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.
116414 Commits
37 Branches
400 Tags
474 MiB
Tag: Branch: Tree: 60b1750134
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '60b1750134'
${ noResults }
FFmpeg/libavcodec/riscv/vp8dsp_rvv.S

535 lines
15 KiB
Raw Normal View History Unescape

lavc/vp8dsp: R-V V vp8_idct_dc_add c908: vp8_idct_dc_add_c: 102.2 vp8_idct_dc_add_rvv_i32: 42.0 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
11 months ago
/*
* Copyright (c) 2024 Institue of Software Chinese Academy of Sciences (ISCAS).
lavc/vp8dsp: R-V V vp8_luma_dc_wht This is not great as transposition is poorly supported, but it works: vp8_luma_dc_wht_c: 2.5 vp8_luma_dc_wht_rvv_i32: 1.7
7 months ago
* Copyright © 2024 Rémi Denis-Courmont.
lavc/vp8dsp: R-V V vp8_idct_dc_add c908: vp8_idct_dc_add_c: 102.2 vp8_idct_dc_add_rvv_i32: 42.0 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
11 months ago
*
* 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/riscv/asm.S"
lavc/vp8dsp: R-V V put_bilin_h v C908: vp8_put_bilin4_h_c: 367.0 vp8_put_bilin4_h_rvv_i32: 137.7 vp8_put_bilin4_v_c: 377.0 vp8_put_bilin4_v_rvv_i32: 137.7 vp8_put_bilin8_h_c: 1431.0 vp8_put_bilin8_h_rvv_i32: 297.5 vp8_put_bilin8_v_c: 1449.0 vp8_put_bilin8_v_rvv_i32: 297.5 vp8_put_bilin16_h_c: 2839.0 vp8_put_bilin16_h_rvv_i32: 344.7 vp8_put_bilin16_v_c: 2857.0 vp8_put_bilin16_v_rvv_i32: 344.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.macro vsetvlstatic8 len
.if \len <= 4
vsetivli zero, \len, e8, mf4, ta, ma
.elseif \len <= 8
vsetivli zero, \len, e8, mf2, ta, ma
.elseif \len <= 16
vsetivli zero, \len, e8, m1, ta, ma
.elseif \len <= 31
vsetivli zero, \len, e8, m2, ta, ma
.endif
.endm
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.macro vsetvlstatic16 len
.if \len <= 4
vsetivli zero, \len, e16, mf2, ta, ma
.elseif \len <= 8
vsetivli zero, \len, e16, m1, ta, ma
.elseif \len <= 16
vsetivli zero, \len, e16, m2, ta, ma
.endif
.endm
lavc/vp8dsp: R-V V vp8_luma_dc_wht This is not great as transposition is poorly supported, but it works: vp8_luma_dc_wht_c: 2.5 vp8_luma_dc_wht_rvv_i32: 1.7
7 months ago
#if __riscv_xlen >= 64
func ff_vp8_luma_dc_wht_rvv, zve64x
vsetivli zero, 1, e64, m1, ta, ma
vlseg4e64.v v4, (a1)
vsetivli zero, 4, e16, mf2, ta, ma
vwadd.vv v1, v5, v6
addi t1, sp, -48
vwadd.vv v0, v4, v7
addi t2, sp, -32
vwsub.vv v2, v5, v6
addi t3, sp, -16
vwsub.vv v3, v4, v7
addi sp, sp, -64
vsetvli zero, zero, e32, m1, ta, ma
vadd.vv v4, v0, v1
vadd.vv v5, v3, v2
vse32.v v4, (sp)
vsub.vv v6, v0, v1
vse32.v v5, (t1)
vsub.vv v7, v3, v2
vse32.v v6, (t2)
vse32.v v7, (t3)
vlseg4e32.v v4, (sp)
vadd.vv v0, v4, v7
sd zero, (a1)
vadd.vv v1, v5, v6
sd zero, 8(a1)
vsub.vv v2, v5, v6
sd zero, 16(a1)
vsub.vv v3, v4, v7
sd zero, 24(a1)
vadd.vi v0, v0, 3 # rounding mode not supported, do it manually
li t0, 4 * 16 * 2
vadd.vi v3, v3, 3
addi t1, a0, 16 * 2
vadd.vv v4, v0, v1
addi t2, a0, 16 * 2 * 2
vadd.vv v5, v3, v2
