Torque3D/Engine/source/math/isa/neon/float4.cpp

#include "float4_dispatch.h"
#include <arm_neon.h> // NEON intrinsics

namespace
{
   typedef float32x4_t f32x4;

   // Load 4 floats from memory into a SIMD register
   inline f32x4 v_load(const float* p) { return vld1q_f32(p); }

   // Store 4 floats from SIMD register back to memory
   inline void v_store(float* dst, f32x4 v) { vst1q_f32(dst, v); }

   // Broadcast a single float across all 4 lanes
   inline f32x4 v_set1(float s) { return vdupq_n_f32(s); }

   // Element-wise multiply
   inline f32x4 v_mul(f32x4 a, f32x4 b) { return vmulq_f32(a, b); }

   // Element-wise divide (approximate fast reciprocal)
   inline f32x4 v_div(f32x4 a, f32x4 b)
   {
      float32x4_t rcp = vrecpeq_f32(b);
      // Refine reciprocal for better precision
      rcp = vmulq_f32(vrecpsq_f32(b, rcp), rcp);
      return vmulq_f32(a, rcp);
   }

   // Element-wise add
   inline f32x4 v_add(f32x4 a, f32x4 b) { return vaddq_f32(a, b); }

   // Element-wise subtract
   inline f32x4 v_sub(f32x4 a, f32x4 b) { return vsubq_f32(a, b); }

   // Horizontal sum of all 4 elements (for dot product, length, etc.)
   inline float v_hadd4(f32x4 a)
   {
      float32x2_t sum_pair = vadd_f32(vget_low_f32(a), vget_high_f32(a)); // add pairs [a0+a2, a1+a3]
      float32x2_t sum = vpadd_f32(sum_pair, sum_pair);                    // horizontal add: total sum
      return vget_lane_f32(sum, 0);
   }

   // Optimized cross product for float4 (w component preserved)
   inline f32x4 v_cross(f32x4 a, f32x4 b)
   {
      // Extract xyz as separate registers
      float32x4_t a_yzx = vextq_f32(a, a, 1); // rotate left: y,z,x,w
      float32x4_t b_yzx = vextq_f32(b, b, 1);
      
      float32x4_t mul1 = vmulq_f32(a, b_yzx);
      float32x4_t mul2 = vmulq_f32(a_yzx, b);
      
      float32x4_t c = vsubq_f32(mul1, mul2);
      
      // Rotate back to x,y,z and keep w from original 'a'
      float32x4_t xyz = vextq_f32(c, c, 3);   // x,y,z in lanes 0..2
      float32x4_t result = vsetq_lane_f32(vgetq_lane_f32(a, 3), xyz, 3); // preserve w
      return result;
   }

}

#include "float4_impl.inl"

namespace math_backend::float4::dispatch
{
    // Install NEON64 backend
    void install_neon()
    {
        gFloat4.add          = float4_add_impl;
        gFloat4.sub          = float4_sub_impl;
        gFloat4.mul          = float4_mul_impl;
        gFloat4.mul_scalar   = float4_mul_scalar_impl;
        gFloat4.div          = float4_div_impl;
        gFloat4.div_scalar   = float4_div_scalar_impl;
        gFloat4.dot          = float4_dot_impl;
        gFloat4.length       = float4_length_impl;
        gFloat4.lengthSquared = float4_length_squared_impl;
        gFloat4.normalize    = float4_normalize_impl;
        gFloat4.normalize_mag = float4_normalize_mag_impl;
        gFloat4.lerp         = float4_lerp_impl;
        gFloat4.cross        = float4_cross_impl;
    }
}