michael@0: /*
michael@0:  * jcsample.c
michael@0:  *
michael@0:  * Copyright (C) 1991-1996, Thomas G. Lane.
michael@0:  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0:  * This file is part of the Independent JPEG Group's software.
michael@0:  * For conditions of distribution and use, see the accompanying README file.
michael@0:  *
michael@0:  * This file contains downsampling routines.
michael@0:  *
michael@0:  * Downsampling input data is counted in "row groups".  A row group
michael@0:  * is defined to be max_v_samp_factor pixel rows of each component,
michael@0:  * from which the downsampler produces v_samp_factor sample rows.
michael@0:  * A single row group is processed in each call to the downsampler module.
michael@0:  *
michael@0:  * The downsampler is responsible for edge-expansion of its output data
michael@0:  * to fill an integral number of DCT blocks horizontally.  The source buffer
michael@0:  * may be modified if it is helpful for this purpose (the source buffer is
michael@0:  * allocated wide enough to correspond to the desired output width).
michael@0:  * The caller (the prep controller) is responsible for vertical padding.
michael@0:  *
michael@0:  * The downsampler may request "context rows" by setting need_context_rows
michael@0:  * during startup.  In this case, the input arrays will contain at least
michael@0:  * one row group's worth of pixels above and below the passed-in data;
michael@0:  * the caller will create dummy rows at image top and bottom by replicating
michael@0:  * the first or last real pixel row.
michael@0:  *
michael@0:  * An excellent reference for image resampling is
michael@0:  *   Digital Image Warping, George Wolberg, 1990.
michael@0:  *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
michael@0:  *
michael@0:  * The downsampling algorithm used here is a simple average of the source
michael@0:  * pixels covered by the output pixel.  The hi-falutin sampling literature
michael@0:  * refers to this as a "box filter".  In general the characteristics of a box
michael@0:  * filter are not very good, but for the specific cases we normally use (1:1
michael@0:  * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
michael@0:  * nearly so bad.  If you intend to use other sampling ratios, you'd be well
michael@0:  * advised to improve this code.
michael@0:  *
michael@0:  * A simple input-smoothing capability is provided.  This is mainly intended
michael@0:  * for cleaning up color-dithered GIF input files (if you find it inadequate,
michael@0:  * we suggest using an external filtering program such as pnmconvol).  When
michael@0:  * enabled, each input pixel P is replaced by a weighted sum of itself and its
michael@0:  * eight neighbors.  P's weight is 1-8*SF and each neighbor's weight is SF,
michael@0:  * where SF = (smoothing_factor / 1024).
michael@0:  * Currently, smoothing is only supported for 2h2v sampling factors.
michael@0:  */
michael@0: 
michael@0: #define JPEG_INTERNALS
michael@0: #include "jinclude.h"
michael@0: #include "jpeglib.h"
michael@0: #include "jsimd.h"
michael@0: 
michael@0: 
michael@0: /* Pointer to routine to downsample a single component */
michael@0: typedef JMETHOD(void, downsample1_ptr,
michael@0: 		(j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		 JSAMPARRAY input_data, JSAMPARRAY output_data));
michael@0: 
michael@0: /* Private subobject */
michael@0: 
michael@0: typedef struct {
michael@0:   struct jpeg_downsampler pub;	/* public fields */
michael@0: 
michael@0:   /* Downsampling method pointers, one per component */
michael@0:   downsample1_ptr methods[MAX_COMPONENTS];
michael@0: } my_downsampler;
michael@0: 
michael@0: typedef my_downsampler * my_downsample_ptr;
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Initialize for a downsampling pass.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: start_pass_downsample (j_compress_ptr cinfo)
michael@0: {
michael@0:   /* no work for now */
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Expand a component horizontally from width input_cols to width output_cols,
michael@0:  * by duplicating the rightmost samples.
