Invert directions again, increasing speed but adding bugs.

diff --git a/qscale.c b/qscale.c
index c847a12122cc38e1deb80fe3e4c75a60c6d27428..8fecd666a972b1fad7a35b302dc5f8445677aca0 100644 (file)
--- a/qscale.c
+++ b/qscale.c
@@ -25,25 +25,22 @@ double lanczos_tap(double x)
                return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
 }
 
-struct filter {
-       struct pix_desc *pd;
-       float *coeffs;
-};
+
 struct pix_desc {
        unsigned start, end;
        unsigned startcoeff;
 };
 
-void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned nw)
+void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned dstride)
 {
        struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
        int size_coeffs = 8;
        float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
        int num_coeffs = 0;
-       int x, sx;
+       int x, y, sx;
        double sf = (double)w / (double)nw;
        double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
-       
+
        /* calculate the filter */
        for (x = 0; x < nw; ++x) {
                int start = ceil(x * sf - support);
@@ -77,39 +74,35 @@ void hscale_calc_filter(struct filter *filter, unsigned w, unsigned h, unsigned
                        coeffs[pd[x].startcoeff + sx - start] /= sum;
                }
        }
-
-       filter->pd = pd;
-       filter->coeffs = coeffs;
-}
-
-void hscale(struct filter *filter, unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
-{
-       struct pix_desc *pd = filter->pd;
-       float *coeffs = filter->coeffs;
-       int y;
-
+       
        for (y = 0; y < h; ++y) {
-               unsigned char *sptr = pix + y*sstride;
-               float *dptr = npix + y*dstride;
-               float acc = 0.0;
-               int x;
+               float *sptr = pix + y*w;
+               unsigned char *dptr = npix + y*dstride;
+               unsigned char ch = 0;
                for (x = 0; x < nw; ++x) {
+                       float acc = 0.0;
                        float *cf = &coeffs[pd[x].startcoeff];
                        unsigned sx;
-                       acc = 0.0;
                        
                        for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
                                acc += sptr[sx] * *cf++;
                        }
-                       *dptr++ = acc;
+
+                       if (acc < 0.0)
+                               ch = 0;
+                       else if (acc > 255.0)
+                               ch = 255;
+                       else
+                               ch = (unsigned char)acc;
+                       *dptr++ = ch;
                }
                for ( ; x < dstride; ++x) {
-                       *dptr++ = acc;
+                       *dptr++ = ch;
                }
        }
 }
 
-void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
+void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 {
        struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
        int size_coeffs = 8;
@@ -155,8 +148,8 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh
 
 #if CACHE_LINE_FACTOR > 1
        for (x = 0; x < w; x += CACHE_LINE_FACTOR) {
-               float *sptr = pix + x;
-               unsigned char *dptr = npix + x;
+               unsigned char *sptr = pix + x;
+               float *dptr = npix + x;
                for (y = 0; y < nh; ++y) {
                        int i;
                        float acc[CACHE_LINE_FACTOR];
@@ -166,6 +159,7 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh
                        unsigned sy;
                        
                        for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
+                               //asm volatile ("prefetcht0 %0" :: "m" (sptr[(sy+1) * w]));
                                for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
                                        acc[i] += sptr[sy * w + i] * *cf;
                                }
@@ -173,14 +167,7 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh
                        }
 
                        for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
-                               unsigned char ch;
-                               if (acc[i] < 0.0)
-                                       ch = 0;
-                               else if (acc[i] > 255.0)
-                                       ch = 255;
-                               else
-                                       ch = (unsigned char)acc[i];
-                               dptr[i] = ch;
+                               dptr[i] = acc[i];
                        }
                        dptr += dstride;
                }
@@ -189,8 +176,8 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh
 #else
        for (x = 0; x < w; ++x) {
 #endif
-               float *sptr = pix + x;
-               unsigned char *dptr = npix + x;
+               unsigned char *sptr = pix + x;
+               float *dptr = npix + x;
                for (y = 0; y < nh; ++y) {
                        float acc = 0.0;
                        float *cf = &coeffs[pd[y].startcoeff];
@@ -200,14 +187,7 @@ void vscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nh
                                acc += sptr[sy * w] * *cf++;
                        }
 
