Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
112860 Commits
37 Branches
400 Tags
465 MiB
Tag: Branch: Tree: 5f87a68cf7
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '5f87a68cf7'
${ noResults }
FFmpeg/libavcodec/mips/vp3dsp_idct_msa.c

599 lines
20 KiB
Raw Normal View History Unescape

avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
/*
* Copyright (c) 2018 gxw <guxiwei-hf@loongson.cn>
*
* 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 "vp3dsp_mips.h"
#include "libavutil/mips/generic_macros_msa.h"
#include "libavutil/intreadwrite.h"
#include "libavcodec/rnd_avg.h"
static void idct_msa(uint8_t *dst, int stride, int16_t *input, int type)
{
v8i16 r0, r1, r2, r3, r4, r5, r6, r7, sign;
v4i32 r0_r, r0_l, r1_r, r1_l, r2_r, r2_l, r3_r, r3_l,
r4_r, r4_l, r5_r, r5_l, r6_r, r6_l, r7_r, r7_l;
v4i32 A, B, C, D, Ad, Bd, Cd, Dd, E, F, G, H;
v4i32 Ed, Gd, Add, Bdd, Fd, Hd;
v16u8 sign_l;
v16i8 d0, d1, d2, d3, d4, d5, d6, d7;
v4i32 c0, c1, c2, c3, c4, c5, c6, c7;
v4i32 f0, f1, f2, f3, f4, f5, f6, f7;
v4i32 sign_t;
v16i8 zero = {0};
v16i8 mask = {0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0};
v4i32 cnst64277w = {64277, 64277, 64277, 64277};
v4i32 cnst60547w = {60547, 60547, 60547, 60547};
v4i32 cnst54491w = {54491, 54491, 54491, 54491};
v4i32 cnst46341w = {46341, 46341, 46341, 46341};
v4i32 cnst36410w = {36410, 36410, 36410, 36410};
v4i32 cnst25080w = {25080, 25080, 25080, 25080};
v4i32 cnst12785w = {12785, 12785, 12785, 12785};
v4i32 cnst8w = {8, 8, 8, 8};
v4i32 cnst2048w = {2048, 2048, 2048, 2048};
v4i32 cnst128w = {128, 128, 128, 128};
/* Extended input data */
LD_SH8(input, 8, r0, r1, r2, r3, r4, r5, r6, r7);
sign = __msa_clti_s_h(r0, 0);
r0_r = (v4i32) __msa_ilvr_h(sign, r0);
r0_l = (v4i32) __msa_ilvl_h(sign, r0);
sign = __msa_clti_s_h(r1, 0);
r1_r = (v4i32) __msa_ilvr_h(sign, r1);
r1_l = (v4i32) __msa_ilvl_h(sign, r1);
sign = __msa_clti_s_h(r2, 0);
r2_r = (v4i32) __msa_ilvr_h(sign, r2);
r2_l = (v4i32) __msa_ilvl_h(sign, r2);
sign = __msa_clti_s_h(r3, 0);
r3_r = (v4i32) __msa_ilvr_h(sign, r3);
r3_l = (v4i32) __msa_ilvl_h(sign, r3);
sign = __msa_clti_s_h(r4, 0);
r4_r = (v4i32) __msa_ilvr_h(sign, r4);
r4_l = (v4i32) __msa_ilvl_h(sign, r4);
sign = __msa_clti_s_h(r5, 0);
r5_r = (v4i32) __msa_ilvr_h(sign, r5);
r5_l = (v4i32) __msa_ilvl_h(sign, r5);
sign = __msa_clti_s_h(r6, 0);
r6_r = (v4i32) __msa_ilvr_h(sign, r6);
r6_l = (v4i32) __msa_ilvl_h(sign, r6);
sign = __msa_clti_s_h(r7, 0);
r7_r = (v4i32) __msa_ilvr_h(sign, r7);
r7_l = (v4i32) __msa_ilvl_h(sign, r7);
/* Right part */
A = ((r1_r * cnst64277w) >> 16) + ((r7_r * cnst12785w) >> 16);
B = ((r1_r * cnst12785w) >> 16) - ((r7_r * cnst64277w) >> 16);
C = ((r3_r * cnst54491w) >> 16) + ((r5_r * cnst36410w) >> 16);
D = ((r5_r * cnst54491w) >> 16) - ((r3_r * cnst36410w) >> 16);
Ad = ((A - C) * cnst46341w) >> 16;
Bd = ((B - D) * cnst46341w) >> 16;
Cd = A + C;
Dd = B + D;
E = ((r0_r + r4_r) * cnst46341w) >> 16;
F = ((r0_r - r4_r) * cnst46341w) >> 16;
G = ((r2_r * cnst60547w) >> 16) + ((r6_r * cnst25080w) >> 16);
H = ((r2_r * cnst25080w) >> 16) - ((r6_r * cnst60547w) >> 16);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
