Home | History | Annotate | Download | only in libjpeg_turbo
      1 /*
      2  * jdcoefct.c
      3  *
      4  * This file was part of the Independent JPEG Group's software:
      5  * Copyright (C) 1994-1997, Thomas G. Lane.
      6  * libjpeg-turbo Modifications:
      7  * Copyright (C) 2010, D. R. Commander.
      8  * For conditions of distribution and use, see the accompanying README file.
      9  *
     10  * This file contains the coefficient buffer controller for decompression.
     11  * This controller is the top level of the JPEG decompressor proper.
     12  * The coefficient buffer lies between entropy decoding and inverse-DCT steps.
     13  *
     14  * In buffered-image mode, this controller is the interface between
     15  * input-oriented processing and output-oriented processing.
     16  * Also, the input side (only) is used when reading a file for transcoding.
     17  */
     18 
     19 #define JPEG_INTERNALS
     20 #include "jinclude.h"
     21 #include "jpeglib.h"
     22 #include "jpegcomp.h"
     23 
     24 /* Block smoothing is only applicable for progressive JPEG, so: */
     25 #ifndef D_PROGRESSIVE_SUPPORTED
     26 #undef BLOCK_SMOOTHING_SUPPORTED
     27 #endif
     28 
     29 /* Private buffer controller object */
     30 
     31 typedef struct {
     32   struct jpeg_d_coef_controller pub; /* public fields */
     33 
     34   /* These variables keep track of the current location of the input side. */
     35   /* cinfo->input_iMCU_row is also used for this. */
     36   JDIMENSION MCU_ctr;		/* counts MCUs processed in current row */
     37   int MCU_vert_offset;		/* counts MCU rows within iMCU row */
     38   int MCU_rows_per_iMCU_row;	/* number of such rows needed */
     39 
     40   /* The output side's location is represented by cinfo->output_iMCU_row. */
     41 
     42   /* In single-pass modes, it's sufficient to buffer just one MCU.
     43    * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
     44    * and let the entropy decoder write into that workspace each time.
     45    * (On 80x86, the workspace is FAR even though it's not really very big;
     46    * this is to keep the module interfaces unchanged when a large coefficient
     47    * buffer is necessary.)
     48    * In multi-pass modes, this array points to the current MCU's blocks
     49    * within the virtual arrays; it is used only by the input side.
     50    */
     51   JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
     52 
     53   /* Temporary workspace for one MCU */
     54   JCOEF * workspace;
     55 
     56 #ifdef D_MULTISCAN_FILES_SUPPORTED
     57   /* In multi-pass modes, we need a virtual block array for each component. */
     58   jvirt_barray_ptr whole_image[MAX_COMPONENTS];
     59 #endif
     60 
     61 #ifdef BLOCK_SMOOTHING_SUPPORTED
     62   /* When doing block smoothing, we latch coefficient Al values here */
     63   int * coef_bits_latch;
     64 #define SAVED_COEFS  6		/* we save coef_bits[0..5] */
     65 #endif
     66 } my_coef_controller;
     67 
     68 typedef my_coef_controller * my_coef_ptr;
     69 
     70 /* Forward declarations */
     71 METHODDEF(int) decompress_onepass
     72 	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
     73 #ifdef D_MULTISCAN_FILES_SUPPORTED
     74 METHODDEF(int) decompress_data
     75 	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
     76 #endif
     77 #ifdef BLOCK_SMOOTHING_SUPPORTED
     78 LOCAL(boolean) smoothing_ok JPP((j_decompress_ptr cinfo));
     79 METHODDEF(int) decompress_smooth_data
     80 	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
     81 #endif
     82 
     83 
     84 LOCAL(void)
     85 start_iMCU_row (j_decompress_ptr cinfo)
     86 /* Reset within-iMCU-row counters for a new row (input side) */
     87 {
     88   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
     89 
     90   /* In an interleaved scan, an MCU row is the same as an iMCU row.
     91    * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
     92    * But at the bottom of the image, process only what's left.