addi t3, a0, 16 * 2 * 3
vsub.vv v6, v0, v1
vsub.vv v7, v3, v2
vsetvli zero, zero, e16, mf2, ta, ma
vnsra.wi v0, v4, 3
addi sp, sp, 64
vnsra.wi v1, v5, 3
vsse16.v v0, (a0), t0
vnsra.wi v2, v6, 3
vsse16.v v1, (t1), t0
vnsra.wi v3, v7, 3
vsse16.v v2, (t2), t0
vsse16.v v3, (t3), t0
ret
endfunc
#endif
lavc/vp8dsp: R-V V vp8_idct_add T-Head C908 (cycles): vp8_idct_add_c: 312.2 vp8_idct_add_rvv_i32: 117.0
7 months ago
func ff_vp8_idct_add_rvv, zve32x
csrwi vxrm, 0
vsetivli zero, 4, e16, mf2, ta, ma
addi a3, a1, 1 * 4 * 2
addi a4, a1, 2 * 4 * 2
addi a5, a1, 3 * 4 * 2
li t1, 20091
li t2, 35468
jal t0, 1f
vsseg4e16.v v0, (a1)
jal t0, 1f
vlsseg4e8.v v4, (a0), a2
vssra.vi v0, v0, 3
sd zero, (a1)
vssra.vi v1, v1, 3
sd zero, 8(a1)
vssra.vi v2, v2, 3
sd zero, 16(a1)
vssra.vi v3, v3, 3
sd zero, 24(a1)
vsetvli zero, zero, e8, mf4, ta, ma
vwaddu.wv v0, v0, v4
vwaddu.wv v1, v1, v5
vwaddu.wv v2, v2, v6
vwaddu.wv v3, v3, v7
vsetvli zero, zero, e16, mf2, ta, ma
vmax.vx v0, v0, zero
vmax.vx v1, v1, zero
vmax.vx v2, v2, zero
vmax.vx v3, v3, zero
vsetvli zero, zero, e8, mf4, ta, ma
vnclipu.wi v4, v0, 0
vnclipu.wi v5, v1, 0
vnclipu.wi v6, v2, 0
vnclipu.wi v7, v3, 0
vssseg4e8.v v4, (a0), a2
ret
1:
vle16.v v0, (a1)
vle16.v v2, (a4)
vle16.v v1, (a3)
vle16.v v3, (a5)
vadd.vv v4, v0, v2 # t0
vsub.vv v5, v0, v2 # t1
vmulhsu.vx v8, v3, t1
vmulhsu.vx v6, v1, t2
vadd.vv v8, v8, v3
vmulhsu.vx v7, v1, t1
vmulhsu.vx v9, v3, t2
vadd.vv v7, v7, v1
vsub.vv v6, v6, v8 # t2
vadd.vv v7, v7, v9 # t3
vadd.vv v1, v5, v6
vsub.vv v2, v5, v6
vadd.vv v0, v4, v7
vsub.vv v3, v4, v7
jr t0
endfunc
lavc/vp8dsp: add R-V V vp7_idct_dc_add This just computes the direct coefficient and hands over to code shared with VP8. Accordingly the bulk of changes are just rewriting the VP8 code to share. Nothing to write home about: vp7_idct_dc_add_c: 1.7 vp7_idct_dc_add_rvv_i32: 1.2
7 months ago
func ff_vp8_idct_dc_add_rvv, zve32x
lh a3, (a1)
addi a3, a3, 4
srai a3, a3, 3
# fall through
endfunc
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
# a3 = DC
lavc/vp8dsp: add R-V V vp7_idct_dc_add This just computes the direct coefficient and hands over to code shared with VP8. Accordingly the bulk of changes are just rewriting the VP8 code to share. Nothing to write home about: vp7_idct_dc_add_c: 1.7 vp7_idct_dc_add_rvv_i32: 1.2
7 months ago
func ff_vp78_idct_dc_add_rvv, zve32x
csrwi vxrm, 0
vsetivli zero, 4, e8, mf4, ta, ma
sh zero, (a1)
vlse32.v v8, (a0), a2
vsetivli zero, 16, e16, m2, ta, ma
vzext.vf2 v16, v8
vadd.vx v16, v16, a3
vmax.vx v16, v16, zero
vsetvli zero, zero, e8, m1, ta, ma
vnclipu.wi v8, v16, 0
vsetivli zero, 4, e8, mf4, ta, ma
vsse32.v v8, (a0), a2
ret
endfunc
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
func ff_vp8_idct_dc_add4y_rvv, zve32x
li t0, 32
vsetivli zero, 4, e16, mf2, ta, ma
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
li t1, 4 - (128 << 3)
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
