Researching bottlenecks.temp-texture-painting-gpu

4 files changed, 54 insertions, 34 deletions

diff --git a/source/blender/gpu/shaders/sculpt_paint/infos/sculpt_paint_image_info.hh b/source/blender/gpu/shaders/sculpt_paint/infos/sculpt_paint_image_info.hh
index 1051abc63e9..86003867f02 100644
--- a/source/blender/gpu/shaders/sculpt_paint/infos/sculpt_paint_image_info.hh
+++ b/source/blender/gpu/shaders/sculpt_paint/infos/sculpt_paint_image_info.hh
@@ -8,7 +8,7 @@
 #include "gpu_shader_create_info.hh"
 
 GPU_SHADER_CREATE_INFO(sculpt_paint_sub_tiles)
-    .storage_buf(0, Qualifier::READ, "PaintTileData", "paint_tile_buf[]")
+    .storage_buf(0, Qualifier::READ_WRITE, "PaintTileData", "paint_tile_buf[]")
     .push_constant(Type::INT, "paint_tile_buf_len")
     .define("SUB_TILE_SIZE", "1024");
 
@@ -28,7 +28,7 @@ GPU_SHADER_CREATE_INFO(sculpt_paint_image_compute)
     .typedef_source("GPU_sculpt_shader_shared.h");
 
 GPU_SHADER_CREATE_INFO(sculpt_paint_image_merge_compute)
-    .local_group_size(1, 1, 1)
+    .local_group_size(32, 1, 1)
     .image(0, GPU_RGBA16F, Qualifier::READ, ImageType::FLOAT_3D, "paint_tiles_img")
     .image(1, GPU_RGBA16F, Qualifier::READ_WRITE, ImageType::FLOAT_2D, "texture_img")
     .push_constant(Type::INT, "layer_id")
diff --git a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_comp.glsl b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_comp.glsl
index 86c90de2d90..f4af1509233 100644
--- a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_comp.glsl
+++ b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_comp.glsl
@@ -20,9 +20,6 @@ void main()
 {
   PackedPixelRow row = pixel_row_buf[gl_GlobalInvocationID.x + pixel_row_offset];
   TrianglePaintInput triangle = paint_input[PIXEL_ROW_PRIM_INDEX(row)];
-  PaintTileData paint_tile;
-  ivec3 image_coord = paint_tile_coord_from_udim(
-      1001, PIXEL_ROW_START_IMAGE_COORD(row), paint_tile);
 
   uint row_len = PIXEL_ROW_LEN(row);
 
@@ -34,34 +31,42 @@ void main()
   vec3 delta;
   SCULPT_get_row_pos_and_delta(co1, co2, co3, triangle, row, pos, delta);
 
-  for (int x = 0; x < row_len; x++) {
-    /* TODO: Do clipping test. */
-    vec4 color;
-    bool color_read = false;
+  int x = 0;
+  while (x < row_len) {
+    PaintTileData paint_tile;
+    ivec3 image_coord = paint_tile_coord_from_udim(
+        1001, PIXEL_ROW_START_IMAGE_COORD(row) + int2(x, 0), paint_tile);
+    bool in_use = false;
+    for (; x < row_len && image_coord.x < SUB_TILE_SIZE; x++, image_coord.x++, pos += delta) {
+      /* TODO: Do clipping test. */
+      vec4 color;
+      bool color_read = false;
 
