1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-11-21 10:55:51 +02:00
FFmpeg/libavcodec/arm/vc1dsp_neon.S

1925 lines
91 KiB
ArmAsm
Raw Normal View History

/*
* VC1 NEON optimisations
*
* Copyright (c) 2010 Rob Clark <rob@ti.com>
* Copyright (c) 2011 Mans Rullgard <mans@mansr.com>
*
* This file is part of FFmpeg.
*
* FFmpeg 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.
*
* FFmpeg 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 FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/arm/asm.S"
#include "neon.S"
#include "config.h"
@ Transpose rows into columns of a matrix of 16-bit elements. For 4x4, pass
@ double-word registers, for 8x4, pass quad-word registers.
.macro transpose16 r0, r1, r2, r3
@ At this point:
@ row[0] r0
@ row[1] r1
@ row[2] r2
@ row[3] r3
vtrn.16 \r0, \r1 @ first and second row
vtrn.16 \r2, \r3 @ third and fourth row
vtrn.32 \r0, \r2 @ first and third row
vtrn.32 \r1, \r3 @ second and fourth row
@ At this point, if registers are quad-word:
@ column[0] d0
@ column[1] d2
@ column[2] d4
@ column[3] d6
@ column[4] d1
@ column[5] d3
@ column[6] d5
@ column[7] d7
@ At this point, if registers are double-word:
@ column[0] d0
@ column[1] d1
@ column[2] d2
@ column[3] d3
.endm
@ ff_vc1_inv_trans_{4,8}x{4,8}_neon and overflow: The input values in the file
@ are supposed to be in a specific range as to allow for 16-bit math without
@ causing overflows, but sometimes the input values are just big enough to
@ barely cause overflow in vadd instructions like:
@
@ vadd.i16 q0, q8, q10
@ vshr.s16 q0, q0, #\rshift
@
@ To prevent these borderline cases from overflowing, we just need one more
@ bit of precision, which is accomplished by replacing the sequence above with:
@
@ vhadd.s16 q0, q8, q10
@ vshr.s16 q0, q0, #(\rshift -1)
@
@ This works because vhadd is a single instruction that adds, then shifts to
@ the right once, all before writing the result to the destination register.
@
@ Even with this workaround, there were still some files that caused overflows
@ in ff_vc1_inv_trans_8x8_neon. See the comments in ff_vc1_inv_trans_8x8_neon
@ for the additional workaround.
@ Takes 4 columns of 8 values each and operates on it. Modeled after the first
@ for loop in vc1_inv_trans_4x8_c.
@ Input columns: q0 q1 q2 q3
@ Output columns: q0 q1 q2 q3
@ Trashes: r12 q8 q9 q10 q11 q12 q13
.macro vc1_inv_trans_4x8_helper add rshift
@ Compute temp1, temp2 and setup scalar #17, #22, #10
vadd.i16 q12, q0, q2 @ temp1 = src[0] + src[2]
movw r12, #17
vsub.i16 q13, q0, q2 @ temp2 = src[0] - src[2]
movt r12, #22
vmov.32 d0[0], r12
movw r12, #10
vmov.16 d1[0], r12
vmov.i16 q8, #\add @ t1 will accumulate here
vmov.i16 q9, #\add @ t2 will accumulate here
vmul.i16 q10, q1, d0[1] @ t3 = 22 * (src[1])
vmul.i16 q11, q3, d0[1] @ t4 = 22 * (src[3])
vmla.i16 q8, q12, d0[0] @ t1 = 17 * (temp1) + 4
vmla.i16 q9, q13, d0[0] @ t2 = 17 * (temp2) + 4
vmla.i16 q10, q3, d1[0] @ t3 += 10 * src[3]
vmls.i16 q11, q1, d1[0] @ t4 -= 10 * src[1]
vhadd.s16 q0, q8, q10 @ dst[0] = (t1 + t3) >> 1
vhsub.s16 q3, q8, q10 @ dst[3] = (t1 - t3) >> 1
vhsub.s16 q1, q9, q11 @ dst[1] = (t2 - t4) >> 1
vhadd.s16 q2, q9, q11 @ dst[2] = (t2 + t4) >> 1
@ Halving add/sub above already did one shift
vshr.s16 q0, q0, #(\rshift - 1) @ dst[0] >>= (rshift - 1)
vshr.s16 q3, q3, #(\rshift - 1) @ dst[3] >>= (rshift - 1)
vshr.s16 q1, q1, #(\rshift - 1) @ dst[1] >>= (rshift - 1)
vshr.s16 q2, q2, #(\rshift - 1) @ dst[2] >>= (rshift - 1)
.endm
@ Takes 8 columns of 4 values each and operates on it. Modeled after the second
@ for loop in vc1_inv_trans_4x8_c.
@ Input columns: d0 d2 d4 d6 d1 d3 d5 d7
@ Output columns: d16 d17 d18 d19 d21 d20 d23 d22
@ Trashes all NEON registers (and r12) except for: q4 q5 q6 q7
.macro vc1_inv_trans_8x4_helper add add1beforeshift rshift
@ At this point:
@ src[0] d0 overwritten later
@ src[8] d2
@ src[16] d4 overwritten later
@ src[24] d6
@ src[32] d1 overwritten later
@ src[40] d3
@ src[48] d5 overwritten later
@ src[56] d7
movw r12, #12
vmov.i16 q14, #\add @ t1|t2 will accumulate here
movt r12, #6
vadd.i16 d20, d0, d1 @ temp1 = src[0] + src[32]
vsub.i16 d21, d0, d1 @ temp2 = src[0] - src[32]
vmov.i32 d0[0], r12 @ 16-bit: d0[0] = #12, d0[1] = #6
vshl.i16 q15, q2, #4 @ t3|t4 = 16 * (src[16]|src[48])
vswp d4, d5 @ q2 = src[48]|src[16]
vmla.i16 q14, q10, d0[0] @ t1|t2 = 12 * (temp1|temp2) + 64
movw r12, #15
movt r12, #9
vmov.i32 d0[1], r12 @ 16-bit: d0[2] = #15, d0[3] = #9
vneg.s16 d31, d31 @ t4 = -t4
vmla.i16 q15, q2, d0[1] @ t3|t4 += 6 * (src[48]|src[16])
@ At this point:
@ d0[2] #15
@ d0[3] #9
@ q1 src[8]|src[40]
@ q3 src[24]|src[56]
@ q14 old t1|t2
@ q15 old t3|t4
vshl.i16 q8, q1, #4 @ t1|t2 = 16 * (src[8]|src[40])
vswp d2, d3 @ q1 = src[40]|src[8]
vshl.i16 q12, q3, #4 @ temp3a|temp4a = 16 * src[24]|src[56]
vswp d6, d7 @ q3 = src[56]|src[24]
vshl.i16 q13, q1, #2 @ temp3b|temp4b = 4 * (src[40]|src[8])
vshl.i16 q2, q3, #2 @ temp1|temp2 = 4 * (src[56]|src[24])
vswp d3, d6 @ q1 = src[40]|src[56], q3 = src[8]|src[24]
vsub.i16 q9, q13, q12 @ t3|t4 = - (temp3a|temp4a) + (temp3b|temp4b)
vadd.i16 q8, q8, q2 @ t1|t2 += temp1|temp2
vmul.i16 q12, q3, d0[3] @ temp3|temp4 = 9 * src[8]|src[24]
vmla.i16 q8, q1, d0[3] @ t1|t2 += 9 * (src[40]|src[56])
vswp d6, d7 @ q3 = src[24]|src[8]
vswp d2, d3 @ q1 = src[56]|src[40]
vsub.i16 q11, q14, q15 @ t8|t7 = old t1|t2 - old t3|t4
vadd.i16 q10, q14, q15 @ t5|t6 = old t1|t2 + old t3|t4
.if \add1beforeshift
vmov.i16 q15, #1
.endif
vadd.i16 d18, d18, d24 @ t3 += temp3
vsub.i16 d19, d19, d25 @ t4 -= temp4
vswp d22, d23 @ q11 = t7|t8
vneg.s16 d17, d17 @ t2 = -t2
vmla.i16 q9, q1, d0[2] @ t3|t4 += 15 * src[56]|src[40]
vmla.i16 q8, q3, d0[2] @ t1|t2 += 15 * src[24]|src[8]
@ At this point:
@ t1 d16
@ t2 d17
@ t3 d18
@ t4 d19
@ t5 d20
@ t6 d21
@ t7 d22
@ t8 d23
@ #1 q15
.if \add1beforeshift
vadd.i16 q3, q15, q10 @ line[7,6] = t5|t6 + 1
vadd.i16 q2, q15, q11 @ line[5,4] = t7|t8 + 1
.endif
@ Sometimes this overflows, so to get one additional bit of precision, use
@ a single instruction that both adds and shifts right (halving).
vhadd.s16 q1, q9, q11 @ line[2,3] = (t3|t4 + t7|t8) >> 1
vhadd.s16 q0, q8, q10 @ line[0,1] = (t1|t2 + t5|t6) >> 1
.if \add1beforeshift
vhsub.s16 q2, q2, q9 @ line[5,4] = (t7|t8 - t3|t4 + 1) >> 1
vhsub.s16 q3, q3, q8 @ line[7,6] = (t5|t6 - t1|t2 + 1) >> 1
.else
vhsub.s16 q2, q11, q9 @ line[5,4] = (t7|t8 - t3|t4) >> 1
vhsub.s16 q3, q10, q8 @ line[7,6] = (t5|t6 - t1|t2) >> 1
.endif
vshr.s16 q9, q1, #(\rshift - 1) @ one shift is already done by vhadd/vhsub above
vshr.s16 q8, q0, #(\rshift - 1)
vshr.s16 q10, q2, #(\rshift - 1)
vshr.s16 q11, q3, #(\rshift - 1)
@ At this point:
@ dst[0] d16
@ dst[1] d17
@ dst[2] d18
@ dst[3] d19
@ dst[4] d21
@ dst[5] d20
@ dst[6] d23
@ dst[7] d22
.endm
@ This is modeled after the first and second for loop in vc1_inv_trans_8x8_c.
@ Input columns: q8, q9, q10, q11, q12, q13, q14, q15
@ Output columns: q8, q9, q10, q11, q12, q13, q14, q15
@ Trashes all NEON registers (and r12) except for: q4 q5 q6 q7
.macro vc1_inv_trans_8x8_helper add add1beforeshift rshift
@ This actually computes half of t1, t2, t3, t4, as explained below
@ near `tNhalf`.
vmov.i16 q0, #(6 / 2) @ q0 = #6/2
vshl.i16 q1, q10, #3 @ t3 = 16/2 * src[16]
vshl.i16 q3, q14, #3 @ temp4 = 16/2 * src[48]
vmul.i16 q2, q10, q0 @ t4 = 6/2 * src[16]
vmla.i16 q1, q14, q0 @ t3 += 6/2 * src[48]
@ unused: q0, q10, q14
vmov.i16 q0, #(12 / 2) @ q0 = #12/2
vadd.i16 q10, q8, q12 @ temp1 = src[0] + src[32]
vsub.i16 q14, q8, q12 @ temp2 = src[0] - src[32]
@ unused: q8, q12
vmov.i16 q8, #(\add / 2) @ t1 will accumulate here
vmov.i16 q12, #(\add / 2) @ t2 will accumulate here
movw r12, #15
vsub.i16 q2, q2, q3 @ t4 = 6/2 * src[16] - 16/2 * src[48]
movt r12, #9
@ unused: q3
vmla.i16 q8, q10, q0 @ t1 = 12/2 * temp1 + add
vmla.i16 q12, q14, q0 @ t2 = 12/2 * temp2 + add
vmov.i32 d0[0], r12
@ unused: q3, q10, q14
@ At this point:
@ q0 d0=#15|#9
@ q1 old t3
@ q2 old t4
@ q3
@ q8 old t1
@ q9 src[8]
@ q10
@ q11 src[24]
@ q12 old t2
@ q13 src[40]
@ q14
@ q15 src[56]
@ unused: q3, q10, q14
movw r12, #16
vshl.i16 q3, q9, #4 @ t1 = 16 * src[8]
movt r12, #4
vshl.i16 q10, q9, #2 @ t4 = 4 * src[8]
vmov.i32 d1[0], r12
vmul.i16 q14, q9, d0[0] @ t2 = 15 * src[8]
vmul.i16 q9, q9, d0[1] @ t3 = 9 * src[8]
@ unused: none
vmla.i16 q3, q11, d0[0] @ t1 += 15 * src[24]
vmls.i16 q10, q11, d0[1] @ t4 -= 9 * src[24]
vmls.i16 q14, q11, d1[1] @ t2 -= 4 * src[24]
vmls.i16 q9, q11, d1[0] @ t3 -= 16 * src[24]
@ unused: q11
vmla.i16 q3, q13, d0[1] @ t1 += 9 * src[40]
vmla.i16 q10, q13, d0[0] @ t4 += 15 * src[40]
vmls.i16 q14, q13, d1[0] @ t2 -= 16 * src[40]
vmla.i16 q9, q13, d1[1] @ t3 += 4 * src[40]
@ unused: q11, q13
@ Compute t5, t6, t7, t8 from old t1, t2, t3, t4. Actually, it computes
@ half of t5, t6, t7, t8 since t1, t2, t3, t4 are halved.
vadd.i16 q11, q8, q1 @ t5 = t1 + t3
vsub.i16 q1, q8, q1 @ t8 = t1 - t3
vadd.i16 q13, q12, q2 @ t6 = t2 + t4
vsub.i16 q2, q12, q2 @ t7 = t2 - t4
@ unused: q8, q12
.if \add1beforeshift
vmov.i16 q12, #1
.endif
@ unused: q8
vmla.i16 q3, q15, d1[1] @ t1 += 4 * src[56]
vmls.i16 q14, q15, d0[1] @ t2 -= 9 * src[56]
vmla.i16 q9, q15, d0[0] @ t3 += 15 * src[56]
vmls.i16 q10, q15, d1[0] @ t4 -= 16 * src[56]
@ unused: q0, q8, q15
@ At this point:
@ t1 q3
@ t2 q14
@ t3 q9
@ t4 q10
@ t5half q11
@ t6half q13
@ t7half q2
@ t8half q1
@ #1 q12
@
@ tNhalf is half of the value of tN (as described in vc1_inv_trans_8x8_c).