vlse16.v v8, (a1), t0
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
vadd.vx v8, v8, t1
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
vsra.vi v8, v8, 3
# fall through
endfunc
.variant_cc ff_vp78_idct_dc_add4y_rvv
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
# v8 = [dc0 - 128, dc1 - 128, dc2 - 128, dc3 - 128]
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
func ff_vp78_idct_dc_add4y_rvv, zve32x
vsetivli zero, 16, e16, m2, ta, ma
vid.v v4
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
li a4, 4
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
vsrl.vi v4, v4, 2
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
li t1, 128
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
vrgather.vv v0, v8, v4 # replicate each DC four times
vsetvli zero, zero, e8, m1, ta, ma
1:
vle8.v v8, (a0)
addi a4, a4, -1
vwaddu.wv v16, v0, v8
sh zero, (a1)
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
vnclip.wi v8, v16, 0
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
addi a1, a1, 32
lavc/vp8dsp: rework R-V V idct_dc_add4y DCT-related FFmpeg functions often add an unsigned 8-bit sample to a signed 16-bit coefficient, then clip the result back to an unsigned 8-bit value. RISC-V has no signed 16-bit to unsigned 8-bit clip, so instead our most common sequence is: VWADDU.WV set SEW to 16 bits VMAX.VV zero # clip negative values to 0 set SEW to 8 bits VNCLIPU.WI # clip values over 255 to 255 and narrow Here we use a different sequence which does not require toggling the vector type. This assumes that the wide addend vector is biased by -128: VWADDU.WV VNCLIP.WI # clip values to signed 8-bit and narrow VXOR.VX 0x80 # flip sign bit (convert signed to unsigned) Also the VMAX is effectively replaced by a VXOR of half-width. In this function, this comes for free as we anyway add a constant to the wide vector in the prologue. On C908, this has no observable effects. On X60, this improves microbenchmarks by about 20%.
7 months ago
vxor.vx v8, v8, t1
lavc/vp8dsp: add R-V V vp7_idct_dc_add4y As with idct_dc_add, most of the code is shared with, and replaces, the previous VP8 function. To improve performance, we break down the 16x4 matrix into 4 rows, rather than 4 squares. Thus strided loads and stores are avoided, and the 4 DC calculations are vectored. Unfortunately this requires a vector gather to splat the DC values, but overall this is still a win for performance: T-Head C908: vp7_idct_dc_add4y_c: 7.2 vp7_idct_dc_add4y_rvv_i32: 2.2 vp8_idct_dc_add4y_c: 6.2 vp8_idct_dc_add4y_rvv_i32: 2.2 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 SpacemiT X60: vp7_idct_dc_add4y_c: 6.2 vp7_idct_dc_add4y_rvv_i32: 2.0 vp8_idct_dc_add4y_c: 5.5 vp8_idct_dc_add4y_rvv_i32: 2.5 (before) vp8_idct_dc_add4y_rvv_i32: 1.7 I also tried to provision the DC values using indexed loads. It ends up slower overall, especially for VP7, as we then have to compute 16 DC's instead of just 4.
