In their excellent paper "A Cross-Platform Framework for Interactive Ray Tracing"[1] Markus Geimer and Stefan Müller propose a branchless SIMD friendly variation of the slab test. It's extremely fast but there's a catch, under some conditions (say (box_min-pos) == 0 while inv_dir = inf) a NaN is produced and due to the way SSE min/max work, it ends up with a bogus result. Here's an attempt to fix that corner case while keeping things tight. It also works with flat voxels and take around 33 cycles on a good day (compared to the 17 cycles of the original version), if you have a decent compiler (wink wink, nudge nudge).
Thierry Berger-Perrin.
PS: if you don't need the intersection points, just discard them.
PPS: scheduling is left as an exercise for the compiler.
// can you say "barebone"?
struct vec_t { float x,y,z,pad; };
struct aabb_t {
vec_t min;
vec_t max;
};
struct ray_t {
vec_t pos;
vec_t inv_dir;
};
struct ray_segment_t {
float t_near,t_far;
};
#include
#include
#include
#include "cotd.h"
#ifdef __GNUC__
#define _MM_ALIGN16 __attribute__ ((aligned (16)))
#endif
// turn those verbose intrinsics into something readable.
#define loadps(mem) _mm_load_ps((const float * const)(mem))
#define storess(ss,mem) _mm_store_ss((float * const)(mem),(ss))
#define minss _mm_min_ss
#define maxss _mm_max_ss
#define minps _mm_min_ps
#define maxps _mm_max_ps
#define mulps _mm_mul_ps
#define subps _mm_sub_ps
#define rotatelps(ps) _mm_shuffle_ps((ps),(ps), 0x39) // a,b,c,d -> b,c,d,a
#define muxhps(low,high) _mm_movehl_ps((low),(high)) // low{a,b,c,d}|high{e,f,g,h} = {c,d,g,h}
static const float flt_plus_inf = -logf(0); // let's keep C and C++ compilers happy.
static const float _MM_ALIGN16
ps_cst_plus_inf[4] = { flt_plus_inf, flt_plus_inf, flt_plus_inf, flt_plus_inf },
ps_cst_minus_inf[4] = { -flt_plus_inf, -flt_plus_inf, -flt_plus_inf, -flt_plus_inf };
static bool ray_box_intersect(const aabb_t &box, const ray_t &ray, ray_segment_t &rs) {
// you may already have those values hanging around somewhere
const __m128
plus_inf = loadps(ps_cst_plus_inf),
minus_inf = loadps(ps_cst_minus_inf);
// use whatever's apropriate to load.
const __m128
box_min = loadps(&box.min),
box_max = loadps(&box.max),
pos = loadps(&ray.pos),
inv_dir = loadps(&ray.inv_dir);
// use a div if inverted directions aren't available
const __m128 l1 = mulps(subps(box_min, pos), inv_dir);
const __m128 l2 = mulps(subps(box_max, pos), inv_dir);
// the order we use for those min/max is vital to filter out
// NaNs that happens when an inv_dir is +/- inf and
// (box_min - pos) is 0. inf * 0 = NaN
const __m128 filtered_l1a = minps(l1, plus_inf);
const __m128 filtered_l2a = minps(l2, plus_inf);
const __m128 filtered_l1b = maxps(l1, minus_inf);
const __m128 filtered_l2b = maxps(l2, minus_inf);
// now that we're back on our feet, test those slabs.
__m128 lmax = maxps(filtered_l1a, filtered_l2a);
__m128 lmin = minps(filtered_l1b, filtered_l2b);
// unfold back. try to hide the latency of the shufps & co.
const __m128 lmax0 = rotatelps(lmax);
const __m128 lmin0 = rotatelps(lmin);
lmax = minss(lmax, lmax0);
lmin = maxss(lmin, lmin0);
const __m128 lmax1 = muxhps(lmax,lmax);
const __m128 lmin1 = muxhps(lmin,lmin);
lmax = minss(lmax, lmax1);
lmin = maxss(lmin, lmin1);
const bool ret = _mm_comige_ss(lmax, _mm_setzero_ps()) & _mm_comige_ss(lmax,lmin);
storess(lmin, &rs.t_near);
storess(lmax, &rs.t_far);
return ret;
}
void checkpointcharlie() {
// let's keep things simple.
// the ray is right on the edge, aimed straight into Z
const aabb_t _MM_ALIGN16 box = { -1,-1,-1, 0, 1,1,1, 0 };
const ray_t _MM_ALIGN16 ray = { -1,-1,-1, 0, flt_plus_inf,flt_plus_inf,-1, 0 };
ray_segment_t rs;
const bool rc = ray_box_intersect(box, ray, rs);
}
