oxigaf-render 0.1.2

Differentiable 3D Gaussian Splatting rasterizer (wgpu)
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
//! CPU reference rasterizer for gradient verification.
//!
//! This module provides a pure-CPU implementation of 3D Gaussian Splatting
//! rasterization. It matches the GPU implementation logic exactly but runs
//! entirely on the CPU for easier debugging and gradient verification.
//!
//! The CPU rasterizer serves as ground truth for validating GPU gradients
//! through finite-difference approximation.

use nalgebra as na;
use rayon::prelude::*;

use crate::config::RasterConfig;
use crate::gaussian::GaussianModel;
use crate::RenderError;

/// Spherical Harmonics constants (matching WGSL shader)
const SH_C0: f32 = 0.282_094_8;
const SH_C1: f32 = 0.488_602_52;
const SH_C2_0: f32 = 1.092_548_5;
const SH_C2_1: f32 = -1.092_548_5;
const SH_C2_2: f32 = 0.315_391_57;
const SH_C2_3: f32 = -1.092_548_5;
const SH_C2_4: f32 = 0.546_274_24;
const SH_C3_0: f32 = -0.590_043_6;
const SH_C3_1: f32 = 2.890_611_4;
const SH_C3_2: f32 = -0.457_045_8;
const SH_C3_3: f32 = 0.373_176_34;
const SH_C3_4: f32 = -0.457_045_8;
const SH_C3_5: f32 = 1.445_305_7;
const SH_C3_6: f32 = -0.590_043_6;

/// Camera parameters for CPU rasterization.
#[derive(Debug, Clone)]
pub struct CpuCamera {
    /// View matrix (world to camera).
    pub view: na::Matrix4<f32>,
    /// Projection matrix.
    pub proj: na::Matrix4<f32>,
    /// Camera position in world space.
    pub position: na::Vector3<f32>,
    /// Focal lengths (fx, fy) in pixels.
    pub focal: na::Vector2<f32>,
}

/// Projected Gaussian in 2D screen space.
///
/// Deliberately carries no source-Gaussian index. One was stored here and read
/// by nobody, held alive by `#[allow(dead_code)]`; the ordering it looked like
/// it existed for is supplied by *position* in the projected vector instead —
/// see the depth sort in [`CpuRasterizer::render`], which relies on these being
/// pushed in ascending model index.
#[derive(Debug, Clone)]
struct ProjectedGaussian {
    /// 2D mean (pixel coordinates).
    mean2d: na::Vector2<f32>,
    /// 2D conic matrix (inverse of covariance): [a, b, c] for [[a, b], [b, c]].
    conic: na::Vector3<f32>,
    /// RGB color (evaluated from SH).
    color: na::Vector3<f32>,
    /// Opacity (after sigmoid).
    opacity: f32,
    /// Depth (distance from camera).
    depth: f32,
}

/// CPU reference rasterizer output.
#[derive(Debug, Clone)]
pub struct CpuRenderOutput {
    /// RGBA color data [H * W * 4] as f32.
    pub color_data: Vec<f32>,
    /// Depth data [H * W].
    pub depth_data: Vec<f32>,
    /// Image width.
    pub width: u32,
    /// Image height.
    pub height: u32,
}

/// Pure-CPU rasterizer for gradient verification.
pub struct CpuRasterizer {
    config: RasterConfig,
}

impl CpuRasterizer {
    /// Create a new CPU rasterizer.
    pub fn new(config: RasterConfig) -> Self {
        Self { config }
    }

    /// Render Gaussians to an image using CPU.
    ///
    /// This implements the same algorithm as the GPU rasterizer:
    /// 1. Project 3D Gaussians to 2D screen space
    /// 2. Sort by depth (painter's algorithm)
    /// 3. For each pixel, blend Gaussians front-to-back
    /// 4. Accumulate RGB with alpha blending
    pub fn render(
        &self,
        model: &GaussianModel,
        camera: &CpuCamera,
    ) -> Result<CpuRenderOutput, RenderError> {
        let width = self.config.image_width;
        let height = self.config.image_height;
        let num_pixels = (width * height) as usize;

        // Step 1: Project all Gaussians to 2D
        let mut projected = Vec::new();
        for i in 0..model.len() {
            if let Some(g) = self.project_gaussian(model, i, camera)? {
                projected.push(g);
            }
        }

