speed up gfx, colored lighting is out for now

pxl8d/build.rs

@@ -53,8 +53,11 @@ fn main() {
         .blocklist_type("pxl8_vec3")
         .blocklist_type("pxl8_vec4")
         .blocklist_type("pxl8_mat4")
+        .blocklist_item(".*_simd.*")
+        .blocklist_item("PXL8_SIMD.*")
+        .blocklist_type("__m128.*")
+        .blocklist_type(".*32x4_t|.*16x8_t")
         .raw_line("pub use crate::math::{pxl8_vec2, pxl8_vec3, pxl8_vec4, pxl8_mat4};")
-        .clang_arg("-DPXL8_NO_SIMD")
         .use_core()
         .rustified_enum(".*")
         .generate()

pxl8d/src/bsp.rs

@@ -3,7 +3,7 @@ extern crate alloc;
 use alloc::boxed::Box;
 use alloc::vec::Vec;
 
-use crate::math::{Vec3, VEC3_ZERO};
+use crate::math::Vec3;
 use crate::pxl8::*;
 
 pub type Vertex = pxl8_bsp_vertex;
@@ -31,8 +31,8 @@ impl Default for Face {
             side: 0,
             styles: [0; 4],
             material_id: 0,
-            aabb_min: VEC3_ZERO,
+            aabb_min: Vec3::ZERO,
-            aabb_max: VEC3_ZERO,
+            aabb_max: Vec3::ZERO,
         }
     }
 }
@@ -68,7 +68,7 @@ impl Default for Plane {
     fn default() -> Self {
         Self {
             dist: 0.0,
-            normal: VEC3_ZERO,
+            normal: Vec3::ZERO,
             type_: 0,
         }
     }
@@ -76,7 +76,7 @@ impl Default for Plane {
 
 impl Default for Vertex {
     fn default() -> Self {
-        Self { position: VEC3_ZERO }
+        Self { position: Vec3::ZERO }
     }
 }
 

pxl8d/src/math.rs

@@ -1,14 +1,14 @@
 use core::ops::{Add, Mul, Sub};
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
+#[derive(Debug, Copy, Clone, Default)]
 pub struct Vec2 {
     pub x: f32,
     pub y: f32,
 }
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
+#[derive(Debug, Copy, Clone, Default)]
 pub struct Vec3 {
     pub x: f32,
     pub y: f32,
@@ -16,7 +16,7 @@ pub struct Vec3 {
 }
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
+#[derive(Debug, Copy, Clone, Default)]
 pub struct Vec4 {
     pub x: f32,
     pub y: f32,
@@ -25,7 +25,7 @@ pub struct Vec4 {
 }
 
 #[repr(C)]
-#[derive(Debug, Copy, Clone)]
+#[derive(Debug, Copy, Clone, Default)]
 pub struct Mat4 {
     pub m: [f32; 16],
 }
@@ -39,31 +39,40 @@ pub type pxl8_vec4 = Vec4;
 #[allow(non_camel_case_types)]
 pub type pxl8_mat4 = Mat4;
 
-pub const VEC3_ZERO: Vec3 = Vec3 { x: 0.0, y: 0.0, z: 0.0 };
+impl Vec3 {
-pub const VEC3_Y: Vec3 = Vec3 { x: 0.0, y: 1.0, z: 0.0 };
+    pub const ZERO: Vec3 = Vec3 { x: 0.0, y: 0.0, z: 0.0 };
+    pub const Y: Vec3 = Vec3 { x: 0.0, y: 1.0, z: 0.0 };
 
-pub trait Vec3Ext {
+    pub fn new(x: f32, y: f32, z: f32) -> Self {
-    fn new(x: f32, y: f32, z: f32) -> Self;
-    fn dot(self, rhs: Self) -> f32;
-}
-
-impl Vec3Ext for pxl8_vec3 {
-    fn new(x: f32, y: f32, z: f32) -> Self {
         Self { x, y, z }
     }
 
-    fn dot(self, rhs: Self) -> f32 {
+    pub fn dot(self, rhs: Self) -> f32 {
         self.x * rhs.x + self.y * rhs.y + self.z * rhs.z
     }
-}
 
-impl Default for pxl8_vec3 {
+    pub fn cross(self, rhs: Self) -> Self {
-    fn default() -> Self {
+        Self {
-        VEC3_ZERO
+            x: self.y * rhs.z - self.z * rhs.y,
+            y: self.z * rhs.x - self.x * rhs.z,
+            z: self.x * rhs.y - self.y * rhs.x,
         }
     }
 
-impl Add for pxl8_vec3 {
+    pub fn length(self) -> f32 {
+        libm::sqrtf(self.dot(self))
+    }
+
+    pub fn normalize(self) -> Self {
+        let len_sq = self.dot(self);
+        if len_sq < 1e-12 {
+            return Self::ZERO;
+        }
+        self * (1.0 / libm::sqrtf(len_sq))
+    }
+}
+
+impl Add for Vec3 {
     type Output = Self;
     fn add(self, rhs: Self) -> Self {
         Self {
@@ -74,7 +83,7 @@ impl Add for pxl8_vec3 {
     }
 }
 
-impl Sub for pxl8_vec3 {
+impl Sub for Vec3 {
     type Output = Self;
     fn sub(self, rhs: Self) -> Self {
         Self {
@@ -85,7 +94,7 @@ impl Sub for pxl8_vec3 {
     }
 }
 
-impl Mul<f32> for pxl8_vec3 {
+impl Mul<f32> for Vec3 {
     type Output = Self;
     fn mul(self, rhs: f32) -> Self {
         Self {

pxl8d/src/procgen.rs

@@ -5,8 +5,7 @@ use alloc::vec::Vec;
 use libm::sqrtf;
 
 use crate::bsp::{Bsp, BspBuilder, CellPortals, Edge, Face, Leaf, Node, Plane, Portal, Vertex};
-use crate::math::{Vec3, Vec3Ext};
+use crate::math::Vec3;
-use crate::pxl8::{pxl8_vec3_cross, pxl8_vec3_dot, pxl8_vec3_normalize, pxl8_vec3_scale, pxl8_vec3_add, pxl8_vec3_sub};
 
 pub const CELL_SIZE: f32 = 64.0;
 pub const WALL_HEIGHT: f32 = 128.0;
@@ -408,11 +407,11 @@ const AO_RAY_LENGTH: f32 = 48.0;
 
 fn generate_hemisphere_samples(normal: Vec3) -> [Vec3; AO_NUM_SAMPLES] {
     let tangent = if normal.y.abs() < 0.9 {
-        unsafe { pxl8_vec3_normalize(pxl8_vec3_cross(normal, Vec3::new(0.0, 1.0, 0.0))) }
+        normal.cross(Vec3::new(0.0, 1.0, 0.0)).normalize()
     } else {
-        unsafe { pxl8_vec3_normalize(pxl8_vec3_cross(normal, Vec3::new(1.0, 0.0, 0.0))) }
+        normal.cross(Vec3::new(1.0, 0.0, 0.0)).normalize()
     };
-    let bitangent = unsafe { pxl8_vec3_cross(normal, tangent) };
+    let bitangent = normal.cross(tangent);
 
     let mut samples = [Vec3::new(0.0, 0.0, 0.0); AO_NUM_SAMPLES];
     for i in 0..AO_NUM_SAMPLES {
@@ -425,46 +424,44 @@ fn generate_hemisphere_samples(normal: Vec3) -> [Vec3; AO_NUM_SAMPLES] {
         let local_y = cos_theta;
         let local_z = sin_theta * sin_phi;
 
-        unsafe {
+        let t_contrib = tangent * local_x;
-            let t_contrib = pxl8_vec3_scale(tangent, local_x);
+        let n_contrib = normal * local_y;
-            let n_contrib = pxl8_vec3_scale(normal, local_y);
+        let b_contrib = bitangent * local_z;
-            let b_contrib = pxl8_vec3_scale(bitangent, local_z);
+        samples[i] = (t_contrib + n_contrib + b_contrib).normalize();
-            samples[i] = pxl8_vec3_normalize(pxl8_vec3_add(pxl8_vec3_add(t_contrib, n_contrib), b_contrib));
-        }
     }
     samples
 }
 
 fn ray_triangle_intersect(origin: Vec3, dir: Vec3, v0: Vec3, v1: Vec3, v2: Vec3, max_dist: f32) -> bool {
-    let edge1 = unsafe { pxl8_vec3_sub(v1, v0) };
+    let edge1 = v1 - v0;
-    let edge2 = unsafe { pxl8_vec3_sub(v2, v0) };
+    let edge2 = v2 - v0;
-    let h = unsafe { pxl8_vec3_cross(dir, edge2) };
+    let h = dir.cross(edge2);
-    let a = unsafe { pxl8_vec3_dot(edge1, h) };
+    let a = edge1.dot(h);
 
     if a > -0.0001 && a < 0.0001 {
         return false;
     }
 
     let f = 1.0 / a;
-    let s = unsafe { pxl8_vec3_sub(origin, v0) };
+    let s = origin - v0;
-    let u = f * unsafe { pxl8_vec3_dot(s, h) };
+    let u = f * s.dot(h);
     if u < 0.0 || u > 1.0 {
         return false;
     }
 
