Fix details mask tiling issues

If rawprepare or demosaic run in tiled mode they can't write the details mask, that would require a rewrite
of those modules with a special tiled process implementation.

As we set the piece->pipe->want_detail_mask while committing params and that comes *after* the mask writing
modules we must check via _piece_may_tile() if we must a) fallback to cpu code or b) process CPU code but at
least leave a note.
Used a flag for that check instead of string compare.

src/develop/imageop.h

@@ -119,7 +119,8 @@ typedef enum dt_iop_flags_t
   IOP_FLAGS_GUIDES_SPECIAL_DRAW = 1 << 14, // handle the grid drawing directly
   IOP_FLAGS_GUIDES_WIDGET = 1 << 15,     // require the guides widget
   IOP_FLAGS_CROP_EXPOSER = 1 << 16,      // offers crop exposing
-  IOP_FLAGS_EXPAND_ROI_IN = 1 << 17      // we might have to take special care about roi expansion
+  IOP_FLAGS_EXPAND_ROI_IN = 1 << 17,     // we might have to take special care about roi expansion
+  IOP_FLAGS_WRITE_DETAILS = 1 << 18
 } dt_iop_flags_t;
 
 /** status of a module*/

src/develop/pixelpipe_hb.c

@@ -1071,6 +1071,13 @@ static gboolean _request_color_pick(dt_dev_pixelpipe_t *pipe,
     && module->request_color_pick != DT_REQUEST_COLORPICK_OFF;
 }
 
+// Is it worth to use a dt_iop_flags_t for this ?
+static inline gboolean _piece_may_tile(const dt_dev_pixelpipe_iop_t *piece)
+{
+  return piece->process_tiling_ready
+        && !(piece->pipe->want_detail_mask && piece->module->flags() & IOP_FLAGS_WRITE_DETAILS);
+}
+
 static void _collect_histogram_on_CPU(dt_dev_pixelpipe_t *pipe,
                                       dt_develop_t *dev,
                                       float *input,
@@ -1185,7 +1192,7 @@ static gboolean _pixelpipe_process_on_CPU(dt_dev_pixelpipe_t *pipe,
                      roi_in->width, roi_in->height, in_bpp,
                      TRUE, dt_dev_pixelpipe_type_to_str(piece->pipe->type));
 
-  if(!fitting && piece->process_tiling_ready)
+  if(!fitting && _piece_may_tile(piece))
   {
     dt_print_pipe(DT_DEBUG_PIPE,
                         "process tiles",
@@ -1214,7 +1221,8 @@ static gboolean _pixelpipe_process_on_CPU(dt_dev_pixelpipe_t *pipe,
 
     // this code section is for simplistic benchmarking via --bench-module
     if((piece->pipe->type & (DT_DEV_PIXELPIPE_FULL | DT_DEV_PIXELPIPE_EXPORT))
-       && darktable.bench_module)
+       && darktable.bench_module
+       && fitting)
     {
       if(dt_str_commasubstring(darktable.bench_module, module->op))
       {
@@ -1712,7 +1720,7 @@ static gboolean _dev_pixelpipe_process_rec(dt_dev_pixelpipe_t *pipe,
 
     if(possible_cl && !fits_on_device)
     {
-      if(!piece->process_tiling_ready)
+      if(!_piece_may_tile(piece))
         possible_cl = FALSE;
 
       const float advantage = darktable.opencl->dev[pipe->devid].advantage;

src/iop/demosaic.c

@@ -300,7 +300,7 @@ int default_group()
 
 int flags()
 {
-  return IOP_FLAGS_ALLOW_TILING | IOP_FLAGS_ONE_INSTANCE | IOP_FLAGS_FENCE;
+  return IOP_FLAGS_ALLOW_TILING | IOP_FLAGS_ONE_INSTANCE | IOP_FLAGS_FENCE | IOP_FLAGS_WRITE_DETAILS;
 }
 
 dt_iop_colorspace_type_t default_colorspace(dt_iop_module_t *self,

src/iop/rawprepare.c

@@ -107,7 +107,7 @@ int operation_tags()
 int flags()
 {
   return IOP_FLAGS_ALLOW_TILING | IOP_FLAGS_TILING_FULL_ROI | IOP_FLAGS_ONE_INSTANCE
-    | IOP_FLAGS_UNSAFE_COPY;
+    | IOP_FLAGS_UNSAFE_COPY | IOP_FLAGS_WRITE_DETAILS;
 }
 
 int default_group()