arm: Add VFP-accelerated version of dca_lfe_fir

               Before           After
               Mean    StdDev   Mean    StdDev  Change
This function    868.2  33.5      436.0  27.0   +99.1%
Overall        15973.0 223.2    15577.5  83.2    +2.5%

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

1 files changed, 220 insertions, 0 deletions

diff --git a/libavcodec/arm/dcadsp_vfp.S b/libavcodec/arm/dcadsp_vfp.S
new file mode 100644
index 0000000000..57e16196f7
--- /dev/null
+++ b/libavcodec/arm/dcadsp_vfp.S
@@ -0,0 +1,220 @@
+/*
+ * Copyright (c) 2013 RISC OS Open Ltd
+ * Author: Ben Avison <bavison@riscosopen.org>
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/arm/asm.S"
+
+POUT          .req    a1
+PIN           .req    a2
+PCOEF         .req    a3
+DECIFACTOR    .req    a4
+OLDFPSCR      .req    a4
+COUNTER       .req    ip
+
+SCALE32       .req    s28  @ use vector of 4 in place of 9th scalar when decifactor=32 / JMAX=8
+SCALE64       .req    s0   @ spare register in scalar bank when decifactor=64 / JMAX=4
+IN0           .req    s4
+IN1           .req    s5
+IN2           .req    s6
+IN3           .req    s7
+IN4           .req    s0
+IN5           .req    s1
+IN6           .req    s2
+IN7           .req    s3
+COEF0         .req    s8   @ coefficient elements
+COEF1         .req    s9
+COEF2         .req    s10
+COEF3         .req    s11
+COEF4         .req    s12
+COEF5         .req    s13
+COEF6         .req    s14
+COEF7         .req    s15
+ACCUM0        .req    s16  @ double-buffered multiply-accumulate results
+ACCUM4        .req    s20
+POST0         .req    s24  @ do long-latency post-multiply in this vector in parallel
+POST1         .req    s25
+POST2         .req    s26
+POST3         .req    s27
+
+
+.macro inner_loop  decifactor, dir, tail, head
+ .ifc "\dir","up"
+  .set X, 0
+  .set Y, 4
+ .else
+  .set X, 4*JMAX*4 - 4
+  .set Y, -4
+ .endif
+ .ifnc "\head",""
+        vldr    COEF0, [PCOEF, #X + (0*JMAX + 0) * Y]
+        vldr    COEF1, [PCOEF, #X + (1*JMAX + 0) * Y]
+        vldr    COEF2, [PCOEF, #X + (2*JMAX + 0) * Y]
+        vldr    COEF3, [PCOEF, #X + (3*JMAX + 0) * Y]
+ .endif
+ .ifnc "\tail",""
+        vadd.f  POST0, ACCUM0, ACCUM4   @ vector operation
+ .endif
+ .ifnc "\head",""
+        vmul.f  ACCUM0, COEF0, IN0      @ vector = vector * scalar
+        vldr    COEF4, [PCOEF, #X + (0*JMAX + 1) * Y]
+        vldr    COEF5, [PCOEF, #X + (1*JMAX + 1) * Y]
+        vldr    COEF6, [PCOEF, #X + (2*JMAX + 1) * Y]
+ .endif
+ .ifnc "\tail",""
+        vmul.f  POST0, POST0, SCALE\decifactor  @ vector operation (SCALE may be scalar)
+ .endif
+ .ifnc "\head",""
+        vldr    COEF7, [PCOEF, #X + (3*JMAX + 1) * Y]
+   .ifc "\tail",""
+        vmul.f  ACCUM4, COEF4, IN1      @ vector operation
+   .endif
+        vldr    COEF0, [PCOEF, #X + (0*JMAX + 2) * Y]
+        vldr    COEF1, [PCOEF, #X + (1*JMAX + 2) * Y]
+   .ifnc "\tail",""
+        vmul.f  ACCUM4, COEF4, IN1      @ vector operation
+   .endif
+        vldr    COEF2, [PCOEF, #X + (2*JMAX + 2) * Y]
+        vldr    COEF3, [PCOEF, #X + (3*JMAX + 2) * Y]
+ .endif
+ .ifnc "\tail",""
+        vstmia  POUT!, {POST0-POST3}
+ .endif
+ .ifnc "\head",""
+        vmla.f  ACCUM0, COEF0, IN2      @ vector = vector * scalar
+        vldr    COEF4, [PCOEF, #X + (0*JMAX + 3) * Y]
+        vldr    COEF5, [PCOEF, #X + (1*JMAX + 3) * Y]
+        vldr    COEF6, [PCOEF, #X + (2*JMAX + 3) * Y]
+        vldr    COEF7, [PCOEF, #X + (3*JMAX + 3) * Y]
+        vmla.f  ACCUM4, COEF4, IN3      @ vector = vector * scalar
+  .if \decifactor == 32