@ This is done because sometimes files have input that causes tN + tM to
@ overflow. To avoid this overflow, we compute tNhalf, then compute
@ tNhalf + tM (which doesn't overflow), and then we use vhadd to compute
@ (tNhalf + (tNhalf + tM)) >> 1 which does not overflow because it is
@ one instruction.
@ For each pair of tN and tM, do:
@ lineA = t5half + t1
@ if add1beforeshift: t1 -= 1
@ lineA = (t5half + lineA) >> 1
@ lineB = t5half - t1
@ lineB = (t5half + lineB) >> 1
@ lineA >>= rshift - 1
@ lineB >>= rshift - 1
vadd.i16 q8, q11, q3 @ q8 = t5half + t1
.if \add1beforeshift
vsub.i16 q3, q3, q12 @ q3 = t1 - 1
.endif
vadd.i16 q0, q13, q14 @ q0 = t6half + t2
.if \add1beforeshift
vsub.i16 q14, q14, q12 @ q14 = t2 - 1
.endif
vadd.i16 q15, q2, q9 @ q15 = t7half + t3
.if \add1beforeshift
vsub.i16 q9, q9, q12 @ q9 = t3 - 1
.endif
@ unused: none
vhadd.s16 q8, q11, q8 @ q8 = (t5half + t5half + t1) >> 1
vsub.i16 q3, q11, q3 @ q3 = t5half - t1 + 1
vhadd.s16 q0, q13, q0 @ q0 = (t6half + t6half + t2) >> 1
vsub.i16 q14, q13, q14 @ q14 = t6half - t2 + 1
vhadd.s16 q15, q2, q15 @ q15 = (t7half + t7half + t3) >> 1
vsub.i16 q9, q2, q9 @ q9 = t7half - t3 + 1
vhadd.s16 q3, q11, q3 @ q3 = (t5half + t5half - t1 + 1) >> 1
@ unused: q11
vadd.i16 q11, q1, q10 @ q11 = t8half + t4
.if \add1beforeshift
vsub.i16 q10, q10, q12 @ q10 = t4 - 1
.endif
@ unused: q12
vhadd.s16 q14, q13, q14 @ q14 = (t6half + t6half - t2 + 1) >> 1
@ unused: q12, q13
vhadd.s16 q13, q2, q9 @ q9 = (t7half + t7half - t3 + 1) >> 1
@ unused: q12, q2, q9
vsub.i16 q10, q1, q10 @ q10 = t8half - t4 + 1
vhadd.s16 q11, q1, q11 @ q11 = (t8half + t8half + t4) >> 1
vshr.s16 q8, q8, #(\rshift - 1) @ q8 = line[0]
vhadd.s16 q12, q1, q10 @ q12 = (t8half + t8half - t4 + 1) >> 1
vshr.s16 q9, q0, #(\rshift - 1) @ q9 = line[1]
vshr.s16 q10, q15, #(\rshift - 1) @ q10 = line[2]
vshr.s16 q11, q11, #(\rshift - 1) @ q11 = line[3]
vshr.s16 q12, q12, #(\rshift - 1) @ q12 = line[4]
vshr.s16 q13, q13, #(\rshift - 1) @ q13 = line[5]
vshr.s16 q14, q14, #(\rshift - 1) @ q14 = line[6]
vshr.s16 q15, q3, #(\rshift - 1) @ q15 = line[7]
.endm
@ (int16_t *block [r0])
function ff_vc1_inv_trans_8x8_neon, export=1
vld1.64 {q8-q9}, [r0,:128]!
vld1.64 {q10-q11}, [r0,:128]!
vld1.64 {q12-q13}, [r0,:128]!
vld1.64 {q14-q15}, [r0,:128]
sub r0, r0, #(16 * 2 * 3) @ restore r0
@ At this point:
@ src[0] q8
@ src[8] q9
@ src[16] q10
@ src[24] q11
@ src[32] q12
@ src[40] q13
@ src[48] q14
@ src[56] q15
vc1_inv_trans_8x8_helper add=4, add1beforeshift=0, rshift=3
@ Transpose result matrix of 8x8
swap4 d17, d19, d21, d23, d24, d26, d28, d30
transpose16_4x4 q8, q9, q10, q11, q12, q13, q14, q15
vc1_inv_trans_8x8_helper add=64, add1beforeshift=1, rshift=7
vst1.64 {q8-q9}, [r0,:128]!
vst1.64 {q10-q11}, [r0,:128]!
vst1.64 {q12-q13}, [r0,:128]!
vst1.64 {q14-q15}, [r0,:128]
bx lr
endfunc
@ (uint8_t *dest [r0], ptrdiff_t stride [r1], int16_t *block [r2])
function ff_vc1_inv_trans_8x4_neon, export=1
vld1.64 {q0-q1}, [r2,:128]! @ load 8 * 4 * 2 = 64 bytes / 16 bytes per quad = 4 quad registers
vld1.64 {q2-q3}, [r2,:128]
transpose16 q0, q1, q2, q3 @ transpose rows to columns
@ At this point:
@ src[0] d0
@ src[1] d2
@ src[2] d4
@ src[3] d6
@ src[4] d1
@ src[5] d3
@ src[6] d5
@ src[7] d7
vc1_inv_trans_8x4_helper add=4, add1beforeshift=0, rshift=3
@ Move output to more standardized registers
vmov d0, d16
vmov d2, d17
vmov d4, d18
vmov d6, d19
vmov d1, d21
vmov d3, d20
vmov d5, d23
vmov d7, d22
@ At this point:
@ dst[0] d0
@ dst[1] d2
@ dst[2] d4
@ dst[3] d6
@ dst[4] d1
@ dst[5] d3
@ dst[6] d5
@ dst[7] d7
transpose16 q0, q1, q2, q3 @ turn columns into rows
@ At this point:
@ row[0] q0
@ row[1] q1
@ row[2] q2
@ row[3] q3
vc1_inv_trans_4x8_helper add=64, rshift=7
@ At this point:
@ line[0].l d0
@ line[0].h d1
@ line[1].l d2
@ line[1].h d3
@ line[2].l d4
@ line[2].h d5
@ line[3].l d6
@ line[3].h d7
@ unused registers: q12, q13, q14, q15
vld1.64 {d28}, [r0,:64], r1 @ read dest
vld1.64 {d29}, [r0,:64], r1
vld1.64 {d30}, [r0,:64], r1
vld1.64 {d31}, [r0,:64], r1
sub r0, r0, r1, lsl #2 @ restore original r0 value
vaddw.u8 q0, q0, d28 @ line[0] += dest[0]
vaddw.u8 q1, q1, d29 @ line[1] += dest[1]
vaddw.u8 q2, q2, d30 @ line[2] += dest[2]
vaddw.u8 q3, q3, d31 @ line[3] += dest[3]
vqmovun.s16 d0, q0 @ line[0]
vqmovun.s16 d1, q1 @ line[1]
vqmovun.s16 d2, q2 @ line[2]
vqmovun.s16 d3, q3 @ line[3]
vst1.64 {d0}, [r0,:64], r1 @ write dest
vst1.64 {d1}, [r0,:64], r1
vst1.64 {d2}, [r0,:64], r1
vst1.64 {d3}, [r0,:64]
bx lr
endfunc
@ (uint8_t *dest [r0], ptrdiff_t stride [r1], int16_t *block [r2])
function ff_vc1_inv_trans_4x8_neon, export=1
mov r12, #(8 * 2) @ 8 elements per line, each element 2 bytes
vld4.16 {d0[], d2[], d4[], d6[]}, [r2,:64], r12 @ read each column into a q register
vld4.16 {d0[1], d2[1], d4[1], d6[1]}, [r2,:64], r12
vld4.16 {d0[2], d2[2], d4[2], d6[2]}, [r2,:64], r12
vld4.16 {d0[3], d2[3], d4[3], d6[3]}, [r2,:64], r12
vld4.16 {d1[], d3[], d5[], d7[]}, [r2,:64], r12
vld4.16 {d1[1], d3[1], d5[1], d7[1]}, [r2,:64], r12
vld4.16 {d1[2], d3[2], d5[2], d7[2]}, [r2,:64], r12
vld4.16 {d1[3], d3[3], d5[3], d7[3]}, [r2,:64]
vc1_inv_trans_4x8_helper add=4, rshift=3
@ At this point:
@ dst[0] = q0
@ dst[1] = q1
@ dst[2] = q2
@ dst[3] = q3
transpose16 q0, q1, q2, q3 @ Transpose rows (registers) into columns
vc1_inv_trans_8x4_helper add=64, add1beforeshift=1, rshift=7
vld1.32 {d28[]}, [r0,:32], r1 @ read dest
vld1.32 {d28[1]}, [r0,:32], r1
vld1.32 {d29[]}, [r0,:32], r1
vld1.32 {d29[1]}, [r0,:32], r1
vld1.32 {d30[]}, [r0,:32], r1
vld1.32 {d30[0]}, [r0,:32], r1
vld1.32 {d31[]}, [r0,:32], r1
vld1.32 {d31[0]}, [r0,:32], r1
sub r0, r0, r1, lsl #3 @ restore original r0 value
vaddw.u8 q8, q8, d28 @ line[0,1] += dest[0,1]
vaddw.u8 q9, q9, d29 @ line[2,3] += dest[2,3]
vaddw.u8 q10, q10, d30 @ line[5,4] += dest[5,4]
vaddw.u8 q11, q11, d31 @ line[7,6] += dest[7,6]
vqmovun.s16 d16, q8 @ clip(line[0,1])
vqmovun.s16 d18, q9 @ clip(line[2,3])
vqmovun.s16 d20, q10 @ clip(line[5,4])
vqmovun.s16 d22, q11 @ clip(line[7,6])
vst1.32 {d16[0]}, [r0,:32], r1 @ write dest
vst1.32 {d16[1]}, [r0,:32], r1
vst1.32 {d18[0]}, [r0,:32], r1
vst1.32 {d18[1]}, [r0,:32], r1
vst1.32 {d20[1]}, [r0,:32], r1
vst1.32 {d20[0]}, [r0,:32], r1
vst1.32 {d22[1]}, [r0,:32], r1
vst1.32 {d22[0]}, [r0,:32]
bx lr
endfunc
@ Setup constants in registers which are used by vc1_inv_trans_4x4_helper
.macro vc1_inv_trans_4x4_helper_setup
vmov.i16 q13, #17
vmov.i16 q14, #22
vmov.i16 d30, #10 @ only need double-word, not quad-word
.endm
@ This is modeled after the first for loop in vc1_inv_trans_4x4_c.
