asrael/pxl8
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

wip sw renderer...bsp is borked, also lots of other things

Browse source
This commit is contained in:
asrael 2026-01-21 23:19:50 -06:00
parent 415d424057
commit c771fa665d
56 changed files with 8151 additions and 1261 deletions
Download patch file Download diff file Expand all files Collapse all files

44
src/math/pxl8_math.h
View file

@ -4,16 +4,49 @@
#include "pxl8_types.h"
#if defined(__x86_64__) || defined(_M_X64)
#include <xmmintrin.h>
#elif defined(__aarch64__) || defined(_M_ARM64)
#include <arm_neon.h>
#endif
#define PXL8_PI 3.14159265358979323846f
#define PXL8_TAU (PXL8_PI * 2.0f)
static inline f32 pxl8_fast_inv_sqrt(f32 x) {
#if defined(__x86_64__) || defined(_M_X64)
__m128 v = _mm_set_ss(x);
v = _mm_rsqrt_ss(v);
return _mm_cvtss_f32(v);
#elif defined(__aarch64__) || defined(_M_ARM64)
float32x2_t v = vdup_n_f32(x);
float32x2_t est = vrsqrte_f32(v);
est = vmul_f32(est, vrsqrts_f32(vmul_f32(v, est), est));
return vget_lane_f32(est, 0);
#else
f32 half = 0.5f * x;
i32 i = *(i32*)&x;
i = 0x5f3759df - (i >> 1);
x = *(f32*)&i;
x = x * (1.5f - half * x * x);
return x;
#endif
}
static inline f32 pxl8_fast_rcp(f32 x) {
#if defined(__x86_64__) || defined(_M_X64)
__m128 v = _mm_set_ss(x);
__m128 rcp = _mm_rcp_ss(v);
rcp = _mm_add_ss(rcp, _mm_mul_ss(rcp, _mm_sub_ss(_mm_set_ss(1.0f), _mm_mul_ss(v, rcp))));
return _mm_cvtss_f32(rcp);
#elif defined(__aarch64__) || defined(_M_ARM64)
float32x2_t v = vdup_n_f32(x);
float32x2_t est = vrecpe_f32(v);
est = vmul_f32(est, vrecps_f32(v, est));
return vget_lane_f32(est, 0);
#else
return 1.0f / x;
#endif
}
typedef struct pxl8_vec2 {
@ -33,18 +66,27 @@ typedef struct pxl8_mat4 {
} pxl8_mat4;
typedef struct pxl8_plane {
pxl8_vec3 normal;
f32 distance;
pxl8_vec3 normal;
} pxl8_plane;
typedef struct pxl8_frustum {
pxl8_plane planes[6];
} pxl8_frustum;
typedef struct pxl8_projected_point {
f32 depth;
i32 x;
i32 y;
bool visible;
} pxl8_projected_point;
#ifdef __cplusplus
extern "C" {
#endif
u32 pxl8_hash32(u32 x);
pxl8_vec2 pxl8_vec2_add(pxl8_vec2 a, pxl8_vec2 b);
f32 pxl8_vec2_dot(pxl8_vec2 a, pxl8_vec2 b);
f32 pxl8_vec2_length(pxl8_vec2 v);