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

speed up gfx, colored lighting is out for now

Browse source
This commit is contained in:
asrael 2026-02-05 02:42:58 -06:00
parent 3c3e961995
commit 01e6059dd1
17 changed files with 1055 additions and 250 deletions
Download patch file Download diff file Expand all files Collapse all files

139
src/gfx/shaders/cpu/lit.c
View file

@ -2,15 +2,105 @@
#include "pxl8_shader.h"
#include "pxl8_shader_builtins.h"
u8 pxl8_shader_lit(
pxl8_shader_ctx* ctx,
void pxl8_shader_lit(
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;
if (strength <= 0.0f) continue;
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;
f32 strength = l->intensity * falloff * ndotl;
light += 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);
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;
colors_out[0] = pxl8_colormap_lookup(bindings, tex_idx, light_u8);
#endif
}