Mercurial > libavcodec.hg
diff i386/dsputil_mmx_avg.h @ 2979:bfabfdf9ce55 libavcodec
COSMETICS: tabs --> spaces, some prettyprinting
| author | diego |
|---|---|
| date | Thu, 22 Dec 2005 01:10:11 +0000 |
| parents | ef2149182f1c |
| children | 0b546eab515d |
line wrap: on
line diff
--- a/i386/dsputil_mmx_avg.h Wed Dec 21 17:50:40 2005 +0000 +++ b/i386/dsputil_mmx_avg.h Thu Dec 22 01:10:11 2005 +0000 @@ -28,522 +28,522 @@ static void DEF(put_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), %%mm1 \n\t" - PAVGB" 1(%1), %%mm0 \n\t" - PAVGB" 1(%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 (%1), %%mm0 \n\t" - "movq (%1, %3), %%mm1 \n\t" - PAVGB" 1(%1), %%mm0 \n\t" - PAVGB" 1(%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"); + "lea (%3, %3), %%"REG_a" \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%1, %3), %%mm1 \n\t" + PAVGB" 1(%1), %%mm0 \n\t" + PAVGB" 1(%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 (%1), %%mm0 \n\t" + "movq (%1, %3), %%mm1 \n\t" + PAVGB" 1(%1), %%mm0 \n\t" + PAVGB" 1(%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(put_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movd (%1), %%mm0 \n\t" - "movd (%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $4, %2 \n\t" - PAVGB" %%mm1, %%mm0 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movd (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movd (%1), %%mm1 \n\t" - "movd (%2), %%mm2 \n\t" - "movd 4(%2), %%mm3 \n\t" - "add %4, %1 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movd %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "movd (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movd (%1), %%mm1 \n\t" - "movd 8(%2), %%mm2 \n\t" - "movd 12(%2), %%mm3 \n\t" - "add %4, %1 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movd %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "add $16, %2 \n\t" - "subl $4, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movd (%1), %%mm0 \n\t" + "movd (%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $4, %2 \n\t" + PAVGB" %%mm1, %%mm0 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movd (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movd (%1), %%mm1 \n\t" + "movd (%2), %%mm2 \n\t" + "movd 4(%2), %%mm3 \n\t" + "add %4, %1 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movd %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "movd (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movd (%1), %%mm1 \n\t" + "movd 8(%2), %%mm2 \n\t" + "movd 12(%2), %%mm3 \n\t" + "add %4, %1 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movd %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "add $16, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); } static void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq (%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $8, %2 \n\t" - PAVGB" %%mm1, %%mm0 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" 16(%2), %%mm0 \n\t" - PAVGB" 24(%2), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $4, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $8, %2 \n\t" + PAVGB" %%mm1, %%mm0 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" 16(%2), %%mm0 \n\t" + PAVGB" 24(%2), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } static void DEF(put_no_rnd_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "pcmpeqb %%mm6, %%mm6 \n\t" - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq (%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $8, %2 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - PAVGB" %%mm1, %%mm0 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - "movq (%2), %%mm2 \n\t" - "movq 8(%2), %%mm3 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "pxor %%mm6, %%mm2 \n\t" - "pxor %%mm6, %%mm3 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - "movq 16(%2), %%mm2 \n\t" - "movq 24(%2), %%mm3 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "pxor %%mm6, %%mm2 \n\t" - "pxor %%mm6, %%mm3 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $4, %0 \n\t" - "jnz 1b \n\t" + "pcmpeqb %%mm6, %%mm6 \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $8, %2 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + PAVGB" %%mm1, %%mm0 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + "movq (%2), %%mm2 \n\t" + "movq 8(%2), %%mm3 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "pxor %%mm6, %%mm2 \n\t" + "pxor %%mm6, %%mm3 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + "movq 16(%2), %%mm2 \n\t" + "movq 24(%2), %%mm3 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "pxor %%mm6, %%mm2 \n\t" + "pxor %%mm6, %%mm3 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } static void DEF(avg_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movd (%1), %%mm0 \n\t" - "movd (%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $4, %2 \n\t" - PAVGB" %%mm1, %%mm0 