-    let q = unsafe { pxl8_vec3_cross(s, edge1) };
+    let q = s.cross(edge1);
-    let v = f * unsafe { pxl8_vec3_dot(dir, q) };
+    let v = f * dir.dot(q);
     if v < 0.0 || u + v > 1.0 {
         return false;
     }
 
-    let t = f * unsafe { pxl8_vec3_dot(edge2, q) };
+    let t = f * edge2.dot(q);
     t > 0.001 && t < max_dist
 }
 
 fn compute_vertex_ao(bsp: &BspBuilder, pos: Vec3, normal: Vec3) -> f32 {
     let samples = generate_hemisphere_samples(normal);
-    let offset_pos = unsafe { pxl8_vec3_add(pos, pxl8_vec3_scale(normal, 0.5)) };
+    let offset_pos = pos + normal * 0.5;
 
     let mut occluded = 0;
 
@@ -526,16 +523,15 @@ fn compute_vertex_light(
     let mut total = 0.0;
 
     for light in lights {
-        let to_light = unsafe { pxl8_vec3_sub(light.position, pos) };
+        let to_light = light.position - pos;
-        let dist = unsafe { pxl8_vec3_dot(to_light, to_light) };
+        let dist = sqrtf(to_light.dot(to_light)).max(1.0);
-        let dist = sqrtf(dist).max(1.0);
 
         if dist > light.radius {
             continue;
         }
 
-        let light_dir = unsafe { pxl8_vec3_normalize(to_light) };
+        let light_dir = to_light.normalize();
-        let ndotl = unsafe { pxl8_vec3_dot(normal, light_dir) }.max(0.0);
+        let ndotl = normal.dot(light_dir).max(0.0);
 
         let attenuation = (1.0 - dist / light.radius).max(0.0);
         let attenuation = attenuation * attenuation;

pxl8d/src/sim.rs

@@ -2,7 +2,7 @@ extern crate alloc;
 
 use alloc::vec::Vec;
 
-use crate::math::{Vec3, Vec3Ext, VEC3_ZERO};
+use crate::math::Vec3;
 use crate::pxl8::*;
 use crate::voxel::VoxelWorld;
 use crate::world::World;
@@ -17,8 +17,8 @@ const MAX_ENTITIES: usize = 1024;
 impl Default for Entity {
     fn default() -> Self {
         Self {
-            pos: VEC3_ZERO,
+            pos: Vec3::ZERO,
-            vel: VEC3_ZERO,
+            vel: Vec3::ZERO,
             yaw: 0.0,
             pitch: 0.0,
             flags: 0,

src/core/pxl8.c

@@ -340,6 +340,7 @@ pxl8_result pxl8_update(pxl8* sys) {
 
     if (game->fps_accumulator >= 1.0f) {
         game->fps = (f32)game->fps_frame_count / game->fps_accumulator;
+        pxl8_info("FPS: %.1f", game->fps);
         game->fps_accumulator = 0.0f;
         game->fps_frame_count = 0;
     }

src/gfx/pxl8_lights.c

@@ -34,50 +34,15 @@ void pxl8_lights_destroy(pxl8_lights* lights) {
 }
 
 void pxl8_lights_add(pxl8_lights* lights, f32 x, f32 y, f32 z, u8 r, u8 g, u8 b, u8 intensity, f32 radius) {
+    (void)r; (void)g; (void)b;
     if (!lights || lights->count >= lights->capacity) return;
 
     f32 radius_sq = radius * radius;
     pxl8_light* l = &lights->data[lights->count++];
-    l->position.x = x;
+    l->position = (pxl8_vec3){{x, y, z}};
-    l->position.y = y;
-    l->position.z = z;
-    l->r = r;
-    l->g = g;
-    l->b = b;
-    l->intensity = intensity;
-    l->radius = radius;
     l->radius_sq = radius_sq;
     l->inv_radius_sq = radius_sq > 0.0f ? 1.0f / radius_sq : 0.0f;
-
+    l->intensity = (f32)intensity / 255.0f;
-    l->constant = 1.0f;
-    if (radius <= 7.0f) {
-        l->linear = 0.7f;
-        l->quadratic = 1.8f;
-    } else if (radius <= 13.0f) {
-        l->linear = 0.35f;
-        l->quadratic = 0.44f;
-    } else if (radius <= 20.0f) {
-        l->linear = 0.22f;
-        l->quadratic = 0.20f;
-    } else if (radius <= 32.0f) {
-        l->linear = 0.14f;
-        l->quadratic = 0.07f;
-    } else if (radius <= 50.0f) {
-        l->linear = 0.09f;
-        l->quadratic = 0.032f;
-    } else if (radius <= 65.0f) {
-        l->linear = 0.07f;
-        l->quadratic = 0.017f;
-    } else if (radius <= 100.0f) {
-        l->linear = 0.045f;
-        l->quadratic = 0.0075f;
-    } else if (radius <= 160.0f) {
-        l->linear = 0.027f;
-        l->quadratic = 0.0028f;
-    } else {
-        l->linear = 0.022f;
-        l->quadratic = 0.0019f;
-    }
 }
 
 void pxl8_lights_clear(pxl8_lights* lights) {

src/gfx/pxl8_lights.h

@@ -7,14 +7,9 @@
 
 typedef struct pxl8_light {
     pxl8_vec3 position;
-    f32 inv_radius_sq;
-    u8 r, g, b;
-    u8 intensity;
-    f32 radius;
     f32 radius_sq;
-    f32 constant;
+    f32 inv_radius_sq;
-    f32 linear;
+    f32 intensity;
-    f32 quadratic;
 } pxl8_light;
 
 typedef struct pxl8_lights pxl8_lights;

src/gfx/pxl8_render.c

@@ -296,7 +296,6 @@ static void rasterize_triangle(
     const tri_setup* setup,
     u8* fb,
     u16* zb,
-    u32* light_accum,
     u32 fb_width,
     pxl8_shader_fn shader,
     const pxl8_gfx_pipeline_desc* pipeline,
@@ -425,7 +424,6 @@ static void rasterize_triangle(
         u32 row_start = (u32)y * fb_width;
         u8* prow = fb + row_start;
         u16* zrow = zb + row_start;
-        u32* lrow = light_accum ? light_accum + row_start : NULL;
 
         i32 x = x_start;
         while (x <= x_end) {
@@ -433,8 +431,8 @@ static void rasterize_triangle(
             if (span_end > x_end) span_end = x_end;
             i32 span_len = span_end - x + 1;
 
-            f32 pw_start = pxl8_fast_rcp(wr);
+            f32 pw_start = 1.0f / wr;
-            f32 pw_end = pxl8_fast_rcp(wr + dwr * (f32)span_len);
+            f32 pw_end = 1.0f / (wr + dwr * (f32)span_len);
 
             f32 u_start = uw * pw_start;
             f32 v_start = vw * pw_start;
@@ -471,7 +469,80 @@ static void rasterize_triangle(
             f32 wy_a = wy_start;
             f32 wz_a = wz_start;
 