     93    */
     94   if (cinfo->comps_in_scan > 1) {
     95     coef->MCU_rows_per_iMCU_row = 1;
     96   } else {
     97     if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows-1))
     98       coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
     99     else
    100       coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
    101   }
    102 
    103   coef->MCU_ctr = 0;
    104   coef->MCU_vert_offset = 0;
    105 }
    106 
    107 
    108 /*
    109  * Initialize for an input processing pass.
    110  */
    111 
    112 METHODDEF(void)
    113 start_input_pass (j_decompress_ptr cinfo)
    114 {
    115   cinfo->input_iMCU_row = 0;
    116   start_iMCU_row(cinfo);
    117 }
    118 
    119 
    120 /*
    121  * Initialize for an output processing pass.
    122  */
    123 
    124 METHODDEF(void)
    125 start_output_pass (j_decompress_ptr cinfo)
    126 {
    127 #ifdef BLOCK_SMOOTHING_SUPPORTED
    128   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    129 
    130   /* If multipass, check to see whether to use block smoothing on this pass */
    131   if (coef->pub.coef_arrays != NULL) {
    132     if (cinfo->do_block_smoothing && smoothing_ok(cinfo))
    133       coef->pub.decompress_data = decompress_smooth_data;
    134     else
    135       coef->pub.decompress_data = decompress_data;
    136   }
    137 #endif
    138   cinfo->output_iMCU_row = 0;
    139 }
    140 
    141 
    142 /*
    143  * Decompress and return some data in the single-pass case.
    144  * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
    145  * Input and output must run in lockstep since we have only a one-MCU buffer.
    146  * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
    147  *
    148  * NB: output_buf contains a plane for each component in image,
    149  * which we index according to the component's SOF position.
    150  */
    151 
    152 METHODDEF(int)
    153 decompress_onepass (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
    154 {
    155   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    156   JDIMENSION MCU_col_num;	/* index of current MCU within row */
    157   JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
    158   JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
    159   int blkn, ci, xindex, yindex, yoffset, useful_width;
    160   JSAMPARRAY output_ptr;
    161   JDIMENSION start_col, output_col;
    162   jpeg_component_info *compptr;
    163   inverse_DCT_method_ptr inverse_DCT;
    164 
    165   /* Loop to process as much as one whole iMCU row */
    166   for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
    167        yoffset++) {
    168     for (MCU_col_num = coef->MCU_ctr; MCU_col_num <= last_MCU_col;
    169 	 MCU_col_num++) {
    170       /* Try to fetch an MCU.  Entropy decoder expects buffer to be zeroed. */
    171       jzero_far((void FAR *) coef->MCU_buffer[0],
    172 		(size_t) (cinfo->blocks_in_MCU * SIZEOF(JBLOCK)));
    173       if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
    174 	/* Suspension forced; update state counters and exit */
    175 	coef->MCU_vert_offset = yoffset;
    176 	coef->MCU_ctr = MCU_col_num;
    177 	return JPEG_SUSPENDED;
    178       }
    179       /* Determine where data should go in output_buf and do the IDCT thing.
    180        * We skip dummy blocks at the right and bottom edges (but blkn gets
    181        * incremented past them!).  Note the inner loop relies on having
    182        * allocated the MCU_buffer[] blocks sequentially.