.macro vc1_inv_trans_4x4_helper add rshift
vmov.i16 q2, #\add @ t1|t2 will accumulate here
vadd.i16 d16, d0, d1 @ temp1 = src[0] + src[2]
vsub.i16 d17, d0, d1 @ temp2 = src[0] - src[2]
vmul.i16 q3, q14, q1 @ t3|t4 = 22 * (src[1]|src[3])
vmla.i16 q2, q13, q8 @ t1|t2 = 17 * (temp1|temp2) + add
vmla.i16 d6, d30, d3 @ t3 += 10 * src[3]
vmls.i16 d7, d30, d2 @ t4 -= 10 * src[1]
vadd.i16 q0, q2, q3 @ dst[0,2] = (t1|t2 + t3|t4)
vsub.i16 q1, q2, q3 @ dst[3,1] = (t1|t2 - t3|t4)
vshr.s16 q0, q0, #\rshift @ dst[0,2] >>= rshift
vshr.s16 q1, q1, #\rshift @ dst[3,1] >>= rshift
.endm
@ (uint8_t *dest [r0], ptrdiff_t stride [r1], int16_t *block [r2])
function ff_vc1_inv_trans_4x4_neon, export=1
mov r12, #(8 * 2) @ 8 elements per line, each element 2 bytes
vld4.16 {d0[], d1[], d2[], d3[]}, [r2,:64], r12 @ read each column into a register
vld4.16 {d0[1], d1[1], d2[1], d3[1]}, [r2,:64], r12
vld4.16 {d0[2], d1[2], d2[2], d3[2]}, [r2,:64], r12
vld4.16 {d0[3], d1[3], d2[3], d3[3]}, [r2,:64]
vswp d1, d2 @ so that we can later access column 1 and column 3 as a single q1 register
vc1_inv_trans_4x4_helper_setup
@ At this point:
@ src[0] = d0
@ src[1] = d2
@ src[2] = d1
@ src[3] = d3
vc1_inv_trans_4x4_helper add=4, rshift=3 @ compute t1, t2, t3, t4 and combine them into dst[0-3]
@ At this point:
@ dst[0] = d0
@ dst[1] = d3
@ dst[2] = d1
@ dst[3] = d2
transpose16 d0, d3, d1, d2 @ Transpose rows (registers) into columns
@ At this point:
@ src[0] = d0
@ src[8] = d3
@ src[16] = d1
@ src[24] = d2
vswp d2, d3 @ so that we can later access column 1 and column 3 in order as a single q1 register
@ At this point:
@ src[0] = d0
@ src[8] = d2
@ src[16] = d1
@ src[24] = d3
vc1_inv_trans_4x4_helper add=64, rshift=7 @ compute t1, t2, t3, t4 and combine them into dst[0-3]
@ At this point:
@ line[0] = d0
@ line[1] = d3
@ line[2] = d1
@ line[3] = d2
vld1.32 {d18[]}, [r0,:32], r1 @ read dest
vld1.32 {d19[]}, [r0,:32], r1
vld1.32 {d18[1]}, [r0,:32], r1
vld1.32 {d19[0]}, [r0,:32], r1
sub r0, r0, r1, lsl #2 @ restore original r0 value
vaddw.u8 q0, q0, d18 @ line[0,2] += dest[0,2]
vaddw.u8 q1, q1, d19 @ line[3,1] += dest[3,1]
vqmovun.s16 d0, q0 @ clip(line[0,2])
vqmovun.s16 d1, q1 @ clip(line[3,1])
vst1.32 {d0[0]}, [r0,:32], r1 @ write dest
vst1.32 {d1[1]}, [r0,:32], r1
vst1.32 {d0[1]}, [r0,:32], r1
vst1.32 {d1[0]}, [r0,:32]
bx lr
endfunc
@ The absolute value of multiplication constants from vc1_mspel_filter and vc1_mspel_{ver,hor}_filter_16bits.
@ The sign is embedded in the code below that carries out the multiplication (mspel_filter{,.16}).
#define MSPEL_MODE_1_MUL_CONSTANTS 4, 53, 18, 3
#define MSPEL_MODE_2_MUL_CONSTANTS 1, 9, 9, 1
#define MSPEL_MODE_3_MUL_CONSTANTS 3, 18, 53, 4
@ These constants are from reading the source code of vc1_mspel_mc and determining the value that
@ is added to `rnd` to result in the variable `r`, and the value of the variable `shift`.
#define MSPEL_MODES_11_ADDSHIFT_CONSTANTS 15, 5
#define MSPEL_MODES_12_ADDSHIFT_CONSTANTS 3, 3
#define MSPEL_MODES_13_ADDSHIFT_CONSTANTS 15, 5
#define MSPEL_MODES_21_ADDSHIFT_CONSTANTS MSPEL_MODES_12_ADDSHIFT_CONSTANTS
#define MSPEL_MODES_22_ADDSHIFT_CONSTANTS 0, 1
#define MSPEL_MODES_23_ADDSHIFT_CONSTANTS 3, 3
#define MSPEL_MODES_31_ADDSHIFT_CONSTANTS MSPEL_MODES_13_ADDSHIFT_CONSTANTS
#define MSPEL_MODES_32_ADDSHIFT_CONSTANTS MSPEL_MODES_23_ADDSHIFT_CONSTANTS
#define MSPEL_MODES_33_ADDSHIFT_CONSTANTS 15, 5
@ The addition and shift constants from vc1_mspel_filter.
#define MSPEL_MODE_1_ADDSHIFT_CONSTANTS 32, 6
#define MSPEL_MODE_2_ADDSHIFT_CONSTANTS 8, 4
#define MSPEL_MODE_3_ADDSHIFT_CONSTANTS 32, 6
@ Setup constants in registers for a subsequent use of mspel_filter{,.16}.
.macro mspel_constants typesize reg_a reg_b reg_c reg_d filter_a filter_b filter_c filter_d reg_add filter_add_register
@ Typesize should be i8 or i16.
@ Only set the register if the value is not 1 and unique
.if \filter_a != 1
vmov.\typesize \reg_a, #\filter_a @ reg_a = filter_a
.endif
vmov.\typesize \reg_b, #\filter_b @ reg_b = filter_b
.if \filter_b != \filter_c
vmov.\typesize \reg_c, #\filter_c @ reg_c = filter_c
.endif
.if \filter_d != 1
vmov.\typesize \reg_d, #\filter_d @ reg_d = filter_d
.endif
@ vdup to double the size of typesize
.ifc \typesize,i8
vdup.16 \reg_add, \filter_add_register @ reg_add = filter_add_register
.else
vdup.32 \reg_add, \filter_add_register @ reg_add = filter_add_register
.endif
.endm
@ After mspel_constants has been used, do the filtering.
.macro mspel_filter acc dest src0 src1 src2 src3 filter_a filter_b filter_c filter_d reg_a reg_b reg_c reg_d reg_add filter_shift narrow=1
.if \filter_a != 1
@ If filter_a != 1, then we need a move and subtract instruction
vmov \acc, \reg_add @ acc = reg_add
vmlsl.u8 \acc, \reg_a, \src0 @ acc -= filter_a * src[-stride]
.else
@ If filter_a is 1, then just subtract without an extra move
vsubw.u8 \acc, \reg_add, \src0 @ acc = reg_add - src[-stride] @ since filter_a == 1
.endif
vmlal.u8 \acc, \reg_b, \src1 @ acc += filter_b * src[0]
.if \filter_b != \filter_c
vmlal.u8 \acc, \reg_c, \src2 @ acc += filter_c * src[stride]
.else
@ If filter_b is the same as filter_c, use the same reg_b register
vmlal.u8 \acc, \reg_b, \src2 @ acc += filter_c * src[stride] @ where filter_c == filter_b
.endif
.if \filter_d != 1
@ If filter_d != 1, then do a multiply accumulate
vmlsl.u8 \acc, \reg_d, \src3 @ acc -= filter_d * src[stride * 2]
.else
@ If filter_d is 1, then just do a subtract
vsubw.u8 \acc, \acc, \src3 @ acc -= src[stride * 2] @ since filter_d == 1
.endif
.if \narrow
vqshrun.s16 \dest, \acc, #\filter_shift @ dest = clip_uint8(acc >> filter_shift)
.else
vshr.s16 \dest, \acc, #\filter_shift @ dest = acc >> filter_shift
.endif
.endm
@ This is similar to mspel_filter, but the input is 16-bit instead of 8-bit and narrow=0 is not supported.
.macro mspel_filter.16 acc0 acc1 acc0_0 acc0_1 dest src0 src1 src2 src3 src4 src5 src6 src7 filter_a filter_b filter_c filter_d reg_a reg_b reg_c reg_d reg_add filter_shift
.if \filter_a != 1
vmov \acc0, \reg_add
vmov \acc1, \reg_add
vmlsl.s16 \acc0, \reg_a, \src0
vmlsl.s16 \acc1, \reg_a, \src1
.else
vsubw.s16 \acc0, \reg_add, \src0
vsubw.s16 \acc1, \reg_add, \src1
.endif
vmlal.s16 \acc0, \reg_b, \src2
vmlal.s16 \acc1, \reg_b, \src3
.if \filter_b != \filter_c
vmlal.s16 \acc0, \reg_c, \src4
vmlal.s16 \acc1, \reg_c, \src5
.else
vmlal.s16 \acc0, \reg_b, \src4
vmlal.s16 \acc1, \reg_b, \src5
.endif
.if \filter_d != 1
vmlsl.s16 \acc0, \reg_d, \src6
vmlsl.s16 \acc1, \reg_d, \src7
.else
vsubw.s16 \acc0, \acc0, \src6
vsubw.s16 \acc1, \acc1, \src7
.endif
@ Use acc0_0 and acc0_1 as temp space
vqshrun.s32 \acc0_0, \acc0, #\filter_shift @ Shift and narrow with saturation from s32 to u16
vqshrun.s32 \acc0_1, \acc1, #\filter_shift
vqmovn.u16 \dest, \acc0 @ Narrow with saturation from u16 to u8
.endm
@ Register usage for put_vc1_mspel_mc functions. Registers marked 'hv' are only used in put_vc1_mspel_mc_hv.
@
@ r0 adjusted dst
@ r1 adjusted src
@ r2 stride
@ r3 adjusted rnd
@ r4 [hv] tmp
@ r11 [hv] sp saved
@ r12 loop counter
@ d0 src[-stride]
@ d1 src[0]
@ d2 src[stride]
@ d3 src[stride * 2]
@ q0 [hv] src[-stride]
@ q1 [hv] src[0]
@ q2 [hv] src[stride]
@ q3 [hv] src[stride * 2]
@ d21 often result from mspel_filter
@ q11 accumulator 0
@ q12 [hv] accumulator 1
@ q13 accumulator initial value
@ d28 filter_a
@ d29 filter_b
@ d30 filter_c
@ d31 filter_d
@ (uint8_t *dst [r0], const uint8_t *src [r1], ptrdiff_t stride [r2], int rnd [r3])
.macro put_vc1_mspel_mc_hv hmode vmode filter_h_a filter_h_b filter_h_c filter_h_d filter_v_a filter_v_b filter_v_c filter_v_d filter_add filter_shift
function ff_put_vc1_mspel_mc\hmode\()\vmode\()_neon, export=1
push {r4, r11, lr}
mov r11, sp @ r11 = stack pointer before realignmnet
A bic sp, sp, #15 @ sp = round down to multiple of 16 bytes
T bic r4, r11, #15
T mov sp, r4
sub sp, sp, #(8*2*16) @ make space for 8 rows * 2 byte per element * 16 elements per row (to fit 11 actual elements per row)
mov r4, sp @ r4 = int16_t tmp[8 * 16]
sub r1, r1, #1 @ src -= 1
.if \filter_add != 0
add r3, r3, #\filter_add @ r3 = filter_add + rnd
.endif
mov r12, #8 @ loop counter
sub r1, r1, r2 @ r1 = &src[-stride] @ slide back
@ Do vertical filtering from src into tmp
mspel_constants i8, d28, d29, d30, d31, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, q13, r3
vld1.64 {d0,d1}, [r1], r2
vld1.64 {d2,d3}, [r1], r2
vld1.64 {d4,d5}, [r1], r2
1:
subs r12, r12, #4
vld1.64 {d6,d7}, [r1], r2
mspel_filter q11, q11, d0, d2, d4, d6, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
mspel_filter q12, q12, d1, d3, d5, d7, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
vst1.64 {q11,q12}, [r4,:128]! @ store and increment
vld1.64 {d0,d1}, [r1], r2
mspel_filter q11, q11, d2, d4, d6, d0, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
mspel_filter q12, q12, d3, d5, d7, d1, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
vst1.64 {q11,q12}, [r4,:128]! @ store and increment
vld1.64 {d2,d3}, [r1], r2
mspel_filter q11, q11, d4, d6, d0, d2, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
mspel_filter q12, q12, d5, d7, d1, d3, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
vst1.64 {q11,q12}, [r4,:128]! @ store and increment
vld1.64 {d4,d5}, [r1], r2
mspel_filter q11, q11, d6, d0, d2, d4, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
mspel_filter q12, q12, d7, d1, d3, d5, \filter_v_a, \filter_v_b, \filter_v_c, \filter_v_d, d28, d29, d30, d31, q13, \filter_shift, narrow=0
vst1.64 {q11,q12}, [r4,:128]! @ store and increment
bne 1b
rsb r3, r3, #(64 + \filter_add) @ r3 = (64 + filter_add) - r3
mov r12, #8 @ loop counter
mov r4, sp @ r4 = tmp
@ Do horizontal filtering from temp to dst
mspel_constants i16, d28, d29, d30, d31, \filter_h_a, \filter_h_b, \filter_h_c, \filter_h_d, q13, r3
2:
subs r12, r12, #1
vld1.64 {q0,q1}, [r4,:128]! @ read one line of tmp
vext.16 q2, q0, q1, #2
vext.16 q3, q0, q1, #3
vext.16 q1, q0, q1, #1 @ do last because it writes to q1 which is read by the other vext instructions
mspel_filter.16 q11, q12, d22, d23, d21, d0, d1, d2, d3, d4, d5, d6, d7, \filter_h_a, \filter_h_b, \filter_h_c, \filter_h_d, d28, d29, d30, d31, q13, 7
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
bne 2b
mov sp, r11
pop {r4, r11, pc}
endfunc
.endm
@ Use C preprocessor and assembler macros to expand to functions for horizontal and vertical filtering.
