[FFmpeg-devel] [PATCH v1 1/3] swscale/la: Optimize hscale functions with lasx.

[Shiyou Yin]

> +
> +void planar_rgb_to_uv_lasx(uint8_t *_dstU, uint8_t *_dstV, const uint8_t *src[4],
> +                           int width, int32_t *rgb2yuv)
> +{
> +    int i;
> +    uint16_t *dstU   = (uint16_t *)_dstU;
> +    uint16_t *dstV   = (uint16_t *)_dstV;
> +    int set          = 0x4001 << (RGB2YUV_SHIFT - 7);
> +    int len          = width - 15;
> +    int32_t tem_ru   = rgb2yuv[RU_IDX], tem_gu = rgb2yuv[GU_IDX];
> +    int32_t tem_bu = rgb2yuv[BU_IDX], tem_rv   = rgb2yuv[RV_IDX];
> +    int32_t tem_gv = rgb2yuv[GV_IDX], tem_bv = rgb2yuv[BV_IDX];
> +    int shift        = RGB2YUV_SHIFT - 6;
> +    const uint8_t *src0 = src[0], *src1 = src[1], *src2 = src[2];
> +    __m256i ru, gu, bu, rv, gv, bv;
> +    __m256i mask = {0x0D0C090805040100, 0x1D1C191815141110,
> +                    0x0D0C090805040100, 0x1D1C191815141110};
> +    __m256i temp = __lasx_xvreplgr2vr_w(set);
> +    __m256i sra  = __lasx_xvreplgr2vr_w(shift);
> +
> +    ru = __lasx_xvreplgr2vr_w(tem_ru);
> +    gu = __lasx_xvreplgr2vr_w(tem_gu);
> +    bu = __lasx_xvreplgr2vr_w(tem_bu);
> +    rv = __lasx_xvreplgr2vr_w(tem_rv);
> +    gv = __lasx_xvreplgr2vr_w(tem_gv);
> +    bv = __lasx_xvreplgr2vr_w(tem_bv);
> +    for (i = 0; i < len; i += 16) {
> +        __m256i _g, _b, _r;
> +        __m256i g_l, g_h, b_l, b_h, r_l, r_h;
> +        __m256i v_l, v_h, u_l, u_h, u_lh, v_lh;
> +
> +        DUP2_ARG2(__lasx_xvld, src0 + i, 0, src1 + i, 0, _g, _b);
> +        _r   = __lasx_xvld(src2 + i, 0);

In this case, it’s more readable not to use DUP2.
The following extensions of r,g,b are the same.

> +        g_l = __lasx_vext2xv_wu_bu(_g);
> +        DUP2_ARG1(__lasx_vext2xv_wu_bu, _b, _r, b_l, r_l);
> +        _g   = __lasx_xvpermi_d(_g, 0x01);
> +        _b   = __lasx_xvpermi_d(_b, 0x01);
> +        _r   = __lasx_xvpermi_d(_r, 0x01);
> +        g_h = __lasx_vext2xv_wu_bu(_g);
> +        DUP2_ARG1(__lasx_vext2xv_wu_bu, _b, _r, b_h, r_h);
> +        u_l  = __lasx_xvmadd_w(temp, ru, r_l);
> +        u_h  = __lasx_xvmadd_w(temp, ru, r_h);
> +        v_l  = __lasx_xvmadd_w(temp, rv, r_l);
> +        v_h  = __lasx_xvmadd_w(temp, rv, r_h);
> +        u_l  = __lasx_xvmadd_w(u_l, gu, g_l);
> +        u_l  = __lasx_xvmadd_w(u_l, bu, b_l);
> +        u_h  = __lasx_xvmadd_w(u_h, gu, g_h);
> +        u_h  = __lasx_xvmadd_w(u_h, bu, b_h);
> +        v_l  = __lasx_xvmadd_w(v_l, gv, g_l);
> +        v_l  = __lasx_xvmadd_w(v_l, bv, b_l);
> +        v_h  = __lasx_xvmadd_w(v_h, gv, g_h);
> +        v_h  = __lasx_xvmadd_w(v_h, bv, b_h);
> +        u_l  = __lasx_xvsra_w(u_l, sra);
> +        u_h  = __lasx_xvsra_w(u_h, sra);
> +        v_l  = __lasx_xvsra_w(v_l, sra);
> +        v_h  = __lasx_xvsra_w(v_h, sra);
> +        DUP2_ARG3(__lasx_xvshuf_b, u_h, u_l, mask, v_h, v_l, mask, u_lh, v_lh);
> +        u_lh = __lasx_xvpermi_d(u_lh, 0xD8);
> +        v_lh = __lasx_xvpermi_d(v_lh, 0xD8);
> +        __lasx_xvst(u_lh, (dstU + i), 0);
> +        __lasx_xvst(v_lh, (dstV + i), 0);
> +    }
> +    if (width - i >= 8) {
> +        __m256i _g, _b, _r;
> +        __m256i g_l, b_l, r_l;
> +        __m256i v_l, u_l, u, v;
> +
> +        _g  = __lasx_xvldrepl_d((src0 + i), 0);
> +        _b  = __lasx_xvldrepl_d((src1 + i), 0);
> +        _r  = __lasx_xvldrepl_d((src2 + i), 0);
> +        g_l = __lasx_vext2xv_wu_bu(_g);
> +        DUP2_ARG1(__lasx_vext2xv_wu_bu, _b, _r, b_l, r_l);
> +        u_l = __lasx_xvmadd_w(temp, ru, r_l);
> +        v_l = __lasx_xvmadd_w(temp, rv, r_l);
> +        u_l = __lasx_xvmadd_w(u_l, gu, g_l);
> +        u_l = __lasx_xvmadd_w(u_l, bu, b_l);
> +        v_l = __lasx_xvmadd_w(v_l, gv, g_l);
> +        v_l = __lasx_xvmadd_w(v_l, bv, b_l);
> +        u_l = __lasx_xvsra_w(u_l, sra);
> +        v_l = __lasx_xvsra_w(v_l, sra);
> +        DUP2_ARG3(__lasx_xvshuf_b, u_l, u_l, mask, v_l, v_l, mask, u, v);
> +        __lasx_xvstelm_d(u, (dstU + i), 0, 0);
> +        __lasx_xvstelm_d(u, (dstU + i), 8, 2);
> +        __lasx_xvstelm_d(v, (dstV + i), 0, 0);
> +        __lasx_xvstelm_d(v, (dstV + i), 8, 2);
> +        i += 8;
> +    }
> +    for (; i < width; i++) {
> +        int g = src[0][i];
> +        int b = src[1][i];
> +        int r = src[2][i];
> +
> +        dstU[i] = (tem_ru * r + tem_gu * g + tem_bu * b + set) >> shift;
> +        dstV[i] = (tem_rv * r + tem_gv * g + tem_bv * b + set) >> shift;
> +    }

Suggest to use following condition controls instead.
While(i=width; i>=16; i-=16) {}
if(i>=8) {}
while(I--){}