michael@0:  */
michael@0: 
michael@0: LOCAL(void)
michael@0: expand_right_edge (JSAMPARRAY image_data, int num_rows,
michael@0: 		   JDIMENSION input_cols, JDIMENSION output_cols)
michael@0: {
michael@0:   register JSAMPROW ptr;
michael@0:   register JSAMPLE pixval;
michael@0:   register int count;
michael@0:   int row;
michael@0:   int numcols = (int) (output_cols - input_cols);
michael@0: 
michael@0:   if (numcols > 0) {
michael@0:     for (row = 0; row < num_rows; row++) {
michael@0:       ptr = image_data[row] + input_cols;
michael@0:       pixval = ptr[-1];		/* don't need GETJSAMPLE() here */
michael@0:       for (count = numcols; count > 0; count--)
michael@0: 	*ptr++ = pixval;
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Do downsampling for a whole row group (all components).
michael@0:  *
michael@0:  * In this version we simply downsample each component independently.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: sep_downsample (j_compress_ptr cinfo,
michael@0: 		JSAMPIMAGE input_buf, JDIMENSION in_row_index,
michael@0: 		JSAMPIMAGE output_buf, JDIMENSION out_row_group_index)
michael@0: {
michael@0:   my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample;
michael@0:   int ci;
michael@0:   jpeg_component_info * compptr;
michael@0:   JSAMPARRAY in_ptr, out_ptr;
michael@0: 
michael@0:   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
michael@0:        ci++, compptr++) {
michael@0:     in_ptr = input_buf[ci] + in_row_index;
michael@0:     out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor);
michael@0:     (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * One row group is processed per call.
michael@0:  * This version handles arbitrary integral sampling ratios, without smoothing.
michael@0:  * Note that this version is not actually used for customary sampling ratios.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: int_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
michael@0:   JDIMENSION outcol, outcol_h;	/* outcol_h == outcol*h_expand */
michael@0:   JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
michael@0:   JSAMPROW inptr, outptr;
michael@0:   INT32 outvalue;
michael@0: 
michael@0:   h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor;
michael@0:   v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor;
michael@0:   numpix = h_expand * v_expand;
michael@0:   numpix2 = numpix/2;
michael@0: 
michael@0:   /* Expand input data enough to let all the output samples be generated
michael@0:    * by the standard loop.  Special-casing padded output would be more
michael@0:    * efficient.
michael@0:    */
michael@0:   expand_right_edge(input_data, cinfo->max_v_samp_factor,
michael@0: 		    cinfo->image_width, output_cols * h_expand);
michael@0: 
michael@0:   inrow = 0;
michael@0:   for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
michael@0:     outptr = output_data[outrow];
michael@0:     for (outcol = 0, outcol_h = 0; outcol < output_cols;
michael@0: 	 outcol++, outcol_h += h_expand) {
michael@0:       outvalue = 0;
michael@0:       for (v = 0; v < v_expand; v++) {
michael@0: 	inptr = input_data[inrow+v] + outcol_h;
michael@0: 	for (h = 0; h < h_expand; h++) {
michael@0: 	  outvalue += (INT32) GETJSAMPLE(*inptr++);
michael@0: 	}
michael@0:       }
michael@0:       *outptr++ = (JSAMPLE) ((outvalue + numpix2) / numpix);
michael@0:     }
michael@0:     inrow += v_expand;
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * This version handles the special case of a full-size component,
michael@0:  * without smoothing.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: fullsize_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		     JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   /* Copy the data */
michael@0:   jcopy_sample_rows(input_data, 0, output_data, 0,
michael@0: 		    cinfo->max_v_samp_factor, cinfo->image_width);
michael@0:   /* Edge-expand */
michael@0:   expand_right_edge(output_data, cinfo->max_v_samp_factor,
michael@0: 		    cinfo->image_width, compptr->width_in_blocks * DCTSIZE);
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * This version handles the common case of 2:1 horizontal and 1:1 vertical,
michael@0:  * without smoothing.
michael@0:  *
michael@0:  * A note about the "bias" calculations: when rounding fractional values to
michael@0:  * integer, we do not want to always round 0.5 up to the next integer.
michael@0:  * If we did that, we'd introduce a noticeable bias towards larger values.
michael@0:  * Instead, this code is arranged so that 0.5 will be rounded up or down at
michael@0:  * alternate pixel locations (a simple ordered dither pattern).
