[FFmpeg-cvslog] aarch64: NEON float (i)MDCT

Janne Grunau git at videolan.org
Tue Apr 22 23:33:45 CEST 2014


ffmpeg | branch: master | Janne Grunau <janne-libav at jannau.net> | Tue Apr 15 18:35:57 2014 +0200| [ee2bc5974fe64fd214f52574400ae01c85f4b855] | committer: Janne Grunau

aarch64: NEON float (i)MDCT

Approximately as fast as the ARM NEON version on Apple's A7.

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=ee2bc5974fe64fd214f52574400ae01c85f4b855
---

 libavcodec/aarch64/Makefile           |    1 +
 libavcodec/aarch64/fft_init_aarch64.c |   10 +
 libavcodec/aarch64/mdct_neon.S        |  323 +++++++++++++++++++++++++++++++++
 3 files changed, 334 insertions(+)

diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
index 01f618f..ec1fd1e 100644
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -18,3 +18,4 @@ NEON-OBJS-$(CONFIG_H264DSP)             += aarch64/h264dsp_neon.o              \
 NEON-OBJS-$(CONFIG_H264QPEL)            += aarch64/h264qpel_neon.o             \
                                            aarch64/hpeldsp_neon.o
 NEON-OBJS-$(CONFIG_HPELDSP)             += aarch64/hpeldsp_neon.o
+NEON-OBJS-$(CONFIG_MDCT)                += aarch64/mdct_neon.o
diff --git a/libavcodec/aarch64/fft_init_aarch64.c b/libavcodec/aarch64/fft_init_aarch64.c
index caa5a0d..589e82d 100644
--- a/libavcodec/aarch64/fft_init_aarch64.c
+++ b/libavcodec/aarch64/fft_init_aarch64.c
@@ -26,6 +26,10 @@
 void ff_fft_permute_neon(FFTContext *s, FFTComplex *z);
 void ff_fft_calc_neon(FFTContext *s, FFTComplex *z);
 
