Improve NavigationLink3DGizmoPlugin::redraw performance

[kiroxas]

Reduced sin/cos calls by 4x.
Correctly resize the vector to only have one allocation.

Spotted during the code review here.
Did not increase the number of points for the circle, as the bottleneck mentionned in the PR by @smix8 may be on the rendering side (too many points).

[Ivorforce]

You are making a mistake where you are treating resize as reserve. If you use resize you'll need to set elements in-place instead of calling push_back. Otherwise, you'll end up with twice the elements :)

[smix8]

If we resize() instead of reserve() could as well use a ptrw() instead of push_back() which should also be slightly faster with the lines CoW Vector. Looking at the debug render PR I did that there as well so it would fit with the update.

[kiroxas]

You are making a mistake where you are treating resize as reserve. If you use resize you'll need to set elements in-place instead of calling push_back. Otherwise, you'll end up with twice the elements :)

I cannot find a reserve in Vector though. :(. I will manually add in place.

[kiroxas]

If we resize() instead of reserve() could as well use a ptrw() instead of push_back() which should also be slightly faster with the lines CoW Vector. Looking at the debug render PR I did that there as well so it would fit with the update.

I will do that

[Calinou]

Note that many other gizmos use similar (but not identical) circle-drawing code. If you search for (float)i across the whole codebase, you'll find most of them:

editor/plugins/gizmos/audio_stream_player_3d_gizmo_plugin.cpp:
  88  	for (int i = 0; i < 180; i++) {
  89: 		float a = Math::deg_to_rad((float)i);
  90  		float an = Math::deg_to_rad((float)(i + 1));

editor/plugins/gizmos/collision_shape_3d_gizmo_plugin.cpp:
  364  		for (int i = 0; i <= 360; i++) {
  365: 			float ra = Math::deg_to_rad((float)i);
  366: 			float rb = Math::deg_to_rad((float)i + 1);
  367  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

  430  		for (int i = 0; i < 360; i++) {
  431: 			float ra = Math::deg_to_rad((float)i);
  432: 			float rb = Math::deg_to_rad((float)i + 1);
  433  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * radius;

  501  		for (int i = 0; i < 360; i++) {
  502: 			float ra = Math::deg_to_rad((float)i);
  503: 			float rb = Math::deg_to_rad((float)i + 1);
  504  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * radius;

editor/plugins/gizmos/gpu_particles_collision_3d_gizmo_plugin.cpp:
  175  		for (int i = 0; i <= 360; i++) {
  176: 			float ra = Math::deg_to_rad((float)i);
  177: 			float rb = Math::deg_to_rad((float)i + 1);
  178  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

editor/plugins/gizmos/joint_3d_gizmo_plugin.cpp:
  243  	for (int i = 0; i < 360; i += 10) {
  244: 		float ra = Math::deg_to_rad((float)i);
  245: 		float rb = Math::deg_to_rad((float)i + 10);
  246  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * w;

  265  	for (int i = 0; i < int(ts); i += 5) {
  266: 		float ra = Math::deg_to_rad((float)i);
  267: 		float rb = Math::deg_to_rad((float)i + 5);
  268  		float c = i / 720.0;

editor/plugins/gizmos/vehicle_body_3d_gizmo_plugin.cpp:
  61  	for (int i = 0; i <= 360; i += skip) {
  62: 		float ra = Math::deg_to_rad((float)i);
  63: 		float rb = Math::deg_to_rad((float)i + skip);
  64  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

Additionally, some primitive mesh generation code uses it too:

scene/resources/3d/capsule_shape_3d.cpp:
  43  	for (int i = 0; i < 360; i++) {
  44: 		float ra = Math::deg_to_rad((float)i);
  45: 		float rb = Math::deg_to_rad((float)i + 1);
  46  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * c_radius;

scene/resources/3d/cylinder_shape_3d.cpp:
  43  	for (int i = 0; i < 360; i++) {
  44: 		float ra = Math::deg_to_rad((float)i);
  45: 		float rb = Math::deg_to_rad((float)i + 1);
  46  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * c_radius;

scene/resources/3d/sphere_shape_3d.cpp:
  41  	for (int i = 0; i <= 360; i++) {
  42: 		float ra = Math::deg_to_rad((float)i);
  43: 		float rb = Math::deg_to_rad((float)i + 1);
  44  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

Could you check if these could benefit from similar optimizations?