7 months ago
vse8.v v8, (a0)
add a0, a0, a2
bnez a4, 1b
ret
endfunc
lavc/vp7dsp: add R-V V vp7_idct_dc_add4uv This is almost the same story as vp7_idct_add4y. We just have to use strided loads of 2 64-bit elements to account for the different data layout in memory. T-Head C908: vp7_idct_dc_add4uv_c: 7.5 vp7_idct_dc_add4uv_rvv_i64: 2.0 vp8_idct_dc_add4uv_c: 6.2 vp8_idct_dc_add4uv_rvv_i32: 2.2 (before) vp8_idct_dc_add4uv_rvv_i64: 2.0 SpacemiT X60: vp7_idct_dc_add4uv_c: 6.7 vp7_idct_dc_add4uv_rvv_i64: 2.2 vp8_idct_dc_add4uv_c: 5.7 vp8_idct_dc_add4uv_rvv_i32: 2.5 (before) vp8_idct_dc_add4uv_rvv_i64: 2.0
7 months ago
func ff_vp8_idct_dc_add4uv_rvv, zve32x
li t0, 32
vsetivli zero, 4, e16, mf2, ta, ma
li t1, 4 - (128 << 3)
vlse16.v v8, (a1), t0
vadd.vx v8, v8, t1
vsra.vi v8, v8, 3
# fall through
endfunc
.variant_cc ff_vp78_idct_dc_add4uv_rvv
func ff_vp78_idct_dc_add4uv_rvv, zve64x
vsetivli zero, 16, e16, m2, ta, ma
vid.v v4
li a4, 4
vsrl.vi v4, v4, 2
li t1, 128
vrgather.vv v0, v8, v4 # replicate each DC four times
slli t2, a2, 2
vsetivli zero, 2, e64, m1, ta, ma
1:
vlse64.v v8, (a0), t2
addi a4, a4, -1
vsetivli zero, 16, e8, m1, ta, ma
vwaddu.wv v16, v0, v8
sh zero, (a1)
vnclip.wi v8, v16, 0
addi a1, a1, 32
vxor.vx v8, v8, t1
vsetivli zero, 2, e64, m1, ta, ma
vsse64.v v8, (a0), t2
add a0, a0, a2
bnez a4, 1b
ret
endfunc
lavc/vp8dsp: R-V V bilin_load to bilin_load_h Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.macro bilin_load_h dst mn
lavc/vp8dsp: R-V V put_bilin_h v C908: vp8_put_bilin4_h_c: 367.0 vp8_put_bilin4_h_rvv_i32: 137.7 vp8_put_bilin4_v_c: 377.0 vp8_put_bilin4_v_rvv_i32: 137.7 vp8_put_bilin8_h_c: 1431.0 vp8_put_bilin8_h_rvv_i32: 297.5 vp8_put_bilin8_v_c: 1449.0 vp8_put_bilin8_v_rvv_i32: 297.5 vp8_put_bilin16_h_c: 2839.0 vp8_put_bilin16_h_rvv_i32: 344.7 vp8_put_bilin16_v_c: 2857.0 vp8_put_bilin16_v_rvv_i32: 344.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
addi t5, a2, 1
vle8.v \dst, (a2)
vle8.v v2, (t5)
vwmulu.vx v28, \dst, t1
vwmaccu.vx v28, \mn, v2
vwaddu.wx v24, v28, t4
vnsra.wi \dst, v24, 3
.endm
lavc/vp8dsp: factor R-V V bilin functions For a given type, only the first VSETVLI instruction varies depending on the size.
