4 files changed, 27 insertions, 8 deletions
diff --git a/drivers/gles3/rasterizer_canvas_gles3.cpp b/drivers/gles3/rasterizer_canvas_gles3.cpp
index 7f381b3f3e..7f67651e62 100644
--- a/drivers/gles3/rasterizer_canvas_gles3.cpp
+++ b/drivers/gles3/rasterizer_canvas_gles3.cpp
@@ -920,7 +920,7 @@ void RasterizerCanvasGLES3::_record_item_commands(const Item *p_item, RID p_rend
 					}
 
 					if (rect->flags & CANVAS_RECT_TRANSPOSE) {
-						dst_rect.size.x *= -1; // Encoding in the dst_rect.z uniform
+						state.instance_data_array[r_index].flags |= FLAGS_TRANSPOSE_RECT;
 					}
 
 					if (rect->flags & CANVAS_RECT_CLIP_UV) {
@@ -1128,6 +1128,12 @@ void RasterizerCanvasGLES3::_record_item_commands(const Item *p_item, RID p_rend
 					state.canvas_instance_batches[state.current_batch_index].tex = mm->texture;
 					state.canvas_instance_batches[state.current_batch_index].shader_variant = CanvasShaderGLES3::MODE_INSTANCED;
 
+					if (GLES3::MeshStorage::get_singleton()->multimesh_uses_colors(mm->multimesh)) {
+						state.instance_data_array[r_index].flags |= FLAGS_INSTANCING_HAS_COLORS;
+					}
+					if (GLES3::MeshStorage::get_singleton()->multimesh_uses_custom_data(mm->multimesh)) {
+						state.instance_data_array[r_index].flags |= FLAGS_INSTANCING_HAS_CUSTOM_DATA;
+					}
 				} else if (c->type == Item::Command::TYPE_PARTICLES) {
 					GLES3::ParticlesStorage *particles_storage = GLES3::ParticlesStorage::get_singleton();
 					GLES3::TextureStorage *texture_storage = GLES3::TextureStorage::get_singleton();
@@ -2684,6 +2690,7 @@ RasterizerCanvasGLES3::RasterizerCanvasGLES3() {
 // Default CanvasGroup shader.
 
 shader_type canvas_item;
+render_mode unshaded;
 
 uniform sampler2D screen_texture : hint_screen_texture, repeat_disable, filter_nearest;
 
@@ -2711,6 +2718,7 @@ void fragment() {
 // Default clip children shader.
 
 shader_type canvas_item;
+render_mode unshaded;
 
 uniform sampler2D screen_texture : hint_screen_texture, repeat_disable, filter_nearest;
 
diff --git a/drivers/gles3/rasterizer_gles3.cpp b/drivers/gles3/rasterizer_gles3.cpp
index 672dcacd88..2e3e6263ed 100644
--- a/drivers/gles3/rasterizer_gles3.cpp
+++ b/drivers/gles3/rasterizer_gles3.cpp
@@ -179,7 +179,8 @@ typedef void (*DEBUGPROCARB)(GLenum source,
 typedef void (*DebugMessageCallbackARB)(DEBUGPROCARB callback, const void *userParam);
 
 void RasterizerGLES3::initialize() {
-	print_line("OpenGL Renderer: " + RS::get_singleton()->get_video_adapter_name());
+	// NVIDIA suffixes all GPU model names with "/PCIe/SSE2" in OpenGL (but not Vulkan). This isn't necessary to display nowadays, so it can be trimmed.
+	print_line(vformat("OpenGL API %s - Compatibility - Using Device: %s - %s", RS::get_singleton()->get_video_adapter_api_version(), RS::get_singleton()->get_video_adapter_vendor(), RS::get_singleton()->get_video_adapter_name().trim_suffix("/PCIe/SSE2")));
 }
 
 void RasterizerGLES3::finalize() {
diff --git a/drivers/gles3/shaders/canvas.glsl b/drivers/gles3/shaders/canvas.glsl
index 1631c65385..ea0a0b660d 100644
--- a/drivers/gles3/shaders/canvas.glsl
+++ b/drivers/gles3/shaders/canvas.glsl
@@ -162,9 +162,13 @@ void main() {
 	vec2 uv = uv_attrib;
 