r0_r = Gd + Cd;
r7_r = Gd - Cd;
r1_r = Add + Hd;
r2_r = Add - Hd;
r3_r = Ed + Dd;
r4_r = Ed - Dd;
r5_r = Fd + Bdd;
r6_r = Fd - Bdd;
/* Left part */
A = ((r1_l * cnst64277w) >> 16) + ((r7_l * cnst12785w) >> 16);
B = ((r1_l * cnst12785w) >> 16) - ((r7_l * cnst64277w) >> 16);
C = ((r3_l * cnst54491w) >> 16) + ((r5_l * cnst36410w) >> 16);
D = ((r5_l * cnst54491w) >> 16) - ((r3_l * cnst36410w) >> 16);
Ad = ((A - C) * cnst46341w) >> 16;
Bd = ((B - D) * cnst46341w) >> 16;
Cd = A + C;
Dd = B + D;
E = ((r0_l + r4_l) * cnst46341w) >> 16;
F = ((r0_l - r4_l) * cnst46341w) >> 16;
G = ((r2_l * cnst60547w) >> 16) + ((r6_l * cnst25080w) >> 16);
H = ((r2_l * cnst25080w) >> 16) - ((r6_l * cnst60547w) >> 16);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
r0_l = Gd + Cd;
r7_l = Gd - Cd;
r1_l = Add + Hd;
r2_l = Add - Hd;
r3_l = Ed + Dd;
r4_l = Ed - Dd;
r5_l = Fd + Bdd;
r6_l = Fd - Bdd;
/* Row 0 to 3 */
TRANSPOSE4x4_SW_SW(r0_r, r1_r, r2_r, r3_r,
r0_r, r1_r, r2_r, r3_r);
TRANSPOSE4x4_SW_SW(r0_l, r1_l, r2_l, r3_l,
r0_l, r1_l, r2_l, r3_l);
A = ((r1_r * cnst64277w) >> 16) + ((r3_l * cnst12785w) >> 16);
B = ((r1_r * cnst12785w) >> 16) - ((r3_l * cnst64277w) >> 16);
C = ((r3_r * cnst54491w) >> 16) + ((r1_l * cnst36410w) >> 16);
D = ((r1_l * cnst54491w) >> 16) - ((r3_r * cnst36410w) >> 16);
Ad = ((A - C) * cnst46341w) >> 16;
Bd = ((B - D) * cnst46341w) >> 16;
Cd = A + C;
Dd = B + D;
E = ((r0_r + r0_l) * cnst46341w) >> 16;
E += cnst8w;
F = ((r0_r - r0_l) * cnst46341w) >> 16;
F += cnst8w;
if (type == 1) { // HACK
E += cnst2048w;
F += cnst2048w;
}
G = ((r2_r * cnst60547w) >> 16) + ((r2_l * cnst25080w) >> 16);
H = ((r2_r * cnst25080w) >> 16) - ((r2_l * cnst60547w) >> 16);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
A = (Gd + Cd) >> 4;
B = (Gd - Cd) >> 4;
C = (Add + Hd) >> 4;
D = (Add - Hd) >> 4;
E = (Ed + Dd) >> 4;
F = (Ed - Dd) >> 4;
G = (Fd + Bdd) >> 4;
H = (Fd - Bdd) >> 4;
if (type != 1) {
LD_SB8(dst, stride, d0, d1, d2, d3, d4, d5, d6, d7);
ILVR_B4_SW(zero, d0, zero, d1, zero, d2, zero, d3,
f0, f1, f2, f3);
ILVR_B4_SW(zero, d4, zero, d5, zero, d6, zero, d7,
f4, f5, f6, f7);
ILVR_H4_SW(zero, f0, zero, f1, zero, f2, zero, f3,
c0, c1, c2, c3);
ILVR_H4_SW(zero, f4, zero, f5, zero, f6, zero, f7,
c4, c5, c6, c7);
A += c0;
B += c7;
C += c1;
D += c2;
E += c3;
F += c4;
G += c5;
H += c6;
}
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(A, B, C, D, E, F, G, H);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
sign_l = __msa_or_v((v16u8)r1_r, (v16u8)r2_r);
sign_l = __msa_or_v(sign_l, (v16u8)r3_r);
sign_l = __msa_or_v(sign_l, (v16u8)r0_l);
sign_l = __msa_or_v(sign_l, (v16u8)r1_l);
sign_l = __msa_or_v(sign_l, (v16u8)r2_l);
sign_l = __msa_or_v(sign_l, (v16u8)r3_l);
sign_t = __msa_ceqi_w((v4i32)sign_l, 0);
Add = ((r0_r * cnst46341w) + (8 << 16)) >> 20;
if (type == 1) {
Bdd = Add + cnst128w;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW_0_255(Bdd);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
Ad = Bdd;
Bd = Bdd;
Cd = Bdd;
Dd = Bdd;
Ed = Bdd;
Fd = Bdd;
Gd = Bdd;
Hd = Bdd;
} else {
Ad = Add + c0;
Bd = Add + c1;
Cd = Add + c2;
Dd = Add + c3;
Ed = Add + c4;
