concinnity-core 0.19.16

Runtime vocabulary for the Concinnity engine: GPU layouts, ECS components, registry, CPU kernels
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
// The forward main pass's shading model and stage interface, spliced at a
// shader's MAIN_SHADING marker. The second half of the main-pass splice:
// `main_types.slang` declares the records a binding names, this is the body
// that reads the bound resources.
//
// `main_bindless.slang` drives it from the GPU-culled object buffer and the
// bindless texture pool. Everything a binding decides is reached through an
// accessor the including shader defines ahead of this splice, so nothing here
// spells one:
//
//   VIEW / LIGHTS / SHADOW_UNI / CLUSTER      the bound uniform blocks
//   LOCAL_LIGHTS / CLUSTER_LIST /
//   SPOT_SHADOWS / AREA_LIGHTS                the bound storage buffers
//   pool_sample                               one albedo / normal / map texel
//   shadow_map_cmp / shadow_map_size          the cascade depth array
//   spot_shadow_cmp / spot_shadow_map_size    the spot depth array
//   ssao_sample / ssao_size                   the blurred occlusion buffer
//   irradiance_sample /
//   prefilter_sample_level0 /
//   prefilter_sample_bias                     the environment cubes
//   environment_specular                      the reflection tap: probe set
//                                             where one is bound, else the
//                                             prefilter cube
//   ltc_matrix_sample / ltc_magnitude_sample  the area-light lookup tables

// ---- Constants ----

static const float PI = 3.14159265359;

// Per-cluster light-list stride: MAX_LIGHTS_PER_CLUSTER + 1 (slot 0 is the
// count). Matches CLUSTER_LIGHT_LIST_STRIDE in render_types.rs.
static const uint CLUSTER_LIGHT_LIST_STRIDE = 64u;

// Surfaces rougher than this get no SSR / RT reflection; the forward fade
// ramps in below it. Matches the resolve gloss gate (SSR_ROUGH_CUT /
// RT_ROUGH_CUT).
static const float REFLECTION_ROUGHNESS_CUT = 0.6;

// Edge of the LTC lookup tables, and the scale / bias that map [0, 1] onto
// texel centres. Must match LTC_LUT_SIZE in `core::render`'s ltc module.
static const float LTC_LUT_SIZE  = 64.0;
static const float LTC_LUT_SCALE = (LTC_LUT_SIZE - 1.0) / LTC_LUT_SIZE;
static const float LTC_LUT_BIAS  = 0.5 / LTC_LUT_SIZE;

static const float3 SKY_ZENITH  = float3(0.110, 0.322, 0.726);
static const float3 SKY_HORIZON = float3(0.765, 0.863, 0.941);

// ---- Stage interface ----

struct VertexIn
{
    [[vk::location(0)]] float3 pos     : POSITION;
    [[vk::location(1)]] float3 normal  : NORMAL;
    [[vk::location(2)]] float3 tangent : TANGENT;
    [[vk::location(3)]] float3 color   : COLOR0;
    [[vk::location(4)]] float2 uv      : TEXCOORD0;
};

struct VertexOut
{
    float4 position : SV_Position;
    [[vk::location(0)]] float3 world_pos  : TEXCOORD1;
    [[vk::location(1)]] float3 normal     : TEXCOORD2;
    [[vk::location(2)]] float3 tangent    : TEXCOORD3;
    [[vk::location(3)]] float3 bitangent  : TEXCOORD4;
    [[vk::location(4)]] float2 uv         : TEXCOORD5;
    [[vk::location(5)]] float view_depth  : TEXCOORD6;
    [[vk::location(6)]] float3 color      : TEXCOORD7;
    // The object id is needed in the fragment stage too (the instance index is
    // a vertex-only built-in), so it rides a flat varying.
    [[vk::location(7)]] nointerpolation uint object_id : TEXCOORD8;
};