-            for (i32 px = x; px <= span_end; px++) {
+            i32 px = x;
+
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+            if (depth_test && depth_compare == PXL8_GFX_COMPARE_LESS && !blend_enabled) {
+                pxl8_f32_simd dz4_simd = pxl8_f32_simd_set(dz * 4.0f);
+                pxl8_f32_simd half = pxl8_f32_simd_set(0.5f);
+                pxl8_f32_simd one = pxl8_f32_simd_set(1.0f);
+                pxl8_f32_simd zero = pxl8_f32_simd_zero();
+                pxl8_f32_simd scale65535 = pxl8_f32_simd_set(65535.0f);
+                pxl8_f32_simd z4 = pxl8_f32_simd_set4(z_a, z_a + dz, z_a + dz * 2.0f, z_a + dz * 3.0f);
+
+                pxl8_f32_simd offsets = pxl8_f32_simd_set4(0.0f, 1.0f, 2.0f, 3.0f);
+                f32 du4 = du * 4.0f, dv4 = dv * 4.0f, dl4 = dl * 4.0f, dc4 = dc * 4.0f;
+                f32 dz4 = dz * 4.0f, dwx4 = dwx * 4.0f, dwy4 = dwy * 4.0f, dwz4 = dwz * 4.0f;
+
+                for (; px + 3 <= span_end; px += 4) {
+                    pxl8_f32_simd depth_norm = pxl8_f32_simd_clamp(pxl8_f32_simd_mul(pxl8_f32_simd_add(z4, one), half), zero, one);
+                    pxl8_i32_simd z16_4 = pxl8_f32_simd_to_i32(pxl8_f32_simd_mul(depth_norm, scale65535));
+                    pxl8_i32_simd zbuf = pxl8_i32_simd_set4((i32)zrow[px], (i32)zrow[px+1], (i32)zrow[px+2], (i32)zrow[px+3]);
+                    i32 mask = pxl8_i32_simd_movemask(pxl8_i32_simd_cmpgt(zbuf, z16_4));
+
+                    STATS_INC(stats, depth_tests, 4);
+
+                    if (mask) {
+                        pxl8_shader_ctx frag_ctx = {
+                            .color_count = 4,
+                            .x = pxl8_i32_simd_set4(px, px + 1, px + 2, px + 3),
+                            .y = pxl8_i32_simd_set(y),
+                            .v_uv = {
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(u_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(du), offsets)),
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(v_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dv), offsets))
+                            },
+                            .v_world = {
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(wx_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dwx), offsets)),
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(wy_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dwy), offsets)),
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(wz_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dwz), offsets))
+                            },
+                            .v_normal = pxl8_vec3_simd_set(setup->normal),
+                            .v_light = pxl8_f32_simd_mul(
+                                pxl8_f32_simd_add(pxl8_f32_simd_set(l_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dl), offsets)),
+                                pxl8_f32_simd_set(1.0f / 255.0f)
+                            ),
+                            .v_color = pxl8_f32_simd_add(pxl8_f32_simd_set(c_a), pxl8_f32_simd_mul(pxl8_f32_simd_set(dc), offsets)),
+                            .v_depth = z4,
+                        };
+
+                        u8 colors[4];
+                        shader(&frag_ctx, bindings, uniforms, colors);
+                        STATS_INC(stats, shader_calls, 1);
+
+                        i32 z16_arr[4];
+                        pxl8_i32_simd_store(z16_arr, z16_4);
+
+                        for (i32 i = 0; i < 4; i++) {
+                            if (!(mask & (0x8 << (i * 4)))) continue;
+                            STATS_INC(stats, depth_passes, 1);
+
+                            u8 color = colors[i];
+                            if (!(alpha_test && color <= alpha_ref) && color != 0) {
+                                prow[px + i] = color;
+                                if (depth_write) zrow[px + i] = (u16)z16_arr[i];
+                                STATS_INC(stats, pixels_written, 1);
+                            }
+                        }
+                    }
+
+                    u_a += du4; v_a += dv4; l_a += dl4; c_a += dc4;
+                    z_a += dz4; wx_a += dwx4; wy_a += dwy4; wz_a += dwz4;
+                    z4 = pxl8_f32_simd_add(z4, dz4_simd);
+                }
+            }
+#endif
+
+            for (; px <= span_end; px++) {
                 f32 depth_norm = pxl8_clamp((z_a + 1.0f) * 0.5f, 0.0f, 1.0f);
                 u16 z16 = (u16)(depth_norm * 65535.0f);
 
@@ -480,18 +551,19 @@ static void rasterize_triangle(
                 if (depth_pass) {
                     STATS_INC(stats, depth_passes, 1);
                     pxl8_shader_ctx frag_ctx = {
-                        .x = px,
+                        .color_count = 1,
-                        .y = y,
+                        .x = pxl8_i32_simd_set(px),
-                        .v_uv = { { u_a, v_a } },
+                        .y = pxl8_i32_simd_set(y),
-                        .v_world = { wx_a, wy_a, wz_a },
+                        .v_uv = { pxl8_f32_simd_set(u_a), pxl8_f32_simd_set(v_a) },
-                        .v_normal = setup->normal,
+                        .v_world = { pxl8_f32_simd_set(wx_a), pxl8_f32_simd_set(wy_a), pxl8_f32_simd_set(wz_a) },
-                        .v_light = l_a / 255.0f,
+                        .v_normal = pxl8_vec3_simd_set(setup->normal),
-                        .v_color = c_a,
+                        .v_light = pxl8_f32_simd_set(l_a / 255.0f),
-                        .v_depth = z_a,
+                        .v_color = pxl8_f32_simd_set(c_a),
-                        .out_light_color = 0,
+                        .v_depth = pxl8_f32_simd_set(z_a),
                     };
 
-                    u8 color = shader(&frag_ctx, bindings, uniforms);
+                    u8 color;
+                    shader(&frag_ctx, bindings, uniforms, &color);
                     STATS_INC(stats, shader_calls, 1);
 
                     if (!(alpha_test && color <= alpha_ref)) {
@@ -506,10 +578,6 @@ static void rasterize_triangle(
                                 zrow[px] = z16;
                             }
                             STATS_INC(stats, pixels_written, 1);
-                            if (lrow && frag_ctx.out_light_color != 0) {
-                                lrow[px] = frag_ctx.out_light_color;
-                                STATS_INC(stats, light_writes, 1);
-                            }
                         }
                     }
                 }
@@ -1080,12 +1148,6 @@ static void execute_draw(
         return;
     }
 
-    u32* light_accum = NULL;
-    if (VALID_TEX(r, pass->desc.light_accum.texture)) {
-        texture_slot* light_tex = &r->textures[SLOT_INDEX(pass->desc.light_accum.texture.id)];
-        light_accum = light_tex->data;
-    }
-
     const pxl8_vertex* vertices = vb->data;
     const u16* indices = use_indices ? ib->data : NULL;
 
@@ -1239,7 +1301,7 @@ static void execute_draw(
                 }
 
                 u64 raster_start = STATS_START();
-                rasterize_triangle(&setup, fb, zb, light_accum, fb_w, shader, &pip->desc,
+                rasterize_triangle(&setup, fb, zb, fb_w, shader, &pip->desc,
                     &shader_bindings, &shader_uniforms, &r->stats);
                 STATS_ADD(&r->stats, raster_ns, raster_start);
             }
@@ -1462,41 +1524,27 @@ void pxl8_resolve_to_rgba(pxl8_renderer* r, pxl8_gfx_texture color, pxl8_gfx_tex
     u8* fb = cs->data;
     u32 w = cs->width;
     u32 h = cs->height;
+    u32 total = w * h;
 
-    u32* light_data = NULL;
+    (void)light_accum;
-    if (light_accum.id != PXL8_GFX_INVALID_ID && VALID_TEX(r, light_accum)) {
+
-        light_data = r->textures[SLOT_INDEX(light_accum.id)].data;
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+    pxl8_i32_simd alpha_mask = pxl8_i32_simd_set((i32)0xFF000000);
+    u32 i = 0;
+    for (; i + 4 <= total; i += 4) {
+        pxl8_i32_simd base = pxl8_i32_simd_set4(
+            (i32)palette[fb[i + 0]], (i32)palette[fb[i + 1]],
+            (i32)palette[fb[i + 2]], (i32)palette[fb[i + 3]]
+        );
+        base = pxl8_i32_simd_or(base, alpha_mask);
+        pxl8_i32_simd_store((i32*)&output[i], base);
     }
-
+    for (; i < total; i++) {
-    for (u32 i = 0; i < w * h; i++) {
+        output[i] = palette[fb[i]] | 0xFF000000;
-        u8 idx = fb[i];
-        u32 base = palette[idx];
-
-        if (light_data) {
-            u32 lv = light_data[i];
-            u32 la = lv >> 24;
-            if (la > 0) {
-                i32 br = base & 0xFF;
-                i32 bg = (base >> 8) & 0xFF;
-                i32 bb = (base >> 16) & 0xFF;
-
-                i32 lr = lv & 0xFF;
-                i32 lg = (lv >> 8) & 0xFF;
-                i32 lb = (lv >> 16) & 0xFF;
-
-                f32 t = (f32)la / 255.0f;
-                br += (i32)((f32)(lr - 128) * t * 2.0f);
-                bg += (i32)((f32)(lg - 128) * t * 2.0f);
-                bb += (i32)((f32)(lb - 128) * t * 2.0f);
-
-                br = pxl8_clamp_byte(br);
-                bg = pxl8_clamp_byte(bg);
-                bb = pxl8_clamp_byte(bb);
-
-                base = (u32)br | ((u32)bg << 8) | ((u32)bb << 16) | 0xFF000000;
     }
+#else
+    for (u32 i = 0; i < total; i++) {
+        output[i] = palette[fb[i]] | 0xFF000000;
     }
-
+#endif
-        output[i] = base | 0xFF000000;
-    }
 }

src/gfx/pxl8_shader.h

@@ -59,7 +59,20 @@ typedef struct pxl8_shader_bindings {
     };
 } pxl8_shader_bindings;
 