#define PUT_VC1_MSPEL_MC_HV(hmode, vmode) \
put_vc1_mspel_mc_hv hmode, vmode, \
MSPEL_MODE_ ## hmode ## _MUL_CONSTANTS, \
MSPEL_MODE_ ## vmode ## _MUL_CONSTANTS, \
MSPEL_MODES_ ## hmode ## vmode ## _ADDSHIFT_CONSTANTS
PUT_VC1_MSPEL_MC_HV(1, 1)
PUT_VC1_MSPEL_MC_HV(1, 2)
PUT_VC1_MSPEL_MC_HV(1, 3)
PUT_VC1_MSPEL_MC_HV(2, 1)
PUT_VC1_MSPEL_MC_HV(2, 2)
PUT_VC1_MSPEL_MC_HV(2, 3)
PUT_VC1_MSPEL_MC_HV(3, 1)
PUT_VC1_MSPEL_MC_HV(3, 2)
PUT_VC1_MSPEL_MC_HV(3, 3)
#undef PUT_VC1_MSPEL_MC_HV
.macro put_vc1_mspel_mc_h_only hmode filter_a filter_b filter_c filter_d filter_add filter_shift
function ff_put_vc1_mspel_mc\hmode\()0_neon, export=1
rsb r3, r3, #\filter_add @ r3 = filter_add - r = filter_add - rnd
mov r12, #8 @ loop counter
sub r1, r1, #1 @ slide back, using immediate
mspel_constants i8, d28, d29, d30, d31, \filter_a, \filter_b, \filter_c, \filter_d, q13, r3
1:
subs r12, r12, #1
vld1.64 {d0,d1}, [r1], r2 @ read 16 bytes even though we only need 11, also src += stride
vext.8 d2, d0, d1, #2
vext.8 d3, d0, d1, #3
vext.8 d1, d0, d1, #1 @ do last because it writes to d1 which is read by the other vext instructions
mspel_filter q11, d21, d0, d1, d2, d3, \filter_a, \filter_b, \filter_c, \filter_d, d28, d29, d30, d31, q13, \filter_shift
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
bne 1b
bx lr
endfunc
.endm
@ Use C preprocessor and assembler macros to expand to functions for horizontal only filtering.
#define PUT_VC1_MSPEL_MC_H_ONLY(hmode) \
put_vc1_mspel_mc_h_only hmode, MSPEL_MODE_ ## hmode ## _MUL_CONSTANTS, MSPEL_MODE_ ## hmode ## _ADDSHIFT_CONSTANTS
PUT_VC1_MSPEL_MC_H_ONLY(1)
PUT_VC1_MSPEL_MC_H_ONLY(2)
PUT_VC1_MSPEL_MC_H_ONLY(3)
#undef PUT_VC1_MSPEL_MC_H_ONLY
@ (uint8_t *dst [r0], const uint8_t *src [r1], ptrdiff_t stride [r2], int rnd [r3])
.macro put_vc1_mspel_mc_v_only vmode filter_a filter_b filter_c filter_d filter_add filter_shift
function ff_put_vc1_mspel_mc0\vmode\()_neon, export=1
add r3, r3, #\filter_add - 1 @ r3 = filter_add - r = filter_add - (1 - rnd) = filter_add - 1 + rnd
mov r12, #8 @ loop counter
sub r1, r1, r2 @ r1 = &src[-stride] @ slide back
mspel_constants i8, d28, d29, d30, d31, \filter_a, \filter_b, \filter_c, \filter_d, q13, r3
vld1.64 {d0}, [r1], r2 @ d0 = src[-stride]
vld1.64 {d1}, [r1], r2 @ d1 = src[0]
vld1.64 {d2}, [r1], r2 @ d2 = src[stride]
1:
subs r12, r12, #4
vld1.64 {d3}, [r1], r2 @ d3 = src[stride * 2]
mspel_filter q11, d21, d0, d1, d2, d3, \filter_a, \filter_b, \filter_c, \filter_d, d28, d29, d30, d31, q13, \filter_shift
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
vld1.64 {d0}, [r1], r2 @ d0 = next line
mspel_filter q11, d21, d1, d2, d3, d0, \filter_a, \filter_b, \filter_c, \filter_d, d28, d29, d30, d31, q13, \filter_shift
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
vld1.64 {d1}, [r1], r2 @ d1 = next line
mspel_filter q11, d21, d2, d3, d0, d1, \filter_a, \filter_b, \filter_c, \filter_d, d28, d29, d30, d31, q13, \filter_shift
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
vld1.64 {d2}, [r1], r2 @ d2 = next line
mspel_filter q11, d21, d3, d0, d1, d2, \filter_a, \filter_b, \filter_c, \filter_d, d28, d29, d30, d31, q13, \filter_shift
vst1.64 {d21}, [r0,:64], r2 @ store and increment dst
bne 1b
bx lr
endfunc
.endm
@ Use C preprocessor and assembler macros to expand to functions for vertical only filtering.
#define PUT_VC1_MSPEL_MC_V_ONLY(vmode) \
put_vc1_mspel_mc_v_only vmode, MSPEL_MODE_ ## vmode ## _MUL_CONSTANTS, MSPEL_MODE_ ## vmode ## _ADDSHIFT_CONSTANTS
PUT_VC1_MSPEL_MC_V_ONLY(1)
PUT_VC1_MSPEL_MC_V_ONLY(2)
PUT_VC1_MSPEL_MC_V_ONLY(3)
#undef PUT_VC1_MSPEL_MC_V_ONLY
function ff_put_pixels8x8_neon, export=1
vld1.64 {d0}, [r1], r2
vld1.64 {d1}, [r1], r2
vld1.64 {d2}, [r1], r2
vld1.64 {d3}, [r1], r2
vld1.64 {d4}, [r1], r2
vld1.64 {d5}, [r1], r2
vld1.64 {d6}, [r1], r2
vld1.64 {d7}, [r1]
vst1.64 {d0}, [r0,:64], r2
vst1.64 {d1}, [r0,:64], r2
vst1.64 {d2}, [r0,:64], r2
vst1.64 {d3}, [r0,:64], r2
vst1.64 {d4}, [r0,:64], r2
vst1.64 {d5}, [r0,:64], r2
vst1.64 {d6}, [r0,:64], r2
vst1.64 {d7}, [r0,:64]
bx lr
endfunc
function ff_vc1_inv_trans_8x8_dc_neon, export=1
ldrsh r2, [r2] @ int dc = block[0];
vld1.64 {d0}, [r0,:64], r1
vld1.64 {d1}, [r0,:64], r1
vld1.64 {d4}, [r0,:64], r1
vld1.64 {d5}, [r0,:64], r1
add r2, r2, r2, lsl #1 @ dc = (3 * dc + 1) >> 1;
vld1.64 {d6}, [r0,:64], r1
add r2, r2, #1
vld1.64 {d7}, [r0,:64], r1
vld1.64 {d16}, [r0,:64], r1
vld1.64 {d17}, [r0,:64], r1
asr r2, r2, #1
sub r0, r0, r1, lsl #3 @ restore r0 to original value
add r2, r2, r2, lsl #1 @ dc = (3 * dc + 16) >> 5;
add r2, r2, #16
asr r2, r2, #5
vdup.16 q1, r2 @ dc
vaddw.u8 q9, q1, d0
vaddw.u8 q10, q1, d1
vaddw.u8 q11, q1, d4
vaddw.u8 q12, q1, d5
vqmovun.s16 d0, q9
vqmovun.s16 d1, q10
vqmovun.s16 d4, q11
vst1.64 {d0}, [r0,:64], r1
vqmovun.s16 d5, q12
vst1.64 {d1}, [r0,:64], r1
vaddw.u8 q13, q1, d6
vst1.64 {d4}, [r0,:64], r1
vaddw.u8 q14, q1, d7
vst1.64 {d5}, [r0,:64], r1
vaddw.u8 q15, q1, d16
vaddw.u8 q1, q1, d17 @ this destroys q1
vqmovun.s16 d6, q13
vqmovun.s16 d7, q14
vqmovun.s16 d16, q15
vqmovun.s16 d17, q1
vst1.64 {d6}, [r0,:64], r1
vst1.64 {d7}, [r0,:64], r1
vst1.64 {d16}, [r0,:64], r1
vst1.64 {d17}, [r0,:64]
bx lr
endfunc
function ff_vc1_inv_trans_8x4_dc_neon, export=1
ldrsh r2, [r2] @ int dc = block[0];
vld1.64 {d0}, [r0,:64], r1
vld1.64 {d1}, [r0,:64], r1
vld1.64 {d4}, [r0,:64], r1
vld1.64 {d5}, [r0,:64], r1
add r2, r2, r2, lsl #1 @ dc = ( 3 * dc + 1) >> 1;
sub r0, r0, r1, lsl #2 @ restore r0 to original value
add r2, r2, #1
asr r2, r2, #1
add r2, r2, r2, lsl #4 @ dc = (17 * dc + 64) >> 7;
add r2, r2, #64
asr r2, r2, #7
vdup.16 q1, r2 @ dc
vaddw.u8 q3, q1, d0
vaddw.u8 q8, q1, d1
vaddw.u8 q9, q1, d4
vaddw.u8 q10, q1, d5
vqmovun.s16 d0, q3
vqmovun.s16 d1, q8
vqmovun.s16 d4, q9
vst1.64 {d0}, [r0,:64], r1
vqmovun.s16 d5, q10
vst1.64 {d1}, [r0,:64], r1
vst1.64 {d4}, [r0,:64], r1
vst1.64 {d5}, [r0,:64]
bx lr
endfunc
function ff_vc1_inv_trans_4x8_dc_neon, export=1
ldrsh r2, [r2] @ int dc = block[0];
vld1.32 {d0[]}, [r0,:32], r1
vld1.32 {d1[]}, [r0,:32], r1
vld1.32 {d0[1]}, [r0,:32], r1
vld1.32 {d1[1]}, [r0,:32], r1
add r2, r2, r2, lsl #4 @ dc = (17 * dc + 4) >> 3;
vld1.32 {d4[]}, [r0,:32], r1
add r2, r2, #4
vld1.32 {d5[]}, [r0,:32], r1
vld1.32 {d4[1]}, [r0,:32], r1
asr r2, r2, #3
vld1.32 {d5[1]}, [r0,:32], r1
add r2, r2, r2, lsl #1 @ dc = (12 * dc + 64) >> 7;
sub r0, r0, r1, lsl #3 @ restore r0 to original value
lsl r2, r2, #2
add r2, r2, #64
asr r2, r2, #7
vdup.16 q1, r2 @ dc
vaddw.u8 q3, q1, d0
vaddw.u8 q8, q1, d1
vaddw.u8 q9, q1, d4
vaddw.u8 q10, q1, d5
vqmovun.s16 d0, q3
vst1.32 {d0[0]}, [r0,:32], r1
vqmovun.s16 d1, q8
vst1.32 {d1[0]}, [r0,:32], r1
vqmovun.s16 d4, q9
vst1.32 {d0[1]}, [r0,:32], r1
vqmovun.s16 d5, q10
vst1.32 {d1[1]}, [r0,:32], r1
vst1.32 {d4[0]}, [r0,:32], r1
vst1.32 {d5[0]}, [r0,:32], r1
vst1.32 {d4[1]}, [r0,:32], r1
vst1.32 {d5[1]}, [r0,:32]
bx lr
endfunc
function ff_vc1_inv_trans_4x4_dc_neon, export=1
ldrsh r2, [r2] @ int dc = block[0];
vld1.32 {d0[]}, [r0,:32], r1
vld1.32 {d1[]}, [r0,:32], r1
vld1.32 {d0[1]}, [r0,:32], r1
vld1.32 {d1[1]}, [r0,:32], r1
add r2, r2, r2, lsl #4 @ dc = (17 * dc + 4) >> 3;
sub r0, r0, r1, lsl #2 @ restore r0 to original value
add r2, r2, #4
asr r2, r2, #3
add r2, r2, r2, lsl #4 @ dc = (17 * dc + 64) >> 7;
add r2, r2, #64
asr r2, r2, #7
vdup.16 q1, r2 @ dc
vaddw.u8 q2, q1, d0
vaddw.u8 q3, q1, d1
vqmovun.s16 d0, q2
vst1.32 {d0[0]}, [r0,:32], r1
vqmovun.s16 d1, q3
vst1.32 {d1[0]}, [r0,:32], r1
vst1.32 {d0[1]}, [r0,:32], r1
vst1.32 {d1[1]}, [r0,:32]
bx lr
endfunc
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
@ VC-1 in-loop deblocking filter for 4 pixel pairs at boundary of vertically-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of lower block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_v_loop_filter4_neon, export=1
sub r3, r0, r1, lsl #2
