librempeg/libavutil/riscv/float_dsp_rvv.S
Rémi Denis-Courmont e49f41fb27 lavu/float_dsp: optimise R-V V fmul_reverse & fmul_window
Roll the loop to avoid slow gathers.

Before:
vector_fmul_reverse_c:       1561.7
vector_fmul_reverse_rvv_f32: 2410.2
vector_fmul_window_c:        2068.2
vector_fmul_window_rvv_f32:  1879.5

After:
vector_fmul_reverse_c:       1561.7
vector_fmul_reverse_rvv_f32:  916.2
vector_fmul_window_c:        2068.2
vector_fmul_window_rvv_f32:  1202.5
2023-11-23 18:57:18 +02:00

252 lines
6.5 KiB
ArmAsm

/*
* Copyright © 2022 Rémi Denis-Courmont.
*
* 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 "asm.S"
// (a0) = (a1) * (a2) [0..a3-1]
func ff_vector_fmul_rvv, zve32f
1:
vsetvli t0, a3, e32, m8, ta, ma
vle32.v v16, (a1)
sub a3, a3, t0
vle32.v v24, (a2)
sh2add a1, t0, a1
vfmul.vv v16, v16, v24
sh2add a2, t0, a2
vse32.v v16, (a0)
sh2add a0, t0, a0
bnez a3, 1b
ret
endfunc
// (a0) += (a1) * fa0 [0..a2-1]
func ff_vector_fmac_scalar_rvv, zve32f
NOHWF fmv.w.x fa0, a2
NOHWF mv a2, a3
1:
vsetvli t0, a2, e32, m8, ta, ma
slli t1, t0, 2
vle32.v v24, (a1)
sub a2, a2, t0
vle32.v v16, (a0)
sh2add a1, t0, a1
vfmacc.vf v16, fa0, v24
vse32.v v16, (a0)
sh2add a0, t0, a0
bnez a2, 1b
ret
endfunc
// (a0) = (a1) * fa0 [0..a2-1]
func ff_vector_fmul_scalar_rvv, zve32f
NOHWF fmv.w.x fa0, a2
NOHWF mv a2, a3
1:
vsetvli t0, a2, e32, m8, ta, ma
vle32.v v16, (a1)
sub a2, a2, t0
vfmul.vf v16, v16, fa0
sh2add a1, t0, a1
vse32.v v16, (a0)
sh2add a0, t0, a0
bnez a2, 1b
ret
endfunc
func ff_vector_fmul_window_rvv, zve32f
// a0: dst, a1: src0, a2: src1, a3: window, a4: length
// e16/m2 and e32/m4 are possible but slower due to gather.
vsetvli t0, zero, e16, m1, ta, ma
sh2add a2, a4, a2
vid.v v0
sh3add t3, a4, a3
vadd.vi v0, v0, 1
sh3add t0, a4, a0
1:
vsetvli t2, a4, e16, m1, ta, ma
slli t4, t2, 2
vrsub.vx v2, v0, t2
sub t3, t3, t4
vsetvli zero, zero, e32, m2, ta, ma
sub a2, a2, t4
vle32.v v8, (t3)
sub t0, t0, t4
vle32.v v4, (a2)
sub a4, a4, t2
vrgatherei16.vv v28, v8, v2
vle32.v v16, (a1)
add a1, a1, t4
vrgatherei16.vv v20, v4, v2
vle32.v v24, (a3)
add a3, a3, t4
vfmul.vv v12, v16, v28
vfmul.vv v16, v16, v24
vfnmsac.vv v12, v20, v24
vfmacc.vv v16, v20, v28
vrgatherei16.vv v8, v16, v2
vse32.v v12, (a0)
add a0, a0, t4
vse32.v v8, (t0)
bnez a4, 1b
ret
endfunc
// (a0) = (a1) * (a2) + (a3) [0..a4-1]
func ff_vector_fmul_add_rvv, zve32f
1:
vsetvli t0, a4, e32, m8, ta, ma
