arm: Add VFP-accelerated version of fft16

               Before           After
               Mean    StdDev   Mean    StdDev  Change
This function   1389.3  4.2       967.8  35.1   +43.6%
Overall        15577.5 83.2     15400.0 336.4    +1.2%

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

3 files changed, 301 insertions, 3 deletions

diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
index 9bb8795bf7..e941aaa806 100644
--- a/libavcodec/arm/Makefile
+++ b/libavcodec/arm/Makefile
@@ -54,6 +54,7 @@ ARMV6-OBJS-$(CONFIG_VP8_DECODER)       += arm/vp8_armv6.o               \
 
 VFP-OBJS-$(CONFIG_DCA_DECODER)         += arm/dcadsp_vfp.o              \
                                           arm/synth_filter_vfp.o
+VFP-OBJS-$(CONFIG_FFT)                 += arm/fft_vfp.o
 VFP-OBJS-$(CONFIG_MDCT)                += arm/mdct_vfp.o
 VFP-OBJS-$(HAVE_ARMV6)                 += arm/fmtconvert_vfp.o
 
diff --git a/libavcodec/arm/fft_vfp.S b/libavcodec/arm/fft_vfp.S
new file mode 100644
index 0000000000..7845ebb3d2
--- /dev/null
+++ b/libavcodec/arm/fft_vfp.S
@@ -0,0 +1,298 @@
+/*
+ * 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"
+
+@ TODO: * FFTs wider than 16
+@       * dispatch code
+
+function fft4_vfp
+        vldr    d0, [a1, #0*2*4]   @ s0,s1   = z[0]
+        vldr    d4, [a1, #1*2*4]   @ s8,s9   = z[1]
+        vldr    d1, [a1, #2*2*4]   @ s2,s3   = z[2]
+        vldr    d5, [a1, #3*2*4]   @ s10,s11 = z[3]
+        @ stall
+        vadd.f  s12, s0, s8        @ i0
+        vadd.f  s13, s1, s9        @ i1
+        vadd.f  s14, s2, s10       @ i2
+        vadd.f  s15, s3, s11       @ i3
+        vsub.f  s8, s0, s8         @ i4
+        vsub.f  s9, s1, s9         @ i5
+        vsub.f  s10, s2, s10       @ i6
+        vsub.f  s11, s3, s11       @ i7
+        @ stall
+        @ stall
+        vadd.f  s0, s12, s14       @ z[0].re
+        vsub.f  s4, s12, s14       @ z[2].re
+        vadd.f  s1, s13, s15       @ z[0].im
+        vsub.f  s5, s13, s15       @ z[2].im
+        vadd.f  s7, s9, s10        @ z[3].im
+        vsub.f  s3, s9, s10        @ z[1].im
+        vadd.f  s2, s8, s11        @ z[1].re
+        vsub.f  s6, s8, s11        @ z[3].re
+        @ stall
+        @ stall
+        vstr    d0, [a1, #0*2*4]
+        vstr    d2, [a1, #2*2*4]
+        @ stall
+        @ stall
+        vstr    d1, [a1, #1*2*4]
+        vstr    d3, [a1, #3*2*4]
+
+        bx      lr
+endfunc
+
+.macro macro_fft8_head
+        @ FFT4
+        vldr    d4, [a1, #0 * 2*4]
+        vldr    d6, [a1, #1 * 2*4]
+        vldr    d5, [a1, #2 * 2*4]
+        vldr    d7, [a1, #3 * 2*4]
+            @ BF
+            vldr    d12, [a1, #4 * 2*4]
+        vadd.f  s16, s8, s12    @ vector op
+            vldr    d14, [a1, #5 * 2*4]
+            vldr    d13, [a1, #6 * 2*4]