vldr d0, .Lcoeffs
vld1.32 {d1[0]}, [r0], r1 @ P5
vld1.32 {d2[0]}, [r3], r1 @ P1
vld1.32 {d3[0]}, [r3], r1 @ P2
vld1.32 {d4[0]}, [r0], r1 @ P6
vld1.32 {d5[0]}, [r3], r1 @ P3
vld1.32 {d6[0]}, [r0], r1 @ P7
vld1.32 {d7[0]}, [r3] @ P4
vld1.32 {d16[0]}, [r0] @ P8
vshll.u8 q9, d1, #1 @ 2*P5
vdup.16 d17, r2 @ pq
vshll.u8 q10, d2, #1 @ 2*P1
vmovl.u8 q11, d3 @ P2
vmovl.u8 q1, d4 @ P6
vmovl.u8 q12, d5 @ P3
vmls.i16 d20, d22, d0[1] @ 2*P1-5*P2
vmovl.u8 q11, d6 @ P7
vmls.i16 d18, d2, d0[1] @ 2*P5-5*P6
vshll.u8 q2, d5, #1 @ 2*P3
vmovl.u8 q3, d7 @ P4
vmla.i16 d18, d22, d0[1] @ 2*P5-5*P6+5*P7
vmovl.u8 q11, d16 @ P8
vmla.u16 d20, d24, d0[1] @ 2*P1-5*P2+5*P3
vmovl.u8 q12, d1 @ P5
vmls.u16 d4, d6, d0[1] @ 2*P3-5*P4
vmls.u16 d18, d22, d0[0] @ 2*P5-5*P6+5*P7-2*P8
vsub.i16 d1, d6, d24 @ P4-P5
vmls.i16 d20, d6, d0[0] @ 2*P1-5*P2+5*P3-2*P4
vmla.i16 d4, d24, d0[1] @ 2*P3-5*P4+5*P5
vmls.i16 d4, d2, d0[0] @ 2*P3-5*P4+5*P5-2*P6
vabs.s16 d2, d1
vrshr.s16 d3, d18, #3
vrshr.s16 d5, d20, #3
vshr.s16 d2, d2, #1 @ clip
vrshr.s16 d4, d4, #3
vabs.s16 d3, d3 @ a2
vshr.s16 d1, d1, #8 @ clip_sign
vabs.s16 d5, d5 @ a1
vceq.i16 d7, d2, #0 @ test clip == 0
vabs.s16 d16, d4 @ a0
vshr.s16 d4, d4, #8 @ a0_sign
vcge.s16 d18, d5, d3 @ test a1 >= a2
vcge.s16 d17, d16, d17 @ test a0 >= pq
vbsl d18, d3, d5 @ a3
vsub.i16 d1, d1, d4 @ clip_sign - a0_sign
vorr d3, d7, d17 @ test clip == 0 || a0 >= pq
vqsub.u16 d4, d16, d18 @ a0 >= a3 ? a0-a3 : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 d5, d18, d16 @ test a3 >= a0
vmul.i16 d0, d4, d0[1] @ a0 >= a3 ? 5*(a0-a3) : 0
vorr d4, d3, d5 @ test clip == 0 || a0 >= pq || a3 >= a0
vmov.32 r0, d4[1] @ move to gp reg
vshr.u16 d0, d0, #3 @ a0 >= a3 ? (5*(a0-a3))>>3 : 0
vcge.s16 d4, d0, d2
tst r0, #1
bne 1f @ none of the 4 pixel pairs should be updated if this one is not filtered
vbsl d4, d2, d0 @ FFMIN(d, clip)
vbic d0, d4, d3 @ set each d to zero if it should not be filtered because clip == 0 || a0 >= pq (a3 > a0 case already zeroed by saturating sub)
vmls.i16 d6, d0, d1 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P4
vmla.i16 d24, d0, d1 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P5
vqmovun.s16 d0, q3
vqmovun.s16 d1, q12
vst1.32 {d0[0]}, [r3], r1
vst1.32 {d1[0]}, [r3]
1: bx lr
endfunc
@ VC-1 in-loop deblocking filter for 4 pixel pairs at boundary of horizontally-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of right block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_h_loop_filter4_neon, export=1
sub r3, r0, #4 @ where to start reading
vldr d0, .Lcoeffs
vld1.32 {d2}, [r3], r1
sub r0, r0, #1 @ where to start writing
vld1.32 {d4}, [r3], r1
vld1.32 {d3}, [r3], r1
vld1.32 {d5}, [r3]
vdup.16 d1, r2 @ pq
vtrn.8 q1, q2
vtrn.16 d2, d3 @ P1, P5, P3, P7
vtrn.16 d4, d5 @ P2, P6, P4, P8
vshll.u8 q3, d2, #1 @ 2*P1, 2*P5
vmovl.u8 q8, d4 @ P2, P6
vmovl.u8 q9, d3 @ P3, P7
vmovl.u8 q2, d5 @ P4, P8
vmls.i16 q3, q8, d0[1] @ 2*P1-5*P2, 2*P5-5*P6
vshll.u8 q10, d3, #1 @ 2*P3, 2*P7
vmovl.u8 q1, d2 @ P1, P5
vmla.i16 q3, q9, d0[1] @ 2*P1-5*P2+5*P3, 2*P5-5*P6+5*P7
vmls.i16 q3, q2, d0[0] @ 2*P1-5*P2+5*P3-2*P4, 2*P5-5*P6+5*P7-2*P8
vmov d2, d3 @ needs to be in an even-numbered vector for when we come to narrow it later
vmls.i16 d20, d4, d0[1] @ 2*P3-5*P4
vmla.i16 d20, d3, d0[1] @ 2*P3-5*P4+5*P5
vsub.i16 d3, d4, d2 @ P4-P5
vmls.i16 d20, d17, d0[0] @ 2*P3-5*P4+5*P5-2*P6
vrshr.s16 q3, q3, #3
vabs.s16 d5, d3
vshr.s16 d3, d3, #8 @ clip_sign
vrshr.s16 d16, d20, #3
vabs.s16 q3, q3 @ a1, a2
vshr.s16 d5, d5, #1 @ clip
vabs.s16 d17, d16 @ a0
vceq.i16 d18, d5, #0 @ test clip == 0
vshr.s16 d16, d16, #8 @ a0_sign
vcge.s16 d19, d6, d7 @ test a1 >= a2
vcge.s16 d1, d17, d1 @ test a0 >= pq
vsub.i16 d16, d3, d16 @ clip_sign - a0_sign
vbsl d19, d7, d6 @ a3
vorr d1, d18, d1 @ test clip == 0 || a0 >= pq
vqsub.u16 d3, d17, d19 @ a0 >= a3 ? a0-a3 : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 d6, d19, d17 @ test a3 >= a0 @
vmul.i16 d0, d3, d0[1] @ a0 >= a3 ? 5*(a0-a3) : 0
vorr d3, d1, d6 @ test clip == 0 || a0 >= pq || a3 >= a0
vmov.32 r2, d3[1] @ move to gp reg
vshr.u16 d0, d0, #3 @ a0 >= a3 ? (5*(a0-a3))>>3 : 0
vcge.s16 d3, d0, d5
tst r2, #1
bne 1f @ none of the 4 pixel pairs should be updated if this one is not filtered
vbsl d3, d5, d0 @ FFMIN(d, clip)
vbic d0, d3, d1 @ set each d to zero if it should not be filtered because clip == 0 || a0 >= pq (a3 > a0 case already zeroed by saturating sub)
vmla.i16 d2, d0, d16 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P5
vmls.i16 d4, d0, d16 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P4
vqmovun.s16 d1, q1
vqmovun.s16 d0, q2
vst2.8 {d0[0], d1[0]}, [r0], r1
vst2.8 {d0[1], d1[1]}, [r0], r1
vst2.8 {d0[2], d1[2]}, [r0], r1
vst2.8 {d0[3], d1[3]}, [r0]
1: bx lr
endfunc
@ VC-1 in-loop deblocking filter for 8 pixel pairs at boundary of vertically-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of lower block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_v_loop_filter8_neon, export=1
sub r3, r0, r1, lsl #2
vldr d0, .Lcoeffs
vld1.32 {d1}, [r0, :64], r1 @ P5
vld1.32 {d2}, [r3, :64], r1 @ P1
vld1.32 {d3}, [r3, :64], r1 @ P2
vld1.32 {d4}, [r0, :64], r1 @ P6
vld1.32 {d5}, [r3, :64], r1 @ P3
vld1.32 {d6}, [r0, :64], r1 @ P7
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vshll.u8 q8, d1, #1 @ 2*P5
vshll.u8 q9, d2, #1 @ 2*P1
vld1.32 {d7}, [r3, :64] @ P4
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vmovl.u8 q1, d3 @ P2
vld1.32 {d20}, [r0, :64] @ P8
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vmovl.u8 q11, d4 @ P6
vdup.16 q12, r2 @ pq
vmovl.u8 q13, d5 @ P3
vmls.i16 q9, q1, d0[1] @ 2*P1-5*P2
vmovl.u8 q1, d6 @ P7
vshll.u8 q2, d5, #1 @ 2*P3
vmls.i16 q8, q11, d0[1] @ 2*P5-5*P6
vmovl.u8 q3, d7 @ P4
vmovl.u8 q10, d20 @ P8
vmla.i16 q8, q1, d0[1] @ 2*P5-5*P6+5*P7
vmovl.u8 q1, d1 @ P5
vmla.i16 q9, q13, d0[1] @ 2*P1-5*P2+5*P3
vsub.i16 q13, q3, q1 @ P4-P5
vmls.i16 q2, q3, d0[1] @ 2*P3-5*P4
vmls.i16 q8, q10, d0[0] @ 2*P5-5*P6+5*P7-2*P8
vabs.s16 q10, q13
vshr.s16 q13, q13, #8 @ clip_sign
vmls.i16 q9, q3, d0[0] @ 2*P1-5*P2+5*P3-2*P4
vshr.s16 q10, q10, #1 @ clip
vmla.i16 q2, q1, d0[1] @ 2*P3-5*P4+5*P5
vrshr.s16 q8, q8, #3
vmls.i16 q2, q11, d0[0] @ 2*P3-5*P4+5*P5-2*P6
vceq.i16 q11, q10, #0 @ test clip == 0
vrshr.s16 q9, q9, #3
vabs.s16 q8, q8 @ a2
vabs.s16 q9, q9 @ a1
vrshr.s16 q2, q2, #3
vcge.s16 q14, q9, q8 @ test a1 >= a2
vabs.s16 q15, q2 @ a0
vshr.s16 q2, q2, #8 @ a0_sign
vbsl q14, q8, q9 @ a3
vcge.s16 q8, q15, q12 @ test a0 >= pq
vsub.i16 q2, q13, q2 @ clip_sign - a0_sign
vqsub.u16 q9, q15, q14 @ a0 >= a3 ? a0-a3 : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 q12, q14, q15 @ test a3 >= a0
vorr q8, q11, q8 @ test clip == 0 || a0 >= pq
vmul.i16 q0, q9, d0[1] @ a0 >= a3 ? 5*(a0-a3) : 0
vorr q9, q8, q12 @ test clip == 0 || a0 >= pq || a3 >= a0
vshl.i64 q11, q9, #16
vmov.32 r0, d18[1] @ move to gp reg
vshr.u16 q0, q0, #3 @ a0 >= a3 ? (5*(a0-a3))>>3 : 0
vmov.32 r2, d19[1]
vshr.s64 q9, q11, #48
vcge.s16 q11, q0, q10
vorr q8, q8, q9
and r0, r0, r2
vbsl q11, q10, q0 @ FFMIN(d, clip)
tst r0, #1
bne 1f @ none of the 8 pixel pairs should be updated in this case
vbic q0, q11, q8 @ set each d to zero if it should not be filtered
vmls.i16 q3, q0, q2 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P4
vmla.i16 q1, q0, q2 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P5
vqmovun.s16 d0, q3
vqmovun.s16 d1, q1
vst1.32 {d0}, [r3, :64], r1
vst1.32 {d1}, [r3, :64]
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
1: bx lr
endfunc
.align 5
.Lcoeffs:
.quad 0x00050002
@ VC-1 in-loop deblocking filter for 8 pixel pairs at boundary of horizontally-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of right block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_h_loop_filter8_neon, export=1
push {lr}
sub r3, r0, #4 @ where to start reading
vldr d0, .Lcoeffs
vld1.32 {d2}, [r3], r1 @ P1[0], P2[0]...
