vf_fspp.c

来自「君正早期ucos系统(只有早期的才不没有打包成库),MPLAYER,文件系统,图」· C语言 代码 · 共 2,126 行 · 第 1/5 页

	"movq %%mm0, %%mm6             \n\t"
	"punpckldq %%mm2, %%mm0        \n\t" //0

	"punpckhdq %%mm2, %%mm6        \n\t" //1
	"movq %%mm0, %%mm5             \n\t"

	"punpckhwd %%mm3, %%mm7        \n\t"
	"psubw %%mm6, %%mm0            \n\t"

	"pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
	"movq %%mm4, %%mm2             \n\t"

	"punpckldq %%mm7, %%mm4        \n\t" //2
	"paddw %%mm6, %%mm5            \n\t"

	"punpckhdq %%mm7, %%mm2        \n\t" //3
	"movq %%mm4, %%mm1             \n\t"

	"psllw $2, %%mm0              \n\t"
	"paddw %%mm2, %%mm4            \n\t" //t10

	"movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
	"psubw %%mm2, %%mm1            \n\t" //t11

	"movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
	"psubw %%mm5, %%mm0            \n\t"

	"movq %%mm4, %%mm6             \n\t"
	"paddw %%mm5, %%mm4            \n\t" //t0

	"psubw %%mm5, %%mm6            \n\t" //t3
	"movq %%mm1, %%mm7             \n\t"

	"movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
	"paddw %%mm0, %%mm1            \n\t" //t1

	"movq %%mm4, "MANGLE(temps)"+0*8       \n\t" //t0
	"movq %%mm3, %%mm4             \n\t"

	"movq %%mm6, "MANGLE(temps)"+1*8       \n\t" //t3
	"punpcklwd %%mm2, %%mm3        \n\t"

	//transpose 4x4    
	"movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
	"punpckhwd %%mm2, %%mm4        \n\t"

	"movq %%mm5, %%mm2             \n\t"
	"punpcklwd %%mm6, %%mm5        \n\t"

	"psubw %%mm0, %%mm7            \n\t" //t2    
	"punpckhwd %%mm6, %%mm2        \n\t"

	"movq %%mm3, %%mm0             \n\t"
	"punpckldq %%mm5, %%mm3        \n\t" //4

	"punpckhdq %%mm5, %%mm0        \n\t" //5
	"movq %%mm4, %%mm5             \n\t"

	//
	"movq %%mm3, %%mm6             \n\t"
	"punpckldq %%mm2, %%mm4        \n\t" //6

	"psubw %%mm0, %%mm3            \n\t" //z10
	"punpckhdq %%mm2, %%mm5        \n\t" //7     

	"paddw %%mm0, %%mm6            \n\t" //z13
	"movq %%mm4, %%mm2             \n\t"

	"movq %%mm3, %%mm0             \n\t"
	"psubw %%mm5, %%mm4            \n\t" //z12    

	"pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //-
	"paddw %%mm4, %%mm3            \n\t"

	"pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5
	"paddw %%mm5, %%mm2            \n\t" //z11  >

	"pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t"
	"movq %%mm2, %%mm5             \n\t"

	"psubw %%mm6, %%mm2            \n\t"
	"paddw %%mm6, %%mm5            \n\t" //t7

	"pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11    
	"paddw %%mm3, %%mm0            \n\t" //t12

	"psllw $3, %%mm0              \n\t"
	"psubw %%mm3, %%mm4            \n\t" //t10    

	"movq "MANGLE(temps)"+0*8, %%mm6       \n\t"
	"movq %%mm1, %%mm3             \n\t"

	"psllw $3, %%mm4              \n\t"
	"psubw %%mm5, %%mm0            \n\t" //t6

	"psllw $3, %%mm2              \n\t"
	"paddw %%mm0, %%mm1            \n\t" //d1

	"psubw %%mm0, %%mm2            \n\t" //t5
	"psubw %%mm0, %%mm3            \n\t" //d6         

	"paddw %%mm2, %%mm4            \n\t" //t4
	"movq %%mm7, %%mm0             \n\t"

	"paddw %%mm2, %%mm7            \n\t" //d2
	"psubw %%mm2, %%mm0            \n\t" //d5