// The whole varying block has to be read by a fragment entry, whichever
// subset the world's `shade` uses: slangc drops an unread input from the
// SPIR-V interface, and Vulkan then reports the vertex stage's matching output
// as unconsumed. An entry folds this sum into its result behind a condition no
// input can meet.
float varyings_read(VertexOut in)
{
    return in.world_pos.x + in.normal.x + in.tangent.x + in.bitangent.x + in.uv.x
         + in.view_depth + in.color.x;
}

// ---- Vertex ----

// Project one model-space vertex through `model`. Normals ride the cofactor
// matrix, so they stay perpendicular under non-uniform scale; the normalize()
// at every use site absorbs the determinant scale it carries.
VertexOut project_vertex(float4x4 model, float3 pos, float3 normal, float3 tangent,
                         float3 color, float2 uv)
{
    VertexOut o;
    float4 world = mul(model, float4(pos, 1.0));
    o.world_pos = world.xyz;

    float3x3 nm = normal_matrix(model);
    o.normal    = normalize(mul(nm, normal));
    o.tangent   = normalize(mul(nm, tangent));
    o.bitangent = cross(o.normal, o.tangent);

    o.uv    = uv;
    o.color = color;

    o.view_depth = -mul(VIEW.view_mat, world).z;
    o.position   = mul(VIEW.vp, world);

    // Skybox sentinel (blue channel 2.0): pin to the far plane so the sky is
    // never clipped by the camera far plane and always renders behind scene
    // geometry. Every forward vertex path needs it.
    if (color.b > 1.5)
    {
        o.position.z = o.position.w * (1.0 - 1e-6);
    }
    return o;
}

// ---- Fragment helpers ----

float distribution_ggx(float3 N, float3 H, float roughness)
{
    float a  = roughness * roughness;
    float a2 = a * a;
    float NdH  = max(dot(N, H), 0.0);
    float NdH2 = NdH * NdH;
    float denom = NdH2 * (a2 - 1.0) + 1.0;
    return a2 / (PI * denom * denom + 0.0001);
}

float geometry_schlick_ggx(float NdV, float roughness)
{
    float r = roughness + 1.0;
    float k = (r * r) / 8.0;
    return NdV / (NdV * (1.0 - k) + k);
}

float geometry_smith(float3 N, float3 V, float3 L, float roughness)
{
    float NdV = max(dot(N, V), 0.0);
    float NdL = max(dot(N, L), 0.0);
    return geometry_schlick_ggx(NdV, roughness) * geometry_schlick_ggx(NdL, roughness);
}

float3 fresnel_schlick(float cosTheta, float3 F0)
{
    return F0 + (1.0 - F0) * pow(clamp(1.0 - cosTheta, 0.0, 1.0), 5.0);
}

float2 env_brdf_approx(float NdV, float rough)
{
    const float4 c0 = float4(-1.0, -0.0275, -0.572, 0.022);
    const float4 c1 = float4( 1.0,  0.0425,  1.040, -0.040);
    float4 r = rough * c0 + c1;
    float a004 = min(r.x * r.x, exp2(-9.28 * NdV)) * r.x + r.y;
    return float2(-1.04, 1.04) * a004 + r.zw;
}

// Decode a tangent-space normal map texel. Only X and Y are read; Z is
// reconstructed from them, so a two-channel source (BC5) decodes the same as
// an RGBA8 one and normal maps can ship as BC5 blocks.
float3 decode_normal_map(float2 encoded)
{
    float2 nxy = encoded * 2.0 - 1.0;
    return float3(nxy, sqrt(clamp(1.0 - dot(nxy, nxy), 0.0, 1.0)));
}

