nuclear@26: /*
nuclear@26:  * jcdctmgr.c
nuclear@26:  *
nuclear@26:  * Copyright (C) 1994-1996, Thomas G. Lane.
nuclear@26:  * This file is part of the Independent JPEG Group's software.
nuclear@26:  * For conditions of distribution and use, see the accompanying README file.
nuclear@26:  *
nuclear@26:  * This file contains the forward-DCT management logic.
nuclear@26:  * This code selects a particular DCT implementation to be used,
nuclear@26:  * and it performs related housekeeping chores including coefficient
nuclear@26:  * quantization.
nuclear@26:  */
nuclear@26: 
nuclear@26: #define JPEG_INTERNALS
nuclear@26: #include "jinclude.h"
nuclear@26: #include "jpeglib.h"
nuclear@26: #include "jdct.h"		/* Private declarations for DCT subsystem */
nuclear@26: 
nuclear@26: 
nuclear@26: /* Private subobject for this module */
nuclear@26: 
nuclear@26: typedef struct {
nuclear@26:   struct jpeg_forward_dct pub;	/* public fields */
nuclear@26: 
nuclear@26:   /* Pointer to the DCT routine actually in use */
nuclear@26:   forward_DCT_method_ptr do_dct;
nuclear@26: 
nuclear@26:   /* The actual post-DCT divisors --- not identical to the quant table
nuclear@26:    * entries, because of scaling (especially for an unnormalized DCT).
nuclear@26:    * Each table is given in normal array order.
nuclear@26:    */
nuclear@26:   DCTELEM * divisors[NUM_QUANT_TBLS];
nuclear@26: 
nuclear@26: #ifdef DCT_FLOAT_SUPPORTED
nuclear@26:   /* Same as above for the floating-point case. */
nuclear@26:   float_DCT_method_ptr do_float_dct;
nuclear@26:   FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
nuclear@26: #endif
nuclear@26: } my_fdct_controller;
nuclear@26: 
nuclear@26: typedef my_fdct_controller * my_fdct_ptr;
nuclear@26: 
nuclear@26: 
nuclear@26: /*
nuclear@26:  * Initialize for a processing pass.
nuclear@26:  * Verify that all referenced Q-tables are present, and set up
nuclear@26:  * the divisor table for each one.
nuclear@26:  * In the current implementation, DCT of all components is done during
nuclear@26:  * the first pass, even if only some components will be output in the
nuclear@26:  * first scan.  Hence all components should be examined here.
nuclear@26:  */
nuclear@26: 
nuclear@26: METHODDEF(void)
nuclear@26: start_pass_fdctmgr (j_compress_ptr cinfo)
nuclear@26: {
nuclear@26:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
nuclear@26:   int ci, qtblno, i;
nuclear@26:   jpeg_component_info *compptr;
nuclear@26:   JQUANT_TBL * qtbl;
nuclear@26:   DCTELEM * dtbl;
nuclear@26: 
nuclear@26:   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
nuclear@26:        ci++, compptr++) {
nuclear@26:     qtblno = compptr->quant_tbl_no;
nuclear@26:     /* Make sure specified quantization table is present */
nuclear@26:     if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
nuclear@26: 	cinfo->quant_tbl_ptrs[qtblno] == NULL)
nuclear@26:       ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
nuclear@26:     qtbl = cinfo->quant_tbl_ptrs[qtblno];
nuclear@26:     /* Compute divisors for this quant table */
nuclear@26:     /* We may do this more than once for same table, but it's not a big deal */
nuclear@26:     switch (cinfo->dct_method) {
nuclear@26: #ifdef DCT_ISLOW_SUPPORTED
nuclear@26:     case JDCT_ISLOW:
nuclear@26:       /* For LL&M IDCT method, divisors are equal to raw quantization
nuclear@26:        * coefficients multiplied by 8 (to counteract scaling).
nuclear@26:        */
nuclear@26:       if (fdct->divisors[qtblno] == NULL) {
nuclear@26: 	fdct->divisors[qtblno] = (DCTELEM *)
nuclear@26: 	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
nuclear@26: 				      DCTSIZE2 * SIZEOF(DCTELEM));
nuclear@26:       }
nuclear@26:       dtbl = fdct->divisors[qtblno];
nuclear@26:       for (i = 0; i < DCTSIZE2; i++) {
nuclear@26: 	dtbl[i] = ((DCTELEM) qtbl->quantval[i]) << 3;
nuclear@26:       }
nuclear@26:       break;
nuclear@26: #endif
nuclear@26: #ifdef DCT_IFAST_SUPPORTED
nuclear@26:     case JDCT_IFAST:
nuclear@26:       {
nuclear@26: 	/* For AA&N IDCT method, divisors are equal to quantization
nuclear@26: 	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
nuclear@26: 	 *   scalefactor[0] = 1
nuclear@26: 	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
nuclear@26: 	 * We apply a further scale factor of 8.
