Metal: GPU_PRIM_LINE_LOOP alternative implementations.

Prefer using immVertex3f when 3D shaders are used for 2D rendering due to overhead of vertex padding in hardware. CPU overhead is negligible.

Authored by Apple: Michael Parkin-White

Ref T96261

Reviewed By: fclem

Maniphest Tasks: T96261

Differential Revision: https://developer.blender.org/D14494

8 files changed, 282 insertions, 76 deletions

diff --git a/source/blender/blenlib/intern/math_base_inline.c b/source/blender/blenlib/intern/math_base_inline.c
index bb6bc0db00d..a983821f15e 100644
--- a/source/blender/blenlib/intern/math_base_inline.c
+++ b/source/blender/blenlib/intern/math_base_inline.c
@@ -524,6 +524,15 @@ MINLINE uint max_uu(uint a, uint b)
   return (b < a) ? a : b;
 }
 
+MINLINE unsigned long long min_ulul(unsigned long long a, unsigned long long b)
+{
+  return (a < b) ? a : b;
+}
+MINLINE unsigned long long max_ulul(unsigned long long a, unsigned long long b)
+{
+  return (b < a) ? a : b;
+}
+
 MINLINE float min_fff(float a, float b, float c)
 {
   return min_ff(min_ff(a, b), c);
diff --git a/source/blender/editors/gizmo_library/gizmo_draw_utils.c b/source/blender/editors/gizmo_library/gizmo_draw_utils.c
index 8d2aec18451..c6303c197e7 100644
--- a/source/blender/editors/gizmo_library/gizmo_draw_utils.c
+++ b/source/blender/editors/gizmo_library/gizmo_draw_utils.c
@@ -82,9 +82,35 @@ void wm_gizmo_vec_draw(
     const float color[4], const float (*verts)[3], uint vert_count, uint pos, uint primitive_type)
 {
   immUniformColor4fv(color);
-  immBegin(primitive_type, vert_count);
-  for (int i = 0; i < vert_count; i++) {
-    immVertex3fv(pos, verts[i]);
+
+  if (primitive_type == GPU_PRIM_LINE_LOOP) {
+    /* Line loop alternative for Metal/Vulkan. */
+    immBegin(GPU_PRIM_LINES, vert_count * 2);
+    immVertex3fv(pos, verts[0]);
+    for (int i = 1; i < vert_count; i++) {
+      immVertex3fv(pos, verts[i]);
+      immVertex3fv(pos, verts[i]);
+    }
+    immVertex3fv(pos, verts[0]);
+    immEnd();
+  }
+  else if (primitive_type == GPU_PRIM_TRI_FAN) {
+    /* Note(Metal): Tri-fan alternative for Metal. Triangle List is more efficient for small
+     * primitive counts. */
+    int tri_count = vert_count - 2;
+    immBegin(GPU_PRIM_TRIS, tri_count * 3);
+    for (int i = 0; i < tri_count; i++) {
+      immVertex3fv(pos, verts[0]);
+      immVertex3fv(pos, verts[i + 1]);
+      immVertex3fv(pos, verts[i + 2]);
+    }
+    immEnd();
+  }
+  else {
+    immBegin(primitive_type, vert_count);
+    for (int i = 0; i < vert_count; i++) {
+      immVertex3fv(pos, verts[i]);
+    }
+    immEnd();
   }
-  immEnd();
 }
diff --git a/source/blender/editors/gizmo_library/gizmo_types/button2d_gizmo.c b/source/blender/editors/gizmo_library/gizmo_types/button2d_gizmo.c
index 5e20cc73f1a..b326d6d1859 100644
--- a/source/blender/editors/gizmo_library/gizmo_types/button2d_gizmo.c
+++ b/source/blender/editors/gizmo_library/gizmo_types/button2d_gizmo.c
@@ -74,20 +74,21 @@ static void button2d_geom_draw_backdrop(const wmGizmo *gz,
   GPU_viewport_size_get_f(viewport);
 
   GPUVertFormat *format = immVertexFormat();
-  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT);
+  /* Note(Metal): Prefer 3D coordinate for 2D rendering when using 3D shader. */
+  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 3, GPU_FETCH_FLOAT);
 