// Geometric specular antialiasing (Kaplanyan et al. 2016, as in Filament):
// widen the NDF by the screen-space variance of the shading normal so an
// undersampled high-frequency normal map at a distance does not alias into
// specular fireflies. A no-op where the normal is smooth (close up), so the
// surface detail is preserved.
float specular_aa_roughness(float3 N, float perceptual_roughness)
{
    const float VARIANCE  = 0.25;
    const float THRESHOLD = 0.18;
    float3 dndx = ddx(N);
    float3 dndy = ddy(N);
    float variance = VARIANCE * (dot(dndx, dndx) + dot(dndy, dndy));
    float alpha = perceptual_roughness * perceptual_roughness;
    float kernel = min(2.0 * variance, THRESHOLD);
    float filtered_alpha2 = clamp(alpha * alpha + kernel, 0.0, 1.0);
    return sqrt(sqrt(filtered_alpha2));
}

float hash_rotation(float2 p)
{
    float h = frac(sin(dot(p, float2(12.9898, 78.233))) * 43758.5453);
    return h * 6.2831853;
}

{SHADOW_BIAS}

// 3x3 hash-rotated PCF of one spot shadow slice. Returns [0, 1] (1.0 fully
// lit), and 1.0 outside the cone's light frustum so an unshadowed region is
// never darkened. A smaller kernel than the cascade PCF: a spot slice covers
// far less world area per texel.
float sample_spot_shadow(int shadow_index, float3 world_pos, float3 normal, float2 screen_xy)
{
    SpotShadowData sd = SPOT_SHADOWS[shadow_index];
    // Offsetting along the normal before projecting pushes the sample off
    // surfaces near-parallel to the light, where depth slope causes acne.
    float3 biased = world_pos + normal * sd.normal_bias;
    float4 light_clip = mul(sd.light_vp, float4(biased, 1.0));
    if (light_clip.w <= 0.0)
    {
        return 1.0;
    }
    float3 ndc = light_clip.xyz / light_clip.w;
    // Flip Y to match the negative-height viewport the spot pass renders with,
    // exactly as the cascade PCF does.
    float2 uv = float2(ndc.x * 0.5 + 0.5, -ndc.y * 0.5 + 0.5);
    if (uv.x < 0.0 || uv.x > 1.0 || uv.y < 0.0 || uv.y > 1.0 || ndc.z < 0.0 || ndc.z > 1.0)
    {
        return 1.0;
    }

    float ref = ndc.z - sd.depth_bias;
    float angle = hash_rotation(screen_xy);
    float ca = cos(angle);
    float sa = sin(angle);
    float2 tex_size = 1.0 / spot_shadow_map_size();

    float sum = 0.0;
    const int RADIUS = 1; // 3x3
    const float SAMPLES = float((2 * RADIUS + 1) * (2 * RADIUS + 1));
    for (int dy = -RADIUS; dy <= RADIUS; dy++)
    {
        for (int dx = -RADIUS; dx <= RADIUS; dx++)
        {
            float2 off = float2(float(dx), float(dy));
            float2 rot = float2(off.x * ca - off.y * sa, off.x * sa + off.y * ca);
            sum += spot_shadow_cmp(
                float3(uv + rot * tex_size, float(shadow_index)), ref);
        }
    }
    return sum / SAMPLES;
}

// Clip a quad against the horizon plane z = 0, keeping the part above it.
// Sutherland-Hodgman rather than the usual hardcoded 16-case table: a quad cut
// by one plane yields at most 5 vertices, and the loop form cannot be got
// wrong case by case. Mirrors clip_quad_to_horizon in `core::render`'s
// ltc::polygon, which is unit-tested against brute-force integration.
int clip_quad_to_horizon(float3 quad[4], out float3 clipped[5])
{
    clipped = { float3(0.0), float3(0.0), float3(0.0), float3(0.0), float3(0.0) };
    int n = 0;
    for (int i = 0; i < 4; i++)
    {
        float3 current  = quad[i];
        float3 previous = quad[(i + 3) % 4];
        bool current_in  = current.z > 0.0;
        bool previous_in = previous.z > 0.0;
        if (current_in != previous_in)
        {
            float t = previous.z / (previous.z - current.z);
            clipped[n++] = float3(previous.xy + t * (current.xy - previous.xy), 0.0);
        }
        if (current_in)
        {
            clipped[n++] = current;
        }
    }
    return n;
}

