ffmpeg/libavcodec/aarch64/vp9itxfm_16bpp_neon.S

2018 lines
80 KiB
ArmAsm
Raw Normal View History

2017-01-03 12:35:54 +00:00
/*
* Copyright (c) 2017 Google Inc.
*
* 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/aarch64/asm.S"
#include "neon.S"
const itxfm4_coeffs, align=4
.short 11585, 0, 6270, 15137
iadst4_coeffs:
.short 5283, 15212, 9929, 13377
endconst
const iadst8_coeffs, align=4
.short 16305, 1606, 14449, 7723, 10394, 12665, 4756, 15679
idct_coeffs:
.short 11585, 0, 6270, 15137, 3196, 16069, 13623, 9102
.short 1606, 16305, 12665, 10394, 7723, 14449, 15679, 4756
.short 804, 16364, 12140, 11003, 7005, 14811, 15426, 5520
.short 3981, 15893, 14053, 8423, 9760, 13160, 16207, 2404
endconst
const iadst16_coeffs, align=4
.short 16364, 804, 15893, 3981, 11003, 12140, 8423, 14053
.short 14811, 7005, 13160, 9760, 5520, 15426, 2404, 16207
endconst
.macro transpose_4x4s r0, r1, r2, r3, r4, r5, r6, r7
trn1 \r4\().4s, \r0\().4s, \r1\().4s
trn2 \r5\().4s, \r0\().4s, \r1\().4s
trn1 \r6\().4s, \r2\().4s, \r3\().4s
trn2 \r7\().4s, \r2\().4s, \r3\().4s
trn1 \r0\().2d, \r4\().2d, \r6\().2d
trn2 \r2\().2d, \r4\().2d, \r6\().2d
trn1 \r1\().2d, \r5\().2d, \r7\().2d
trn2 \r3\().2d, \r5\().2d, \r7\().2d
.endm
// Transpose a 8x8 matrix of 32 bit elements, where each row is spread out
// over two registers.
.macro transpose_8x8s r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15, t0, t1, t2, t3
transpose_4x4s \r0, \r2, \r4, \r6, \t0, \t1, \t2, \t3
transpose_4x4s \r9, \r11, \r13, \r15, \t0, \t1, \t2, \t3
// Do 4x4 transposes of r1,r3,r5,r7 and r8,r10,r12,r14
// while swapping the two 4x4 matrices between each other
// First step of the 4x4 transpose of r1-r7, into t0-t3
trn1 \t0\().4s, \r1\().4s, \r3\().4s
trn2 \t1\().4s, \r1\().4s, \r3\().4s
trn1 \t2\().4s, \r5\().4s, \r7\().4s
trn2 \t3\().4s, \r5\().4s, \r7\().4s
// First step of the 4x4 transpose of r8-r12, into r1-r7
trn1 \r1\().4s, \r8\().4s, \r10\().4s
trn2 \r3\().4s, \r8\().4s, \r10\().4s
trn1 \r5\().4s, \r12\().4s, \r14\().4s
trn2 \r7\().4s, \r12\().4s, \r14\().4s
// Second step of the 4x4 transpose of r1-r7 (now in t0-r3), into r8-r12
trn1 \r8\().2d, \t0\().2d, \t2\().2d
trn2 \r12\().2d, \t0\().2d, \t2\().2d
trn1 \r10\().2d, \t1\().2d, \t3\().2d
trn2 \r14\().2d, \t1\().2d, \t3\().2d
// Second step of the 4x4 transpose of r8-r12 (now in r1-r7), in place as far as possible
trn1 \t0\().2d, \r1\().2d, \r5\().2d
trn2 \r5\().2d, \r1\().2d, \r5\().2d
trn1 \t1\().2d, \r3\().2d, \r7\().2d
trn2 \r7\().2d, \r3\().2d, \r7\().2d
// Move the outputs of trn1 back in place
mov \r1\().16b, \t0\().16b
mov \r3\().16b, \t1\().16b
.endm
// out1 = ((in1 + in2) * d0[0] + (1 << 13)) >> 14
// out2 = ((in1 - in2) * d0[0] + (1 << 13)) >> 14
// in/out are .4s registers; this can do with 4 temp registers, but is
// more efficient if 6 temp registers are available.
.macro dmbutterfly0 out1, out2, in1, in2, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, neg=0
.if \neg > 0
neg \tmp4\().4s, v0.4s
.endif
add \tmp1\().4s, \in1\().4s, \in2\().4s
sub \tmp2\().4s, \in1\().4s, \in2\().4s
.if \neg > 0
smull \tmp3\().2d, \tmp1\().2s, \tmp4\().s[0]
smull2 \tmp4\().2d, \tmp1\().4s, \tmp4\().s[0]
.else
smull \tmp3\().2d, \tmp1\().2s, v0.s[0]
smull2 \tmp4\().2d, \tmp1\().4s, v0.s[0]
.endif
.ifb \tmp5
rshrn \out1\().2s, \tmp3\().2d, #14
rshrn2 \out1\().4s, \tmp4\().2d, #14
smull \tmp3\().2d, \tmp2\().2s, v0.s[0]
smull2 \tmp4\().2d, \tmp2\().4s, v0.s[0]
rshrn \out2\().2s, \tmp3\().2d, #14
rshrn2 \out2\().4s, \tmp4\().2d, #14
.else
smull \tmp5\().2d, \tmp2\().2s, v0.s[0]
smull2 \tmp6\().2d, \tmp2\().4s, v0.s[0]
rshrn \out1\().2s, \tmp3\().2d, #14
rshrn2 \out1\().4s, \tmp4\().2d, #14
rshrn \out2\().2s, \tmp5\().2d, #14
rshrn2 \out2\().4s, \tmp6\().2d, #14
.endif
.endm
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
// Same as dmbutterfly0 above, but treating the input in in2 as zero,
// writing the same output into both out1 and out2.
.macro dmbutterfly0_h out1, out2, in1, in2, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6
smull \tmp1\().2d, \in1\().2s, v0.s[0]
smull2 \tmp2\().2d, \in1\().4s, v0.s[0]
rshrn \out1\().2s, \tmp1\().2d, #14
rshrn2 \out1\().4s, \tmp2\().2d, #14
rshrn \out2\().2s, \tmp1\().2d, #14
rshrn2 \out2\().4s, \tmp2\().2d, #14
.endm
2017-01-03 12:35:54 +00:00
// out1,out2 = in1 * coef1 - in2 * coef2
// out3,out4 = in1 * coef2 + in2 * coef1
// out are 4 x .2d registers, in are 2 x .4s registers
.macro dmbutterfly_l out1, out2, out3, out4, in1, in2, coef1, coef2
smull \out1\().2d, \in1\().2s, \coef1
smull2 \out2\().2d, \in1\().4s, \coef1
smull \out3\().2d, \in1\().2s, \coef2
smull2 \out4\().2d, \in1\().4s, \coef2
smlsl \out1\().2d, \in2\().2s, \coef2
smlsl2 \out2\().2d, \in2\().4s, \coef2
smlal \out3\().2d, \in2\().2s, \coef1
smlal2 \out4\().2d, \in2\().4s, \coef1
.endm
// inout1 = (inout1 * coef1 - inout2 * coef2 + (1 << 13)) >> 14
// inout2 = (inout1 * coef2 + inout2 * coef1 + (1 << 13)) >> 14
// inout are 2 x .4s registers
.macro dmbutterfly inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4, neg=0
dmbutterfly_l \tmp1, \tmp2, \tmp3, \tmp4, \inout1, \inout2, \coef1, \coef2
.if \neg > 0
neg \tmp3\().2d, \tmp3\().2d
neg \tmp4\().2d, \tmp4\().2d
.endif
rshrn \inout1\().2s, \tmp1\().2d, #14
rshrn2 \inout1\().4s, \tmp2\().2d, #14
rshrn \inout2\().2s, \tmp3\().2d, #14
rshrn2 \inout2\().4s, \tmp4\().2d, #14
.endm
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
// Same as dmbutterfly above, but treating the input in inout2 as zero
.macro dmbutterfly_h1 inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4
smull \tmp1\().2d, \inout1\().2s, \coef1
smull2 \tmp2\().2d, \inout1\().4s, \coef1
smull \tmp3\().2d, \inout1\().2s, \coef2
smull2 \tmp4\().2d, \inout1\().4s, \coef2
rshrn \inout1\().2s, \tmp1\().2d, #14
rshrn2 \inout1\().4s, \tmp2\().2d, #14
rshrn \inout2\().2s, \tmp3\().2d, #14
rshrn2 \inout2\().4s, \tmp4\().2d, #14
.endm
// Same as dmbutterfly above, but treating the input in inout1 as zero
.macro dmbutterfly_h2 inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4
smull \tmp1\().2d, \inout2\().2s, \coef2
smull2 \tmp2\().2d, \inout2\().4s, \coef2
smull \tmp3\().2d, \inout2\().2s, \coef1
smull2 \tmp4\().2d, \inout2\().4s, \coef1
neg \tmp1\().2d, \tmp1\().2d
neg \tmp2\().2d, \tmp2\().2d
rshrn \inout2\().2s, \tmp3\().2d, #14
rshrn2 \inout2\().4s, \tmp4\().2d, #14
rshrn \inout1\().2s, \tmp1\().2d, #14
rshrn2 \inout1\().4s, \tmp2\().2d, #14
.endm
.macro dsmull_h out1, out2, in, coef
smull \out1\().2d, \in\().2s, \coef
smull2 \out2\().2d, \in\().4s, \coef
.endm
.macro drshrn_h out, in1, in2, shift
rshrn \out\().2s, \in1\().2d, \shift
rshrn2 \out\().4s, \in2\().2d, \shift
.endm
2017-01-03 12:35:54 +00:00
// out1 = in1 + in2
// out2 = in1 - in2
.macro butterfly_4s out1, out2, in1, in2
add \out1\().4s, \in1\().4s, \in2\().4s
sub \out2\().4s, \in1\().4s, \in2\().4s
.endm
// out1 = in1 - in2
// out2 = in1 + in2
.macro butterfly_4s_r out1, out2, in1, in2
sub \out1\().4s, \in1\().4s, \in2\().4s
add \out2\().4s, \in1\().4s, \in2\().4s
.endm
// out1 = (in1,in2 + in3,in4 + (1 << 13)) >> 14
// out2 = (in1,in2 - in3,in4 + (1 << 13)) >> 14
// out are 2 x .4s registers, in are 4 x .2d registers
.macro dbutterfly_n out1, out2, in1, in2, in3, in4, tmp1, tmp2, tmp3, tmp4
add \tmp1\().2d, \in1\().2d, \in3\().2d
add \tmp2\().2d, \in2\().2d, \in4\().2d
sub \tmp3\().2d, \in1\().2d, \in3\().2d
sub \tmp4\().2d, \in2\().2d, \in4\().2d
rshrn \out1\().2s, \tmp1\().2d, #14
rshrn2 \out1\().4s, \tmp2\().2d, #14
rshrn \out2\().2s, \tmp3\().2d, #14
rshrn2 \out2\().4s, \tmp4\().2d, #14
.endm
.macro iwht4_10 c0, c1, c2, c3
add \c0\().4s, \c0\().4s, \c1\().4s