Fd = Add + c5;
Gd = Add + c6;
Hd = Add + c7;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(Ad, Bd, Cd, Dd, Ed, Fd, Gd, Hd);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
}
Ad = (v4i32)__msa_and_v((v16u8)Ad, (v16u8)sign_t);
Bd = (v4i32)__msa_and_v((v16u8)Bd, (v16u8)sign_t);
Cd = (v4i32)__msa_and_v((v16u8)Cd, (v16u8)sign_t);
Dd = (v4i32)__msa_and_v((v16u8)Dd, (v16u8)sign_t);
Ed = (v4i32)__msa_and_v((v16u8)Ed, (v16u8)sign_t);
Fd = (v4i32)__msa_and_v((v16u8)Fd, (v16u8)sign_t);
Gd = (v4i32)__msa_and_v((v16u8)Gd, (v16u8)sign_t);
Hd = (v4i32)__msa_and_v((v16u8)Hd, (v16u8)sign_t);
sign_t = __msa_ceqi_w(sign_t, 0);
A = (v4i32)__msa_and_v((v16u8)A, (v16u8)sign_t);
B = (v4i32)__msa_and_v((v16u8)B, (v16u8)sign_t);
C = (v4i32)__msa_and_v((v16u8)C, (v16u8)sign_t);
D = (v4i32)__msa_and_v((v16u8)D, (v16u8)sign_t);
E = (v4i32)__msa_and_v((v16u8)E, (v16u8)sign_t);
F = (v4i32)__msa_and_v((v16u8)F, (v16u8)sign_t);
G = (v4i32)__msa_and_v((v16u8)G, (v16u8)sign_t);
H = (v4i32)__msa_and_v((v16u8)H, (v16u8)sign_t);
r0_r = Ad + A;
r1_r = Bd + C;
r2_r = Cd + D;
r3_r = Dd + E;
r0_l = Ed + F;
r1_l = Fd + G;
r2_l = Gd + H;
r3_l = Hd + B;
/* Row 4 to 7 */
TRANSPOSE4x4_SW_SW(r4_r, r5_r, r6_r, r7_r,
r4_r, r5_r, r6_r, r7_r);
TRANSPOSE4x4_SW_SW(r4_l, r5_l, r6_l, r7_l,
r4_l, r5_l, r6_l, r7_l);
A = ((r5_r * cnst64277w) >> 16) + ((r7_l * cnst12785w) >> 16);
B = ((r5_r * cnst12785w) >> 16) - ((r7_l * cnst64277w) >> 16);
C = ((r7_r * cnst54491w) >> 16) + ((r5_l * cnst36410w) >> 16);
D = ((r5_l * cnst54491w) >> 16) - ((r7_r * cnst36410w) >> 16);
Ad = ((A - C) * cnst46341w) >> 16;
Bd = ((B - D) * cnst46341w) >> 16;
Cd = A + C;
Dd = B + D;
E = ((r4_r + r4_l) * cnst46341w) >> 16;
E += cnst8w;
F = ((r4_r - r4_l) * cnst46341w) >> 16;
F += cnst8w;
if (type == 1) { // HACK
E += cnst2048w;
F += cnst2048w;
}
G = ((r6_r * cnst60547w) >> 16) + ((r6_l * cnst25080w) >> 16);
H = ((r6_r * cnst25080w) >> 16) - ((r6_l * cnst60547w) >> 16);
Ed = E - G;
Gd = E + G;
Add = F + Ad;
Bdd = Bd - H;
Fd = F - Ad;
Hd = Bd + H;
A = (Gd + Cd) >> 4;
B = (Gd - Cd) >> 4;
C = (Add + Hd) >> 4;
D = (Add - Hd) >> 4;
E = (Ed + Dd) >> 4;
F = (Ed - Dd) >> 4;
G = (Fd + Bdd) >> 4;
H = (Fd - Bdd) >> 4;
if (type != 1) {
ILVL_H4_SW(zero, f0, zero, f1, zero, f2, zero, f3,
c0, c1, c2, c3);
ILVL_H4_SW(zero, f4, zero, f5, zero, f6, zero, f7,
c4, c5, c6, c7);
A += c0;
B += c7;
C += c1;
D += c2;
E += c3;
F += c4;
G += c5;
H += c6;
}
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(A, B, C, D, E, F, G, H);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
sign_l = __msa_or_v((v16u8)r5_r, (v16u8)r6_r);
sign_l = __msa_or_v(sign_l, (v16u8)r7_r);
sign_l = __msa_or_v(sign_l, (v16u8)r4_l);
sign_l = __msa_or_v(sign_l, (v16u8)r5_l);
sign_l = __msa_or_v(sign_l, (v16u8)r6_l);
sign_l = __msa_or_v(sign_l, (v16u8)r7_l);
sign_t = __msa_ceqi_w((v4i32)sign_l, 0);
Add = ((r4_r * cnst46341w) + (8 << 16)) >> 20;
if (type == 1) {
Bdd = Add + cnst128w;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW_0_255(Bdd);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
Ad = Bdd;
Bd = Bdd;
Cd = Bdd;
Dd = Bdd;
Ed = Bdd;
Fd = Bdd;
Gd = Bdd;
Hd = Bdd;
} else {
