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

	    tmp7 = (z11 + z13)>>2; //+2 !
	    tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562);
	    z5 =    MULTIPLY16H(z10 + z12, FIX_1_847759065);
	    tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
	    tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !!

	    tmp6 = tmp12 - tmp7;
	    tmp5 = tmp11 - tmp6;
	    tmp4 = tmp10 + tmp5;

	    wsptr[DCTSIZE*0]+=  (tmp0 + tmp7);
	    wsptr[DCTSIZE*1]+=  (tmp1 + tmp6);
	    wsptr[DCTSIZE*2]+=  (tmp2 + tmp5);
	    wsptr[DCTSIZE*3]+=  (tmp3 - tmp4);
	    wsptr[DCTSIZE*4]+=  (tmp3 + tmp4);
	    wsptr[DCTSIZE*5]+=  (tmp2 - tmp5);
	    wsptr[DCTSIZE*6]=  (tmp1 - tmp6);
	    wsptr[DCTSIZE*7]=  (tmp0 - tmp7);
	    //
	    dataptr++; //next column
	    wsptr++;
	    threshold++;
	}
	dataptr+=8; //skip each second start pos
	wsptr  +=8;       
    }
}

#else /* HAVE_MMX */

static void column_fidct_mmx(int16_t* thr_adr,  DCTELEM *data,  DCTELEM *output,  int cnt)
{
    asm volatile(
	ASMALIGN(4)
	"1:                   \n\t"
	"movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
	//
	"movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
	"movq %%mm1, %%mm0             \n\t"

	"paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0    
	"movq %%mm7, %%mm3             \n\t"

	"paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
	"movq %%mm1, %%mm5             \n\t"

	"movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
	"psubw %%mm7, %%mm1            \n\t" //t13

	"movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
	"movq %%mm6, %%mm4             \n\t"

	"paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
	"paddw %%mm7, %%mm5            \n\t" //t10

	"paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
	"movq %%mm6, %%mm7             \n\t"

	"paddw %%mm2, %%mm6            \n\t" //t11    
	"psubw %%mm2, %%mm7            \n\t" //t12

	"movq %%mm5, %%mm2             \n\t"
	"paddw %%mm6, %%mm5            \n\t" //d0
	// i0 t13 t12 i3 i1 d0 - d4
	"psubw %%mm6, %%mm2            \n\t" //d4      
	"paddw %%mm1, %%mm7            \n\t"

	"movq  4*16(%%"REG_d"), %%mm6      \n\t"
	"psllw $2, %%mm7              \n\t"

	"psubw 0*16(%%"REG_d"), %%mm5      \n\t"
	"psubw %%mm6, %%mm2            \n\t"

	"paddusw 0*16(%%"REG_d"), %%mm5    \n\t"
	"paddusw %%mm6, %%mm2          \n\t"

	"pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
	//
	"paddw 0*16(%%"REG_d"), %%mm5      \n\t"
	"paddw %%mm6, %%mm2            \n\t"

	"psubusw 0*16(%%"REG_d"), %%mm5    \n\t"
	"psubusw %%mm6, %%mm2          \n\t"

//This func is totally compute-bound,  operates at huge speed. So,  DC shortcut
// at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
//However,  typical numbers: nondc - 29%%,  dc - 46%%,  zero - 25%%. All <> 0 case is very rare.
	"paddw "MANGLE(MM_2)", %%mm5            \n\t"
	"movq %%mm2, %%mm6             \n\t"

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

	"movq %%mm1, %%mm6             \n\t"
	"paddw %%mm7, %%mm1            \n\t" //d2

	"psubw 2*16(%%"REG_d"), %%mm1      \n\t"
	"psubw %%mm7, %%mm6            \n\t" //d6

	"movq 6*16(%%"REG_d"), %%mm7       \n\t"
	"psraw $2, %%mm5              \n\t"

	"paddusw 2*16(%%"REG_d"), %%mm1    \n\t"
	"psubw %%mm7, %%mm6            \n\t"
	// t7 d2 /t11 t4 t6 - d6 /t10     

	"paddw 2*16(%%"REG_d"), %%mm1      \n\t"
	"paddusw %%mm7, %%mm6          \n\t"

	"psubusw 2*16(%%"REG_d"), %%mm1    \n\t"
	"paddw %%mm7, %%mm6            \n\t"

	"psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
	"psubusw %%mm7, %%mm6          \n\t"

	//movq [edi+"DCTSIZE_S"*2*2], mm1
	//movq [edi+"DCTSIZE_S"*6*2], mm6     
	"movq %%mm1, %%mm7             \n\t"
	"psraw $2, %%mm2              \n\t"

	"psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
	"psubw %%mm6, %%mm1            \n\t"

	"psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
	"paddw %%mm7, %%mm6            \n\t" //'t13

	"psraw $2, %%mm6              \n\t" //paddw mm6, MM_2 !!    ---
	"movq %%mm2, %%mm7             \n\t"