// Twice the contribution of one edge of the spherical polygon. The cross
// product's z carries the sign, so a reversed winding flips the whole sum,
// which is what tells a front-facing polygon from a back-facing one.
float integrate_edge(float3 v1, float3 v2)
{
    float cos_theta = clamp(dot(v1, v2), -1.0, 1.0);
    float theta     = acos(cos_theta);
    float sin_theta = sqrt(max(1.0 - cos_theta * cos_theta, 0.0));
    float ratio     = (sin_theta > 1e-4) ? (theta / sin_theta) : 1.0;
    return cross(v1, v2).z * ratio;
}

// Fraction of the clamped-cosine distribution the quad covers, in [0, 1].
// `m_inv` is the LTC inverse transform (rows follow mul(v, M) convention), or
// the identity for the diffuse term.
float ltc_evaluate(float3 N, float3 V, float3 P, float3x3 m_inv, float3 corners[4], bool two_sided)
{
    // Shading frame with the normal on +z and the first tangent in the view
    // plane, matching how the table was fitted.
    float3 t1 = normalize(V - N * dot(V, N));
    float3 t2 = cross(N, t1);

    float3 quad[4];
    for (int i = 0; i < 4; i++)
    {
        float3 d = corners[i] - P;
        float3 local = float3(dot(t1, d), dot(t2, d), dot(N, d));
        quad[i] = mul(local, m_inv);
    }

    float3 clipped[5];
    int n = clip_quad_to_horizon(quad, clipped);
    if (n < 3)
    {
        return 0.0;
    }
    for (int k = 0; k < n; k++)
    {
        clipped[k] = normalize(clipped[k]);
    }

    float sum = 0.0;
    for (int e = 0; e < n; e++)
    {
        sum += integrate_edge(clipped[e], clipped[(e + 1) % n]);
    }

    // The edge sum is twice the irradiance; dividing by pi normalises the
    // clamped cosine, so the covered fraction is sum / (2 * pi).
    float form_factor = sum / (2.0 * PI);
    return two_sided ? abs(form_factor) : max(-form_factor, 0.0);
}

// 5x5 hash-rotated PCF of a single cascade. Returns the shadow factor in
// [0, 1] (1.0 fully lit), or 1.0 when the fragment lies outside this
// cascade's light frustum.
float sample_cascade_pcf(int cascade, float3 world_pos, float2 screen_xy)
{
    float4 lc = mul(SHADOW_UNI.light_vps[cascade], float4(world_pos, 1.0));
    float3 ndc = lc.xyz / lc.w;
    float2 uv = float2(ndc.x * 0.5 + 0.5, -ndc.y * 0.5 + 0.5);
    if (uv.x < 0.0 || uv.x > 1.0 || uv.y < 0.0 || uv.y > 1.0 || ndc.z < 0.0 || ndc.z > 1.0)
    {
        return 1.0;
    }

    float ref = ndc.z - cascade_depth_bias(cascade);

    float angle = hash_rotation(screen_xy);
    float ca = cos(angle);
    float sa = sin(angle);

    float2 tex_size = 1.0 / shadow_map_size();

    float sum = 0.0;
    const int RADIUS = 2;
    const float SAMPLES = float((2 * RADIUS + 1) * (2 * RADIUS + 1));
    for (int dy = -RADIUS; dy <= RADIUS; dy++)
    {
        for (int dx = -RADIUS; dx <= RADIUS; dx++)
        {
            float2 off = float2(float(dx), float(dy));
            float2 rot = float2(off.x * ca - off.y * sa, off.x * sa + off.y * ca);
            float2 sample_uv = uv + rot * tex_size;
            sum += shadow_map_cmp(float3(sample_uv, float(cascade)), ref);
        }
    }
    return sum / SAMPLES;
}