Ad = Add + c0;
Bd = Add + c1;
Cd = Add + c2;
Dd = Add + c3;
Ed = Add + c4;
Fd = Add + c5;
Gd = Add + c6;
Hd = Add + c7;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(Ad, Bd, Cd, Dd, Ed, Fd, Gd, Hd);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
}
Ad = (v4i32)__msa_and_v((v16u8)Ad, (v16u8)sign_t);
Bd = (v4i32)__msa_and_v((v16u8)Bd, (v16u8)sign_t);
Cd = (v4i32)__msa_and_v((v16u8)Cd, (v16u8)sign_t);
Dd = (v4i32)__msa_and_v((v16u8)Dd, (v16u8)sign_t);
Ed = (v4i32)__msa_and_v((v16u8)Ed, (v16u8)sign_t);
Fd = (v4i32)__msa_and_v((v16u8)Fd, (v16u8)sign_t);
Gd = (v4i32)__msa_and_v((v16u8)Gd, (v16u8)sign_t);
Hd = (v4i32)__msa_and_v((v16u8)Hd, (v16u8)sign_t);
sign_t = __msa_ceqi_w(sign_t, 0);
A = (v4i32)__msa_and_v((v16u8)A, (v16u8)sign_t);
B = (v4i32)__msa_and_v((v16u8)B, (v16u8)sign_t);
C = (v4i32)__msa_and_v((v16u8)C, (v16u8)sign_t);
D = (v4i32)__msa_and_v((v16u8)D, (v16u8)sign_t);
E = (v4i32)__msa_and_v((v16u8)E, (v16u8)sign_t);
F = (v4i32)__msa_and_v((v16u8)F, (v16u8)sign_t);
G = (v4i32)__msa_and_v((v16u8)G, (v16u8)sign_t);
H = (v4i32)__msa_and_v((v16u8)H, (v16u8)sign_t);
r4_r = Ad + A;
r5_r = Bd + C;
r6_r = Cd + D;
r7_r = Dd + E;
r4_l = Ed + F;
r5_l = Fd + G;
r6_l = Gd + H;
r7_l = Hd + B;
VSHF_B2_SB(r0_r, r4_r, r1_r, r5_r, mask, mask, d0, d1);
VSHF_B2_SB(r2_r, r6_r, r3_r, r7_r, mask, mask, d2, d3);
VSHF_B2_SB(r0_l, r4_l, r1_l, r5_l, mask, mask, d4, d5);
VSHF_B2_SB(r2_l, r6_l, r3_l, r7_l, mask, mask, d6, d7);
/* Final sequence of operations over-write original dst */
avutil/mips: refactor msa load and store macros. Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}. The old macros are difficult to use because they don't follow the same parameter passing rules. Changing details as following: 1. remove LD4x4_SH. 2. replace ST2x4_UB with ST_H4. 3. replace ST4x2_UB with ST_W2. 4. replace ST4x4_UB with ST_W4. 5. replace ST4x8_UB with ST_W8. 6. replace ST6x4_UB with ST_W2 and ST_H2. 7. replace ST8x1_UB with ST_D1. 8. replace ST8x2_UB with ST_D2. 9. replace ST8x4_UB with ST_D4. 10. replace ST8x8_UB with ST_D8. 11. replace ST12x4_UB with ST_D4 and ST_W4. Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride) ST_H4 store four half-word elements in vector 'in' to pdst with stride. About the macro name: 1) 'ST' means store operation. 2) 'H/W/D' means type of vector element is 'half-word/word/double-word'. 3) Number '1/2/4/8' means how many elements will be stored. About the macro parameter: 1) 'in0, in1...' 128-bits vector. 2) 'idx0, idx1...' elements index. 3) 'pdst' destination pointer to store to 4) 'stride' stride of each store operation. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
ST_D1(d0, 0, dst);
ST_D1(d1, 0, dst + stride);
ST_D1(d2, 0, dst + 2 * stride);
ST_D1(d3, 0, dst + 3 * stride);
ST_D1(d4, 0, dst + 4 * stride);
ST_D1(d5, 0, dst + 5 * stride);
ST_D1(d6, 0, dst + 6 * stride);
ST_D1(d7, 0, dst + 7 * stride);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
}
void ff_vp3_idct_put_msa(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
{
idct_msa(dest, line_size, block, 1);
memset(block, 0, sizeof(*block) * 64);
}