sub v17.4s, \c2\().4s, \c3\().4s
sub v16.4s, \c0\().4s, v17.4s
sshr v16.4s, v16.4s, #1
sub \c2\().4s, v16.4s, \c1\().4s
sub \c1\().4s, v16.4s, \c3\().4s
add \c3\().4s, v17.4s, \c2\().4s
sub \c0\().4s, \c0\().4s, \c1\().4s
.endm
.macro iwht4_12 c0, c1, c2, c3
iwht4_10 \c0, \c1, \c2, \c3
.endm
.macro idct4_10 c0, c1, c2, c3
mul v22.4s, \c1\().4s, v0.s[3]
mul v20.4s, \c1\().4s, v0.s[2]
add v16.4s, \c0\().4s, \c2\().4s
sub v17.4s, \c0\().4s, \c2\().4s
mla v22.4s, \c3\().4s, v0.s[2]
mul v18.4s, v16.4s, v0.s[0]
mul v24.4s, v17.4s, v0.s[0]
mls v20.4s, \c3\().4s, v0.s[3]
srshr v22.4s, v22.4s, #14
srshr v18.4s, v18.4s, #14
srshr v24.4s, v24.4s, #14
srshr v20.4s, v20.4s, #14
add \c0\().4s, v18.4s, v22.4s
sub \c3\().4s, v18.4s, v22.4s
add \c1\().4s, v24.4s, v20.4s
sub \c2\().4s, v24.4s, v20.4s
.endm
.macro idct4_12 c0, c1, c2, c3
smull v22.2d, \c1\().2s, v0.s[3]
smull2 v23.2d, \c1\().4s, v0.s[3]
smull v20.2d, \c1\().2s, v0.s[2]
smull2 v21.2d, \c1\().4s, v0.s[2]
add v16.4s, \c0\().4s, \c2\().4s
sub v17.4s, \c0\().4s, \c2\().4s
smlal v22.2d, \c3\().2s, v0.s[2]
smlal2 v23.2d, \c3\().4s, v0.s[2]
smull v18.2d, v16.2s, v0.s[0]
smull2 v19.2d, v16.4s, v0.s[0]
smull v24.2d, v17.2s, v0.s[0]
smull2 v25.2d, v17.4s, v0.s[0]
smlsl v20.2d, \c3\().2s, v0.s[3]
smlsl2 v21.2d, \c3\().4s, v0.s[3]
rshrn v22.2s, v22.2d, #14
rshrn2 v22.4s, v23.2d, #14
rshrn v18.2s, v18.2d, #14
rshrn2 v18.4s, v19.2d, #14
rshrn v24.2s, v24.2d, #14
rshrn2 v24.4s, v25.2d, #14
rshrn v20.2s, v20.2d, #14
rshrn2 v20.4s, v21.2d, #14
add \c0\().4s, v18.4s, v22.4s
sub \c3\().4s, v18.4s, v22.4s
add \c1\().4s, v24.4s, v20.4s
sub \c2\().4s, v24.4s, v20.4s
.endm
.macro iadst4_10 c0, c1, c2, c3
mul v16.4s, \c0\().4s, v1.s[0]
mla v16.4s, \c2\().4s, v1.s[1]
mla v16.4s, \c3\().4s, v1.s[2]
mul v18.4s, \c0\().4s, v1.s[2]
mls v18.4s, \c2\().4s, v1.s[0]
sub \c0\().4s, \c0\().4s, \c2\().4s
mls v18.4s, \c3\().4s, v1.s[1]
add \c0\().4s, \c0\().4s, \c3\().4s
mul v22.4s, \c1\().4s, v1.s[3]
mul v20.4s, \c0\().4s, v1.s[3]
add v24.4s, v16.4s, v22.4s
add v26.4s, v18.4s, v22.4s
srshr \c0\().4s, v24.4s, #14
add v16.4s, v16.4s, v18.4s
srshr \c1\().4s, v26.4s, #14
sub v16.4s, v16.4s, v22.4s
srshr \c2\().4s, v20.4s, #14
srshr \c3\().4s, v16.4s, #14
.endm
.macro iadst4_12 c0, c1, c2, c3
smull v16.2d, \c0\().2s, v1.s[0]
smull2 v17.2d, \c0\().4s, v1.s[0]
smlal v16.2d, \c2\().2s, v1.s[1]
smlal2 v17.2d, \c2\().4s, v1.s[1]
smlal v16.2d, \c3\().2s, v1.s[2]
smlal2 v17.2d, \c3\().4s, v1.s[2]
smull v18.2d, \c0\().2s, v1.s[2]
smull2 v19.2d, \c0\().4s, v1.s[2]
smlsl v18.2d, \c2\().2s, v1.s[0]
smlsl2 v19.2d, \c2\().4s, v1.s[0]
sub \c0\().4s, \c0\().4s, \c2\().4s
smlsl v18.2d, \c3\().2s, v1.s[1]
smlsl2 v19.2d, \c3\().4s, v1.s[1]
add \c0\().4s, \c0\().4s, \c3\().4s
smull v22.2d, \c1\().2s, v1.s[3]
smull2 v23.2d, \c1\().4s, v1.s[3]
smull v20.2d, \c0\().2s, v1.s[3]
smull2 v21.2d, \c0\().4s, v1.s[3]
add v24.2d, v16.2d, v22.2d
add v25.2d, v17.2d, v23.2d
add v26.2d, v18.2d, v22.2d
add v27.2d, v19.2d, v23.2d
rshrn \c0\().2s, v24.2d, #14
rshrn2 \c0\().4s, v25.2d, #14
add v16.2d, v16.2d, v18.2d
add v17.2d, v17.2d, v19.2d
rshrn \c1\().2s, v26.2d, #14
rshrn2 \c1\().4s, v27.2d, #14
sub v16.2d, v16.2d, v22.2d
sub v17.2d, v17.2d, v23.2d
rshrn \c2\().2s, v20.2d, #14
rshrn2 \c2\().4s, v21.2d, #14
rshrn \c3\().2s, v16.2d, #14
rshrn2 \c3\().4s, v17.2d, #14
.endm
// The public functions in this file have got the following signature:
// void itxfm_add(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
.macro itxfm_func4x4 txfm1, txfm2, bpp
function ff_vp9_\txfm1\()_\txfm2\()_4x4_add_\bpp\()_neon, export=1
.ifc \txfm1,\txfm2
.ifc \txfm1,idct
movrel x4, itxfm4_coeffs
ld1 {v0.4h}, [x4]
sxtl v0.4s, v0.4h
.endif
.ifc \txfm1,iadst
movrel x4, iadst4_coeffs
ld1 {v0.d}[1], [x4]
sxtl2 v1.4s, v0.8h
.endif
.else
movrel x4, itxfm4_coeffs
ld1 {v0.8h}, [x4]
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
.endif
movi v30.4s, #0
movi v31.4s, #0
.ifc \txfm1\()_\txfm2,idct_idct
cmp w3, #1
b.ne 1f
// DC-only for idct/idct
ld1 {v2.s}[0], [x2]
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
st1 {v31.s}[0], [x2]
dup v4.4s, v2.s[0]
mov v5.16b, v4.16b
mov v6.16b, v4.16b
mov v7.16b, v4.16b
b 2f
.endif
1:
ld1 {v4.4s,v5.4s,v6.4s,v7.4s}, [x2]
st1 {v30.4s,v31.4s}, [x2], #32
.ifc \txfm1,iwht
sshr v4.4s, v4.4s, #2
sshr v5.4s, v5.4s, #2
sshr v6.4s, v6.4s, #2
sshr v7.4s, v7.4s, #2
.endif
\txfm1\()4_\bpp v4, v5, v6, v7
st1 {v30.4s,v31.4s}, [x2], #32
// Transpose 4x4 with 32 bit elements
transpose_4x4s v4, v5, v6, v7, v16, v17, v18, v19
\txfm2\()4_\bpp v4, v5, v6, v7
2:
mvni v31.8h, #((0xff << (\bpp - 8)) & 0xff), lsl #8
ld1 {v0.4h}, [x0], x1
ld1 {v1.4h}, [x0], x1
.ifnc \txfm1,iwht
srshr v4.4s, v4.4s, #4
srshr v5.4s, v5.4s, #4
srshr v6.4s, v6.4s, #4
srshr v7.4s, v7.4s, #4
.endif
uaddw v4.4s, v4.4s, v0.4h
uaddw v5.4s, v5.4s, v1.4h
ld1 {v2.4h}, [x0], x1
ld1 {v3.4h}, [x0], x1
sqxtun v0.4h, v4.4s
sqxtun2 v0.8h, v5.4s
sub x0, x0, x1, lsl #2
uaddw v6.4s, v6.4s, v2.4h
umin v0.8h, v0.8h, v31.8h
uaddw v7.4s, v7.4s, v3.4h
st1 {v0.4h}, [x0], x1
sqxtun v2.4h, v6.4s
sqxtun2 v2.8h, v7.4s
umin v2.8h, v2.8h, v31.8h
st1 {v0.d}[1], [x0], x1
st1 {v2.4h}, [x0], x1
st1 {v2.d}[1], [x0], x1
ret
endfunc
.endm
.macro itxfm_funcs4x4 bpp
itxfm_func4x4 idct, idct, \bpp
itxfm_func4x4 iadst, idct, \bpp
itxfm_func4x4 idct, iadst, \bpp
itxfm_func4x4 iadst, iadst, \bpp
itxfm_func4x4 iwht, iwht, \bpp
.endm
itxfm_funcs4x4 10
itxfm_funcs4x4 12
function idct8x8_dc_add_neon
movrel x4, idct_coeffs
ld1 {v0.4h}, [x4]
movi v1.4h, #0
sxtl v0.4s, v0.4h
ld1 {v2.s}[0], [x2]
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
st1 {v1.s}[0], [x2]
dup v2.4s, v2.s[0]
srshr v2.4s, v2.4s, #5
mov x4, #8
mov x3, x0
dup v31.8h, w5
1:
// Loop to add the constant from v2 into all 8x8 outputs
subs x4, x4, #2
ld1 {v3.8h}, [x0], x1
ld1 {v4.8h}, [x0], x1
uaddw v16.4s, v2.4s, v3.4h
uaddw2 v17.4s, v2.4s, v3.8h
uaddw v18.4s, v2.4s, v4.4h
uaddw2 v19.4s, v2.4s, v4.8h
sqxtun v3.4h, v16.4s
sqxtun2 v3.8h, v17.4s
sqxtun v4.4h, v18.4s
sqxtun2 v4.8h, v19.4s
umin v3.8h, v3.8h, v31.8h
umin v4.8h, v4.8h, v31.8h
st1 {v3.8h}, [x3], x1
st1 {v4.8h}, [x3], x1
b.ne 1b
ret
endfunc
.macro idct8 r0, r1, r2, r3, r4, r5, r6, r7, t0, t1, t2, t3, t4, t5
dmbutterfly0 \r0, \r4, \r0, \r4, \t0, \t1, \t2, \t3, \t4, \t5 // r0 = t0a, r4 = t1a
dmbutterfly \r2, \r6, v0.s[2], v0.s[3], \t0, \t1, \t2, \t3 // r2 = t2a, r6 = t3a
dmbutterfly \r1, \r7, v1.s[0], v1.s[1], \t0, \t1, \t2, \t3 // r1 = t4a, r7 = t7a
dmbutterfly \r5, \r3, v1.s[2], v1.s[3], \t0, \t1, \t2, \t3 // r5 = t5a, r3 = t6a
butterfly_4s \t0, \t1, \r0, \r6 // t0 = t0, t1 = t3
butterfly_4s \t2, \r5, \r1, \r5 // t2 = t4, r5 = t5a
butterfly_4s \t3, \r6, \r7, \r3 // t3 = t7, r6 = t6a
butterfly_4s \r7, \r4, \r4, \r2 // r7 = t1, r4 = t2
dmbutterfly0 \r6, \r5, \r6, \r5, \r0, \r1, \r2, \r3, \t4, \t5 // r6 = t6, r5 = t5
butterfly_4s \r1, \r6, \r7, \r6 // r1 = out[1], r6 = out[6]
butterfly_4s \r0, \r7, \t0, \t3 // r0 = out[0], r7 = out[7]
butterfly_4s \r2, \r5, \r4, \r5 // r2 = out[2], r5 = out[5]
butterfly_4s \r3, \r4, \t1, \t2 // r3 = out[3], r4 = out[4]
.endm
.macro iadst8 r0, r1, r2, r3, r4, r5, r6, r7, t0, t1, t2, t3, t4, t5
dmbutterfly_l \t2, \t3, \t0, \t1, \r7, \r0, v2.s[1], v2.s[0] // t2,t3 = t1a, t0,t1 = t0a
dmbutterfly_l \r0, \r7, \t4, \t5, \r3, \r4, v3.s[1], v3.s[0] // r0,r7 = t5a, t4,t5 = t4a
dbutterfly_n \r3, \t0, \t0, \t1, \t4, \t5, \r3, \r4, \t0, \t1 // r3 = t0, t0 = t4
dbutterfly_n \r4, \t1, \t2, \t3, \r0, \r7, \r4, \t1, \t4, \t5 // r4 = t1, t1 = t5
dmbutterfly_l \t4, \t5, \t2, \t3, \r5, \r2, v2.s[3], v2.s[2] // t4,t5 = t3a, t2,t3 = t2a
dmbutterfly_l \r2, \r5, \r0, \r7, \r1, \r6, v3.s[3], v3.s[2] // r2,r5 = t7a, r0,r7 = t6a