-                       unsigned char ch;
-                       if (acc < 0.0)
-                               ch = 0;
-                       else if (acc > 255.0)
-                               ch = 255;
-                       else
-                               ch = (unsigned char)acc;
-                       *dptr = ch;
+                       *dptr = acc;
                        dptr += dstride;
                }
        }
@@ -244,38 +224,15 @@ int main(int argc, char **argv)
        fprintf(stderr, "  Y component: %ux%u -> %ux%u\n", dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nw0, nh0);
        fprintf(stderr, "  Cb component: %ux%u -> %ux%u\n", dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nw1, nh1);
        fprintf(stderr, "  Cr component: %ux%u -> %ux%u\n", dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nw2, nh2);
-               
-       float *npix_y  = (float *)malloc((dinfo.comp_info[0].height_in_blocks + dinfo.comp_info[0].v_samp_factor - 1) * DCTSIZE * stride0 * sizeof(float));
-       float *npix_cb = (float *)malloc((dinfo.comp_info[1].height_in_blocks + dinfo.comp_info[1].v_samp_factor - 1) * DCTSIZE * stride1 * sizeof(float));
-       float *npix_cr = (float *)malloc((dinfo.comp_info[2].height_in_blocks + dinfo.comp_info[2].v_samp_factor - 1) * DCTSIZE * stride2 * sizeof(float));
-       JSAMPLE *data_y  = (unsigned char *)malloc(nh0 * stride0);
-       JSAMPLE *data_cb = (unsigned char *)malloc(nh1 * stride1);
-       JSAMPLE *data_cr = (unsigned char *)malloc(nh2 * stride2);
-
-       struct filter filt0, filt1, filt2;
-
-       hscale_calc_filter(&filt0,
-               /* w=       */ dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor,
-               /* h=       */ dinfo.comp_info[0].v_samp_factor * DCTSIZE,
-               /* nw=      */ nw0
-       );
-       hscale_calc_filter(&filt1,
-               /* w=       */ dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor,
-               /* h=       */ dinfo.comp_info[1].v_samp_factor * DCTSIZE,
-               /* nw=      */ nw1
-       );
-       hscale_calc_filter(&filt2,
-               /* w=       */ dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor,
-               /* h=       */ dinfo.comp_info[2].v_samp_factor * DCTSIZE,
-               /* nw=      */ nw2
-       );
+
+       JSAMPLE *data_y  = (JSAMPLE*)malloc(dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_cb = (JSAMPLE*)malloc(dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_cr = (JSAMPLE*)malloc(dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
+       JSAMPLE *data_ny, *data_ncb, *data_ncr;
 
        int total_lines = 0, blocks = 0;
        while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
                unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
-               JSAMPLE tmp_y [(dinfo.comp_info[0].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[0].v_samp_factor)];
-               JSAMPLE tmp_cb[(dinfo.comp_info[1].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[1].v_samp_factor)];
-               JSAMPLE tmp_cr[(dinfo.comp_info[2].width_in_blocks * DCTSIZE) * (DCTSIZE * dinfo.comp_info[2].v_samp_factor)];
 
                JSAMPROW y_row_ptrs[max_lines];
                JSAMPROW cb_row_ptrs[max_lines];
@@ -284,49 +241,39 @@ int main(int argc, char **argv)
                int i;
 
                for (i = 0; i < max_lines; ++i) {
-                       y_row_ptrs[i]  = tmp_y  + i * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
-                       cb_row_ptrs[i] = tmp_cb + i * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
-                       cr_row_ptrs[i] = tmp_cr + i * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
+                       y_row_ptrs[i]  = data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
+                       cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
+                       cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
                }
                