// Cascade-aware PCF with cross-cascade blending. Selects the cascade whose far
// split exceeds the fragment's view-space depth, then blends into the next
// cascade across a band at the far edge of that cascade's depth range: each
// cascade places the shadow edge slightly differently, and a hard switch would
// sweep across the world as the camera moves.
float shadow_factor_cascaded(float3 world_pos, float view_depth, float2 screen_xy)
{
    int cascade = 4;
    if      (view_depth < SHADOW_UNI.cascade_splits[0]) cascade = 0;
    else if (view_depth < SHADOW_UNI.cascade_splits[1]) cascade = 1;
    else if (view_depth < SHADOW_UNI.cascade_splits[2]) cascade = 2;
    else if (view_depth < SHADOW_UNI.cascade_splits[3]) cascade = 3;
    if (cascade >= int(SHADOW_UNI.active_cascades))
    {
        return 1.0;
    }

    float shade = sample_cascade_pcf(cascade, world_pos, screen_xy);

    if (cascade + 1 < int(SHADOW_UNI.active_cascades))
    {
        float split_far  = SHADOW_UNI.cascade_splits[cascade];
        float split_near = (cascade == 0) ? 0.0 : SHADOW_UNI.cascade_splits[cascade - 1];
        float band = (split_far - split_near) * 0.15;
        float t = (view_depth - (split_far - band)) / max(band, 1e-4);
        if (t > 0.0)
        {
            float next = sample_cascade_pcf(cascade + 1, world_pos, screen_xy);
            shade = lerp(shade, next, clamp(t, 0.0, 1.0));
        }
    }
    return shade;
}

// ---- The shading model ----

// The forward shading model. `od` is the object record the GPU-driven pass
// reads straight out of the per-frame buffer.
float4 shade_surface(VertexOut in, GpuObjectData od)
{
    float roughness = od.tint_roughness.w;
    float metallic  = od.emissive_metallic.w;
    float3 tint     = od.tint_roughness.xyz;
    float3 emissive = od.emissive_metallic.xyz;

    float3 cam_pos = float3(VIEW.cam_x, VIEW.cam_y, VIEW.cam_z);
    bool ibl_enabled = VIEW.prefilter_mip_count > 0.5;

    // Skybox sentinel (blue channel 2.0): sky colour from the view direction.
    if (in.color.b > 1.5)
    {
        float3 view_dir = normalize(in.world_pos - cam_pos);
        float3 sky;
        if (ibl_enabled)
        {
            sky = prefilter_sample_level0(view_dir);
        }
        else
        {
            float t = max(0.0, view_dir.y);
            sky = lerp(SKY_HORIZON, SKY_ZENITH, t);
        }
        return float4(sky, 1.0);
    }

    // The record is per-object, so a fragment wave that straddles two objects
    // of one indirect draw carries two pool indices. That makes every pool
    // index non-uniform on the descriptor-indexing targets (pool_sample
    // annotates it there).
    float4 albedo_samp = pool_sample(od.albedo_index, in.uv);
    // Alpha cutout: punch the texel out entirely so foliage and decal cards
    // stay in the opaque pass. Disabled at cutoff 0.
    float alpha_cutoff = od.bb_max_alpha_cutoff.w;
    if (alpha_cutoff > 0.0 && albedo_samp.a < alpha_cutoff)
    {
        discard;
    }
    float3 albedo = albedo_samp.rgb * in.color * tint;

    // Unlit view mode: the surface's base color, no lighting.
    if (VIEW.shade_mode > 0.5)
    {
        return float4(albedo, 1.0);
    }

    // Per-material emissive texture carries the colour (the scalar factor is a
    // uniform strength when a map is bound). Slot 0 is the "no map" sentinel.
    if (od.emissive_map_index != 0u)
    {
        emissive *= pool_sample(od.emissive_map_index, in.uv).rgb;
    }