    183        */
    184       blkn = 0;			/* index of current DCT block within MCU */
    185       for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    186 	compptr = cinfo->cur_comp_info[ci];
    187 	/* Don't bother to IDCT an uninteresting component. */
    188 	if (! compptr->component_needed) {
    189 	  blkn += compptr->MCU_blocks;
    190 	  continue;
    191 	}
    192 	inverse_DCT = cinfo->idct->inverse_DCT[compptr->component_index];
    193 	useful_width = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
    194 						    : compptr->last_col_width;
    195 	output_ptr = output_buf[compptr->component_index] +
    196 	  yoffset * compptr->_DCT_scaled_size;
    197 	start_col = MCU_col_num * compptr->MCU_sample_width;
    198 	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
    199 	  if (cinfo->input_iMCU_row < last_iMCU_row ||
    200 	      yoffset+yindex < compptr->last_row_height) {
    201 	    output_col = start_col;
    202 	    for (xindex = 0; xindex < useful_width; xindex++) {
    203 	      (*inverse_DCT) (cinfo, compptr,
    204 			      (JCOEFPTR) coef->MCU_buffer[blkn+xindex],
    205 			      output_ptr, output_col);
    206 	      output_col += compptr->_DCT_scaled_size;
    207 	    }
    208 	  }
    209 	  blkn += compptr->MCU_width;
    210 	  output_ptr += compptr->_DCT_scaled_size;
    211 	}
    212       }
    213     }
    214     /* Completed an MCU row, but perhaps not an iMCU row */
    215     coef->MCU_ctr = 0;
    216   }
    217   /* Completed the iMCU row, advance counters for next one */
    218   cinfo->output_iMCU_row++;
    219   if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
    220     start_iMCU_row(cinfo);
    221     return JPEG_ROW_COMPLETED;
    222   }
    223   /* Completed the scan */
    224   (*cinfo->inputctl->finish_input_pass) (cinfo);
    225   return JPEG_SCAN_COMPLETED;
    226 }
    227 
    228 
    229 /*
    230  * Dummy consume-input routine for single-pass operation.
    231  */
    232 
    233 METHODDEF(int)
    234 dummy_consume_data (j_decompress_ptr cinfo)
    235 {
    236   return JPEG_SUSPENDED;	/* Always indicate nothing was done */
    237 }
    238 
    239 
    240 #ifdef D_MULTISCAN_FILES_SUPPORTED
    241 
    242 /*
    243  * Consume input data and store it in the full-image coefficient buffer.
    244  * We read as much as one fully interleaved MCU row ("iMCU" row) per call,
    245  * ie, v_samp_factor block rows for each component in the scan.
    246  * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
    247  */
    248 
    249 METHODDEF(int)
    250 consume_data (j_decompress_ptr cinfo)
    251 {
    252   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    253   JDIMENSION MCU_col_num;	/* index of current MCU within row */
    254   int blkn, ci, xindex, yindex, yoffset;
    255   JDIMENSION start_col;
    256   JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
    257   JBLOCKROW buffer_ptr;
    258   jpeg_component_info *compptr;
    259 
    260   /* Align the virtual buffers for the components used in this scan. */
    261   for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    262     compptr = cinfo->cur_comp_info[ci];
    263     buffer[ci] = (*cinfo->mem->access_virt_barray)
    264       ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
    265        cinfo->input_iMCU_row * compptr->v_samp_factor,
    266        (JDIMENSION) compptr->v_samp_factor, TRUE);
    267     /* Note: entropy decoder expects buffer to be zeroed,
    268      * but this is handled automatically by the memory manager
    269      * because we requested a pre-zeroed array.
    270      */
    271   }
    272 
    273   /* Loop to process one whole iMCU row */
    274   for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
    275        yoffset++) {
    276     for (MCU_col_num = coef->MCU_ctr; MCU_col_num < cinfo->MCUs_per_row;
    277 	 MCU_col_num++) {
    278       /* Construct list of pointers to DCT blocks belonging to this MCU */
    279       blkn = 0;			/* index of current DCT block within MCU */
    280       for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    281 	compptr = cinfo->cur_comp_info[ci];
    282 	start_col = MCU_col_num * compptr->MCU_width;
    283 	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
    284 	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
    285 	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
    286 	    coef->MCU_buffer[blkn++] = buffer_ptr++;
    287 	  }
    288 	}
    289       }
    290       /* Try to fetch the MCU. */
    291       if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
    292 	/* Suspension forced; update state counters and exit */
    293 	coef->MCU_vert_offset = yoffset;
    294 	coef->MCU_ctr = MCU_col_num;
    295 	return JPEG_SUSPENDED;
    296       }
    297     }
    298     /* Completed an MCU row, but perhaps not an iMCU row */
    299     coef->MCU_ctr = 0;
    300   }
    301   /* Completed the iMCU row, advance counters for next one */
    302   if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
    303     start_iMCU_row(cinfo);
    304     return JPEG_ROW_COMPLETED;
    305   }
    306   /* Completed the scan */
    307   (*cinfo->inputctl->finish_input_pass) (cinfo);
    308   return JPEG_SCAN_COMPLETED;
    309 }
    310 
    311 
    312 /*
    313  * Decompress and return some data in the multi-pass case.