7 months ago
.macro put_vp8_bilin_h_v type mn
func ff_put_vp8_bilin4_\type\()_rvv, zve32x
vsetvlstatic8 4
.Lbilin_\type:
lavc/vp8dsp: R-V V put_bilin_h v C908: vp8_put_bilin4_h_c: 367.0 vp8_put_bilin4_h_rvv_i32: 137.7 vp8_put_bilin4_v_c: 377.0 vp8_put_bilin4_v_rvv_i32: 137.7 vp8_put_bilin8_h_c: 1431.0 vp8_put_bilin8_h_rvv_i32: 297.5 vp8_put_bilin8_v_c: 1449.0 vp8_put_bilin8_v_rvv_i32: 297.5 vp8_put_bilin16_h_c: 2839.0 vp8_put_bilin16_h_rvv_i32: 344.7 vp8_put_bilin16_v_c: 2857.0 vp8_put_bilin16_v_rvv_i32: 344.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
li t1, 8
li t4, 4
sub t1, t1, \mn
1:
lavc/vp8dsp: R-V V put_bilin_h v unroll Since len < 64, the registers are sufficient, so it can be directly unrolled (a4 is even). Another benefit of unrolling is that it reduces one load operation vertically compared to horizontally. old new C908 X60 C908 X60 vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
add t0, a2, a3
add t2, a0, a1
addi a4, a4, -2
.ifc \type,v
add t3, t0, a3
.else
addi t5, a2, 1
addi t3, t0, 1
vle8.v v2, (t5)
.endif
vle8.v v0, (a2)
vle8.v v4, (t0)
vle8.v v6, (t3)
vwmulu.vx v28, v0, t1
vwmulu.vx v26, v4, t1
.ifc \type,v
vwmaccu.vx v28, \mn, v4
.else
vwmaccu.vx v28, \mn, v2
.endif
vwmaccu.vx v26, \mn, v6
vwaddu.wx v24, v28, t4
vwaddu.wx v22, v26, t4
vnsra.wi v30, v24, 3
vnsra.wi v0, v22, 3
vse8.v v30, (a0)
vse8.v v0, (t2)
add a2, t0, a3
add a0, t2, a1
lavc/vp8dsp: R-V V put_bilin_h v C908: vp8_put_bilin4_h_c: 367.0 vp8_put_bilin4_h_rvv_i32: 137.7 vp8_put_bilin4_v_c: 377.0 vp8_put_bilin4_v_rvv_i32: 137.7 vp8_put_bilin8_h_c: 1431.0 vp8_put_bilin8_h_rvv_i32: 297.5 vp8_put_bilin8_v_c: 1449.0 vp8_put_bilin8_v_rvv_i32: 297.5 vp8_put_bilin16_h_c: 2839.0 vp8_put_bilin16_h_rvv_i32: 344.7 vp8_put_bilin16_v_c: 2857.0 vp8_put_bilin16_v_rvv_i32: 344.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
bnez a4, 1b
ret
endfunc
.endm
lavc/vp8dsp: factor R-V V bilin functions For a given type, only the first VSETVLI instruction varies depending on the size.
7 months ago
put_vp8_bilin_h_v h a5
put_vp8_bilin_h_v v a6
func ff_put_vp8_bilin4_hv_rvv, zve32x
vsetvlstatic8 4
.Lbilin_hv:
lavc/vp8dsp: R-V V put_bilin_hv C908: vp8_put_bilin4_hv_c: 561.0 vp8_put_bilin4_hv_rvv_i32: 232.7 vp8_put_bilin8_hv_c: 2162.7 vp8_put_bilin8_hv_rvv_i32: 506.7 vp8_put_bilin16_hv_c: 4769.7 vp8_put_bilin16_hv_rvv_i32: 556.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
li t3, 8
sub t1, t3, a5
sub t2, t3, a6
li t4, 4
lavc/vp8dsp: R-V V bilin_load to bilin_load_h Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
bilin_load_h v4, a5
lavc/vp8dsp: R-V V put_bilin_hv C908: vp8_put_bilin4_hv_c: 561.0 vp8_put_bilin4_hv_rvv_i32: 232.7 vp8_put_bilin8_hv_c: 2162.7 vp8_put_bilin8_hv_rvv_i32: 506.7 vp8_put_bilin16_hv_c: 4769.7 vp8_put_bilin16_hv_rvv_i32: 556.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
add a2, a2, a3
1:
addi a4, a4, -1
vwmulu.vx v20, v4, t2
lavc/vp8dsp: R-V V bilin_load to bilin_load_h Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
bilin_load_h v4, a5
lavc/vp8dsp: R-V V put_bilin_hv C908: vp8_put_bilin4_hv_c: 561.0 vp8_put_bilin4_hv_rvv_i32: 232.7 vp8_put_bilin8_hv_c: 2162.7 vp8_put_bilin8_hv_rvv_i32: 506.7 vp8_put_bilin16_hv_c: 4769.7 vp8_put_bilin16_hv_rvv_i32: 556.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vwmaccu.vx v20, a6, v4
vwaddu.wx v24, v20, t4
vnsra.wi v0, v24, 3
vse8.v v0, (a0)
add a2, a2, a3
add a0, a0, a1
bnez a4, 1b
ret
endfunc
lavc/vp8dsp: factor R-V V bilin functions For a given type, only the first VSETVLI instruction varies depending on the size.