   /* TODO: other draw styles. */
   if (color[3] == 1.0 && fill_alpha == 1.0 && select == false) {
     immBindBuiltinProgram(GPU_SHADER_3D_UNIFORM_COLOR);
     immUniformColor4fv(color);
-    imm_draw_circle_fill_2d(pos, 0, 0, 1.0f, CIRCLE_RESOLUTION);
+    imm_draw_circle_fill_3d(pos, 0.0f, 0.0f, 1.0f, CIRCLE_RESOLUTION);
     immUnbindProgram();
 
     immBindBuiltinProgram(GPU_SHADER_3D_POLYLINE_UNIFORM_COLOR);
     immUniform2fv("viewportSize", &viewport[2]);
     immUniform1f("lineWidth", gz->line_width * U.pixelsize);
     immUniformColor4fv(color);
-    imm_draw_circle_wire_2d(pos, 0, 0, 1.0f, CIRCLE_RESOLUTION);
+    imm_draw_circle_wire_3d(pos, 0.0f, 0.0f, 1.0f, CIRCLE_RESOLUTION);
     immUnbindProgram();
   }
   else {
@@ -96,7 +97,7 @@ static void button2d_geom_draw_backdrop(const wmGizmo *gz,
       const float fill_color[4] = {UNPACK3(color), fill_alpha * color[3]};
       immBindBuiltinProgram(GPU_SHADER_3D_UNIFORM_COLOR);
       immUniformColor4fv(fill_color);
-      imm_draw_circle_fill_2d(pos, 0, 0, 1.0f, CIRCLE_RESOLUTION);
+      imm_draw_circle_fill_3d(pos, 0.0f, 0.0f, 1.0f, CIRCLE_RESOLUTION);
       immUnbindProgram();
     }
 
@@ -106,7 +107,7 @@ static void button2d_geom_draw_backdrop(const wmGizmo *gz,
       immUniform2fv("viewportSize", &viewport[2]);
       immUniform1f("lineWidth", gz->line_width * U.pixelsize);
       immUniformColor4fv(color);
-      imm_draw_circle_wire_2d(pos, 0, 0, 1.0f, CIRCLE_RESOLUTION);
+      imm_draw_circle_wire_3d(pos, 0.0f, 0.0f, 1.0f, CIRCLE_RESOLUTION);
       immUnbindProgram();
     }
   }
diff --git a/source/blender/editors/gizmo_library/gizmo_types/cage2d_gizmo.c b/source/blender/editors/gizmo_library/gizmo_types/cage2d_gizmo.c
index 4c54aa10c33..e4cb6d149f5 100644
--- a/source/blender/editors/gizmo_library/gizmo_types/cage2d_gizmo.c
+++ b/source/blender/editors/gizmo_library/gizmo_types/cage2d_gizmo.c
@@ -99,7 +99,8 @@ static void cage2d_draw_box_corners(const rctf *r,
                                     const float color[3],
                                     const float line_width)
 {
-  uint pos = GPU_vertformat_attr_add(immVertexFormat(), "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT);
+  /* Note(Metal): Prefer using 3D coordinates with 3D shader, even if rendering 2D gizmo's. */
+  uint pos = GPU_vertformat_attr_add(immVertexFormat(), "pos", GPU_COMP_F32, 3, GPU_FETCH_FLOAT);
 
   immBindBuiltinProgram(GPU_SHADER_3D_POLYLINE_UNIFORM_COLOR);
   immUniformColor3fv(color);
@@ -112,25 +113,25 @@ static void cage2d_draw_box_corners(const rctf *r,
 
   immBegin(GPU_PRIM_LINES, 16);
 
-  immVertex2f(pos, r->xmin, r->ymin + margin[1]);
-  immVertex2f(pos, r->xmin, r->ymin);
-  immVertex2f(pos, r->xmin, r->ymin);
-  immVertex2f(pos, r->xmin + margin[0], r->ymin);
+  immVertex3f(pos, r->xmin, r->ymin + margin[1], 0.0f);
+  immVertex3f(pos, r->xmin, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmin, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmin + margin[0], r->ymin, 0.0f);
 