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
 typedef struct pxl8_shader_ctx {
+    u32 color_count;
+    pxl8_i32_simd x, y;
+    pxl8_vec2_simd v_uv;
+    pxl8_vec3_simd v_world;
+    pxl8_vec3_simd v_normal;
+    pxl8_f32_simd v_color;
+    pxl8_f32_simd v_light;
+    pxl8_f32_simd v_depth;
+} pxl8_shader_ctx;
+#else
+typedef struct pxl8_shader_ctx {
+    u32 color_count;
     i32 x, y;
     pxl8_vec2 v_uv;
     pxl8_vec3 v_world;
@@ -67,13 +80,14 @@ typedef struct pxl8_shader_ctx {
     f32 v_color;
     f32 v_light;
     f32 v_depth;
-    u32 out_light_color;
 } pxl8_shader_ctx;
+#endif
 
-typedef u8 (*pxl8_shader_fn)(
+typedef void (*pxl8_shader_fn)(
-    pxl8_shader_ctx* ctx,
+    const pxl8_shader_ctx* ctx,
     const pxl8_shader_bindings* bindings,
-    const pxl8_shader_uniforms* uniforms
+    const pxl8_shader_uniforms* uniforms,
+    u8* colors_out
 );
 
 #ifdef __cplusplus

src/gfx/pxl8_shader_builtins.h

@@ -51,30 +51,6 @@ static inline u8 pxl8_colormap_lookup(const pxl8_shader_bindings* b, u8 color, u
     return cm[(row << 8) | (u32)color];
 }
 
-static inline f32 pxl8_light_falloff(const pxl8_shader_ctx* ctx, const pxl8_shader_uniforms* u, u32 light_idx) {
-    if (!u || light_idx >= u->lights_count) return 0.0f;
-    const pxl8_light* light = &u->lights[light_idx];
-    f32 dx = light->position.x - ctx->v_world.x;
-    f32 dy = light->position.y - ctx->v_world.y;
-    f32 dz = light->position.z - ctx->v_world.z;
-    f32 dist_sq = dx * dx + dy * dy + dz * dz;
-    if (dist_sq >= light->radius_sq) return 0.0f;
-    return 1.0f - dist_sq * light->inv_radius_sq;
-}
-
-static inline u32 pxl8_light_color(const pxl8_shader_uniforms* u, u32 light_idx) {
-    if (!u || light_idx >= u->lights_count) return 0;
-    const pxl8_light* light = &u->lights[light_idx];
-    return (u32)light->r | ((u32)light->g << 8) | ((u32)light->b << 16);
-}
-
-static inline void pxl8_set_light_tint(pxl8_shader_ctx* ctx, u32 color, f32 strength) {
-    if (strength <= 0.0f) return;
-    if (strength > 1.0f) strength = 1.0f;
-    u32 alpha = (u32)(strength * 255.0f);
-    ctx->out_light_color = (color & 0x00FFFFFF) | (alpha << 24);
-}
-
 #ifdef __cplusplus
 }
 #endif

src/gfx/pxl8_shader_math.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include "pxl8_math.h"
+
+#if defined(PXL8_SIMD_NEON) || defined(PXL8_SIMD_SSE)
+
+#undef f32
+#define f32 pxl8_f32_simd
+
+typedef pxl8_vec2_simd vec2;
+typedef pxl8_vec3_simd vec3;
+typedef pxl8_vec4_simd vec4;
+
+#define set pxl8_f32_simd_set
+#define add pxl8_f32_simd_add
+#define sub pxl8_f32_simd_sub
+#define mul pxl8_f32_simd_mul
+#define div pxl8_f32_simd_div
+#define inversesqrt pxl8_f32_simd_rsqrt
+#define rcp pxl8_f32_simd_rcp
+#define max pxl8_f32_simd_max
+#define min pxl8_f32_simd_min
+#define clamp pxl8_f32_simd_clamp
+
+#define vec3_set pxl8_vec3_simd_set
+#define vec3_add pxl8_vec3_simd_add
+#define vec3_sub pxl8_vec3_simd_sub
+#define vec3_scale pxl8_vec3_simd_scale
+#define dot pxl8_vec3_simd_dot
+#define normalize pxl8_vec3_simd_normalize
+
+#else
+
+typedef pxl8_vec2 vec2;
+typedef pxl8_vec3 vec3;
+typedef pxl8_vec4 vec4;
+
+#define set(x) (x)
+#define add(a, b) ((a) + (b))
+#define sub(a, b) ((a) - (b))
+#define mul(a, b) ((a) * (b))
+#define div(a, b) ((a) / (b))
+#define inversesqrt pxl8_fast_inv_sqrt
+#define rcp(x) (1.0f / (x))
+#define max pxl8_max
+#define min pxl8_min
+#define clamp pxl8_clamp
+
+#define vec3_set(v) (v)
+#define vec3_add pxl8_vec3_add
+#define vec3_sub pxl8_vec3_sub
+#define vec3_scale pxl8_vec3_scale
+#define dot pxl8_vec3_dot
+#define normalize pxl8_vec3_normalize
+
+#endif

src/gfx/shaders/cpu/lit.c

@@ -2,15 +2,105 @@
 #include "pxl8_shader.h"
 #include "pxl8_shader_builtins.h"
 