sub r0, r0, #1 @ where to start writing
vld1.32 {d4}, [r3], r1
add r12, r0, r1, lsl #2
vld1.32 {d3}, [r3], r1
vld1.32 {d5}, [r3], r1
vld1.32 {d6}, [r3], r1
vld1.32 {d16}, [r3], r1
vld1.32 {d7}, [r3], r1
vld1.32 {d17}, [r3]
vtrn.8 q1, q2 @ P1[0], P1[1], P3[0]... P1[2], P1[3], P3[2]... P2[0], P2[1], P4[0]... P2[2], P2[3], P4[2]...
vdup.16 q9, r2 @ pq
vtrn.16 d2, d3 @ P1[0], P1[1], P1[2], P1[3], P5[0]... P3[0], P3[1], P3[2], P3[3], P7[0]...
vtrn.16 d4, d5 @ P2[0], P2[1], P2[2], P2[3], P6[0]... P4[0], P4[1], P4[2], P4[3], P8[0]...
vtrn.8 q3, q8 @ P1[4], P1[5], P3[4]... P1[6], P1[7], P3[6]... P2[4], P2[5], P4[4]... P2[6], P2[7], P4[6]...
vtrn.16 d6, d7 @ P1[4], P1[5], P1[6], P1[7], P5[4]... P3[4], P3[5], P3[5], P3[7], P7[4]...
vtrn.16 d16, d17 @ P2[4], P2[5], P2[6], P2[7], P6[4]... P4[4], P4[5], P4[6], P4[7], P8[4]...
vtrn.32 d2, d6 @ P1, P5
vtrn.32 d4, d16 @ P2, P6
vtrn.32 d3, d7 @ P3, P7
vtrn.32 d5, d17 @ P4, P8
vshll.u8 q10, d2, #1 @ 2*P1
vshll.u8 q11, d6, #1 @ 2*P5
vmovl.u8 q12, d4 @ P2
vmovl.u8 q13, d16 @ P6
vmovl.u8 q14, d3 @ P3
vmls.i16 q10, q12, d0[1] @ 2*P1-5*P2
vmovl.u8 q12, d7 @ P7
vshll.u8 q1, d3, #1 @ 2*P3
vmls.i16 q11, q13, d0[1] @ 2*P5-5*P6
vmovl.u8 q2, d5 @ P4
vmovl.u8 q8, d17 @ P8
vmla.i16 q11, q12, d0[1] @ 2*P5-5*P6+5*P7
vmovl.u8 q3, d6 @ P5
vmla.i16 q10, q14, d0[1] @ 2*P1-5*P2+5*P3
vsub.i16 q12, q2, q3 @ P4-P5
vmls.i16 q1, q2, d0[1] @ 2*P3-5*P4
vmls.i16 q11, q8, d0[0] @ 2*P5-5*P6+5*P7-2*P8
vabs.s16 q8, q12
vshr.s16 q12, q12, #8 @ clip_sign
vmls.i16 q10, q2, d0[0] @ 2*P1-5*P2+5*P3-2*P4
vshr.s16 q8, q8, #1 @ clip
vmla.i16 q1, q3, d0[1] @ 2*P3-5*P4+5*P5
vrshr.s16 q11, q11, #3
vmls.i16 q1, q13, d0[0] @ 2*P3-5*P4+5*P5-2*P6
vceq.i16 q13, q8, #0 @ test clip == 0
vrshr.s16 q10, q10, #3
vabs.s16 q11, q11 @ a2
vabs.s16 q10, q10 @ a1
vrshr.s16 q1, q1, #3
vcge.s16 q14, q10, q11 @ test a1 >= a2
vabs.s16 q15, q1 @ a0
vshr.s16 q1, q1, #8 @ a0_sign
vbsl q14, q11, q10 @ a3
vcge.s16 q9, q15, q9 @ test a0 >= pq
vsub.i16 q1, q12, q1 @ clip_sign - a0_sign
vqsub.u16 q10, q15, q14 @ a0 >= a3 ? a0-a3 : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 q11, q14, q15 @ test a3 >= a0
vorr q9, q13, q9 @ test clip == 0 || a0 >= pq
vmul.i16 q0, q10, d0[1] @ a0 >= a3 ? 5*(a0-a3) : 0
vorr q10, q9, q11 @ test clip == 0 || a0 >= pq || a3 >= a0
vmov.32 r2, d20[1] @ move to gp reg
vshr.u16 q0, q0, #3 @ a0 >= a3 ? (5*(a0-a3))>>3 : 0
vmov.32 r3, d21[1]
vcge.s16 q10, q0, q8
and r14, r2, r3
vbsl q10, q8, q0 @ FFMIN(d, clip)
tst r14, #1
bne 2f @ none of the 8 pixel pairs should be updated in this case
vbic q0, q10, q9 @ set each d to zero if it should not be filtered because clip == 0 || a0 >= pq (a3 > a0 case already zeroed by saturating sub)
vmla.i16 q3, q0, q1 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P5
vmls.i16 q2, q0, q1 @ invert d depending on clip_sign & a0_sign, or zero it if they match, and accumulate into P4
vqmovun.s16 d1, q3
vqmovun.s16 d0, q2
tst r2, #1
bne 1f @ none of the first 4 pixel pairs should be updated if so
vst2.8 {d0[0], d1[0]}, [r0], r1
vst2.8 {d0[1], d1[1]}, [r0], r1
vst2.8 {d0[2], d1[2]}, [r0], r1
vst2.8 {d0[3], d1[3]}, [r0]
1: tst r3, #1
bne 2f @ none of the second 4 pixel pairs should be updated if so
vst2.8 {d0[4], d1[4]}, [r12], r1
vst2.8 {d0[5], d1[5]}, [r12], r1
vst2.8 {d0[6], d1[6]}, [r12], r1
vst2.8 {d0[7], d1[7]}, [r12]
2: pop {pc}
endfunc
@ VC-1 in-loop deblocking filter for 16 pixel pairs at boundary of vertically-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of lower block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_v_loop_filter16_neon, export=1
vpush {d8-d15}
sub r3, r0, r1, lsl #2
vldr d0, .Lcoeffs
vld1.64 {q1}, [r0, :128], r1 @ P5
vld1.64 {q2}, [r3, :128], r1 @ P1
vld1.64 {q3}, [r3, :128], r1 @ P2
vld1.64 {q4}, [r0, :128], r1 @ P6
vld1.64 {q5}, [r3, :128], r1 @ P3
vld1.64 {q6}, [r0, :128], r1 @ P7
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vshll.u8 q7, d2, #1 @ 2*P5[0..7]
vshll.u8 q8, d4, #1 @ 2*P1[0..7]
vld1.64 {q9}, [r3, :128] @ P4
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vmovl.u8 q10, d6 @ P2[0..7]
vld1.64 {q11}, [r0, :128] @ P8
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
vmovl.u8 q12, d8 @ P6[0..7]
vdup.16 q13, r2 @ pq
vshll.u8 q2, d5, #1 @ 2*P1[8..15]
vmls.i16 q8, q10, d0[1] @ 2*P1[0..7]-5*P2[0..7]
vshll.u8 q10, d3, #1 @ 2*P5[8..15]
vmovl.u8 q3, d7 @ P2[8..15]
vmls.i16 q7, q12, d0[1] @ 2*P5[0..7]-5*P6[0..7]
vmovl.u8 q4, d9 @ P6[8..15]
vmovl.u8 q14, d10 @ P3[0..7]
vmovl.u8 q15, d12 @ P7[0..7]
vmls.i16 q2, q3, d0[1] @ 2*P1[8..15]-5*P2[8..15]
vshll.u8 q3, d10, #1 @ 2*P3[0..7]
vmls.i16 q10, q4, d0[1] @ 2*P5[8..15]-5*P6[8..15]
vmovl.u8 q6, d13 @ P7[8..15]
vmla.i16 q8, q14, d0[1] @ 2*P1[0..7]-5*P2[0..7]+5*P3[0..7]
vmovl.u8 q14, d18 @ P4[0..7]
vmovl.u8 q9, d19 @ P4[8..15]
vmla.i16 q7, q15, d0[1] @ 2*P5[0..7]-5*P6[0..7]+5*P7[0..7]
vmovl.u8 q15, d11 @ P3[8..15]
vshll.u8 q5, d11, #1 @ 2*P3[8..15]
vmls.i16 q3, q14, d0[1] @ 2*P3[0..7]-5*P4[0..7]
vmla.i16 q2, q15, d0[1] @ 2*P1[8..15]-5*P2[8..15]+5*P3[8..15]
vmovl.u8 q15, d22 @ P8[0..7]
vmovl.u8 q11, d23 @ P8[8..15]
vmla.i16 q10, q6, d0[1] @ 2*P5[8..15]-5*P6[8..15]+5*P7[8..15]
vmovl.u8 q6, d2 @ P5[0..7]
vmovl.u8 q1, d3 @ P5[8..15]
vmls.i16 q5, q9, d0[1] @ 2*P3[8..15]-5*P4[8..15]
vmls.i16 q8, q14, d0[0] @ 2*P1[0..7]-5*P2[0..7]+5*P3[0..7]-2*P4[0..7]
vmls.i16 q7, q15, d0[0] @ 2*P5[0..7]-5*P6[0..7]+5*P7[0..7]-2*P8[0..7]
vsub.i16 q15, q14, q6 @ P4[0..7]-P5[0..7]
vmla.i16 q3, q6, d0[1] @ 2*P3[0..7]-5*P4[0..7]+5*P5[0..7]
vrshr.s16 q8, q8, #3
vmls.i16 q2, q9, d0[0] @ 2*P1[8..15]-5*P2[8..15]+5*P3[8..15]-2*P4[8..15]
vrshr.s16 q7, q7, #3
vmls.i16 q10, q11, d0[0] @ 2*P5[8..15]-5*P6[8..15]+5*P7[8..15]-2*P8[8..15]
vabs.s16 q11, q15
vabs.s16 q8, q8 @ a1[0..7]
vmla.i16 q5, q1, d0[1] @ 2*P3[8..15]-5*P4[8..15]+5*P5[8..15]
vshr.s16 q15, q15, #8 @ clip_sign[0..7]
vrshr.s16 q2, q2, #3
vmls.i16 q3, q12, d0[0] @ 2*P3[0..7]-5*P4[0..7]+5*P5[0..7]-2*P6[0..7]
vabs.s16 q7, q7 @ a2[0..7]
vrshr.s16 q10, q10, #3
vsub.i16 q12, q9, q1 @ P4[8..15]-P5[8..15]
vshr.s16 q11, q11, #1 @ clip[0..7]
vmls.i16 q5, q4, d0[0] @ 2*P3[8..15]-5*P4[8..15]+5*P5[8..15]-2*P6[8..15]
vcge.s16 q4, q8, q7 @ test a1[0..7] >= a2[0..7]
vabs.s16 q2, q2 @ a1[8..15]
vrshr.s16 q3, q3, #3
vabs.s16 q10, q10 @ a2[8..15]
vbsl q4, q7, q8 @ a3[0..7]
vabs.s16 q7, q12
vshr.s16 q8, q12, #8 @ clip_sign[8..15]
vrshr.s16 q5, q5, #3
vcge.s16 q12, q2, q10 @ test a1[8..15] >= a2[8.15]
vshr.s16 q7, q7, #1 @ clip[8..15]
vbsl q12, q10, q2 @ a3[8..15]
vabs.s16 q2, q3 @ a0[0..7]
vceq.i16 q10, q11, #0 @ test clip[0..7] == 0
vshr.s16 q3, q3, #8 @ a0_sign[0..7]
vsub.i16 q3, q15, q3 @ clip_sign[0..7] - a0_sign[0..7]
vcge.s16 q15, q2, q13 @ test a0[0..7] >= pq
vorr q10, q10, q15 @ test clip[0..7] == 0 || a0[0..7] >= pq