-  immVertex2f(pos, r->xmax, r->ymin + margin[1]);
-  immVertex2f(pos, r->xmax, r->ymin);
-  immVertex2f(pos, r->xmax, r->ymin);
-  immVertex2f(pos, r->xmax - margin[0], r->ymin);
+  immVertex3f(pos, r->xmax, r->ymin + margin[1], 0.0f);
+  immVertex3f(pos, r->xmax, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmax, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmax - margin[0], r->ymin, 0.0f);
 
-  immVertex2f(pos, r->xmax, r->ymax - margin[1]);
-  immVertex2f(pos, r->xmax, r->ymax);
-  immVertex2f(pos, r->xmax, r->ymax);
-  immVertex2f(pos, r->xmax - margin[0], r->ymax);
+  immVertex3f(pos, r->xmax, r->ymax - margin[1], 0.0f);
+  immVertex3f(pos, r->xmax, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmax, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmax - margin[0], r->ymax, 0.0f);
 
-  immVertex2f(pos, r->xmin, r->ymax - margin[1]);
-  immVertex2f(pos, r->xmin, r->ymax);
-  immVertex2f(pos, r->xmin, r->ymax);
-  immVertex2f(pos, r->xmin + margin[0], r->ymax);
+  immVertex3f(pos, r->xmin, r->ymax - margin[1], 0.0f);
+  immVertex3f(pos, r->xmin, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmin, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmin + margin[0], r->ymax, 0.0f);
 
   immEnd();
 
@@ -440,12 +441,35 @@ static void cage2d_draw_box_interaction(const float color[4],
 static void imm_draw_point_aspect_2d(
     uint pos, float x, float y, float rad_x, float rad_y, bool solid)
 {
-  immBegin(solid ? GPU_PRIM_TRI_FAN : GPU_PRIM_LINE_LOOP, 4);
-  immVertex2f(pos, x - rad_x, y - rad_y);
-  immVertex2f(pos, x - rad_x, y + rad_y);
-  immVertex2f(pos, x + rad_x, y + rad_y);
-  immVertex2f(pos, x + rad_x, y - rad_y);
-  immEnd();
+  if (solid) {
+    /* Note(Metal/AMD): Small Triangle-list primitives more optimal for GPU HW than Trianglestrip.
+     */
+    immBegin(GPU_PRIM_TRIS, 6);
+    immVertex2f(pos, x - rad_x, y - rad_y);
+    immVertex2f(pos, x - rad_x, y + rad_y);
+    immVertex2f(pos, x + rad_x, y + rad_y);
+
+    immVertex2f(pos, x - rad_x, y - rad_y);
+    immVertex2f(pos, x + rad_x, y + rad_y);
+    immVertex2f(pos, x + rad_x, y - rad_y);
+    immEnd();
+  }
+  else {
+    /* Note(Metal/AMD): Small Line-list primitives more optimal for GPU HW than Linestrip. */
+    immBegin(GPU_PRIM_LINES, 8);
+    immVertex2f(pos, x - rad_x, y - rad_y);
+    immVertex2f(pos, x - rad_x, y + rad_y);
+
+    immVertex2f(pos, x - rad_x, y + rad_y);
+    immVertex2f(pos, x + rad_x, y + rad_y);
+
+    immVertex2f(pos, x + rad_x, y + rad_y);
+    immVertex2f(pos, x + rad_x, y - rad_y);
+
+    immVertex2f(pos, x + rad_x, y - rad_y);
+    immVertex2f(pos, x - rad_x, y - rad_y);
+    immEnd();
+  }
 }
 
 static void cage2d_draw_circle_wire(const rctf *r,
@@ -455,7 +479,9 @@ static void cage2d_draw_circle_wire(const rctf *r,
                                     const int draw_options,
                                     const float line_width)
 {
-  uint pos = GPU_vertformat_attr_add(immVertexFormat(), "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT);
+  /* Note(Metal): Prefer using 3D coordinates with 3D shader input, even if rendering 2D gizmo's.
+   */
+  uint pos = GPU_vertformat_attr_add(immVertexFormat(), "pos", GPU_COMP_F32, 3, GPU_FETCH_FLOAT);
 
   immBindBuiltinProgram(GPU_SHADER_3D_POLYLINE_UNIFORM_COLOR);
   immUniformColor3fv(color);
@@ -465,17 +491,28 @@ static void cage2d_draw_circle_wire(const rctf *r,
   immUniform2fv("viewportSize", &viewport[2]);
   immUniform1f("lineWidth", line_width * U.pixelsize);
 
