dsputil_mmx_avg.h

ffmpeg源码分析

/* GL: this function does incorrect rounding if overflow */
static void DEF(put_no_rnd_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    MOVQ_BONE(mm6);
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "1:                             \n\t"
        "movq (%1), %%mm0               \n\t"
        "movq (%1, %3), %%mm2           \n\t"
        "movq 1(%1), %%mm1              \n\t"
        "movq 1(%1, %3), %%mm3          \n\t"
        "add %%"REG_a", %1              \n\t"
        "psubusb %%mm6, %%mm0           \n\t"
        "psubusb %%mm6, %%mm2           \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm3, %%mm2             \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "movq (%1), %%mm0               \n\t"
        "movq 1(%1), %%mm1              \n\t"
        "movq (%1, %3), %%mm2           \n\t"
        "movq 1(%1, %3), %%mm3          \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        "psubusb %%mm6, %%mm0           \n\t"
        "psubusb %%mm6, %%mm2           \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm3, %%mm2             \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D"(block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

static void DEF(put_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "movq (%1), %%mm0               \n\t"
        "sub %3, %2                     \n\t"
        "1:                             \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm2    \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm2, %%mm1             \n\t"
        "movq %%mm0, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" %%mm1, %%mm2             \n\t"
        PAVGB" %%mm0, %%mm1             \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D" (block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

/* GL: this function does incorrect rounding if overflow */
static void DEF(put_no_rnd_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    MOVQ_BONE(mm6);
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "movq (%1), %%mm0               \n\t"
        "sub %3, %2                     \n\t"
        "1:                             \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm2    \n\t"
        "add %%"REG_a", %1              \n\t"
        "psubusb %%mm6, %%mm1           \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm2, %%mm1             \n\t"
        "movq %%mm0, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        "psubusb %%mm6, %%mm1           \n\t"
        PAVGB" %%mm1, %%mm2             \n\t"
        PAVGB" %%mm0, %%mm1             \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D" (block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

static void DEF(avg_pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "1:                             \n\t"
        "movq (%2), %%mm0               \n\t"
        "movq (%2, %3), %%mm1           \n\t"
        PAVGB" (%1), %%mm0              \n\t"
        PAVGB" (%1, %3), %%mm1          \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm1, (%2, %3)           \n\t"
        "add %%"REG_a", %1              \n\t"
        "add %%"REG_a", %2              \n\t"
        "movq (%2), %%mm0               \n\t"
        "movq (%2, %3), %%mm1           \n\t"
        PAVGB" (%1), %%mm0              \n\t"
        PAVGB" (%1, %3), %%mm1          \n\t"
        "add %%"REG_a", %1              \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm1, (%2, %3)           \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D"(block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

static void DEF(avg_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "1:                             \n\t"
        "movq (%1), %%mm0               \n\t"
        "movq (%1, %3), %%mm2           \n\t"
        PAVGB" 1(%1), %%mm0             \n\t"
        PAVGB" 1(%1, %3), %%mm2         \n\t"
        PAVGB" (%2), %%mm0              \n\t"
        PAVGB" (%2, %3), %%mm2          \n\t"
        "add %%"REG_a", %1              \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "movq (%1), %%mm0               \n\t"
        "movq (%1, %3), %%mm2           \n\t"
        PAVGB" 1(%1), %%mm0             \n\t"
        PAVGB" 1(%1, %3), %%mm2         \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" (%2), %%mm0              \n\t"
        PAVGB" (%2, %3), %%mm2          \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D"(block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