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		 JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   int outrow;
michael@0:   JDIMENSION outcol;
michael@0:   JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
michael@0:   register JSAMPROW inptr, outptr;
michael@0:   register int bias;
michael@0: 
michael@0:   /* Expand input data enough to let all the output samples be generated
michael@0:    * by the standard loop.  Special-casing padded output would be more
michael@0:    * efficient.
michael@0:    */
michael@0:   expand_right_edge(input_data, cinfo->max_v_samp_factor,
michael@0: 		    cinfo->image_width, output_cols * 2);
michael@0: 
michael@0:   for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
michael@0:     outptr = output_data[outrow];
michael@0:     inptr = input_data[outrow];
michael@0:     bias = 0;			/* bias = 0,1,0,1,... for successive samples */
michael@0:     for (outcol = 0; outcol < output_cols; outcol++) {
michael@0:       *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1])
michael@0: 			      + bias) >> 1);
michael@0:       bias ^= 1;		/* 0=>1, 1=>0 */
michael@0:       inptr += 2;
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
michael@0:  * without smoothing.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 		 JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   int inrow, outrow;
michael@0:   JDIMENSION outcol;
michael@0:   JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
michael@0:   register JSAMPROW inptr0, inptr1, outptr;
michael@0:   register int bias;
michael@0: 
michael@0:   /* Expand input data enough to let all the output samples be generated
michael@0:    * by the standard loop.  Special-casing padded output would be more
michael@0:    * efficient.
michael@0:    */
michael@0:   expand_right_edge(input_data, cinfo->max_v_samp_factor,
michael@0: 		    cinfo->image_width, output_cols * 2);
michael@0: 
michael@0:   inrow = 0;
michael@0:   for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
michael@0:     outptr = output_data[outrow];
michael@0:     inptr0 = input_data[inrow];
michael@0:     inptr1 = input_data[inrow+1];
michael@0:     bias = 1;			/* bias = 1,2,1,2,... for successive samples */
michael@0:     for (outcol = 0; outcol < output_cols; outcol++) {
michael@0:       *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
michael@0: 			      GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1])
michael@0: 			      + bias) >> 2);
michael@0:       bias ^= 3;		/* 1=>2, 2=>1 */
michael@0:       inptr0 += 2; inptr1 += 2;
michael@0:     }
michael@0:     inrow += 2;
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: #ifdef INPUT_SMOOTHING_SUPPORTED
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
michael@0:  * with smoothing.  One row of context is required.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0: 			JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   int inrow, outrow;
michael@0:   JDIMENSION colctr;
michael@0:   JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
michael@0:   register JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
michael@0:   INT32 membersum, neighsum, memberscale, neighscale;
michael@0: 
michael@0:   /* Expand input data enough to let all the output samples be generated
michael@0:    * by the standard loop.  Special-casing padded output would be more
michael@0:    * efficient.
michael@0:    */
michael@0:   expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
michael@0: 		    cinfo->image_width, output_cols * 2);
michael@0: 
michael@0:   /* We don't bother to form the individual "smoothed" input pixel values;
michael@0:    * we can directly compute the output which is the average of the four
michael@0:    * smoothed values.  Each of the four member pixels contributes a fraction
michael@0:    * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
michael@0:    * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
michael@0:    * output.  The four corner-adjacent neighbor pixels contribute a fraction
michael@0:    * SF to just one smoothed pixel, or SF/4 to the final output; while the
michael@0:    * eight edge-adjacent neighbors contribute SF to each of two smoothed
michael@0:    * pixels, or SF/2 overall.  In order to use integer arithmetic, these
michael@0:    * factors are scaled by 2^16 = 65536.
michael@0:    * Also recall that SF = smoothing_factor / 1024.
