Implement re-ordering for PLS atomic draws

Atomic mode needs a barrier between overlapping draws. Implement re-ordering of non-overlapping draws to minimize barriers.

Add a massively SIMD "IntersectionBoard" class that assigns a minimal "groupIndex" to each draw based on its bounding box. Sort draws based on groupIndex, plus lower priority items like draw type and texture hash. Issue draws in the new order with barriers between different groupIndex values.

Discard draws ahead of time that are empty or offscreen. (IntersectionBoard doesn't support these.)

Now that the draw logic is getting more complex, store PLSDraws in a smart pointer that guarantees we never miss a call to releaseRefs().

New simd features:
  - reduce_add, reduce_min, reduce_max, reduce_and, reduce_or
  - >2 arguments to join
  - zip()
  - more typedefs

Diffs=
d67aeac4d Implement re-ordering for PLS atomic draws (#6417)

Co-authored-by: Chris Dalton <[email protected]>

.rive_head

@@ -1 +1 @@
-a3788ed8ad14d8046faec27d34c9483ddb7d7cb3
+d67aeac4d7643cdb60f7d22c9b2c0005e7838ed9

include/rive/math/aabb.hpp

@@ -2,7 +2,6 @@
 #define _RIVE_AABB_HPP_
 
 #include "rive/span.hpp"
-#include "rive/math/mat2d.hpp"
 #include "rive/math/vec2d.hpp"
 #include <cstddef>
 #include <limits>
@@ -67,6 +66,7 @@ class AABB
     AABB offset(float dx, float dy) const { return {minX + dx, minY + dy, maxX + dx, maxY + dy}; }
 
     IAABB round() const;
+    IAABB roundOut() const; // Rounds out to integer bounds that fully contain the rectangle.
 
     ///
     /// Initialize an AABB to values that represent an invalid/collapsed

include/rive/math/mat2d.hpp

@@ -1,6 +1,7 @@
 #ifndef _RIVE_MAT2D_HPP_
 #define _RIVE_MAT2D_HPP_
 
+#include "rive/math/aabb.hpp"
 #include "rive/math/vec2d.hpp"
 #include <array>
 #include <cstddef>
@@ -44,6 +45,11 @@ class Mat2D
     // Sets dst[i] = M * pts[i] for i in 0..n-1.
     void mapPoints(Vec2D dst[], const Vec2D pts[], size_t n) const;
 
+    // Computes a bounding box that would tightly contain the given points if they were to all be
+    // transformed by this matrix.
+    AABB mapBoundingBox(const Vec2D pts[], size_t n) const;
+    AABB mapBoundingBox(const AABB&) const;
+
     // If returns true, result holds the inverse.
     // If returns false, result is unchnaged.
     bool invert(Mat2D* result) const;

include/rive/math/math_types.hpp

@@ -16,6 +16,7 @@ namespace rive
 namespace math
 {
 constexpr float PI = 3.14159265f;
+constexpr float SQRT2 = 1.41421356f;
 constexpr float EPSILON = 1.f / (1 << 12); // Common threshold for detecting values near zero.
 
 RIVE_MAYBE_UNUSED inline bool nearly_zero(float a, float tolerance = EPSILON)