michael@0: /*
michael@0:  * jcdctmgr.c
michael@0:  *
michael@0:  * This file was part of the Independent JPEG Group's software:
michael@0:  * Copyright (C) 1994-1996, Thomas G. Lane.
michael@0:  * libjpeg-turbo Modifications:
michael@0:  * Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0:  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0:  * Copyright (C) 2011 D. R. Commander
michael@0:  * For conditions of distribution and use, see the accompanying README file.
michael@0:  *
michael@0:  * This file contains the forward-DCT management logic.
michael@0:  * This code selects a particular DCT implementation to be used,
michael@0:  * and it performs related housekeeping chores including coefficient
michael@0:  * quantization.
michael@0:  */
michael@0: 
michael@0: #define JPEG_INTERNALS
michael@0: #include "jinclude.h"
michael@0: #include "jpeglib.h"
michael@0: #include "jdct.h"		/* Private declarations for DCT subsystem */
michael@0: #include "jsimddct.h"
michael@0: 
michael@0: 
michael@0: /* Private subobject for this module */
michael@0: 
michael@0: typedef JMETHOD(void, forward_DCT_method_ptr, (DCTELEM * data));
michael@0: typedef JMETHOD(void, float_DCT_method_ptr, (FAST_FLOAT * data));
michael@0: 
michael@0: typedef JMETHOD(void, convsamp_method_ptr,
michael@0:                 (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0:                  DCTELEM * workspace));
michael@0: typedef JMETHOD(void, float_convsamp_method_ptr,
michael@0:                 (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0:                  FAST_FLOAT *workspace));
michael@0: 
michael@0: typedef JMETHOD(void, quantize_method_ptr,
michael@0:                 (JCOEFPTR coef_block, DCTELEM * divisors,
michael@0:                  DCTELEM * workspace));
michael@0: typedef JMETHOD(void, float_quantize_method_ptr,
michael@0:                 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
michael@0:                  FAST_FLOAT * workspace));
michael@0: 
michael@0: METHODDEF(void) quantize (JCOEFPTR, DCTELEM *, DCTELEM *);
michael@0: 
michael@0: typedef struct {
michael@0:   struct jpeg_forward_dct pub;	/* public fields */
michael@0: 
michael@0:   /* Pointer to the DCT routine actually in use */
michael@0:   forward_DCT_method_ptr dct;
michael@0:   convsamp_method_ptr convsamp;
michael@0:   quantize_method_ptr quantize;
michael@0: 
michael@0:   /* The actual post-DCT divisors --- not identical to the quant table
michael@0:    * entries, because of scaling (especially for an unnormalized DCT).
michael@0:    * Each table is given in normal array order.
michael@0:    */
michael@0:   DCTELEM * divisors[NUM_QUANT_TBLS];
michael@0: 
michael@0:   /* work area for FDCT subroutine */
michael@0:   DCTELEM * workspace;
michael@0: 
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:   /* Same as above for the floating-point case. */
michael@0:   float_DCT_method_ptr float_dct;
michael@0:   float_convsamp_method_ptr float_convsamp;
michael@0:   float_quantize_method_ptr float_quantize;
michael@0:   FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
michael@0:   FAST_FLOAT * float_workspace;
michael@0: #endif
michael@0: } my_fdct_controller;
michael@0: 
michael@0: typedef my_fdct_controller * my_fdct_ptr;
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Find the highest bit in an integer through binary search.
michael@0:  */
michael@0: LOCAL(int)
michael@0: flss (UINT16 val)
michael@0: {
michael@0:   int bit;
michael@0: 
michael@0:   bit = 16;
michael@0: 
michael@0:   if (!val)
michael@0:     return 0;
michael@0: 
michael@0:   if (!(val & 0xff00)) {
michael@0:     bit -= 8;
michael@0:     val <<= 8;
michael@0:   }
michael@0:   if (!(val & 0xf000)) {
michael@0:     bit -= 4;
michael@0:     val <<= 4;
michael@0:   }
michael@0:   if (!(val & 0xc000)) {
michael@0:     bit -= 2;
michael@0:     val <<= 2;
michael@0:   }
michael@0:   if (!(val & 0x8000)) {
michael@0:     bit -= 1;
michael@0:     val <<= 1;
michael@0:   }
michael@0: 
michael@0:   return bit;
michael@0: }
michael@0: 
michael@0: /*
michael@0:  * Compute values to do a division using reciprocal.
