RepoMirrors/ffmpeg
1
0
Fork 0
mirror of https://git.ffmpeg.org/ffmpeg.git synced 2024-12-25 16:52:31 +00:00
Code Issues Projects Releases Wiki Activity
581c7f0e12
ffmpeg/libavcodec/arm/mdct_vfp.S

348 lines
11 KiB
ArmAsm
Raw Normal View History

arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
/*
* 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"
CONTEXT .req a1
ORIGOUT .req a2
IN .req a3
OUT .req v1
REVTAB .req v2
TCOS .req v3
TSIN .req v4
OLDFPSCR .req v5
J0 .req a2
J1 .req a4
J2 .req ip
J3 .req lr
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
REVTAB_HI .req v5
IN_HI .req v6
OUT_HI .req v6
TCOS_HI .req sl
TSIN_HI .req fp
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
.macro prerotation_innerloop
.set trig_lo, k
.set trig_hi, n4 - k - 2
.set in_lo, trig_lo * 2
.set in_hi, trig_hi * 2
vldr d8, [TCOS, #trig_lo*4] @ s16,s17
vldr d9, [TCOS, #trig_hi*4] @ s18,s19
vldr s0, [IN, #in_hi*4 + 12]
vldr s1, [IN, #in_hi*4 + 4]
vldr s2, [IN, #in_lo*4 + 12]
vldr s3, [IN, #in_lo*4 + 4]
vmul.f s8, s0, s16 @ vector operation
vldr d10, [TSIN, #trig_lo*4] @ s20,s21
vldr d11, [TSIN, #trig_hi*4] @ s22,s23
vldr s4, [IN, #in_lo*4]
vldr s5, [IN, #in_lo*4 + 8]
vldr s6, [IN, #in_hi*4]
vldr s7, [IN, #in_hi*4 + 8]
ldr J0, [REVTAB, #trig_lo*2]
vmul.f s12, s0, s20 @ vector operation
ldr J2, [REVTAB, #trig_hi*2]
mov J1, J0, lsr #16
and J0, J0, #255 @ halfword value will be < n4
vmls.f s8, s4, s20 @ vector operation
mov J3, J2, lsr #16
and J2, J2, #255 @ halfword value will be < n4
add J0, OUT, J0, lsl #3
vmla.f s12, s4, s16 @ vector operation
add J1, OUT, J1, lsl #3
add J2, OUT, J2, lsl #3
add J3, OUT, J3, lsl #3
vstr s8, [J0]
vstr s9, [J1]
vstr s10, [J2]
vstr s11, [J3]
vstr s12, [J0, #4]
vstr s13, [J1, #4]
vstr s14, [J2, #4]
vstr s15, [J3, #4]
.set k, k + 2
.endm
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
.macro prerotation_innerloop_rolled
vldmia TCOS!, {s16,s17}
vldmdb TCOS_HI!, {s18,s19}
vldr s0, [IN_HI, #-4]
vldr s1, [IN_HI, #-12]
vldr s2, [IN, #12]
vldr s3, [IN, #4]
vmul.f s8, s0, s16 @ vector operation
vldmia TSIN!, {s20,s21}
vldmdb TSIN_HI!, {s22,s23}
vldr s4, [IN]
vldr s5, [IN, #8]
vldr s6, [IN_HI, #-16]
vldr s7, [IN_HI, #-8]
vmul.f s12, s0, s20 @ vector operation
add IN, IN, #16
sub IN_HI, IN_HI, #16
ldrh J0, [REVTAB], #2
ldrh J1, [REVTAB], #2
vmls.f s8, s4, s20 @ vector operation
ldrh J3, [REVTAB_HI, #-2]!
ldrh J2, [REVTAB_HI, #-2]!