-  immBegin(GPU_PRIM_LINE_LOOP, 4);
-  immVertex2f(pos, r->xmin, r->ymin);
-  immVertex2f(pos, r->xmax, r->ymin);
-  immVertex2f(pos, r->xmax, r->ymax);
-  immVertex2f(pos, r->xmin, r->ymax);
+  /* Small 'lines' primitives more efficient for hardware processing than linestrip. */
+  immBegin(GPU_PRIM_LINES, 8);
+  immVertex3f(pos, r->xmin, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmax, r->ymin, 0.0f);
+
+  immVertex3f(pos, r->xmax, r->ymin, 0.0f);
+  immVertex3f(pos, r->xmax, r->ymax, 0.0f);
+
+  immVertex3f(pos, r->xmax, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmin, r->ymax, 0.0f);
+
+  immVertex3f(pos, r->xmin, r->ymax, 0.0f);
+  immVertex3f(pos, r->xmin, r->ymin, 0.0f);
   immEnd();
 
   if (transform_flag & ED_GIZMO_CAGE2D_XFORM_FLAG_ROTATE) {
-    immBegin(GPU_PRIM_LINE_LOOP, 2);
-    immVertex2f(pos, BLI_rctf_cent_x(r), r->ymax);
-    immVertex2f(pos, BLI_rctf_cent_x(r), r->ymax + margin[1]);
+    immBegin(GPU_PRIM_LINES, 4);
+    immVertex3f(pos, BLI_rctf_cent_x(r), r->ymax, 0.0f);
+    immVertex3f(pos, BLI_rctf_cent_x(r), r->ymax + margin[1], 0.0f);
+
+    immVertex3f(pos, BLI_rctf_cent_x(r), r->ymax + margin[1], 0.0f);
+    immVertex3f(pos, BLI_rctf_cent_x(r), r->ymax, 0.0f);
     immEnd();
   }
 
@@ -485,10 +522,10 @@ static void cage2d_draw_circle_wire(const rctf *r,
       const float center[2] = {BLI_rctf_cent_x(r), BLI_rctf_cent_y(r)};
 
       immBegin(GPU_PRIM_LINES, 4);
-      immVertex2f(pos, center[0] - rad[0], center[1] - rad[1]);
-      immVertex2f(pos, center[0] + rad[0], center[1] + rad[1]);
-      immVertex2f(pos, center[0] + rad[0], center[1] - rad[1]);
-      immVertex2f(pos, center[0] - rad[0], center[1] + rad[1]);
+      immVertex3f(pos, center[0] - rad[0], center[1] - rad[1], 0.0f);
+      immVertex3f(pos, center[0] + rad[0], center[1] + rad[1], 0.0f);
+      immVertex3f(pos, center[0] + rad[0], center[1] - rad[1], 0.0f);
+      immVertex3f(pos, center[0] - rad[0], center[1] + rad[1], 0.0f);
       immEnd();
     }
   }
diff --git a/source/blender/editors/gizmo_library/gizmo_types/dial3d_gizmo.c b/source/blender/editors/gizmo_library/gizmo_types/dial3d_gizmo.c
index e1f584bf9e4..a76242404ba 100644
--- a/source/blender/editors/gizmo_library/gizmo_types/dial3d_gizmo.c
+++ b/source/blender/editors/gizmo_library/gizmo_types/dial3d_gizmo.c
@@ -96,7 +96,8 @@ static void dial_geom_draw(const float color[4],
                                                  ED_GIZMO_DIAL_DRAW_FLAG_FILL)));
 