    314  * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
    315  * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
    316  *
    317  * NB: output_buf contains a plane for each component in image.
    318  */
    319 
    320 METHODDEF(int)
    321 decompress_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
    322 {
    323   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    324   JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
    325   JDIMENSION block_num;
    326   int ci, block_row, block_rows;
    327   JBLOCKARRAY buffer;
    328   JBLOCKROW buffer_ptr;
    329   JSAMPARRAY output_ptr;
    330   JDIMENSION output_col;
    331   jpeg_component_info *compptr;
    332   inverse_DCT_method_ptr inverse_DCT;
    333 
    334   /* Force some input to be done if we are getting ahead of the input. */
    335   while (cinfo->input_scan_number < cinfo->output_scan_number ||
    336 	 (cinfo->input_scan_number == cinfo->output_scan_number &&
    337 	  cinfo->input_iMCU_row <= cinfo->output_iMCU_row)) {
    338     if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
    339       return JPEG_SUSPENDED;
    340   }
    341 
    342   /* OK, output from the virtual arrays. */
    343   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
    344        ci++, compptr++) {
    345     /* Don't bother to IDCT an uninteresting component. */
    346     if (! compptr->component_needed)
    347       continue;
    348     /* Align the virtual buffer for this component. */
    349     buffer = (*cinfo->mem->access_virt_barray)
    350       ((j_common_ptr) cinfo, coef->whole_image[ci],
    351        cinfo->output_iMCU_row * compptr->v_samp_factor,
    352        (JDIMENSION) compptr->v_samp_factor, FALSE);
    353     /* Count non-dummy DCT block rows in this iMCU row. */
    354     if (cinfo->output_iMCU_row < last_iMCU_row)
    355       block_rows = compptr->v_samp_factor;
    356     else {
    357       /* NB: can't use last_row_height here; it is input-side-dependent! */
    358       block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
    359       if (block_rows == 0) block_rows = compptr->v_samp_factor;
    360     }
    361     inverse_DCT = cinfo->idct->inverse_DCT[ci];
    362     output_ptr = output_buf[ci];
    363     /* Loop over all DCT blocks to be processed. */
    364     for (block_row = 0; block_row < block_rows; block_row++) {
    365       buffer_ptr = buffer[block_row];
    366       output_col = 0;
    367       for (block_num = 0; block_num < compptr->width_in_blocks; block_num++) {
    368 	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) buffer_ptr,
    369 			output_ptr, output_col);
    370 	buffer_ptr++;
    371 	output_col += compptr->_DCT_scaled_size;
    372       }
    373       output_ptr += compptr->_DCT_scaled_size;
    374     }
    375   }
    376 
    377   if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
    378     return JPEG_ROW_COMPLETED;
    379   return JPEG_SCAN_COMPLETED;
    380 }
    381 
    382 #endif /* D_MULTISCAN_FILES_SUPPORTED */
    383 
    384 
    385 #ifdef BLOCK_SMOOTHING_SUPPORTED
    386 
    387 /*
    388  * This code applies interblock smoothing as described by section K.8
    389  * of the JPEG standard: the first 5 AC coefficients are estimated from
    390  * the DC values of a DCT block and its 8 neighboring blocks.
    391  * We apply smoothing only for progressive JPEG decoding, and only if
    392  * the coefficients it can estimate are not yet known to full precision.
    393  */
    394 
    395 /* Natural-order array positions of the first 5 zigzag-order coefficients */
    396 #define Q01_POS  1
    397 #define Q10_POS  8
    398 #define Q20_POS  16
    399 #define Q11_POS  9
    400 #define Q02_POS  2
    401 
    402 /*
    403  * Determine whether block smoothing is applicable and safe.