+void ff_imdct_calc_neon(FFTContext *s, FFTSample *output, const FFTSample *input);
+void ff_imdct_half_neon(FFTContext *s, FFTSample *output, const FFTSample *input);
+void ff_mdct_calc_neon(FFTContext *s, FFTSample *output, const FFTSample *input);
+
 av_cold void ff_fft_init_aarch64(FFTContext *s)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -33,5 +37,11 @@ av_cold void ff_fft_init_aarch64(FFTContext *s)
     if (have_neon(cpu_flags)) {
         s->fft_permute  = ff_fft_permute_neon;
         s->fft_calc     = ff_fft_calc_neon;
+#if CONFIG_MDCT
+        s->imdct_calc   = ff_imdct_calc_neon;
+        s->imdct_half   = ff_imdct_half_neon;
+        s->mdct_calc    = ff_mdct_calc_neon;
+        s->mdct_permutation = FF_MDCT_PERM_INTERLEAVE;
+#endif
     }
 }
diff --git a/libavcodec/aarch64/mdct_neon.S b/libavcodec/aarch64/mdct_neon.S
new file mode 100644
index 0000000..bccd832
--- /dev/null
+++ b/libavcodec/aarch64/mdct_neon.S
@@ -0,0 +1,323 @@
+/*
+ * AArch64 NEON optimised MDCT
+ * Copyright (c) 2009 Mans Rullgard <mans at mansr.com>
+ * Copyright (c) 2014 Janne Grunau <janne-libav at jannau.net>
+ *
+ * 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/aarch64/asm.S"
+
+function ff_imdct_half_neon, export=1
+        sub             sp,  sp,  #32
+        stp             x19, x20, [sp]
+        str             x30, [sp, #16]
+        mov             x12, #1
+        ldr             w14, [x0, #28]          // mdct_bits
+        ldr             x4,  [x0, #32]          // tcos
+        ldr             x3,  [x0, #8]           // revtab
+        lsl             x12, x12, x14           // n  = 1 << nbits
+        lsr             x14, x12, #2            // n4 = n >> 2
+        add             x7,  x2,  x12,  lsl #1
+        mov             x12, #-16
+        sub             x7,  x7,  #16
+
+        ld2             {v16.2s,v17.2s}, [x7], x12 // d16=x,n1 d17=x,n0
+        ld2             {v0.2s,v1.2s},   [x2], #16 // d0 =m0,x d1 =m1,x
+        rev64           v17.2s, v17.2s
+        ld2             {v2.2s,v3.2s},   [x4], #16 // d2=c0,c1 d3=s0,s2
+        fmul            v6.2s,  v17.2s, v2.2s
+        fmul            v7.2s,  v0.2s,  v2.2s
+1:
+        subs            x14, x14, #2
+        ldr             w6,  [x3], #4
+        fmul            v4.2s,  v0.2s,  v3.2s
+        fmul            v5.2s,  v17.2s, v3.2s
+        fsub            v4.2s,  v6.2s,  v4.2s
+        fadd            v5.2s,  v5.2s,  v7.2s
+        ubfm            x8,  x6,  #16, #31
+        ubfm            x6,  x6,  #0,  #15
+        add             x8,  x1,  x8,  lsl #3
+        add             x6,  x1,  x6,  lsl #3
+        b.eq            2f
+        ld2             {v16.2s,v17.2s}, [x7], x12
+        ld2             {v0.2s,v1.2s},   [x2], #16
+        rev64           v17.2s, v17.2s
+        ld2             {v2.2s,v3.2s},   [x4], #16    // d2=c0,c1 d3=s0,s2
+        fmul            v6.2s,  v17.2s, v2.2s
+        fmul            v7.2s,  v0.2s,  v2.2s
+        st2             {v4.s,v5.s}[0], [x6]
+        st2             {v4.s,v5.s}[1], [x8]
+        b               1b
+2:
+        st2             {v4.s,v5.s}[0], [x6]
+        st2             {v4.s,v5.s}[1], [x8]
+
+        mov             x19, x0
+        mov             x20, x1
+        bl              X(ff_fft_calc_neon)
+
+        mov             x12, #1
+        ldr             w14, [x19, #28]          // mdct_bits
+        ldr             x4,  [x19, #32]          // tcos
+        lsl             x12, x12, x14            // n  = 1 << nbits
+        lsr             x14, x12, #3             // n8 = n >> 3
+
+        add             x4,  x4,  x14, lsl #3
+        add             x6,  x20, x14, lsl #3