add J0, OUT, J0, lsl #3
vmla.f s12, s4, s16 @ vector operation
add J1, OUT, J1, lsl #3
add J2, OUT, J2, lsl #3
add J3, OUT, J3, lsl #3
vstr s8, [J0]
vstr s9, [J1]
vstr s10, [J2]
vstr s11, [J3]
vstr s12, [J0, #4]
vstr s13, [J1, #4]
vstr s14, [J2, #4]
vstr s15, [J3, #4]
.endm
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
.macro postrotation_innerloop tail, head
.set trig_lo_head, n8 - k - 2
.set trig_hi_head, n8 + k
.set out_lo_head, trig_lo_head * 2
.set out_hi_head, trig_hi_head * 2
.set trig_lo_tail, n8 - (k - 2) - 2
.set trig_hi_tail, n8 + (k - 2)
.set out_lo_tail, trig_lo_tail * 2
.set out_hi_tail, trig_hi_tail * 2
.if (k & 2) == 0
TCOS_D0_HEAD .req d10 @ s20,s21
TCOS_D1_HEAD .req d11 @ s22,s23
TCOS_S0_TAIL .req s24
.else
TCOS_D0_HEAD .req d12 @ s24,s25
TCOS_D1_HEAD .req d13 @ s26,s27
TCOS_S0_TAIL .req s20
.endif
.ifnc "\tail",""
vmls.f s8, s0, TCOS_S0_TAIL @ vector operation
.endif
.ifnc "\head",""
vldr d8, [TSIN, #trig_lo_head*4] @ s16,s17
vldr d9, [TSIN, #trig_hi_head*4] @ s18,s19
vldr TCOS_D0_HEAD, [TCOS, #trig_lo_head*4]
.endif
.ifnc "\tail",""
vmla.f s12, s4, TCOS_S0_TAIL @ vector operation
.endif
.ifnc "\head",""
vldr s0, [OUT, #out_lo_head*4]
vldr s1, [OUT, #out_lo_head*4 + 8]
vldr s2, [OUT, #out_hi_head*4]
vldr s3, [OUT, #out_hi_head*4 + 8]
vldr s4, [OUT, #out_lo_head*4 + 4]
vldr s5, [OUT, #out_lo_head*4 + 12]
vldr s6, [OUT, #out_hi_head*4 + 4]
vldr s7, [OUT, #out_hi_head*4 + 12]
.endif
.ifnc "\tail",""
vstr s8, [OUT, #out_lo_tail*4]
vstr s9, [OUT, #out_lo_tail*4 + 8]
vstr s10, [OUT, #out_hi_tail*4]
vstr s11, [OUT, #out_hi_tail*4 + 8]
.endif
.ifnc "\head",""
vmul.f s8, s4, s16 @ vector operation
.endif
.ifnc "\tail",""
vstr s12, [OUT, #out_hi_tail*4 + 12]
vstr s13, [OUT, #out_hi_tail*4 + 4]
vstr s14, [OUT, #out_lo_tail*4 + 12]
vstr s15, [OUT, #out_lo_tail*4 + 4]
.endif
.ifnc "\head",""
vmul.f s12, s0, s16 @ vector operation
vldr TCOS_D1_HEAD, [TCOS, #trig_hi_head*4]
.endif
.unreq TCOS_D0_HEAD
.unreq TCOS_D1_HEAD
.unreq TCOS_S0_TAIL
.ifnc "\head",""
.set k, k + 2
.endif
.endm
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
.macro postrotation_innerloop_rolled tail, head, tcos_s0_head, tcos_s1_head, tcos_s2_head, tcos_s3_head, tcos_s0_tail, out_offset_head, out_offset_tail
.ifnc "\tail",""
vmls.f s8, s0, \tcos_s0_tail @ vector operation
.endif
.ifnc "\head",""
vldmia TSIN!, {s16,s17}
vldmdb TSIN_HI!, {s18,s19}
vldmia TCOS!, {\tcos_s0_head,\tcos_s1_head}
.endif
.ifnc "\tail",""
vmla.f s12, s4, \tcos_s0_tail @ vector operation
.endif
.ifnc "\head",""
vldr s0, [OUT, #+\out_offset_head+0]
vldr s1, [OUT, #+\out_offset_head+8]
vldr s2, [OUT_HI, #-\out_offset_head-16]
vldr s3, [OUT_HI, #-\out_offset_head-8]
vldr s4, [OUT, #+\out_offset_head+4]
vldr s5, [OUT, #+\out_offset_head+12]
vldr s6, [OUT_HI, #-\out_offset_head-12]
vldr s7, [OUT_HI, #-\out_offset_head-4]
.endif
.ifnc "\tail",""
vstr s8, [OUT, #+\out_offset_tail+0]
vstr s9, [OUT, #+\out_offset_tail+8]
vstr s10, [OUT_HI, #-\out_offset_tail-16]
vstr s11, [OUT_HI, #-\out_offset_tail-8]
.endif
.ifnc "\head",""
vmul.f s8, s4, s16 @ vector operation
.endif
.ifnc "\tail",""
vstr s12, [OUT_HI, #-\out_offset_tail-4]
vstr s13, [OUT_HI, #-\out_offset_tail-12]
vstr s14, [OUT, #+\out_offset_tail+12]
vstr s15, [OUT, #+\out_offset_tail+4]
.endif
.ifnc "\head",""
vmul.f s12, s0, s16 @ vector operation
vldmdb TCOS_HI!, {\tcos_s2_head,\tcos_s3_head}
.endif
.endm
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
/* void ff_imdct_half_vfp(FFTContext *s,
* FFTSample *output,
* const FFTSample *input)
*/
function ff_imdct_half_vfp, export=1
ldr ip, [CONTEXT, #5*4] @ mdct_bits
teq ip, #6
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
bne 10f
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
.set n, 1<<6
.set n2, n/2
.set n4, n/4
.set n8, n/8
push {v1-v5,lr}
vpush {s16-s27}