    // Occlusion-roughness-metallic map: green carries roughness, blue carries
    // metallic (glTF convention). Slot 0 is the "no map" sentinel.
    if (od.orm_map_index != 0u)
    {
        float3 orm = pool_sample(od.orm_map_index, in.uv).rgb;
        roughness = orm.g;
        metallic  = orm.b;
    }

    float3 norm_samp = decode_normal_map(pool_sample(od.normal_index, in.uv).rg);
    // Tangent frame as rows so mul(v, M) applies the column-basis transform.
    float3x3 TBN = float3x3(
        normalize(in.tangent),
        normalize(in.bitangent),
        normalize(in.normal));
    float3 N = normalize(mul(norm_samp, TBN));

    // Geometric specular antialiasing on the normal map. Minification aliasing
    // is handled by the texture's mip chain (trilinear + anisotropic
    // sampling); this widens the specular NDF for residual sub-pixel normal
    // variance.
    roughness = specular_aa_roughness(N, roughness);

    float3 V   = normalize(cam_pos - in.world_pos);
    float NdV  = max(dot(N, V), 0.0);

    float3 F0 = lerp(float3(0.04), albedo, metallic);

    float2 screen_xy = in.position.xy;
    float shadow = shadow_factor_cascaded(in.world_pos, in.view_depth, screen_xy);

    float2 ab        = env_brdf_approx(NdV, roughness);
    float ess        = ab.x + ab.y;
    float3 energy_ms = 1.0 + F0 * (1.0 / max(ess, 0.001) - 1.0);

    float3 Lo = float3(0.0);

    for (int i = 0; i < LIGHTS.num_dir; i++)
    {
        float3 L = normalize(LIGHTS.dir[i].dir_i.xyz);
        float intensity = LIGHTS.dir[i].dir_i.w;
        float3 radiance = LIGHTS.dir[i].col.xyz * intensity;

        float3 H = normalize(V + L);
        float NdL = max(dot(N, L), 0.0);

        float D = distribution_ggx(N, H, roughness);
        float G = geometry_smith(N, V, L, roughness);
        float3 F = fresnel_schlick(max(dot(H, V), 0.0), F0);

        float3 kd = (1.0 - F) * (1.0 - metallic);
        float3 spec = (D * G * F) / max(4.0 * NdV * NdL, 0.001) * energy_ms;
        float3 diff = kd * albedo / PI;

        float s = (i == 0) ? shadow : 1.0;
        Lo += (diff + spec) * radiance * NdL * s;
    }

    // Clustered light iteration: when clustering is active (the main camera),
    // map this fragment to its froxel cluster and shade only that cluster's
    // binned lights. Planar / probe re-renders bind use_clusters = 0 (their
    // viewpoint differs from the grid the main camera binned) and fall back
    // to iterating every local light.
    uint cluster_base = 0u;
    int  local_count;
    if (CLUSTER.use_clusters != 0u)
    {
        uint cx = min(uint(screen_xy.x / CLUSTER.screen_w * float(CLUSTER.grid_x)),
                      CLUSTER.grid_x - 1u);
        uint cy = min(uint(screen_xy.y / CLUSTER.screen_h * float(CLUSTER.grid_y)),
                      CLUSTER.grid_y - 1u);
        float zd = max(in.view_depth, CLUSTER.cam_pos_znear.w);
        uint cz = min(uint(log(zd / CLUSTER.cam_pos_znear.w) / log(CLUSTER.view_forward_zfar.w / CLUSTER.cam_pos_znear.w)
                           * float(CLUSTER.grid_z)),
                      CLUSTER.grid_z - 1u);
        uint cid = cx + cy * CLUSTER.grid_x + cz * CLUSTER.grid_x * CLUSTER.grid_y;
        cluster_base = cid * CLUSTER_LIGHT_LIST_STRIDE;
        local_count = int(CLUSTER_LIST[cluster_base]);
    }
    else
    {
        local_count = LIGHTS.num_local_lights;
    }