michael@0:    */
michael@0: 
michael@0:   memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
michael@0:   neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
michael@0: 
michael@0:   inrow = 0;
michael@0:   for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
michael@0:     outptr = output_data[outrow];
michael@0:     inptr0 = input_data[inrow];
michael@0:     inptr1 = input_data[inrow+1];
michael@0:     above_ptr = input_data[inrow-1];
michael@0:     below_ptr = input_data[inrow+2];
michael@0: 
michael@0:     /* Special case for first column: pretend column -1 is same as column 0 */
michael@0:     membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
michael@0: 		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
michael@0:     neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
michael@0: 	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
michael@0: 	       GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
michael@0: 	       GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
michael@0:     neighsum += neighsum;
michael@0:     neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
michael@0: 		GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
michael@0:     membersum = membersum * memberscale + neighsum * neighscale;
michael@0:     *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0:     inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
michael@0: 
michael@0:     for (colctr = output_cols - 2; colctr > 0; colctr--) {
michael@0:       /* sum of pixels directly mapped to this output element */
michael@0:       membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
michael@0: 		  GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
michael@0:       /* sum of edge-neighbor pixels */
michael@0:       neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
michael@0: 		 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
michael@0: 		 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
michael@0: 		 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
michael@0:       /* The edge-neighbors count twice as much as corner-neighbors */
michael@0:       neighsum += neighsum;
michael@0:       /* Add in the corner-neighbors */
michael@0:       neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
michael@0: 		  GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
michael@0:       /* form final output scaled up by 2^16 */
michael@0:       membersum = membersum * memberscale + neighsum * neighscale;
michael@0:       /* round, descale and output it */
michael@0:       *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0:       inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
michael@0:     }
michael@0: 
michael@0:     /* Special case for last column */
michael@0:     membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
michael@0: 		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
michael@0:     neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
michael@0: 	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
michael@0: 	       GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
michael@0: 	       GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
michael@0:     neighsum += neighsum;
michael@0:     neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
michael@0: 		GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
michael@0:     membersum = membersum * memberscale + neighsum * neighscale;
michael@0:     *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0: 
michael@0:     inrow += 2;
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Downsample pixel values of a single component.
michael@0:  * This version handles the special case of a full-size component,
michael@0:  * with smoothing.  One row of context is required.
michael@0:  */
michael@0: 
michael@0: METHODDEF(void)
michael@0: fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
michael@0: 			    JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0: {
michael@0:   int outrow;
michael@0:   JDIMENSION colctr;
michael@0:   JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
michael@0:   register JSAMPROW inptr, above_ptr, below_ptr, outptr;
michael@0:   INT32 membersum, neighsum, memberscale, neighscale;
michael@0:   int colsum, lastcolsum, nextcolsum;
michael@0: 
michael@0:   /* Expand input data enough to let all the output samples be generated
michael@0:    * by the standard loop.  Special-casing padded output would be more
michael@0:    * efficient.
michael@0:    */
michael@0:   expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
michael@0: 		    cinfo->image_width, output_cols);
michael@0: 
michael@0:   /* Each of the eight neighbor pixels contributes a fraction SF to the
michael@0:    * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
michael@0:    * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
michael@0:    * Also recall that SF = smoothing_factor / 1024.