-u8 pxl8_shader_lit(
+void pxl8_shader_lit(
-    pxl8_shader_ctx* ctx,
+    const pxl8_shader_ctx* ctx,
     const pxl8_shader_bindings* bindings,
-    const pxl8_shader_uniforms* uniforms
+    const pxl8_shader_uniforms* uniforms,
+    u8* colors_out
 ) {
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+    f32 uv_x[4], uv_y[4], color_f[4];
+    i32 px[4], py[4];
+    pxl8_f32_simd_store(uv_x, ctx->v_uv.x);
+    pxl8_f32_simd_store(uv_y, ctx->v_uv.y);
+    pxl8_f32_simd_store(color_f, ctx->v_color);
+    pxl8_i32_simd_store(px, ctx->x);
+    pxl8_i32_simd_store(py, ctx->y);
+
+    u8 tex_idx[4];
+    for (u32 i = 0; i < ctx->color_count; i++) {
+        if (bindings && bindings->atlas) {
+            tex_idx[i] = pxl8_sample_indexed(bindings, (pxl8_vec2){{ uv_x[i], uv_y[i] }});
+        } else {
+            if (uniforms && uniforms->dither) {
+                tex_idx[i] = pxl8_gfx_dither(color_f[i], (u32)px[i], (u32)py[i]);
+            } else {
+                f32 clamped = pxl8_clamp(color_f[i], 0.0f, 255.0f);
+                tex_idx[i] = (u8)(clamped);
+            }
+        }
+    }
+
+    pxl8_f32_simd light = ctx->v_light;
+
+    if (uniforms) {
+        pxl8_f32_simd ambient = pxl8_f32_simd_set((f32)uniforms->ambient / 255.0f);
+        light = pxl8_f32_simd_max(light, ambient);
+
+        if (uniforms->celestial_intensity > 0.0f) {
+            pxl8_vec3_simd cel_dir = pxl8_vec3_simd_set(uniforms->celestial_dir);
+            pxl8_f32_simd ndotl = pxl8_f32_simd_sub(
+                pxl8_f32_simd_zero(),
+                pxl8_vec3_simd_dot(ctx->v_normal, cel_dir)
+            );
+            pxl8_f32_simd cel_contrib = pxl8_f32_simd_mul(
+                pxl8_f32_simd_max(ndotl, pxl8_f32_simd_zero()),
+                pxl8_f32_simd_set(uniforms->celestial_intensity)
+            );
+            light = pxl8_f32_simd_add(light, cel_contrib);
+        }
+
+        for (u32 i = 0; i < uniforms->lights_count; i++) {
+            const pxl8_light* l = &uniforms->lights[i];
+            pxl8_vec3_simd light_pos = pxl8_vec3_simd_set(l->position);
+            pxl8_vec3_simd to_light = pxl8_vec3_simd_sub(light_pos, ctx->v_world);
+            pxl8_f32_simd dist_sq = pxl8_vec3_simd_dot(to_light, to_light);
+
+            pxl8_f32_simd in_range = pxl8_f32_simd_cmpgt(
+                pxl8_f32_simd_set(l->radius_sq), dist_sq
+            );
+            if (!pxl8_f32_simd_movemask(in_range)) continue;
+
+            pxl8_f32_simd inv_dist = pxl8_f32_simd_rsqrt(dist_sq);
+            pxl8_vec3_simd light_dir = pxl8_vec3_simd_scale(to_light, inv_dist);
+            pxl8_f32_simd ndotl = pxl8_vec3_simd_dot(ctx->v_normal, light_dir);
+            ndotl = pxl8_f32_simd_max(ndotl, pxl8_f32_simd_zero());
+
+            pxl8_f32_simd falloff = pxl8_f32_simd_sub(
+                pxl8_f32_simd_set(1.0f),
+                pxl8_f32_simd_mul(dist_sq, pxl8_f32_simd_set(l->inv_radius_sq))
+            );
+            falloff = pxl8_f32_simd_max(falloff, pxl8_f32_simd_zero());
+
+            pxl8_f32_simd strength = pxl8_f32_simd_mul(
+                pxl8_f32_simd_mul(pxl8_f32_simd_set(l->intensity), falloff),
+                ndotl
+            );
+
+            light = pxl8_f32_simd_add(light, strength);
+        }
+    }
+
+    light = pxl8_f32_simd_clamp(light, pxl8_f32_simd_zero(), pxl8_f32_simd_set(1.0f));
+    pxl8_f32_simd light_f = pxl8_f32_simd_mul(light, pxl8_f32_simd_set(255.0f));
+
+    f32 light_arr[4];
+    pxl8_f32_simd_store(light_arr, light_f);
+
+    for (u32 i = 0; i < ctx->color_count; i++) {
+        u8 light_u8;
+        if (uniforms && uniforms->dither) {
+            light_u8 = pxl8_gfx_dither(light_arr[i], (u32)px[i], (u32)py[i]);
+        } else {
+            light_u8 = (u8)light_arr[i];
+        }
+        colors_out[i] = pxl8_colormap_lookup(bindings, tex_idx[i], light_u8);
+    }
+
+#else
     u8 tex_idx = 0;
     if (bindings && bindings->atlas) {
         tex_idx = pxl8_sample_indexed(bindings, ctx->v_uv);
-        if (pxl8_unlikely(tex_idx == 0)) return 0;
     } else {
         if (uniforms && uniforms->dither) {
             tex_idx = pxl8_gfx_dither(ctx->v_color, (u32)ctx->x, (u32)ctx->y);
@@ -35,11 +125,6 @@ u8 pxl8_shader_lit(
             }
         }
 
-        f32 dyn_strength = 0.0f;
-        f32 dyn_r = 0.0f;
-        f32 dyn_g = 0.0f;
-        f32 dyn_b = 0.0f;
-
         for (u32 i = 0; i < uniforms->lights_count; i++) {
             const pxl8_light* l = &uniforms->lights[i];
             f32 lx = l->position.x - ctx->v_world.x;
@@ -58,28 +143,9 @@ u8 pxl8_shader_lit(
 
             f32 falloff = 1.0f - dist_sq * l->inv_radius_sq;
             if (falloff <= 0.0f) continue;
-            if (uniforms->dither && falloff < 0.33f) {
-                f32 threshold = (PXL8_BAYER_4X4[((u32)ctx->y & 3) * 4 + ((u32)ctx->x & 3)] + 0.5f) * (1.0f / 16.0f);
-                if (falloff < threshold * 0.33f) continue;
-            }
 
-            f32 strength = ((f32)l->intensity / 255.0f) * falloff * ndotl;
+            f32 strength = l->intensity * falloff * ndotl;
-            if (strength <= 0.0f) continue;
+            light += strength;
-
-            dyn_strength += strength;
-            dyn_r += strength * (f32)l->r;
-            dyn_g += strength * (f32)l->g;
-            dyn_b += strength * (f32)l->b;
-        }
-
-        if (dyn_strength > 0.0f) {
-            f32 inv = pxl8_fast_rcp(dyn_strength);
-            u8 r = (u8)pxl8_clamp(dyn_r * inv, 0.0f, 255.0f);
-            u8 g = (u8)pxl8_clamp(dyn_g * inv, 0.0f, 255.0f);
-            u8 b = (u8)pxl8_clamp(dyn_b * inv, 0.0f, 255.0f);
-            u8 a = (u8)pxl8_clamp(dyn_strength * 255.0f, 0.0f, 255.0f);
-            ctx->out_light_color = (u32)r | ((u32)g << 8) | ((u32)b << 16) | ((u32)a << 24);
-            light += dyn_strength;
         }
     }
 
@@ -92,15 +158,6 @@ u8 pxl8_shader_lit(
         light_u8 = pxl8_gfx_dither(light_f, (u32)ctx->x, (u32)ctx->y);
     }
 
-    u8 shaded = pxl8_colormap_lookup(bindings, tex_idx, light_u8);
+    colors_out[0] = pxl8_colormap_lookup(bindings, tex_idx, light_u8);
-
+#endif
-    if (uniforms && uniforms->emissive) {
-        u32 rgb = 0x00FFFFFF;
-        if (bindings && bindings->palette) {
-            rgb = bindings->palette[tex_idx] & 0x00FFFFFF;
-        }
-        pxl8_set_light_tint(ctx, rgb, 1.0f);
-    }
-
-    return shaded;
 }

src/gfx/shaders/cpu/unlit.c

@@ -1,15 +1,37 @@
 #include "pxl8_shader.h"
 #include "pxl8_shader_builtins.h"
 
-u8 pxl8_shader_unlit(
+void pxl8_shader_unlit(
-    pxl8_shader_ctx* ctx,
+    const pxl8_shader_ctx* ctx,
     const pxl8_shader_bindings* bindings,
-    const pxl8_shader_uniforms* uniforms
+    const pxl8_shader_uniforms* uniforms,
+    u8* colors_out
 ) {
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+    f32 uv_x[4], uv_y[4], color_f[4];
+    i32 px[4], py[4];
+    pxl8_f32_simd_store(uv_x, ctx->v_uv.x);
+    pxl8_f32_simd_store(uv_y, ctx->v_uv.y);
+    pxl8_f32_simd_store(color_f, ctx->v_color);
+    pxl8_i32_simd_store(px, ctx->x);
+    pxl8_i32_simd_store(py, ctx->y);
+
+    for (u32 i = 0; i < ctx->color_count; i++) {
+        if (bindings && bindings->atlas) {
+            colors_out[i] = pxl8_sample_indexed(bindings, (pxl8_vec2){{ uv_x[i], uv_y[i] }});
+        } else {
+            if (uniforms && uniforms->dither) {
+                colors_out[i] = pxl8_gfx_dither(color_f[i], (u32)px[i], (u32)py[i]);
+            } else {
+                f32 clamped = pxl8_clamp(color_f[i], 0.0f, 255.0f);
+                colors_out[i] = (u8)(clamped);
+            }
+        }
+    }
+#else
     u8 tex_idx = 0;
     if (bindings && bindings->atlas) {
         tex_idx = pxl8_sample_indexed(bindings, ctx->v_uv);
-        if (tex_idx == 0) return 0;
     } else {
         if (uniforms && uniforms->dither) {
             tex_idx = pxl8_gfx_dither(ctx->v_color, (u32)ctx->x, (u32)ctx->y);
@@ -18,14 +40,6 @@ u8 pxl8_shader_unlit(
             tex_idx = (u8)(clamped);
         }
     }
-
+    colors_out[0] = tex_idx;
-    if (uniforms && uniforms->emissive) {
+#endif
-        u32 rgb = 0x00FFFFFF;
-        if (bindings && bindings->palette) {
-            rgb = bindings->palette[tex_idx] & 0x00FFFFFF;
-        }
-        pxl8_set_light_tint(ctx, rgb, 1.0f);
-    }
-
-    return tex_idx;
 }

src/math/pxl8_math.h

@@ -4,12 +4,25 @@
 
 #include "pxl8_types.h"
 