    for (int jj = 0; jj < local_count; jj++)
    {
        int i = (CLUSTER.use_clusters != 0u)
              ? int(CLUSTER_LIST[cluster_base + 1u + uint(jj)])
              : jj;
        float3 pos_w  = LOCAL_LIGHTS[i].position_range.xyz;
        float  range  = LOCAL_LIGHTS[i].position_range.w;
        float3 col    = LOCAL_LIGHTS[i].color_intensity.xyz;
        float  intens = LOCAL_LIGHTS[i].color_intensity.w;

        // Area lights integrate the whole panel rather than a single
        // direction, so they replace the point / spot BRDF evaluation.
        if (light_kind(LOCAL_LIGHTS[i]) == LIGHT_KIND_AREA)
        {
            int ai = LOCAL_LIGHTS[i].data_index;
            if (ai < 0)
            {
                continue;
            }
            float3 centre = pos_w;
            float3 right  = AREA_LIGHTS[ai].right_two_sided.xyz;
            float3 up     = AREA_LIGHTS[ai].up_pad.xyz;
            bool two_sided = asuint(AREA_LIGHTS[ai].right_two_sided.w) != 0u;

            // Range is a cutoff measured from the panel centre, matching the
            // sphere the clustered cull bins this light with. The physical
            // falloff is already in the form factor: the panel subtends a
            // smaller solid angle further away.
            float centre_dist = length(centre - in.world_pos);
            float window = clamp(1.0 - centre_dist / range, 0.0, 1.0);
            window = window * window;
            if (window <= 0.0)
            {
                continue;
            }

            float3 corners[4];
            corners[0] = centre - right - up;
            corners[1] = centre + right - up;
            corners[2] = centre + right + up;
            corners[3] = centre - right + up;

            // Diffuse needs no lookup: it is the polygon integral under the
            // plain clamped cosine, i.e. an identity transform.
            float3x3 identity = float3x3(
                float3(1.0, 0.0, 0.0),
                float3(0.0, 1.0, 0.0),
                float3(0.0, 0.0, 1.0));
            float diffuse_ff = ltc_evaluate(N, V, in.world_pos, identity, corners, two_sided);

            // Specular applies the fitted transform before the same integral.
            float2 lut_uv = float2(roughness, sqrt(clamp(1.0 - NdV, 0.0, 1.0)));
            lut_uv = lut_uv * LTC_LUT_SCALE + LTC_LUT_BIAS;
            float4 t1 = ltc_matrix_sample(lut_uv);
            float2 t2 = ltc_magnitude_sample(lut_uv);
            // The table stores the inverse normalised so its middle entry is
            // 1, packed as (m00, m20, m02, m22). Rows here follow the
            // mul(v, M) convention, matching the GLSL / MSL column form.
            float3x3 m_inv = float3x3(
                float3(t1.x, 0.0, t1.y),
                float3(0.0,  1.0, 0.0),
                float3(t1.z, 0.0, t1.w));
            float specular_ff = ltc_evaluate(N, V, in.world_pos, m_inv, corners, two_sided);
            // Schlick split baked into the table: t2.x weights the base
            // reflectance, t2.y the grazing response.
            float3 area_spec = F0 * t2.x + (1.0 - F0) * t2.y;

            float3 area_radiance = col * intens * window;
            float3 area_kd = (1.0 - F0) * (1.0 - metallic);
            Lo += area_radiance * (area_kd * albedo * diffuse_ff
                                   + area_spec * specular_ff);
            continue;
        }