void ff_vp3_idct_add_msa(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
{
idct_msa(dest, line_size, block, 2);
memset(block, 0, sizeof(*block) * 64);
}
void ff_vp3_idct_dc_add_msa(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
{
int i = (block[0] + 15) >> 5;
v4i32 dc = {i, i, i, i};
v16i8 d0, d1, d2, d3, d4, d5, d6, d7;
v4i32 c0, c1, c2, c3, c4, c5, c6, c7;
v4i32 e0, e1, e2, e3, e4, e5, e6, e7;
v4i32 r0, r1, r2, r3, r4, r5, r6, r7;
v16i8 mask = {0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0};
v16i8 zero = {0};
LD_SB8(dest, line_size, d0, d1, d2, d3, d4, d5, d6, d7);
ILVR_B4_SW(zero, d0, zero, d1, zero, d2, zero, d3,
c0, c1, c2, c3);
ILVR_B4_SW(zero, d4, zero, d5, zero, d6, zero, d7,
c4, c5, c6, c7);
/* Right part */
ILVR_H4_SW(zero, c0, zero, c1, zero, c2, zero, c3,
e0, e1, e2, e3);
ILVR_H4_SW(zero, c4, zero, c5, zero, c6, zero, c7,
e4, e5, e6, e7);
e0 += dc;
e1 += dc;
e2 += dc;
e3 += dc;
e4 += dc;
e5 += dc;
e6 += dc;
e7 += dc;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(e0, e1, e2, e3, e4, e5, e6, e7);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
/* Left part */
ILVL_H4_SW(zero, c0, zero, c1, zero, c2, zero, c3,
r0, r1, r2, r3);
ILVL_H4_SW(zero, c4, zero, c5, zero, c6, zero, c7,
r4, r5, r6, r7);
r0 += dc;
r1 += dc;
r2 += dc;
r3 += dc;
r4 += dc;
r5 += dc;
r6 += dc;
r7 += dc;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW8_0_255(r0, r1, r2, r3, r4, r5, r6, r7);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
VSHF_B2_SB(e0, r0, e1, r1, mask, mask, d0, d1);
VSHF_B2_SB(e2, r2, e3, r3, mask, mask, d2, d3);
VSHF_B2_SB(e4, r4, e5, r5, mask, mask, d4, d5);
VSHF_B2_SB(e6, r6, e7, r7, mask, mask, d6, d7);
/* Final sequence of operations over-write original dst */
avutil/mips: refactor msa load and store macros. Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}. The old macros are difficult to use because they don't follow the same parameter passing rules. Changing details as following: 1. remove LD4x4_SH. 2. replace ST2x4_UB with ST_H4. 3. replace ST4x2_UB with ST_W2. 4. replace ST4x4_UB with ST_W4. 5. replace ST4x8_UB with ST_W8. 6. replace ST6x4_UB with ST_W2 and ST_H2. 7. replace ST8x1_UB with ST_D1. 8. replace ST8x2_UB with ST_D2. 9. replace ST8x4_UB with ST_D4. 10. replace ST8x8_UB with ST_D8. 11. replace ST12x4_UB with ST_D4 and ST_W4. Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride) ST_H4 store four half-word elements in vector 'in' to pdst with stride. About the macro name: 1) 'ST' means store operation. 2) 'H/W/D' means type of vector element is 'half-word/word/double-word'. 3) Number '1/2/4/8' means how many elements will be stored. About the macro parameter: 1) 'in0, in1...' 128-bits vector. 2) 'idx0, idx1...' elements index. 3) 'pdst' destination pointer to store to 4) 'stride' stride of each store operation. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
ST_D1(d0, 0, dest);
ST_D1(d1, 0, dest + line_size);
ST_D1(d2, 0, dest + 2 * line_size);
ST_D1(d3, 0, dest + 3 * line_size);
ST_D1(d4, 0, dest + 4 * line_size);
ST_D1(d5, 0, dest + 5 * line_size);
ST_D1(d6, 0, dest + 6 * line_size);