+            vldr    d15, [a1, #7 * 2*4]
+        vsub.f  s20, s8, s12    @ vector op
+        vadd.f  s0, s16, s18
+        vsub.f  s2, s16, s18
+        vadd.f  s1, s17, s19
+        vsub.f  s3, s17, s19
+        vadd.f  s7, s21, s22
+        vsub.f  s5, s21, s22
+        vadd.f  s4, s20, s23
+        vsub.f  s6, s20, s23
+            vsub.f  s20, s24, s28   @ vector op
+        vstr    d0, [a1, #0 * 2*4]  @ transfer s0-s7 to s24-s31 via memory
+        vstr    d1, [a1, #1 * 2*4]
+        vldr    s0, cos1pi4
+            vadd.f  s16, s24, s28   @ vector op
+        vstr    d2, [a1, #2 * 2*4]
+        vstr    d3, [a1, #3 * 2*4]
+        vldr    d12, [a1, #0 * 2*4]
+            @ TRANSFORM
+            vmul.f  s20, s20, s0    @ vector x scalar op
+        vldr    d13, [a1, #1 * 2*4]
+        vldr    d14, [a1, #2 * 2*4]
+        vldr    d15, [a1, #3 * 2*4]
+        @ BUTTERFLIES
+        vadd.f  s0, s18, s16
+        vadd.f  s1, s17, s19
+        vsub.f  s2, s17, s19
+        vsub.f  s3, s18, s16
+            vadd.f  s4, s21, s20
+            vsub.f  s5, s21, s20
+            vadd.f  s6, s22, s23
+            vsub.f  s7, s22, s23
+        vadd.f  s8, s0, s24         @ vector op
+        vstr    d0, [a1, #0 * 2*4]  @ transfer s0-s3 to s12-s15 via memory
+        vstr    d1, [a1, #1 * 2*4]
+        vldr    d6, [a1, #0 * 2*4]
+        vldr    d7, [a1, #1 * 2*4]
+            vadd.f  s1, s5, s6
+            vadd.f  s0, s7, s4
+            vsub.f  s2, s5, s6
+            vsub.f  s3, s7, s4
+        vsub.f  s12, s24, s12       @ vector op
+            vsub.f  s5, s29, s1
+            vsub.f  s4, s28, s0
+            vsub.f  s6, s30, s2
+            vsub.f  s7, s31, s3
+            vadd.f  s16, s0, s28    @ vector op
+        vstr    d6, [a1, #4 * 2*4]
+        vstr    d7, [a1, #6 * 2*4]
+        vstr    d4, [a1, #0 * 2*4]
+        vstr    d5, [a1, #2 * 2*4]
+             vstr    d2, [a1, #5 * 2*4]
+             vstr    d3, [a1, #7 * 2*4]
+.endm
+
+.macro macro_fft8_tail
+             vstr    d8, [a1, #1 * 2*4]
+             vstr    d9, [a1, #3 * 2*4]
+.endm
+
+function fft8_vfp
+        ldr     a3, =0x03030000     @ RunFast mode, vector length 4, stride 1
+        fmrx    a2, FPSCR
+        fmxr    FPSCR, a3
+        vpush   {s16-s31}
+
+        macro_fft8_head
+        macro_fft8_tail
+
+        vpop    {s16-s31}
+        fmxr    FPSCR, a2
+        bx      lr
+endfunc
+
+.align 3
+cos1pi4:    @ cos(1*pi/4) = sqrt(2)
+        .float  0.707106769084930419921875
+cos1pi8:    @ cos(1*pi/8) = sqrt(2+sqrt(2))/2
+        .float  0.92387950420379638671875
+cos3pi8:    @ cos(2*pi/8) = sqrt(2-sqrt(2))/2
+        .float  0.3826834261417388916015625
+
+function ff_fft16_vfp, export=1
+        ldr     a3, =0x03030000     @ RunFast mode, vector length 4, stride 1
+        fmrx    a2, FPSCR
+        fmxr    FPSCR, a3
