📄 dsputil_mmx_avg.h.svn-base
字号:
"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 cannot be used :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)#else :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)#endif :"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");*/}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");}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"#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cannot be used :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)#else :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)#endif :"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");*/}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"#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cannot be used :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)#else :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)#endif :"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");*/}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"#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cannot be used :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)#else :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)#endif :"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");*/}/* 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"
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -