Re: [PATCH v4 02/12] riscv: Add vector extension XOR implementation

["Rémi Denis-Courmont" <remi@remlab.net>]

Le tiistaina 11. heinäkuuta 2023, 18.37.33 EEST Heiko Stuebner a écrit :
> diff --git a/arch/riscv/lib/xor.S b/arch/riscv/lib/xor.S
> new file mode 100644
> index 000000000000..3bc059e18171
> --- /dev/null
> +++ b/arch/riscv/lib/xor.S
> @@ -0,0 +1,81 @@
> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> +/*
> + * Copyright (C) 2021 SiFive
> + */
> +#include <linux/linkage.h>
> +#include <asm-generic/export.h>
> +#include <asm/asm.h>
> +
> +ENTRY(xor_regs_2_)
> +	vsetvli a3, a0, e8, m8, ta, ma

AFAICT, so far, Linux only uses `vsetvli` to save/restore/flush vectors, and 
that's, of course, with LMUL=8, so that's not really telling much anything. 
This function could be the first actual vector optimisation in kernel if/when 
it gets merged.

Should the same group multiplier be used for "actual" vector loops throughout 
the kernel? I've seen conflicting advises or opinions here. Should kernel code 
always use the maximum possible LMUL, depending on register pressure of the 
loop? Or will that just increase latency with no bandwidth gains compared to, 
say, LMUL=1 or LMUL=2?

> +	vle8.v v0, (a1)
> +	vle8.v v8, (a2)
> +	sub a0, a0, a3
> +	vxor.vv v16, v0, v8
> +	add a2, a2, a3
> +	vse8.v v16, (a1)
> +	add a1, a1, a3
> +	bnez a0, xor_regs_2_
> +	ret
> +END(xor_regs_2_)
> +EXPORT_SYMBOL(xor_regs_2_)
> +
> +ENTRY(xor_regs_3_)
> +	vsetvli a4, a0, e8, m8, ta, ma
> +	vle8.v v0, (a1)
> +	vle8.v v8, (a2)
> +	sub a0, a0, a4
> +	vxor.vv v0, v0, v8
> +	vle8.v v16, (a3)
> +	add a2, a2, a4
> +	vxor.vv v16, v0, v16
> +	add a3, a3, a4
> +	vse8.v v16, (a1)
> +	add a1, a1, a4
> +	bnez a0, xor_regs_3_
> +	ret
> +END(xor_regs_3_)
> +EXPORT_SYMBOL(xor_regs_3_)
> +
> +ENTRY(xor_regs_4_)
> +	vsetvli a5, a0, e8, m8, ta, ma
> +	vle8.v v0, (a1)
> +	vle8.v v8, (a2)
> +	sub a0, a0, a5
> +	vxor.vv v0, v0, v8
> +	vle8.v v16, (a3)
> +	add a2, a2, a5
> +	vxor.vv v0, v0, v16
> +	vle8.v v24, (a4)
> +	add a3, a3, a5
> +	vxor.vv v16, v0, v24
> +	add a4, a4, a5
> +	vse8.v v16, (a1)
> +	add a1, a1, a5
> +	bnez a0, xor_regs_4_
> +	ret
> +END(xor_regs_4_)
> +EXPORT_SYMBOL(xor_regs_4_)
> +
> +ENTRY(xor_regs_5_)
> +	vsetvli a6, a0, e8, m8, ta, ma
> +	vle8.v v0, (a1)
> +	vle8.v v8, (a2)
> +	sub a0, a0, a6
> +	vxor.vv v0, v0, v8
> +	vle8.v v16, (a3)
> +	add a2, a2, a6
> +	vxor.vv v0, v0, v16
> +	vle8.v v24, (a4)
> +	add a3, a3, a6
> +	vxor.vv v0, v0, v24
> +	vle8.v v8, (a5)
> +	add a4, a4, a6
> +	vxor.vv v16, v0, v8
> +	add a5, a5, a6
> +	vse8.v v16, (a1)
> +	add a1, a1, a6
> +	bnez a0, xor_regs_5_
> +	ret
> +END(xor_regs_5_)
> +EXPORT_SYMBOL(xor_regs_5_)


-- 
レミ・デニ-クールモン
http://www.remlab.net/