+        sub             x1,  x4,  #16
+        sub             x3,  x6,  #16
+
+        mov             x7,  #-16
+        mov             x8,  x6
+        mov             x0,  x3
+
+        ld2             {v0.2s,v1.2s},  [x3], x7 // d0 =i1,r1 d1 =i0,r0
+        ld2             {v20.2s,v21.2s},[x6], #16 // d20=i2,r2 d21=i3,r3
+        ld2             {v16.2s,v17.2s},[x1], x7 // d16=c1,c0 d18=s1,s0
+3:
+        subs            x14, x14, #2
+        fmul            v7.2s,  v0.2s,  v17.2s
+        ld2             {v18.2s,v19.2s},[x4], #16    // d17=c2,c3 d19=s2,s3
+        fmul            v4.2s,  v1.2s,  v17.2s
+        fmul            v6.2s,  v21.2s, v19.2s
+        fmul            v5.2s,  v20.2s, v19.2s
+        fmul            v22.2s, v1.2s,  v16.2s
+        fmul            v23.2s, v21.2s, v18.2s
+        fmul            v24.2s, v0.2s,  v16.2s
+        fmul            v25.2s, v20.2s, v18.2s
+        fadd            v7.2s,  v7.2s,  v22.2s
+        fadd            v5.2s,  v5.2s,  v23.2s
+        fsub            v4.2s,  v4.2s,  v24.2s
+        fsub            v6.2s,  v6.2s,  v25.2s
+        b.eq            4f
+        ld2             {v0.2s,v1.2s},  [x3], x7
+        ld2             {v20.2s,v21.2s},[x6], #16
+        ld2             {v16.2s,v17.2s},[x1], x7 // d16=c1,c0 d18=s1,s0
+        rev64           v5.2s,  v5.2s
+        rev64           v7.2s,  v7.2s
+        st2             {v4.2s,v5.2s},  [x0], x7
+        st2             {v6.2s,v7.2s},  [x8], #16
+        b               3b
+4:
+        rev64           v5.2s,  v5.2s
+        rev64           v7.2s,  v7.2s
+        st2             {v4.2s,v5.2s},  [x0]
+        st2             {v6.2s,v7.2s},  [x8]
+
+        ldp             x19, x20, [sp]
+        ldr             x30, [sp, #16]
+        add             sp,  sp,  #32
+
+        ret
+endfunc
+
+function ff_imdct_calc_neon, export=1
+        sub             sp,  sp,  #32
+        stp             x19, x20, [sp]
+        str             x30, [sp, #16]
+        ldr             w3,  [x0, #28]          // mdct_bits
+        mov             x19, #1
+        mov             x20, x1
+        lsl             x19, x19, x3
+        add             x1,  x1,  x19
+
+        bl              X(ff_imdct_half_neon)
+
+        add             x0,  x20, x19,  lsl #2
+        add             x1,  x20, x19,  lsl #1
+        sub             x0,  x0,  #8
+        sub             x2,  x1,  #16
+        mov             x3,  #-16
+        mov             x6,  #-8
+1:
+        ld1             {v0.4s}, [x2], x3
+        prfum           pldl1keep, [x0, #-16]
+        rev64           v0.4s, v0.4s
+        ld1             {v2.2s,v3.2s}, [x1], #16
+        fneg            v4.4s,  v0.4s
+        prfum           pldl1keep, [x2, #-16]
+        rev64           v2.2s, v2.2s
+        rev64           v3.2s, v3.2s
+        ext             v4.16b, v4.16b, v4.16b, #8
+        st1             {v2.2s}, [x0], x6
+        st1             {v3.2s}, [x0], x6
+        st1             {v4.4s}, [x20], #16
+        subs            x19, x19,  #16
+        b.gt            1b
+
+        ldp             x19, x20, [sp], #16
+        ldr             x30, [sp], #16
+
+        ret
+endfunc
+
+
+function ff_mdct_calc_neon, export=1
+        sub             sp,  sp,  #32
+        stp             x19, x20, [sp]
+        str             x30, [sp, #16]
+
+        mov             x12, #1
+        ldr             w14, [x0, #28]          // mdct_bits
+        ldr             x4,  [x0, #32]          // tcos
+        ldr             x3,  [x0, #8]           // revtab
+        lsl             x14, x12, x14           // n  = 1 << nbits
+        add             x7,  x2,  x14           // in4u
+        sub             x9,  x7,  #16           // in4d