ST_D1(d7, 0, dest + 7 * line_size);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
block[0] = 0;
}
void ff_vp3_v_loop_filter_msa(uint8_t *first_pixel, ptrdiff_t stride,
int *bounding_values)
{
int nstride = -stride;
v4i32 e0, e1, f0, f1, g0, g1;
v16i8 zero = {0};
v16i8 d0, d1, d2, d3;
v8i16 c0, c1, c2, c3;
v8i16 r0;
v8i16 cnst3h = {3, 3, 3, 3, 3, 3, 3, 3},
cnst4h = {4, 4, 4, 4, 4, 4, 4, 4};
v16i8 mask = {0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0};
int16_t temp_16[8];
int temp_32[8];
LD_SB4(first_pixel + nstride * 2, stride, d0, d1, d2, d3);
ILVR_B4_SH(zero, d0, zero, d1, zero, d2, zero, d3,
c0, c1, c2, c3);
r0 = (c0 - c3) + (c2 - c1) * cnst3h;
r0 += cnst4h;
r0 = r0 >> 3;
/* Get filter_value from bounding_values one by one */
ST_SH(r0, temp_16);
for (int i = 0; i < 8; i++)
temp_32[i] = bounding_values[temp_16[i]];
LD_SW2(temp_32, 4, e0, e1);
ILVR_H2_SW(zero, c1, zero, c2, f0, g0);
ILVL_H2_SW(zero, c1, zero, c2, f1, g1);
f0 += e0;
f1 += e1;
g0 -= e0;
g1 -= e1;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW4_0_255(f0, f1, g0, g1);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
VSHF_B2_SB(f0, f1, g0, g1, mask, mask, d1, d2);
/* Final move to first_pixel */
avutil/mips: refactor msa load and store macros. Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}. The old macros are difficult to use because they don't follow the same parameter passing rules. Changing details as following: 1. remove LD4x4_SH. 2. replace ST2x4_UB with ST_H4. 3. replace ST4x2_UB with ST_W2. 4. replace ST4x4_UB with ST_W4. 5. replace ST4x8_UB with ST_W8. 6. replace ST6x4_UB with ST_W2 and ST_H2. 7. replace ST8x1_UB with ST_D1. 8. replace ST8x2_UB with ST_D2. 9. replace ST8x4_UB with ST_D4. 10. replace ST8x8_UB with ST_D8. 11. replace ST12x4_UB with ST_D4 and ST_W4. Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride) ST_H4 store four half-word elements in vector 'in' to pdst with stride. About the macro name: 1) 'ST' means store operation. 2) 'H/W/D' means type of vector element is 'half-word/word/double-word'. 3) Number '1/2/4/8' means how many elements will be stored. About the macro parameter: 1) 'in0, in1...' 128-bits vector. 2) 'idx0, idx1...' elements index. 3) 'pdst' destination pointer to store to 4) 'stride' stride of each store operation. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
ST_D1(d1, 0, first_pixel + nstride);
ST_D1(d2, 0, first_pixel);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
}
void ff_vp3_h_loop_filter_msa(uint8_t *first_pixel, ptrdiff_t stride,
int *bounding_values)
{
v16i8 d0, d1, d2, d3, d4, d5, d6, d7;
v8i16 c0, c1, c2, c3, c4, c5, c6, c7;
v8i16 r0;
v4i32 e0, e1, f0, f1, g0, g1;
v16i8 zero = {0};
v8i16 cnst3h = {3, 3, 3, 3, 3, 3, 3, 3},
cnst4h = {4, 4, 4, 4, 4, 4, 4, 4};
v16i8 mask = {0, 16, 4, 20, 8, 24, 12, 28, 0, 0, 0, 0, 0, 0, 0, 0};
int16_t temp_16[8];
int temp_32[8];
LD_SB8(first_pixel - 2, stride, d0, d1, d2, d3, d4, d5, d6, d7);
ILVR_B4_SH(zero, d0, zero, d1, zero, d2, zero, d3,
c0, c1, c2, c3);
ILVR_B4_SH(zero, d4, zero, d5, zero, d6, zero, d7,
c4, c5, c6, c7);
TRANSPOSE8x8_SH_SH(c0, c1, c2, c3, c4, c5, c6, c7,
c0, c1, c2, c3, c4, c5, c6, c7);