dbutterfly_n \r1, \t2, \t2, \t3, \r0, \r7, \r1, \r6, \t2, \t3 // r1 = t2, t2 = t6
dbutterfly_n \r0, \t4, \t4, \t5, \r2, \r5, \r0, \r7, \t4, \t5 // r0 = t3, t4 = t7
butterfly_4s \r7, \r4, \r4, \r0 // r7 = -out[7], r4 = t3
neg \r7\().4s, \r7\().4s // r7 = out[7]
butterfly_4s \r0, \r1, \r3, \r1 // r0 = out[0], r1 = t2
dmbutterfly_l \r2, \r3, \t3, \t5, \t0, \t1, v0.s[2], v0.s[3] // r2,r3 = t5a, t3,t5 = t4a
dmbutterfly_l \t0, \t1, \r5, \r6, \t4, \t2, v0.s[3], v0.s[2] // t0,t1 = t6a, r5,r6 = t7a
dbutterfly_n \r6, \t2, \r2, \r3, \r5, \r6, \t2, \t4, \r2, \r3 // r6 = out[6], t2 = t7
dmbutterfly0 \r3, \r4, \r1, \r4, \t4, \r5, \r1, \r2 // r3 = -out[3], r4 = out[4]
neg \r3\().4s, \r3\().4s // r3 = out[3]
dbutterfly_n \r1, \t0, \t3, \t5, \t0, \t1, \r1, \r2, \t0, \t1 // r1 = -out[1], t0 = t6
neg \r1\().4s, \r1\().4s // r1 = out[1]
dmbutterfly0 \r2, \r5, \t0, \t2, \t1, \t3, \t4, \t5 // r2 = out[2], r5 = -out[5]
neg \r5\().4s, \r5\().4s // r5 = out[5]
.endm
.macro itxfm_func8x8 txfm1, txfm2
function vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
.ifc \txfm1\()_\txfm2,idct_idct
cmp w3, #1
b.eq idct8x8_dc_add_neon
.endif
// The iadst also uses a few coefficients from
// idct, so those always need to be loaded.
.ifc \txfm1\()_\txfm2,idct_idct
movrel x4, idct_coeffs
.else
movrel x4, iadst8_coeffs
ld1 {v1.8h}, [x4], #16
stp d8, d9, [sp, #-0x10]!
sxtl2 v3.4s, v1.8h
sxtl v2.4s, v1.4h
.endif
ld1 {v0.8h}, [x4]
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
movi v4.4s, #0
movi v5.4s, #0
movi v6.4s, #0
movi v7.4s, #0
1:
ld1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x2], #64
ld1 {v20.4s,v21.4s,v22.4s,v23.4s}, [x2], #64
ld1 {v24.4s,v25.4s,v26.4s,v27.4s}, [x2], #64
ld1 {v28.4s,v29.4s,v30.4s,v31.4s}, [x2], #64
sub x2, x2, #256
st1 {v4.4s,v5.4s,v6.4s,v7.4s}, [x2], #64
st1 {v4.4s,v5.4s,v6.4s,v7.4s}, [x2], #64
st1 {v4.4s,v5.4s,v6.4s,v7.4s}, [x2], #64
st1 {v4.4s,v5.4s,v6.4s,v7.4s}, [x2], #64
.ifc \txfm1\()_\txfm2,idct_idct
idct8 v16, v18, v20, v22, v24, v26, v28, v30, v2, v3, v4, v5, v6, v7
idct8 v17, v19, v21, v23, v25, v27, v29, v31, v2, v3, v4, v5, v6, v7
.else
\txfm1\()8 v16, v18, v20, v22, v24, v26, v28, v30, v4, v5, v6, v7, v8, v9
\txfm1\()8 v17, v19, v21, v23, v25, v27, v29, v31, v4, v5, v6, v7, v8, v9
.endif
// Transpose 8x8 with 16 bit elements
transpose_8x8s v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v4, v5, v6, v7
.ifc \txfm1\()_\txfm2,idct_idct
idct8 v16, v18, v20, v22, v24, v26, v28, v30, v2, v3, v4, v5, v6, v7
idct8 v17, v19, v21, v23, v25, v27, v29, v31, v2, v3, v4, v5, v6, v7
.else
\txfm2\()8 v16, v18, v20, v22, v24, v26, v28, v30, v4, v5, v6, v7, v8, v9
\txfm2\()8 v17, v19, v21, v23, v25, v27, v29, v31, v4, v5, v6, v7, v8, v9
.endif
2:
mov x3, x0
// Add into the destination
ld1 {v0.8h}, [x0], x1
srshr v16.4s, v16.4s, #5
srshr v17.4s, v17.4s, #5
ld1 {v1.8h}, [x0], x1
srshr v18.4s, v18.4s, #5
srshr v19.4s, v19.4s, #5
ld1 {v2.8h}, [x0], x1
srshr v20.4s, v20.4s, #5
srshr v21.4s, v21.4s, #5
uaddw v16.4s, v16.4s, v0.4h
uaddw2 v17.4s, v17.4s, v0.8h
ld1 {v3.8h}, [x0], x1
srshr v22.4s, v22.4s, #5
srshr v23.4s, v23.4s, #5
uaddw v18.4s, v18.4s, v1.4h
uaddw2 v19.4s, v19.4s, v1.8h
ld1 {v4.8h}, [x0], x1
srshr v24.4s, v24.4s, #5
srshr v25.4s, v25.4s, #5
uaddw v20.4s, v20.4s, v2.4h
uaddw2 v21.4s, v21.4s, v2.8h
sqxtun v0.4h, v16.4s
sqxtun2 v0.8h, v17.4s
dup v16.8h, w5
ld1 {v5.8h}, [x0], x1
srshr v26.4s, v26.4s, #5
srshr v27.4s, v27.4s, #5
uaddw v22.4s, v22.4s, v3.4h
uaddw2 v23.4s, v23.4s, v3.8h
sqxtun v1.4h, v18.4s
sqxtun2 v1.8h, v19.4s
umin v0.8h, v0.8h, v16.8h
ld1 {v6.8h}, [x0], x1
srshr v28.4s, v28.4s, #5
srshr v29.4s, v29.4s, #5
uaddw v24.4s, v24.4s, v4.4h
uaddw2 v25.4s, v25.4s, v4.8h
sqxtun v2.4h, v20.4s
sqxtun2 v2.8h, v21.4s
umin v1.8h, v1.8h, v16.8h
ld1 {v7.8h}, [x0], x1
srshr v30.4s, v30.4s, #5
srshr v31.4s, v31.4s, #5
uaddw v26.4s, v26.4s, v5.4h
uaddw2 v27.4s, v27.4s, v5.8h
sqxtun v3.4h, v22.4s
sqxtun2 v3.8h, v23.4s
umin v2.8h, v2.8h, v16.8h
st1 {v0.8h}, [x3], x1
uaddw v28.4s, v28.4s, v6.4h
uaddw2 v29.4s, v29.4s, v6.8h
st1 {v1.8h}, [x3], x1
sqxtun v4.4h, v24.4s
sqxtun2 v4.8h, v25.4s
umin v3.8h, v3.8h, v16.8h
st1 {v2.8h}, [x3], x1
uaddw v30.4s, v30.4s, v7.4h
uaddw2 v31.4s, v31.4s, v7.8h
st1 {v3.8h}, [x3], x1
sqxtun v5.4h, v26.4s
sqxtun2 v5.8h, v27.4s
umin v4.8h, v4.8h, v16.8h
st1 {v4.8h}, [x3], x1
sqxtun v6.4h, v28.4s
sqxtun2 v6.8h, v29.4s
umin v5.8h, v5.8h, v16.8h
st1 {v5.8h}, [x3], x1
sqxtun v7.4h, v30.4s
sqxtun2 v7.8h, v31.4s
umin v6.8h, v6.8h, v16.8h
st1 {v6.8h}, [x3], x1
umin v7.8h, v7.8h, v16.8h
st1 {v7.8h}, [x3], x1
.ifnc \txfm1\()_\txfm2,idct_idct
ldp d8, d9, [sp], 0x10
.endif
ret
endfunc
function ff_vp9_\txfm1\()_\txfm2\()_8x8_add_10_neon, export=1
mov x5, #0x03ff
b vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
endfunc
function ff_vp9_\txfm1\()_\txfm2\()_8x8_add_12_neon, export=1
mov x5, #0x0fff
b vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
endfunc
.endm
itxfm_func8x8 idct, idct
itxfm_func8x8 iadst, idct
itxfm_func8x8 idct, iadst
itxfm_func8x8 iadst, iadst
function idct16x16_dc_add_neon
movrel x4, idct_coeffs
ld1 {v0.4h}, [x4]
sxtl v0.4s, v0.4h
movi v1.4h, #0
ld1 {v2.s}[0], [x2]
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
st1 {v1.s}[0], [x2]
dup v2.4s, v2.s[0]
srshr v0.4s, v2.4s, #6
mov x3, x0
mov x4, #16
dup v31.8h, w13
1:
// Loop to add the constant from v2 into all 16x16 outputs
subs x4, x4, #2
ld1 {v1.8h,v2.8h}, [x0], x1
uaddw v16.4s, v0.4s, v1.4h
uaddw2 v17.4s, v0.4s, v1.8h
ld1 {v3.8h,v4.8h}, [x0], x1
uaddw v18.4s, v0.4s, v2.4h
uaddw2 v19.4s, v0.4s, v2.8h
uaddw v20.4s, v0.4s, v3.4h
uaddw2 v21.4s, v0.4s, v3.8h
uaddw v22.4s, v0.4s, v4.4h
uaddw2 v23.4s, v0.4s, v4.8h
sqxtun v1.4h, v16.4s
sqxtun2 v1.8h, v17.4s
sqxtun v2.4h, v18.4s
sqxtun2 v2.8h, v19.4s
sqxtun v3.4h, v20.4s
sqxtun2 v3.8h, v21.4s
sqxtun v4.4h, v22.4s
sqxtun2 v4.8h, v23.4s
umin v1.8h, v1.8h, v31.8h
umin v2.8h, v2.8h, v31.8h
st1 {v1.8h,v2.8h}, [x3], x1
umin v3.8h, v3.8h, v31.8h
umin v4.8h, v4.8h, v31.8h
st1 {v3.8h,v4.8h}, [x3], x1
b.ne 1b
ret
endfunc
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.macro idct16_end
butterfly_4s v18, v7, v4, v7 // v18 = t0a, v7 = t7a
butterfly_4s v19, v22, v5, v22 // v19 = t1a, v22 = t6
butterfly_4s v4, v26, v20, v26 // v4 = t2a, v26 = t5
butterfly_4s v5, v6, v28, v6 // v5 = t3a, v6 = t4
butterfly_4s v20, v28, v16, v24 // v20 = t8a, v28 = t11a
butterfly_4s v24, v21, v23, v21 // v24 = t9, v21 = t10
butterfly_4s v23, v27, v25, v27 // v23 = t14, v27 = t13
butterfly_4s v25, v29, v29, v17 // v25 = t15a, v29 = t12a
dmbutterfly0 v8, v9, v27, v21, v8, v9, v16, v17, v30, v31 // v8 = t13a, v9 = t10a
dmbutterfly0 v28, v27, v29, v28, v21, v29, v16, v17, v30, v31 // v28 = t12, v27 = t11
butterfly_4s v16, v31, v18, v25 // v16 = out[0], v31 = out[15]
butterfly_4s v17, v30, v19, v23 // v17 = out[1], v30 = out[14]
butterfly_4s_r v25, v22, v22, v24 // v25 = out[9], v22 = out[6]
butterfly_4s v23, v24, v7, v20 // v23 = out[7], v24 = out[8]
butterfly_4s v18, v29, v4, v8 // v18 = out[2], v29 = out[13]
butterfly_4s v19, v28, v5, v28 // v19 = out[3], v28 = out[12]
butterfly_4s v20, v27, v6, v27 // v20 = out[4], v27 = out[11]
butterfly_4s v21, v26, v26, v9 // v21 = out[5], v26 = out[10]
ret
.endm
function idct16
2017-01-03 12:35:54 +00:00
dmbutterfly0 v16, v24, v16, v24, v4, v5, v6, v7, v8, v9 // v16 = t0a, v24 = t1a
dmbutterfly v20, v28, v0.s[2], v0.s[3], v4, v5, v6, v7 // v20 = t2a, v28 = t3a
dmbutterfly v18, v30, v1.s[0], v1.s[1], v4, v5, v6, v7 // v18 = t4a, v30 = t7a
dmbutterfly v26, v22, v1.s[2], v1.s[3], v4, v5, v6, v7 // v26 = t5a, v22 = t6a
dmbutterfly v17, v31, v2.s[0], v2.s[1], v4, v5, v6, v7 // v17 = t8a, v31 = t15a