michael@0:  *
michael@0:  * This implementation is based on an algorithm described in
michael@0:  *   "How to optimize for the Pentium family of microprocessors"
michael@0:  *   (http://www.agner.org/assem/).
michael@0:  * More information about the basic algorithm can be found in
michael@0:  * the paper "Integer Division Using Reciprocals" by Robert Alverson.
michael@0:  *
michael@0:  * The basic idea is to replace x/d by x * d^-1. In order to store
michael@0:  * d^-1 with enough precision we shift it left a few places. It turns
michael@0:  * out that this algoright gives just enough precision, and also fits
michael@0:  * into DCTELEM:
michael@0:  *
michael@0:  *   b = (the number of significant bits in divisor) - 1
michael@0:  *   r = (word size) + b
michael@0:  *   f = 2^r / divisor
michael@0:  *
michael@0:  * f will not be an integer for most cases, so we need to compensate
michael@0:  * for the rounding error introduced:
michael@0:  *
michael@0:  *   no fractional part:
michael@0:  *
michael@0:  *       result = input >> r
michael@0:  *
michael@0:  *   fractional part of f < 0.5:
michael@0:  *
michael@0:  *       round f down to nearest integer
michael@0:  *       result = ((input + 1) * f) >> r
michael@0:  *
michael@0:  *   fractional part of f > 0.5:
michael@0:  *
michael@0:  *       round f up to nearest integer
michael@0:  *       result = (input * f) >> r
michael@0:  *
michael@0:  * This is the original algorithm that gives truncated results. But we
michael@0:  * want properly rounded results, so we replace "input" with
michael@0:  * "input + divisor/2".
michael@0:  *
michael@0:  * In order to allow SIMD implementations we also tweak the values to
michael@0:  * allow the same calculation to be made at all times:
michael@0:  * 
michael@0:  *   dctbl[0] = f rounded to nearest integer
michael@0:  *   dctbl[1] = divisor / 2 (+ 1 if fractional part of f < 0.5)
michael@0:  *   dctbl[2] = 1 << ((word size) * 2 - r)
michael@0:  *   dctbl[3] = r - (word size)
michael@0:  *
michael@0:  * dctbl[2] is for stupid instruction sets where the shift operation
michael@0:  * isn't member wise (e.g. MMX).
michael@0:  *
michael@0:  * The reason dctbl[2] and dctbl[3] reduce the shift with (word size)
michael@0:  * is that most SIMD implementations have a "multiply and store top
michael@0:  * half" operation.
michael@0:  *
michael@0:  * Lastly, we store each of the values in their own table instead
michael@0:  * of in a consecutive manner, yet again in order to allow SIMD
michael@0:  * routines.
michael@0:  */
michael@0: LOCAL(int)
michael@0: compute_reciprocal (UINT16 divisor, DCTELEM * dtbl)
michael@0: {
michael@0:   UDCTELEM2 fq, fr;
michael@0:   UDCTELEM c;
michael@0:   int b, r;
michael@0: 
michael@0:   b = flss(divisor) - 1;
michael@0:   r  = sizeof(DCTELEM) * 8 + b;
michael@0: 
michael@0:   fq = ((UDCTELEM2)1 << r) / divisor;
michael@0:   fr = ((UDCTELEM2)1 << r) % divisor;
michael@0: 
michael@0:   c = divisor / 2; /* for rounding */
michael@0: 
michael@0:   if (fr == 0) { /* divisor is power of two */
michael@0:     /* fq will be one bit too large to fit in DCTELEM, so adjust */
michael@0:     fq >>= 1;
michael@0:     r--;
michael@0:   } else if (fr <= (divisor / 2U)) { /* fractional part is < 0.5 */
michael@0:     c++;
michael@0:   } else { /* fractional part is > 0.5 */
michael@0:     fq++;
michael@0:   }
michael@0: 
michael@0:   dtbl[DCTSIZE2 * 0] = (DCTELEM) fq;      /* reciprocal */
michael@0:   dtbl[DCTSIZE2 * 1] = (DCTELEM) c;       /* correction + roundfactor */
michael@0:   dtbl[DCTSIZE2 * 2] = (DCTELEM) (1 << (sizeof(DCTELEM)*8*2 - r));  /* scale */
michael@0:   dtbl[DCTSIZE2 * 3] = (DCTELEM) r - sizeof(DCTELEM)*8; /* shift */
michael@0: 
michael@0:   if(r <= 16) return 0;
michael@0:   else return 1;
michael@0: }
michael@0: 
michael@0: /*
michael@0:  * Initialize for a processing pass.