	"movq "MANGLE(MM_DESCALE_RND)", %%mm2   \n\t" //4
	"psubw %%mm5, %%mm6            \n\t" //d7

	"paddw "MANGLE(temps)"+0*8, %%mm5      \n\t" //d0
	"paddw %%mm2, %%mm1            \n\t"

	"paddw %%mm2, %%mm5            \n\t"
	"psraw $3, %%mm1              \n\t"

	"paddw %%mm2, %%mm7            \n\t"
	"psraw $3, %%mm5              \n\t"

	"paddw (%%"REG_D"), %%mm5          \n\t"
	"psraw $3, %%mm7              \n\t"

	"paddw (%%"REG_D",%%"REG_a",), %%mm1    \n\t"
	"paddw %%mm2, %%mm0            \n\t"

	"paddw (%%"REG_D",%%"REG_a",2), %%mm7   \n\t"
	"paddw %%mm2, %%mm3            \n\t"

	"movq %%mm5, (%%"REG_D")           \n\t"
	"paddw %%mm2, %%mm6            \n\t"

	"movq %%mm1, (%%"REG_D",%%"REG_a",)     \n\t"
	"psraw $3, %%mm0              \n\t"

	"movq %%mm7, (%%"REG_D",%%"REG_a",2)    \n\t"
	"add %%"REG_d", %%"REG_D"             \n\t" //3*ls

	"movq "MANGLE(temps)"+1*8, %%mm5       \n\t" //t3
	"psraw $3, %%mm3              \n\t"

	"paddw (%%"REG_D",%%"REG_a",2), %%mm0   \n\t"
	"psubw %%mm4, %%mm5            \n\t" //d3

	"paddw (%%"REG_D",%%"REG_d",), %%mm3    \n\t"
	"psraw $3, %%mm6              \n\t"

	"paddw "MANGLE(temps)"+1*8, %%mm4      \n\t" //d4        
	"paddw %%mm2, %%mm5            \n\t"

	"paddw (%%"REG_D",%%"REG_a",4), %%mm6   \n\t"
	"paddw %%mm2, %%mm4            \n\t"

	"movq %%mm0, (%%"REG_D",%%"REG_a",2)    \n\t"
	"psraw $3, %%mm5              \n\t"

	"paddw (%%"REG_D"), %%mm5          \n\t"
	"psraw $3, %%mm4              \n\t"

	"paddw (%%"REG_D",%%"REG_a",), %%mm4    \n\t"
	"add $"DCTSIZE_S"*2*4, %%"REG_S"      \n\t" //4 rows

	"movq %%mm3, (%%"REG_D",%%"REG_d",)     \n\t"
	"movq %%mm6, (%%"REG_D",%%"REG_a",4)    \n\t"
	"movq %%mm5, (%%"REG_D")           \n\t"
	"movq %%mm4, (%%"REG_D",%%"REG_a",)     \n\t"

	"sub %%"REG_d", %%"REG_D"             \n\t"
	"add $8, %%"REG_D"               \n\t"
	"dec %%"REG_c"                   \n\t"
	"jnz 1b                  \n\t"

	: "+S"(workspace), "+D"(output_adr), "+c"(cnt) //input regs
	: "a"(output_stride*sizeof(short))
	: "%"REG_d
	);
}

#endif // HAVE_MMX

#ifndef HAVE_MMX

static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt)
{
    int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    int_simd16_t tmp10, tmp11, tmp12, tmp13;
    int_simd16_t z1, z2, z3, z4, z5, z11, z13;
    DCTELEM *dataptr;  
  
    cnt*=4;
    // Pass 1: process rows. 
  
    dataptr = data;
    for (; cnt > 0; cnt--) {    
	tmp0 = pixels[line_size*0] + pixels[line_size*7];
	tmp7 = pixels[line_size*0] - pixels[line_size*7];
	tmp1 = pixels[line_size*1] + pixels[line_size*6];
	tmp6 = pixels[line_size*1] - pixels[line_size*6];
	tmp2 = pixels[line_size*2] + pixels[line_size*5];
	tmp5 = pixels[line_size*2] - pixels[line_size*5];
	tmp3 = pixels[line_size*3] + pixels[line_size*4];
	tmp4 = pixels[line_size*3] - pixels[line_size*4];
    