[kiroxas]

Note that many other gizmos use similar (but not identical) circle-drawing code. If you search for (float)i across the whole codebase, you'll find most of them:

editor/plugins/gizmos/audio_stream_player_3d_gizmo_plugin.cpp:
  88  	for (int i = 0; i < 180; i++) {
  89: 		float a = Math::deg_to_rad((float)i);
  90  		float an = Math::deg_to_rad((float)(i + 1));

editor/plugins/gizmos/collision_shape_3d_gizmo_plugin.cpp:
  364  		for (int i = 0; i <= 360; i++) {
  365: 			float ra = Math::deg_to_rad((float)i);
  366: 			float rb = Math::deg_to_rad((float)i + 1);
  367  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

  430  		for (int i = 0; i < 360; i++) {
  431: 			float ra = Math::deg_to_rad((float)i);
  432: 			float rb = Math::deg_to_rad((float)i + 1);
  433  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * radius;

  501  		for (int i = 0; i < 360; i++) {
  502: 			float ra = Math::deg_to_rad((float)i);
  503: 			float rb = Math::deg_to_rad((float)i + 1);
  504  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * radius;

editor/plugins/gizmos/gpu_particles_collision_3d_gizmo_plugin.cpp:
  175  		for (int i = 0; i <= 360; i++) {
  176: 			float ra = Math::deg_to_rad((float)i);
  177: 			float rb = Math::deg_to_rad((float)i + 1);
  178  			Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

editor/plugins/gizmos/joint_3d_gizmo_plugin.cpp:
  243  	for (int i = 0; i < 360; i += 10) {
  244: 		float ra = Math::deg_to_rad((float)i);
  245: 		float rb = Math::deg_to_rad((float)i + 10);
  246  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * w;

  265  	for (int i = 0; i < int(ts); i += 5) {
  266: 		float ra = Math::deg_to_rad((float)i);
  267: 		float rb = Math::deg_to_rad((float)i + 5);
  268  		float c = i / 720.0;

editor/plugins/gizmos/vehicle_body_3d_gizmo_plugin.cpp:
  61  	for (int i = 0; i <= 360; i += skip) {
  62: 		float ra = Math::deg_to_rad((float)i);
  63: 		float rb = Math::deg_to_rad((float)i + skip);
  64  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

Additionally, some primitive mesh generation code uses it too:

scene/resources/3d/capsule_shape_3d.cpp:
  43  	for (int i = 0; i < 360; i++) {
  44: 		float ra = Math::deg_to_rad((float)i);
  45: 		float rb = Math::deg_to_rad((float)i + 1);
  46  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * c_radius;

scene/resources/3d/cylinder_shape_3d.cpp:
  43  	for (int i = 0; i < 360; i++) {
  44: 		float ra = Math::deg_to_rad((float)i);
  45: 		float rb = Math::deg_to_rad((float)i + 1);
  46  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * c_radius;

scene/resources/3d/sphere_shape_3d.cpp:
  41  	for (int i = 0; i <= 360; i++) {
  42: 		float ra = Math::deg_to_rad((float)i);
  43: 		float rb = Math::deg_to_rad((float)i + 1);
  44  		Point2 a = Vector2(Math::sin(ra), Math::cos(ra)) * r;

Could you check if these could benefit from similar optimizations?

I definitely can, but the hardest part is checking how it looks ( to check I did not break anything ) and I have no idea where to look for most of them.

[kiroxas]

Ok, I found a way to speed it up greatly again. We're now down from inital 320 cos/sin calls to 80 to now ... 8.
This is probably fast enough now, the limiting code is probably the copy into the gizmo now.

[kiroxas]

Reduced the cos use again to only 4.
Reduced from 19 400 cycles to 4100 (on average after 3000 iterations) now. So bewteen 4-5x speed up. (Measured only the output vertices of the two circles to the vector, not the whole function)

[Ivorforce]

Your improvement is impressive! I'd still love to see a final benchmark, but it's pretty clear to me this will be substantially faster than the old implementation.
I have some style comments that should probably be addressed before a merge, but generally I think it's ready (assuming you tested it).

[kiroxas]

Your improvement is impressive! I'd still love to see a final benchmark, but it's pretty clear to me this will be substantially faster than the old implementation. I have some style comments that should probably be addressed before a merge, but generally I think it's ready (assuming you tested it).

I did, here :

Reduced from 19 400 cycles to 4100 (on average after 3000 iterations) now. So bewteen 4-5x speed up. (Measured only the output vertices of the two circles to the vector, not the whole function)

[akien-mga]

Thanks!