dmbutterfly v25, v23, v2.s[2], v2.s[3], v4, v5, v6, v7 // v25 = t9a, v23 = t14a
dmbutterfly v21, v27, v3.s[0], v3.s[1], v4, v5, v6, v7 // v21 = t10a, v27 = t13a
dmbutterfly v29, v19, v3.s[2], v3.s[3], v4, v5, v6, v7 // v29 = t11a, v19 = t12a
butterfly_4s v4, v28, v16, v28 // v4 = t0, v28 = t3
butterfly_4s v5, v20, v24, v20 // v5 = t1, v20 = t2
butterfly_4s v6, v26, v18, v26 // v6 = t4, v26 = t5
butterfly_4s v7, v22, v30, v22 // v7 = t7, v22 = t6
butterfly_4s v16, v25, v17, v25 // v16 = t8, v25 = t9
butterfly_4s v24, v21, v29, v21 // v24 = t11, v21 = t10
butterfly_4s v17, v27, v19, v27 // v17 = t12, v27 = t13
butterfly_4s v29, v23, v31, v23 // v29 = t15, v23 = t14
dmbutterfly0 v22, v26, v22, v26, v8, v9, v18, v19, v30, v31 // v22 = t6a, v26 = t5a
dmbutterfly v23, v25, v0.s[2], v0.s[3], v18, v19, v30, v31 // v23 = t9a, v25 = t14a
dmbutterfly v27, v21, v0.s[2], v0.s[3], v18, v19, v30, v31, neg=1 // v27 = t13a, v21 = t10a
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
idct16_end
endfunc
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
function idct16_half
dmbutterfly0_h v16, v24, v16, v24, v4, v5, v6, v7, v8, v9 // v16 = t0a, v24 = t1a
dmbutterfly_h1 v20, v28, v0.s[2], v0.s[3], v4, v5, v6, v7 // v20 = t2a, v28 = t3a
dmbutterfly_h1 v18, v30, v1.s[0], v1.s[1], v4, v5, v6, v7 // v18 = t4a, v30 = t7a
dmbutterfly_h2 v26, v22, v1.s[2], v1.s[3], v4, v5, v6, v7 // v26 = t5a, v22 = t6a
dmbutterfly_h1 v17, v31, v2.s[0], v2.s[1], v4, v5, v6, v7 // v17 = t8a, v31 = t15a
dmbutterfly_h2 v25, v23, v2.s[2], v2.s[3], v4, v5, v6, v7 // v25 = t9a, v23 = t14a
dmbutterfly_h1 v21, v27, v3.s[0], v3.s[1], v4, v5, v6, v7 // v21 = t10a, v27 = t13a
dmbutterfly_h2 v29, v19, v3.s[2], v3.s[3], v4, v5, v6, v7 // v29 = t11a, v19 = t12a
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
butterfly_4s v4, v28, v16, v28 // v4 = t0, v28 = t3
butterfly_4s v5, v20, v24, v20 // v5 = t1, v20 = t2
butterfly_4s v6, v26, v18, v26 // v6 = t4, v26 = t5
butterfly_4s v7, v22, v30, v22 // v7 = t7, v22 = t6
butterfly_4s v16, v25, v17, v25 // v16 = t8, v25 = t9
butterfly_4s v24, v21, v29, v21 // v24 = t11, v21 = t10
butterfly_4s v17, v27, v19, v27 // v17 = t12, v27 = t13
butterfly_4s v29, v23, v31, v23 // v29 = t15, v23 = t14
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
dmbutterfly0 v22, v26, v22, v26, v8, v9, v18, v19, v30, v31 // v22 = t6a, v26 = t5a
dmbutterfly v23, v25, v0.s[2], v0.s[3], v18, v19, v30, v31 // v23 = t9a, v25 = t14a
dmbutterfly v27, v21, v0.s[2], v0.s[3], v18, v19, v30, v31, neg=1 // v27 = t13a, v21 = t10a
idct16_end
endfunc
function idct16_quarter
dsmull_h v24, v25, v19, v3.s[3]
dsmull_h v4, v5, v17, v2.s[0]
dsmull_h v7, v6, v18, v1.s[1]
dsmull_h v30, v31, v18, v1.s[0]
neg v24.2d, v24.2d
neg v25.2d, v25.2d
dsmull_h v29, v28, v17, v2.s[1]
dsmull_h v26, v27, v19, v3.s[2]
dsmull_h v22, v23, v16, v0.s[0]
drshrn_h v24, v24, v25, #14
drshrn_h v16, v4, v5, #14
drshrn_h v7, v7, v6, #14
drshrn_h v6, v30, v31, #14
drshrn_h v29, v29, v28, #14
drshrn_h v17, v26, v27, #14
drshrn_h v28, v22, v23, #14
dmbutterfly_l v20, v21, v22, v23, v17, v24, v0.s[2], v0.s[3]
dmbutterfly_l v18, v19, v30, v31, v29, v16, v0.s[2], v0.s[3]
neg v22.2d, v22.2d
neg v23.2d, v23.2d
drshrn_h v27, v20, v21, #14
drshrn_h v21, v22, v23, #14
drshrn_h v23, v18, v19, #14
drshrn_h v25, v30, v31, #14
mov v4.16b, v28.16b
mov v5.16b, v28.16b
dmbutterfly0 v22, v26, v7, v6, v18, v19, v30, v31
mov v20.16b, v28.16b
idct16_end
endfunc
2017-01-03 12:35:54 +00:00
function iadst16
2017-01-03 12:35:54 +00:00
ld1 {v0.8h,v1.8h}, [x11]
sxtl v2.4s, v1.4h
sxtl2 v3.4s, v1.8h
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
dmbutterfly_l v6, v7, v4, v5, v31, v16, v0.s[1], v0.s[0] // v6,v7 = t1, v4,v5 = t0
dmbutterfly_l v10, v11, v8, v9, v23, v24, v1.s[1], v1.s[0] // v10,v11 = t9, v8,v9 = t8
dbutterfly_n v31, v24, v6, v7, v10, v11, v12, v13, v10, v11 // v31 = t1a, v24 = t9a
dmbutterfly_l v14, v15, v12, v13, v29, v18, v0.s[3], v0.s[2] // v14,v15 = t3, v12,v13 = t2
dbutterfly_n v16, v23, v4, v5, v8, v9, v6, v7, v8, v9 // v16 = t0a, v23 = t8a
dmbutterfly_l v6, v7, v4, v5, v21, v26, v1.s[3], v1.s[2] // v6,v7 = t11, v4,v5 = t10
dbutterfly_n v29, v26, v14, v15, v6, v7, v8, v9, v6, v7 // v29 = t3a, v26 = t11a
dmbutterfly_l v10, v11, v8, v9, v27, v20, v2.s[1], v2.s[0] // v10,v11 = t5, v8,v9 = t4
dbutterfly_n v18, v21, v12, v13, v4, v5, v6, v7, v4, v5 // v18 = t2a, v21 = t10a
dmbutterfly_l v14, v15, v12, v13, v19, v28, v3.s[1], v3.s[0] // v14,v15 = t13, v12,v13 = t12
dbutterfly_n v20, v28, v10, v11, v14, v15, v4, v5, v14, v15 // v20 = t5a, v28 = t13a
dmbutterfly_l v6, v7, v4, v5, v25, v22, v2.s[3], v2.s[2] // v6,v7 = t7, v4,v5 = t6
dbutterfly_n v27, v19, v8, v9, v12, v13, v10, v11, v12, v13 // v27 = t4a, v19 = t12a
dmbutterfly_l v10, v11, v8, v9, v17, v30, v3.s[3], v3.s[2] // v10,v11 = t15, v8,v9 = t14
ld1 {v0.8h}, [x10]
dbutterfly_n v22, v30, v6, v7, v10, v11, v12, v13, v10, v11 // v22 = t7a, v30 = t15a
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
dmbutterfly_l v14, v15, v12, v13, v23, v24, v1.s[0], v1.s[1] // v14,v15 = t9, v12,v13 = t8
dbutterfly_n v25, v17, v4, v5, v8, v9, v6, v7, v8, v9 // v25 = t6a, v17 = t14a
dmbutterfly_l v4, v5, v6, v7, v28, v19, v1.s[1], v1.s[0] // v4,v5 = t12, v6,v7 = t13
dbutterfly_n v23, v19, v12, v13, v4, v5, v8, v9, v4, v5 // v23 = t8a, v19 = t12a
dmbutterfly_l v10, v11, v8, v9, v21, v26, v1.s[2], v1.s[3] // v10,v11 = t11, v8,v9 = t10
butterfly_4s_r v4, v27, v16, v27 // v4 = t4, v27 = t0
dbutterfly_n v24, v28, v14, v15, v6, v7, v12, v13, v6, v7 // v24 = t9a, v28 = t13a
dmbutterfly_l v12, v13, v14, v15, v30, v17, v1.s[3], v1.s[2] // v12,v13 = t14, v14,v15 = t15
butterfly_4s_r v5, v20, v31, v20 // v5 = t5, v20 = t1
dbutterfly_n v21, v17, v8, v9, v12, v13, v6, v7, v12, v13 // v21 = t10a, v17 = t14a
dbutterfly_n v26, v30, v10, v11, v14, v15, v8, v9, v14, v15 // v26 = t11a, v30 = t15a
butterfly_4s_r v6, v25, v18, v25 // v6 = t6, v25 = t2
butterfly_4s_r v7, v22, v29, v22 // v7 = t7, v22 = t3
dmbutterfly_l v10, v11, v8, v9, v19, v28, v0.s[2], v0.s[3] // v10,v11 = t13, v8,v9 = t12
dmbutterfly_l v12, v13, v14, v15, v30, v17, v0.s[3], v0.s[2] // v12,v13 = t14, v14,v15 = t15
dbutterfly_n v18, v30, v8, v9, v12, v13, v16, v17, v12, v13 // v18 = out[2], v30 = t14a
dbutterfly_n v29, v17, v10, v11, v14, v15, v12, v13, v14, v15 // v29 = -out[13], v17 = t15a
neg v29.4s, v29.4s // v29 = out[13]
dmbutterfly_l v10, v11, v8, v9, v4, v5, v0.s[2], v0.s[3] // v10,v11 = t5a, v8,v9 = t4a
dmbutterfly_l v12, v13, v14, v15, v7, v6, v0.s[3], v0.s[2] // v12,v13 = t6a, v14,v15 = t7a
butterfly_4s v2, v6, v27, v25 // v2 = out[0], v6 = t2a
butterfly_4s v3, v7, v23, v21 // v3 =-out[1], v7 = t10
dbutterfly_n v19, v31, v8, v9, v12, v13, v4, v5, v8, v9 // v19 = -out[3], v31 = t6
neg v19.4s, v19.4s // v19 = out[3]
dbutterfly_n v28, v16, v10, v11, v14, v15, v4, v5, v10, v11 // v28 = out[12], v16 = t7
butterfly_4s v5, v8, v20, v22 // v5 =-out[15],v8 = t3a
butterfly_4s v4, v9, v24, v26 // v4 = out[14],v9 = t11
dmbutterfly0 v23, v24, v6, v8, v10, v11, v12, v13, v14, v15, 1 // v23 = out[7], v24 = out[8]
dmbutterfly0 v21, v26, v30, v17, v10, v11, v12, v13, v14, v15, 1 // v21 = out[5], v26 = out[10]
dmbutterfly0 v20, v27, v16, v31, v10, v11, v12, v13, v14, v15 // v20 = out[4], v27 = out[11]
dmbutterfly0 v22, v25, v9, v7, v10, v11, v12, v13, v14, v15 // v22 = out[6], v25 = out[9]
neg v31.4s, v5.4s // v31 = out[15]
neg v17.4s, v3.4s // v17 = out[1]
mov v16.16b, v2.16b
mov v30.16b, v4.16b
ret
endfunc
2017-01-03 12:35:54 +00:00
// Helper macros; we can't use these expressions directly within
// e.g. .irp due to the extra concatenation \(). Therefore wrap
// them in macros to allow using .irp below.