7 months ago
.irp len,16,8
func ff_put_vp8_bilin\len\()_h_rvv, zve32x
vsetvlstatic8 \len
j .Lbilin_h
endfunc
func ff_put_vp8_bilin\len\()_v_rvv, zve32x
vsetvlstatic8 \len
j .Lbilin_v
endfunc
func ff_put_vp8_bilin\len\()_hv_rvv, zve32x
vsetvlstatic8 \len
j .Lbilin_hv
endfunc
.endr
lavc/vp8dsp: R-V V put_bilin_hv C908: vp8_put_bilin4_hv_c: 561.0 vp8_put_bilin4_hv_rvv_i32: 232.7 vp8_put_bilin8_hv_c: 2162.7 vp8_put_bilin8_hv_rvv_i32: 506.7 vp8_put_bilin16_hv_c: 4769.7 vp8_put_bilin16_hv_rvv_i32: 556.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
const subpel_filters
.byte 0, -6, 123, 12, -1, 0
.byte 2, -11, 108, 36, -8, 1
.byte 0, -9, 93, 50, -6, 0
.byte 3, -16, 77, 77, -16, 3
.byte 0, -6, 50, 93, -9, 0
.byte 1, -8, 36, 108, -11, 2
.byte 0, -1, 12, 123, -6, 0
endconst
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.macro epel_filter size type regtype
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.ifc \type,v
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
addi \regtype\()0, a6, -1
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.else
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
addi \regtype\()0, a5, -1
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endif
lavc/vp8dsp: avoid one multiplication on RISC-V Use shifts rather than multiply, and save one instruction.
7 months ago
lla \regtype\()2, subpel_filters
sh1add \regtype\()0, \regtype\()0, \regtype\()0
sh1add \regtype\()0, \regtype\()0, \regtype\()2
lavc/vp8dsp: fix .irp use with LLVM as
7 months ago
.irp n,1,2,3,4
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
lb \regtype\n, \n(\regtype\()0)
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endr
.ifc \size,6
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
lb \regtype\()5, 5(\regtype\()0)
lb \regtype\()0, (\regtype\()0)
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endif
.endm
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.macro epel_load dst len size type from_mem regtype
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.ifc \type,v
lavc/vp8dsp: save one R-V GPR This saves one instruction and frees up A5, which will be repurposed in later changes. Unfortunately, we need to add quite a lot of alternative code for this.
7 months ago
sub t6, a2, a3
add a7, a2, a3
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.else
lavc/vp8dsp: save one R-V GPR This saves one instruction and frees up A5, which will be repurposed in later changes. Unfortunately, we need to add quite a lot of alternative code for this.
7 months ago
addi t6, a2, -1
addi a7, a2, 1
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endif
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.if \from_mem
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vle8.v v24, (a2)
vle8.v v22, (t6)
vle8.v v26, (a7)
lavc/vp8dsp: save one R-V GPR This saves one instruction and frees up A5, which will be repurposed in later changes. Unfortunately, we need to add quite a lot of alternative code for this.
7 months ago
.ifc \type,v
add a7, a7, a3
.else
addi a7, a7, 1
.endif
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vle8.v v28, (a7)
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
vwmulu.vx v16, v24, \regtype\()2
vwmulu.vx v20, v26, \regtype\()3
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.ifc \size,6
lavc/vp8dsp: save one R-V GPR This saves one instruction and frees up A5, which will be repurposed in later changes. Unfortunately, we need to add quite a lot of alternative code for this.
