ffmpeg/libavcodec/riscv/h264qpel_rvv.S

465 lines
16 KiB
ArmAsm

/*
* SPDX-License-Identifier: BSD-2-Clause
*
* Copyright (c) 2024 Niklas Haas
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include "libavutil/riscv/asm.S"
.macro lx rd, addr
#if (__riscv_xlen == 32)
lw \rd, \addr
#elif (__riscv_xlen == 64)
ld \rd, \addr
#else
lq \rd, \addr
#endif
.endm
.macro sx rd, addr
#if (__riscv_xlen == 32)
sw \rd, \addr
#elif (__riscv_xlen == 64)
sd \rd, \addr
#else
sq \rd, \addr
#endif
.endm
/* output is unclipped; clobbers v26-v31 plus t0 and t02 */
.macro lowpass_h vdst, src
addi t4, \src, 3
lbu t5, 2(\src)
vle8.v v31, (t4)
lbu t4, 1(\src)
vslide1up.vx v30, v31, t5
lbu t5, 0(\src)
vslide1up.vx v29, v30, t4
lbu t4, -1(\src)
vslide1up.vx v28, v29, t5
lbu t5, -2(\src)
vslide1up.vx v27, v28, t4
vslide1up.vx v26, v27, t5
vwaddu.vv \vdst, v26, v31
vwmaccu.vx \vdst, t6, v28
vwmaccu.vx \vdst, t6, v29
vwmaccsu.vx \vdst, a7, v27
vwmaccsu.vx \vdst, a7, v30
.endm
/* output is unclipped */
.macro lowpass_v vdst, vsrc0, vsrc1, vsrc2, vsrc3, vsrc4, vsrc5, signed=0
.if \signed
vwadd.vv \vdst, \vsrc0, \vsrc5
vwmacc.vx \vdst, t6, \vsrc2
vwmacc.vx \vdst, t6, \vsrc3
vwmacc.vx \vdst, a7, \vsrc1
vwmacc.vx \vdst, a7, \vsrc4
.else
vwaddu.vv \vdst, \vsrc0, \vsrc5
vwmaccu.vx \vdst, t6, \vsrc2
vwmaccu.vx \vdst, t6, \vsrc3
vwmaccsu.vx \vdst, a7, \vsrc1
vwmaccsu.vx \vdst, a7, \vsrc4
.endif
.endm
.macro qpel_mc00 op, dst, src, stride, size
func ff_\op\()_h264_qpel_pixels, zve32x
1: add t1, a2, a1
add t2, a2, t1
add t3, a2, t2
vle8.v v0, (a1)
vle8.v v1, (t1)
vle8.v v2, (t2)
vle8.v v3, (t3)
addi a4, a4, -4
add a1, a2, t3
add t1, a2, a0
add t2, a2, t1
add t3, a2, t2
.ifc \op, avg
vle8.v v4, (a0)
vle8.v v5, (t1)
vle8.v v6, (t2)
vle8.v v7, (t3)
vaaddu.vv v0, v0, v4
vaaddu.vv v1, v1, v5
vaaddu.vv v2, v2, v6
vaaddu.vv v3, v3, v7
.endif
vse8.v v0, (a0)
vse8.v v1, (t1)
vse8.v v2, (t2)
vse8.v v3, (t3)
add a0, a2, t3
bnez a4, 1b
jr t0
endfunc
.endm
qpel_mc00 put, a0, a1, a2, a4
qpel_mc00 avg, a0, a1, a2, a4
.macro qpel_lowpass op, ext, lmul, lmul2
func ff_\op\()_h264_qpel_h_lowpass_\lmul\ext, zve32x
1: add t1, a3, a1
add t2, a3, t1
add t3, a3, t2
lowpass_h v0, a1
lowpass_h v2, t1
lowpass_h v4, t2
lowpass_h v6, t3
add a1, a3, t3
addi a4, a4, -4
vsetvli zero, zero, e16, \lmul2, ta, ma
vmax.vx v0, v0, zero
vmax.vx v2, v2, zero