	"pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
	"paddw %%mm6, %%mm2            \n\t" //'t0

	"movq %%mm2, "MANGLE(temps)"+0*8       \n\t" //!
	"psubw %%mm6, %%mm7            \n\t" //'t3

	"movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
	"psubw %%mm6, %%mm1            \n\t" //'t12        

	"psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
	"movq %%mm5, %%mm6             \n\t"

	"movq %%mm7, "MANGLE(temps)"+3*8       \n\t"
	"paddw %%mm2, %%mm3            \n\t" //t10

	"paddw %%mm4, %%mm2            \n\t" //t11
	"paddw %%mm0, %%mm4            \n\t" //t12

	"movq %%mm3, %%mm7             \n\t"
	"psubw %%mm4, %%mm3            \n\t"

	"psllw $2, %%mm3              \n\t"
	"psllw $2, %%mm7              \n\t" //opt for P6

	"pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
	"psllw $2, %%mm4              \n\t"

	"pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
	"psllw $2, %%mm2              \n\t"

	"pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
	"paddw %%mm1, %%mm5            \n\t" //'t1

	"pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
	"psubw %%mm1, %%mm6            \n\t" //'t2
	// t7 't12 't11 t4 t6 - 't13 't10   ---

	"paddw %%mm3, %%mm7            \n\t" //z2        

	"movq %%mm5, "MANGLE(temps)"+1*8       \n\t"
	"paddw %%mm3, %%mm4            \n\t" //z4

	"movq 3*16(%%"REG_d"), %%mm3       \n\t"
	"movq %%mm0, %%mm1             \n\t"

	"movq %%mm6, "MANGLE(temps)"+2*8       \n\t"
	"psubw %%mm2, %%mm1            \n\t" //z13            

//===
	"paddw %%mm2, %%mm0            \n\t" //z11 
	"movq %%mm1, %%mm5             \n\t"

	"movq 5*16(%%"REG_d"), %%mm2       \n\t"
	"psubw %%mm7, %%mm1            \n\t" //d3

	"paddw %%mm7, %%mm5            \n\t" //d5
	"psubw %%mm3, %%mm1            \n\t"

	"movq 1*16(%%"REG_d"), %%mm7       \n\t"
	"psubw %%mm2, %%mm5            \n\t"

	"movq %%mm0, %%mm6             \n\t"
	"paddw %%mm4, %%mm0            \n\t" //d1    

	"paddusw %%mm3, %%mm1          \n\t"
	"psubw %%mm4, %%mm6            \n\t" //d7  

	// d1 d3 - - - d5 d7 -    
	"movq 7*16(%%"REG_d"), %%mm4       \n\t"
	"psubw %%mm7, %%mm0            \n\t"

	"psubw %%mm4, %%mm6            \n\t"
	"paddusw %%mm2, %%mm5          \n\t"

	"paddusw %%mm4, %%mm6          \n\t"
	"paddw %%mm3, %%mm1            \n\t"

	"paddw %%mm2, %%mm5            \n\t"
	"paddw %%mm4, %%mm6            \n\t"

	"psubusw %%mm3, %%mm1          \n\t"
	"psubusw %%mm2, %%mm5          \n\t"

	"psubusw %%mm4, %%mm6          \n\t"
	"movq %%mm1, %%mm4             \n\t"

	"por %%mm5, %%mm4              \n\t"
	"paddusw %%mm7, %%mm0          \n\t"

	"por %%mm6, %%mm4              \n\t"
	"paddw %%mm7, %%mm0            \n\t"

	"packssdw %%mm4, %%mm4         \n\t"
	"psubusw %%mm7, %%mm0          \n\t"

	"movd %%mm4, %%"REG_a"             \n\t"
	"or %%"REG_a", %%"REG_a"              \n\t"
	"jnz 2f                 \n\t"
	//movq [edi+"DCTSIZE_S"*3*2], mm1
	//movq [edi+"DCTSIZE_S"*5*2], mm5
	//movq [edi+"DCTSIZE_S"*1*2], mm0
	//movq [edi+"DCTSIZE_S"*7*2], mm6
	// t4 t5 - - - t6 t7 -
	//--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
//Typical numbers: nondc - 19%%,  dc - 26%%,  zero - 55%%. zero case alone isn't worthwhile
	"movq "MANGLE(temps)"+0*8, %%mm4       \n\t"
	"movq %%mm0, %%mm1             \n\t"

	"pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
	"movq %%mm1, %%mm2             \n\t"

	"movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
	"movq %%mm2, %%mm3             \n\t"

	"pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
	"paddw %%mm4, %%mm5            \n\t"

	"movq "MANGLE(temps)"+1*8, %%mm6       \n\t"
	//paddw mm3, MM_2
	"psraw $2, %%mm3              \n\t" //tmp7     

	"pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
	"psubw %%mm3, %%mm4            \n\t"

	"movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
	"paddw %%mm3, %%mm5            \n\t"

	"movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
	"paddw %%mm6, %%mm7            \n\t"