static void DEF(avg_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "movq (%1), %%mm0               \n\t"
        "sub %3, %2                     \n\t"
        "1:                             \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm2    \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm2, %%mm1             \n\t"
        "movq (%2, %3), %%mm3           \n\t"
        "movq (%2, %%"REG_a"), %%mm4    \n\t"
        PAVGB" %%mm3, %%mm0             \n\t"
        PAVGB" %%mm4, %%mm1             \n\t"
        "movq %%mm0, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        PAVGB" %%mm1, %%mm2             \n\t"
        PAVGB" %%mm0, %%mm1             \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        "movq (%2, %3), %%mm3           \n\t"
        "movq (%2, %%"REG_a"), %%mm4    \n\t"
        PAVGB" %%mm3, %%mm2             \n\t"
        PAVGB" %%mm4, %%mm1             \n\t"
        "movq %%mm2, (%2, %3)           \n\t"
        "movq %%mm1, (%2, %%"REG_a")    \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D"(block)
        :"r" ((long)line_size)
        :"%"REG_a, "memory");
}

// Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter
static void DEF(avg_pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
    MOVQ_BONE(mm6);
    __asm __volatile(
        "lea (%3, %3), %%"REG_a"        \n\t"
        "movq (%1), %%mm0               \n\t"
        PAVGB" 1(%1), %%mm0             \n\t"
        ".balign 8                      \n\t"
        "1:                             \n\t"
        "movq (%1, %%"REG_a"), %%mm2    \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "psubusb %%mm6, %%mm2           \n\t"
        PAVGB" 1(%1, %3), %%mm1         \n\t"
        PAVGB" 1(%1, %%"REG_a"), %%mm2  \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm2, %%mm1             \n\t"
        PAVGB" (%2), %%mm0              \n\t"
        PAVGB" (%2, %3), %%mm1          \n\t"
        "movq %%mm0, (%2)               \n\t"
        "movq %%mm1, (%2, %3)           \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        PAVGB" 1(%1, %3), %%mm1         \n\t"
        PAVGB" 1(%1, %%"REG_a"), %%mm0  \n\t"
        "add %%"REG_a", %2              \n\t"
        "add %%"REG_a", %1              \n\t"
        PAVGB" %%mm1, %%mm2             \n\t"
        PAVGB" %%mm0, %%mm1             \n\t"
        PAVGB" (%2), %%mm2              \n\t"
        PAVGB" (%2, %3), %%mm1          \n\t"
        "movq %%mm2, (%2)               \n\t"
        "movq %%mm1, (%2, %3)           \n\t"
        "add %%"REG_a", %2              \n\t"
        "subl $4, %0                    \n\t"
        "jnz 1b                         \n\t"
        :"+g"(h), "+S"(pixels), "+D"(block)
        :"r" ((long)line_size)
        :"%"REG_a,  "memory");
}

//FIXME the following could be optimized too ...
static void DEF(put_no_rnd_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(put_no_rnd_pixels8_x2)(block  , pixels  , line_size, h);
    DEF(put_no_rnd_pixels8_x2)(block+8, pixels+8, line_size, h);
}
static void DEF(put_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(put_pixels8_y2)(block  , pixels  , line_size, h);
    DEF(put_pixels8_y2)(block+8, pixels+8, line_size, h);
}
static void DEF(put_no_rnd_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(put_no_rnd_pixels8_y2)(block  , pixels  , line_size, h);
    DEF(put_no_rnd_pixels8_y2)(block+8, pixels+8, line_size, h);
}
static void DEF(avg_pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(avg_pixels8)(block  , pixels  , line_size, h);
    DEF(avg_pixels8)(block+8, pixels+8, line_size, h);
}
static void DEF(avg_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(avg_pixels8_x2)(block  , pixels  , line_size, h);
    DEF(avg_pixels8_x2)(block+8, pixels+8, line_size, h);
}
static void DEF(avg_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(avg_pixels8_y2)(block  , pixels  , line_size, h);
    DEF(avg_pixels8_y2)(block+8, pixels+8, line_size, h);
}
static void DEF(avg_pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    DEF(avg_pixels8_xy2)(block  , pixels  , line_size, h);
    DEF(avg_pixels8_xy2)(block+8, pixels+8, line_size, h);
}