-#ifndef PXL8_NO_SIMD
 #if defined(__x86_64__) || defined(_M_X64)
+    #define PXL8_SIMD_SSE
     #include <xmmintrin.h>
+    #include <emmintrin.h>
+    typedef __m128 pxl8_f32_simd;
+    typedef __m128i pxl8_i32_simd;
+    typedef __m128i pxl8_i16x8_simd;
 #elif defined(__aarch64__) || defined(_M_ARM64)
+    #define PXL8_SIMD_NEON
     #include <arm_neon.h>
+    typedef float32x4_t pxl8_f32_simd;
+    typedef int32x4_t pxl8_i32_simd;
+    typedef int16x8_t pxl8_i16x8_simd;
 #endif
+
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+typedef struct pxl8_vec2_simd { pxl8_f32_simd x, y; } pxl8_vec2_simd;
+typedef struct pxl8_vec3_simd { pxl8_f32_simd x, y, z; } pxl8_vec3_simd;
+typedef struct pxl8_vec4_simd { pxl8_f32_simd x, y, z, w; } pxl8_vec4_simd;
 #endif
 
 #define PXL8_PI  3.14159265358979323846f
@@ -122,3 +135,7 @@ bool pxl8_frustum_test_aabb(const pxl8_frustum* frustum, pxl8_vec3 min, pxl8_vec
 #ifdef __cplusplus
 }
 #endif
+
+#if defined(PXL8_SIMD_SSE) || defined(PXL8_SIMD_NEON)
+#include "pxl8_math_simd.h"
+#endif

src/math/pxl8_math_scalar.h

@@ -0,0 +1,265 @@
+#pragma once
+
+static inline f32 pxl8_fast_inv_sqrt(f32 x) {
+    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;
+}
+
+static inline u32 pxl8_hash32(u32 x) {
+    x ^= x >> 16;
+    x *= 0x85EBCA6Bu;
+    x ^= x >> 13;
+    x *= 0xC2B2AE35u;
+    x ^= x >> 16;
+    return x;
+}
+
+static inline u64 pxl8_hash64(u64 x) {
+    x ^= x >> 33;
+    x *= 0xff51afd7ed558ccdULL;
+    x ^= x >> 33;
+    x *= 0xc4ceb9fe1a85ec53ULL;
+    x ^= x >> 33;
+    return x;
+}
+
+static inline f32 pxl8_smoothstep(f32 t) {
+    return t * t * (3.0f - 2.0f * t);
+}
+
+static inline pxl8_vec2 pxl8_vec2_add(pxl8_vec2 a, pxl8_vec2 b) {
+    return (pxl8_vec2){ .x = a.x + b.x, .y = a.y + b.y };
+}
+
+static inline pxl8_vec2 pxl8_vec2_sub(pxl8_vec2 a, pxl8_vec2 b) {
+    return (pxl8_vec2){ .x = a.x - b.x, .y = a.y - b.y };
+}
+
+static inline pxl8_vec2 pxl8_vec2_scale(pxl8_vec2 v, f32 s) {
+    return (pxl8_vec2){ .x = v.x * s, .y = v.y * s };
+}
+
+static inline f32 pxl8_vec2_cross(pxl8_vec2 a, pxl8_vec2 b) {
+    return a.x * b.y - a.y * b.x;
+}
+
+static inline f32 pxl8_vec2_dot(pxl8_vec2 a, pxl8_vec2 b) {
+    return a.x * b.x + a.y * b.y;
+}
+
+static inline f32 pxl8_vec2_length(pxl8_vec2 v) {
+    return sqrtf(v.x * v.x + v.y * v.y);
+}
+
+static inline pxl8_vec2 pxl8_vec2_normalize(pxl8_vec2 v) {
+    f32 len_sq = pxl8_vec2_dot(v, v);
+    if (len_sq < 1e-12f) return (pxl8_vec2){0};
+    return pxl8_vec2_scale(v, pxl8_fast_inv_sqrt(len_sq));
+}
+
+static inline pxl8_vec3 pxl8_vec3_add(pxl8_vec3 a, pxl8_vec3 b) {
+    return (pxl8_vec3){ .x = a.x + b.x, .y = a.y + b.y, .z = a.z + b.z };
+}
+
+static inline pxl8_vec3 pxl8_vec3_sub(pxl8_vec3 a, pxl8_vec3 b) {
+    return (pxl8_vec3){ .x = a.x - b.x, .y = a.y - b.y, .z = a.z - b.z };
+}
+
+static inline pxl8_vec3 pxl8_vec3_scale(pxl8_vec3 v, f32 s) {
+    return (pxl8_vec3){ .x = v.x * s, .y = v.y * s, .z = v.z * s };
+}
+
+static inline f32 pxl8_vec3_dot(pxl8_vec3 a, pxl8_vec3 b) {
+    return a.x * b.x + a.y * b.y + a.z * b.z;
+}
+
+static inline pxl8_vec3 pxl8_vec3_cross(pxl8_vec3 a, pxl8_vec3 b) {
+    return (pxl8_vec3){
+        .x = a.y * b.z - a.z * b.y,
+        .y = a.z * b.x - a.x * b.z,
+        .z = a.x * b.y - a.y * b.x,
+    };
+}
+
+static inline f32 pxl8_vec3_length(pxl8_vec3 v) {
+    return sqrtf(pxl8_vec3_dot(v, v));
+}
+
+static inline pxl8_vec3 pxl8_vec3_lerp(pxl8_vec3 a, pxl8_vec3 b, f32 t) {
+    return (pxl8_vec3){
+        a.x + (b.x - a.x) * t,
+        a.y + (b.y - a.y) * t,
+        a.z + (b.z - a.z) * t
+    };
+}
+
+static inline pxl8_vec3 pxl8_vec3_normalize(pxl8_vec3 v) {
+    f32 len_sq = pxl8_vec3_dot(v, v);
+    if (len_sq < 1e-12f) return (pxl8_vec3){0};
+    return pxl8_vec3_scale(v, pxl8_fast_inv_sqrt(len_sq));
+}
+
+static inline pxl8_mat4 pxl8_mat4_identity(void) {
+    pxl8_mat4 mat = {0};
+    mat.m[0] = mat.m[5] = mat.m[10] = mat.m[15] = 1.0f;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_multiply(pxl8_mat4 a, pxl8_mat4 b) {
+    pxl8_mat4 mat = {0};
+    for (i32 col = 0; col < 4; col++) {
+        for (i32 row = 0; row < 4; row++) {
+            mat.m[col * 4 + row] =
+                a.m[0 * 4 + row] * b.m[col * 4 + 0] +
+                a.m[1 * 4 + row] * b.m[col * 4 + 1] +
+                a.m[2 * 4 + row] * b.m[col * 4 + 2] +
+                a.m[3 * 4 + row] * b.m[col * 4 + 3];
+        }
+    }
+    return mat;
+}
+
+static inline pxl8_vec4 pxl8_mat4_multiply_vec4(pxl8_mat4 m, pxl8_vec4 v) {
+    return (pxl8_vec4){
+        .x = m.m[0] * v.x + m.m[4] * v.y + m.m[8] * v.z + m.m[12] * v.w,
+        .y = m.m[1] * v.x + m.m[5] * v.y + m.m[9] * v.z + m.m[13] * v.w,
+        .z = m.m[2] * v.x + m.m[6] * v.y + m.m[10] * v.z + m.m[14] * v.w,
+        .w = m.m[3] * v.x + m.m[7] * v.y + m.m[11] * v.z + m.m[15] * v.w,
+    };
+}
+
+static inline pxl8_vec3 pxl8_mat4_multiply_vec3(pxl8_mat4 m, pxl8_vec3 v) {