.macro load i, src, inc
ld1 {v\i\().4s}, [\src], \inc
.endm
.macro store i, dst, inc
st1 {v\i\().4s}, [\dst], \inc
.endm
.macro movi_v i, size, imm
movi v\i\()\size, \imm
.endm
.macro load_clear i, src, inc
ld1 {v\i\().4s}, [\src]
st1 {v4.4s}, [\src], \inc
.endm
.macro load_add_store coef0, coef1, coef2, coef3, coef4, coef5, coef6, coef7
srshr \coef0, \coef0, #6
ld1 {v4.4h}, [x0], x1
srshr \coef1, \coef1, #6
ld1 {v4.d}[1], [x3], x1
srshr \coef2, \coef2, #6
ld1 {v5.4h}, [x0], x1
srshr \coef3, \coef3, #6
uaddw \coef0, \coef0, v4.4h
ld1 {v5.d}[1], [x3], x1
srshr \coef4, \coef4, #6
uaddw2 \coef1, \coef1, v4.8h
ld1 {v6.4h}, [x0], x1
srshr \coef5, \coef5, #6
uaddw \coef2, \coef2, v5.4h
ld1 {v6.d}[1], [x3], x1
sqxtun v4.4h, \coef0
srshr \coef6, \coef6, #6
uaddw2 \coef3, \coef3, v5.8h
ld1 {v7.4h}, [x0], x1
sqxtun2 v4.8h, \coef1
srshr \coef7, \coef7, #6
uaddw \coef4, \coef4, v6.4h
ld1 {v7.d}[1], [x3], x1
umin v4.8h, v4.8h, v8.8h
sub x0, x0, x1, lsl #2
sub x3, x3, x1, lsl #2
sqxtun v5.4h, \coef2
uaddw2 \coef5, \coef5, v6.8h
st1 {v4.4h}, [x0], x1
sqxtun2 v5.8h, \coef3
uaddw \coef6, \coef6, v7.4h
st1 {v4.d}[1], [x3], x1
umin v5.8h, v5.8h, v8.8h
sqxtun v6.4h, \coef4
uaddw2 \coef7, \coef7, v7.8h
st1 {v5.4h}, [x0], x1
sqxtun2 v6.8h, \coef5
st1 {v5.d}[1], [x3], x1
umin v6.8h, v6.8h, v8.8h
sqxtun v7.4h, \coef6
st1 {v6.4h}, [x0], x1
sqxtun2 v7.8h, \coef7
st1 {v6.d}[1], [x3], x1
umin v7.8h, v7.8h, v8.8h
st1 {v7.4h}, [x0], x1
st1 {v7.d}[1], [x3], x1
.endm
2017-01-03 12:35:54 +00:00
// Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
// transpose into a horizontal 16x4 slice and store.
// x0 = dst (temp buffer)
// x1 = slice offset
// x2 = src
// x9 = input stride
.macro itxfm16_1d_funcs txfm
function \txfm\()16_1d_4x16_pass1_neon
mov x14, x30
2017-01-03 12:35:54 +00:00
movi v4.4s, #0
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
load_clear \i, x2, x9
.endr
bl \txfm\()16
2017-01-03 12:35:54 +00:00
// Do four 4x4 transposes. Originally, v16-v31 contain the
// 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
// contain the four transposed 4x4 blocks.
transpose_4x4s v16, v17, v18, v19, v4, v5, v6, v7
transpose_4x4s v20, v21, v22, v23, v4, v5, v6, v7
transpose_4x4s v24, v25, v26, v27, v4, v5, v6, v7
transpose_4x4s v28, v29, v30, v31, v4, v5, v6, v7
// Store the transposed 4x4 blocks horizontally.
2017-01-03 12:35:54 +00:00
cmp x1, #12
b.eq 1f
.irp i, 16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31
store \i, x0, #16
.endr
ret x14
2017-01-03 12:35:54 +00:00
1:
// Special case: For the last input column (x1 == 12),
// which would be stored as the last row in the temp buffer,
// don't store the first 4x4 block, but keep it in registers
// for the first slice of the second pass (where it is the
// last 4x4 block).
add x0, x0, #16
st1 {v20.4s}, [x0], #16
st1 {v24.4s}, [x0], #16
st1 {v28.4s}, [x0], #16
2017-01-03 12:35:54 +00:00
add x0, x0, #16
st1 {v21.4s}, [x0], #16
st1 {v25.4s}, [x0], #16
st1 {v29.4s}, [x0], #16
2017-01-03 12:35:54 +00:00
add x0, x0, #16
st1 {v22.4s}, [x0], #16
st1 {v26.4s}, [x0], #16
st1 {v30.4s}, [x0], #16
2017-01-03 12:35:54 +00:00
add x0, x0, #16
st1 {v23.4s}, [x0], #16
st1 {v27.4s}, [x0], #16
st1 {v31.4s}, [x0], #16
2017-01-03 12:35:54 +00:00
mov v28.16b, v16.16b
mov v29.16b, v17.16b
mov v30.16b, v18.16b
mov v31.16b, v19.16b
ret x14
2017-01-03 12:35:54 +00:00
endfunc
// Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
// load the destination pixels (from a similar 4x16 slice), add and store back.
// x0 = dst
// x1 = dst stride
// x2 = src (temp buffer)
// x3 = slice offset
// x9 = temp buffer stride
function \txfm\()16_1d_4x16_pass2_neon
mov x14, x30
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27
load \i, x2, x9
.endr
cbz x3, 1f
.irp i, 28, 29, 30, 31
load \i, x2, x9
.endr
1:
add x3, x0, x1
lsl x1, x1, #1
bl \txfm\()16
2017-01-03 12:35:54 +00:00
dup v8.8h, w13
load_add_store v16.4s, v17.4s, v18.4s, v19.4s, v20.4s, v21.4s, v22.4s, v23.4s
load_add_store v24.4s, v25.4s, v26.4s, v27.4s, v28.4s, v29.4s, v30.4s, v31.4s
ret x14
2017-01-03 12:35:54 +00:00
endfunc
.endm
itxfm16_1d_funcs idct
itxfm16_1d_funcs iadst
// This is the minimum eob value for each subpartition, in increments of 4
const min_eob_idct_idct_16, align=4
.short 0, 10, 38, 89
endconst
.macro itxfm_func16x16 txfm1, txfm2
function vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
.ifc \txfm1\()_\txfm2,idct_idct
cmp w3, #1
b.eq idct16x16_dc_add_neon
.endif
mov x15, x30
// iadst16 requires clobbering v8-v15, idct16 only clobbers v8-v9.
.ifnc \txfm1\()_\txfm2,idct_idct
stp d14, d15, [sp, #-0x10]!
stp d12, d13, [sp, #-0x10]!
stp d10, d11, [sp, #-0x10]!