+        add             x2,  x7,  x14, lsl #1   // in3u
+        add             x8,  x9,  x14, lsl #1   // in3d
+        add             x5,  x4,  x14, lsl #1
+        sub             x5,  x5,  #16
+        sub             x3,  x3,  #4
+        mov             x12, #-16
+        lsr             x13, x14, #1
+
+        ld2             {v16.2s,v17.2s}, [x9], x12  // in0u0,in0u1 in4d1,in4d0
+        ld2             {v18.2s,v19.2s}, [x8], x12  // in2u0,in2u1 in3d1,in3d0
+        ld2             {v0.2s, v1.2s},  [x7], #16  // in4u0,in4u1 in2d1,in2d0
+        rev64           v17.2s, v17.2s              // in4d0,in4d1 in3d0,in3d1
+        rev64           v19.2s, v19.2s              // in4d0,in4d1 in3d0,in3d1
+        ld2             {v2.2s, v3.2s},  [x2], #16  // in3u0,in3u1 in1d1,in1d0
+        fsub            v0.2s,  v17.2s, v0.2s       // in4d-in4u      I
+        ld2             {v20.2s,v21.2s}, [x4], #16  // c0,c1 s0,s1
+        rev64           v1.2s,  v1.2s               // in2d0,in2d1 in1d0,in1d1
+        rev64           v3.2s,  v3.2s               // in2d0,in2d1 in1d0,in1d1
+        ld2             {v30.2s,v31.2s}, [x5], x12  // c2,c3 s2,s3
+        fadd            v2.2s,  v2.2s,  v19.2s      // in3u+in3d     -R
+        fsub            v16.2s, v16.2s, v1.2s       // in0u-in2d      R
+        fadd            v18.2s, v18.2s, v3.2s       // in2u+in1d     -I
+1:
+        fmul            v7.2s,  v0.2s,  v21.2s      //  I*s
+        ldr             w10, [x3, x13]
+        fmul            v6.2s,  v2.2s,  v20.2s      // -R*c
+        ldr             w6,  [x3, #4]!
+        fmul            v4.2s,  v2.2s,  v21.2s      // -R*s
+        fmul            v5.2s,  v0.2s,  v20.2s      //  I*c
+        fmul            v24.2s, v16.2s, v30.2s      //  R*c
+        fmul            v25.2s, v18.2s, v31.2s      // -I*s
+        fmul            v22.2s, v16.2s, v31.2s      //  R*s
+        fmul            v23.2s, v18.2s, v30.2s      //  I*c
+        subs            x14, x14, #16
+        subs            x13, x13, #8
+        fsub            v6.2s,  v6.2s,  v7.2s       // -R*c-I*s
+        fadd            v7.2s,  v4.2s,  v5.2s       // -R*s+I*c
+        fsub            v24.2s, v25.2s, v24.2s      // I*s-R*c
+        fadd            v25.2s, v22.2s, v23.2s      // R*s-I*c
+        b.eq            1f
+        mov             x12, #-16
+        ld2             {v16.2s,v17.2s}, [x9], x12  // in0u0,in0u1 in4d1,in4d0
+        ld2             {v18.2s,v19.2s}, [x8], x12  // in2u0,in2u1 in3d1,in3d0
+        fneg            v7.2s,  v7.2s               //  R*s-I*c
+        ld2             {v0.2s, v1.2s},  [x7], #16  // in4u0,in4u1 in2d1,in2d0
+        rev64           v17.2s, v17.2s              // in4d0,in4d1 in3d0,in3d1
+        rev64           v19.2s, v19.2s              // in4d0,in4d1 in3d0,in3d1
+        ld2             {v2.2s, v3.2s},  [x2], #16  // in3u0,in3u1 in1d1,in1d0
+        fsub            v0.2s,  v17.2s, v0.2s       // in4d-in4u      I
+        ld2             {v20.2s,v21.2s}, [x4], #16  // c0,c1 s0,s1
+        rev64           v1.2s,  v1.2s               // in2d0,in2d1 in1d0,in1d1
+        rev64           v3.2s,  v3.2s               // in2d0,in2d1 in1d0,in1d1
+        ld2             {v30.2s,v31.2s}, [x5], x12  // c2,c3 s2,s3
+        fadd            v2.2s,  v2.2s,  v19.2s      // in3u+in3d     -R
+        fsub            v16.2s, v16.2s, v1.2s       // in0u-in2d      R
+        fadd            v18.2s, v18.2s, v3.2s       // in2u+in1d     -I
+        ubfm            x12, x6,  #16, #31
+        ubfm            x6,  x6,  #0,  #15
+        add             x12, x1,  x12, lsl #3
+        add             x6,  x1,  x6,  lsl #3