fmrx OLDFPSCR, FPSCR
ldr lr, =0x03030000 @ RunFast mode, short vectors of length 4, stride 1
fmxr FPSCR, lr
mov OUT, ORIGOUT
ldr REVTAB, [CONTEXT, #2*4]
ldr TCOS, [CONTEXT, #6*4]
ldr TSIN, [CONTEXT, #7*4]
.set k, 0
.rept n8/2
prerotation_innerloop
.endr
fmxr FPSCR, OLDFPSCR
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
mov a1, OUT
arm: Mangle external symbols properly in new vfp assembly files Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-22 09:24:43 +00:00
bl X(ff_fft16_vfp)
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
ldr lr, =0x03030000 @ RunFast mode, short vectors of length 4, stride 1
fmxr FPSCR, lr
.set k, 0
postrotation_innerloop , head
.rept n8/2 - 1
postrotation_innerloop tail, head
.endr
postrotation_innerloop tail
fmxr FPSCR, OLDFPSCR
vpop {s16-s27}
pop {v1-v5,pc}
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
10:
push {v1-v6,sl,fp,lr}
vpush {s16-s27}
fmrx OLDFPSCR, FPSCR
ldr lr, =0x03030000 @ RunFast mode, short vectors of length 4, stride 1
fmxr FPSCR, lr
mov lr, #1
mov OUT, ORIGOUT
ldr REVTAB, [CONTEXT, #2*4]
ldr TCOS, [CONTEXT, #6*4]
ldr TSIN, [CONTEXT, #7*4]
mov lr, lr, lsl ip
push {CONTEXT,OLDFPSCR}
add IN_HI, IN, lr, lsl #1
add REVTAB_HI, REVTAB, lr, lsr #1
add TCOS_HI, TCOS, lr
add TSIN_HI, TSIN, lr
0: prerotation_innerloop_rolled
teq IN, IN_HI
bne 0b
ldmia sp, {CONTEXT,OLDFPSCR}
mov ORIGOUT, OUT
fmxr FPSCR, OLDFPSCR
ldr ip, [CONTEXT, #9*4]
blx ip @ s->fft_calc(s, output)
pop {CONTEXT,OLDFPSCR}
ldr lr, =0x03030000 @ RunFast mode, short vectors of length 4, stride 1
ldr ip, [CONTEXT, #5*4] @ mdct_bits
fmxr FPSCR, lr
mov lr, #1
mov lr, lr, lsl ip
sub TCOS, TCOS, lr, lsr #1
sub TSIN, TSIN, lr, lsr #1
add OUT_HI, OUT, lr, lsl #1
add TCOS_HI, TCOS, lr
add TSIN_HI, TSIN, lr
postrotation_innerloop_rolled , head, s20, s21, s22, s23,, 0
b 1f
0: add OUT, OUT, #32
sub OUT_HI, OUT_HI, #32
postrotation_innerloop_rolled tail, head, s20, s21, s22, s23, s24, 0, -16
1: postrotation_innerloop_rolled tail, head, s24, s25, s26, s27, s20, 16, 0
teq TSIN, TSIN_HI
bne 0b
postrotation_innerloop_rolled tail,,,,,, s24,, 16
fmxr FPSCR, OLDFPSCR
vpop {s16-s27}
pop {v1-v6,sl,fp,pc}
arm: Add VFP-accelerated version of imdct_half Before After Mean StdDev Mean StdDev Change This function 2653.0 28.5 1108.8 51.4 +139.3% Overall 17049.5 408.2 15973.0 223.2 +6.7% Signed-off-by: Martin Storsjö <martin@martin.st>
2013-07-19 07:59:17 +00:00
endfunc
.unreq CONTEXT
.unreq ORIGOUT
.unreq IN
.unreq OUT
.unreq REVTAB
.unreq TCOS
.unreq TSIN
.unreq OLDFPSCR
.unreq J0
.unreq J1
.unreq J2
.unreq J3
armv6: Accelerate ff_imdct_half for general case (mdct_bits != 6) The previous implementation targeted DTS Coherent Acoustics, which only requires mdct_bits == 6. This relatively small size lent itself to unrolling the loops a small number of times, and encoding offsets calculated at assembly time within the load/store instructions of each iteration. In the more general case (codecs such as AAC and AC3) much larger arrays are used - mdct_bits == [8, 9, 11]. The old method does not scale for these cases, so more integer registers are used with non-unrolled versions of the loops (and with some stack spillage). The postrotation filter loop is still unrolled by a factor of 2 to permit the double-buffering of some VFP registers to facilitate overlap of neighbouring iterations. 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 ff_imdct_half_c / ff_imdct_half_vfp, for the same example AAC stream: Before After Mean StdDev Mean StdDev Confidence Change aac_decode_frame 2368.1 35.8 2117.2 35.3 100.0% +11.8% ff_imdct_half_* 457.5 22.4 251.2 16.2 100.0% +82.1% Signed-off-by: Martin Storsjö <martin@martin.st>
2014-07-10 23:12:31 +00:00
.unreq REVTAB_HI
.unreq IN_HI
.unreq OUT_HI
.unreq TCOS_HI
.unreq TSIN_HI
Reference in New Issue Copy Permalink