[FFmpeg-devel,2/2] lavc/lpc: R-V V compute_autocorr

The loop iterates over the length of the vector, not the order. This is
to avoid reloading the same data for each lag value. However this means
the loop only works if the maximum order is no larger than VLENB.

The loop is roughly equivalent to:

    for (size_t j = 0; j < lag; j++)
        autoc[j] = 1.;

    while (len > lag) {
        for (ptrdiff_t j = 0; j < lag; j++)
            autoc[j] += data[j] * *data;
        data++;
        len--;
    }

    while (len > 0) {
        for (ptrdiff_t j = 0; j < len; j++)
            autoc[j] += data[j] * *data;
        data++;
        len--;
    }

Since register pressure is only at 50%, it should be possible to implement
the same loop for order up to 2xVLENB. But this is left for future work.

Performance numbers are all over the place from ~1.25x to ~4x speedups,
but at least they are always noticeably better than nothing.
---
 libavcodec/riscv/lpc_init.c |  8 +++++++-
 libavcodec/riscv/lpc_rvv.S  | 29 +++++++++++++++++++++++++++++
 2 files changed, 36 insertions(+), 1 deletion(-)

Le tiistaina 12. joulukuuta 2023, 23.02.40 EET Rémi Denis-Courmont a écrit :
> The loop iterates over the length of the vector, not the order. This is
> to avoid reloading the same data for each lag value. However this means
> the loop only works if the maximum order is no larger than VLENB.
> 
> The loop is roughly equivalent to:
> 
>     for (size_t j = 0; j < lag; j++)
>         autoc[j] = 1.;
> 
>     while (len > lag) {
>         for (ptrdiff_t j = 0; j < lag; j++)
>             autoc[j] += data[j] * *data;
>         data++;
>         len--;
>     }
> 
>     while (len > 0) {
>         for (ptrdiff_t j = 0; j < len; j++)
>             autoc[j] += data[j] * *data;
>         data++;
>         len--;
>     }
> 
> Since register pressure is only at 50%, it should be possible to implement
> the same loop for order up to 2xVLENB. But this is left for future work.
> 
> Performance numbers are all over the place from ~1.25x to ~4x speedups,
> but at least they are always noticeably better than nothing.
> ---
>  libavcodec/riscv/lpc_init.c |  8 +++++++-
>  libavcodec/riscv/lpc_rvv.S  | 29 +++++++++++++++++++++++++++++
>  2 files changed, 36 insertions(+), 1 deletion(-)
> 
> diff --git a/libavcodec/riscv/lpc_init.c b/libavcodec/riscv/lpc_init.c
> index c16e5745f0..ab91956f2d 100644
> --- a/libavcodec/riscv/lpc_init.c
> +++ b/libavcodec/riscv/lpc_init.c
> @@ -22,16 +22,22 @@
> 
>  #include "libavutil/attributes.h"
>  #include "libavutil/cpu.h"
> +#include "libavutil/riscv/cpu.h"
>  #include "libavcodec/lpc.h"
> 
>  void ff_lpc_apply_welch_window_rvv(const int32_t *, ptrdiff_t, double *);
> +void ff_lpc_compute_autocorr_rvv(const double *, ptrdiff_t, int, double *);
> 
>  av_cold void ff_lpc_init_riscv(LPCContext *c)
>  {
>  #if HAVE_RVV && (__riscv_xlen >= 64)
>      int flags = av_get_cpu_flags();
> 
> -    if ((flags & AV_CPU_FLAG_RVV_F64) && (flags & AV_CPU_FLAG_RVB_ADDR))
> +    if ((flags & AV_CPU_FLAG_RVV_F64) && (flags & AV_CPU_FLAG_RVB_ADDR)) {
>          c->lpc_apply_welch_window = ff_lpc_apply_welch_window_rvv;
> +
> +        if (ff_get_rv_vlenb() >= c->max_order)
> +            c->lpc_compute_autocorr = ff_lpc_compute_autocorr_rvv;
> +    }
>  #endif
>  }
> diff --git a/libavcodec/riscv/lpc_rvv.S b/libavcodec/riscv/lpc_rvv.S
> index f81a2392c1..654156bf12 100644
> --- a/libavcodec/riscv/lpc_rvv.S
> +++ b/libavcodec/riscv/lpc_rvv.S
> @@ -85,4 +85,33 @@ func ff_lpc_apply_welch_window_rvv, zve64d
> 
>          ret
>  endfunc
> +
> +func ff_lpc_compute_autocorr_rvv, zve64d
> +        li        t0, 1
> +        vsetvli   t1, a2, e64, m8, ta, ma

t1 is unused and should be zero. This is leftover from incomplete attempt to 
unroll.

> +        fcvt.d.l  ft0, t0
> +        vle64.v   v0, (a0)
> +        sh3add    a0, a2, a0   # data += lag
> +        vfmv.v.f  v16, ft0
> +        bge       a2, a1, 2f
> +1:
> +        vfmv.f.s  ft0, v0
> +        fld       ft1, (a0)    # ft1 = data[lag + i]
> +        vfmacc.vf v16, ft0, v0 # v16[j] += data[i] * data[i + j]
> +        addi      a1, a1, -1
> +        vfslide1down.vf v0, v0, ft1
> +        addi      a0, a0, 8
> +        bgt       a1, a2, 1b   # while (len > lag);
> +2:
> +        vfmv.f.s  ft0, v0
> +        vsetvli   zero, a1, e64, m8, tu, ma
> +        vfmacc.vf v16, ft0, v0
> +        addi      a1, a1, -1
> +        vslide1down.vx v0, v0, zero
> +        bnez      a1, 2b       # while (len > 0);
> +
> +        vsetvli   zero, a2, e64, m8, ta, ma
> +        vse64.v   v16, (a3)
> +        ret
> +endfunc
>  #endif