nuclear@26: 	 */
nuclear@26: #define CONST_BITS 14
nuclear@26: 	static const INT16 aanscales[DCTSIZE2] = {
nuclear@26: 	  /* precomputed values scaled up by 14 bits */
nuclear@26: 	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
nuclear@26: 	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
nuclear@26: 	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
nuclear@26: 	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
nuclear@26: 	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
nuclear@26: 	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
nuclear@26: 	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
nuclear@26: 	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
nuclear@26: 	};
nuclear@26: 	SHIFT_TEMPS
nuclear@26: 
nuclear@26: 	if (fdct->divisors[qtblno] == NULL) {
nuclear@26: 	  fdct->divisors[qtblno] = (DCTELEM *)
nuclear@26: 	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
nuclear@26: 					DCTSIZE2 * SIZEOF(DCTELEM));
nuclear@26: 	}
nuclear@26: 	dtbl = fdct->divisors[qtblno];
nuclear@26: 	for (i = 0; i < DCTSIZE2; i++) {
nuclear@26: 	  dtbl[i] = (DCTELEM)
nuclear@26: 	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
nuclear@26: 				  (INT32) aanscales[i]),
nuclear@26: 		    CONST_BITS-3);
nuclear@26: 	}
nuclear@26:       }
nuclear@26:       break;
nuclear@26: #endif
nuclear@26: #ifdef DCT_FLOAT_SUPPORTED
nuclear@26:     case JDCT_FLOAT:
nuclear@26:       {
nuclear@26: 	/* For float AA&N IDCT method, divisors are equal to quantization
nuclear@26: 	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
nuclear@26: 	 *   scalefactor[0] = 1
nuclear@26: 	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
nuclear@26: 	 * We apply a further scale factor of 8.
nuclear@26: 	 * What's actually stored is 1/divisor so that the inner loop can
nuclear@26: 	 * use a multiplication rather than a division.
nuclear@26: 	 */
nuclear@26: 	FAST_FLOAT * fdtbl;
nuclear@26: 	int row, col;
nuclear@26: 	static const double aanscalefactor[DCTSIZE] = {
nuclear@26: 	  1.0, 1.387039845, 1.306562965, 1.175875602,
nuclear@26: 	  1.0, 0.785694958, 0.541196100, 0.275899379
nuclear@26: 	};
nuclear@26: 
nuclear@26: 	if (fdct->float_divisors[qtblno] == NULL) {
nuclear@26: 	  fdct->float_divisors[qtblno] = (FAST_FLOAT *)
nuclear@26: 	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
nuclear@26: 					DCTSIZE2 * SIZEOF(FAST_FLOAT));
nuclear@26: 	}
nuclear@26: 	fdtbl = fdct->float_divisors[qtblno];
nuclear@26: 	i = 0;
nuclear@26: 	for (row = 0; row < DCTSIZE; row++) {
nuclear@26: 	  for (col = 0; col < DCTSIZE; col++) {
nuclear@26: 	    fdtbl[i] = (FAST_FLOAT)
nuclear@26: 	      (1.0 / (((double) qtbl->quantval[i] *
nuclear@26: 		       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
nuclear@26: 	    i++;
nuclear@26: 	  }
nuclear@26: 	}
nuclear@26:       }
nuclear@26:       break;
nuclear@26: #endif
nuclear@26:     default:
nuclear@26:       ERREXIT(cinfo, JERR_NOT_COMPILED);
nuclear@26:       break;
nuclear@26:     }
nuclear@26:   }
nuclear@26: }
nuclear@26: 
nuclear@26: 
nuclear@26: /*
nuclear@26:  * Perform forward DCT on one or more blocks of a component.
nuclear@26:  *
nuclear@26:  * The input samples are taken from the sample_data[] array starting at
nuclear@26:  * position start_row/start_col, and moving to the right for any additional
nuclear@26:  * blocks. The quantized coefficients are returned in coef_blocks[].