+        st2             {v6.s,v7.s}[0],   [x6]
+        st2             {v6.s,v7.s}[1],   [x12]
+        ubfm            x6,  x10, #16, #31
+        ubfm            x10, x10, #0,  #15
+        add             x6 , x1,  x6,  lsl #3
+        add             x10, x1,  x10, lsl #3
+        st2             {v24.s,v25.s}[0], [x10]
+        st2             {v24.s,v25.s}[1], [x6]
+        b               1b
+1:
+        fneg            v7.2s,  v7.2s           //  R*s-I*c
+        ubfm            x12, x6,  #16, #31
+        ubfm            x6,  x6,  #0,  #15
+        add             x12, x1,  x12, lsl #3
+        add             x6,  x1,  x6,  lsl #3
+        st2             {v6.s,v7.s}[0],   [x6]
+        st2             {v6.s,v7.s}[1],   [x12]
+        ubfm            x6,  x10, #16, #31
+        ubfm            x10, x10, #0,  #15
+        add             x6 , x1,  x6,  lsl #3
+        add             x10, x1,  x10, lsl #3
+        st2             {v24.s,v25.s}[0], [x10]
+        st2             {v24.s,v25.s}[1], [x6]
+
+        mov             x19, x0
+        mov             x20, x1
+        bl              X(ff_fft_calc_neon)
+
+        mov             x12, #1
+        ldr             w14, [x19, #28]         // mdct_bits
+        ldr             x4,  [x19, #32]         // tcos
+        lsl             x12, x12, x14           // n  = 1 << nbits
+        lsr             x14, x12, #3            // n8 = n >> 3
+
+        add             x4,  x4,  x14, lsl #3
+        add             x6,  x20, x14, lsl #3
+        sub             x1,  x4,  #16
+        sub             x3,  x6,  #16
+
+        mov             x7,  #-16
+        mov             x8,  x6
+        mov             x0,  x3
+
+        ld2             {v0.2s,v1.2s},   [x3], x7   // d0 =r1,i1 d1 =r0,i0
+        ld2             {v20.2s,v21.2s}, [x6], #16  // d20=r2,i2 d21=r3,i3
+        ld2             {v16.2s,v17.2s}, [x1], x7   // c1,c0 s1,s0
+1:
+        subs            x14, x14, #2
+        fmul            v7.2s,  v0.2s,  v17.2s      // r1*s1,r0*s0
+        ld2             {v18.2s,v19.2s}, [x4], #16  // c2,c3 s2,s3
+        fmul            v4.2s,  v1.2s,  v17.2s      // i1*s1,i0*s0
+        fmul            v6.2s,  v21.2s, v19.2s      // i2*s2,i3*s3
+        fmul            v5.2s,  v20.2s, v19.2s      // r2*s2,r3*s3
+        fmul            v24.2s, v0.2s,  v16.2s      // r1*c1,r0*c0
+        fmul            v25.2s, v20.2s, v18.2s      // r2*c2,r3*c3
+        fmul            v22.2s, v21.2s, v18.2s      // i2*c2,i3*c3
+        fmul            v23.2s, v1.2s,  v16.2s      // i1*c1,i0*c0
+        fadd            v4.2s,  v4.2s,  v24.2s      // i1*s1+r1*c1,i0*s0+r0*c0
+        fadd            v6.2s,  v6.2s,  v25.2s      // i2*s2+r2*c2,i3*s3+r3*c3
+        fsub            v5.2s,  v22.2s, v5.2s       // i2*c2-r2*s2,i3*c3-r3*s3
+        fsub            v7.2s,  v23.2s, v7.2s       // i1*c1-r1*s1,i0*c0-r0*s0
+        fneg            v4.2s,  v4.2s
+        fneg            v6.2s,  v6.2s
+        b.eq            1f
+        ld2             {v0.2s, v1.2s},  [x3], x7
+        ld2             {v20.2s,v21.2s}, [x6], #16
+        ld2             {v16.2s,v17.2s}, [x1], x7   // c1,c0 s1,s0
+        rev64           v5.2s,  v5.2s
+        rev64           v7.2s,  v7.2s
+        st2             {v4.2s,v5.2s},  [x0], x7
+        st2             {v6.2s,v7.2s},  [x8], #16
+        b               1b
+1:
+        rev64           v5.2s,  v5.2s
+        rev64           v7.2s,  v7.2s
+        st2             {v4.2s,v5.2s},  [x0]
+        st2             {v6.2s,v7.2s},  [x8]
+
+        ldp             x19, x20, [sp], #16
+        ldr             x30, [sp], #16
+        ret
+endfunc



More information about the ffmpeg-cvslog mailing list