dsputil_mmx.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

      "add %2,%0\n"
      "subl $2, %%ecx\n"
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      "1:\n"
      
      SUM(%%mm4, %%mm5, %%mm0, %%mm1)
      
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      
      "subl $2, %%ecx\n"
      "jnz 1b\n"

      "movd %%mm6,%1\n"
      : "+r" (pix), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp;
}
#undef SUM

static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    int tmp;
    
    assert( (((int)pix1) & 7) == 0);
    assert( (((int)pix2) & 7) == 0);
    assert((line_size &7) ==0);
    
#define SUM(in0, in1, out0, out1) \
      "movq (%0),%%mm2\n"\
      "movq (%1)," #out0 "\n"\
      "movq 8(%0),%%mm3\n"\
      "movq 8(%1)," #out1 "\n"\
      "add %3,%0\n"\
      "add %3,%1\n"\
      "psubb " #out0 ", %%mm2\n"\
      "psubb " #out1 ", %%mm3\n"\
      "pxor %%mm7, %%mm2\n"\
      "pxor %%mm7, %%mm3\n"\
      "movq %%mm2, " #out0 "\n"\
      "movq %%mm3, " #out1 "\n"\
      "psubusb " #in0 ", %%mm2\n"\
      "psubusb " #in1 ", %%mm3\n"\
      "psubusb " #out0 ", " #in0 "\n"\
      "psubusb " #out1 ", " #in1 "\n"\
      "por %%mm2, " #in0 "\n"\
      "por %%mm3, " #in1 "\n"\
      "movq " #in0 ", %%mm2\n"\
      "movq " #in1 ", %%mm3\n"\
      "punpcklbw %%mm7, " #in0 "\n"\
      "punpcklbw %%mm7, " #in1 "\n"\
      "punpckhbw %%mm7, %%mm2\n"\
      "punpckhbw %%mm7, %%mm3\n"\
      "paddw " #in1 ", " #in0 "\n"\
      "paddw %%mm3, %%mm2\n"\
      "paddw %%mm2, " #in0 "\n"\
      "paddw " #in0 ", %%mm6\n"

    
  asm volatile (
      "movl %4,%%ecx\n"
      "pxor %%mm6,%%mm6\n"
      "pcmpeqw %%mm7,%%mm7\n"
      "psllw $15, %%mm7\n"
      "packsswb %%mm7, %%mm7\n"
      "movq (%0),%%mm0\n"
      "movq (%1),%%mm2\n"
      "movq 8(%0),%%mm1\n"
      "movq 8(%1),%%mm3\n"
      "add %3,%0\n"
      "add %3,%1\n"
      "subl $2, %%ecx\n"
      "psubb %%mm2, %%mm0\n"
      "psubb %%mm3, %%mm1\n"
      "pxor %%mm7, %%mm0\n"
      "pxor %%mm7, %%mm1\n"
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      "1:\n"
      
      SUM(%%mm4, %%mm5, %%mm0, %%mm1)
      
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      
      "subl $2, %%ecx\n"
      "jnz 1b\n"

      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddw %%mm6,%%mm0\n"
      "movq %%mm0,%%mm6\n"
      "psrlq $16, %%mm0\n"
      "paddw %%mm6,%%mm0\n"
      "movd %%mm0,%2\n"
      : "+r" (pix1), "+r" (pix2), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp & 0x7FFF;
}
#undef SUM

static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    int tmp;
    
    assert( (((int)pix1) & 7) == 0);
    assert( (((int)pix2) & 7) == 0);
    assert((line_size &7) ==0);
    
#define SUM(in0, in1, out0, out1) \
      "movq (%0)," #out0 "\n"\
      "movq (%1),%%mm2\n"\
      "movq 8(%0)," #out1 "\n"\
      "movq 8(%1),%%mm3\n"\
      "add %3,%0\n"\
      "add %3,%1\n"\
      "psubb %%mm2, " #out0 "\n"\
      "psubb %%mm3, " #out1 "\n"\
      "pxor %%mm7, " #out0 "\n"\
      "pxor %%mm7, " #out1 "\n"\
      "psadbw " #out0 ", " #in0 "\n"\
      "psadbw " #out1 ", " #in1 "\n"\
      "paddw " #in1 ", " #in0 "\n"\
      "paddw " #in0 ", %%mm6\n"

  asm volatile (
      "movl %4,%%ecx\n"
      "pxor %%mm6,%%mm6\n"
      "pcmpeqw %%mm7,%%mm7\n"
      "psllw $15, %%mm7\n"
      "packsswb %%mm7, %%mm7\n"
      "movq (%0),%%mm0\n"
      "movq (%1),%%mm2\n"
      "movq 8(%0),%%mm1\n"
      "movq 8(%1),%%mm3\n"
      "add %3,%0\n"
      "add %3,%1\n"
      "subl $2, %%ecx\n"
      "psubb %%mm2, %%mm0\n"
      "psubb %%mm3, %%mm1\n"
      "pxor %%mm7, %%mm0\n"
      "pxor %%mm7, %%mm1\n"
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      "1:\n"
      