    404  * We also latch the current states of the coef_bits[] entries for the
    405  * AC coefficients; otherwise, if the input side of the decompressor
    406  * advances into a new scan, we might think the coefficients are known
    407  * more accurately than they really are.
    408  */
    409 
    410 LOCAL(boolean)
    411 smoothing_ok (j_decompress_ptr cinfo)
    412 {
    413   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    414   boolean smoothing_useful = FALSE;
    415   int ci, coefi;
    416   jpeg_component_info *compptr;
    417   JQUANT_TBL * qtable;
    418   int * coef_bits;
    419   int * coef_bits_latch;
    420 
    421   if (! cinfo->progressive_mode || cinfo->coef_bits == NULL)
    422     return FALSE;
    423 
    424   /* Allocate latch area if not already done */
    425   if (coef->coef_bits_latch == NULL)
    426     coef->coef_bits_latch = (int *)
    427       (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
    428 				  cinfo->num_components *
    429 				  (SAVED_COEFS * SIZEOF(int)));
    430   coef_bits_latch = coef->coef_bits_latch;
    431 
    432   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
    433        ci++, compptr++) {
    434     /* All components' quantization values must already be latched. */
    435     if ((qtable = compptr->quant_table) == NULL)
    436       return FALSE;
    437     /* Verify DC & first 5 AC quantizers are nonzero to avoid zero-divide. */
    438     if (qtable->quantval[0] == 0 ||
    439 	qtable->quantval[Q01_POS] == 0 ||
    440 	qtable->quantval[Q10_POS] == 0 ||
    441 	qtable->quantval[Q20_POS] == 0 ||
    442 	qtable->quantval[Q11_POS] == 0 ||
    443 	qtable->quantval[Q02_POS] == 0)
    444       return FALSE;
    445     /* DC values must be at least partly known for all components. */
    446     coef_bits = cinfo->coef_bits[ci];
    447     if (coef_bits[0] < 0)
    448       return FALSE;
    449     /* Block smoothing is helpful if some AC coefficients remain inaccurate. */
    450     for (coefi = 1; coefi <= 5; coefi++) {
    451       coef_bits_latch[coefi] = coef_bits[coefi];
    452       if (coef_bits[coefi] != 0)
    453 	smoothing_useful = TRUE;
    454     }
    455     coef_bits_latch += SAVED_COEFS;
    456   }
    457 
    458   return smoothing_useful;
    459 }
    460 
    461 
    462 /*
    463  * Variant of decompress_data for use when doing block smoothing.
    464  */
    465 
    466 METHODDEF(int)
    467 decompress_smooth_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
    468 {
    469   my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
    470   JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
    471   JDIMENSION block_num, last_block_column;
    472   int ci, block_row, block_rows, access_rows;
    473   JBLOCKARRAY buffer;
    474   JBLOCKROW buffer_ptr, prev_block_row, next_block_row;
    475   JSAMPARRAY output_ptr;
    476   JDIMENSION output_col;
    477   jpeg_component_info *compptr;
    478   inverse_DCT_method_ptr inverse_DCT;
    479   boolean first_row, last_row;
    480   JCOEF * workspace;
    481   int *coef_bits;
    482   JQUANT_TBL *quanttbl;
    483   INT32 Q00,Q01,Q02,Q10,Q11,Q20, num;
    484   int DC1,DC2,DC3,DC4,DC5,DC6,DC7,DC8,DC9;
    485   int Al, pred;
    486 
    487   /* Keep a local variable to avoid looking it up more than once */
    488   workspace = coef->workspace;
    489 
    490   /* Force some input to be done if we are getting ahead of the input. */
    491   while (cinfo->input_scan_number <= cinfo->output_scan_number &&
    492 	 ! cinfo->inputctl->eoi_reached) {
    493     if (cinfo->input_scan_number == cinfo->output_scan_number) {
    494       /* If input is working on current scan, we ordinarily want it to
    495        * have completed the current row.  But if input scan is DC,
    496        * we want it to keep one row ahead so that next block row's DC
    497        * values are up to date.