   GPUVertFormat *format = immVertexFormat();
-  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT);
+  /* Note(Metal): Prefer using 3D coordinates with 3D shader, even if rendering 2D gizmo's. */
+  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 3, GPU_FETCH_FLOAT);
 
   if (clip_plane) {
     immBindBuiltinProgram(filled ? GPU_SHADER_3D_CLIPPED_UNIFORM_COLOR :
@@ -114,18 +115,19 @@ static void dial_geom_draw(const float color[4],
   if (filled) {
     if (arc_partial_angle == 0.0f) {
       if (arc_inner_factor == 0.0f) {
-        imm_draw_circle_fill_2d(pos, 0, 0, 1.0, DIAL_RESOLUTION);
+        imm_draw_circle_fill_3d(pos, 0.0f, 0.0f, 1.0f, DIAL_RESOLUTION);
       }
       else {
-        imm_draw_disk_partial_fill_2d(
-            pos, 0, 0, arc_inner_factor, 1.0f, DIAL_RESOLUTION, 0, RAD2DEGF(M_PI * 2));
+        imm_draw_disk_partial_fill_3d(
+            pos, 0.0f, 0.0f, 0.0f, arc_inner_factor, 1.0f, DIAL_RESOLUTION, 0, RAD2DEGF(M_PI * 2));
       }
     }
     else {
       float arc_partial_deg = RAD2DEGF((M_PI * 2) - arc_partial_angle);
-      imm_draw_disk_partial_fill_2d(pos,
-                                    0,
-                                    0,
+      imm_draw_disk_partial_fill_3d(pos,
+                                    0.0f,
+                                    0.0f,
+                                    0.0f,
                                     arc_inner_factor,
                                     1.0f,
                                     DIAL_RESOLUTION,
@@ -140,15 +142,15 @@ static void dial_geom_draw(const float color[4],
     immUniform1f("lineWidth", line_width * U.pixelsize);
 
     if (arc_partial_angle == 0.0f) {
-      imm_draw_circle_wire_2d(pos, 0, 0, 1.0, DIAL_RESOLUTION);
+      imm_draw_circle_wire_3d(pos, 0.0f, 0.0f, 1.0f, DIAL_RESOLUTION);
       if (arc_inner_factor != 0.0f) {
-        imm_draw_circle_wire_2d(pos, 0, 0, arc_inner_factor, DIAL_RESOLUTION);
+        imm_draw_circle_wire_3d(pos, 0.0f, 0.0f, arc_inner_factor, DIAL_RESOLUTION);
       }
     }
     else {
       float arc_partial_deg = RAD2DEGF((M_PI * 2) - arc_partial_angle);
-      imm_draw_circle_partial_wire_2d(
-          pos, 0, 0, 1.0, DIAL_RESOLUTION, -arc_partial_deg / 2, arc_partial_deg);
+      imm_draw_circle_partial_wire_3d(
+          pos, 0.0f, 0.0f, 0.0f, 1.0f, DIAL_RESOLUTION, -arc_partial_deg / 2, arc_partial_deg);
 #  if 0
       if (arc_inner_factor != 0.0f) {
         BLI_assert(0);
@@ -186,7 +188,7 @@ static void dial_ghostarc_draw_helpline(const float angle,
   immUniformColor4fv(color);
 
   immBegin(GPU_PRIM_LINE_STRIP, 2);
-  immVertex3f(pos, 0.0f, 0, 0.0f);
+  immVertex3f(pos, 0.0f, 0.0f, 0.0f);
   immVertex3fv(pos, co_outer);
   immEnd();
 
diff --git a/source/blender/editors/gizmo_library/gizmo_types/move3d_gizmo.c b/source/blender/editors/gizmo_library/gizmo_types/move3d_gizmo.c
index 447fe1005a1..5fb1173521a 100644
--- a/source/blender/editors/gizmo_library/gizmo_types/move3d_gizmo.c
+++ b/source/blender/editors/gizmo_library/gizmo_types/move3d_gizmo.c
@@ -98,7 +98,8 @@ static void move_geom_draw(const wmGizmo *gz,
                                                  ED_GIZMO_MOVE_DRAW_FLAG_FILL)));
 