+        vldr    COEF0, [PCOEF, #X + (0*JMAX + 4) * Y]
+        vldr    COEF1, [PCOEF, #X + (1*JMAX + 4) * Y]
+        vldr    COEF2, [PCOEF, #X + (2*JMAX + 4) * Y]
+        vldr    COEF3, [PCOEF, #X + (3*JMAX + 4) * Y]
+        vmla.f  ACCUM0, COEF0, IN4      @ vector = vector * scalar
+        vldr    COEF4, [PCOEF, #X + (0*JMAX + 5) * Y]
+        vldr    COEF5, [PCOEF, #X + (1*JMAX + 5) * Y]
+        vldr    COEF6, [PCOEF, #X + (2*JMAX + 5) * Y]
+        vldr    COEF7, [PCOEF, #X + (3*JMAX + 5) * Y]
+        vmla.f  ACCUM4, COEF4, IN5      @ vector = vector * scalar
+        vldr    COEF0, [PCOEF, #X + (0*JMAX + 6) * Y]
+        vldr    COEF1, [PCOEF, #X + (1*JMAX + 6) * Y]
+        vldr    COEF2, [PCOEF, #X + (2*JMAX + 6) * Y]
+        vldr    COEF3, [PCOEF, #X + (3*JMAX + 6) * Y]
+        vmla.f  ACCUM0, COEF0, IN6      @ vector = vector * scalar
+        vldr    COEF4, [PCOEF, #X + (0*JMAX + 7) * Y]
+        vldr    COEF5, [PCOEF, #X + (1*JMAX + 7) * Y]
+        vldr    COEF6, [PCOEF, #X + (2*JMAX + 7) * Y]
+        vldr    COEF7, [PCOEF, #X + (3*JMAX + 7) * Y]
+        vmla.f  ACCUM4, COEF4, IN7      @ vector = vector * scalar
+  .endif
+ .endif
+.endm
+
+.macro dca_lfe_fir  decifactor
+ .if \decifactor == 32
+  .set JMAX, 8
+        vpush   {s16-s31}
+        vmov    SCALE32, s0             @ duplicate scalar across vector
+        vldr    IN4, [PIN, #-4*4]
+        vldr    IN5, [PIN, #-5*4]
+        vldr    IN6, [PIN, #-6*4]
+        vldr    IN7, [PIN, #-7*4]
+ .else
+  .set JMAX, 4
+        vpush   {s16-s27}
+ .endif
+
+        mov     COUNTER, #\decifactor/4 - 1
+        inner_loop  \decifactor, up,, head
+1:      add     PCOEF, PCOEF, #4*JMAX*4
+        subs    COUNTER, COUNTER, #1
+        inner_loop  \decifactor, up, tail, head
+        bne     1b
+        inner_loop  \decifactor, up, tail
+
+        mov     COUNTER, #\decifactor/4 - 1
+        inner_loop  \decifactor, down,, head
+1:      sub     PCOEF, PCOEF, #4*JMAX*4
+        subs    COUNTER, COUNTER, #1
+        inner_loop  \decifactor, down, tail, head
+        bne     1b
+        inner_loop  \decifactor, down, tail
+
+ .if \decifactor == 32
+        vpop    {s16-s31}
+ .else
+        vpop    {s16-s27}
+ .endif
+        fmxr    FPSCR, OLDFPSCR
+        bx      lr
+.endm
+
+
+/* void ff_dca_lfe_fir_vfp(float *out, const float *in, const float *coefs,
+ *                         int decifactor, float scale)
+ */
+function ff_dca_lfe_fir_vfp, export=1
+        teq     DECIFACTOR, #32
+        fmrx    OLDFPSCR, FPSCR
+        ldr     ip, =0x03030000         @ RunFast mode, short vectors of length 4, stride 1
+        fmxr    FPSCR, ip
+NOVFP   vldr    s0, [sp]
+        vldr    IN0, [PIN, #-0*4]
+        vldr    IN1, [PIN, #-1*4]
+        vldr    IN2, [PIN, #-2*4]
+        vldr    IN3, [PIN, #-3*4]
+        beq     32f
+64:     dca_lfe_fir  64
+ .ltorg
+32:     dca_lfe_fir  32
+endfunc
+
+        .unreq  POUT
+        .unreq  PIN
+        .unreq  PCOEF
+        .unreq  DECIFACTOR
+        .unreq  OLDFPSCR
+        .unreq  COUNTER
+
+        .unreq  SCALE32
+        .unreq  SCALE64
+        .unreq  IN0
+        .unreq  IN1
+        .unreq  IN2
+        .unreq  IN3
+        .unreq  IN4
+        .unreq  IN5
+        .unreq  IN6
+        .unreq  IN7
+        .unreq  COEF0
+        .unreq  COEF1
+        .unreq  COEF2
+        .unreq  COEF3
+        .unreq  COEF4
+        .unreq  COEF5
+        .unreq  COEF6
+        .unreq  COEF7
+        .unreq  ACCUM0
+        .unreq  ACCUM4
+        .unreq  POST0
+        .unreq  POST1
+        .unreq  POST2
+        .unreq  POST3