  [IN, #(0*8+0)*4]
+        vldr    s10, [IN, #(0*8+1)*4]
+        vldr    s12, [IN, #(0*8+2)*4]
+        vldr    s14, [IN, #(0*8+3)*4]
+        vldr    s16, [IN, #(0*8+4)*4]
+        vldr    s18, [IN, #(0*8+5)*4]
+        vldr    s20, [IN, #(0*8+6)*4]
+        vldr    s22, [IN, #(0*8+7)*4]
+        vneg.f  s8, s8
+        vldr    s9,  zero
+        vldr    s11, zero
+        vldr    s13, zero
+        vldr    s15, zero
+        vneg.f  s16, s16
+        vldr    s17, zero
+        vldr    s19, zero
+        vldr    s21, zero
+        vldr    s23, zero
+        vstr    d4,  [BUF, #(0*32+0)*4]
+        vstr    d5,  [BUF, #(1*32+0)*4]
+        vstr    d6,  [BUF, #(2*32+0)*4]
+        vstr    d7,  [BUF, #(3*32+0)*4]
+        vstr    d8,  [BUF, #(4*32+0)*4]
+        vstr    d9,  [BUF, #(5*32+0)*4]
+        vstr    d10, [BUF, #(6*32+0)*4]
+        vstr    d11, [BUF, #(7*32+0)*4]
+        add     BUF, BUF, #2*4
+        sub     COUNT, COUNT, #1
+        b       4f
+3:      @ sb_act was n*4+2 or n*4+3, so do the first 2
+        vldr    s8,  [IN, #(0*8+0)*4]
+        vldr    s10, [IN, #(0*8+1)*4]
+        vldr    s12, [IN, #(0*8+2)*4]
+        vldr    s14, [IN, #(0*8+3)*4]
+        vldr    s16, [IN, #(0*8+4)*4]
+        vldr    s18, [IN, #(0*8+5)*4]
+        vldr    s20, [IN, #(0*8+6)*4]
+        vldr    s22, [IN, #(0*8+7)*4]
+        vneg.f  s8, s8
+        vldr    s9,  [IN, #(1*8+0)*4]
+        vldr    s11, [IN, #(1*8+1)*4]
+        vldr    s13, [IN, #(1*8+2)*4]
+        vldr    s15, [IN, #(1*8+3)*4]
+        vneg.f  s16, s16
+        vldr    s17, [IN, #(1*8+4)*4]
+        vldr    s19, [IN, #(1*8+5)*4]
+        vldr    s21, [IN, #(1*8+6)*4]
+        vldr    s23, [IN, #(1*8+7)*4]
+        vstr    d4,  [BUF, #(0*32+0)*4]
+        vstr    d5,  [BUF, #(1*32+0)*4]
+        vstr    d6,  [BUF, #(2*32+0)*4]
+        vstr    d7,  [BUF, #(3*32+0)*4]
+        vstr    d8,  [BUF, #(4*32+0)*4]
+        vstr    d9,  [BUF, #(5*32+0)*4]
+        vstr    d10, [BUF, #(6*32+0)*4]
+        vstr    d11, [BUF, #(7*32+0)*4]
+        add     BUF, BUF, #2*4
+        sub     COUNT, COUNT, #(2 << 5) + 1
+        bics    lr, COUNT, #0x1F
+        bne     4f
+        @ sb_act was n*4+3
+        vldr    s8,  [IN, #(2*8+0)*4]
+        vldr    s10, [IN, #(2*8+1)*4]
+        vldr    s12, [IN, #(2*8+2)*4]
+        vldr    s14, [IN, #(2*8+3)*4]
+        vldr    s16, [IN, #(2*8+4)*4]
+        vldr    s18, [IN, #(2*8+5)*4]
+        vldr    s20, [IN, #(2*8+6)*4]
+        vldr    s22, [IN, #(2*8+7)*4]
+        vldr    s9,  zero
+        vldr    s11, zero
+        vldr    s13, zero
+        vldr    s15, zero
+        vldr    s17, zero
+        vldr    s19, zero
+        vldr    s21, zero
+        vldr    s23, zero
+        vstr    d4,  [BUF, #(0*32+0)*4]
+        vstr    d5,  [BUF, #(1*32+0)*4]
+        vstr    d6,  [BUF, #(2*32+0)*4]
+        vstr    d7,  [BUF, #(3*32+0)*4]
+        vstr    d8,  [BUF, #(4*32+0)*4]
+        vstr    d9,  [BUF, #(5*32+0)*4]
+        vstr    d10, [BUF, #(6*32+0)*4]
+        vstr    d11, [BUF, #(7*32+0)*4]
+        add     BUF, BUF, #2*4
+        sub     COUNT, COUNT, #1
+4:      @ Now fill the remainder with 0
+        vldr    s8, zero
+        vldr    s9, zero
+        ands    COUNT, COUNT, #0x1F
+        beq     6f
+5:      vstr    d4, [BUF, #(0*32+0)*4]
+        vstr    d4, [BUF, #(1*32+0)*4]
+        vstr    d4, [BUF, #(2*32+0)*4]
+        vstr    d4, [BUF, #(3*32+0)*4]
+        vstr    d4, [BUF, #(4*32+0)*4]
+        vstr    d4, [BUF, #(5*32+0)*4]
+        vstr    d4, [BUF, #(6*32+0)*4]
+        vstr    d4, [BUF, #(7*32+0)*4]
+        add     BUF, BUF, #2*4
+        subs    COUNT, COUNT, #1
+        bne     5b
+6:
+        fmxr    FPSCR, OLDFPSCR
+        ldr     WINDOW, [fp, #3*4]
+        ldr     OUT, [fp, #4*4]
+        sub     BUF, BUF, #32*4
+NOVFP   ldr     SCALEINT, [fp, #6*4]
+        mov     COUNT, #8
+VFP     vpush   {SCALE}
+VFP     sub     sp, sp, #3*4
+NOVFP   sub     sp, sp, #4*4
+7:
+VFP     ldr     a1, [fp, #-7*4]     @ imdct
+NOVFP   ldr     a1, [fp, #-8*4]
+        ldmia   fp, {a2-a4}
+VFP     stmia   sp, {WINDOW, OUT, BUF}
+NOVFP   stmia   sp, {WINDOW, OUT, BUF, SCALEINT}
+VFP     vldr    SCALE, [sp, #3*4]
+        bl      ff_synth_filter_float_vfp
+        add     OUT, OUT, #32*4
+        add     BUF, BUF, #32*4
+        subs    COUNT, COUNT, #1
+        bne     7b
+
+A       sub     sp, fp, #(8+8)*4
+T       sub     fp, fp, #(8+8)*4
+T       mov     sp, fp
+        vpop    {s16-s23}
+VFP     pop     {a3-a4,v1-v3,v5,fp,pc}
+NOVFP   pop     {a4,v1-v5,fp,pc}
+endfunc
+
+        .unreq  IN
+        .unreq  SBACT
+        .unreq  OLDFPSCR
+        .unreq  IMDCT
+        .unreq  WINDOW
+        .unreq  OUT
+        .unreq  BUF
+        .unreq  SCALEINT
+        .unreq  COUNT
+
+        .unreq  SCALE
+
+        .align 2
+zero:   .word   0