vle32.v v8, (a1)
sub a4, a4, t0
vle32.v v16, (a2)
sh2add a1, t0, a1
vle32.v v24, (a3)
sh2add a2, t0, a2
vfmadd.vv v8, v16, v24
sh2add a3, t0, a3
vse32.v v8, (a0)
sh2add a0, t0, a0
bnez a4, 1b
ret
endfunc
// TODO factor vrsub, separate last iteration?
// (a0) = (a1) * reverse(a2) [0..a3-1]
func ff_vector_fmul_reverse_rvv, zve32f
// e16/m4 and e32/m8 are possible but slower due to gather.
vsetvli t0, zero, e16, m1, ta, ma
sh2add a2, a3, a2
vid.v v0
vadd.vi v0, v0, 1
1:
vsetvli t0, a3, e16, m1, ta, ma
slli t1, t0, 2
vrsub.vx v4, v0, t0 // v4[i] = [VL-1, VL-2... 1, 0]
sub a2, a2, t1
vsetvli zero, zero, e32, m2, ta, ma
vle32.v v8, (a2)
sub a3, a3, t0
vle32.v v16, (a1)
add a1, a1, t1
vrgatherei16.vv v24, v8, v4 // v24 = reverse(v8)
vfmul.vv v16, v16, v24
vse32.v v16, (a0)
add a0, a0, t1
bnez a3, 1b
ret
endfunc
// (a0) = (a0) + (a1), (a1) = (a0) - (a1) [0..a2-1]
func ff_butterflies_float_rvv, zve32f
1:
vsetvli t0, a2, e32, m8, ta, ma
vle32.v v16, (a0)
sub a2, a2, t0
vle32.v v24, (a1)
vfadd.vv v0, v16, v24
vfsub.vv v8, v16, v24
vse32.v v0, (a0)
sh2add a0, t0, a0
vse32.v v8, (a1)
sh2add a1, t0, a1
bnez a2, 1b
ret
endfunc
// a0 = (a0).(a1) [0..a2-1]
func ff_scalarproduct_float_rvv, zve32f
vsetvli t0, zero, e32, m8, ta, ma
vmv.v.x v8, zero
vmv.s.x v0, zero
1:
vsetvli t0, a2, e32, m8, tu, ma
vle32.v v16, (a0)
sub a2, a2, t0
vle32.v v24, (a1)
sh2add a0, t0, a0
vfmacc.vv v8, v16, v24
sh2add a1, t0, a1
bnez a2, 1b
vsetvli t0, zero, e32, m8, ta, ma
vfredusum.vs v0, v8, v0
vfmv.f.s fa0, v0
NOHWF fmv.x.w a0, fa0
ret
endfunc
// (a0) = (a1) * (a2) [0..a3-1]
func ff_vector_dmul_rvv, zve64d
1:
vsetvli t0, a3, e64, m8, ta, ma
vle64.v v16, (a1)
sub a3, a3, t0
vle64.v v24, (a2)
sh3add a1, t0, a1
vfmul.vv v16, v16, v24
sh3add a2, t0, a2
vse64.v v16, (a0)
sh3add a0, t0, a0
bnez a3, 1b
ret
endfunc
// (a0) += (a1) * fa0 [0..a2-1]
func ff_vector_dmac_scalar_rvv, zve64d
NOHWD fmv.d.x fa0, a2
NOHWD mv a2, a3
1:
vsetvli t0, a2, e64, m8, ta, ma
vle64.v v24, (a1)
sub a2, a2, t0
vle64.v v16, (a0)
sh3add a1, t0, a1
vfmacc.vf v16, fa0, v24
vse64.v v16, (a0)
sh3add a0, t0, a0
bnez a2, 1b
ret
endfunc
// (a0) = (a1) * fa0 [0..a2-1]
func ff_vector_dmul_scalar_rvv, zve64d
NOHWD fmv.d.x fa0, a2
NOHWD mv a2, a3
1:
vsetvli t0, a2, e64, m8, ta, ma
vle64.v v16, (a1)
sub a2, a2, t0
vfmul.vf v16, v16, fa0
sh3add a1, t0, a1
vse64.v v16, (a0)
sh3add a0, t0, a0
bnez a2, 1b
ret
endfunc