comparison i386/dsputilenc_mmx.c @ 7220:a94b2cf78a2e libavcodec

Make LOAD4/STORE4 macros more generic. Patch by Victor Pollex victor pollex web de Original thread: [PATCH] mmx implementation of vc-1 inverse transformations Date: 06/21/2008 03:37 PM

7219:f72ef5b28253

    "movdqa (%1), %%xmm2              \n\t"\
    MMABS_SUM(%%xmm2, %%xmm7, %%xmm1)\
    "paddusw %%xmm1, %%xmm0           \n\t"
#endif

#define LOAD4(o, a, b, c, d)\
    "movq "#o"(%1),    "#a"           \n\t"\
    "movq "#o"+8(%1),  "#b"           \n\t"\
    "movq "#o"+16(%1), "#c"           \n\t"\
    "movq "#o"+24(%1), "#d"           \n\t"\

#define STORE4(o, a, b, c, d)\
    "movq "#a", "#o"(%1)              \n\t"\
    "movq "#b", "#o"+8(%1)            \n\t"\
    "movq "#c", "#o"+16(%1)           \n\t"\
    "movq "#d", "#o"+24(%1)           \n\t"\

/* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to
 * about 100k on extreme inputs. But that's very unlikely to occur in natural video,
 * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */
#define HSUM_MMX(a, t, dst)\
    "movq "#a", "#t"                  \n\t"\

        HADAMARD48\
\
        "movq %%mm7, 96(%1)             \n\t"\
\
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
        STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)\
\
        "movq 96(%1), %%mm7             \n\t"\
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
        STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)\
\
        : "=r" (sum)\
        : "r"(temp)\
    );\
\

        HADAMARD48\
\
        "movq %%mm7, 96(%1)             \n\t"\
\
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
        STORE4(32, %%mm0, %%mm3, %%mm7, %%mm2)\
\
        "movq 96(%1), %%mm7             \n\t"\
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
        "movq %%mm7, %%mm5              \n\t"/*FIXME remove*/\
        "movq %%mm6, %%mm7              \n\t"\
        "movq %%mm0, %%mm6              \n\t"\
\
        LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)\
\
        HADAMARD48\
        "movq %%mm7, 64(%1)             \n\t"\
        MMABS(%%mm0, %%mm7)\
        MMABS(%%mm1, %%mm7)\

        "movq 64(%1), %%mm2             \n\t"\
        MMABS_SUM(%%mm2, %%mm7, %%mm1)\
        "paddusw %%mm1, %%mm0           \n\t"\
        "movq %%mm0, 64(%1)             \n\t"\
\
        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)\
        LOAD4(32, %%mm4, %%mm5, %%mm6, %%mm7)\
\
        HADAMARD48\
        "movq %%mm7, (%1)               \n\t"\
        MMABS(%%mm0, %%mm7)\
        MMABS(%%mm1, %%mm7)\

7220:a94b2cf78a2e

    "movdqa (%1), %%xmm2              \n\t"\
    MMABS_SUM(%%xmm2, %%xmm7, %%xmm1)\
    "paddusw %%xmm1, %%xmm0           \n\t"
#endif

/* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to
 * about 100k on extreme inputs. But that's very unlikely to occur in natural video,
 * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */
#define HSUM_MMX(a, t, dst)\
    "movq "#a", "#t"                  \n\t"\

        HADAMARD48\
\
        "movq %%mm7, 96(%1)             \n\t"\
\
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
        STORE4(8,  0(%1), %%mm0, %%mm3, %%mm7, %%mm2)\
\
        "movq 96(%1), %%mm7             \n\t"\
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
        STORE4(8, 64(%1), %%mm4, %%mm7, %%mm0, %%mm6)\
\
        : "=r" (sum)\
        : "r"(temp)\
    );\
\

        HADAMARD48\
\
        "movq %%mm7, 96(%1)             \n\t"\
\
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
        STORE4(8, 32(%1), %%mm0, %%mm3, %%mm7, %%mm2)\
\
        "movq 96(%1), %%mm7             \n\t"\
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
        "movq %%mm7, %%mm5              \n\t"/*FIXME remove*/\
        "movq %%mm6, %%mm7              \n\t"\
        "movq %%mm0, %%mm6              \n\t"\
\
        LOAD4(8, 64(%1), %%mm0, %%mm1, %%mm2, %%mm3)\
\
        HADAMARD48\
        "movq %%mm7, 64(%1)             \n\t"\
        MMABS(%%mm0, %%mm7)\
        MMABS(%%mm1, %%mm7)\

        "movq 64(%1), %%mm2             \n\t"\
        MMABS_SUM(%%mm2, %%mm7, %%mm1)\
        "paddusw %%mm1, %%mm0           \n\t"\
        "movq %%mm0, 64(%1)             \n\t"\
\
        LOAD4(8,  0(%1), %%mm0, %%mm1, %%mm2, %%mm3)\
        LOAD4(8, 32(%1), %%mm4, %%mm5, %%mm6, %%mm7)\
\
        HADAMARD48\
        "movq %%mm7, (%1)               \n\t"\
        MMABS(%%mm0, %%mm7)\
        MMABS(%%mm1, %%mm7)\