-               int lines = jpeg_read_raw_data(&dinfo, ptrs, max_lines);
-               if (lines == 0)
-                       break;
-
-               hscale(/* filter=  */ &filt0,
-                      /* from=    */ tmp_y, 
-                      /* to=      */ npix_y + blocks * DCTSIZE * dinfo.comp_info[0].v_samp_factor * stride0,
-                      /* w=       */ dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor,
-                      /* h=       */ dinfo.comp_info[0].v_samp_factor * DCTSIZE,
-                      /* nw=      */ nw0,
-                      /* sstride= */ dinfo.comp_info[0].width_in_blocks * DCTSIZE,
-                      /* dstride= */ stride0);
-
-               hscale(/* filter=  */ &filt1,
-                      /* from=    */ tmp_cb, 
-                      /* to=      */ npix_cb + blocks * DCTSIZE * dinfo.comp_info[1].v_samp_factor * stride1,
-                      /* w=       */ dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor,
-                      /* h=       */ dinfo.comp_info[1].v_samp_factor * DCTSIZE,
-                      /* nw=      */ nw1,
-                      /* sstride= */ dinfo.comp_info[1].width_in_blocks * DCTSIZE,
-                      /* dstride= */ stride1);
-
-               hscale(/* filter=  */ &filt2,
-                      /* from=    */ tmp_cr, 
-                      /* to=      */ npix_cr + blocks * DCTSIZE * dinfo.comp_info[2].v_samp_factor * stride2,
-                      /* w=       */ dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor,
-                      /* h=       */ dinfo.comp_info[2].v_samp_factor * DCTSIZE,
-                      /* nw=      */ nw2,
-                      /* sstride= */ dinfo.comp_info[2].width_in_blocks * DCTSIZE,
-                      /* dstride= */ stride2);
-               
                total_lines += max_lines;
                ++blocks;
+
+               if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
+                       break;
        }
 
-       vscale(npix_y,  data_y,  stride0, dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor, nh0, stride0);
-       vscale(npix_cb, data_cb, stride1, dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor, nh1, stride1);
-       vscale(npix_cr, data_cr, stride2, dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor, nh2, stride2);
+       {
+               float *npix = (float*)malloc(dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));       
+               vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
+               data_ny = (unsigned char *)malloc(nw0 * stride0);
+               hscale(npix, data_ny, dinfo.comp_info[0].width_in_blocks * DCTSIZE, nh0, nw0, stride0);
+               free(npix);
+       }
+       {
+               float *npix = (float*)malloc(dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));       
+               vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
+               data_ncr = (unsigned char *)malloc(nw1 * stride1);
+               hscale(npix, data_ncr, dinfo.comp_info[1].width_in_blocks * DCTSIZE, nh1, nw1, stride1);
+               free(npix);
+       }
+       {
+               float *npix = (float*)malloc(dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));       
+               vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
+               data_ncb = (unsigned char *)malloc(nw2 * stride2);
+               hscale(npix, data_ncb, dinfo.comp_info[2].width_in_blocks * DCTSIZE, nh2, nw2, stride2);
+               free(npix);
+       }
        jpeg_destroy_decompress(&dinfo);
        
        struct jpeg_compress_struct cinfo;
@@ -373,9 +320,9 @@ int main(int argc, char **argv)
                        if (crline > nh2 - 1)
                                crline = nh2 - 1;
 
-                       y_row_ptrs[i]  = data_y  + yline * stride0;
-                       cb_row_ptrs[i] = data_cb + cbline * stride1;
-                       cr_row_ptrs[i] = data_cr + crline * stride2;
+                       y_row_ptrs[i]  = data_ny  + yline * stride0;
+                       cb_row_ptrs[i] = data_ncb + cbline * stride1;
+                       cr_row_ptrs[i] = data_ncr + crline * stride2;
                }
                
                total_lines += max_lines;