7 months ago
.ifc \type,v
sub t6, t6, a3
add a7, a7, a3
.else
addi t6, t6, -1
addi a7, a7, 1
.endif
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vle8.v v24, (t6)
vle8.v v26, (a7)
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
vwmaccu.vx v16, \regtype\()0, v24
vwmaccu.vx v16, \regtype\()5, v26
.endif
vwmaccsu.vx v16, \regtype\()1, v22
vwmaccsu.vx v16, \regtype\()4, v28
.else
vwmulu.vx v16, v4, \regtype\()2
vwmulu.vx v20, v6, \regtype\()3
.ifc \size,6
vwmaccu.vx v16, \regtype\()0, v0
vwmaccu.vx v16, \regtype\()5, v10
.endif
vwmaccsu.vx v16, \regtype\()1, v2
vwmaccsu.vx v16, \regtype\()4, v8
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endif
li t6, 64
vwadd.wx v16, v16, t6
vsetvlstatic16 \len
vwadd.vv v24, v16, v20
vnsra.wi v24, v24, 7
vmax.vx v24, v24, zero
vsetvlstatic8 \len
vnclipu.wi \dst, v24, 0
.endm
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.macro epel_load_inc dst len size type from_mem regtype
epel_load \dst \len \size \type \from_mem \regtype
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
add a2, a2, a3
.endm
.macro epel len size type
func ff_put_vp8_epel\len\()_\type\()\size\()_rvv, zve32x
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
epel_filter \size \type t
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vsetvlstatic8 \len
1:
addi a4, a4, -1
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
epel_load_inc v30 \len \size \type 1 t
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
vse8.v v30, (a0)
add a0, a0, a1
bnez a4, 1b
ret
endfunc
.endm
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
.macro epel_hv len hsize vsize
func ff_put_vp8_epel\len\()_h\hsize\()v\vsize\()_rvv, zve32x
#if __riscv_xlen == 64
addi sp, sp, -48
.irp n,0,1,2,3,4,5
sd s\n, \n\()<<3(sp)
.endr
#else
addi sp, sp, -24
.irp n,0,1,2,3,4,5
sw s\n, \n\()<<2(sp)
.endr
#endif
sub a2, a2, a3
epel_filter \hsize h t
epel_filter \vsize v s
vsetvlstatic8 \len
.if \hsize == 6 || \vsize == 6
sub a2, a2, a3
epel_load_inc v0 \len \hsize h 1 t
.endif
epel_load_inc v2 \len \hsize h 1 t
epel_load_inc v4 \len \hsize h 1 t
epel_load_inc v6 \len \hsize h 1 t
epel_load_inc v8 \len \hsize h 1 t
.if \hsize == 6 || \vsize == 6
epel_load_inc v10 \len \hsize h 1 t
.endif
addi a4, a4, -1
1:
addi a4, a4, -1
epel_load v30 \len \vsize v 0 s
vse8.v v30, (a0)
.if \hsize == 6 || \vsize == 6
vmv.v.v v0, v2
.endif
vmv.v.v v2, v4
vmv.v.v v4, v6
vmv.v.v v6, v8
.if \hsize == 6 || \vsize == 6
vmv.v.v v8, v10
epel_load_inc v10 \len \hsize h 1 t
.else
epel_load_inc v8 \len 4 h 1 t
.endif
add a0, a0, a1
bnez a4, 1b
epel_load v30 \len \vsize v 0 s
vse8.v v30, (a0)
#if __riscv_xlen == 64
.irp n,0,1,2,3,4,5
ld s\n, \n\()<<3(sp)
.endr
addi sp, sp, 48
#else
.irp n,0,1,2,3,4,5
lw s\n, \n\()<<2(sp)
.endr
addi sp, sp, 24
#endif
ret
endfunc
