jcdctmgr.c

jpeg编解码器

{
  /* This routine is heavily used, so it's worth coding it tightly. */
  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
  forward_DCT_method_ptr do_dct = fdct->do_dct;
  DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
  DCTELEM workspace[DCTSIZE2];	/* work area for FDCT subroutine */
  JDIMENSION bi;

  sample_data += start_row;	/* fold in the vertical offset once */

  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
    /* Load data into workspace, applying unsigned->signed conversion */
    { register DCTELEM *workspaceptr;
      register JSAMPROW elemptr;
      register int elemr;

      workspaceptr = workspace;
      for (elemr = 0; elemr < DCTSIZE; elemr++) 
	  {
		  elemptr = sample_data[elemr] + start_col;
#if DCTSIZE == 8		/* unroll the inner loop */
		  workspaceptr[0] = GETJSAMPLE(elemptr[0]) - CENTERJSAMPLE;
		  workspaceptr[1] = GETJSAMPLE(elemptr[1]) - CENTERJSAMPLE;
		  workspaceptr[2] = GETJSAMPLE(elemptr[2]) - CENTERJSAMPLE;
		  workspaceptr[3] = GETJSAMPLE(elemptr[3]) - CENTERJSAMPLE;
		  workspaceptr[4] = GETJSAMPLE(elemptr[4]) - CENTERJSAMPLE;
		  workspaceptr[5] = GETJSAMPLE(elemptr[5]) - CENTERJSAMPLE;
		  workspaceptr[6] = GETJSAMPLE(elemptr[6]) - CENTERJSAMPLE;
		  workspaceptr[7] = GETJSAMPLE(elemptr[7]) - CENTERJSAMPLE;
		  workspaceptr += 8; elemptr += 8;
#else
		  { 
			  register int elemc;
			  for (elemc = DCTSIZE; elemc > 0; elemc--) {
				  *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
			  }
		  }
#endif
      }
    }

    /* Perform the DCT */
    (*do_dct) (workspace);

	jcquant_int( workspace, coef_blocks[bi], 
				 fdct->divisors[compptr->quant_tbl_no] );
  }
}


#if DCTSIZE == 8 && defined(HAVE_MMX_ATT_MNEMONICS)	
METHODDEF(void)
	forward_DCT_x86float32 (j_compress_ptr cinfo, jpeg_component_info * compptr,
							JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
							JDIMENSION start_row, JDIMENSION start_col,
							JDIMENSION num_blocks)
/* This version is used for integer DCT implementations. */
{
	/* This routine is heavily used, so it's worth coding it tightly. */
	my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
	forward_DCT_method_ptr do_dct = fdct->do_dct;
	float32_quant_method_ptr do_quant = fdct->do_float32_quant;
	DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
	DCTELEM workspace[DCTSIZE2] ATTR_ALIGN(SIMD_ALIGN);	/* work area for FDCT subroutine */
	JDIMENSION bi;

	sample_data += start_row;	/* fold in the vertical offset once */

	for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) 
	{
		register DCTELEM *workspaceptr;
		register JSAMPROW elemptr;
		register int elemr;

		workspaceptr = workspace;
		elemr = 0;
		elemptr = sample_data[elemr] + start_col;
	  
		pxor_r2r(mm7,mm7);
		movq_m2r( *(mmx_t*)&int16_centrejsample, mm6 );
		while(elemr < DCTSIZE) 
		{
			movq_m2r( *(mmx_t*)elemptr, mm0 );
			elemptr = sample_data[elemr+1] + start_col;
			movq_m2r( *(mmx_t*)elemptr, mm1 );

			movq_r2r( mm0, mm2 );
			punpcklbw_r2r( mm7, mm0 );
			movq_r2r( mm1, mm3 );
			punpcklbw_r2r( mm7, mm1 );
			psubw_r2r( mm6, mm0 );
			psubw_r2r( mm6, mm1 );
			elemr += 2;
			punpckhbw_r2r( mm7, mm2 );
			punpckhbw_r2r( mm7, mm3 );
			elemptr = sample_data[elemr] + start_col;
			psubw_r2r( mm6, mm2 );
			psubw_r2r( mm6, mm3 );

			movq_r2m( mm0, *(mmx_t*)(&workspaceptr[0]) );
			movq_r2m( mm2, *(mmx_t*)(&workspaceptr[4]) );
			movq_r2m( mm1, *(mmx_t*)(&workspaceptr[8]) );
			movq_r2m( mm3, *(mmx_t*)(&workspaceptr[12]) );
			workspaceptr += 16;