vmax.vx v4, v4, zero
vmax.vx v6, v6, zero
vsetvli zero, zero, e8, \lmul, ta, ma
vnclipu.wi v0, v0, 5
vnclipu.wi v2, v2, 5
vnclipu.wi v4, v4, 5
vnclipu.wi v6, v6, 5
.ifc \ext, _l2
add t1, a6, a5
add t2, a6, t1
add t3, a6, t2
vle8.v v8, (a5)
vle8.v v10, (t1)
vle8.v v12, (t2)
vle8.v v14, (t3)
add a5, a2, t3
vaaddu.vv v0, v0, v8
vaaddu.vv v2, v2, v10
vaaddu.vv v4, v4, v12
vaaddu.vv v6, v6, v14
.endif
add t1, a2, a0
add t2, a2, t1
add t3, a2, t2
.ifc \op, avg
vle8.v v1, (a0)
vle8.v v3, (t1)
vle8.v v5, (t2)
vle8.v v7, (t3)
vaaddu.vv v0, v0, v1
vaaddu.vv v2, v2, v3
vaaddu.vv v4, v4, v5
vaaddu.vv v6, v6, v7
.endif
vse8.v v0, (a0)
vse8.v v2, (t1)
vse8.v v4, (t2)
vse8.v v6, (t3)
add a0, a2, t3
bnez a4, 1b
jr t0
endfunc
func ff_\op\()_h264_qpel_v_lowpass_\lmul\ext, zve32x
sub t1, a1, a3
sub t2, t1, a3
vle8.v v2, (a1)
vle8.v v1, (t1)
vle8.v v0, (t2)
add t1, a1, a3
add t2, t1, a3
add a1, t2, a3
vle8.v v3, (t1)
vle8.v v4, (t2)
1: add t1, a3, a1
add t2, a3, t1
add t3, a3, t2
vle8.v v5, (a1)
vle8.v v6, (t1)
vle8.v v7, (t2)
vle8.v v8, (t3)
add a1, a3, t3
lowpass_v v24, v0, v1, v2, v3, v4, v5
lowpass_v v26, v1, v2, v3, v4, v5, v6
lowpass_v v28, v2, v3, v4, v5, v6, v7
lowpass_v v30, v3, v4, v5, v6, v7, v8
addi a4, a4, -4
vsetvli zero, zero, e16, \lmul2, ta, ma
vmax.vx v24, v24, zero
vmax.vx v26, v26, zero
vmax.vx v28, v28, zero
vmax.vx v30, v30, zero
vsetvli zero, zero, e8, \lmul, ta, ma
vnclipu.wi v24, v24, 5
vnclipu.wi v26, v26, 5
vnclipu.wi v28, v28, 5
vnclipu.wi v30, v30, 5
.ifc \ext, _l2
add t1, a6, a5
add t2, a6, t1
add t3, a6, t2
vle8.v v9, (a5)
vle8.v v10, (t1)
vle8.v v11, (t2)
vle8.v v12, (t3)
add a5, a6, t3
vaaddu.vv v24, v24, v9
vaaddu.vv v26, v26, v10
vaaddu.vv v28, v28, v11
vaaddu.vv v30, v30, v12
.endif
add t1, a2, a0
add t2, a2, t1
add t3, a2, t2
.ifc \op, avg
vle8.v v9, (a0)
vle8.v v10, (t1)
vle8.v v11, (t2)
vle8.v v12, (t3)
vaaddu.vv v24, v24, v9
vaaddu.vv v26, v26, v10
vaaddu.vv v28, v28, v11
vaaddu.vv v30, v30, v12
.endif
vse8.v v24, (a0)
vse8.v v26, (t1)
vse8.v v28, (t2)
vse8.v v30, (t3)
add a0, a2, t3
vmv.v.v v0, v4
vmv.v.v v1, v5
vmv.v.v v2, v6
vmv.v.v v3, v7
vmv.v.v v4, v8
bnez a4, 1b
jr t0
endfunc
func ff_\op\()_h264_qpel_hv_lowpass_\lmul\ext, zve32x
sub t1, a1, a3
sub t2, t1, a3
lowpass_h v4, a1
lowpass_h v2, t1
lowpass_h v0, t2
add t1, a1, a3
add t2, t1, a3
add a1, t2, a3
lowpass_h v6, t1