+        vpush   {s16-s31}
+
+        macro_fft8_head
+        @ FFT4(z+8)
+        vldr    d10, [a1, #8 * 2*4]
+        vldr    d12, [a1, #9 * 2*4]
+        vldr    d11, [a1, #10 * 2*4]
+        vldr    d13, [a1, #11 * 2*4]
+        macro_fft8_tail
+        vadd.f  s16, s20, s24   @ vector op
+            @ FFT4(z+12)
+            vldr    d4, [a1, #12 * 2*4]
+            vldr    d6, [a1, #13 * 2*4]
+            vldr    d5, [a1, #14 * 2*4]
+        vsub.f  s20, s20, s24   @ vector op
+            vldr    d7, [a1, #15 * 2*4]
+        vadd.f  s0, s16, s18
+        vsub.f  s4, s16, s18
+        vadd.f  s1, s17, s19
+        vsub.f  s5, s17, s19
+        vadd.f  s7, s21, s22
+        vsub.f  s3, s21, s22
+        vadd.f  s2, s20, s23
+        vsub.f  s6, s20, s23
+            vadd.f  s16, s8, s12    @ vector op
+        vstr    d0, [a1, #8 * 2*4]
+        vstr    d2, [a1, #10 * 2*4]
+        vstr    d1, [a1, #9 * 2*4]
+            vsub.f  s20, s8, s12
+        vstr    d3, [a1, #11 * 2*4]
+        @ TRANSFORM(z[2],z[6],z[10],z[14],cos1pi4,cos1pi4)
+        vldr    d12, [a1, #10 * 2*4]
+            vadd.f  s0, s16, s18
+            vadd.f  s1, s17, s19
+            vsub.f  s6, s16, s18
+            vsub.f  s7, s17, s19
+            vsub.f  s3, s21, s22
+            vadd.f  s2, s20, s23
+            vadd.f  s5, s21, s22
+            vsub.f  s4, s20, s23
+            vstr    d0, [a1, #12 * 2*4]
+        vmov    s0, s6
+          @ TRANSFORM(z[1],z[5],z[9],z[13],cos1pi8,cos3pi8)
+          vldr    d6, [a1, #9 * 2*4]
+            vstr    d1, [a1, #13 * 2*4]
+        vldr    d1, cos1pi4 @ s2 = cos1pi4, s3 = cos1pi8
+            vstr    d2, [a1, #15 * 2*4]
+          vldr    d7, [a1, #13 * 2*4]
+        vadd.f  s4, s25, s24
+        vsub.f  s5, s25, s24
+        vsub.f  s6, s0, s7
+        vadd.f  s7, s0, s7
+          vmul.f  s20, s12, s3  @ vector op
+            @ TRANSFORM(z[3],z[7],z[11],z[15],cos3pi8,cos1pi8)
+            vldr    d4, [a1, #11 * 2*4]
+            vldr    d5, [a1, #15 * 2*4]
+            vldr    s1, cos3pi8
+        vmul.f  s24, s4, s2     @ vector * scalar op
+          vmul.f  s28, s12, s1  @ vector * scalar op
+            vmul.f  s12, s8, s1 @ vector * scalar op
+          vadd.f  s4, s20, s29
+          vsub.f  s5, s21, s28
+          vsub.f  s6, s22, s31
+          vadd.f  s7, s23, s30
+            vmul.f  s8, s8, s3  @ vector * scalar op
+          vldr    d8, [a1, #1 * 2*4]
+          vldr    d9, [a1, #5 * 2*4]
+            vldr    d10, [a1, #3 * 2*4]
+            vldr    d11, [a1, #7 * 2*4]
+        vldr    d14, [a1, #2 * 2*4]
+          vadd.f  s0, s6, s4
+          vadd.f  s1, s5, s7
+          vsub.f  s2, s5, s7
+          vsub.f  s3, s6, s4
+            vadd.f  s4, s12, s9
+            vsub.f  s5, s13, s8
+            vsub.f  s6, s14, s11