    498        */
    499       JDIMENSION delta = (cinfo->Ss == 0) ? 1 : 0;
    500       if (cinfo->input_iMCU_row > cinfo->output_iMCU_row+delta)
    501 	break;
    502     }
    503     if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
    504       return JPEG_SUSPENDED;
    505   }
    506 
    507   /* OK, output from the virtual arrays. */
    508   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
    509        ci++, compptr++) {
    510     /* Don't bother to IDCT an uninteresting component. */
    511     if (! compptr->component_needed)
    512       continue;
    513     /* Count non-dummy DCT block rows in this iMCU row. */
    514     if (cinfo->output_iMCU_row < last_iMCU_row) {
    515       block_rows = compptr->v_samp_factor;
    516       access_rows = block_rows * 2; /* this and next iMCU row */
    517       last_row = FALSE;
    518     } else {
    519       /* NB: can't use last_row_height here; it is input-side-dependent! */
    520       block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
    521       if (block_rows == 0) block_rows = compptr->v_samp_factor;
    522       access_rows = block_rows; /* this iMCU row only */
    523       last_row = TRUE;
    524     }
    525     /* Align the virtual buffer for this component. */
    526     if (cinfo->output_iMCU_row > 0) {
    527       access_rows += compptr->v_samp_factor; /* prior iMCU row too */
    528       buffer = (*cinfo->mem->access_virt_barray)
    529 	((j_common_ptr) cinfo, coef->whole_image[ci],
    530 	 (cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
    531 	 (JDIMENSION) access_rows, FALSE);
    532       buffer += compptr->v_samp_factor;	/* point to current iMCU row */
    533       first_row = FALSE;
    534     } else {
    535       buffer = (*cinfo->mem->access_virt_barray)
    536 	((j_common_ptr) cinfo, coef->whole_image[ci],
    537 	 (JDIMENSION) 0, (JDIMENSION) access_rows, FALSE);
    538       first_row = TRUE;
    539     }
    540     /* Fetch component-dependent info */
    541     coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
    542     quanttbl = compptr->quant_table;
    543     Q00 = quanttbl->quantval[0];
    544     Q01 = quanttbl->quantval[Q01_POS];
    545     Q10 = quanttbl->quantval[Q10_POS];
    546     Q20 = quanttbl->quantval[Q20_POS];
    547     Q11 = quanttbl->quantval[Q11_POS];
    548     Q02 = quanttbl->quantval[Q02_POS];
    549     inverse_DCT = cinfo->idct->inverse_DCT[ci];
    550     output_ptr = output_buf[ci];
    551     /* Loop over all DCT blocks to be processed. */
    552     for (block_row = 0; block_row < block_rows; block_row++) {
    553       buffer_ptr = buffer[block_row];
    554       if (first_row && block_row == 0)
    555 	prev_block_row = buffer_ptr;
    556       else
    557 	prev_block_row = buffer[block_row-1];
    558       if (last_row && block_row == block_rows-1)
    559 	next_block_row = buffer_ptr;
    560       else
    561 	next_block_row = buffer[block_row+1];
    562       /* We fetch the surrounding DC values using a sliding-register approach.
    563        * Initialize all nine here so as to do the right thing on narrow pics.
    564        */
    565       DC1 = DC2 = DC3 = (int) prev_block_row[0][0];
    566       DC4 = DC5 = DC6 = (int) buffer_ptr[0][0];
    567       DC7 = DC8 = DC9 = (int) next_block_row[0][0];
    568       output_col = 0;
    569       last_block_column = compptr->width_in_blocks - 1;
    570       for (block_num = 0; block_num <= last_block_column; block_num++) {
    571 	/* Fetch current DCT block into workspace so we can modify it. */
    572 	jcopy_block_row(buffer_ptr, (JBLOCKROW) workspace, (JDIMENSION) 1);
    573 	/* Update DC values */
    574 	if (block_num < last_block_column) {
    575 	  DC3 = (int) prev_block_row[1][0];
    576 	  DC6 = (int) buffer_ptr[1][0];
    577 	  DC9 = (int) next_block_row[1][0];
    578 	}
    579 	/* Compute coefficient estimates per K.8.