lowpass_h v8, t2
1: add t1, a3, a1
add t2, a3, t1
add t3, a3, t2
lowpass_h v10, a1
lowpass_h v12, t1
lowpass_h v14, t2
lowpass_h v16, t3
vsetvli zero, zero, e16, \lmul2, ta, ma
addi a4, a4, -4
lowpass_v v20, v0, v2, v4, v6, v8, v10, signed=1
lowpass_v v24, v2, v4, v6, v8, v10, v12, signed=1
lowpass_v v28, v4, v6, v8, v10, v12, v14, signed=1
vnclip.wi v0, v20, 10
lowpass_v v20, v6, v8, v10, v12, v14, v16, signed=1
vnclip.wi v2, v24, 10
vnclip.wi v4, v28, 10
vnclip.wi v6, v20, 10
vmax.vx v18, v0, zero
vmax.vx v20, v2, zero
vmax.vx v22, v4, zero
vmax.vx v24, v6, zero
vmv.v.v v0, v8
vmv.v.v v2, v10
vmv.v.v v4, v12
vmv.v.v v6, v14
vmv.v.v v8, v16
add a1, a3, t3
vsetvli zero, zero, e8, \lmul, ta, ma
vnclipu.wi v18, v18, 0
vnclipu.wi v20, v20, 0
vnclipu.wi v22, v22, 0
vnclipu.wi v24, v24, 0
.ifc \ext, _l2
add t1, a6, a5
add t2, a6, t1
add t3, a6, t2
vle8.v v26, (a5)
vle8.v v27, (t1)
vle8.v v28, (t2)
vle8.v v29, (t3)
add a5, a6, t3
vaaddu.vv v18, v18, v26
vaaddu.vv v20, v20, v27
vaaddu.vv v22, v22, v28
vaaddu.vv v24, v24, v29
.endif
add t1, a2, a0
add t2, a2, t1
add t3, a2, t2
.ifc \op, avg
vle8.v v26, (a0)
vle8.v v27, (t1)
vle8.v v28, (t2)
vle8.v v29, (t3)
vaaddu.vv v18, v18, v26
vaaddu.vv v20, v20, v27
vaaddu.vv v22, v22, v28
vaaddu.vv v24, v24, v29
.endif
vse8.v v18, (a0)
vse8.v v20, (t1)
vse8.v v22, (t2)
vse8.v v24, (t3)
add a0, a2, t3
bnez a4, 1b
jr t0
endfunc
.endm
/* Note: We could possibly specialize for the width 8 / width 4 cases by
loading 32 bit integers, but this makes the convolutions more complicated
to implement, so it's not necessarily any faster. */
.macro h264_qpel lmul, lmul2
qpel_lowpass put, , \lmul, \lmul2
qpel_lowpass put, _l2, \lmul, \lmul2
qpel_lowpass avg, , \lmul, \lmul2
qpel_lowpass avg, _l2, \lmul, \lmul2
.endm
h264_qpel m1, m2
h264_qpel mf2, m1
h264_qpel mf4, mf2
h264_qpel mf8, mf4
.macro h264_qpel_1pass op, case, lmul, size, ext=rvv, dir, offset
func ff_\op\()_h264_qpel\size\()_\case\()_\ext, zve32x
lpad 0
vsetivli zero, \size, e8, \lmul, ta, ma
csrwi vxrm, 0
li a4, \size
li t6, 20
li a7, -5
mv a3, a2
mv t0, ra
.ifnb \offset
.ifc \dir, v
add a5, a1, \offset
.else
addi a5, a1, \offset
.endif
mv a6, a3
j ff_\op\()_h264_qpel_\dir\()_lowpass_\lmul\()_l2
.else
j ff_\op\()_h264_qpel_\dir\()_lowpass_\lmul\()
.endif
endfunc
.endm