	// Even part 
    
	tmp10 = tmp0 + tmp3;    
	tmp13 = tmp0 - tmp3;
	tmp11 = tmp1 + tmp2;
	tmp12 = tmp1 - tmp2;
	//Even columns are written first, this leads to different order of columns 
	//in column_fidct(), but they are processed independently, so all ok.
	//Later in the row_idct() columns readed at the same order.
	dataptr[2] = tmp10 + tmp11; 
	dataptr[3] = tmp10 - tmp11;
    
	z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781);
	dataptr[0] = tmp13 + z1;    
	dataptr[1] = tmp13 - z1;
    
	// Odd part 

	tmp10 = (tmp4 + tmp5) <<2;  
	tmp11 = (tmp5 + tmp6) <<2;
	tmp12 = (tmp6 + tmp7) <<2;

	z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
	z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
	z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
	z3 = MULTIPLY16H(tmp11, FIX_0_707106781);

	z11 = tmp7 + z3;
	z13 = tmp7 - z3;

	dataptr[4] = z13 + z2;
	dataptr[5] = z13 - z2;
	dataptr[6] = z11 + z4;
	dataptr[7] = z11 - z4;

	pixels++;               // advance pointer to next column
	dataptr += DCTSIZE;         
    }
}

#else /* HAVE_MMX */

static void row_fdct_mmx(DCTELEM *data,  const uint8_t *pixels,  int line_size,  int cnt)
{
    asm volatile(
	"lea (%%"REG_a",%%"REG_a",2), %%"REG_d"    \n\t"
	"6:                     \n\t"
	"movd (%%"REG_S"), %%mm0           \n\t"
	"pxor %%mm7, %%mm7             \n\t"

	"movd (%%"REG_S",%%"REG_a",), %%mm1     \n\t"
	"punpcklbw %%mm7, %%mm0        \n\t"

	"movd (%%"REG_S",%%"REG_a",2), %%mm2    \n\t"
	"punpcklbw %%mm7, %%mm1        \n\t"

	"punpcklbw %%mm7, %%mm2        \n\t"
	"add %%"REG_d", %%"REG_S"             \n\t"

	"movq %%mm0, %%mm5             \n\t"
	//       

	"movd (%%"REG_S",%%"REG_a",4), %%mm3    \n\t" //7  ;prefetch!
	"movq %%mm1, %%mm6             \n\t"

	"movd (%%"REG_S",%%"REG_d",), %%mm4     \n\t" //6
	"punpcklbw %%mm7, %%mm3        \n\t"

	"psubw %%mm3, %%mm5            \n\t"
	"punpcklbw %%mm7, %%mm4        \n\t"

	"paddw %%mm3, %%mm0            \n\t"
	"psubw %%mm4, %%mm6            \n\t"

	"movd (%%"REG_S",%%"REG_a",2), %%mm3    \n\t" //5
	"paddw %%mm4, %%mm1            \n\t"

	"movq %%mm5, "MANGLE(temps)"+0*8       \n\t" //t7
	"punpcklbw %%mm7, %%mm3        \n\t"

	"movq %%mm6, "MANGLE(temps)"+1*8       \n\t" //t6
	"movq %%mm2, %%mm4             \n\t"

	"movd (%%"REG_S"), %%mm5           \n\t" //3
	"paddw %%mm3, %%mm2            \n\t"

	"movd (%%"REG_S",%%"REG_a",), %%mm6     \n\t" //4
	"punpcklbw %%mm7, %%mm5        \n\t"

	"psubw %%mm3, %%mm4            \n\t"
	"punpcklbw %%mm7, %%mm6        \n\t"

	"movq %%mm5, %%mm3             \n\t"
	"paddw %%mm6, %%mm5            \n\t" //t3

	"psubw %%mm6, %%mm3            \n\t" //t4  ; t0 t1 t2 t4 t5 t3 - -
	"movq %%mm0, %%mm6             \n\t"

	"movq %%mm1, %%mm7             \n\t"
	"psubw %%mm5, %%mm0            \n\t" //t13

	"psubw %%mm2, %%mm1            \n\t"
	"paddw %%mm2, %%mm7            \n\t" //t11    

	"paddw %%mm0, %%mm1            \n\t"
	"movq %%mm7, %%mm2             \n\t"