   GPUVertFormat *format = immVertexFormat();
-  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT);
+  /* Note(Metal): Prefer using 3D coordinates with 3D shader, even if rendering 2D gizmo's. */
+  uint pos = GPU_vertformat_attr_add(format, "pos", GPU_COMP_F32, 3, GPU_FETCH_FLOAT);
 
   immBindBuiltinProgram(filled ? GPU_SHADER_3D_UNIFORM_COLOR :
                                  GPU_SHADER_3D_POLYLINE_UNIFORM_COLOR);
@@ -115,20 +116,20 @@ static void move_geom_draw(const wmGizmo *gz,
 
   if (draw_style == ED_GIZMO_MOVE_STYLE_RING_2D) {
     if (filled) {
-      imm_draw_circle_fill_2d(pos, 0, 0, radius, DIAL_RESOLUTION);
+      imm_draw_circle_fill_3d(pos, 0.0f, 0.0f, radius, DIAL_RESOLUTION);
     }
     else {
-      imm_draw_circle_wire_2d(pos, 0, 0, radius, DIAL_RESOLUTION);
+      imm_draw_circle_wire_3d(pos, 0.0f, 0.0f, radius, DIAL_RESOLUTION);
     }
   }
   else if (draw_style == ED_GIZMO_MOVE_STYLE_CROSS_2D) {
     const float radius_diag = M_SQRT1_2 * radius;
     immBegin(GPU_PRIM_LINES, 4);
-    immVertex2f(pos, radius_diag, radius_diag);
-    immVertex2f(pos, -radius_diag, -radius_diag);
+    immVertex3f(pos, radius_diag, radius_diag, 0.0f);
+    immVertex3f(pos, -radius_diag, -radius_diag, 0.0f);
 
-    immVertex2f(pos, -radius_diag, radius_diag);
-    immVertex2f(pos, radius_diag, -radius_diag);
+    immVertex3f(pos, -radius_diag, radius_diag, 0.0f);
+    immVertex3f(pos, radius_diag, -radius_diag, 0.0f);
     immEnd();
   }
   else {
diff --git a/source/blender/gpu/GPU_immediate_util.h b/source/blender/gpu/GPU_immediate_util.h
index 924a735af79..cb7ca877b1d 100644
--- a/source/blender/gpu/GPU_immediate_util.h
+++ b/source/blender/gpu/GPU_immediate_util.h
@@ -80,6 +80,8 @@ void imm_draw_circle_fill_3d(uint pos, float x, float y, float radius, int nsegm
  */
 void imm_draw_circle_partial_wire_2d(
     uint pos, float x, float y, float radius, int nsegments, float start, float sweep);
+void imm_draw_circle_partial_wire_3d(
+    uint pos, float x, float y, float z, float radius, int nsegments, float start, float sweep);
 
 /**
  * Draw a filled arc with the given inner and outer radius.
@@ -104,6 +106,15 @@ void imm_draw_disk_partial_fill_2d(uint pos,
                                    int nsegments,
                                    float start,
                                    float sweep);
+void imm_draw_disk_partial_fill_3d(uint pos,
+                                   float x,
+                                   float y,
+                                   float z,
+                                   float rad_inner,
+                                   float rad_outer,
+                                   int nsegments,
+                                   float start,
+                                   float sweep);
 
 /**
  * Draw a lined box.
diff --git a/source/blender/gpu/intern/gpu_immediate_util.c b/source/blender/gpu/intern/gpu_immediate_util.c
index cf8837ab26e..67035853594 100644
--- a/source/blender/gpu/intern/gpu_immediate_util.c
+++ b/source/blender/gpu/intern/gpu_immediate_util.c
@@ -142,12 +142,27 @@ static void imm_draw_circle(GPUPrimType prim_type,
                             float radius_y,
                             int nsegments)
 {
-  immBegin(prim_type, nsegments);
-  for (int i = 0; i < nsegments; i++) {
-    const float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
-    immVertex2f(shdr_pos, x + (radius_x * cosf(angle)), y + (radius_y * sinf(angle)));
+  if (prim_type == GPU_PRIM_LINE_LOOP) {
+    /* Note(Metal/AMD): For small primitives, line list more efficient than line strip.. */
+    immBegin(GPU_PRIM_LINES, nsegments * 2);
+
+    immVertex2f(shdr_pos, x + (radius_x * cosf(0.0f)), y + (radius_y * sinf(0.0f)));
+    for (int i = 1; i < nsegments; i++) {
+      const float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
+      immVertex2f(shdr_pos, x + (radius_x * cosf(angle)), y + (radius_y * sinf(angle)));
+      immVertex2f(shdr_pos, x + (radius_x * cosf(angle)), y + (radius_y * sinf(angle)));
+    }
+    immVertex2f(shdr_pos, x + (radius_x * cosf(0.0f)), y + (radius_y * sinf(0.0f)));
+    immEnd();
+  }
+  else {
+    immBegin(prim_type, nsegments);
+    for (int i = 0; i < nsegments; i++) {
+      const float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
+      immVertex2f(shdr_pos, x + (radius_x * cosf(angle)), y + (radius_y * sinf(angle)));
+    }
+    immEnd();
   }
-  immEnd();
 }
 