      SUM(%%mm4, %%mm5, %%mm0, %%mm1)
      
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      
      "subl $2, %%ecx\n"
      "jnz 1b\n"

      "movd %%mm6,%2\n"
      : "+r" (pix1), "+r" (pix2), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp;
}
#undef SUM

static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
    long i=0;
    asm volatile(
        "1:				\n\t"
        "movq  (%2, %0), %%mm0		\n\t"
        "movq  (%1, %0), %%mm1		\n\t"
        "psubb %%mm0, %%mm1		\n\t"
        "movq %%mm1, (%3, %0)		\n\t"
        "movq 8(%2, %0), %%mm0		\n\t"
        "movq 8(%1, %0), %%mm1		\n\t"
        "psubb %%mm0, %%mm1		\n\t"
        "movq %%mm1, 8(%3, %0)		\n\t"
        "add $16, %0			\n\t"
        "cmp %4, %0			\n\t"
        " jb 1b				\n\t"
        : "+r" (i)
        : "r"(src1), "r"(src2), "r"(dst), "r"((long)w-15)
    );
    for(; i<w; i++)
        dst[i+0] = src1[i+0]-src2[i+0];
}

static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){
    long i=0;
    uint8_t l, lt;
    
    asm volatile(
        "1:				\n\t"
        "movq  -1(%1, %0), %%mm0	\n\t" // LT
        "movq  (%1, %0), %%mm1		\n\t" // T
        "movq  -1(%2, %0), %%mm2	\n\t" // L
        "movq  (%2, %0), %%mm3		\n\t" // X
        "movq %%mm2, %%mm4		\n\t" // L
        "psubb %%mm0, %%mm2		\n\t"
        "paddb %%mm1, %%mm2		\n\t" // L + T - LT
        "movq %%mm4, %%mm5		\n\t" // L
        "pmaxub %%mm1, %%mm4		\n\t" // max(T, L)
        "pminub %%mm5, %%mm1		\n\t" // min(T, L)
        "pminub %%mm2, %%mm4		\n\t" 
        "pmaxub %%mm1, %%mm4		\n\t"
        "psubb %%mm4, %%mm3		\n\t" // dst - pred
        "movq %%mm3, (%3, %0)		\n\t"
        "add $8, %0			\n\t"
        "cmp %4, %0			\n\t"
        " jb 1b				\n\t"
        : "+r" (i)
        : "r"(src1), "r"(src2), "r"(dst), "r"((long)w)
    );

    l= *left;
    lt= *left_top;
    
    dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF);
    
    *left_top= src1[w-1];
    *left    = src2[w-1];
}

#define LBUTTERFLY2(a1,b1,a2,b2)\
    "paddw " #b1 ", " #a1 "		\n\t"\
    "paddw " #b2 ", " #a2 "		\n\t"\
    "paddw " #b1 ", " #b1 "		\n\t"\
    "paddw " #b2 ", " #b2 "		\n\t"\
    "psubw " #a1 ", " #b1 "		\n\t"\
    "psubw " #a2 ", " #b2 "		\n\t"

#define HADAMARD48\
        LBUTTERFLY2(%%mm0, %%mm1, %%mm2, %%mm3)\
        LBUTTERFLY2(%%mm4, %%mm5, %%mm6, %%mm7)\
        LBUTTERFLY2(%%mm0, %%mm2, %%mm1, %%mm3)\
        LBUTTERFLY2(%%mm4, %%mm6, %%mm5, %%mm7)\
        LBUTTERFLY2(%%mm0, %%mm4, %%mm1, %%mm5)\
        LBUTTERFLY2(%%mm2, %%mm6, %%mm3, %%mm7)\

#define MMABS(a,z)\
    "pxor " #z ", " #z "		\n\t"\
    "pcmpgtw " #a ", " #z "		\n\t"\
    "pxor " #z ", " #a "		\n\t"\
    "psubw " #z ", " #a "		\n\t"

#define MMABS_SUM(a,z, sum)\
    "pxor " #z ", " #z "		\n\t"\
    "pcmpgtw " #a ", " #z "		\n\t"\
    "pxor " #z ", " #a "		\n\t"\
    "psubw " #z ", " #a "		\n\t"\
    "paddusw " #a ", " #sum "		\n\t"

