[FFmpeg-devel] [PATCH] lavc/lpc: R-V V apply_welch_window

[Anton Khirnov]

Quoting Rémi Denis-Courmont (2023-12-08 18:46:51)
> +#if __riscv_xlen >= 64
> +func ff_lpc_apply_welch_window_rvv, zve64d
> +        vsetvli t0, zero, e64, m8, ta, ma
> +        vid.v   v0
> +        addi    t2, a1, -1
> +        vfcvt.f.xu.v v0, v0
> +        li      t3, 2
> +        fcvt.d.l ft2, t2
> +        srai    t1, a1, 1
> +        fcvt.d.l ft3, t3
> +        li      t4, 1
> +        fdiv.d  ft0, ft3, ft2    # ft0 = c = 2. / (len - 1)
> +        fcvt.d.l fa1, t4         # fa1 = 1.
> +        fsub.d  ft1, ft0, fa1
> +        vfrsub.vf v0, v0, ft1    # v0[i] = c - i - 1.
> +1:
> +        vsetvli t0, t1, e64, m8, ta, ma
> +        vfmul.vv v16, v0, v0  # no fused multipy-add as v0 is reused
> +        sub     t1, t1, t0
> +        vle32.v v8, (a0)
> +        fcvt.d.l ft2, t0
> +        vfrsub.vf v16, v16, fa1  # v16 = 1. - w * w
> +        sh2add  a0, t0, a0
> +        vsetvli zero, zero, e32, m4, ta, ma
> +        vfwcvt.f.x.v v24, v8
> +        vsetvli zero, zero, e64, m8, ta, ma
> +        vfsub.vf v0, v0, ft2     # v0 -= vl
> +        vfmul.vv v8, v24, v16
> +        vse64.v v8, (a2)
> +        sh3add  a2, t0, a2
> +        bnez    t1, 1b
> +
> +        andi    t1, a1, 1
> +        beqz    t1, 2f
> +
> +        sd      zero, (a2)
> +        addi    a0, a0, 4
> +        addi    a2, a2, 8
> +2:
> +        vsetvli t0, zero, e64, m8, ta, ma
> +        vid.v   v0
> +        srai    t1, a1, 1
> +        vfcvt.f.xu.v v0, v0
> +        fcvt.d.l ft1, t1
> +        fsub.d  ft1, ft0, ft1    # ft1 = c - (len / 2)
> +        vfadd.vf v0, v0, ft1     # v0[i] = c - (len / 2) + i
> +3:
> +        vsetvli t0, t1, e64, m8, ta, ma
> +        vfmul.vv v16, v0, v0
> +        sub     t1, t1, t0
> +        vle32.v v8, (a0)
> +        fcvt.d.l ft2, t0
> +        vfrsub.vf v16, v16, fa1  # v16 = 1. - w * w
> +        sh2add  a0, t0, a0
> +        vsetvli zero, zero, e32, m4, ta, ma
> +        vfwcvt.f.x.v v24, v8
> +        vsetvli zero, zero, e64, m8, ta, ma
> +        vfadd.vf v0, v0, ft2     # v0 += vl
> +        vfmul.vv v8, v24, v16
> +        vse64.v v8, (a2)
> +        sh3add  a2, t0, a2
> +        bnez    t1, 3b

I think it'd look a lot less like base64 < /dev/random if you vertically
aligned the first operands.

-- 
Anton Khirnov