+            vadd.f  s7, s15, s10
+          vadd.f  s12, s0, s16  @ vector op
+          vstr    d0, [a1, #1 * 2*4]
+          vstr    d1, [a1, #5 * 2*4]
+          vldr    d4, [a1, #1 * 2*4]
+          vldr    d5, [a1, #5 * 2*4]
+            vadd.f  s0, s6, s4
+            vadd.f  s1, s5, s7
+            vsub.f  s2, s5, s7
+            vsub.f  s3, s6, s4
+          vsub.f  s8, s16, s8   @ vector op
+          vstr    d6, [a1, #1 * 2*4]
+          vstr    d7, [a1, #5 * 2*4]
+        vldr    d15, [a1, #6 * 2*4]
+            vsub.f  s4, s20, s0
+            vsub.f  s5, s21, s1
+            vsub.f  s6, s22, s2
+            vsub.f  s7, s23, s3
+            vadd.f  s20, s0, s20    @ vector op
+          vstr    d4, [a1, #9 * 2*4]
+              @ TRANSFORM_ZERO(z[0],z[4],z[8],z[12])
+              vldr    d6, [a1, #8 * 2*4]
+          vstr    d5, [a1, #13 * 2*4]
+              vldr    d7, [a1, #12 * 2*4]
+          vstr    d2, [a1, #11 * 2*4]
+              vldr    d8, [a1, #0 * 2*4]
+          vstr    d3, [a1, #15 * 2*4]
+              vldr    d9, [a1, #4 * 2*4]
+        vadd.f  s0, s26, s24
+        vadd.f  s1, s25, s27
+        vsub.f  s2, s25, s27
+        vsub.f  s3, s26, s24
+              vadd.f  s4, s14, s12
+              vadd.f  s5, s13, s15
+              vsub.f  s6, s13, s15
+              vsub.f  s7, s14, s12
+        vadd.f  s8, s0, s28 @ vector op
+        vstr    d0, [a1, #3 * 2*4]
+        vstr    d1, [a1, #7 * 2*4]
+        vldr    d6, [a1, #3 * 2*4]
+        vldr    d7, [a1, #7 * 2*4]
+              vsub.f  s0, s16, s4
+              vsub.f  s1, s17, s5
+              vsub.f  s2, s18, s6
+              vsub.f  s3, s19, s7
+        vsub.f  s12, s28, s12       @ vector op
+              vadd.f  s16, s4, s16  @ vector op
+            vstr    d10, [a1, #3 * 2*4]
+            vstr    d11, [a1, #7 * 2*4]
+        vstr    d4, [a1, #2 * 2*4]
+        vstr    d5, [a1, #6 * 2*4]
+              vstr    d0, [a1, #8 * 2*4]
+              vstr    d1, [a1, #12 * 2*4]
+        vstr    d6, [a1, #10 * 2*4]
+        vstr    d7, [a1, #14 * 2*4]
+              vstr    d8, [a1, #0 * 2*4]
+              vstr    d9, [a1, #4 * 2*4]
+
+        vpop    {s16-s31}
+        fmxr    FPSCR, a2
+        bx      lr
+endfunc
diff --git a/libavcodec/arm/mdct_vfp.S b/libavcodec/arm/mdct_vfp.S
index 7413a41c66..0623e9618e 100644
--- a/libavcodec/arm/mdct_vfp.S
+++ b/libavcodec/arm/mdct_vfp.S
@@ -174,9 +174,8 @@ function ff_imdct_half_vfp, export=1
  .endr
 
         fmxr    FPSCR, OLDFPSCR
-        mov     ORIGOUT, OUT
-        ldr     ip, [CONTEXT, #9*4]
-        blx     ip                          @ s->fft_calc(s, output)
+        mov     a1, OUT
+        bl      ff_fft16_vfp
         ldr     lr, =0x03030000             @ RunFast mode, short vectors of length 4, stride 1
         fmxr    FPSCR, lr