 #ifdef USE_INSTANCING
-	vec4 instance_color = vec4(unpackHalf2x16(instance_color_custom_data.x), unpackHalf2x16(instance_color_custom_data.y));
-	color *= instance_color;
-	instance_custom = vec4(unpackHalf2x16(instance_color_custom_data.z), unpackHalf2x16(instance_color_custom_data.w));
+	if (bool(read_draw_data_flags & FLAGS_INSTANCING_HAS_COLORS)) {
+		vec4 instance_color = vec4(unpackHalf2x16(instance_color_custom_data.x), unpackHalf2x16(instance_color_custom_data.y));
+		color *= instance_color;
+	}
+	if (bool(read_draw_data_flags & FLAGS_INSTANCING_HAS_CUSTOM_DATA)) {
+		instance_custom = vec4(unpackHalf2x16(instance_color_custom_data.z), unpackHalf2x16(instance_color_custom_data.w));
+	}
 #endif
 
 #else
@@ -645,7 +649,7 @@ void main() {
 
 #ifdef MODE_LIGHT_ONLY
 	color = vec4(0.0);
-#else
+#elif !defined(MODE_UNSHADED)
 	color *= canvas_modulation;
 #endif
 
diff --git a/drivers/gles3/storage/mesh_storage.cpp b/drivers/gles3/storage/mesh_storage.cpp
index 5ba0c5a09c..36b34dd8a2 100644
--- a/drivers/gles3/storage/mesh_storage.cpp
+++ b/drivers/gles3/storage/mesh_storage.cpp
@@ -491,6 +491,8 @@ void MeshStorage::mesh_set_custom_aabb(RID p_mesh, const AABB &p_aabb) {
 	Mesh *mesh = mesh_owner.get_or_null(p_mesh);
 	ERR_FAIL_COND(!mesh);
 	mesh->custom_aabb = p_aabb;
+
+	mesh->dependency.changed_notify(Dependency::DEPENDENCY_CHANGED_AABB);
 }
 
 AABB MeshStorage::mesh_get_custom_aabb(RID p_mesh) const {
@@ -1833,8 +1835,12 @@ void MeshStorage::multimesh_set_visible_instances(RID p_multimesh, int p_visible
 	}
 
 	if (multimesh->data_cache.size()) {
-		//there is a data cache..
+		// There is a data cache, but we may need to update some sections.
 		_multimesh_mark_all_dirty(multimesh, false, true);
+		int start = multimesh->visible_instances >= 0 ? multimesh->visible_instances : multimesh->instances;
+		for (int i = start; i < p_visible; i++) {
+			_multimesh_mark_dirty(multimesh, i, true);
+		}
 	}
 
 	multimesh->visible_instances = p_visible;
@@ -1866,7 +1872,7 @@ void MeshStorage::_update_dirty_multimeshes() {
 				if (multimesh->data_cache_used_dirty_regions > 32 || multimesh->data_cache_used_dirty_regions > visible_region_count / 2) {
 					// If there too many dirty regions, or represent the majority of regions, just copy all, else transfer cost piles up too much
 					glBindBuffer(GL_ARRAY_BUFFER, multimesh->buffer);
-					glBufferData(GL_ARRAY_BUFFER, MIN(visible_region_count * region_size, multimesh->instances * multimesh->stride_cache * sizeof(float)), data, GL_STATIC_DRAW);
+					glBufferSubData(GL_ARRAY_BUFFER, 0, MIN(visible_region_count * region_size, multimesh->instances * multimesh->stride_cache * sizeof(float)), data);
 					glBindBuffer(GL_ARRAY_BUFFER, 0);
 				} else {
 					// Not that many regions? update them all