r0 = (c0 - c3) + (c2 - c1) * cnst3h;
r0 += cnst4h;
r0 = r0 >> 3;
/* Get filter_value from bounding_values one by one */
ST_SH(r0, temp_16);
for (int i = 0; i < 8; i++)
temp_32[i] = bounding_values[temp_16[i]];
LD_SW2(temp_32, 4, e0, e1);
ILVR_H2_SW(zero, c1, zero, c2, f0, g0);
ILVL_H2_SW(zero, c1, zero, c2, f1, g1);
f0 += e0;
f1 += e1;
g0 -= e0;
g1 -= e1;
avutil/mips: refine msa macros CLIP_*. Changing details as following: 1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in source vector. 2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'. Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x). Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x). Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x). 3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255' instead, because there are no difference in the effect of this two macros. Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
CLIP_SW4_0_255(f0, f1, g0, g1);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
VSHF_B2_SB(f0, g0, f1, g1, mask, mask, d1, d2);
/* Final move to first_pixel */
avutil/mips: refactor msa load and store macros. Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}. The old macros are difficult to use because they don't follow the same parameter passing rules. Changing details as following: 1. remove LD4x4_SH. 2. replace ST2x4_UB with ST_H4. 3. replace ST4x2_UB with ST_W2. 4. replace ST4x4_UB with ST_W4. 5. replace ST4x8_UB with ST_W8. 6. replace ST6x4_UB with ST_W2 and ST_H2. 7. replace ST8x1_UB with ST_D1. 8. replace ST8x2_UB with ST_D2. 9. replace ST8x4_UB with ST_D4. 10. replace ST8x8_UB with ST_D8. 11. replace ST12x4_UB with ST_D4 and ST_W4. Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride) ST_H4 store four half-word elements in vector 'in' to pdst with stride. About the macro name: 1) 'ST' means store operation. 2) 'H/W/D' means type of vector element is 'half-word/word/double-word'. 3) Number '1/2/4/8' means how many elements will be stored. About the macro parameter: 1) 'in0, in1...' 128-bits vector. 2) 'idx0, idx1...' elements index. 3) 'pdst' destination pointer to store to 4) 'stride' stride of each store operation. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
ST_H4(d1, 0, 1, 2, 3, first_pixel - 1, stride);
ST_H4(d2, 0, 1, 2, 3, first_pixel - 1 + 4 * stride, stride);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
}
void ff_put_no_rnd_pixels_l2_msa(uint8_t *dst, const uint8_t *src1,
const uint8_t *src2, ptrdiff_t stride, int h)
{
if (h == 8) {
v16i8 d0, d1, d2, d3, d4, d5, d6, d7;
v16i8 c0, c1, c2, c3;
v4i32 a0, a1, a2, a3, b0, b1, b2, b3;
v4i32 e0, e1, e2;
v4i32 f0, f1, f2;
v4u32 t0, t1, t2, t3;
v16i8 mask = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
int32_t value = 0xfefefefe;
v4i32 fmask = {value, value, value, value};
LD_SB8(src1, stride, d0, d1, d2, d3, d4, d5, d6, d7);
VSHF_B2_SB(d0, d1, d2, d3, mask, mask, c0, c1);
VSHF_B2_SB(d4, d5, d6, d7, mask, mask, c2, c3);
a0 = (v4i32) __msa_pckev_d((v2i64)c1, (v2i64)c0);
a2 = (v4i32) __msa_pckod_d((v2i64)c1, (v2i64)c0);
a1 = (v4i32) __msa_pckev_d((v2i64)c3, (v2i64)c2);