michael@0:  * Verify that all referenced Q-tables are present, and set up
michael@0:  * the divisor table for each one.
michael@0:  * In the current implementation, DCT of all components is done during
michael@0:  * the first pass, even if only some components will be output in the
michael@0:  * first scan.  Hence all components should be examined here.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: start_pass_fdctmgr (j_compress_ptr cinfo)
michael@0: {
michael@0:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
michael@0:   int ci, qtblno, i;
michael@0:   jpeg_component_info *compptr;
michael@0:   JQUANT_TBL * qtbl;
michael@0:   DCTELEM * dtbl;
michael@0: 
michael@0:   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
michael@0:        ci++, compptr++) {
michael@0:     qtblno = compptr->quant_tbl_no;
michael@0:     /* Make sure specified quantization table is present */
michael@0:     if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
michael@0: 	cinfo->quant_tbl_ptrs[qtblno] == NULL)
michael@0:       ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
michael@0:     qtbl = cinfo->quant_tbl_ptrs[qtblno];
michael@0:     /* Compute divisors for this quant table */
michael@0:     /* We may do this more than once for same table, but it's not a big deal */
michael@0:     switch (cinfo->dct_method) {
michael@0: #ifdef DCT_ISLOW_SUPPORTED
michael@0:     case JDCT_ISLOW:
michael@0:       /* For LL&M IDCT method, divisors are equal to raw quantization
michael@0:        * coefficients multiplied by 8 (to counteract scaling).
michael@0:        */
michael@0:       if (fdct->divisors[qtblno] == NULL) {
michael@0: 	fdct->divisors[qtblno] = (DCTELEM *)
michael@0: 	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 				      (DCTSIZE2 * 4) * SIZEOF(DCTELEM));
michael@0:       }
michael@0:       dtbl = fdct->divisors[qtblno];
michael@0:       for (i = 0; i < DCTSIZE2; i++) {
michael@0: 	if(!compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i])
michael@0: 	  && fdct->quantize == jsimd_quantize)
michael@0: 	  fdct->quantize = quantize;
michael@0:       }
michael@0:       break;
michael@0: #endif
michael@0: #ifdef DCT_IFAST_SUPPORTED
michael@0:     case JDCT_IFAST:
michael@0:       {
michael@0: 	/* For AA&N IDCT method, divisors are equal to quantization
michael@0: 	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
michael@0: 	 *   scalefactor[0] = 1
michael@0: 	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
michael@0: 	 * We apply a further scale factor of 8.
michael@0: 	 */
michael@0: #define CONST_BITS 14
michael@0: 	static const INT16 aanscales[DCTSIZE2] = {
michael@0: 	  /* precomputed values scaled up by 14 bits */
michael@0: 	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
michael@0: 	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
michael@0: 	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
michael@0: 	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
michael@0: 	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
michael@0: 	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
michael@0: 	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
michael@0: 	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
michael@0: 	};
michael@0: 	SHIFT_TEMPS
michael@0: 
michael@0: 	if (fdct->divisors[qtblno] == NULL) {
michael@0: 	  fdct->divisors[qtblno] = (DCTELEM *)
michael@0: 	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 					(DCTSIZE2 * 4) * SIZEOF(DCTELEM));
michael@0: 	}
michael@0: 	dtbl = fdct->divisors[qtblno];
michael@0: 	for (i = 0; i < DCTSIZE2; i++) {
michael@0: 	  if(!compute_reciprocal(
michael@0: 	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
michael@0: 				  (INT32) aanscales[i]),
michael@0: 		    CONST_BITS-3), &dtbl[i])
michael@0: 	    && fdct->quantize == jsimd_quantize)
michael@0: 	    fdct->quantize = quantize;
michael@0: 	}
michael@0:       }
michael@0:       break;
michael@0: #endif
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:     case JDCT_FLOAT:
michael@0:       {
michael@0: 	/* For float AA&N IDCT method, divisors are equal to quantization
michael@0: 	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
michael@0: 	 *   scalefactor[0] = 1
michael@0: 	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
michael@0: 	 * We apply a further scale factor of 8.
michael@0: 	 * What's actually stored is 1/divisor so that the inner loop can
michael@0: 	 * use a multiplication rather than a division.