vqsub.u16 q15, q2, q4 @ a0[0..7] >= a3[0..7] ? a0[0..7]-a3[0..7] : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 q2, q4, q2 @ test a3[0..7] >= a0[0..7]
vabs.s16 q4, q5 @ a0[8..15]
vshr.s16 q5, q5, #8 @ a0_sign[8..15]
vmul.i16 q15, q15, d0[1] @ a0[0..7] >= a3[0..7] ? 5*(a0[0..7]-a3[0..7]) : 0
vcge.s16 q13, q4, q13 @ test a0[8..15] >= pq
vorr q2, q10, q2 @ test clip[0..7] == 0 || a0[0..7] >= pq || a3[0..7] >= a0[0..7]
vsub.i16 q5, q8, q5 @ clip_sign[8..15] - a0_sign[8..15]
vceq.i16 q8, q7, #0 @ test clip[8..15] == 0
vshr.u16 q15, q15, #3 @ a0[0..7] >= a3[0..7] ? (5*(a0[0..7]-a3[0..7]))>>3 : 0
vmov.32 r0, d4[1] @ move to gp reg
vorr q8, q8, q13 @ test clip[8..15] == 0 || a0[8..15] >= pq
vqsub.u16 q13, q4, q12 @ a0[8..15] >= a3[8..15] ? a0[8..15]-a3[8..15] : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vmov.32 r2, d5[1]
vcge.s16 q4, q12, q4 @ test a3[8..15] >= a0[8..15]
vshl.i64 q2, q2, #16
vcge.s16 q12, q15, q11
vmul.i16 q0, q13, d0[1] @ a0[8..15] >= a3[8..15] ? 5*(a0[8..15]-a3[8..15]) : 0
vorr q4, q8, q4 @ test clip[8..15] == 0 || a0[8..15] >= pq || a3[8..15] >= a0[8..15]
vshr.s64 q2, q2, #48
and r0, r0, r2
vbsl q12, q11, q15 @ FFMIN(d[0..7], clip[0..7])
vshl.i64 q11, q4, #16
vmov.32 r2, d8[1]
vshr.u16 q0, q0, #3 @ a0[8..15] >= a3[8..15] ? (5*(a0[8..15]-a3[8..15]))>>3 : 0
vorr q2, q10, q2
vmov.32 r12, d9[1]
vshr.s64 q4, q11, #48
vcge.s16 q10, q0, q7
vbic q2, q12, q2 @ set each d[0..7] to zero if it should not be filtered because clip[0..7] == 0 || a0[0..7] >= pq (a3 > a0 case already zeroed by saturating sub)
vorr q4, q8, q4
and r2, r2, r12
vbsl q10, q7, q0 @ FFMIN(d[8..15], clip[8..15])
vmls.i16 q14, q2, q3 @ invert d[0..7] depending on clip_sign[0..7] & a0_sign[0..7], or zero it if they match, and accumulate into P4[0..7]
and r0, r0, r2
vbic q0, q10, q4 @ set each d[8..15] to zero if it should not be filtered because clip[8..15] == 0 || a0[8..15] >= pq (a3 > a0 case already zeroed by saturating sub)
tst r0, #1
bne 1f @ none of the 16 pixel pairs should be updated in this case
vmla.i16 q6, q2, q3 @ invert d[0..7] depending on clip_sign[0..7] & a0_sign[0..7], or zero it if they match, and accumulate into P5[0..7]
vmls.i16 q9, q0, q5 @ invert d[8..15] depending on clip_sign[8..15] & a0_sign[8..15], or zero it if they match, and accumulate into P4[8..15]
vqmovun.s16 d4, q14
vmla.i16 q1, q0, q5 @ invert d[8..15] depending on clip_sign[8..15] & a0_sign[8..15], or zero it if they match, and accumulate into P5[8..15]
vqmovun.s16 d0, q6
vqmovun.s16 d5, q9
vqmovun.s16 d1, q1
vst1.64 {q2}, [r3, :128], r1
vst1.64 {q0}, [r3, :128]
avcodec/vc1: Arm 32-bit NEON deblocking filter fast paths checkasm benchmarks on 1.5 GHz Cortex-A72 are as follows. Note that the C version can still outperform the NEON version in specific cases. The balance between different code paths is stream-dependent, but in practice the best case happens about 5% of the time, the worst case happens about 40% of the time, and the complexity of the remaining cases fall somewhere in between. Therefore, taking the average of the best and worst case timings is probably a conservative estimate of the degree by which the NEON code improves performance. vc1dsp.vc1_h_loop_filter4_bestcase_c: 19.0 vc1dsp.vc1_h_loop_filter4_bestcase_neon: 48.5 vc1dsp.vc1_h_loop_filter4_worstcase_c: 144.7 vc1dsp.vc1_h_loop_filter4_worstcase_neon: 76.2 vc1dsp.vc1_h_loop_filter8_bestcase_c: 41.0 vc1dsp.vc1_h_loop_filter8_bestcase_neon: 75.0 vc1dsp.vc1_h_loop_filter8_worstcase_c: 294.0 vc1dsp.vc1_h_loop_filter8_worstcase_neon: 102.7 vc1dsp.vc1_h_loop_filter16_bestcase_c: 54.7 vc1dsp.vc1_h_loop_filter16_bestcase_neon: 130.0 vc1dsp.vc1_h_loop_filter16_worstcase_c: 569.7 vc1dsp.vc1_h_loop_filter16_worstcase_neon: 186.7 vc1dsp.vc1_v_loop_filter4_bestcase_c: 20.2 vc1dsp.vc1_v_loop_filter4_bestcase_neon: 47.2 vc1dsp.vc1_v_loop_filter4_worstcase_c: 164.2 vc1dsp.vc1_v_loop_filter4_worstcase_neon: 68.5 vc1dsp.vc1_v_loop_filter8_bestcase_c: 43.5 vc1dsp.vc1_v_loop_filter8_bestcase_neon: 55.2 vc1dsp.vc1_v_loop_filter8_worstcase_c: 316.2 vc1dsp.vc1_v_loop_filter8_worstcase_neon: 72.7 vc1dsp.vc1_v_loop_filter16_bestcase_c: 62.2 vc1dsp.vc1_v_loop_filter16_bestcase_neon: 103.7 vc1dsp.vc1_v_loop_filter16_worstcase_c: 646.5 vc1dsp.vc1_v_loop_filter16_worstcase_neon: 110.7 Signed-off-by: Ben Avison <bavison@riscosopen.org> Signed-off-by: Martin Storsjö <martin@martin.st>
2022-03-31 19:23:47 +02:00
1: vpop {d8-d15}
bx lr
endfunc
@ VC-1 in-loop deblocking filter for 16 pixel pairs at boundary of horizontally-neighbouring blocks
@ On entry:
@ r0 -> top-left pel of right block
@ r1 = row stride, bytes
@ r2 = PQUANT bitstream parameter
function ff_vc1_h_loop_filter16_neon, export=1
push {r4-r6,lr}
vpush {d8-d15}
sub r3, r0, #4 @ where to start reading
vldr d0, .Lcoeffs
vld1.32 {d2}, [r3], r1 @ P1[0], P2[0]...
sub r0, r0, #1 @ where to start writing
vld1.32 {d3}, [r3], r1
add r4, r0, r1, lsl #2
vld1.32 {d10}, [r3], r1
vld1.32 {d11}, [r3], r1
vld1.32 {d16}, [r3], r1
vld1.32 {d4}, [r3], r1
vld1.32 {d8}, [r3], r1
vtrn.8 d2, d3 @ P1[0], P1[1], P3[0]... P2[0], P2[1], P4[0]...
vld1.32 {d14}, [r3], r1
vld1.32 {d5}, [r3], r1
vtrn.8 d10, d11 @ P1[2], P1[3], P3[2]... P2[2], P2[3], P4[2]...
vld1.32 {d6}, [r3], r1
vld1.32 {d12}, [r3], r1
vtrn.8 d16, d4 @ P1[4], P1[5], P3[4]... P2[4], P2[5], P4[4]...
vld1.32 {d13}, [r3], r1
vtrn.16 d2, d10 @ P1[0], P1[1], P1[2], P1[3], P5[0]... P3[0], P3[1], P3[2], P3[3], P7[0]...
vld1.32 {d1}, [r3], r1
vtrn.8 d8, d14 @ P1[6], P1[7], P3[6]... P2[6], P2[7], P4[6]...
vld1.32 {d7}, [r3], r1
vtrn.16 d3, d11 @ P2[0], P2[1], P2[2], P2[3], P6[0]... P4[0], P4[1], P4[2], P4[3], P8[0]...
vld1.32 {d9}, [r3], r1
vtrn.8 d5, d6 @ P1[8], P1[9], P3[8]... P2[8], P2[9], P4[8]...
vld1.32 {d15}, [r3]
vtrn.16 d16, d8 @ P1[4], P1[5], P1[6], P1[7], P5[4]... P3[4], P3[5], P3[6], P3[7], P7[4]...
vtrn.16 d4, d14 @ P2[4], P2[5], P2[6], P2[7], P6[4]... P4[4], P4[5], P4[6], P4[7], P8[4]...
vtrn.8 d12, d13 @ P1[10], P1[11], P3[10]... P2[10], P2[11], P4[10]...
vdup.16 q9, r2 @ pq
vtrn.8 d1, d7 @ P1[12], P1[13], P3[12]... P2[12], P2[13], P4[12]...
vtrn.32 d2, d16 @ P1[0..7], P5[0..7]
vtrn.16 d5, d12 @ P1[8], P1[7], P1[10], P1[11], P5[8]... P3[8], P3[9], P3[10], P3[11], P7[8]...
vtrn.16 d6, d13 @ P2[8], P2[7], P2[10], P2[11], P6[8]... P4[8], P4[9], P4[10], P4[11], P8[8]...
vtrn.8 d9, d15 @ P1[14], P1[15], P3[14]... P2[14], P2[15], P4[14]...
vtrn.32 d3, d4 @ P2[0..7], P6[0..7]
vshll.u8 q10, d2, #1 @ 2*P1[0..7]
vtrn.32 d10, d8 @ P3[0..7], P7[0..7]
vshll.u8 q11, d16, #1 @ 2*P5[0..7]
vtrn.32 d11, d14 @ P4[0..7], P8[0..7]
vtrn.16 d1, d9 @ P1[12], P1[13], P1[14], P1[15], P5[12]... P3[12], P3[13], P3[14], P3[15], P7[12]...
vtrn.16 d7, d15 @ P2[12], P2[13], P2[14], P2[15], P6[12]... P4[12], P4[13], P4[14], P4[15], P8[12]...
