comparison libmpcodecs/vf_fspp.c @ 37086:f19959d0bbc9

Fix libmpcodecs inline asm on ICL. Author: Matt Oliver <[e-mail]> Signed-off-by: Michael Niedermayer <[e-mail]>

37085:aadeccf8159b

        "jnz 1b                \n\t"
        "5:                      \n\t"

        : "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps)
        : "d"(thr_adr)
        : "%"REG_a
        );
}

#endif // HAVE_MMX

        "dec %%"REG_c"                   \n\t"
        "jnz 1b                  \n\t"

        : "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps)
        : "a"(output_stride*sizeof(short))
        : "%"REG_d
        );
}

#endif // HAVE_MMX

        "psubw %%mm4, %%mm6            \n\t"

        "movd (%%"REG_S",%%"REG_a",2), %%mm3    \n\t" //5
        "paddw %%mm4, %%mm1            \n\t"

        "movq %%mm5, 0*8+%3            \n\t" //t7
        "punpcklbw %%mm7, %%mm3        \n\t"

        "movq %%mm6, 1*8+%3            \n\t" //t6
        "movq %%mm2, %%mm4             \n\t"

        "movd (%%"REG_S"), %%mm5           \n\t" //3
        "paddw %%mm3, %%mm2            \n\t"

        "punpckhwd %%mm6, %%mm2        \n\t"

        "psubw %%mm1, %%mm5            \n\t" //d1
        "movq %%mm0, %%mm6             \n\t"

        "movq 1*8+%3, %%mm1            \n\t"
        "punpcklwd %%mm5, %%mm0        \n\t"

        "punpckhwd %%mm5, %%mm6        \n\t"
        "movq %%mm0, %%mm5             \n\t"

        "paddw %%mm1, %%mm4            \n\t"

        "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
        "psllw $2, %%mm3              \n\t" //t10

        "movq 0*8+%3, %%mm2           \n\t"
        "psllw $2, %%mm4              \n\t" //t11

        "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
        "paddw %%mm2, %%mm1            \n\t"

        "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
        "add $"DCTSIZE_S"*2*4, %%"REG_D"      \n\t" //4 rows
        "dec %%"REG_c"                   \n\t"
        "jnz 6b                  \n\t"

        : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps)
        : "a"(line_size)
        : "%"REG_d);
}

#endif // HAVE_MMX

37086:f19959d0bbc9

        "jnz 1b                \n\t"
        "5:                      \n\t"

        : "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps)
        : "d"(thr_adr)
          NAMED_CONSTRAINTS_ADD(ff_MM_FIX_0_707106781,MM_2,MM_FIX_1_414213562_A,MM_FIX_1_414213562,MM_FIX_0_382683433,
          ff_MM_FIX_0_541196100,MM_FIX_1_306562965,MM_FIX_0_847759065)
          NAMED_CONSTRAINTS_ADD(MM_FIX_0_566454497,MM_FIX_0_198912367,MM_FIX_2_613125930,MM_FIX_1_847759065,
          MM_FIX_1_082392200,ff_MM_FIX_0_541196100,MM_FIX_1_306562965)
        : "%"REG_a
        );
}

#endif // HAVE_MMX

        "dec %%"REG_c"                   \n\t"
        "jnz 1b                  \n\t"

        : "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps)
        : "a"(output_stride*sizeof(short))
        NAMED_CONSTRAINTS_ADD(MM_FIX_1_414213562_A,MM_FIX_2_613125930,MM_FIX_1_847759065,MM_FIX_1_082392200,
        MM_FIX_1_414213562,MM_DESCALE_RND)
        : "%"REG_d
        );
}

#endif // HAVE_MMX

        "psubw %%mm4, %%mm6            \n\t"

        "movd (%%"REG_S",%%"REG_a",2), %%mm3    \n\t" //5
        "paddw %%mm4, %%mm1            \n\t"

        "movq %%mm5, %3                \n\t" //t7
        "punpcklbw %%mm7, %%mm3        \n\t"

        "movq %%mm6, %4                \n\t" //t6
        "movq %%mm2, %%mm4             \n\t"

        "movd (%%"REG_S"), %%mm5           \n\t" //3
        "paddw %%mm3, %%mm2            \n\t"

        "punpckhwd %%mm6, %%mm2        \n\t"

        "psubw %%mm1, %%mm5            \n\t" //d1
        "movq %%mm0, %%mm6             \n\t"

        "movq %4, %%mm1                \n\t"
        "punpcklwd %%mm5, %%mm0        \n\t"

        "punpckhwd %%mm5, %%mm6        \n\t"
        "movq %%mm0, %%mm5             \n\t"

        "paddw %%mm1, %%mm4            \n\t"

        "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
        "psllw $2, %%mm3              \n\t" //t10

        "movq %3, %%mm2               \n\t"
        "psllw $2, %%mm4              \n\t" //t11

        "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
        "paddw %%mm2, %%mm1            \n\t"

        "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
        "add $"DCTSIZE_S"*2*4, %%"REG_D"      \n\t" //4 rows
        "dec %%"REG_c"                   \n\t"
        "jnz 6b                  \n\t"

        : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps), "=o"(temps[1])
        : "a"(line_size)
        NAMED_CONSTRAINTS_ADD(ff_MM_FIX_0_707106781,ff_MM_FIX_0_541196100,MM_FIX_0_382683433,MM_FIX_1_306562965)
        : "%"REG_d);
}

#endif // HAVE_MMX