.endif
stp d8, d9, [sp, #-0x10]!
sub sp, sp, #1024
mov x4, x0
mov x5, x1
mov x6, x2
movrel x10, idct_coeffs
.ifnc \txfm1\()_\txfm2,idct_idct
movrel x11, iadst16_coeffs
.endif
.ifc \txfm1,idct
ld1 {v0.8h,v1.8h}, [x10]
sxtl v2.4s, v1.4h
sxtl2 v3.4s, v1.8h
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
.endif
mov x9, #64
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifc \txfm1\()_\txfm2,idct_idct
cmp w3, #10
b.le idct16x16_quarter_add_16_neon
cmp w3, #38
b.le idct16x16_half_add_16_neon
movrel x12, min_eob_idct_idct_16, 2
.endif
2017-01-03 12:35:54 +00:00
.irp i, 0, 4, 8, 12
add x0, sp, #(\i*64)
.ifc \txfm1\()_\txfm2,idct_idct
.if \i > 0
ldrh w1, [x12], #2
cmp w3, w1
mov x1, #(16 - \i)/4
b.le 1f
.endif
.endif
mov x1, #\i
add x2, x6, #(\i*4)
bl \txfm1\()16_1d_4x16_pass1_neon
.endr
.ifc \txfm1\()_\txfm2,iadst_idct
ld1 {v0.8h,v1.8h}, [x10]
sxtl v2.4s, v1.4h
sxtl2 v3.4s, v1.8h
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
.endif
.ifc \txfm1\()_\txfm2,idct_idct
b 3f
1:
// Set v28-v31 to zero, for the in-register passthrough of
// coefficients to pass 2.
movi v28.4s, #0
movi v29.4s, #0
movi v30.4s, #0
movi v31.4s, #0
2:
subs x1, x1, #1
.rept 4
st1 {v28.4s,v29.4s,v30.4s,v31.4s}, [x0], x9
.endr
b.ne 2b
3:
.endif
.irp i, 0, 4, 8, 12
add x0, x4, #(\i*2)
mov x1, x5
add x2, sp, #(\i*4)
mov x3, #\i
bl \txfm2\()16_1d_4x16_pass2_neon
.endr
add sp, sp, #1024
ldp d8, d9, [sp], 0x10
.ifnc \txfm1\()_\txfm2,idct_idct
ldp d10, d11, [sp], 0x10
ldp d12, d13, [sp], 0x10
ldp d14, d15, [sp], 0x10
.endif
ret x15
2017-01-03 12:35:54 +00:00
endfunc
function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_10_neon, export=1
mov x13, #0x03ff
b vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
endfunc
function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_12_neon, export=1
mov x13, #0x0fff
b vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
endfunc
.endm
itxfm_func16x16 idct, idct
itxfm_func16x16 iadst, idct
itxfm_func16x16 idct, iadst
itxfm_func16x16 iadst, iadst
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
function idct16_1d_4x16_pass1_quarter_neon
mov x14, x30
movi v4.4s, #0
.irp i, 16, 17, 18, 19
load_clear \i, x2, x9
.endr
bl idct16_quarter
// Do four 4x4 transposes. Originally, v16-v31 contain the
// 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
// contain the four transposed 4x4 blocks.
transpose_4x4s v16, v17, v18, v19, v4, v5, v6, v7
transpose_4x4s v20, v21, v22, v23, v4, v5, v6, v7
transpose_4x4s v24, v25, v26, v27, v4, v5, v6, v7
transpose_4x4s v28, v29, v30, v31, v4, v5, v6, v7
// Store the transposed 4x4 blocks horizontally.
// The first 4x4 block is kept in registers for the second pass,
// store the rest in the temp buffer.
add x0, x0, #16
st1 {v20.4s}, [x0], #16
st1 {v24.4s}, [x0], #16
st1 {v28.4s}, [x0], #16
add x0, x0, #16
st1 {v21.4s}, [x0], #16
st1 {v25.4s}, [x0], #16
st1 {v29.4s}, [x0], #16
add x0, x0, #16
st1 {v22.4s}, [x0], #16
st1 {v26.4s}, [x0], #16
st1 {v30.4s}, [x0], #16
add x0, x0, #16
st1 {v23.4s}, [x0], #16
st1 {v27.4s}, [x0], #16
st1 {v31.4s}, [x0], #16
ret x14
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
function idct16_1d_4x16_pass2_quarter_neon
mov x14, x30
// Only load the top 4 lines, and only do it for the later slices.
// For the first slice, d16-d19 is kept in registers from the first pass.
cbz x3, 1f
.irp i, 16, 17, 18, 19
load \i, x2, x9
.endr
1:
add x3, x0, x1
lsl x1, x1, #1
bl idct16_quarter
dup v8.8h, w13
load_add_store v16.4s, v17.4s, v18.4s, v19.4s, v20.4s, v21.4s, v22.4s, v23.4s
load_add_store v24.4s, v25.4s, v26.4s, v27.4s, v28.4s, v29.4s, v30.4s, v31.4s
ret x14
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
function idct16_1d_4x16_pass1_half_neon
mov x14, x30
movi v4.4s, #0
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
load_clear \i, x2, x9
.endr
bl idct16_half
// Do four 4x4 transposes. Originally, v16-v31 contain the
// 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
// contain the four transposed 4x4 blocks.
transpose_4x4s v16, v17, v18, v19, v4, v5, v6, v7
transpose_4x4s v20, v21, v22, v23, v4, v5, v6, v7
transpose_4x4s v24, v25, v26, v27, v4, v5, v6, v7
transpose_4x4s v28, v29, v30, v31, v4, v5, v6, v7
// Store the transposed 4x4 blocks horizontally.
cmp x1, #4
b.eq 1f
.irp i, 16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31
store \i, x0, #16
.endr
ret x14
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
1:
// Special case: For the second input column (r1 == 4),
// which would be stored as the second row in the temp buffer,
// don't store the first 4x4 block, but keep it in registers
// for the first slice of the second pass (where it is the
// second 4x4 block).
add x0, x0, #16
st1 {v20.4s}, [x0], #16
st1 {v24.4s}, [x0], #16
st1 {v28.4s}, [x0], #16
add x0, x0, #16
st1 {v21.4s}, [x0], #16
st1 {v25.4s}, [x0], #16
st1 {v29.4s}, [x0], #16
add x0, x0, #16
st1 {v22.4s}, [x0], #16
st1 {v26.4s}, [x0], #16
st1 {v30.4s}, [x0], #16
add x0, x0, #16
st1 {v23.4s}, [x0], #16
st1 {v27.4s}, [x0], #16
st1 {v31.4s}, [x0], #16
mov v20.16b, v16.16b
mov v21.16b, v17.16b
mov v22.16b, v18.16b
mov v23.16b, v19.16b
ret x14
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
function idct16_1d_4x16_pass2_half_neon
mov x14, x30
.irp i, 16, 17, 18, 19
load \i, x2, x9
.endr
cbz x3, 1f
.irp i, 20, 21, 22, 23
load \i, x2, x9
.endr
1:
add x3, x0, x1
lsl x1, x1, #1
bl idct16_half
dup v8.8h, w13
load_add_store v16.4s, v17.4s, v18.4s, v19.4s, v20.4s, v21.4s, v22.4s, v23.4s
load_add_store v24.4s, v25.4s, v26.4s, v27.4s, v28.4s, v29.4s, v30.4s, v31.4s
ret x14
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
.macro idct16_partial size
function idct16x16_\size\()_add_16_neon
add x0, sp, #(0*64)
mov x1, #0
add x2, x6, #(0*4)
bl idct16_1d_4x16_pass1_\size\()_neon
.ifc \size,half
add x0, sp, #(4*64)
mov x1, #4
add x2, x6, #(4*4)
bl idct16_1d_4x16_pass1_\size\()_neon
.endif
.irp i, 0, 4, 8, 12
add x0, x4, #(\i*2)
mov x1, x5
add x2, sp, #(\i*4)
mov x3, #\i
bl idct16_1d_4x16_pass2_\size\()_neon
.endr
add sp, sp, #1024
ldp d8, d9, [sp], 0x10
ret x15
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
.endm
idct16_partial quarter
idct16_partial half
2017-01-03 12:35:54 +00:00
function idct32x32_dc_add_neon
movrel x4, idct_coeffs
ld1 {v0.4h}, [x4]
sxtl v0.4s, v0.4h
movi v1.4h, #0
ld1 {v2.s}[0], [x2]
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
smull v2.2d, v2.2s, v0.s[0]
rshrn v2.2s, v2.2d, #14
st1 {v1.s}[0], [x2]
dup v2.4s, v2.s[0]
srshr v0.4s, v2.4s, #6
mov x3, x0
mov x4, #32
sub x1, x1, #32
dup v31.8h, w13
1:
// Loop to add the constant v0 into all 32x32 outputs
subs x4, x4, #1
ld1 {v1.8h,v2.8h}, [x0], #32
uaddw v16.4s, v0.4s, v1.4h
uaddw2 v17.4s, v0.4s, v1.8h
ld1 {v3.8h,v4.8h}, [x0], x1
uaddw v18.4s, v0.4s, v2.4h
uaddw2 v19.4s, v0.4s, v2.8h
uaddw v20.4s, v0.4s, v3.4h
uaddw2 v21.4s, v0.4s, v3.8h
uaddw v22.4s, v0.4s, v4.4h
uaddw2 v23.4s, v0.4s, v4.8h
sqxtun v1.4h, v16.4s
sqxtun2 v1.8h, v17.4s
sqxtun v2.4h, v18.4s
sqxtun2 v2.8h, v19.4s
sqxtun v3.4h, v20.4s
sqxtun2 v3.8h, v21.4s
sqxtun v4.4h, v22.4s
sqxtun2 v4.8h, v23.4s
umin v1.8h, v1.8h, v31.8h
umin v2.8h, v2.8h, v31.8h
st1 {v1.8h,v2.8h}, [x3], #32
umin v3.8h, v3.8h, v31.8h
umin v4.8h, v4.8h, v31.8h
st1 {v3.8h,v4.8h}, [x3], x1
b.ne 1b
ret
endfunc
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.macro idct32_end
2017-01-03 12:35:54 +00:00
butterfly_4s v16, v5, v4, v5 // v16 = t16a, v5 = t19a
butterfly_4s v17, v20, v23, v20 // v17 = t17, v20 = t18
butterfly_4s v18, v6, v7, v6 // v18 = t23a, v6 = t20a
butterfly_4s v19, v21, v22, v21 // v19 = t22, v21 = t21
butterfly_4s v4, v28, v28, v30 // v4 = t24a, v28 = t27a
butterfly_4s v23, v26, v25, v26 // v23 = t25, v26 = t26
butterfly_4s v7, v8, v29, v31 // v7 = t31a, v3 = t28a
butterfly_4s v22, v27, v24, v27 // v22 = t30, v27 = t29
dmbutterfly v27, v20, v0.s[2], v0.s[3], v24, v25, v30, v31 // v27 = t18a, v20 = t29a
dmbutterfly v8, v5, v0.s[2], v0.s[3], v24, v25, v30, v31 // v3 = t19, v5 = t28
dmbutterfly v28, v6, v0.s[2], v0.s[3], v24, v25, v30, v31, neg=1 // v28 = t27, v6 = t20
dmbutterfly v26, v21, v0.s[2], v0.s[3], v24, v25, v30, v31, neg=1 // v26 = t26a, v21 = t21a
butterfly_4s v31, v24, v7, v4 // v31 = t31, v24 = t24
butterfly_4s v30, v25, v22, v23 // v30 = t30a, v25 = t25a
butterfly_4s_r v23, v16, v16, v18 // v23 = t23, v16 = t16
butterfly_4s_r v22, v17, v17, v19 // v22 = t22a, v17 = t17a
butterfly_4s v18, v21, v27, v21 // v18 = t18, v21 = t21