    580 	 * An estimate is applied only if coefficient is still zero,
    581 	 * and is not known to be fully accurate.
    582 	 */
    583 	/* AC01 */
    584 	if ((Al=coef_bits[1]) != 0 && workspace[1] == 0) {
    585 	  num = 36 * Q00 * (DC4 - DC6);
    586 	  if (num >= 0) {
    587 	    pred = (int) (((Q01<<7) + num) / (Q01<<8));
    588 	    if (Al > 0 && pred >= (1<<Al))
    589 	      pred = (1<<Al)-1;
    590 	  } else {
    591 	    pred = (int) (((Q01<<7) - num) / (Q01<<8));
    592 	    if (Al > 0 && pred >= (1<<Al))
    593 	      pred = (1<<Al)-1;
    594 	    pred = -pred;
    595 	  }
    596 	  workspace[1] = (JCOEF) pred;
    597 	}
    598 	/* AC10 */
    599 	if ((Al=coef_bits[2]) != 0 && workspace[8] == 0) {
    600 	  num = 36 * Q00 * (DC2 - DC8);
    601 	  if (num >= 0) {
    602 	    pred = (int) (((Q10<<7) + num) / (Q10<<8));
    603 	    if (Al > 0 && pred >= (1<<Al))
    604 	      pred = (1<<Al)-1;
    605 	  } else {
    606 	    pred = (int) (((Q10<<7) - num) / (Q10<<8));
    607 	    if (Al > 0 && pred >= (1<<Al))
    608 	      pred = (1<<Al)-1;
    609 	    pred = -pred;
    610 	  }
    611 	  workspace[8] = (JCOEF) pred;
    612 	}
    613 	/* AC20 */
    614 	if ((Al=coef_bits[3]) != 0 && workspace[16] == 0) {
    615 	  num = 9 * Q00 * (DC2 + DC8 - 2*DC5);
    616 	  if (num >= 0) {
    617 	    pred = (int) (((Q20<<7) + num) / (Q20<<8));
    618 	    if (Al > 0 && pred >= (1<<Al))
    619 	      pred = (1<<Al)-1;
    620 	  } else {
    621 	    pred = (int) (((Q20<<7) - num) / (Q20<<8));
    622 	    if (Al > 0 && pred >= (1<<Al))
    623 	      pred = (1<<Al)-1;
    624 	    pred = -pred;
    625 	  }
    626 	  workspace[16] = (JCOEF) pred;
    627 	}
    628 	/* AC11 */
    629 	if ((Al=coef_bits[4]) != 0 && workspace[9] == 0) {
    630 	  num = 5 * Q00 * (DC1 - DC3 - DC7 + DC9);
    631 	  if (num >= 0) {
    632 	    pred = (int) (((Q11<<7) + num) / (Q11<<8));
    633 	    if (Al > 0 && pred >= (1<<Al))
    634 	      pred = (1<<Al)-1;
    635 	  } else {
    636 	    pred = (int) (((Q11<<7) - num) / (Q11<<8));
    637 	    if (Al > 0 && pred >= (1<<Al))
    638 	      pred = (1<<Al)-1;
    639 	    pred = -pred;
    640 	  }
    641 	  workspace[9] = (JCOEF) pred;
    642 	}
    643 	/* AC02 */
    644 	if ((Al=coef_bits[5]) != 0 && workspace[2] == 0) {
    645 	  num = 9 * Q00 * (DC4 + DC6 - 2*DC5);
    646 	  if (num >= 0) {
    647 	    pred = (int) (((Q02<<7) + num) / (Q02<<8));
    648 	    if (Al > 0 && pred >= (1<<Al))
    649 	      pred = (1<<Al)-1;
    650 	  } else {
    651 	    pred = (int) (((Q02<<7) - num) / (Q02<<8));
    652 	    if (Al > 0 && pred >= (1<<Al))
    653 	      pred = (1<<Al)-1;
    654 	    pred = -pred;
    655 	  }
    656 	  workspace[2] = (JCOEF) pred;
    657 	}
    658 	/* OK, do the IDCT */
    659 	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) workspace,
    660 			output_ptr, output_col);
    661 	/* Advance for next column */
    662 	DC1 = DC2; DC2 = DC3;
    663 	DC4 = DC5; DC5 = DC6;
    664 	DC7 = DC8; DC8 = DC9;
    665 	buffer_ptr++, prev_block_row++, next_block_row++;
    666 	output_col += compptr->_DCT_scaled_size;
    667       }
    668       output_ptr += compptr->_DCT_scaled_size;