michael@0:    */
michael@0: 
michael@0:   memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
michael@0:   neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
michael@0: 
michael@0:   for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
michael@0:     outptr = output_data[outrow];
michael@0:     inptr = input_data[outrow];
michael@0:     above_ptr = input_data[outrow-1];
michael@0:     below_ptr = input_data[outrow+1];
michael@0: 
michael@0:     /* Special case for first column */
michael@0:     colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
michael@0: 	     GETJSAMPLE(*inptr);
michael@0:     membersum = GETJSAMPLE(*inptr++);
michael@0:     nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
michael@0: 		 GETJSAMPLE(*inptr);
michael@0:     neighsum = colsum + (colsum - membersum) + nextcolsum;
michael@0:     membersum = membersum * memberscale + neighsum * neighscale;
michael@0:     *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0:     lastcolsum = colsum; colsum = nextcolsum;
michael@0: 
michael@0:     for (colctr = output_cols - 2; colctr > 0; colctr--) {
michael@0:       membersum = GETJSAMPLE(*inptr++);
michael@0:       above_ptr++; below_ptr++;
michael@0:       nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
michael@0: 		   GETJSAMPLE(*inptr);
michael@0:       neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
michael@0:       membersum = membersum * memberscale + neighsum * neighscale;
michael@0:       *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0:       lastcolsum = colsum; colsum = nextcolsum;
michael@0:     }
michael@0: 
michael@0:     /* Special case for last column */
michael@0:     membersum = GETJSAMPLE(*inptr);
michael@0:     neighsum = lastcolsum + (colsum - membersum) + colsum;
michael@0:     membersum = membersum * memberscale + neighsum * neighscale;
michael@0:     *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
michael@0: 
michael@0:   }
michael@0: }
michael@0: 
michael@0: #endif /* INPUT_SMOOTHING_SUPPORTED */
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Module initialization routine for downsampling.
michael@0:  * Note that we must select a routine for each component.
michael@0:  */
michael@0: 
michael@0: GLOBAL(void)
michael@0: jinit_downsampler (j_compress_ptr cinfo)
michael@0: {
michael@0:   my_downsample_ptr downsample;
michael@0:   int ci;
michael@0:   jpeg_component_info * compptr;
michael@0:   boolean smoothok = TRUE;
michael@0: 
michael@0:   downsample = (my_downsample_ptr)
michael@0:     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
michael@0: 				SIZEOF(my_downsampler));
michael@0:   cinfo->downsample = (struct jpeg_downsampler *) downsample;
michael@0:   downsample->pub.start_pass = start_pass_downsample;
michael@0:   downsample->pub.downsample = sep_downsample;
michael@0:   downsample->pub.need_context_rows = FALSE;
michael@0: 
michael@0:   if (cinfo->CCIR601_sampling)
michael@0:     ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
michael@0: 
michael@0:   /* Verify we can handle the sampling factors, and set up method pointers */
michael@0:   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
michael@0:        ci++, compptr++) {
michael@0:     if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
michael@0: 	compptr->v_samp_factor == cinfo->max_v_samp_factor) {
michael@0: #ifdef INPUT_SMOOTHING_SUPPORTED
michael@0:       if (cinfo->smoothing_factor) {
michael@0: 	downsample->methods[ci] = fullsize_smooth_downsample;
michael@0: 	downsample->pub.need_context_rows = TRUE;
michael@0:       } else
michael@0: #endif
michael@0: 	downsample->methods[ci] = fullsize_downsample;
michael@0:     } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
michael@0: 	       compptr->v_samp_factor == cinfo->max_v_samp_factor) {
michael@0:       smoothok = FALSE;
michael@0:       if (jsimd_can_h2v1_downsample())
michael@0:         downsample->methods[ci] = jsimd_h2v1_downsample;
michael@0:       else
michael@0:         downsample->methods[ci] = h2v1_downsample;
michael@0:     } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
michael@0: 	       compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
michael@0: #ifdef INPUT_SMOOTHING_SUPPORTED
michael@0:       if (cinfo->smoothing_factor) {
michael@0: 	downsample->methods[ci] = h2v2_smooth_downsample;
michael@0: 	downsample->pub.need_context_rows = TRUE;
michael@0:       } else
michael@0: #endif
michael@0: 	if (jsimd_can_h2v2_downsample())
michael@0: 	  downsample->methods[ci] = jsimd_h2v2_downsample;
michael@0: 	else
michael@0: 	  downsample->methods[ci] = h2v2_downsample;
michael@0:     } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
michael@0: 	       (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
michael@0:       smoothok = FALSE;
michael@0:       downsample->methods[ci] = int_downsample;
michael@0:     } else
michael@0:       ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
michael@0:   }
michael@0: 
michael@0: #ifdef INPUT_SMOOTHING_SUPPORTED
michael@0:   if (cinfo->smoothing_factor && !smoothok)
michael@0:     TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
michael@0: #endif
michael@0: }