butterfly_4s_r v27, v28, v5, v28 // v27 = t27a, v28 = t28a
butterfly_4s v29, v26, v20, v26 // v29 = t29, v26 = t26
butterfly_4s v19, v20, v8, v6 // v19 = t19a, v20 = t20
dmbutterfly0 v27, v20, v27, v20, v4, v5, v6, v7, v8, v9 // v27 = t27, v20 = t20
dmbutterfly0 v26, v21, v26, v21, v4, v5, v6, v7, v8, v9 // v26 = t26a, v21 = t21a
dmbutterfly0 v25, v22, v25, v22, v4, v5, v6, v7, v8, v9 // v25 = t25, v22 = t22
dmbutterfly0 v24, v23, v24, v23, v4, v5, v6, v7, v8, v9 // v24 = t24a, v23 = t23a
ret
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endm
function idct32_odd
dmbutterfly v16, v31, v10.s[0], v10.s[1], v4, v5, v6, v7 // v16 = t16a, v31 = t31a
dmbutterfly v24, v23, v10.s[2], v10.s[3], v4, v5, v6, v7 // v24 = t17a, v23 = t30a
dmbutterfly v20, v27, v11.s[0], v11.s[1], v4, v5, v6, v7 // v20 = t18a, v27 = t29a
dmbutterfly v28, v19, v11.s[2], v11.s[3], v4, v5, v6, v7 // v28 = t19a, v19 = t28a
dmbutterfly v18, v29, v12.s[0], v12.s[1], v4, v5, v6, v7 // v18 = t20a, v29 = t27a
dmbutterfly v26, v21, v12.s[2], v12.s[3], v4, v5, v6, v7 // v26 = t21a, v21 = t26a
dmbutterfly v22, v25, v13.s[0], v13.s[1], v4, v5, v6, v7 // v22 = t22a, v25 = t25a
dmbutterfly v30, v17, v13.s[2], v13.s[3], v4, v5, v6, v7 // v30 = t23a, v17 = t24a
butterfly_4s v4, v24, v16, v24 // v4 = t16, v24 = t17
butterfly_4s v5, v20, v28, v20 // v5 = t19, v20 = t18
butterfly_4s v6, v26, v18, v26 // v6 = t20, v26 = t21
butterfly_4s v7, v22, v30, v22 // v7 = t23, v22 = t22
butterfly_4s v28, v25, v17, v25 // v28 = t24, v25 = t25
butterfly_4s v30, v21, v29, v21 // v30 = t27, v21 = t26
butterfly_4s v29, v23, v31, v23 // v29 = t31, v23 = t30
butterfly_4s v31, v27, v19, v27 // v31 = t28, v27 = t29
dmbutterfly v23, v24, v1.s[0], v1.s[1], v16, v17, v18, v19 // v23 = t17a, v24 = t30a
dmbutterfly v27, v20, v1.s[0], v1.s[1], v16, v17, v18, v19, neg=1 // v27 = t29a, v20 = t18a
dmbutterfly v21, v26, v1.s[2], v1.s[3], v16, v17, v18, v19 // v21 = t21a, v26 = t26a
dmbutterfly v25, v22, v1.s[2], v1.s[3], v16, v17, v18, v19, neg=1 // v25 = t25a, v22 = t22a
idct32_end
endfunc
function idct32_odd_half
dmbutterfly_h1 v16, v31, v10.s[0], v10.s[1], v4, v5, v6, v7 // v16 = t16a, v31 = t31a
dmbutterfly_h2 v24, v23, v10.s[2], v10.s[3], v4, v5, v6, v7 // v24 = t17a, v23 = t30a
dmbutterfly_h1 v20, v27, v11.s[0], v11.s[1], v4, v5, v6, v7 // v20 = t18a, v27 = t29a
dmbutterfly_h2 v28, v19, v11.s[2], v11.s[3], v4, v5, v6, v7 // v28 = t19a, v19 = t28a
dmbutterfly_h1 v18, v29, v12.s[0], v12.s[1], v4, v5, v6, v7 // v18 = t20a, v29 = t27a
dmbutterfly_h2 v26, v21, v12.s[2], v12.s[3], v4, v5, v6, v7 // v26 = t21a, v21 = t26a
dmbutterfly_h1 v22, v25, v13.s[0], v13.s[1], v4, v5, v6, v7 // v22 = t22a, v25 = t25a
dmbutterfly_h2 v30, v17, v13.s[2], v13.s[3], v4, v5, v6, v7 // v30 = t23a, v17 = t24a
butterfly_4s v4, v24, v16, v24 // v4 = t16, v24 = t17
butterfly_4s v5, v20, v28, v20 // v5 = t19, v20 = t18
butterfly_4s v6, v26, v18, v26 // v6 = t20, v26 = t21
butterfly_4s v7, v22, v30, v22 // v7 = t23, v22 = t22
butterfly_4s v28, v25, v17, v25 // v28 = t24, v25 = t25
butterfly_4s v30, v21, v29, v21 // v30 = t27, v21 = t26
butterfly_4s v29, v23, v31, v23 // v29 = t31, v23 = t30
butterfly_4s v31, v27, v19, v27 // v31 = t28, v27 = t29
dmbutterfly v23, v24, v1.s[0], v1.s[1], v16, v17, v18, v19 // v23 = t17a, v24 = t30a
dmbutterfly v27, v20, v1.s[0], v1.s[1], v16, v17, v18, v19, neg=1 // v27 = t29a, v20 = t18a
dmbutterfly v21, v26, v1.s[2], v1.s[3], v16, v17, v18, v19 // v21 = t21a, v26 = t26a
dmbutterfly v25, v22, v1.s[2], v1.s[3], v16, v17, v18, v19, neg=1 // v25 = t25a, v22 = t22a
idct32_end
endfunc
function idct32_odd_quarter
dsmull_h v4, v5, v16, v10.s[0]
dsmull_h v28, v29, v19, v11.s[3]
dsmull_h v30, v31, v16, v10.s[1]
dsmull_h v22, v23, v17, v13.s[2]
dsmull_h v7, v6, v17, v13.s[3]
dsmull_h v26, v27, v19, v11.s[2]
dsmull_h v20, v21, v18, v12.s[0]
dsmull_h v24, v25, v18, v12.s[1]
neg v28.2d, v28.2d
neg v29.2d, v29.2d
neg v7.2d, v7.2d
neg v6.2d, v6.2d
drshrn_h v4, v4, v5, #14
drshrn_h v5, v28, v29, #14
drshrn_h v29, v30, v31, #14
drshrn_h v28, v22, v23, #14
drshrn_h v7, v7, v6, #14
drshrn_h v31, v26, v27, #14
drshrn_h v6, v20, v21, #14
drshrn_h v30, v24, v25, #14
dmbutterfly_l v16, v17, v18, v19, v29, v4, v1.s[0], v1.s[1]
dmbutterfly_l v27, v26, v20, v21, v31, v5, v1.s[0], v1.s[1]
drshrn_h v23, v16, v17, #14
drshrn_h v24, v18, v19, #14
neg v20.2d, v20.2d
neg v21.2d, v21.2d
drshrn_h v27, v27, v26, #14
drshrn_h v20, v20, v21, #14
dmbutterfly_l v16, v17, v18, v19, v30, v6, v1.s[2], v1.s[3]
drshrn_h v21, v16, v17, #14
drshrn_h v26, v18, v19, #14
dmbutterfly_l v16, v17, v18, v19, v28, v7, v1.s[2], v1.s[3]
drshrn_h v25, v16, v17, #14
neg v18.2d, v18.2d
neg v19.2d, v19.2d
drshrn_h v22, v18, v19, #14
idct32_end
endfunc
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.macro idct32_funcs suffix
2017-01-03 12:35:54 +00:00
// Do an 32-point IDCT of a 4x32 slice out of a 32x32 matrix.
// The 32-point IDCT can be decomposed into two 16-point IDCTs;
// a normal IDCT16 with every other input component (the even ones, with
// each output written twice), followed by a separate 16-point IDCT
// of the odd inputs, added/subtracted onto the outputs of the first idct16.
// x0 = dst (temp buffer)
// x1 = unused
// x2 = src
// x9 = double input stride
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
function idct32_1d_4x32_pass1\suffix\()_neon
mov x14, x30
2017-01-03 12:35:54 +00:00
movi v4.4s, #0
// v16 = IN(0), v17 = IN(2) ... v31 = IN(30)
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifb \suffix
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
load_clear \i, x2, x9
.endr
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
load_clear \i, x2, x9
2017-01-03 12:35:54 +00:00
.endr
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
load_clear \i, x2, x9
.endr
.endif
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
bl idct16\suffix
2017-01-03 12:35:54 +00:00
// Do four 4x4 transposes. Originally, v16-v31 contain the
// 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
// contain the four transposed 4x4 blocks.
transpose_4x4s v16, v17, v18, v19, v4, v5, v6, v7
transpose_4x4s v20, v21, v22, v23, v4, v5, v6, v7
transpose_4x4s v24, v25, v26, v27, v4, v5, v6, v7
transpose_4x4s v28, v29, v30, v31, v4, v5, v6, v7
// Store the registers a, b, c, d horizontally, followed by the
// same registers d, c, b, a mirrored.
.macro store_rev a, b, c, d
// There's no rev128 instruction, but we reverse each 64 bit
// half, and then flip them using an ext with 8 bytes offset.
rev64 v7.4s, \d
st1 {\a}, [x0], #16
2017-01-03 12:35:54 +00:00
ext v7.16b, v7.16b, v7.16b, #8
st1 {\b}, [x0], #16
rev64 v6.4s, \c
st1 {\c}, [x0], #16
2017-01-03 12:35:54 +00:00
ext v6.16b, v6.16b, v6.16b, #8
st1 {\d}, [x0], #16
rev64 v5.4s, \b
2017-01-03 12:35:54 +00:00
st1 {v7.4s}, [x0], #16
ext v5.16b, v5.16b, v5.16b, #8
st1 {v6.4s}, [x0], #16
rev64 v4.4s, \a
2017-01-03 12:35:54 +00:00
st1 {v5.4s}, [x0], #16
ext v4.16b, v4.16b, v4.16b, #8
st1 {v4.4s}, [x0], #16
.endm
store_rev v16.4s, v20.4s, v24.4s, v28.4s
store_rev v17.4s, v21.4s, v25.4s, v29.4s
store_rev v18.4s, v22.4s, v26.4s, v30.4s
store_rev v19.4s, v23.4s, v27.4s, v31.4s
2017-01-03 12:35:54 +00:00
sub x0, x0, #512
.purgem store_rev
// Move x2 back to the start of the input, and move
// to the first odd row
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifb \suffix
2017-01-03 12:35:54 +00:00
sub x2, x2, x9, lsl #4
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endif
.ifc \suffix,_quarter
sub x2, x2, x9, lsl #2
.endif
.ifc \suffix,_half
sub x2, x2, x9, lsl #3
.endif
2017-01-03 12:35:54 +00:00
add x2, x2, #128
movi v4.4s, #0
// v16 = IN(1), v17 = IN(3) ... v31 = IN(31)
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifb \suffix
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
load_clear \i, x2, x9
.endr
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
load_clear \i, x2, x9
2017-01-03 12:35:54 +00:00
.endr
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
load_clear \i, x2, x9
.endr
.endif
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
bl idct32_odd\suffix
2017-01-03 12:35:54 +00:00
transpose_4x4s v31, v30, v29, v28, v4, v5, v6, v7
transpose_4x4s v27, v26, v25, v24, v4, v5, v6, v7
transpose_4x4s v23, v22, v21, v20, v4, v5, v6, v7
transpose_4x4s v19, v18, v17, v16, v4, v5, v6, v7
// Store the registers a, b, c, d horizontally,
// adding into the output first, and the mirrored,
// subtracted from the output.