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movd (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movd (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 4(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - PAVGB" (%3), %%mm1 \n\t" - "movd %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "movd (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movd (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" 8(%2), %%mm0 \n\t" - PAVGB" 12(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movd %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - PAVGB" (%3), %%mm1 \n\t" - "movd %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "add $16, %2 \n\t" - "subl $4, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movd (%1), %%mm0 \n\t" + "movd (%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $4, %2 \n\t" + PAVGB" %%mm1, %%mm0 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movd (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movd (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 4(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + PAVGB" (%3), %%mm1 \n\t" + "movd %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "movd (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movd (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" 8(%2), %%mm0 \n\t" + PAVGB" 12(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movd %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + PAVGB" (%3), %%mm1 \n\t" + "movd %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "add $16, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); } static void DEF(avg_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq (%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $8, %2 \n\t" - PAVGB" %%mm1, %%mm0 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - PAVGB" (%3), %%mm1 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "add %4, %1 \n\t" - "movq (%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" 16(%2), %%mm0 \n\t" - PAVGB" 24(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - "movq %%mm0, (%3) \n\t" - "add %5, %3 \n\t" - PAVGB" (%3), %%mm1 \n\t" - "movq %%mm1, (%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $4, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $8, %2 \n\t" + PAVGB" %%mm1, %%mm0 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + PAVGB" (%3), %%mm1 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "add %4, %1 \n\t" + "movq (%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" 16(%2), %%mm0 \n\t" + PAVGB" 24(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + "movq %%mm0, (%3) \n\t" + "add %5, %3 \n\t" + PAVGB" (%3), %%mm1 \n\t" + "movq %%mm1, (%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } static void DEF(put_pixels16_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), %%mm1 \n\t" - "movq 8(%1), %%mm2 \n\t" - "movq 8(%1, %3), %%mm3 \n\t" - PAVGB" 1(%1), %%mm0 \n\t" - PAVGB" 1(%1, %3), %%mm1 \n\t" - PAVGB" 9(%1), %%mm2 \n\t" - PAVGB" 9(%1, %3), %%mm3 \n\t" - "movq %%mm0, (%2) \n\t" - "movq %%mm1, (%2, %3) \n\t" - "movq %%mm2, 8(%2) \n\t" - "movq %%mm3, 8(%2, %3) \n\t" - "add %%"REG_a", %1 \n\t" - "add %%"REG_a", %2 \n\t" - "movq (%1), %%mm0 \n\t" - "movq (%1, %3), %%mm1 \n\t" - "movq 8(%1), %%mm2 \n\t" - "movq 8(%1, %3), %%mm3 \n\t" - PAVGB" 1(%1), %%mm0 \n\t" - PAVGB" 1(%1, %3), %%mm1 \n\t" - PAVGB" 9(%1), %%mm2 \n\t" - PAVGB" 9(%1, %3), %%mm3 \n\t" - "add %%"REG_a", %1 \n\t" - "movq %%mm0, (%2) \n\t" - "movq %%mm1, (%2, %3) \n\t" - "movq %%mm2, 8(%2) \n\t" - "movq %%mm3, 8(%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"); + "lea (%3, %3), %%"REG_a" \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%1, %3), %%mm1 \n\t" + "movq 8(%1), %%mm2 \n\t" + "movq 8(%1, %3), %%mm3 \n\t" + PAVGB" 1(%1), %%mm0 \n\t" + PAVGB" 1(%1, %3), %%mm1 \n\t" + PAVGB" 9(%1), %%mm2 \n\t" + PAVGB" 9(%1, %3), %%mm3 \n\t" + "movq %%mm0, (%2) \n\t" + "movq %%mm1, (%2, %3) \n\t" + "movq %%mm2, 8(%2) \n\t" + "movq %%mm3, 8(%2, %3) \n\t" + "add %%"REG_a", %1 \n\t" + "add %%"REG_a", %2 \n\t" + "movq (%1), %%mm0 \n\t" + "movq (%1, %3), %%mm1 \n\t" + "movq 8(%1), %%mm2 \n\t" + "movq 8(%1, %3), %%mm3 \n\t" + PAVGB" 1(%1), %%mm0 \n\t" + PAVGB" 1(%1, %3), %%mm1 \n\t" + PAVGB" 9(%1), %%mm2 \n\t" + PAVGB" 9(%1, %3), %%mm3 \n\t" + "add %%"REG_a", %1 \n\t" + "movq %%mm0, (%2) \n\t" + "movq %%mm1, (%2, %3) \n\t" + "movq %%mm2, 8(%2) \n\t" + "movq %%mm3, 8(%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_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $16, %2 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" 16(%2), %%mm0 \n\t" - PAVGB" 24(%2), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $2, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $16, %2 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" 16(%2), %%mm0 \n\t" + PAVGB" 