		}
		emms();

		/* Perform the DCT */
		(*do_dct)(workspace);
		
		(*do_quant)( workspace, coef_blocks[bi], 
					 fdct->float32_divisors[compptr->quant_tbl_no] );
	}
}


METHODDEF(void)
	forward_DCT_mmx (j_compress_ptr cinfo, jpeg_component_info * compptr,
						  JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
						  JDIMENSION start_row, JDIMENSION start_col,
						  JDIMENSION num_blocks)
/* This version is used for integer DCT implementations. */
{
	/* This routine is heavily used, so it's worth coding it tightly. */
	my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
	forward_DCT_method_ptr do_dct = fdct->do_dct;
	float32_quant_method_ptr do_quant = fdct->do_float32_quant;
	DCTELEM workspace[DCTSIZE2] ATTR_ALIGN(SIMD_ALIGN);	/* work area for FDCT subroutine */
	JDIMENSION bi;
	int i;
	sample_data += start_row;	/* fold in the vertical offset once */

	for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) 
	{
		register DCTELEM *workspaceptr;
		register JSAMPROW elemptr;
		register int elemr;

		workspaceptr = workspace;
		elemr = 0;
		elemptr = sample_data[elemr] + start_col;
	  
		pxor_r2r(mm7,mm7);
		movq_m2r( *(mmx_t*)&int16_centrejsample, mm6 );
		while(elemr < DCTSIZE) 
		{
			movq_m2r( *(mmx_t*)elemptr, mm0 );
			elemptr = sample_data[elemr+1] + start_col;
			movq_m2r( *(mmx_t*)elemptr, mm1 );

			movq_r2r( mm0, mm2 );
			punpcklbw_r2r( mm7, mm0 );
			movq_r2r( mm1, mm3 );
			punpcklbw_r2r( mm7, mm1 );
			psubw_r2r( mm6, mm0 );
			psubw_r2r( mm6, mm1 );
			elemr += 2;
			punpckhbw_r2r( mm7, mm2 );
			punpckhbw_r2r( mm7, mm3 );
			elemptr = sample_data[elemr] + start_col;
			psubw_r2r( mm6, mm2 );
			psubw_r2r( mm6, mm3 );

			movq_r2m( mm0, *(mmx_t*)(&workspaceptr[0]) );
			movq_r2m( mm2, *(mmx_t*)(&workspaceptr[4]) );
			movq_r2m( mm1, *(mmx_t*)(&workspaceptr[8]) );
			movq_r2m( mm3, *(mmx_t*)(&workspaceptr[12]) );
			workspaceptr += 16;

		}
		emms();

		/* Perform the DCT */
		(*do_dct)(workspace);
		
		jcquant_mmx( workspace,
					 coef_blocks[bi],  
					 fdct->divisors[compptr->quant_tbl_no],
					 fdct->int16_divisors[compptr->quant_tbl_no],
					 fdct->shift
			);

	}
}

#endif
#ifdef DCT_FLOAT_SUPPORTED

METHODDEF(void)
forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
		   JDIMENSION start_row, JDIMENSION start_col,
		   JDIMENSION num_blocks)
/* This version is used for floating-point DCT implementations. */
{
  /* This routine is heavily used, so it's worth coding it tightly. */
  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
  float_DCT_method_ptr do_dct = fdct->do_float_dct;
  FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
  FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
  JDIMENSION bi;

  sample_data += start_row;	/* fold in the vertical offset once */

  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
    /* Load data into workspace, applying unsigned->signed conversion */
    { register FAST_FLOAT *workspaceptr;
      register JSAMPROW elemptr;
      register int elemr;

      workspaceptr = workspace;
      for (elemr = 0; elemr < DCTSIZE; elemr++) {
	elemptr = sample_data[elemr] + start_col;
#if DCTSIZE == 8		/* unroll the inner loop */
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
#else
	{ register int elemc;
	  for (elemc = DCTSIZE; elemc > 0; elemc--) {
	    *workspaceptr++ = (FAST_FLOAT)
	      (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
	  }
	}
#endif
      }
    }