.macro store_rev a, b, c, d, a16b, b16b
2017-01-03 12:35:54 +00:00
ld1 {v4.4s}, [x0]
rev64 v9.4s, \d
add v4.4s, v4.4s, \a
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
rev64 v8.4s, \c
2017-01-03 12:35:54 +00:00
ld1 {v4.4s}, [x0]
ext v9.16b, v9.16b, v9.16b, #8
add v4.4s, v4.4s, \b
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
ext v8.16b, v8.16b, v8.16b, #8
ld1 {v4.4s}, [x0]
rev64 \b, \b
add v4.4s, v4.4s, \c
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
rev64 \a, \a
2017-01-03 12:35:54 +00:00
ld1 {v4.4s}, [x0]
ext \b16b, \b16b, \b16b, #8
add v4.4s, v4.4s, \d
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
ext \a16b, \a16b, \a16b, #8
2017-01-03 12:35:54 +00:00
ld1 {v4.4s}, [x0]
sub v4.4s, v4.4s, v9.4s
st1 {v4.4s}, [x0], #16
ld1 {v4.4s}, [x0]
sub v4.4s, v4.4s, v8.4s
st1 {v4.4s}, [x0], #16
ld1 {v4.4s}, [x0]
sub v4.4s, v4.4s, \b
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
ld1 {v4.4s}, [x0]
sub v4.4s, v4.4s, \a
2017-01-03 12:35:54 +00:00
st1 {v4.4s}, [x0], #16
.endm
store_rev v31.4s, v27.4s, v23.4s, v19.4s, v31.16b, v27.16b
store_rev v30.4s, v26.4s, v22.4s, v18.4s, v30.16b, v26.16b
store_rev v29.4s, v25.4s, v21.4s, v17.4s, v29.16b, v25.16b
store_rev v28.4s, v24.4s, v20.4s, v16.4s, v28.16b, v24.16b
2017-01-03 12:35:54 +00:00
.purgem store_rev
ret x14
2017-01-03 12:35:54 +00:00
endfunc
// This is mostly the same as 4x32_pass1, but without the transpose,
// and use the source as temp buffer between the two idct passes, and
// add into the destination.
// x0 = dst
// x1 = dst stride
// x2 = src (temp buffer)
// x7 = negative double temp buffer stride
// x9 = double temp buffer stride
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
function idct32_1d_4x32_pass2\suffix\()_neon
mov x14, x30
2017-01-03 12:35:54 +00:00
// v16 = IN(0), v17 = IN(2) ... v31 = IN(30)
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifb \suffix
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
load \i, x2, x9
2017-01-03 12:35:54 +00:00
.endr
sub x2, x2, x9, lsl #4
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
load \i, x2, x9
.endr
sub x2, x2, x9, lsl #2
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
load \i, x2, x9
.endr
sub x2, x2, x9, lsl #3
.endif
2017-01-03 12:35:54 +00:00
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
bl idct16\suffix
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
store \i, x2, x9
2017-01-03 12:35:54 +00:00
.endr
sub x2, x2, x9, lsl #4
add x2, x2, #128
// v16 = IN(1), v17 = IN(3) ... v31 = IN(31)
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.ifb \suffix
2017-01-03 12:35:54 +00:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
load \i, x2, x9
2017-01-03 12:35:54 +00:00
.endr
sub x2, x2, x9, lsl #4
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
load \i, x2, x9
.endr
sub x2, x2, x9, lsl #2
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
load \i, x2, x9
.endr
sub x2, x2, x9, lsl #3
.endif
2017-01-03 12:35:54 +00:00
sub x2, x2, #128
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
bl idct32_odd\suffix
2017-01-03 12:35:54 +00:00
.macro load_acc_store a, b, c, d, neg=0
.if \neg == 0
ld1 {v4.4s}, [x2], x9
ld1 {v5.4s}, [x2], x9
add v4.4s, v4.4s, \a
2017-01-03 12:35:54 +00:00
ld1 {v6.4s}, [x2], x9
add v5.4s, v5.4s, \b
2017-01-03 12:35:54 +00:00
ld1 {v7.4s}, [x2], x9
add v6.4s, v6.4s, \c
add v7.4s, v7.4s, \d
2017-01-03 12:35:54 +00:00
.else
ld1 {v4.4s}, [x2], x7
ld1 {v5.4s}, [x2], x7
sub v4.4s, v4.4s, \a
2017-01-03 12:35:54 +00:00
ld1 {v6.4s}, [x2], x7
sub v5.4s, v5.4s, \b
2017-01-03 12:35:54 +00:00
ld1 {v7.4s}, [x2], x7
sub v6.4s, v6.4s, \c
sub v7.4s, v7.4s, \d
2017-01-03 12:35:54 +00:00
.endif
ld1 {v8.4h}, [x0], x1
ld1 {v8.d}[1], [x0], x1
srshr v4.4s, v4.4s, #6
ld1 {v9.4h}, [x0], x1
srshr v5.4s, v5.4s, #6
uaddw v4.4s, v4.4s, v8.4h
ld1 {v9.d}[1], [x0], x1
srshr v6.4s, v6.4s, #6
uaddw2 v5.4s, v5.4s, v8.8h
srshr v7.4s, v7.4s, #6
sub x0, x0, x1, lsl #2
uaddw v6.4s, v6.4s, v9.4h
sqxtun v4.4h, v4.4s
uaddw2 v7.4s, v7.4s, v9.8h
sqxtun2 v4.8h, v5.4s
umin v4.8h, v4.8h, v15.8h
st1 {v4.4h}, [x0], x1
sqxtun v5.4h, v6.4s
st1 {v4.d}[1], [x0], x1
sqxtun2 v5.8h, v7.4s
umin v5.8h, v5.8h, v15.8h
st1 {v5.4h}, [x0], x1
st1 {v5.d}[1], [x0], x1
.endm
load_acc_store v31.4s, v30.4s, v29.4s, v28.4s
load_acc_store v27.4s, v26.4s, v25.4s, v24.4s
load_acc_store v23.4s, v22.4s, v21.4s, v20.4s
load_acc_store v19.4s, v18.4s, v17.4s, v16.4s
2017-01-03 12:35:54 +00:00
sub x2, x2, x9
load_acc_store v16.4s, v17.4s, v18.4s, v19.4s, 1
load_acc_store v20.4s, v21.4s, v22.4s, v23.4s, 1
load_acc_store v24.4s, v25.4s, v26.4s, v27.4s, 1
load_acc_store v28.4s, v29.4s, v30.4s, v31.4s, 1
2017-01-03 12:35:54 +00:00
.purgem load_acc_store
ret x14
2017-01-03 12:35:54 +00:00
endfunc
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endm
idct32_funcs
idct32_funcs _quarter
idct32_funcs _half
2017-01-03 12:35:54 +00:00
const min_eob_idct_idct_32, align=4
.short 0, 9, 34, 70, 135, 240, 336, 448
endconst
function vp9_idct_idct_32x32_add_16_neon
cmp w3, #1
b.eq idct32x32_dc_add_neon
movrel x10, idct_coeffs
mov x15, x30
stp d8, d9, [sp, #-0x10]!
stp d10, d11, [sp, #-0x10]!
stp d12, d13, [sp, #-0x10]!
stp d14, d15, [sp, #-0x10]!
sub sp, sp, #4096
mov x4, x0
mov x5, x1
mov x6, x2
// Double stride of the input, since we only read every other line
mov x9, #256
neg x7, x9
ld1 {v0.8h,v1.8h}, [x10], #32
sxtl v2.4s, v1.4h
sxtl2 v3.4s, v1.8h
sxtl2 v1.4s, v0.8h
sxtl v0.4s, v0.4h
ld1 {v10.8h,v11.8h}, [x10]
sxtl v12.4s, v11.4h
sxtl2 v13.4s, v11.8h
sxtl2 v11.4s, v10.8h
sxtl v10.4s, v10.4h
dup v15.8h, w13
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
cmp w3, #34
b.le idct32x32_quarter_add_16_neon
cmp w3, #135
b.le idct32x32_half_add_16_neon
movrel x12, min_eob_idct_idct_32, 2
2017-01-03 12:35:54 +00:00
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add x0, sp, #(\i*128)
.if \i > 0
ldrh w1, [x12], #2
cmp w3, w1
mov x1, #(32 - \i)/4
b.le 1f
.endif
add x2, x6, #(\i*4)
bl idct32_1d_4x32_pass1_neon
.endr
b 3f
1:
// Write zeros to the temp buffer for pass 2
movi v16.4s, #0
movi v17.4s, #0
movi v18.4s, #0
movi v19.4s, #0
2:
subs x1, x1, #1
.rept 4
st1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x0], #64
st1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x0], #64
2017-01-03 12:35:54 +00:00
.endr
b.ne 2b
3:
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add x0, x4, #(\i*2)
mov x1, x5
add x2, sp, #(\i*4)
bl idct32_1d_4x32_pass2_neon
.endr
add sp, sp, #4096
ldp d14, d15, [sp], 0x10
ldp d12, d13, [sp], 0x10
ldp d10, d11, [sp], 0x10
ldp d8, d9, [sp], 0x10
ret x15
2017-01-03 12:35:54 +00:00
endfunc
function ff_vp9_idct_idct_32x32_add_10_neon, export=1
mov x13, #0x03ff
b vp9_idct_idct_32x32_add_16_neon
endfunc
function ff_vp9_idct_idct_32x32_add_12_neon, export=1
mov x13, #0x0fff
b vp9_idct_idct_32x32_add_16_neon
endfunc
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.macro idct32_partial size
function idct32x32_\size\()_add_16_neon
.irp i, 0, 4
add x0, sp, #(\i*128)
.ifc \size,quarter
.if \i == 4
cmp w3, #9
b.le 1f
.endif
.endif
add x2, x6, #(\i*4)
bl idct32_1d_4x32_pass1_\size\()_neon
.endr
.ifc \size,half
.irp i, 8, 12
add x0, sp, #(\i*128)
.if \i == 12
cmp w3, #70
b.le 1f
.endif
add x2, x6, #(\i*4)
bl idct32_1d_4x32_pass1_\size\()_neon
.endr
.endif
b 3f
1:
// Write zeros to the temp buffer for pass 2
movi v16.4s, #0
movi v17.4s, #0
movi v18.4s, #0
movi v19.4s, #0
.rept 4
st1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x0], #64
st1 {v16.4s,v17.4s,v18.4s,v19.4s}, [x0], #64
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
.endr
3:
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add x0, x4, #(\i*2)
mov x1, x5
add x2, sp, #(\i*4)
bl idct32_1d_4x32_pass2_\size\()_neon
.endr
add sp, sp, #4096
ldp d14, d15, [sp], 0x10
ldp d12, d13, [sp], 0x10
ldp d10, d11, [sp], 0x10
ldp d8, d9, [sp], 0x10
ret x15
aarch64: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 21512 bytes to 31400 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 284.6 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1902.7 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1903.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 2201.1 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2510.0 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2821.3 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1011.6 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 9716.5 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 9704.9 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 10641.7 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 11555.7 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 12499.8 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13403.7 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14335.8 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15253.6 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16179.5 After: vp9_inv_dct_dct_16x16_sub1_add_10_neon: 282.8 vp9_inv_dct_dct_16x16_sub2_add_10_neon: 1142.4 vp9_inv_dct_dct_16x16_sub4_add_10_neon: 1139.0 vp9_inv_dct_dct_16x16_sub8_add_10_neon: 1772.9 vp9_inv_dct_dct_16x16_sub12_add_10_neon: 2515.2 vp9_inv_dct_dct_16x16_sub16_add_10_neon: 2823.5 vp9_inv_dct_dct_32x32_sub1_add_10_neon: 1012.7 vp9_inv_dct_dct_32x32_sub2_add_10_neon: 6944.4 vp9_inv_dct_dct_32x32_sub4_add_10_neon: 6944.2 vp9_inv_dct_dct_32x32_sub8_add_10_neon: 7609.8 vp9_inv_dct_dct_32x32_sub12_add_10_neon: 9953.4 vp9_inv_dct_dct_32x32_sub16_add_10_neon: 10770.1 vp9_inv_dct_dct_32x32_sub20_add_10_neon: 13418.8 vp9_inv_dct_dct_32x32_sub24_add_10_neon: 14330.7 vp9_inv_dct_dct_32x32_sub28_add_10_neon: 15257.1 vp9_inv_dct_dct_32x32_sub32_add_10_neon: 16190.6 Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-25 22:38:48 +00:00
endfunc
.endm
idct32_partial quarter
idct32_partial half