+    return (pxl8_vec3){
+        .x = m.m[0] * v.x + m.m[4] * v.y + m.m[8] * v.z,
+        .y = m.m[1] * v.x + m.m[5] * v.y + m.m[9] * v.z,
+        .z = m.m[2] * v.x + m.m[6] * v.y + m.m[10] * v.z,
+    };
+}
+
+static inline pxl8_mat4 pxl8_mat4_translate(f32 x, f32 y, f32 z) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    mat.m[12] = x; mat.m[13] = y; mat.m[14] = z;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_rotate_x(f32 angle) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    f32 c = cosf(angle), s = sinf(angle);
+    mat.m[5] = c; mat.m[9] = -s; mat.m[6] = s; mat.m[10] = c;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_rotate_y(f32 angle) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    f32 c = cosf(angle), s = sinf(angle);
+    mat.m[0] = c; mat.m[8] = s; mat.m[2] = -s; mat.m[10] = c;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_rotate_z(f32 angle) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    f32 c = cosf(angle), s = sinf(angle);
+    mat.m[0] = c; mat.m[4] = -s; mat.m[1] = s; mat.m[5] = c;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_scale(f32 x, f32 y, f32 z) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    mat.m[0] = x; mat.m[5] = y; mat.m[10] = z;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_orthographic(f32 left, f32 right, f32 bottom, f32 top, f32 near, f32 far) {
+    pxl8_mat4 mat = {0};
+    mat.m[0] = 2.0f / (right - left);
+    mat.m[5] = 2.0f / (top - bottom);
+    mat.m[10] = -2.0f / (far - near);
+    mat.m[12] = -(right + left) / (right - left);
+    mat.m[13] = -(top + bottom) / (top - bottom);
+    mat.m[14] = -(far + near) / (far - near);
+    mat.m[15] = 1.0f;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_perspective(f32 fov, f32 aspect, f32 near, f32 far) {
+    pxl8_mat4 mat = {0};
+    f32 tan_half_fov = tanf(fov / 2.0f);
+    mat.m[0] = 1.0f / (aspect * tan_half_fov);
+    mat.m[5] = 1.0f / tan_half_fov;
+    mat.m[10] = -(far + near) / (far - near);
+    mat.m[14] = -(2.0f * far * near) / (far - near);
+    mat.m[11] = -1.0f;
+    return mat;
+}
+
+static inline pxl8_mat4 pxl8_mat4_lookat(pxl8_vec3 eye, pxl8_vec3 center, pxl8_vec3 up) {
+    pxl8_mat4 mat = pxl8_mat4_identity();
+    pxl8_vec3 f = pxl8_vec3_normalize(pxl8_vec3_sub(center, eye));
+    pxl8_vec3 s = pxl8_vec3_normalize(pxl8_vec3_cross(f, up));
+    pxl8_vec3 u = pxl8_vec3_cross(s, f);
+    mat.m[0] = s.x; mat.m[4] = s.y; mat.m[8] = s.z;
+    mat.m[1] = u.x; mat.m[5] = u.y; mat.m[9] = u.z;
+    mat.m[2] = -f.x; mat.m[6] = -f.y; mat.m[10] = -f.z;
+    mat.m[12] = -pxl8_vec3_dot(s, eye);
+    mat.m[13] = -pxl8_vec3_dot(u, eye);
+    mat.m[14] = pxl8_vec3_dot(f, eye);
+    return mat;
+}
+
+static inline pxl8_frustum pxl8_frustum_from_matrix(pxl8_mat4 vp) {
+    pxl8_frustum frustum;
+    const f32* m = vp.m;
+    frustum.planes[0].normal.x = m[3] + m[0];
+    frustum.planes[0].normal.y = m[7] + m[4];
+    frustum.planes[0].normal.z = m[11] + m[8];
+    frustum.planes[0].distance = m[15] + m[12];
+    frustum.planes[1].normal.x = m[3] - m[0];
+    frustum.planes[1].normal.y = m[7] - m[4];
+    frustum.planes[1].normal.z = m[11] - m[8];
+    frustum.planes[1].distance = m[15] - m[12];
+    frustum.planes[2].normal.x = m[3] + m[1];
+    frustum.planes[2].normal.y = m[7] + m[5];
+    frustum.planes[2].normal.z = m[11] + m[9];
+    frustum.planes[2].distance = m[15] + m[13];
+    frustum.planes[3].normal.x = m[3] - m[1];
+    frustum.planes[3].normal.y = m[7] - m[5];
+    frustum.planes[3].normal.z = m[11] - m[9];
+    frustum.planes[3].distance = m[15] - m[13];
+    frustum.planes[4].normal.x = m[3] + m[2];
+    frustum.planes[4].normal.y = m[7] + m[6];
+    frustum.planes[4].normal.z = m[11] + m[10];
+    frustum.planes[4].distance = m[15] + m[14];
+    frustum.planes[5].normal.x = m[3] - m[2];
+    frustum.planes[5].normal.y = m[7] - m[6];
+    frustum.planes[5].normal.z = m[11] - m[10];
+    frustum.planes[5].distance = m[15] - m[14];
+    for (i32 i = 0; i < 6; i++) {
+        f32 len = pxl8_vec3_length(frustum.planes[i].normal);
+        if (len > 1e-6f) {
+            f32 inv_len = 1.0f / len;
+            frustum.planes[i].normal = pxl8_vec3_scale(frustum.planes[i].normal, inv_len);
+            frustum.planes[i].distance *= inv_len;
+        }
+    }
+    return frustum;
+}
+
+static inline bool pxl8_frustum_test_aabb(const pxl8_frustum* frustum, pxl8_vec3 min, pxl8_vec3 max) {
+    const f32 FRUSTUM_EPSILON = -75.0f;
+    for (i32 i = 0; i < 6; i++) {
+        pxl8_vec3 normal = frustum->planes[i].normal;
+        f32 d = frustum->planes[i].distance;
+        pxl8_vec3 p_vertex = {
+            (normal.x > 0.0f) ? max.x : min.x,
+            (normal.y > 0.0f) ? max.y : min.y,
+            (normal.z > 0.0f) ? max.z : min.z
+        };
+        f32 p_dist = pxl8_vec3_dot(normal, p_vertex) + d;
+        if (p_dist < FRUSTUM_EPSILON) return false;
+    }
+    return true;
+}

