dsputil_altivec.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.

       pixelsv2 = temp2;
     }
     else
     {
       pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
     }

     pixelsv1 = vec_mergeh(vczero, pixelsv1);
     pixelsv2 = vec_mergeh(vczero, pixelsv2);
     pixelssum2 = vec_add((vector unsigned short)pixelsv1,
                          (vector unsigned short)pixelsv2);
     temp3 = vec_add(pixelssum1, pixelssum2);
     temp3 = vec_sra(temp3, vctwo);
     pixelssum1 = vec_add(pixelssum2, vcone);
     pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
     
     if (rightside)
     {
       blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
     }
     else
     {
       blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
     }
     
     vec_st(blockv, 0, block);
     
     block += line_size;
     pixels += line_size;
   }
   
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* next one assumes that ((line_size % 16) == 0) */
void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int j;
POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
      for (j = 0; j < 4; j++) {
      int i;
      const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
      const uint32_t b =
        (((const struct unaligned_32 *) (pixels + 1))->l);
      uint32_t l0 =
        (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
      uint32_t h0 =
        ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
      uint32_t l1, h1;
      pixels += line_size;
      for (i = 0; i < h; i += 2) {
        uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
        uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
        h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
        a = (((const struct unaligned_32 *) (pixels))->l);
        b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
        h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
      } pixels += 4 - line_size * (h + 1);
      block += 4 - line_size * h;
    }

POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
   register int i;
   register vector unsigned char
     pixelsv1, pixelsv2, pixelsv3, pixelsv4;
   register vector unsigned char
     blockv, temp1, temp2;
   register vector unsigned short
     pixelssum1, pixelssum2, temp3,
     pixelssum3, pixelssum4, temp4;
   register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
   register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);

POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
 
   temp1 = vec_ld(0, pixels);
   temp2 = vec_ld(16, pixels);
   pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F)
   {
     pixelsv2 = temp2;
   }
   else
   {
     pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   }
   pixelsv3 = vec_mergel(vczero, pixelsv1);
   pixelsv4 = vec_mergel(vczero, pixelsv2);
   pixelsv1 = vec_mergeh(vczero, pixelsv1);
   pixelsv2 = vec_mergeh(vczero, pixelsv2);
   pixelssum3 = vec_add((vector unsigned short)pixelsv3,
                        (vector unsigned short)pixelsv4);
   pixelssum3 = vec_add(pixelssum3, vctwo);
   pixelssum1 = vec_add((vector unsigned short)pixelsv1,
                        (vector unsigned short)pixelsv2);
   pixelssum1 = vec_add(pixelssum1, vctwo);
   
   for (i = 0; i < h ; i++) {
     blockv = vec_ld(0, block);

     temp1 = vec_ld(line_size, pixels);
     temp2 = vec_ld(line_size + 16, pixels);
     pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
     if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F)
     {
       pixelsv2 = temp2;
     }
     else
     {
       pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
     }

     pixelsv3 = vec_mergel(vczero, pixelsv1);
     pixelsv4 = vec_mergel(vczero, pixelsv2);
     pixelsv1 = vec_mergeh(vczero, pixelsv1);
     pixelsv2 = vec_mergeh(vczero, pixelsv2);
     
     pixelssum4 = vec_add((vector unsigned short)pixelsv3,
                          (vector unsigned short)pixelsv4);
     pixelssum2 = vec_add((vector unsigned short)pixelsv1,
                          (vector unsigned short)pixelsv2);
     temp4 = vec_add(pixelssum3, pixelssum4);
     temp4 = vec_sra(temp4, vctwo);
     temp3 = vec_add(pixelssum1, pixelssum2);
     temp3 = vec_sra(temp3, vctwo);

     pixelssum3 = vec_add(pixelssum4, vctwo);
     pixelssum1 = vec_add(pixelssum2, vctwo);

     blockv = vec_packsu(temp3, temp4);
     
     vec_st(blockv, 0, block);
     
     block += line_size;
     pixels += line_size;
   }
   
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* next one assumes that ((line_size % 16) == 0) */
void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int j;
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
      for (j = 0; j < 4; j++) {
      int i;
      const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
      const uint32_t b =
        (((const struct unaligned_32 *) (pixels + 1))->l);
      uint32_t l0 =
        (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
      uint32_t h0 =
        ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
      uint32_t l1, h1;
      pixels += line_size;
      for (i = 0; i < h; i += 2) {
        uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
        uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
        h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
        a = (((const struct unaligned_32 *) (pixels))->l);
        b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
        h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
      } pixels += 4 - line_size * (h + 1);
      block += 4 - line_size * h;
    }

POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
   register int i;
   register vector unsigned char
     pixelsv1, pixelsv2, pixelsv3, pixelsv4;
   register vector unsigned char
     blockv, temp1, temp2;
   register vector unsigned short
     pixelssum1, pixelssum2, temp3,
     pixelssum3, pixelssum4, temp4;
   register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
   register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
   register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);

POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
 
   temp1 = vec_ld(0, pixels);
   temp2 = vec_ld(16, pixels);
   pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F)
   {
     pixelsv2 = temp2;
   }
   else
   {
     pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   }
   pixelsv3 = vec_mergel(vczero, pixelsv1);
   pixelsv4 = vec_mergel(vczero, pixelsv2);
   pixelsv1 = vec_mergeh(vczero, pixelsv1);
   pixelsv2 = vec_mergeh(vczero, pixelsv2);
   pixelssum3 = vec_add((vector unsigned short)pixelsv3,
                        (vector unsigned short)pixelsv4);
   pixelssum3 = vec_add(pixelssum3, vcone);
   pixelssum1 = vec_add((vector unsigned short)pixelsv1,
                        (vector unsigned short)pixelsv2);
   pixelssum1 = vec_add(pixelssum1, vcone);
   
   for (i = 0; i < h ; i++) {
     blockv = vec_ld(0, block);

     temp1 = vec_ld(line_size, pixels);
     temp2 = vec_ld(line_size + 16, pixels);
     pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
     if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F)
     {
       pixelsv2 = temp2;
     }
     else
     {
       pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
     }

     pixelsv3 = vec_mergel(vczero, pixelsv1);
     pixelsv4 = vec_mergel(vczero, pixelsv2);
     pixelsv1 = vec_mergeh(vczero, pixelsv1);
     pixelsv2 = vec_mergeh(vczero, pixelsv2);
     
     pixelssum4 = vec_add((vector unsigned short)pixelsv3,
                          (vector unsigned short)pixelsv4);
     pixelssum2 = vec_add((vector unsigned short)pixelsv1,
                          (vector unsigned short)pixelsv2);
     temp4 = vec_add(pixelssum3, pixelssum4);
     temp4 = vec_sra(temp4, vctwo);
     temp3 = vec_add(pixelssum1, pixelssum2);
     temp3 = vec_sra(temp3, vctwo);

     pixelssum3 = vec_add(pixelssum4, vcone);
     pixelssum1 = vec_add(pixelssum2, vcone);

     blockv = vec_packsu(temp3, temp4);
     
     vec_st(blockv, 0, block);
     
     block += line_size;
     pixels += line_size;
   }
   
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

#ifdef CONFIG_DARWIN
int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
  int sum;
POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
  register const_vector unsigned char vzero = (const_vector unsigned char)vec_splat_u8(0);
  register vector signed short temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
  {
    register const_vector signed short vprod1 = (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
    register const_vector signed short vprod2 = (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
    register const_vector signed short vprod3 = (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
    register const_vector unsigned char perm1 = (const_vector unsigned char)
      AVV(0x02, 0x03, 0x00, 0x01,
       0x06, 0x07, 0x04, 0x05,
       0x0A, 0x0B, 0x08, 0x09,
       0x0E, 0x0F, 0x0C, 0x0D);
    register const_vector unsigned char perm2 = (const_vector unsigned char)
      AVV(0x04, 0x05, 0x06, 0x07,
       0x00, 0x01, 0x02, 0x03,
       0x0C, 0x0D, 0x0E, 0x0F,
       0x08, 0x09, 0x0A, 0x0B);
    register const_vector unsigned char perm3 = (const_vector unsigned char)
      AVV(0x08, 0x09, 0x0A, 0x0B,
       0x0C, 0x0D, 0x0E, 0x0F,
       0x00, 0x01, 0x02, 0x03,
       0x04, 0x05, 0x06, 0x07);

#define ONEITERBUTTERFLY(i, res)					\
    {									\
      register vector unsigned char src1, src2, srcO;		       	\
      register vector unsigned char dst1, dst2, dstO;		       	\
      src1 = vec_ld(stride * i, src);					\
      if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8)	\
	src2 = vec_ld((stride * i) + 16, src);				\
      srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src));		\
      dst1 = vec_ld(stride * i, dst);					\
      if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8)	\
	dst2 = vec_ld((stride * i) + 16, dst);				\
      dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst));		\
      /* promote the unsigned chars to signed shorts */			\
      /* we're in the 8x8 function, we only care for the first 8 */	\
      register vector signed short srcV =			       	\
	(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)srcO); \
      register vector signed short dstV =			       	\
	(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)dstO); \
      /* substractions inside the first butterfly */			\
      register vector signed short but0 = vec_sub(srcV, dstV);	       	\
      register vector signed short op1 = vec_perm(but0, but0, perm1);  	\
      register vector signed short but1 = vec_mladd(but0, vprod1, op1);	\
      register vector signed short op2 = vec_perm(but1, but1, perm2);  	\
      register vector signed short but2 = vec_mladd(but1, vprod2, op2);	\
      register vector signed short op3 = vec_perm(but2, but2, perm3);  	\
      res = vec_mladd(but2, vprod3, op3);				\
    }
    ONEITERBUTTERFLY(0, temp0);
    ONEITERBUTTERFLY(1, temp1);
    ONEITERBUTTERFLY(2, temp2);
    ONEITERBUTTERFLY(3, temp3);
    ONEITERBUTTERFLY(4, temp4);
    ONEITERBUTTERFLY(5, temp5);
    ONEITERBUTTERFLY(6, temp6);
    ONEITERBUTTERFLY(7, temp7);
  }
#undef ONEITERBUTTERFLY
  {
    register vector signed int vsum;