24(%2), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $2, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } static void DEF(avg_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - "add %4, %1 \n\t" - "add $16, %2 \n\t" - PAVGB" (%3), %%mm0 \n\t" - PAVGB" 8(%3), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" (%2), %%mm0 \n\t" - PAVGB" 8(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - PAVGB" 8(%3), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - PAVGB" 16(%2), %%mm0 \n\t" - PAVGB" 24(%2), %%mm1 \n\t" - PAVGB" (%3), %%mm0 \n\t" - PAVGB" 8(%3), %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $2, %0 \n\t" - "jnz 1b \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + "add %4, %1 \n\t" + "add $16, %2 \n\t" + PAVGB" (%3), %%mm0 \n\t" + PAVGB" 8(%3), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" (%2), %%mm0 \n\t" + PAVGB" 8(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + PAVGB" 8(%3), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + PAVGB" 16(%2), %%mm0 \n\t" + PAVGB" 24(%2), %%mm1 \n\t" + PAVGB" (%3), %%mm0 \n\t" + PAVGB" 8(%3), %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $2, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } static void DEF(put_no_rnd_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) { __asm __volatile( - "pcmpeqb %%mm6, %%mm6\n\t" - "testl $1, %0 \n\t" - " jz 1f \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "movq (%2), %%mm2 \n\t" - "movq 8(%2), %%mm3 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "pxor %%mm6, %%mm2 \n\t" - "pxor %%mm6, %%mm3 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "add %4, %1 \n\t" - "add $16, %2 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "decl %0 \n\t" - "1: \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - "movq (%2), %%mm2 \n\t" - "movq 8(%2), %%mm3 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "pxor %%mm6, %%mm2 \n\t" - "pxor %%mm6, %%mm3 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "movq (%1), %%mm0 \n\t" - "movq 8(%1), %%mm1 \n\t" - "add %4, %1 \n\t" - "movq 16(%2), %%mm2 \n\t" - "movq 24(%2), %%mm3 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "pxor %%mm6, %%mm2 \n\t" - "pxor %%mm6, %%mm3 \n\t" - PAVGB" %%mm2, %%mm0 \n\t" - PAVGB" %%mm3, %%mm1 \n\t" - "pxor %%mm6, %%mm0 \n\t" - "pxor %%mm6, %%mm1 \n\t" - "movq %%mm0, (%3) \n\t" - "movq %%mm1, 8(%3) \n\t" - "add %5, %3 \n\t" - "add $32, %2 \n\t" - "subl $2, %0 \n\t" - "jnz 1b \n\t" + "pcmpeqb %%mm6, %%mm6 \n\t" + "testl $1, %0 \n\t" + " jz 1f \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "movq (%2), %%mm2 \n\t" + "movq 8(%2), %%mm3 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "pxor %%mm6, %%mm2 \n\t" + "pxor %%mm6, %%mm3 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "add %4, %1 \n\t" + "add $16, %2 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "decl %0 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + "movq (%2), %%mm2 \n\t" + "movq 8(%2), %%mm3 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "pxor %%mm6, %%mm2 \n\t" + "pxor %%mm6, %%mm3 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm1 \n\t" + "add %4, %1 \n\t" + "movq 16(%2), %%mm2 \n\t" + "movq 24(%2), %%mm3 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "pxor %%mm6, %%mm2 \n\t" + "pxor %%mm6, %%mm3 \n\t" + PAVGB" %%mm2, %%mm0 \n\t" + PAVGB" %%mm3, %%mm1 \n\t" + "pxor %%mm6, %%mm0 \n\t" + "pxor %%mm6, %%mm1 \n\t" + "movq %%mm0, (%3) \n\t" + "movq %%mm1, 8(%3) \n\t" + "add %5, %3 \n\t" + "add $32, %2 \n\t" + "subl $2, %0 \n\t" + "jnz 1b \n\t" #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used - :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) #else - :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) + :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) #endif - :"S"((long)src1Stride), "D"((long)dstStride) - :"memory"); + :"S"((long)src1Stride), "D"((long)dstStride) + :"memory"); //the following should be used, though better not with gcc ... -/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) - :"r"(src1Stride), "r"(dstStride) - :"memory");*/ +/* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) + :"r"(src1Stride), "r"(dstStride) + :"memory");*/ } /* GL: this function does incorrect rounding if overflow */ @@ -551,67 +551,67 @@ { 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"); + "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"); + "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 */ @@ -619,131 +619,131 @@ { 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"); + "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"); + "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"); + "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"); + "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 @@ -751,41 +751,41 @@ { 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"); + "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 ...