.macro h264_qpel_2pass op, case, lmul, size, ext=rvv, dir1, dir2, off1=0, off2
func ff_\op\()_h264_qpel\size\()_\case\()_\ext, zve32x
lpad 0
vsetivli zero, \size, e8, \lmul, ta, ma
csrwi vxrm, 0
addi sp, sp, (-(__riscv_xlen >> 2))
li a4, \size
li t6, 20
li a7, -5
sx a0, 0(sp)
sx a1, (__riscv_xlen >> 3)(sp)
.ifc \off1, a2
add a1, a1, \off1
.elseif \off1
addi a1, a1, \off1
.endif
mv a3, a2
.ifc \op, avg
// Use temporary array on stack for the first pass
addi a0, sp, -(\size * \size)
li a2, \size
.endif
jal t0, ff_put_h264_qpel_\dir1\()_lowpass_\lmul
lx a0, 0(sp)
lx a1, (__riscv_xlen >> 3)(sp)
.ifc \op, put
// Directly reuse the first pass output buffer
mv a5, a0
mv a6, a2
.else
addi a5, sp, -(\size * \size)
li a6, \size
mv a2, a3
.endif
.ifnb \off2
addi a1, a1, \off2
.endif
li a4, \size
mv t0, ra
addi sp, sp, 16
j ff_\op\()_h264_qpel_\dir2\()_lowpass_\lmul\()_l2
endfunc
.endm
.macro ff_h264_qpel_fns op, lmul, size, ext=rvv
func ff_\op\()_h264_qpel\size\()_mc00_\ext, zve32x
lpad 0
vsetivli zero, \size, e8, \lmul, ta, ma
csrwi vxrm, 0
li a4, \size
mv t0, ra
j ff_\op\()_h264_qpel_pixels
endfunc
h264_qpel_1pass \op, mc20, \lmul, \size, \ext, h
h264_qpel_1pass \op, mc02, \lmul, \size, \ext, v
h264_qpel_1pass \op, mc10, \lmul, \size, \ext, h, 0
h264_qpel_1pass \op, mc30, \lmul, \size, \ext, h, 1
h264_qpel_1pass \op, mc01, \lmul, \size, \ext, v, zero
h264_qpel_1pass \op, mc03, \lmul, \size, \ext, v, a2
h264_qpel_1pass \op, mc22, \lmul, \size, \ext, hv
h264_qpel_2pass \op, mc11, \lmul, \size, \ext, h, v
h264_qpel_2pass \op, mc21, \lmul, \size, \ext, h, hv
h264_qpel_2pass \op, mc12, \lmul, \size, \ext, v, hv
h264_qpel_2pass \op, mc31, \lmul, \size, \ext, h, v, off2=1
h264_qpel_2pass \op, mc13, \lmul, \size, \ext, h, v, a2
h264_qpel_2pass \op, mc33, \lmul, \size, \ext, h, v, a2, 1
h264_qpel_2pass \op, mc23, \lmul, \size, \ext, h, hv, a2
h264_qpel_2pass \op, mc32, \lmul, \size, \ext, v, hv, 1
.endm
ff_h264_qpel_fns put, mf2, 16, rvv256
ff_h264_qpel_fns put, mf4, 8, rvv256
/* ff_h264_qpel_fns put, mf8, 4, rvv256 */
ff_h264_qpel_fns avg, mf2, 16, rvv256
ff_h264_qpel_fns avg, mf4, 8, rvv256
/* ff_h264_qpel_fns avg, mf8, 4, rvv256 */
ff_h264_qpel_fns put, m1, 16, rvv
ff_h264_qpel_fns put, mf2, 8, rvv
ff_h264_qpel_fns put, mf4, 4, rvv
ff_h264_qpel_fns avg, m1, 16, rvv
ff_h264_qpel_fns avg, mf2, 8, rvv
ff_h264_qpel_fns avg, mf4, 4, rvv