RepoMirrors
/
ffmpeg
mirror of https://git.ffmpeg.org/ffmpeg.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
ffmpeg/libavcodec/arm/fft_vfp.S

531 lines
19 KiB
ArmAsm
Raw Normal View History

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>
2013-07-19 08:23:57 +00:00
/*
* Copyright (c) 2013 RISC OS Open Ltd
* Author: Ben Avison <bavison@riscosopen.org>
*
Merge commit '8b9eba664edaddf9a304d3acbf0388b5c520781d' * commit '8b9eba664edaddf9a304d3acbf0388b5c520781d': arm: Add VFP-accelerated version of fft16 Merged-by: Michael Niedermayer <michaelni@gmx.at>
2013-07-22 10:05:16 +00:00
* This file is part of FFmpeg.
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>
2013-07-19 08:23:57 +00:00
*
Merge commit '8b9eba664edaddf9a304d3acbf0388b5c520781d' * commit '8b9eba664edaddf9a304d3acbf0388b5c520781d': arm: Add VFP-accelerated version of fft16 Merged-by: Michael Niedermayer <michaelni@gmx.at>
2013-07-22 10:05:16 +00:00
* FFmpeg is free software; you can redistribute it and/or
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>
2013-07-19 08:23:57 +00:00
* 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.
*
Merge commit '8b9eba664edaddf9a304d3acbf0388b5c520781d' * commit '8b9eba664edaddf9a304d3acbf0388b5c520781d': arm: Add VFP-accelerated version of fft16 Merged-by: Michael Niedermayer <michaelni@gmx.at>
2013-07-22 10:05:16 +00:00
* FFmpeg is distributed in the hope that it will be useful,
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>
2013-07-19 08:23:57 +00:00
* 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
Merge commit '8b9eba664edaddf9a304d3acbf0388b5c520781d' * commit '8b9eba664edaddf9a304d3acbf0388b5c520781d': arm: Add VFP-accelerated version of fft16 Merged-by: Michael Niedermayer <michaelni@gmx.at>
2013-07-22 10:05:16 +00:00
* License along with FFmpeg; if not, write to the Free Software
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>
2013-07-19 08:23:57 +00:00
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/arm/asm.S"
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
@ The fftx_internal_vfp versions of the functions obey a modified AAPCS:
@ VFP is in RunFast mode, vector length 4, stride 1 thoroughout, and
@ all single-precision VFP registers may be corrupted on exit. The a2
@ register may not be clobbered in these functions, as it holds the
@ stored original FPSCR.
function ff_fft_calc_vfp, export=1
ldr ip, [a1, #0] @ nbits
mov a1, a2
arm: fft_vfp: Unify the behaviour in ff_fft_calc_vfp between arm/thumb Don't include the function pointer table in the code segment in arm mode. This shouldn't have any significant performance effect. It does end up as a few more instructions than before, for ARM, but only at the entry to this function, not within the fft functions themselves. Signed-off-by: Martin Storsjö <martin@martin.st>
2014-11-16 10:46:51 +00:00
movrel a2, (fft_tab_vfp - 8)
ldr pc, [a2, ip, lsl #2]
endfunc
arm: Use .data.rel.ro for const data with relocations Signed-off-by: Martin Storsjö <martin@martin.st>
2014-11-16 10:51:54 +00:00
const fft_tab_vfp, relocate=1
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
.word fft4_vfp
.word fft8_vfp
.word X(ff_fft16_vfp) @ this one alone is exported
.word fft32_vfp
.word fft64_vfp
.word fft128_vfp
.word fft256_vfp
.word fft512_vfp
.word fft1024_vfp
.word fft2048_vfp
.word fft4096_vfp
.word fft8192_vfp
.word fft16384_vfp
.word fft32768_vfp
.word fft65536_vfp
arm: fft_vfp: Unify the behaviour in ff_fft_calc_vfp between arm/thumb Don't include the function pointer table in the code segment in arm mode. This shouldn't have any significant performance effect. It does end up as a few more instructions than before, for ARM, but only at the entry to this function, not within the fft functions themselves. Signed-off-by: Martin Storsjö <martin@martin.st>
2014-11-16 10:46:51 +00:00