    669     }
    670   }
    671 
    672   if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
    673     return JPEG_ROW_COMPLETED;
    674   return JPEG_SCAN_COMPLETED;
    675 }
    676 
    677 #endif /* BLOCK_SMOOTHING_SUPPORTED */
    678 
    679 
    680 /*
    681  * Initialize coefficient buffer controller.
    682  */
    683 
    684 GLOBAL(void)
    685 jinit_d_coef_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
    686 {
    687   my_coef_ptr coef;
    688 
    689   coef = (my_coef_ptr)
    690     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
    691 				SIZEOF(my_coef_controller));
    692   cinfo->coef = (struct jpeg_d_coef_controller *) coef;
    693   coef->pub.start_input_pass = start_input_pass;
    694   coef->pub.start_output_pass = start_output_pass;
    695 #ifdef BLOCK_SMOOTHING_SUPPORTED
    696   coef->coef_bits_latch = NULL;
    697 #endif
    698 
    699   /* Create the coefficient buffer. */
    700   if (need_full_buffer) {
    701 #ifdef D_MULTISCAN_FILES_SUPPORTED
    702     /* Allocate a full-image virtual array for each component, */
    703     /* padded to a multiple of samp_factor DCT blocks in each direction. */
    704     /* Note we ask for a pre-zeroed array. */
    705     int ci, access_rows;
    706     jpeg_component_info *compptr;
    707 
    708     for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
    709 	 ci++, compptr++) {
    710       access_rows = compptr->v_samp_factor;
    711 #ifdef BLOCK_SMOOTHING_SUPPORTED
    712       /* If block smoothing could be used, need a bigger window */
    713       if (cinfo->progressive_mode)
    714 	access_rows *= 3;
    715 #endif
    716       coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
    717 	((j_common_ptr) cinfo, JPOOL_IMAGE, TRUE,
    718 	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
    719 				(long) compptr->h_samp_factor),
    720 	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
    721 				(long) compptr->v_samp_factor),
    722 	 (JDIMENSION) access_rows);
    723     }
    724     coef->pub.consume_data = consume_data;
    725     coef->pub.decompress_data = decompress_data;
    726     coef->pub.coef_arrays = coef->whole_image; /* link to virtual arrays */
    727 #else
    728     ERREXIT(cinfo, JERR_NOT_COMPILED);
    729 #endif
    730   } else {
    731     /* We only need a single-MCU buffer. */
    732     JBLOCKROW buffer;
    733     int i;
    734 
    735     buffer = (JBLOCKROW)
    736       (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
    737 				  D_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
    738     for (i = 0; i < D_MAX_BLOCKS_IN_MCU; i++) {
    739       coef->MCU_buffer[i] = buffer + i;
    740     }
    741     coef->pub.consume_data = dummy_consume_data;
    742     coef->pub.decompress_data = decompress_onepass;
    743     coef->pub.coef_arrays = NULL; /* flag for no virtual arrays */
    744   }
    745 
    746   /* Allocate the workspace buffer */
    747   coef->workspace = (JCOEF *)
    748     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
    749                                 SIZEOF(JCOEF) * DCTSIZE2);
    750 }
    751