    /* Perform the DCT */
    (*do_dct) (workspace);

    /* Quantize/descale the coefficients, and store into coef_blocks[] */
    { register FAST_FLOAT temp;
      register int i;
      register JCOEFPTR output_ptr = coef_blocks[bi];

      for (i = 0; i < DCTSIZE2; i++) {
	/* Apply the quantization and scaling factor */
	temp = workspace[i] * divisors[i];
	/* Round to nearest integer.
	 * Since C does not specify the direction of rounding for negative
	 * quotients, we have to force the dividend positive for portability.
	 * The maximum coefficient size is +-16K (for 12-bit data), so this
	 * code should work for either 16-bit or 32-bit ints.
	 */
	output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
      }
    }
  }
}

#endif /* DCT_FLOAT_SUPPORTED */


/*
 * Initialize FDCT manager.
 */

GLOBAL(void)
jinit_forward_dct (j_compress_ptr cinfo)
{
  my_fdct_ptr fdct;
  int i;

  fdct = (my_fdct_ptr)
    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
				SIZEOF(my_fdct_controller));
  cinfo->fdct = (struct jpeg_forward_dct *) fdct;
  fdct->pub.start_pass = start_pass_fdctmgr;

  switch (cinfo->dct_method) {
#ifdef DCT_ISLOW_SUPPORTED
  case JDCT_ISLOW:
    fdct->pub.forward_DCT = forward_DCT;
    fdct->do_dct = jpeg_fdct_islow;
    break;
#endif
#ifdef DCT_IFAST_SUPPORTED
  case JDCT_IFAST:

  {
#if defined(HAVE_MMX_ATT_MNEMONICS)	  
	  int cpu_flags =  cpu_accel();
	  if( cpu_flags & ACCEL_X86_MMX )
	  {

		  fdct->do_dct = jpeg_fdct_ifast_mmx;
		  if( cpu_flags & ACCEL_X86_SSE )
		  {
			  fdct->fast_quantiser = QUANT_FLOAT32;
			  fdct->pub.forward_DCT = forward_DCT_x86float32;
			  fdct->do_float32_quant = jcquant_sse;
		  }
		  else if( cpu_flags & ACCEL_X86_3DNOW )
		  {
			  fdct->fast_quantiser = QUANT_FLOAT32;
			  fdct->pub.forward_DCT = forward_DCT_x86float32;
			  fdct->do_float32_quant = jcquant_3dnow;
		  }
		  else
		  {
			  fdct->fast_quantiser = QUANT_INT16;
			  fdct->pub.forward_DCT = forward_DCT_mmx;
			  fdct->do_float32_quant = NULL;
		  }
	  }
	  else
#endif
	  {
		  fdct->fast_quantiser = QUANT_INT;
		  fdct->pub.forward_DCT = forward_DCT;
		  fdct->do_dct = jpeg_fdct_ifast;
		  fdct->do_float32_quant = NULL;
	  }

	  break;
  }

#endif
#ifdef DCT_FLOAT_SUPPORTED
  case JDCT_FLOAT:
    fdct->pub.forward_DCT = forward_DCT_float;
    fdct->do_float_dct = jpeg_fdct_float;
    break;
#endif
  default:
    ERREXIT(cinfo, JERR_NOT_COMPILED);
    break;
  }

  /* Mark divisor tables unallocated */
  for (i = 0; i < NUM_QUANT_TBLS; i++) {
    fdct->divisors[i] = NULL;
    fdct->int16_divisors[i] = NULL;
    fdct->float32_divisors[i] = NULL;
#ifdef DCT_FLOAT_SUPPORTED
    fdct->float_divisors[i] = NULL;
#endif
  }
}