michael@0: 	 */
michael@0: 	FAST_FLOAT * fdtbl;
michael@0: 	int row, col;
michael@0: 	static const double aanscalefactor[DCTSIZE] = {
michael@0: 	  1.0, 1.387039845, 1.306562965, 1.175875602,
michael@0: 	  1.0, 0.785694958, 0.541196100, 0.275899379
michael@0: 	};
michael@0: 
michael@0: 	if (fdct->float_divisors[qtblno] == NULL) {
michael@0: 	  fdct->float_divisors[qtblno] = (FAST_FLOAT *)
michael@0: 	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 					DCTSIZE2 * SIZEOF(FAST_FLOAT));
michael@0: 	}
michael@0: 	fdtbl = fdct->float_divisors[qtblno];
michael@0: 	i = 0;
michael@0: 	for (row = 0; row < DCTSIZE; row++) {
michael@0: 	  for (col = 0; col < DCTSIZE; col++) {
michael@0: 	    fdtbl[i] = (FAST_FLOAT)
michael@0: 	      (1.0 / (((double) qtbl->quantval[i] *
michael@0: 		       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
michael@0: 	    i++;
michael@0: 	  }
michael@0: 	}
michael@0:       }
michael@0:       break;
michael@0: #endif
michael@0:     default:
michael@0:       ERREXIT(cinfo, JERR_NOT_COMPILED);
michael@0:       break;
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Load data into workspace, applying unsigned->signed conversion.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, DCTELEM * workspace)
michael@0: {
michael@0:   register DCTELEM *workspaceptr;
michael@0:   register JSAMPROW elemptr;
michael@0:   register int elemr;
michael@0: 
michael@0:   workspaceptr = workspace;
michael@0:   for (elemr = 0; elemr < DCTSIZE; elemr++) {
michael@0:     elemptr = sample_data[elemr] + start_col;
michael@0: 
michael@0: #if DCTSIZE == 8		/* unroll the inner loop */
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0: #else
michael@0:     {
michael@0:       register int elemc;
michael@0:       for (elemc = DCTSIZE; elemc > 0; elemc--)
michael@0:         *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
michael@0:     }
michael@0: #endif
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Quantize/descale the coefficients, and store into coef_blocks[].
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: quantize (JCOEFPTR coef_block, DCTELEM * divisors, DCTELEM * workspace)
michael@0: {
michael@0:   int i;
michael@0:   DCTELEM temp;
michael@0:   UDCTELEM recip, corr, shift;
michael@0:   UDCTELEM2 product;
michael@0:   JCOEFPTR output_ptr = coef_block;
michael@0: 
michael@0:   for (i = 0; i < DCTSIZE2; i++) {
michael@0:     temp = workspace[i];
michael@0:     recip = divisors[i + DCTSIZE2 * 0];
michael@0:     corr =  divisors[i + DCTSIZE2 * 1];
michael@0:     shift = divisors[i + DCTSIZE2 * 3];
michael@0: 
michael@0:     if (temp < 0) {
michael@0:       temp = -temp;
michael@0:       product = (UDCTELEM2)(temp + corr) * recip;
michael@0:       product >>= shift + sizeof(DCTELEM)*8;
michael@0:       temp = product;
michael@0:       temp = -temp;
michael@0:     } else {
michael@0:       product = (UDCTELEM2)(temp + corr) * recip;
michael@0:       product >>= shift + sizeof(DCTELEM)*8;
michael@0:       temp = product;
michael@0:     }
michael@0: 
michael@0:     output_ptr[i] = (JCOEF) temp;
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Perform forward DCT on one or more blocks of a component.
michael@0:  *
michael@0:  * The input samples are taken from the sample_data[] array starting at
michael@0:  * position start_row/start_col, and moving to the right for any additional
michael@0:  * blocks. The quantized coefficients are returned in coef_blocks[].