 void imm_draw_circle_wire_2d(uint shdr_pos, float x, float y, float radius, int nsegments)
@@ -194,12 +209,42 @@ static void imm_draw_circle_partial(GPUPrimType prim_type,
   immEnd();
 }
 
+static void imm_draw_circle_partial_3d(GPUPrimType prim_type,
+                                       uint pos,
+                                       float x,
+                                       float y,
+                                       float z,
+                                       float rad,
+                                       int nsegments,
+                                       float start,
+                                       float sweep)
+{
+  /* shift & reverse angle, increase 'nsegments' to match gluPartialDisk */
+  const float angle_start = -(DEG2RADF(start)) + (float)(M_PI / 2);
+  const float angle_end = -(DEG2RADF(sweep) - angle_start);
+  nsegments += 1;
+  immBegin(prim_type, nsegments);
+  for (int i = 0; i < nsegments; i++) {
+    const float angle = interpf(angle_start, angle_end, ((float)i / (float)(nsegments - 1)));
+    const float angle_sin = sinf(angle);
+    const float angle_cos = cosf(angle);
+    immVertex3f(pos, x + rad * angle_cos, y + rad * angle_sin, z);
+  }
+  immEnd();
+}
+
 void imm_draw_circle_partial_wire_2d(
     uint pos, float x, float y, float radius, int nsegments, float start, float sweep)
 {
   imm_draw_circle_partial(GPU_PRIM_LINE_STRIP, pos, x, y, radius, nsegments, start, sweep);
 }
 
+void imm_draw_circle_partial_wire_3d(
+    uint pos, float x, float y, float z, float rad, int nsegments, float start, float sweep)
+{
+  imm_draw_circle_partial_3d(GPU_PRIM_LINE_STRIP, pos, x, y, z, rad, nsegments, start, sweep);
+}
+
 static void imm_draw_disk_partial(GPUPrimType prim_type,
                                   uint pos,
                                   float x,
@@ -229,6 +274,36 @@ static void imm_draw_disk_partial(GPUPrimType prim_type,
   immEnd();
 }
 
+static void imm_draw_disk_partial_3d(GPUPrimType prim_type,
+                                     uint pos,
+                                     float x,
+                                     float y,
+                                     float z,
+                                     float rad_inner,
+                                     float rad_outer,
+                                     int nsegments,
+                                     float start,
+                                     float sweep)
+{
+  /* to avoid artifacts */
+  const float max_angle = 3 * 360;
+  CLAMP(sweep, -max_angle, max_angle);
+
+  /* shift & reverse angle, increase 'nsegments' to match gluPartialDisk */
+  const float angle_start = -(DEG2RADF(start)) + (float)M_PI_2;
+  const float angle_end = -(DEG2RADF(sweep) - angle_start);
+  nsegments += 1;
+  immBegin(prim_type, nsegments * 2);
+  for (int i = 0; i < nsegments; i++) {
+    const float angle = interpf(angle_start, angle_end, ((float)i / (float)(nsegments - 1)));
+    const float angle_sin = sinf(angle);
+    const float angle_cos = cosf(angle);
+    immVertex3f(pos, x + rad_inner * angle_cos, y + rad_inner * angle_sin, z);
+    immVertex3f(pos, x + rad_outer * angle_cos, y + rad_outer * angle_sin, z);
+  }
+  immEnd();
+}
+
 void imm_draw_disk_partial_fill_2d(uint pos,
                                    float x,
                                    float y,
@@ -241,16 +316,44 @@ void imm_draw_disk_partial_fill_2d(uint pos,
   imm_draw_disk_partial(
       GPU_PRIM_TRI_STRIP, pos, x, y, rad_inner, rad_outer, nsegments, start, sweep);
 }
+void imm_draw_disk_partial_fill_3d(uint pos,
+                                   float x,
+                                   float y,
+                                   float z,
+                                   float rad_inner,
+                                   float rad_outer,
+                                   int nsegments,
+                                   float start,
+                                   float sweep)
+{
+  imm_draw_disk_partial_3d(
+      GPU_PRIM_TRI_STRIP, pos, x, y, z, rad_inner, rad_outer, nsegments, start, sweep);
+}
 