nuclear@26:  */
nuclear@26: 
nuclear@26: METHODDEF(void)
nuclear@26: forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
nuclear@26: 	     JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
nuclear@26: 	     JDIMENSION start_row, JDIMENSION start_col,
nuclear@26: 	     JDIMENSION num_blocks)
nuclear@26: /* This version is used for integer DCT implementations. */
nuclear@26: {
nuclear@26:   /* This routine is heavily used, so it's worth coding it tightly. */
nuclear@26:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
nuclear@26:   forward_DCT_method_ptr do_dct = fdct->do_dct;
nuclear@26:   DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
nuclear@26:   DCTELEM workspace[DCTSIZE2];	/* work area for FDCT subroutine */
nuclear@26:   JDIMENSION bi;
nuclear@26: 
nuclear@26:   sample_data += start_row;	/* fold in the vertical offset once */
nuclear@26: 
nuclear@26:   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
nuclear@26:     /* Load data into workspace, applying unsigned->signed conversion */
nuclear@26:     { register DCTELEM *workspaceptr;
nuclear@26:       register JSAMPROW elemptr;
nuclear@26:       register int elemr;
nuclear@26: 
nuclear@26:       workspaceptr = workspace;
nuclear@26:       for (elemr = 0; elemr < DCTSIZE; elemr++) {
nuclear@26: 	elemptr = sample_data[elemr] + start_col;
nuclear@26: #if DCTSIZE == 8		/* unroll the inner loop */
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: #else
nuclear@26: 	{ register int elemc;
nuclear@26: 	  for (elemc = DCTSIZE; elemc > 0; elemc--) {
nuclear@26: 	    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
nuclear@26: 	  }
nuclear@26: 	}
nuclear@26: #endif
nuclear@26:       }
nuclear@26:     }
nuclear@26: 
nuclear@26:     /* Perform the DCT */
nuclear@26:     (*do_dct) (workspace);
nuclear@26: 
nuclear@26:     /* Quantize/descale the coefficients, and store into coef_blocks[] */
nuclear@26:     { register DCTELEM temp, qval;
nuclear@26:       register int i;
nuclear@26:       register JCOEFPTR output_ptr = coef_blocks[bi];
nuclear@26: 
nuclear@26:       for (i = 0; i < DCTSIZE2; i++) {
nuclear@26: 	qval = divisors[i];
nuclear@26: 	temp = workspace[i];
nuclear@26: 	/* Divide the coefficient value by qval, ensuring proper rounding.
nuclear@26: 	 * Since C does not specify the direction of rounding for negative
nuclear@26: 	 * quotients, we have to force the dividend positive for portability.
nuclear@26: 	 *
nuclear@26: 	 * In most files, at least half of the output values will be zero
nuclear@26: 	 * (at default quantization settings, more like three-quarters...)
nuclear@26: 	 * so we should ensure that this case is fast.  On many machines,
nuclear@26: 	 * a comparison is enough cheaper than a divide to make a special test
nuclear@26: 	 * a win.  Since both inputs will be nonnegative, we need only test
nuclear@26: 	 * for a < b to discover whether a/b is 0.
nuclear@26: 	 * If your machine's division is fast enough, define FAST_DIVIDE.
nuclear@26: 	 */
nuclear@26: #ifdef FAST_DIVIDE
nuclear@26: #define DIVIDE_BY(a,b)	a /= b
nuclear@26: #else
nuclear@26: #define DIVIDE_BY(a,b)	if (a >= b) a /= b; else a = 0
nuclear@26: #endif
nuclear@26: 	if (temp < 0) {
nuclear@26: 	  temp = -temp;
nuclear@26: 	  temp += qval>>1;	/* for rounding */
nuclear@26: 	  DIVIDE_BY(temp, qval);
nuclear@26: 	  temp = -temp;
nuclear@26: 	} else {
nuclear@26: 	  temp += qval>>1;	/* for rounding */
nuclear@26: 	  DIVIDE_BY(temp, qval);
nuclear@26: 	}
nuclear@26: 	output_ptr[i] = (JCOEF) temp;
nuclear@26:       }
nuclear@26:     }
nuclear@26:   }
nuclear@26: }
nuclear@26: 
nuclear@26: 
nuclear@26: #ifdef DCT_FLOAT_SUPPORTED
nuclear@26: 
nuclear@26: METHODDEF(void)
nuclear@26: forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
nuclear@26: 		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
nuclear@26: 		   JDIMENSION start_row, JDIMENSION start_col,
nuclear@26: 		   JDIMENSION num_blocks)
nuclear@26: /* This version is used for floating-point DCT implementations. */
nuclear@26: {