src/math/pxl8_math_simd.h

@@ -0,0 +1,389 @@
+#pragma once
+
+#if defined(PXL8_SIMD_NEON) || defined(PXL8_SIMD_SSE)
+
+static inline pxl8_f32_simd pxl8_f32_simd_set(f32 x) {
+#if defined(PXL8_SIMD_NEON)
+    return vdupq_n_f32(x);
+#else
+    return _mm_set1_ps(x);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_set4(f32 a, f32 b, f32 c, f32 d) {
+#if defined(PXL8_SIMD_NEON)
+    f32 arr[4] = {a, b, c, d};
+    return vld1q_f32(arr);
+#else
+    return _mm_set_ps(d, c, b, a);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_zero(void) {
+#if defined(PXL8_SIMD_NEON)
+    return vdupq_n_f32(0.0f);
+#else
+    return _mm_setzero_ps();
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_load(const f32* ptr) {
+#if defined(PXL8_SIMD_NEON)
+    return vld1q_f32(ptr);
+#else
+    return _mm_loadu_ps(ptr);
+#endif
+}
+
+static inline void pxl8_f32_simd_store(f32* ptr, pxl8_f32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    vst1q_f32(ptr, v);
+#else
+    _mm_storeu_ps(ptr, v);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_add(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vaddq_f32(a, b);
+#else
+    return _mm_add_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_sub(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vsubq_f32(a, b);
+#else
+    return _mm_sub_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_mul(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vmulq_f32(a, b);
+#else
+    return _mm_mul_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_div(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vdivq_f32(a, b);
+#else
+    return _mm_div_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_rsqrt(pxl8_f32_simd x) {
+#if defined(PXL8_SIMD_NEON)
+    float32x4_t est = vrsqrteq_f32(x);
+    est = vmulq_f32(est, vrsqrtsq_f32(vmulq_f32(x, est), est));
+    return est;
+#else
+    return _mm_rsqrt_ps(x);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_rcp(pxl8_f32_simd x) {
+#if defined(PXL8_SIMD_NEON)
+    float32x4_t est = vrecpeq_f32(x);
+    est = vmulq_f32(est, vrecpsq_f32(x, est));
+    return est;
+#else
+    pxl8_f32_simd rcp = _mm_rcp_ps(x);
+    rcp = _mm_add_ps(rcp, _mm_mul_ps(rcp, _mm_sub_ps(_mm_set1_ps(1.0f), _mm_mul_ps(x, rcp))));
+    return rcp;
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_max(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vmaxq_f32(a, b);
+#else
+    return _mm_max_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_min(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vminq_f32(a, b);
+#else
+    return _mm_min_ps(a, b);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_clamp(pxl8_f32_simd x, pxl8_f32_simd lo, pxl8_f32_simd hi) {
+    return pxl8_f32_simd_min(pxl8_f32_simd_max(x, lo), hi);
+}
+
+static inline pxl8_i32_simd pxl8_f32_simd_to_i32(pxl8_f32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    return vcvtq_s32_f32(v);
+#else
+    return _mm_cvtps_epi32(v);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_f32_simd_cmpgt(pxl8_f32_simd a, pxl8_f32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vreinterpretq_f32_u32(vcgtq_f32(a, b));
+#else
+    return _mm_cmpgt_ps(a, b);
+#endif
+}
+
+static inline i32 pxl8_f32_simd_movemask(pxl8_f32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    static const i32 shift_arr[4] = {0, 1, 2, 3};
+    uint32x4_t mask = vreinterpretq_u32_f32(v);
+    int32x4_t shift = vld1q_s32(shift_arr);
+    mask = vshrq_n_u32(mask, 31);
+    mask = vshlq_u32(mask, shift);
+    return (i32)vaddvq_u32(mask);
+#else
+    return _mm_movemask_ps(v);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_set(i32 x) {
+#if defined(PXL8_SIMD_NEON)
+    return vdupq_n_s32(x);
+#else
+    return _mm_set1_epi32(x);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_set4(i32 a, i32 b, i32 c, i32 d) {
+#if defined(PXL8_SIMD_NEON)
+    i32 arr[4] = {a, b, c, d};
+    return vld1q_s32(arr);
+#else
+    return _mm_set_epi32(d, c, b, a);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_zero(void) {
+#if defined(PXL8_SIMD_NEON)
+    return vdupq_n_s32(0);
+#else
+    return _mm_setzero_si128();
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_load(const i32* ptr) {
+#if defined(PXL8_SIMD_NEON)
+    return vld1q_s32(ptr);
+#else
+    return _mm_loadu_si128((const __m128i*)ptr);
+#endif
+}
+
+static inline void pxl8_i32_simd_store(i32* ptr, pxl8_i32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    vst1q_s32(ptr, v);
+#else
+    _mm_storeu_si128((__m128i*)ptr, v);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_add(pxl8_i32_simd a, pxl8_i32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vaddq_s32(a, b);
+#else
+    return _mm_add_epi32(a, b);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_sub(pxl8_i32_simd a, pxl8_i32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vsubq_s32(a, b);
+#else
+    return _mm_sub_epi32(a, b);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_srli(pxl8_i32_simd v, i32 imm) {
+#if defined(PXL8_SIMD_NEON)
+    int32x4_t shift = vdupq_n_s32(-imm);
+    return vreinterpretq_s32_u32(vshlq_u32(vreinterpretq_u32_s32(v), shift));
+#else
+    return _mm_srli_epi32(v, imm);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_or(pxl8_i32_simd a, pxl8_i32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vorrq_s32(a, b);
+#else
+    return _mm_or_si128(a, b);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_and(pxl8_i32_simd a, pxl8_i32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vandq_s32(a, b);
+#else
+    return _mm_and_si128(a, b);
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i32_simd_cmpgt(pxl8_i32_simd a, pxl8_i32_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vreinterpretq_s32_u32(vcgtq_s32(a, b));
+#else
+    return _mm_cmpgt_epi32(a, b);
+#endif
+}
+
+static inline i32 pxl8_i32_simd_movemask(pxl8_i32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    static const i32 shift_arr[4] = {0, 1, 2, 3};
+    uint32x4_t mask = vreinterpretq_u32_s32(v);
+    int32x4_t shift = vld1q_s32(shift_arr);
+    mask = vshrq_n_u32(mask, 31);
+    mask = vshlq_u32(mask, shift);
+    return (i32)vaddvq_u32(mask);
+#else
+    return _mm_movemask_epi8(v);
+#endif
+}
+
+static inline pxl8_f32_simd pxl8_i32_simd_to_f32(pxl8_i32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    return vcvtq_f32_s32(v);
+#else
+    return _mm_cvtepi32_ps(v);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_set(i16 x) {
+#if defined(PXL8_SIMD_NEON)
+    return vdupq_n_s16(x);
+#else
+    return _mm_set1_epi16(x);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_add(pxl8_i16x8_simd a, pxl8_i16x8_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vaddq_s16(a, b);
+#else
+    return _mm_add_epi16(a, b);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_sub(pxl8_i16x8_simd a, pxl8_i16x8_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vsubq_s16(a, b);
+#else
+    return _mm_sub_epi16(a, b);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_mullo(pxl8_i16x8_simd a, pxl8_i16x8_simd b) {
+#if defined(PXL8_SIMD_NEON)
+    return vmulq_s16(a, b);
+#else
+    return _mm_mullo_epi16(a, b);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_srai(pxl8_i16x8_simd v, i32 imm) {
+#if defined(PXL8_SIMD_NEON)
+    int16x8_t shift = vdupq_n_s16((i16)(-imm));
+    return vshlq_s16(v, shift);
+#else
+    return _mm_srai_epi16(v, imm);
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i32_simd_unpacklo8(pxl8_i32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    uint8x16_t bytes = vreinterpretq_u8_s32(v);
+    return vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bytes)));
+#else
+    return _mm_unpacklo_epi8(v, _mm_setzero_si128());
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i32_simd_unpackhi8(pxl8_i32_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    uint8x16_t bytes = vreinterpretq_u8_s32(v);
+    return vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bytes)));
+#else
+    return _mm_unpackhi_epi8(v, _mm_setzero_si128());
+#endif
+}
+
+static inline pxl8_i16x8_simd pxl8_i16x8_simd_shuffle_3333(pxl8_i16x8_simd v) {
+#if defined(PXL8_SIMD_NEON)
+    int16x4_t lo = vget_low_s16(v);
+    int16x4_t hi = vget_high_s16(v);
+    int16x4_t lo_bcast = vdup_lane_s16(lo, 3);
+    int16x4_t hi_bcast = vdup_lane_s16(hi, 3);
+    return vcombine_s16(lo_bcast, hi_bcast);
+#else
+    return _mm_shufflehi_epi16(
+        _mm_shufflelo_epi16(v, _MM_SHUFFLE(3, 3, 3, 3)),
+        _MM_SHUFFLE(3, 3, 3, 3));
+#endif
+}
+
+static inline pxl8_i32_simd pxl8_i16x8_simd_packus(pxl8_i16x8_simd lo, pxl8_i16x8_simd hi) {
+#if defined(PXL8_SIMD_NEON)
+    uint8x8_t lo8 = vqmovun_s16(lo);
+    uint8x8_t hi8 = vqmovun_s16(hi);
+    return vreinterpretq_s32_u8(vcombine_u8(lo8, hi8));
+#else
+    return _mm_packus_epi16(lo, hi);
+#endif
+}
+
+static inline pxl8_vec3_simd pxl8_vec3_simd_set(pxl8_vec3 v) {
+    return (pxl8_vec3_simd){
+        pxl8_f32_simd_set(v.x),
+        pxl8_f32_simd_set(v.y),
+        pxl8_f32_simd_set(v.z),
+    };
+}
+
+static inline pxl8_vec3_simd pxl8_vec3_simd_add(pxl8_vec3_simd a, pxl8_vec3_simd b) {
+    return (pxl8_vec3_simd){
+        pxl8_f32_simd_add(a.x, b.x),
+        pxl8_f32_simd_add(a.y, b.y),
+        pxl8_f32_simd_add(a.z, b.z),
+    };
+}
+
+static inline pxl8_vec3_simd pxl8_vec3_simd_sub(pxl8_vec3_simd a, pxl8_vec3_simd b) {
+    return (pxl8_vec3_simd){
+        pxl8_f32_simd_sub(a.x, b.x),
+        pxl8_f32_simd_sub(a.y, b.y),
+        pxl8_f32_simd_sub(a.z, b.z),
+    };
+}
+
+static inline pxl8_vec3_simd pxl8_vec3_simd_scale(pxl8_vec3_simd v, pxl8_f32_simd s) {
+    return (pxl8_vec3_simd){
+        pxl8_f32_simd_mul(v.x, s),
+        pxl8_f32_simd_mul(v.y, s),
+        pxl8_f32_simd_mul(v.z, s),
+    };
+}
+
+static inline pxl8_f32_simd pxl8_vec3_simd_dot(pxl8_vec3_simd a, pxl8_vec3_simd b) {
+    return pxl8_f32_simd_add(
+        pxl8_f32_simd_add(
+            pxl8_f32_simd_mul(a.x, b.x),
+            pxl8_f32_simd_mul(a.y, b.y)),
+        pxl8_f32_simd_mul(a.z, b.z));
+}
+
+static inline pxl8_vec3_simd pxl8_vec3_simd_normalize(pxl8_vec3_simd v) {
+    pxl8_f32_simd len_sq = pxl8_vec3_simd_dot(v, v);
+    pxl8_f32_simd inv_len = pxl8_f32_simd_rsqrt(len_sq);
+    return pxl8_vec3_simd_scale(v, inv_len);
+}
+
+#endif