 static void imm_draw_circle_3D(
     GPUPrimType prim_type, uint pos, float x, float y, float radius, int nsegments)
 {
-  immBegin(prim_type, nsegments);
-  for (int i = 0; i < nsegments; i++) {
-    float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
-    immVertex3f(pos, x + radius * cosf(angle), y + radius * sinf(angle), 0.0f);
+  if (prim_type == GPU_PRIM_LINE_LOOP) {
+    /* Note(Metal/AMD): For small primitives, line list more efficient than line strip. */
+    immBegin(GPU_PRIM_LINES, nsegments * 2);
+
+    immVertex3f(pos, x + radius * cosf(0.0f), y + radius * sinf(0.0f), 0.0f);
+    for (int i = 1; i < nsegments; i++) {
+      float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
+      immVertex3f(pos, x + radius * cosf(angle), y + radius * sinf(angle), 0.0f);
+      immVertex3f(pos, x + radius * cosf(angle), y + radius * sinf(angle), 0.0f);
+    }
+    immVertex3f(pos, x + radius * cosf(0.0f), y + radius * sinf(0.0f), 0.0f);
+    immEnd();
+  }
+  else {
+    immBegin(prim_type, nsegments);
+    for (int i = 0; i < nsegments; i++) {
+      float angle = (float)(2 * M_PI) * ((float)i / (float)nsegments);
+      immVertex3f(pos, x + radius * cosf(angle), y + radius * sinf(angle), 0.0f);
+    }
+    immEnd();
   }
-  immEnd();
 }
 
 void imm_draw_circle_wire_3d(uint pos, float x, float y, float radius, int nsegments)
@@ -270,22 +373,38 @@ void imm_draw_circle_fill_3d(uint pos, float x, float y, float radius, int nsegm
 
 void imm_draw_box_wire_2d(uint pos, float x1, float y1, float x2, float y2)
 {
-  immBegin(GPU_PRIM_LINE_LOOP, 4);
+  /* Note(Metal/AMD): For small primitives, line list more efficient than line-strip. */
+  immBegin(GPU_PRIM_LINES, 8);
   immVertex2f(pos, x1, y1);
   immVertex2f(pos, x1, y2);
+
+  immVertex2f(pos, x1, y2);
   immVertex2f(pos, x2, y2);
+
+  immVertex2f(pos, x2, y2);
+  immVertex2f(pos, x2, y1);
+
   immVertex2f(pos, x2, y1);
+  immVertex2f(pos, x1, y1);
   immEnd();
 }
 
 void imm_draw_box_wire_3d(uint pos, float x1, float y1, float x2, float y2)
 {
   /* use this version when GPUVertFormat has a vec3 position */
-  immBegin(GPU_PRIM_LINE_LOOP, 4);
+  /* Note(Metal/AMD): For small primitives, line list more efficient than line-strip. */
+  immBegin(GPU_PRIM_LINES, 8);
   immVertex3f(pos, x1, y1, 0.0f);
   immVertex3f(pos, x1, y2, 0.0f);
+
+  immVertex3f(pos, x1, y2, 0.0f);
+  immVertex3f(pos, x2, y2, 0.0f);
+
   immVertex3f(pos, x2, y2, 0.0f);
   immVertex3f(pos, x2, y1, 0.0f);
+
+  immVertex3f(pos, x2, y1, 0.0f);
+  immVertex3f(pos, x1, y1, 0.0f);
   immEnd();
 }