nuclear@26:   /* This routine is heavily used, so it's worth coding it tightly. */
nuclear@26:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
nuclear@26:   float_DCT_method_ptr do_dct = fdct->do_float_dct;
nuclear@26:   FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
nuclear@26:   FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
nuclear@26:   JDIMENSION bi;
nuclear@26: 
nuclear@26:   sample_data += start_row;	/* fold in the vertical offset once */
nuclear@26: 
nuclear@26:   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
nuclear@26:     /* Load data into workspace, applying unsigned->signed conversion */
nuclear@26:     { register FAST_FLOAT *workspaceptr;
nuclear@26:       register JSAMPROW elemptr;
nuclear@26:       register int elemr;
nuclear@26: 
nuclear@26:       workspaceptr = workspace;
nuclear@26:       for (elemr = 0; elemr < DCTSIZE; elemr++) {
nuclear@26: 	elemptr = sample_data[elemr] + start_col;
nuclear@26: #if DCTSIZE == 8		/* unroll the inner loop */
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: #else
nuclear@26: 	{ register int elemc;
nuclear@26: 	  for (elemc = DCTSIZE; elemc > 0; elemc--) {
nuclear@26: 	    *workspaceptr++ = (FAST_FLOAT)
nuclear@26: 	      (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
nuclear@26: 	  }
nuclear@26: 	}
nuclear@26: #endif
nuclear@26:       }
nuclear@26:     }
nuclear@26: 
nuclear@26:     /* Perform the DCT */
nuclear@26:     (*do_dct) (workspace);
nuclear@26: 
nuclear@26:     /* Quantize/descale the coefficients, and store into coef_blocks[] */
nuclear@26:     { register FAST_FLOAT temp;
nuclear@26:       register int i;
nuclear@26:       register JCOEFPTR output_ptr = coef_blocks[bi];
nuclear@26: 
nuclear@26:       for (i = 0; i < DCTSIZE2; i++) {
nuclear@26: 	/* Apply the quantization and scaling factor */
nuclear@26: 	temp = workspace[i] * divisors[i];
nuclear@26: 	/* Round to nearest integer.
nuclear@26: 	 * Since C does not specify the direction of rounding for negative
nuclear@26: 	 * quotients, we have to force the dividend positive for portability.
nuclear@26: 	 * The maximum coefficient size is +-16K (for 12-bit data), so this
nuclear@26: 	 * code should work for either 16-bit or 32-bit ints.
nuclear@26: 	 */
nuclear@26: 	output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
nuclear@26:       }
nuclear@26:     }
nuclear@26:   }
nuclear@26: }
nuclear@26: 
nuclear@26: #endif /* DCT_FLOAT_SUPPORTED */
nuclear@26: 
nuclear@26: 
nuclear@26: /*
nuclear@26:  * Initialize FDCT manager.
nuclear@26:  */
nuclear@26: 
nuclear@26: GLOBAL(void)
nuclear@26: jinit_forward_dct (j_compress_ptr cinfo)
nuclear@26: {
nuclear@26:   my_fdct_ptr fdct;
nuclear@26:   int i;
nuclear@26: 
nuclear@26:   fdct = (my_fdct_ptr)
nuclear@26:     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
nuclear@26: 				SIZEOF(my_fdct_controller));
nuclear@26:   cinfo->fdct = (struct jpeg_forward_dct *) fdct;
nuclear@26:   fdct->pub.start_pass = start_pass_fdctmgr;
nuclear@26: 
nuclear@26:   switch (cinfo->dct_method) {
nuclear@26: #ifdef DCT_ISLOW_SUPPORTED
nuclear@26:   case JDCT_ISLOW:
nuclear@26:     fdct->pub.forward_DCT = forward_DCT;
nuclear@26:     fdct->do_dct = jpeg_fdct_islow;
nuclear@26:     break;
nuclear@26: #endif
nuclear@26: #ifdef DCT_IFAST_SUPPORTED
nuclear@26:   case JDCT_IFAST:
nuclear@26:     fdct->pub.forward_DCT = forward_DCT;
nuclear@26:     fdct->do_dct = jpeg_fdct_ifast;
nuclear@26:     break;
nuclear@26: #endif
nuclear@26: #ifdef DCT_FLOAT_SUPPORTED
nuclear@26:   case JDCT_FLOAT:
nuclear@26:     fdct->pub.forward_DCT = forward_DCT_float;
nuclear@26:     fdct->do_float_dct = jpeg_fdct_float;
nuclear@26:     break;
nuclear@26: #endif
nuclear@26:   default:
nuclear@26:     ERREXIT(cinfo, JERR_NOT_COMPILED);
nuclear@26:     break;
nuclear@26:   }
nuclear@26: 
nuclear@26:   /* Mark divisor tables unallocated */
nuclear@26:   for (i = 0; i < NUM_QUANT_TBLS; i++) {
nuclear@26:     fdct->divisors[i] = NULL;
nuclear@26: #ifdef DCT_FLOAT_SUPPORTED
nuclear@26:     fdct->float_divisors[i] = NULL;
nuclear@26: #endif
nuclear@26:   }
nuclear@26: }