        float3 L    = normalize(pos_w - in.world_pos);
        float  dist = length(pos_w - in.world_pos);
        float atten = clamp(1.0 - (dist / range), 0.0, 1.0);
        atten *= atten;
        // Spot cone: full brightness inside cos_inner, squared fade to black
        // at cos_outer. Point lights leave both at zero and skip this.
        if (light_kind(LOCAL_LIGHTS[i]) == LIGHT_KIND_SPOT)
        {
            float cd = dot(LOCAL_LIGHTS[i].direction_kind.xyz, -L);
            float ci = LOCAL_LIGHTS[i].cos_inner;
            float co = LOCAL_LIGHTS[i].cos_outer;
            float t  = clamp((cd - co) / max(ci - co, 1e-4), 0.0, 1.0);
            atten *= t * t;
            // Only spots that claimed a shadow slice sample the array; the
            // rest keep shadow_index at -1 and light without casting.
            int si = LOCAL_LIGHTS[i].shadow_index;
            if (si >= 0 && atten > 0.0)
            {
                atten *= sample_spot_shadow(si, in.world_pos, N, screen_xy);
            }
        }
        float3 radiance = col * intens * atten;

        float3 H  = normalize(V + L);
        float NdL = max(dot(N, L), 0.0);

        float D = distribution_ggx(N, H, roughness);
        float G = geometry_smith(N, V, L, roughness);
        float3 F = fresnel_schlick(max(dot(H, V), 0.0), F0);

        float3 kd   = (1.0 - F) * (1.0 - metallic);
        float3 spec = (D * G * F) / max(4.0 * NdV * NdL, 0.001) * energy_ms;
        float3 diff = kd * albedo / PI;
        Lo += (diff + spec) * radiance * NdL;
    }

    float3 ambient;
    if (ibl_enabled)
    {
        float3 F_ibl       = fresnel_schlick(NdV, F0);
        float3 kd_ibl      = (1.0 - F_ibl) * (1.0 - metallic);
        float3 irradiance  = irradiance_sample(N);
        float3 diffuse_ibl = kd_ibl * albedo * irradiance / PI;

        float3 R = reflect(-V, N);
        // SampleBias (not SampleLevel) so the reflection vector's screen-space
        // footprint widens the mip at grazing or distant angles. A forced LOD
        // defeats minification filtering and aliases the environment into
        // sparkle on near mirrors; flat close-up pixels have a near-zero
        // footprint, so they keep the plain roughness mip.
        float lod = roughness * (VIEW.prefilter_mip_count - 1.0);
        // Local reflection probes when the host binds a probe set and any are
        // baked (box-parallax partition of unity), else the imported
        // environment prefilter cube.
        float3 prefiltered = environment_specular(in.world_pos, R, lod);
        float3 specular_ibl = prefiltered * (F0 * ab.x + ab.y);

        // When an SSR / RT reflection composite owns the sharp specular for
        // glossy surfaces this frame, fade the forward probe specular for
        // glossy dielectrics so the two do not double-count. Metals keep
        // their full albedo-tinted forward specular (the resolve adds only a
        // faint dielectric term), and surfaces rougher than the cut (which
        // the resolve skips) keep theirs too.
        if (VIEW.reflections_enabled > 0.5)
        {
            float fade = smoothstep(REFLECTION_ROUGHNESS_CUT * 0.7,
                                    REFLECTION_ROUGHNESS_CUT, roughness);
            specular_ibl *= lerp(1.0, fade, 1.0 - metallic);
        }

        ambient = diffuse_ibl + specular_ibl;
    }
    else
    {
        ambient = float3(0.03) * albedo;
    }

    // Authored indirect-fill multiplier (PostProcessConfig.ambient_intensity);
    // 1.0 is a no-op. Lifts shadow fill without touching sun-lit surfaces.
    ambient *= LIGHTS.ambient_intensity;

    // SSAO modulates the indirect (ambient / IBL) term only: direct lighting
    // is unaffected. A 1x1 white view is bound when SSAO is disabled, so this
    // samples a constant 1.0 then.
    float2 ssao_uv = screen_xy / ssao_size();
    ambient *= ssao_sample(ssao_uv);

    float3 color = ambient + Lo + emissive;

    return float4(color, albedo_samp.a);
}