项目记录：SSE/NEON快速指令集优化像素拷贝

记录一下在实验室做了的一些项目与demo

记录其中的知识点与技术要点

SSE/NEON快速指令集优化像素拷贝

项目需求:

• 360全景图划分成42 TILE.
  

• 高纬度按 4:1 下采样. 中纬度按 2:1 下采样. 低纬度 1:1

• 渲染时,高纬度按 1:4 还原. 中纬度按 1:2 还原.

参考资料:

SSE

NEON

具体实现:

// 宏定义: 以空间换时间.. 小心使用..
// ---------PC端:SSE指令集----------------
// 高纬度的1:4 拷贝 4字节->16
#define set4(output_buf_y,input_buf_y) {\
    __m128i itmp1 = _mm_loadl_epi64((__m128i*)(input_buf_y));\
    __m128i itmp2 = _mm_unpacklo_epi8(itmp1, itmp1);\
    __m128i dst = _mm_unpacklo_epi8(itmp2, itmp2);\
    __m128i* otmp = (__m128i*)(output_buf_y);\
    _mm_store_si128(otmp, dst);\
}
// 中纬度的1:2 拷贝 8字节->16
#define set2(output_buf_y,input_buf_y) {\
    __m128i itmp =_mm_loadl_epi64((__m128i*)(input_buf_y));\
    __m128i* otmp = (__m128i*)(output_buf_y);\
    _mm_store_si128(otmp, _mm_unpacklo_epi8(itmp, itmp));\
}
// ---------Android端:NEON指令集----------------
#define NEON_INTRINSICS 1// 1:neon内联函数 0:汇编
#if NEON_INTRINSICS
//1:4 拷贝 8->32
#define set4(output_buf_y,input_buf_y) {\
        uint8x8x4_t v;\
        v.val[0] = vld1_u8(input_buf_y);\
        v.val[1] = v.val[0];\
        v.val[2] = v.val[0];\
        v.val[3] = v.val[0];\
        vst4_u8(output_buf_y,v);\
}
// 1:2 拷贝 8字节->16
#define set2(output_buf_y,input_buf_y) {\
        uint8x8x2_t v;\
        v.val[0] =vld1_u8(input_buf_y);\
        v.val[1] = v.val[0];\
        vst2_u8(output_buf_y,v);\
}
#else
// 汇编版本
void set4(unsigned char * output_buf_y,unsigned char * input_buf_y) {\
    __asm__(\
        "VLD1.8 {d0},[r1] \t\n"\
        "VMOV   d1,d0 \t\n"\
        "VMOV   d2,d0 \t\n"\
        "VMOV   d3,d0 \t\n"\
        "VST4.8 {d0, d1, d2, d3}, [r0]"\
    );\
}
void set2(unsigned char *output_buf_y,unsigned char *input_buf_y) {\
    __asm__(\
        "VLD1.8 {d0},[r1] \t\n"\
        "VMOV   d1,d0 \t\n"\
        "VST2.8 {d0, d1}, [r0]"\
    );\
}
#endif

---