	"movq "MANGLE(temps)"+2*8, %%mm3       \n\t"
	"psubw %%mm0, %%mm6            \n\t"

	"movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
	"paddw %%mm0, %%mm7            \n\t"

	"movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
	"paddw %%mm3, %%mm4            \n\t"

	"movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
	"psubw %%mm1, %%mm3            \n\t"

	"movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
	"paddw %%mm1, %%mm4            \n\t"

	"movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
	"paddw %%mm3, %%mm5            \n\t"

	"movq "MANGLE(temps)"+3*8, %%mm0       \n\t"
	"add $8, %%"REG_S"               \n\t"

	"movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
	"paddw %%mm0, %%mm6            \n\t"

	"movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
	"psubw %%mm2, %%mm0            \n\t"

	"movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
	"paddw %%mm2, %%mm6            \n\t"

	"movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
	"paddw %%mm0, %%mm7            \n\t"

	"movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"

	"movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
	"add $8, %%"REG_D"               \n\t"
	"jmp 4f                  \n\t"

	"2:                    \n\t"
	//--- non DC2
	//psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1  (actually thr1, thr1, thr1-1)
	//psraw mm5, 2              
	//psraw mm0, 2
	//psraw mm6, 2
	"movq %%mm5, %%mm3             \n\t"
	"psubw %%mm1, %%mm5            \n\t"

	"psllw $1, %%mm5              \n\t" //'z10
	"paddw %%mm1, %%mm3            \n\t" //'z13

	"movq %%mm0, %%mm2             \n\t"
	"psubw %%mm6, %%mm0            \n\t"

	"movq %%mm5, %%mm1             \n\t"
	"psllw $1, %%mm0              \n\t" //'z12

	"pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
	"paddw %%mm0, %%mm5            \n\t"

	"pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
	"paddw %%mm6, %%mm2            \n\t" //'z11

	"pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
	"movq %%mm2, %%mm7             \n\t"

	//---
	"movq "MANGLE(temps)"+0*8, %%mm4       \n\t"
	"psubw %%mm3, %%mm2            \n\t"

	"psllw $1, %%mm2              \n\t"
	"paddw %%mm3, %%mm7            \n\t" //'t7

	"pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
	"movq %%mm4, %%mm6             \n\t"
	//paddw mm7, MM_2
	"psraw $2, %%mm7              \n\t"

	"paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
	"psubw %%mm7, %%mm6            \n\t"

	"movq "MANGLE(temps)"+1*8, %%mm3       \n\t"
	"paddw %%mm7, %%mm4            \n\t"

	"movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
	"paddw %%mm5, %%mm1            \n\t" //'t12

	"movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
	"psubw %%mm7, %%mm1            \n\t" //'t6

	"movq "MANGLE(temps)"+2*8, %%mm7       \n\t"
	"psubw %%mm5, %%mm0            \n\t" //'t10

	"movq "MANGLE(temps)"+3*8, %%mm6       \n\t"
	"movq %%mm3, %%mm5             \n\t"

	"paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
	"psubw %%mm1, %%mm5            \n\t"

	"psubw %%mm1, %%mm2            \n\t" //'t5
	"paddw %%mm1, %%mm3            \n\t"

	"movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
	"movq %%mm7, %%mm4             \n\t"

	"paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
	"psubw %%mm2, %%mm4            \n\t"

	"paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
	"paddw %%mm2, %%mm7            \n\t"

	"movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
	"paddw %%mm2, %%mm0            \n\t" //'t4     

	// 't4 't6 't5 - - - - 't7
	"movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
	"movq %%mm6, %%mm1             \n\t"

	"paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
	"psubw %%mm0, %%mm1            \n\t"

	"paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
	"paddw %%mm0, %%mm6            \n\t"

	"movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
	"add $8, %%"REG_S"               \n\t"

	"movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"

	"movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
	"add $8, %%"REG_D"               \n\t"

	"4:                     \n\t"
//=part 2 (the same)===========================================================    
	"movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
	//
	"movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
	"movq %%mm1, %%mm0             \n\t"

	"paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0    
	"movq %%mm7, %%mm3             \n\t"

	"paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
	"movq %%mm1, %%mm5             \n\t"

	"movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
	"psubw %%mm7, %%mm1            \n\t" //t13

	"movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
	"movq %%mm6, %%mm4             \n\t"

	"paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
	"paddw %%mm7, %%mm5            \n\t" //t10

	"paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
	"movq %%mm6, %%mm7             \n\t"

	"paddw %%mm2, %%mm6            \n\t" //t11    
	"psubw %%mm2, %%mm7            \n\t" //t12

	"movq %%mm5, %%mm2             \n\t"
	"paddw %%mm6, %%mm5            \n\t" //d0
	// i0 t13 t12 i3 i1 d0 - d4
	"psubw %%mm6, %%mm2            \n\t" //d4      
	"paddw %%mm1, %%mm7            \n\t"