a3 = (v4i32) __msa_pckod_d((v2i64)c3, (v2i64)c2);
LD_SB8(src2, stride, d0, d1, d2, d3, d4, d5, d6, d7);
VSHF_B2_SB(d0, d1, d2, d3, mask, mask, c0, c1);
VSHF_B2_SB(d4, d5, d6, d7, mask, mask, c2, c3);
b0 = (v4i32) __msa_pckev_d((v2i64)c1, (v2i64)c0);
b2 = (v4i32) __msa_pckod_d((v2i64)c1, (v2i64)c0);
b1 = (v4i32) __msa_pckev_d((v2i64)c3, (v2i64)c2);
b3 = (v4i32) __msa_pckod_d((v2i64)c3, (v2i64)c2);
e0 = (v4i32) __msa_xor_v((v16u8)a0, (v16u8)b0);
e0 = (v4i32) __msa_and_v((v16u8)e0, (v16u8)fmask);
t0 = ((v4u32)e0) >> 1;
e2 = (v4i32) __msa_and_v((v16u8)a0, (v16u8)b0);
t0 = t0 + (v4u32)e2;
e1 = (v4i32) __msa_xor_v((v16u8)a1, (v16u8)b1);
e1 = (v4i32) __msa_and_v((v16u8)e1, (v16u8)fmask);
t1 = ((v4u32)e1) >> 1;
e2 = (v4i32) __msa_and_v((v16u8)a1, (v16u8)b1);
t1 = t1 + (v4u32)e2;
f0 = (v4i32) __msa_xor_v((v16u8)a2, (v16u8)b2);
f0 = (v4i32) __msa_and_v((v16u8)f0, (v16u8)fmask);
t2 = ((v4u32)f0) >> 1;
f2 = (v4i32) __msa_and_v((v16u8)a2, (v16u8)b2);
t2 = t2 + (v4u32)f2;
f1 = (v4i32) __msa_xor_v((v16u8)a3, (v16u8)b3);
f1 = (v4i32) __msa_and_v((v16u8)f1, (v16u8)fmask);
t3 = ((v4u32)f1) >> 1;
f2 = (v4i32) __msa_and_v((v16u8)a3, (v16u8)b3);
t3 = t3 + (v4u32)f2;
avutil/mips: refactor msa load and store macros. Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}. The old macros are difficult to use because they don't follow the same parameter passing rules. Changing details as following: 1. remove LD4x4_SH. 2. replace ST2x4_UB with ST_H4. 3. replace ST4x2_UB with ST_W2. 4. replace ST4x4_UB with ST_W4. 5. replace ST4x8_UB with ST_W8. 6. replace ST6x4_UB with ST_W2 and ST_H2. 7. replace ST8x1_UB with ST_D1. 8. replace ST8x2_UB with ST_D2. 9. replace ST8x4_UB with ST_D4. 10. replace ST8x8_UB with ST_D8. 11. replace ST12x4_UB with ST_D4 and ST_W4. Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride) ST_H4 store four half-word elements in vector 'in' to pdst with stride. About the macro name: 1) 'ST' means store operation. 2) 'H/W/D' means type of vector element is 'half-word/word/double-word'. 3) Number '1/2/4/8' means how many elements will be stored. About the macro parameter: 1) 'in0, in1...' 128-bits vector. 2) 'idx0, idx1...' elements index. 3) 'pdst' destination pointer to store to 4) 'stride' stride of each store operation. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
5 years ago
ST_W8(t0, t1, 0, 1, 2, 3, 0, 1, 2, 3, dst, stride);
ST_W8(t2, t3, 0, 1, 2, 3, 0, 1, 2, 3, dst + 4, stride);
avcodec/mips: [loongson] optimize theora decoding in vp3dsp. Optimize theora decoding with msa in functions: 1. ff_vp3_idct_add_msa 2. ff_vp3_idct_put_msa 3. ff_vp3_idct_dc_add_msa 4. ff_vp3_v_loop_filter_msa 5. ff_vp3_h_loop_filter_msa 6. ff_put_no_rnd_pixels_l2_msa Theora decoding speed improved about 36%(from 22fps to 30fps, Tested on loongson 2K1000). Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
6 years ago
} else {
int i;
for (i = 0; i < h; i++) {
uint32_t a, b;
a = AV_RN32(&src1[i * stride]);
b = AV_RN32(&src2[i * stride]);
AV_WN32A(&dst[i * stride], no_rnd_avg32(a, b));
a = AV_RN32(&src1[i * stride + 4]);
b = AV_RN32(&src2[i * stride + 4]);
AV_WN32A(&dst[i * stride + 4], no_rnd_avg32(a, b));
}
}
}