	"psllw $2, %%mm1              \n\t"
	"paddw %%mm5, %%mm6            \n\t" //t10

	"pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t"
	"paddw %%mm6, %%mm7            \n\t" //d2

	"psubw %%mm2, %%mm6            \n\t" //d3
	"movq %%mm0, %%mm5             \n\t"

	//transpose 4x4
	"movq %%mm7, %%mm2             \n\t"
	"punpcklwd %%mm6, %%mm7        \n\t"

	"paddw %%mm1, %%mm0            \n\t" //d0
	"punpckhwd %%mm6, %%mm2        \n\t"

	"psubw %%mm1, %%mm5            \n\t" //d1                
	"movq %%mm0, %%mm6             \n\t"

	"movq "MANGLE(temps)"+1*8, %%mm1       \n\t"
	"punpcklwd %%mm5, %%mm0        \n\t"

	"punpckhwd %%mm5, %%mm6        \n\t"
	"movq %%mm0, %%mm5             \n\t"

	"punpckldq %%mm7, %%mm0        \n\t" //0
	"paddw %%mm4, %%mm3            \n\t"

	"punpckhdq %%mm7, %%mm5        \n\t" //1
	"movq %%mm6, %%mm7             \n\t"

	"movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
	"punpckldq %%mm2, %%mm6        \n\t" //2     

	"movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
	"punpckhdq %%mm2, %%mm7        \n\t" //3    

	"movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
	"paddw %%mm1, %%mm4            \n\t"

	"movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
	"psllw $2, %%mm3              \n\t" //t10    

	"movq "MANGLE(temps)"+0*8, %%mm2       \n\t"
	"psllw $2, %%mm4              \n\t" //t11

	"pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
	"paddw %%mm2, %%mm1            \n\t"

	"psllw $2, %%mm1              \n\t" //t12
	"movq %%mm3, %%mm0             \n\t"

	"pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t"
	"psubw %%mm1, %%mm3            \n\t"

	"pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
	"movq %%mm2, %%mm5             \n\t"

	"pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
	"psubw %%mm4, %%mm2            \n\t" //z13

	"paddw %%mm4, %%mm5            \n\t" //z11
	"movq %%mm2, %%mm6             \n\t"

	"paddw %%mm3, %%mm0            \n\t" //z2
	"movq %%mm5, %%mm7             \n\t"

	"paddw %%mm0, %%mm2            \n\t" //d4
	"psubw %%mm0, %%mm6            \n\t" //d5    

	"movq %%mm2, %%mm4             \n\t"
	"paddw %%mm3, %%mm1            \n\t" //z4    

	//transpose 4x4
	"punpcklwd %%mm6, %%mm2        \n\t"
	"paddw %%mm1, %%mm5            \n\t" //d6

	"punpckhwd %%mm6, %%mm4        \n\t"
	"psubw %%mm1, %%mm7            \n\t" //d7    

	"movq %%mm5, %%mm6             \n\t"
	"punpcklwd %%mm7, %%mm5        \n\t"

	"punpckhwd %%mm7, %%mm6        \n\t"
	"movq %%mm2, %%mm7             \n\t"

	"punpckldq %%mm5, %%mm2        \n\t" //4
	"sub %%"REG_d", %%"REG_S"             \n\t"

	"punpckhdq %%mm5, %%mm7        \n\t" //5
	"movq %%mm4, %%mm5             \n\t"

	"movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
	"punpckldq %%mm6, %%mm4        \n\t" //6

	"movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
	"punpckhdq %%mm6, %%mm5        \n\t" //7    

	"movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
	"add $4, %%"REG_S"               \n\t"

	"movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
	"add $"DCTSIZE_S"*2*4, %%"REG_D"      \n\t" //4 rows    
	"dec %%"REG_c"                   \n\t"
	"jnz 6b                  \n\t"

	: "+S"(pixels), "+D"(data), "+c"(cnt) //input regs
	: "a"(line_size)
	: "%"REG_d);
}

#endif // HAVE_MMX