vmovl.u8 q1, d3 @ P2[0..7]
vmovl.u8 q12, d4 @ P6[0..7]
vtrn.32 d5, d1 @ P1[8..15], P5[8..15]
vtrn.32 d6, d7 @ P2[8..15], P6[8..15]
vtrn.32 d12, d9 @ P3[8..15], P7[8..15]
vtrn.32 d13, d15 @ P4[8..15], P8[8..15]
vmls.i16 q10, q1, d0[1] @ 2*P1[0..7]-5*P2[0..7]
vmovl.u8 q1, d10 @ P3[0..7]
vshll.u8 q2, d5, #1 @ 2*P1[8..15]
vshll.u8 q13, d1, #1 @ 2*P5[8..15]
vmls.i16 q11, q12, d0[1] @ 2*P5[0..7]-5*P6[0..7]
vmovl.u8 q14, d6 @ P2[8..15]
vmovl.u8 q3, d7 @ P6[8..15]
vmovl.u8 q15, d8 @ P7[0..7]
vmla.i16 q10, q1, d0[1] @ 2*P1[0..7]-5*P2[0..7]+5*P3[0..7]
vmovl.u8 q1, d12 @ P3[8..15]
vmls.i16 q2, q14, d0[1] @ 2*P1[8..15]-5*P2[8..15]
vmovl.u8 q4, d9 @ P7[8..15]
vshll.u8 q14, d10, #1 @ 2*P3[0..7]
vmls.i16 q13, q3, d0[1] @ 2*P5[8..15]-5*P6[8..15]
vmovl.u8 q5, d11 @ P4[0..7]
vmla.i16 q11, q15, d0[1] @ 2*P5[0..7]-5*P6[0..7]+5*P7[0..7]
vshll.u8 q15, d12, #1 @ 2*P3[8..15]
vmovl.u8 q6, d13 @ P4[8..15]
vmla.i16 q2, q1, d0[1] @ 2*P1[8..15]-5*P2[8..15]+5*P3[8..15]
vmovl.u8 q1, d14 @ P8[0..7]
vmovl.u8 q7, d15 @ P8[8..15]
vmla.i16 q13, q4, d0[1] @ 2*P5[8..15]-5*P6[8..15]+5*P7[8..15]
vmovl.u8 q4, d16 @ P5[0..7]
vmovl.u8 q8, d1 @ P5[8..15]
vmls.i16 q14, q5, d0[1] @ 2*P3[0..7]-5*P4[0..7]
vmls.i16 q15, q6, d0[1] @ 2*P3[8..15]-5*P4[8..15]
vmls.i16 q10, q5, d0[0] @ 2*P1[0..7]-5*P2[0..7]+5*P3[0..7]-2*P4[0..7]
vmls.i16 q11, q1, d0[0] @ 2*P5[0..7]-5*P6[0..7]+5*P7[0..7]-2*P8[0..7]
vsub.i16 q1, q5, q4 @ P4[0..7]-P5[0..7]
vmls.i16 q2, q6, d0[0] @ 2*P1[8..15]-5*P2[8..15]+5*P3[8..15]-2*P4[8..15]
vrshr.s16 q10, q10, #3
vmls.i16 q13, q7, d0[0] @ 2*P5[8..15]-5*P6[8..15]+5*P7[8..15]-2*P8[8..15]
vsub.i16 q7, q6, q8 @ P4[8..15]-P5[8..15]
vrshr.s16 q11, q11, #3
vmla.s16 q14, q4, d0[1] @ 2*P3[0..7]-5*P4[0..7]+5*P5[0..7]
vrshr.s16 q2, q2, #3
vmla.i16 q15, q8, d0[1] @ 2*P3[8..15]-5*P4[8..15]+5*P5[8..15]
vabs.s16 q10, q10 @ a1[0..7]
vrshr.s16 q13, q13, #3
vmls.i16 q15, q3, d0[0] @ 2*P3[8..15]-5*P4[8..15]+5*P5[8..15]-2*P6[8..15]
vabs.s16 q3, q11 @ a2[0..7]
vabs.s16 q2, q2 @ a1[8..15]
vmls.i16 q14, q12, d0[0] @ 2*P3[0..7]-5*P4[0..7]+5*P5[0..7]-2*P6[0..7]
vabs.s16 q11, q1
vabs.s16 q12, q13 @ a2[8..15]
vcge.s16 q13, q10, q3 @ test a1[0..7] >= a2[0..7]
vshr.s16 q1, q1, #8 @ clip_sign[0..7]
vrshr.s16 q15, q15, #3
vshr.s16 q11, q11, #1 @ clip[0..7]
vrshr.s16 q14, q14, #3
vbsl q13, q3, q10 @ a3[0..7]
vcge.s16 q3, q2, q12 @ test a1[8..15] >= a2[8.15]
vabs.s16 q10, q15 @ a0[8..15]
vshr.s16 q15, q15, #8 @ a0_sign[8..15]
vbsl q3, q12, q2 @ a3[8..15]
vabs.s16 q2, q14 @ a0[0..7]
vabs.s16 q12, q7
vshr.s16 q7, q7, #8 @ clip_sign[8..15]
vshr.s16 q14, q14, #8 @ a0_sign[0..7]
vshr.s16 q12, q12, #1 @ clip[8..15]
vsub.i16 q7, q7, q15 @ clip_sign[8..15] - a0_sign[8..15]
vqsub.u16 q15, q10, q3 @ a0[8..15] >= a3[8..15] ? a0[8..15]-a3[8..15] : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 q3, q3, q10 @ test a3[8..15] >= a0[8..15]
vcge.s16 q10, q10, q9 @ test a0[8..15] >= pq
vcge.s16 q9, q2, q9 @ test a0[0..7] >= pq
vsub.i16 q1, q1, q14 @ clip_sign[0..7] - a0_sign[0..7]
vqsub.u16 q14, q2, q13 @ a0[0..7] >= a3[0..7] ? a0[0..7]-a3[0..7] : 0 (a0 > a3 in all cases where filtering is enabled, so makes more sense to subtract this way round than the opposite and then taking the abs)
vcge.s16 q2, q13, q2 @ test a3[0..7] >= a0[0..7]
vmul.i16 q13, q15, d0[1] @ a0[8..15] >= a3[8..15] ? 5*(a0[8..15]-a3[8..15]) : 0
vceq.i16 q15, q11, #0 @ test clip[0..7] == 0
vmul.i16 q0, q14, d0[1] @ a0[0..7] >= a3[0..7] ? 5*(a0[0..7]-a3[0..7]) : 0
vorr q9, q15, q9 @ test clip[0..7] == 0 || a0[0..7] >= pq
vceq.i16 q14, q12, #0 @ test clip[8..15] == 0
vshr.u16 q13, q13, #3 @ a0[8..15] >= a3[8..15] ? (5*(a0[8..15]-a3[8..15]))>>3 : 0
vorr q2, q9, q2 @ test clip[0..7] == 0 || a0[0..7] >= pq || a3[0..7] >= a0[0..7]
vshr.u16 q0, q0, #3 @ a0[0..7] >= a3[0..7] ? (5*(a0[0..7]-a3[0..7]))>>3 : 0
vorr q10, q14, q10 @ test clip[8..15] == 0 || a0[8..15] >= pq
vcge.s16 q14, q13, q12
vmov.32 r2, d4[1] @ move to gp reg
vorr q3, q10, q3 @ test clip[8..15] == 0 || a0[8..15] >= pq || a3[8..15] >= a0[8..15]
vmov.32 r3, d5[1]
vcge.s16 q2, q0, q11
vbsl q14, q12, q13 @ FFMIN(d[8..15], clip[8..15])
vbsl q2, q11, q0 @ FFMIN(d[0..7], clip[0..7])
vmov.32 r5, d6[1]
vbic q0, q14, q10 @ set each d[8..15] to zero if it should not be filtered because clip[8..15] == 0 || a0[8..15] >= pq (a3 > a0 case already zeroed by saturating sub)
vmov.32 r6, d7[1]
and r12, r2, r3
vbic q2, q2, q9 @ set each d[0..7] to zero if it should not be filtered because clip[0..7] == 0 || a0[0..7] >= pq (a3 > a0 case already zeroed by saturating sub)
vmls.i16 q6, q0, q7 @ invert d[8..15] depending on clip_sign[8..15] & a0_sign[8..15], or zero it if they match, and accumulate into P4
vmls.i16 q5, q2, q1 @ invert d[0..7] depending on clip_sign[0..7] & a0_sign[0..7], or zero it if they match, and accumulate into P4
and r14, r5, r6
vmla.i16 q4, q2, q1 @ invert d[0..7] depending on clip_sign[0..7] & a0_sign[0..7], or zero it if they match, and accumulate into P5
and r12, r12, r14
vqmovun.s16 d4, q6
vmla.i16 q8, q0, q7 @ invert d[8..15] depending on clip_sign[8..15] & a0_sign[8..15], or zero it if they match, and accumulate into P5
tst r12, #1
bne 4f @ none of the 16 pixel pairs should be updated in this case
vqmovun.s16 d2, q5
vqmovun.s16 d3, q4
vqmovun.s16 d5, q8
tst r2, #1
bne 1f
vst2.8 {d2[0], d3[0]}, [r0], r1
vst2.8 {d2[1], d3[1]}, [r0], r1
vst2.8 {d2[2], d3[2]}, [r0], r1
vst2.8 {d2[3], d3[3]}, [r0]
1: add r0, r4, r1, lsl #2
tst r3, #1
bne 2f
vst2.8 {d2[4], d3[4]}, [r4], r1
vst2.8 {d2[5], d3[5]}, [r4], r1
vst2.8 {d2[6], d3[6]}, [r4], r1
vst2.8 {d2[7], d3[7]}, [r4]
2: add r4, r0, r1, lsl #2
tst r5, #1
bne 3f
vst2.8 {d4[0], d5[0]}, [r0], r1
vst2.8 {d4[1], d5[1]}, [r0], r1
vst2.8 {d4[2], d5[2]}, [r0], r1
vst2.8 {d4[3], d5[3]}, [r0]
3: tst r6, #1
bne 4f
vst2.8 {d4[4], d5[4]}, [r4], r1
vst2.8 {d4[5], d5[5]}, [r4], r1
vst2.8 {d4[6], d5[6]}, [r4], r1
vst2.8 {d4[7], d5[7]}, [r4]
4: vpop {d8-d15}
pop {r4-r6,pc}
endfunc
@ Copy at most the specified number of bytes from source to destination buffer,
@ stopping at a multiple of 16 bytes, none of which are the start of an escape sequence
@ On entry:
@ r0 -> source buffer
@ r1 = max number of bytes to copy
@ r2 -> destination buffer, optimally 8-byte aligned
@ On exit:
@ r0 = number of bytes not copied
function ff_vc1_unescape_buffer_helper_neon, export=1
@ Offset by 48 to screen out cases that are too short for us to handle,
@ and also make it easy to test for loop termination, or to determine
@ whether we need an odd number of half-iterations of the loop.
subs r1, r1, #48
bmi 90f
@ Set up useful constants
vmov.i32 q0, #0x3000000
vmov.i32 q1, #0x30000
tst r1, #16
bne 1f
vld1.8 {q8, q9}, [r0]!
vbic q12, q8, q0
vext.8 q13, q8, q9, #1
vext.8 q14, q8, q9, #2
vext.8 q15, q8, q9, #3
veor q12, q12, q1
vbic q13, q13, q0
vbic q14, q14, q0
vbic q15, q15, q0
vceq.i32 q12, q12, #0
veor q13, q13, q1
veor q14, q14, q1
veor q15, q15, q1
vceq.i32 q13, q13, #0
vceq.i32 q14, q14, #0
vceq.i32 q15, q15, #0
add r1, r1, #16
b 3f
1: vld1.8 {q10, q11}, [r0]!
vbic q12, q10, q0
vext.8 q13, q10, q11, #1
vext.8 q14, q10, q11, #2
vext.8 q15, q10, q11, #3
veor q12, q12, q1
vbic q13, q13, q0
vbic q14, q14, q0
vbic q15, q15, q0
vceq.i32 q12, q12, #0
veor q13, q13, q1
veor q14, q14, q1
veor q15, q15, q1
vceq.i32 q13, q13, #0
vceq.i32 q14, q14, #0
vceq.i32 q15, q15, #0
@ Drop through...
2: vmov q8, q11
vld1.8 {q9}, [r0]!
vorr q13, q12, q13
vorr q15, q14, q15
vbic q12, q8, q0
vorr q3, q13, q15
vext.8 q13, q8, q9, #1
vext.8 q14, q8, q9, #2
vext.8 q15, q8, q9, #3
veor q12, q12, q1
vorr d6, d6, d7
vbic q13, q13, q0
vbic q14, q14, q0
vbic q15, q15, q0
vceq.i32 q12, q12, #0
vmov r3, r12, d6
veor q13, q13, q1
veor q14, q14, q1
veor q15, q15, q1
vceq.i32 q13, q13, #0
vceq.i32 q14, q14, #0
vceq.i32 q15, q15, #0
orrs r3, r3, r12
bne 90f
vst1.64 {q10}, [r2]!
3: vmov q10, q9
vld1.8 {q11}, [r0]!
vorr q13, q12, q13
vorr q15, q14, q15
vbic q12, q10, q0
vorr q3, q13, q15
vext.8 q13, q10, q11, #1
vext.8 q14, q10, q11, #2
vext.8 q15, q10, q11, #3
veor q12, q12, q1
vorr d6, d6, d7
vbic q13, q13, q0
vbic q14, q14, q0
vbic q15, q15, q0
vceq.i32 q12, q12, #0
vmov r3, r12, d6
veor q13, q13, q1
veor q14, q14, q1
veor q15, q15, q1
vceq.i32 q13, q13, #0
vceq.i32 q14, q14, #0
vceq.i32 q15, q15, #0
orrs r3, r3, r12
bne 91f
vst1.64 {q8}, [r2]!
subs r1, r1, #32
bpl 2b
90: add r0, r1, #48
bx lr
91: sub r1, r1, #16
b 90b
endfunc