endconst
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>
2013-07-19 08:23:57 +00:00
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
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
function .Lfft8_internal_vfp
macro_fft8_head
macro_fft8_tail
bx lr
endfunc
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>
2013-07-19 08:23:57 +00:00
function fft8_vfp
ldr a3, =0x03030000 @ RunFast mode, vector length 4, stride 1
fmrx a2, FPSCR
fmxr FPSCR, a3
vpush {s16-s31}
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
mov ip, lr
bl .Lfft8_internal_vfp
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>
2013-07-19 08:23:57 +00:00
vpop {s16-s31}
fmxr FPSCR, a2
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
bx ip
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>
2013-07-19 08:23:57 +00:00
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
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
function .Lfft16_internal_vfp
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>
2013-07-19 08:23:57 +00:00
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]
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
bx lr
endfunc
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}
mov ip, lr
bl .Lfft16_internal_vfp
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>
2013-07-19 08:23:57 +00:00
vpop {s16-s31}
fmxr FPSCR, a2
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
bx ip
endfunc
.macro pass n, z0, z1, z2, z3
add v6, v5, #4*2*\n
@ TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3])
@ TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1])
@ TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0])
@ TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1])
vldr d8, [\z2, #8*(o2+1)] @ s16,s17
vldmdb v6!, {s2}
vldr d9, [\z3, #8*(o3+1)] @ s18,s19
vldmia v5!, {s0,s1} @ s0 is unused
vldr s7, [\z2, #8*o2] @ t1
vmul.f s20, s16, s2 @ vector * scalar
vldr s0, [\z3, #8*o3] @ t5
vldr s6, [\z2, #8*o2+4] @ t2
vldr s3, [\z3, #8*o3+4] @ t6
vmul.f s16, s16, s1 @ vector * scalar
ldr a4, =\n-1
1: add \z0, \z0, #8*2
.if \n*4*2 >= 512
add \z1, \z1, #8*2
.endif
.if \n*4*2 >= 256
add \z2, \z2, #8*2
.endif
.if \n*4*2 >= 512
add \z3, \z3, #8*2
.endif
@ up to 2 stalls (VFP vector issuing / waiting for s0)
@ depending upon whether this is the first iteration and
@ how many add instructions are inserted above
vadd.f s4, s0, s7 @ t5
vadd.f s5, s6, s3 @ t6
vsub.f s6, s6, s3 @ t4
vsub.f s7, s0, s7 @ t3
vldr d6, [\z0, #8*0-8*2] @ s12,s13
vadd.f s0, s16, s21 @ t1
vldr d7, [\z1, #8*o1-8*2] @ s14,s15
vsub.f s1, s18, s23 @ t5
vadd.f s8, s4, s12 @ vector + vector
@ stall (VFP vector issuing)
@ stall (VFP vector issuing)
@ stall (VFP vector issuing)
vsub.f s4, s12, s4
vsub.f s5, s13, s5
vsub.f s6, s14, s6
vsub.f s7, s15, s7
vsub.f s2, s17, s20 @ t2
vadd.f s3, s19, s22 @ t6
vstr d4, [\z0, #8*0-8*2] @ s8,s9
vstr d5, [\z1, #8*o1-8*2] @ s10,s11
@ stall (waiting for s5)
vstr d2, [\z2, #8*o2-8*2] @ s4,s5
vadd.f s4, s1, s0 @ t5
vstr d3, [\z3, #8*o3-8*2] @ s6,s7
vsub.f s7, s1, s0 @ t3
vadd.f s5, s2, s3 @ t6
vsub.f s6, s2, s3 @ t4
vldr d6, [\z0, #8*1-8*2] @ s12,s13
vldr d7, [\z1, #8*(o1+1)-8*2] @ s14,s15
vldr d4, [\z2, #8*o2] @ s8,s9
vldmdb v6!, {s2,s3}
vldr d5, [\z3, #8*o3] @ s10,s11
vadd.f s20, s4, s12 @ vector + vector
vldmia v5!, {s0,s1}
vldr d8, [\z2, #8*(o2+1)] @ s16,s17
@ stall (VFP vector issuing)
vsub.f s4, s12, s4
vsub.f s5, s13, s5
vsub.f s6, s14, s6
vsub.f s7, s15, s7
vmul.f s12, s8, s3 @ vector * scalar
vstr d10, [\z0, #8*1-8*2] @ s20,s21
vldr d9, [\z3, #8*(o3+1)] @ s18,s19
vstr d11, [\z1, #8*(o1+1)-8*2] @ s22,s23
vmul.f s8, s8, s0 @ vector * scalar
vstr d2, [\z2, #8*(o2+1)-8*2] @ s4,s5
@ stall (waiting for s7)
vstr d3, [\z3, #8*(o3+1)-8*2] @ s6,s7
vmul.f s20, s16, s2 @ vector * scalar
@ stall (VFP vector issuing)
@ stall (VFP vector issuing)
@ stall (VFP vector issuing)
