comparison postprocess_altivec_template.c @ 133:51571e34b760 libpostproc

Move array specifiers outside DECLARE_ALIGNED() invocations

132:a65cfe0fe4b2

    this code makes no assumption on src or stride.
    One could remove the recomputation of the perm
    vector by assuming (stride % 16) == 0, unfortunately
    this is not always true.
    */
    DECLARE_ALIGNED(16, short, data[8]) =
                    {
                        ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1,
                        data[0] * 2 + 1,
                        c->QP * 2,
                        c->QP * 4

    */
    uint8_t *src2 = src;
    const vector signed int zero = vec_splat_s32(0);
    const int properStride = (stride % 16);
    const int srcAlign = ((unsigned long)src2 % 16);
    DECLARE_ALIGNED(16, short, qp[8]) = {c->QP};
    vector signed short vqp = vec_ld(0, qp);
    vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
    vector unsigned char vbA0, av_uninit(vbA1), av_uninit(vbA2), av_uninit(vbA3), av_uninit(vbA4), av_uninit(vbA5), av_uninit(vbA6), av_uninit(vbA7), av_uninit(vbA8), vbA9;
    vector unsigned char vbB0, av_uninit(vbB1), av_uninit(vbB2), av_uninit(vbB3), av_uninit(vbB4), av_uninit(vbB5), av_uninit(vbB6), av_uninit(vbB7), av_uninit(vbB8), vbB9;
    vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;

    can be removed by assuming proper alignment of
    src & stride :-(
    */
    uint8_t *src2 = src + stride*3;
    const vector signed int zero = vec_splat_s32(0);
    DECLARE_ALIGNED(16, short, qp[8]) = {8*c->QP};
    vector signed short vqp = vec_splat(
                                (vector signed short)vec_ld(0, qp), 0);

#define LOAD_LINE(i)                                                    \
    const vector unsigned char perm##i =                                \

    this is not always true. Quite a lot of load/stores
    can be removed by assuming proper alignment of
    src & stride :-(
    */
    uint8_t *srcCopy = src;
    DECLARE_ALIGNED(16, uint8_t, dt[16]);
    const vector signed int zero = vec_splat_s32(0);
    vector unsigned char v_dt;
    dt[0] = deringThreshold;
    v_dt = vec_splat(vec_ld(0, dt), 0);

        return;

    v_avg = vec_avg(v_min, v_max);
    }

    DECLARE_ALIGNED(16, signed int, S[8]);
    {
    const vector unsigned short mask1 = (vector unsigned short)
                                        {0x0001, 0x0002, 0x0004, 0x0008,
                                         0x0010, 0x0020, 0x0040, 0x0080};
    const vector unsigned short mask2 = (vector unsigned short)

    }

    /* I'm not sure the following is actually faster
       than straight, unvectorized C code :-( */

    DECLARE_ALIGNED(16, int, tQP2[4]);
    tQP2[0]= c->QP/2 + 1;
    vector signed int vQP2 = vec_ld(0, tQP2);
    vQP2 = vec_splat(vQP2, 0);
    const vector signed int vsint32_8 = vec_splat_s32(8);
    const vector unsigned int vuint32_4 = vec_splat_u32(4);

133:51571e34b760

    this code makes no assumption on src or stride.
    One could remove the recomputation of the perm
    vector by assuming (stride % 16) == 0, unfortunately
    this is not always true.
    */
    DECLARE_ALIGNED(16, short, data)[8] =
                    {
                        ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1,
                        data[0] * 2 + 1,
                        c->QP * 2,
                        c->QP * 4

    */
    uint8_t *src2 = src;
    const vector signed int zero = vec_splat_s32(0);
    const int properStride = (stride % 16);
    const int srcAlign = ((unsigned long)src2 % 16);
    DECLARE_ALIGNED(16, short, qp)[8] = {c->QP};
    vector signed short vqp = vec_ld(0, qp);
    vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
    vector unsigned char vbA0, av_uninit(vbA1), av_uninit(vbA2), av_uninit(vbA3), av_uninit(vbA4), av_uninit(vbA5), av_uninit(vbA6), av_uninit(vbA7), av_uninit(vbA8), vbA9;
    vector unsigned char vbB0, av_uninit(vbB1), av_uninit(vbB2), av_uninit(vbB3), av_uninit(vbB4), av_uninit(vbB5), av_uninit(vbB6), av_uninit(vbB7), av_uninit(vbB8), vbB9;
    vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;

    can be removed by assuming proper alignment of
    src & stride :-(
    */
    uint8_t *src2 = src + stride*3;
    const vector signed int zero = vec_splat_s32(0);
    DECLARE_ALIGNED(16, short, qp)[8] = {8*c->QP};
    vector signed short vqp = vec_splat(
                                (vector signed short)vec_ld(0, qp), 0);

#define LOAD_LINE(i)                                                    \
    const vector unsigned char perm##i =                                \

    this is not always true. Quite a lot of load/stores
    can be removed by assuming proper alignment of
    src & stride :-(
    */
    uint8_t *srcCopy = src;
    DECLARE_ALIGNED(16, uint8_t, dt)[16];
    const vector signed int zero = vec_splat_s32(0);
    vector unsigned char v_dt;
    dt[0] = deringThreshold;
    v_dt = vec_splat(vec_ld(0, dt), 0);

        return;

    v_avg = vec_avg(v_min, v_max);
    }

    DECLARE_ALIGNED(16, signed int, S)[8];
    {
    const vector unsigned short mask1 = (vector unsigned short)
                                        {0x0001, 0x0002, 0x0004, 0x0008,
                                         0x0010, 0x0020, 0x0040, 0x0080};
    const vector unsigned short mask2 = (vector unsigned short)

    }

    /* I'm not sure the following is actually faster
       than straight, unvectorized C code :-( */

    DECLARE_ALIGNED(16, int, tQP2)[4];
    tQP2[0]= c->QP/2 + 1;
    vector signed int vQP2 = vec_ld(0, tQP2);
    vQP2 = vec_splat(vQP2, 0);
    const vector signed int vsint32_8 = vec_splat_s32(8);
    const vector unsigned int vuint32_4 = vec_splat_u32(4);