#define MMABS_MMX2(a,z)\
    "pxor " #z ", " #z "		\n\t"\
    "psubw " #a ", " #z "		\n\t"\
    "pmaxsw " #z ", " #a "		\n\t"

#define MMABS_SUM_MMX2(a,z, sum)\
    "pxor " #z ", " #z "		\n\t"\
    "psubw " #a ", " #z "		\n\t"\
    "pmaxsw " #z ", " #a "		\n\t"\
    "paddusw " #a ", " #sum "		\n\t"
        
#define SBUTTERFLY(a,b,t,n)\
    "movq " #a ", " #t "		\n\t" /* abcd */\
    "punpckl" #n " " #b ", " #a "	\n\t" /* aebf */\
    "punpckh" #n " " #b ", " #t "	\n\t" /* cgdh */\

#define TRANSPOSE4(a,b,c,d,t)\
    SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\
    SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\
    SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\
    SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */

#define LOAD4(o, a, b, c, d)\
        "movq "#o"(%1), " #a "		\n\t"\
        "movq "#o"+16(%1), " #b "	\n\t"\
        "movq "#o"+32(%1), " #c "	\n\t"\
        "movq "#o"+48(%1), " #d "	\n\t"

#define STORE4(o, a, b, c, d)\
        "movq "#a", "#o"(%1)		\n\t"\
        "movq "#b", "#o"+16(%1)		\n\t"\
        "movq "#c", "#o"+32(%1)		\n\t"\
        "movq "#d", "#o"+48(%1)		\n\t"\

static int hadamard8_diff_mmx(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){
    uint64_t temp[16] __align8;
    int sum=0;
    
    assert(h==8);

    diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride);

    asm volatile(
        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
        LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        
        "movq %%mm7, 112(%1)		\n\t"
        
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
        STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)
        
        "movq 112(%1), %%mm7 		\n\t"
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
        STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)

        LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3)
        LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        
        "movq %%mm7, 120(%1)		\n\t"
        
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
        STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2)
        
        "movq 120(%1), %%mm7 		\n\t"
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
        "movq %%mm7, %%mm5		\n\t"//FIXME remove
        "movq %%mm6, %%mm7		\n\t"
        "movq %%mm0, %%mm6		\n\t"
//        STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove
        
        LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)
//        LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        "movq %%mm7, 64(%1)		\n\t"
        MMABS(%%mm0, %%mm7)
        MMABS_SUM(%%mm1, %%mm7, %%mm0)
        MMABS_SUM(%%mm2, %%mm7, %%mm0)
        MMABS_SUM(%%mm3, %%mm7, %%mm0)
        MMABS_SUM(%%mm4, %%mm7, %%mm0)
        MMABS_SUM(%%mm5, %%mm7, %%mm0)
        MMABS_SUM(%%mm6, %%mm7, %%mm0)
        "movq 64(%1), %%mm1		\n\t"
        MMABS_SUM(%%mm1, %%mm7, %%mm0)
        "movq %%mm0, 64(%1)		\n\t"
        
        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
        LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        "movq %%mm7, (%1)		\n\t"
        MMABS(%%mm0, %%mm7)
        MMABS_SUM(%%mm1, %%mm7, %%mm0)
        MMABS_SUM(%%mm2, %%mm7, %%mm0)
        MMABS_SUM(%%mm3, %%mm7, %%mm0)
        MMABS_SUM(%%mm4, %%mm7, %%mm0)
        MMABS_SUM(%%mm5, %%mm7, %%mm0)
        MMABS_SUM(%%mm6, %%mm7, %%mm0)
        "movq (%1), %%mm1		\n\t"
        MMABS_SUM(%%mm1, %%mm7, %%mm0)
        "movq 64(%1), %%mm1		\n\t"
        MMABS_SUM(%%mm1, %%mm7, %%mm0)
        
        "movq %%mm0, %%mm1		\n\t"
        "psrlq $32, %%mm0		\n\t"
        "paddusw %%mm1, %%mm0		\n\t"
        "movq %%mm0, %%mm1		\n\t"
        "psrlq $16, %%mm0		\n\t"
        "paddusw %%mm1, %%mm0		\n\t"
        "movd %%mm0, %0			\n\t"
                
        : "=r" (sum)
        : "r"(temp)
    );
    return sum&0xFFFF;
}

static int hadamard8_diff_mmx2(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){
    uint64_t temp[16] __align8;
    int sum=0;
    
    assert(h==8);

    diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride);

    asm volatile(
        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)
        LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        
        "movq %%mm7, 112(%1)		\n\t"
        
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)
        STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)
        
        "movq 112(%1), %%mm7 		\n\t"
        TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)
        STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)

        LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3)
        LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7)
        
        HADAMARD48
        
        "movq %%mm7, 120(%1)		\n\t"