.endm
lavc/vp8dsp: fix .irp use with LLVM as
7 months ago
.irp len,16,8,4
lavc/vp8dsp: R-V V put_epel h C908: vp8_put_epel4_h4_c: 10.7 vp8_put_epel4_h4_rvv_i32: 5.0 vp8_put_epel4_h6_c: 15.0 vp8_put_epel4_h6_rvv_i32: 6.2 vp8_put_epel8_h4_c: 43.2 vp8_put_epel8_h4_rvv_i32: 11.2 vp8_put_epel8_h6_c: 57.5 vp8_put_epel8_h6_rvv_i32: 13.5 vp8_put_epel16_h4_c: 92.5 vp8_put_epel16_h4_rvv_i32: 13.7 vp8_put_epel16_h6_c: 139.0 vp8_put_epel16_h6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
epel \len 6 h
epel \len 4 h
lavc/vp8dsp: R-V V put_epel v C908: vp8_put_epel4_v4_c: 11.0 vp8_put_epel4_v4_rvv_i32: 5.0 vp8_put_epel4_v6_c: 16.5 vp8_put_epel4_v6_rvv_i32: 6.2 vp8_put_epel8_v4_c: 43.7 vp8_put_epel8_v4_rvv_i32: 11.2 vp8_put_epel8_v6_c: 68.7 vp8_put_epel8_v6_rvv_i32: 13.2 vp8_put_epel16_v4_c: 92.5 vp8_put_epel16_v4_rvv_i32: 13.7 vp8_put_epel16_v6_c: 135.7 vp8_put_epel16_v6_rvv_i32: 16.5 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
epel \len 6 v
epel \len 4 v
lavc/vp8dsp: disable EPEL HV on RV128 RV128 is mostly scifi at this point, so we can just disable it here (the EPEL HV prologue/epilogue do not save 128-bit registers).
7 months ago
#if __riscv_xlen <= 64
lavc/vp8dsp: R-V V put_epel hv C908: vp8_put_epel4_h4v4_c: 20.0 vp8_put_epel4_h4v4_rvv_i32: 11.0 vp8_put_epel4_h4v6_c: 25.2 vp8_put_epel4_h4v6_rvv_i32: 13.5 vp8_put_epel4_h6v4_c: 22.2 vp8_put_epel4_h6v4_rvv_i32: 14.5 vp8_put_epel4_h6v6_c: 29.0 vp8_put_epel4_h6v6_rvv_i32: 15.7 vp8_put_epel8_h4v4_c: 73.0 vp8_put_epel8_h4v4_rvv_i32: 22.2 vp8_put_epel8_h4v6_c: 90.5 vp8_put_epel8_h4v6_rvv_i32: 26.7 vp8_put_epel8_h6v4_c: 85.0 vp8_put_epel8_h6v4_rvv_i32: 27.2 vp8_put_epel8_h6v6_c: 104.7 vp8_put_epel8_h6v6_rvv_i32: 29.5 vp8_put_epel16_h4v4_c: 145.5 vp8_put_epel16_h4v4_rvv_i32: 26.5 vp8_put_epel16_h4v6_c: 190.7 vp8_put_epel16_h4v6_rvv_i32: 47.5 vp8_put_epel16_h6v4_c: 173.7 vp8_put_epel16_h6v4_rvv_i32: 33.2 vp8_put_epel16_h6v6_c: 222.2 vp8_put_epel16_h6v6_rvv_i32: 35.5 Amended to disable unsupported RV128. Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
7 months ago
epel_hv \len 6 6
epel_hv \len 4 4
epel_hv \len 6 4
epel_hv \len 4 6
lavc/vp8dsp: disable EPEL HV on RV128 RV128 is mostly scifi at this point, so we can just disable it here (the EPEL HV prologue/epilogue do not save 128-bit registers).
7 months ago
#endif
lavc/vp8dsp: R-V V put_bilin_h v C908: vp8_put_bilin4_h_c: 367.0 vp8_put_bilin4_h_rvv_i32: 137.7 vp8_put_bilin4_v_c: 377.0 vp8_put_bilin4_v_rvv_i32: 137.7 vp8_put_bilin8_h_c: 1431.0 vp8_put_bilin8_h_rvv_i32: 297.5 vp8_put_bilin8_v_c: 1449.0 vp8_put_bilin8_v_rvv_i32: 297.5 vp8_put_bilin16_h_c: 2839.0 vp8_put_bilin16_h_rvv_i32: 344.7 vp8_put_bilin16_v_c: 2857.0 vp8_put_bilin16_v_rvv_i32: 344.7 Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>
8 months ago
.endr