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 	     JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
michael@0: 	     JDIMENSION start_row, JDIMENSION start_col,
michael@0: 	     JDIMENSION num_blocks)
michael@0: /* This version is used for integer DCT implementations. */
michael@0: {
michael@0:   /* This routine is heavily used, so it's worth coding it tightly. */
michael@0:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
michael@0:   DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
michael@0:   DCTELEM * workspace;
michael@0:   JDIMENSION bi;
michael@0: 
michael@0:   /* Make sure the compiler doesn't look up these every pass */
michael@0:   forward_DCT_method_ptr do_dct = fdct->dct;
michael@0:   convsamp_method_ptr do_convsamp = fdct->convsamp;
michael@0:   quantize_method_ptr do_quantize = fdct->quantize;
michael@0:   workspace = fdct->workspace;
michael@0: 
michael@0:   sample_data += start_row;	/* fold in the vertical offset once */
michael@0: 
michael@0:   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
michael@0:     /* Load data into workspace, applying unsigned->signed conversion */
michael@0:     (*do_convsamp) (sample_data, start_col, workspace);
michael@0: 
michael@0:     /* Perform the DCT */
michael@0:     (*do_dct) (workspace);
michael@0: 
michael@0:     /* Quantize/descale the coefficients, and store into coef_blocks[] */
michael@0:     (*do_quantize) (coef_blocks[bi], divisors, workspace);
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0: 
michael@0: 
michael@0: METHODDEF(void)
michael@0: convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, FAST_FLOAT * workspace)
michael@0: {
michael@0:   register FAST_FLOAT *workspaceptr;
michael@0:   register JSAMPROW elemptr;
michael@0:   register int elemr;
michael@0: 
michael@0:   workspaceptr = workspace;
michael@0:   for (elemr = 0; elemr < DCTSIZE; elemr++) {
michael@0:     elemptr = sample_data[elemr] + start_col;
michael@0: #if DCTSIZE == 8		/* unroll the inner loop */
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0: #else
michael@0:     {
michael@0:       register int elemc;
michael@0:       for (elemc = DCTSIZE; elemc > 0; elemc--)
michael@0:         *workspaceptr++ = (FAST_FLOAT)
michael@0:                           (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
michael@0:     }
michael@0: #endif
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: METHODDEF(void)
michael@0: quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, FAST_FLOAT * workspace)
michael@0: {
michael@0:   register FAST_FLOAT temp;
michael@0:   register int i;
michael@0:   register JCOEFPTR output_ptr = coef_block;
michael@0: 
michael@0:   for (i = 0; i < DCTSIZE2; i++) {
michael@0:     /* Apply the quantization and scaling factor */
michael@0:     temp = workspace[i] * divisors[i];
michael@0: 
michael@0:     /* Round to nearest integer.
michael@0:      * Since C does not specify the direction of rounding for negative
michael@0:      * quotients, we have to force the dividend positive for portability.
michael@0:      * The maximum coefficient size is +-16K (for 12-bit data), so this
michael@0:      * code should work for either 16-bit or 32-bit ints.
michael@0:      */
michael@0:     output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: METHODDEF(void)
michael@0: forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
michael@0: 		   JDIMENSION start_row, JDIMENSION start_col,
michael@0: 		   JDIMENSION num_blocks)
michael@0: /* This version is used for floating-point DCT implementations. */
michael@0: {
michael@0:   /* This routine is heavily used, so it's worth coding it tightly. */
michael@0:   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
michael@0:   FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
michael@0:   FAST_FLOAT * workspace;
michael@0:   JDIMENSION bi;
michael@0: 
michael@0: 
michael@0:   /* Make sure the compiler doesn't look up these every pass */
michael@0:   float_DCT_method_ptr do_dct = fdct->float_dct;
michael@0:   float_convsamp_method_ptr do_convsamp = fdct->float_convsamp;
michael@0:   float_quantize_method_ptr do_quantize = fdct->float_quantize;
michael@0:   workspace = fdct->float_workspace;
michael@0: 
michael@0:   sample_data += start_row;	/* fold in the vertical offset once */
michael@0: 
michael@0:   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
michael@0:     /* Load data into workspace, applying unsigned->signed conversion */
michael@0:     (*do_convsamp) (sample_data, start_col, workspace);
michael@0: 
michael@0:     /* Perform the DCT */
michael@0:     (*do_dct) (workspace);
michael@0: 
michael@0:     /* Quantize/descale the coefficients, and store into coef_blocks[] */
michael@0:     (*do_quantize) (coef_blocks[bi], divisors, workspace);
michael@0:   }
michael@0: }
michael@0: 
michael@0: #endif /* DCT_FLOAT_SUPPORTED */
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Initialize FDCT manager.
