Merge pull request #17689 from victoryforce/accelerate-PFM-loading

[imageio] Radically speed up PFM loading

Author: TurboGit

src/imageio/imageio_pfm.c

@@ -34,21 +34,32 @@
 #include <time.h>
 #include <unistd.h>
 
-dt_imageio_retval_t dt_imageio_open_pfm(dt_image_t *img, const char *filename, dt_mipmap_buffer_t *mbuf)
+dt_imageio_retval_t dt_imageio_open_pfm(dt_image_t *img,
+                                        const char *filename,
+                                        dt_mipmap_buffer_t *mbuf)
 {
   const char *ext = filename + strlen(filename);
-  while(*ext != '.' && ext > filename) ext--;
-  if(strcasecmp(ext, ".pfm")) return DT_IMAGEIO_LOAD_FAILED;
+
+  while(*ext != '.' && ext > filename)
+    ext--;
+
+  if(strcasecmp(ext, ".pfm"))
+    return DT_IMAGEIO_LOAD_FAILED;
 
   FILE *f = g_fopen(filename, "rb");
-  if(!f) return DT_IMAGEIO_FILE_NOT_FOUND;
+  if(!f)
+    return DT_IMAGEIO_FILE_NOT_FOUND;
 
   int ret = 0;
   int cols = 3;
   float scale_factor;
   char head[2] = { 'X', 'X' };
+
   ret = fscanf(f, "%c%c\n", head, head + 1);
-  if(ret != 2 || head[0] != 'P') goto error_corrupt;
+
+  if(ret != 2 || head[0] != 'P')
+    goto error_corrupt;
+
   if(head[1] == 'F')
     cols = 3;
   else if(head[1] == 'f')
@@ -59,63 +70,78 @@ dt_imageio_retval_t dt_imageio_open_pfm(dt_image_t *img, const char *filename, d
   char width_string[10] = { 0 };
   char height_string[10] = { 0 };
   char scale_factor_string[64] = { 0 };
+
   ret = fscanf(f, "%9s %9s %63s%*[^\n]", width_string, height_string, scale_factor_string);
-  if(ret != 3) goto error_corrupt;
+
+  if(ret != 3)
+    goto error_corrupt;
 
   errno = 0;
   img->width = strtol(width_string, NULL, 0);
   img->height = strtol(height_string, NULL, 0);
   scale_factor = g_ascii_strtod(scale_factor_string, NULL);
-  if(errno != 0) goto error_corrupt;
-  if(img->width <= 0 || img->height <= 0 ) goto error_corrupt;
+
+  if(errno != 0)
+    goto error_corrupt;
+  if(img->width <= 0 || img->height <= 0 )
+    goto error_corrupt;
 
   ret = fread(&ret, sizeof(char), 1, f);
-  if(ret != 1) goto error_corrupt;
+  if(ret != 1)
+    goto error_corrupt;
   ret = 0;
 
   int swap_byte_order = (scale_factor >= 0.0) ^ (G_BYTE_ORDER == G_BIG_ENDIAN);
 
   img->buf_dsc.channels = 4;
   img->buf_dsc.datatype = TYPE_FLOAT;
   float *buf = (float *)dt_mipmap_cache_alloc(mbuf, img);
-  if(!buf) goto error_cache_full;
+  if(!buf)
+    goto error_cache_full;
+
+  const size_t npixels = (size_t)img->width * img->height;
+
+  float *readbuf = dt_alloc_align_float(npixels * 4);
+  if(!readbuf)
+    goto error_cache_full;
+
+  union { float f; guint32 i; } value;
 
   if(cols == 3)
   {
-    ret = fread(buf, 3 * sizeof(float), (size_t)img->width * img->height, f);
-    for(size_t i = (size_t)img->width * img->height; i > 0; i--)
-      for(int c = 0; c < 3; c++)
+    ret = fread(readbuf, 3 * sizeof(float), npixels, f);
+
+DT_OMP_FOR(collapse(2))
+    for(size_t j = 0; j < img->height; j++)
+      for(size_t i = 0; i < img->width; i++)
       {
-        union { float f; guint32 i; } v;
-        v.f = buf[3 * (i - 1) + c];
-        if(swap_byte_order) v.i = GUINT32_SWAP_LE_BE(v.i);
-        buf[4 * (i - 1) + c] = v.f;
+        dt_aligned_pixel_t pix = {0.0f, 0.0f, 0.0f, 0.0f};
+        for_three_channels(c)
+        {
+        value.f = readbuf[3 * ((img->height - 1 - j) * img->width + i) + c];
+        if(swap_byte_order) value.i = GUINT32_SWAP_LE_BE(value.i);
+        pix[c] = value.f;
+        }
+        copy_pixel_nontemporal(&buf[4 * (img->width * j + i)], pix);
       }
   }
   else
+  {
+    ret = fread(readbuf, sizeof(float), npixels, f);
+
+DT_OMP_FOR(collapse(2))
     for(size_t j = 0; j < img->height; j++)
       for(size_t i = 0; i < img->width; i++)
       {
-        union { float f; guint32 i; } v;
-        ret = fread(&v.f, sizeof(float), 1, f);
-        if(swap_byte_order) v.i = GUINT32_SWAP_LE_BE(v.i);
+        value.f = readbuf[((img->height - 1 - j) * img->width + i)];
+        if(swap_byte_order) value.i = GUINT32_SWAP_LE_BE(value.i);
         buf[4 * (img->width * j + i) + 2] = buf[4 * (img->width * j + i) + 1]
-            = buf[4 * (img->width * j + i) + 0] = v.f;
+            = buf[4 * (img->width * j + i) + 0] = value.f;
       }
-
-  float *line = (float *)calloc(4 * img->width, sizeof(float));
-  if(line == NULL) goto error_cache_full;
-
-  for(size_t j = 0; j < img->height / 2; j++)
-  {
-    memcpy(line, buf + img->width * j * 4, sizeof(float) * 4 * img->width);
-    memcpy(buf + img->width * j * 4, buf + img->width * (img->height - 1 - j) * 4,
-           sizeof(float) * 4 * img->width);
-    memcpy(buf + img->width * (img->height - 1 - j) * 4, line, sizeof(float) * 4 * img->width);
   }
 
-  free(line);
   fclose(f);
+  dt_free_align(readbuf);
 
   img->buf_dsc.cst = IOP_CS_RGB;
   img->buf_dsc.filters = 0u;