vadd.f s7, s8, s13 @ t1
vsub.f s6, s9, s12 @ t2
vsub.f s0, s10, s15 @ t5
vadd.f s3, s11, s14 @ t6
vmul.f s16, s16, s1 @ vector * scalar
subs a4, a4, #1
bne 1b
@ What remains is identical to the first two indentations of
@ the above, but without the increment of z
vadd.f s4, s0, s7 @ t5
vadd.f s5, s6, s3 @ t6
vsub.f s6, s6, s3 @ t4
vsub.f s7, s0, s7 @ t3
vldr d6, [\z0, #8*0] @ s12,s13
vadd.f s0, s16, s21 @ t1
vldr d7, [\z1, #8*o1] @ s14,s15
vsub.f s1, s18, s23 @ t5
vadd.f s8, s4, s12 @ vector + vector
vsub.f s4, s12, s4
vsub.f s5, s13, s5
vsub.f s6, s14, s6
vsub.f s7, s15, s7
vsub.f s2, s17, s20 @ t2
vadd.f s3, s19, s22 @ t6
vstr d4, [\z0, #8*0] @ s8,s9
vstr d5, [\z1, #8*o1] @ s10,s11
vstr d2, [\z2, #8*o2] @ s4,s5
vadd.f s4, s1, s0 @ t5
vstr d3, [\z3, #8*o3] @ s6,s7
vsub.f s7, s1, s0 @ t3
vadd.f s5, s2, s3 @ t6
vsub.f s6, s2, s3 @ t4
vldr d6, [\z0, #8*1] @ s12,s13
vldr d7, [\z1, #8*(o1+1)] @ s14,s15
vadd.f s20, s4, s12 @ vector + vector
vsub.f s4, s12, s4
vsub.f s5, s13, s5
vsub.f s6, s14, s6
vsub.f s7, s15, s7
vstr d10, [\z0, #8*1] @ s20,s21
vstr d11, [\z1, #8*(o1+1)] @ s22,s23
vstr d2, [\z2, #8*(o2+1)] @ s4,s5
vstr d3, [\z3, #8*(o3+1)] @ s6,s7
.endm
.macro def_fft n, n2, n4
function .Lfft\n\()_internal_vfp
.if \n >= 512
push {v1-v6,lr}
.elseif \n >= 256
push {v1-v2,v5-v6,lr}
.else
push {v1,v5-v6,lr}
.endif
mov v1, a1
bl .Lfft\n2\()_internal_vfp
add a1, v1, #8*(\n/4)*2
bl .Lfft\n4\()_internal_vfp
movrelx v5, X(ff_cos_\n), a1
add a1, v1, #8*(\n/4)*3
bl .Lfft\n4\()_internal_vfp
.if \n >= 512
.set o1, 0*(\n/4/2)
.set o2, 0*(\n/4/2)
.set o3, 0*(\n/4/2)
add v2, v1, #8*2*(\n/4/2)
add v3, v1, #8*4*(\n/4/2)
add v4, v1, #8*6*(\n/4/2)
pass (\n/4/2), v1, v2, v3, v4
pop {v1-v6,pc}
.elseif \n >= 256
.set o1, 2*(\n/4/2)
.set o2, 0*(\n/4/2)
.set o3, 2*(\n/4/2)
add v2, v1, #8*4*(\n/4/2)
pass (\n/4/2), v1, v1, v2, v2
pop {v1-v2,v5-v6,pc}
.else
.set o1, 2*(\n/4/2)
.set o2, 4*(\n/4/2)
.set o3, 6*(\n/4/2)
pass (\n/4/2), v1, v1, v1, v1
pop {v1,v5-v6,pc}
.endif
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>
2013-07-19 08:23:57 +00:00
endfunc
armv6: Accelerate ff_fft_calc for general case (nbits != 4) The previous implementation targeted DTS Coherent Acoustics, which only requires nbits == 4 (fft16()). This case was (and still is) linked directly rather than being indirected through ff_fft_calc_vfp(), but now the full range from radix-4 up to radix-65536 is available. This benefits other codecs such as AAC and AC3. The implementaion is based upon the C version, with each routine larger than radix-16 calling a hierarchy of smaller FFT functions, then performing a post-processing pass. This pass benefits a lot from loop unrolling to counter the long pipelines in the VFP. A relaxed calling standard also reduces the overhead of the call hierarchy, and avoiding the excessive inlining performed by GCC probably helps with I-cache utilisation too. I benchmarked the result by measuring the number of gperftools samples that hit anywhere in the AAC decoder (starting from aac_decode_frame()) or specifically in the FFT routines (fft4() to fft512() and pass()) for the same sample AAC stream: Before After Mean StdDev Mean StdDev Confidence Change Audio decode 2245.5 53.1 1599.6 43.8 100.0% +40.4% FFT routines 940.6 22.0 348.1 20.8 100.0% +170.2% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-16 15:02:01 +00:00
function fft\n\()_vfp
ldr a3, =0x03030000 /* RunFast mode, vector length 4, stride 1 */
fmrx a2, FPSCR
fmxr FPSCR, a3
vpush {s16-s31}
mov ip, lr
bl .Lfft\n\()_internal_vfp
vpop {s16-s31}
fmxr FPSCR, a2
bx ip
endfunc
.ltorg
.endm
def_fft 32, 16, 8
def_fft 64, 32, 16
def_fft 128, 64, 32
def_fft 256, 128, 64
def_fft 512, 256, 128
def_fft 1024, 512, 256
def_fft 2048, 1024, 512
def_fft 4096, 2048, 1024
def_fft 8192, 4096, 2048
def_fft 16384, 8192, 4096
def_fft 32768, 16384, 8192
def_fft 65536, 32768, 16384