michael@0:  */
michael@0: 
michael@0: GLOBAL(void)
michael@0: jinit_forward_dct (j_compress_ptr cinfo)
michael@0: {
michael@0:   my_fdct_ptr fdct;
michael@0:   int i;
michael@0: 
michael@0:   fdct = (my_fdct_ptr)
michael@0:     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 				SIZEOF(my_fdct_controller));
michael@0:   cinfo->fdct = (struct jpeg_forward_dct *) fdct;
michael@0:   fdct->pub.start_pass = start_pass_fdctmgr;
michael@0: 
michael@0:   /* First determine the DCT... */
michael@0:   switch (cinfo->dct_method) {
michael@0: #ifdef DCT_ISLOW_SUPPORTED
michael@0:   case JDCT_ISLOW:
michael@0:     fdct->pub.forward_DCT = forward_DCT;
michael@0:     if (jsimd_can_fdct_islow())
michael@0:       fdct->dct = jsimd_fdct_islow;
michael@0:     else
michael@0:       fdct->dct = jpeg_fdct_islow;
michael@0:     break;
michael@0: #endif
michael@0: #ifdef DCT_IFAST_SUPPORTED
michael@0:   case JDCT_IFAST:
michael@0:     fdct->pub.forward_DCT = forward_DCT;
michael@0:     if (jsimd_can_fdct_ifast())
michael@0:       fdct->dct = jsimd_fdct_ifast;
michael@0:     else
michael@0:       fdct->dct = jpeg_fdct_ifast;
michael@0:     break;
michael@0: #endif
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:   case JDCT_FLOAT:
michael@0:     fdct->pub.forward_DCT = forward_DCT_float;
michael@0:     if (jsimd_can_fdct_float())
michael@0:       fdct->float_dct = jsimd_fdct_float;
michael@0:     else
michael@0:       fdct->float_dct = jpeg_fdct_float;
michael@0:     break;
michael@0: #endif
michael@0:   default:
michael@0:     ERREXIT(cinfo, JERR_NOT_COMPILED);
michael@0:     break;
michael@0:   }
michael@0: 
michael@0:   /* ...then the supporting stages. */
michael@0:   switch (cinfo->dct_method) {
michael@0: #ifdef DCT_ISLOW_SUPPORTED
michael@0:   case JDCT_ISLOW:
michael@0: #endif
michael@0: #ifdef DCT_IFAST_SUPPORTED
michael@0:   case JDCT_IFAST:
michael@0: #endif
michael@0: #if defined(DCT_ISLOW_SUPPORTED) || defined(DCT_IFAST_SUPPORTED)
michael@0:     if (jsimd_can_convsamp())
michael@0:       fdct->convsamp = jsimd_convsamp;
michael@0:     else
michael@0:       fdct->convsamp = convsamp;
michael@0:     if (jsimd_can_quantize())
michael@0:       fdct->quantize = jsimd_quantize;
michael@0:     else
michael@0:       fdct->quantize = quantize;
michael@0:     break;
michael@0: #endif
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:   case JDCT_FLOAT:
michael@0:     if (jsimd_can_convsamp_float())
michael@0:       fdct->float_convsamp = jsimd_convsamp_float;
michael@0:     else
michael@0:       fdct->float_convsamp = convsamp_float;
michael@0:     if (jsimd_can_quantize_float())
michael@0:       fdct->float_quantize = jsimd_quantize_float;
michael@0:     else
michael@0:       fdct->float_quantize = quantize_float;
michael@0:     break;
michael@0: #endif
michael@0:   default:
michael@0:     ERREXIT(cinfo, JERR_NOT_COMPILED);
michael@0:     break;
michael@0:   }
michael@0: 
michael@0:   /* Allocate workspace memory */
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:   if (cinfo->dct_method == JDCT_FLOAT)
michael@0:     fdct->float_workspace = (FAST_FLOAT *)
michael@0:       (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 				  SIZEOF(FAST_FLOAT) * DCTSIZE2);
michael@0:   else
michael@0: #endif
michael@0:     fdct->workspace = (DCTELEM *)
michael@0:       (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 				  SIZEOF(DCTELEM) * DCTSIZE2);
michael@0: 
michael@0:   /* Mark divisor tables unallocated */
michael@0:   for (i = 0; i < NUM_QUANT_TBLS; i++) {
michael@0:     fdct->divisors[i] = NULL;
michael@0: #ifdef DCT_FLOAT_SUPPORTED
michael@0:     fdct->float_divisors[i] = NULL;
michael@0: #endif
michael@0:   }
michael@0: }