-    for (int step_index = paint_step_range[0]; step_index < paint_step_range[1]; step_index++) {
-      PaintStepData step_data = paint_step_buf[step_index];
-
-      float distance;
-      bool test_result = SCULPT_brush_test(paint_brush_buf.test, step_data, pos, distance);
-      if (test_result) {
-        if (!color_read) {
-          color = imageLoad(paint_tiles_img, image_coord);
-          color_read = true;
+      for (int step_index = paint_step_range[0]; step_index < paint_step_range[1]; step_index++) {
+        PaintStepData step_data = paint_step_buf[step_index];
+        float distance;
+        bool test_result = SCULPT_brush_test(paint_brush_buf.test, step_data, pos, distance);
+        if (test_result) {
+          if (!color_read) {
+            color = imageLoad(paint_tiles_img, image_coord);
+            color_read = true;
+            in_use = true;
+          }
+          // TODO: blend with color...
+          float factor = SCULPT_hardness_factor(distance, step_data.hardness, step_data.radius);
+          float curve_factor = SCULPT_curve_strength(factor, paint_brush_buf.falloff_shape);
+          vec4 final_paint_color = SCULPT_blend_color(
+              color, paint_brush_buf.color * curve_factor * step_data.strength);
+          final_paint_color *= paint_brush_buf.alpha;
+          color = SCULPT_blend_color(color, final_paint_color);
         }
-        // TODO: blend with color...
-        float factor = SCULPT_hardness_factor(distance, step_data.hardness, step_data.radius);
-        float curve_factor = SCULPT_curve_strength(factor, paint_brush_buf.falloff_shape);
-        vec4 final_paint_color = SCULPT_blend_color(
-            color, paint_brush_buf.color * curve_factor * step_data.strength);
-        final_paint_color *= paint_brush_buf.alpha;
-        color = SCULPT_blend_color(color, final_paint_color);
+      }
+      if (color_read) {
+        imageStore(paint_tiles_img, image_coord, color);
       }
     }
-    if (color_read) {
-      imageStore(paint_tiles_img, image_coord, color);
+    if (in_use) {
+      paint_tile_mark_used(paint_tile);
     }
-    image_coord.x += 1;
-    pos += delta;
   }
 }
\ No newline at end of file
diff --git a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_merge_comp.glsl b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_merge_comp.glsl
index b3ff3ac3ff3..cec4546f783 100644
--- a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_merge_comp.glsl
+++ b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_image_merge_comp.glsl
@@ -4,6 +4,10 @@ void main()
 {
   PaintTileData paint_tile;
   paint_tile_get_layer(layer_id, paint_tile);
+  if (!paint_tile.in_use_frame) {
+    return;
+  }
+
   ivec3 coord_in = ivec3(gl_GlobalInvocationID.xy, layer_id);
   vec4 paint_color = imageLoad(paint_tiles_img, coord_in);
   paint_color.a = 1.0;
diff --git a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_tile_lib.glsl b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_tile_lib.glsl
index c45664dd97c..88931feecd1 100644
--- a/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_tile_lib.glsl
+++ b/source/blender/gpu/shaders/sculpt_paint/sculpt_paint_tile_lib.glsl
@@ -3,28 +3,39 @@ ivec2 paint_tile_coord_to_sub_tile_id(ivec2 coord)
   return coord / ivec2(SUB_TILE_SIZE);
 }
 
-bool paint_tile_search(int tile_number, ivec2 coord, out PaintTileData r_paint_tile)
+bool paint_tile_search(int tile_number, int2 sub_tile_id, out PaintTileData r_paint_tile)
 {
-  int2 sub_tile_id = paint_tile_coord_to_sub_tile_id(coord);
   for (int i = 0; i < paint_tile_buf_len; i++) {
     if (paint_tile_buf[i].tile_number == tile_number &&
         paint_tile_buf[i].sub_tile_id == sub_tile_id) {
       r_paint_tile = paint_tile_buf[i];
+      r_paint_tile.index = i;
       return true;
     }
   }
   return false;
 }
 
+void paint_tile_mark_used(PaintTileData paint_tile)
+{
+  paint_tile_buf[paint_tile.index].in_use_frame = true;
+}
+
 void paint_tile_get_layer(int layer_id, out PaintTileData r_paint_tile)
 {
   r_paint_tile = paint_tile_buf[layer_id];
 }
 
-ivec3 paint_tile_coord_from_udim(int tile_number, ivec2 coord, out PaintTileData r_paint_tile)
+ivec3 paint_tile_coord_from_paint_tile(ivec2 coord, PaintTileData paint_tile)
 {
-  if (paint_tile_search(tile_number, coord, r_paint_tile)) {
-    return ivec3(coord - r_paint_tile.sub_tile_id * ivec2(SUB_TILE_SIZE), r_paint_tile.layer_id);
+  return ivec3(coord - paint_tile.sub_tile_id * ivec2(SUB_TILE_SIZE), paint_tile.layer_id);
+}
+
+ivec3 paint_tile_coord_from_udim(int tile_number, ivec2 coord, inout PaintTileData r_paint_tile)
+{
+  int2 sub_tile_id = paint_tile_coord_to_sub_tile_id(coord);
+  if (paint_tile_search(tile_number, sub_tile_id, r_paint_tile)) {
+    return paint_tile_coord_from_paint_tile(coord, r_paint_tile);
   }
 
   return ivec3(0);