        // Step 2: Sort by depth (front-to-back).
        //
        // GPU parity for equal depths: `shaders/tile_assign.wgsl` builds the
        // key as `(tile_id, bitcast<u32>(depth))` with no index component, so
        // two Gaussians at exactly the same depth in the same tile are ordered
        // purely by the stability of the radix sort — i.e. by the order their
        // pairs were emitted, which is ascending Gaussian index. `sort_by` is
        // likewise a stable sort and `projected` is filled by the ascending
        // `0..model.len()` loop above, so ties break by ascending model index
        // on both sides. Do not switch this to `sort_unstable_by`: that would
        // silently diverge from the GPU exactly on the coincident-depth case
        // the CPU path exists to cross-check.
        projected.sort_by(|a, b| {
            a.depth
                .partial_cmp(&b.depth)
                .unwrap_or(std::cmp::Ordering::Equal)
        });

        // Step 3: Rasterize per-pixel (parallel)
        let pixel_results: Vec<([f32; 4], f32)> = (0..num_pixels)
            .into_par_iter()
            .map(|pixel_idx| {
                let px = (pixel_idx % width as usize) as f32;
                let py = (pixel_idx / width as usize) as f32;
                self.blend_pixel(px, py, &projected)
            })
            .collect();

        // Convert to output format
        let mut color_data = vec![0.0; num_pixels * 4];
        let mut depth_data = vec![0.0; num_pixels];

        for (i, (pixel_color, pixel_depth)) in pixel_results.iter().enumerate() {
            color_data[i * 4] = pixel_color[0];
            color_data[i * 4 + 1] = pixel_color[1];
            color_data[i * 4 + 2] = pixel_color[2];
            color_data[i * 4 + 3] = pixel_color[3];
            depth_data[i] = *pixel_depth;
        }

        Ok(CpuRenderOutput {
            color_data,
            depth_data,
            width,
            height,
        })
    }

    /// Project a single Gaussian from 3D to 2D.
    ///
    /// Returns `None` if Gaussian is culled (behind camera, off-screen, etc.).
    fn project_gaussian(
        &self,
        model: &GaussianModel,
        idx: usize,
        camera: &CpuCamera,
    ) -> Result<Option<ProjectedGaussian>, RenderError> {
        let gaussian = &model.gaussians[idx];

        // Get position
        let pos = na::Vector3::new(
            gaussian.position[0],
            gaussian.position[1],
            gaussian.position[2],
        );

        // Transform to camera space
        let pos_cam = camera.view.transform_point(&na::Point3::from(pos));

        // Cull against near/far planes, mirroring shaders/preprocess.wgsl
        // exactly (`let depth = -p_view.z; if depth <= near_plane || depth
        // >= far_plane { cull }`), including its boundary strictness — a
        // Gaussian sitting exactly on a clip plane is culled just like on
        // the GPU, and Gaussians beyond `far_plane` are dropped here too
        // (previously only the near plane was checked).
        let depth = -pos_cam.z;
        if depth <= self.config.near_plane || depth >= self.config.far_plane {
            return Ok(None);
        }

        // Project to screen space
        let pos_ndc = camera.proj.transform_point(&pos_cam);

        // Convert to pixel coordinates
        let mean2d = na::Vector2::new(
            (pos_ndc.x * 0.5 + 0.5) * self.config.image_width as f32,
            (pos_ndc.y * 0.5 + 0.5) * self.config.image_height as f32,
        );

        // Build rotation matrix directly from raw quaternion (matching GPU's quat_to_mat
        // which does NOT normalize the quaternion)
        let qx = gaussian.rotation[0];
        let qy = gaussian.rotation[1];
        let qz = gaussian.rotation[2];
        let qw = gaussian.rotation[3];
        let x2 = 2.0 * qx;
        let y2 = 2.0 * qy;
        let z2 = 2.0 * qz;
        let xx = qx * x2;
        let xy = qx * y2;
        let xz = qx * z2;
        let yy = qy * y2;
        let yz = qy * z2;
        let zz = qz * z2;
        let wx = qw * x2;
        let wy = qw * y2;
        let wz = qw * z2;
        let rotation_matrix = na::Matrix3::new(
            1.0 - (yy + zz),
            xy - wz,
            xz + wy,
            xy + wz,
            1.0 - (xx + zz),
            yz - wx,
            xz - wy,
            yz + wx,
            1.0 - (xx + yy),
        );

        let scale = na::Vector3::new(
            gaussian.scale[0].exp(),
            gaussian.scale[1].exp(),
            gaussian.scale[2].exp(),
        );

        // Compute 3D covariance matrix
        let s = na::Matrix3::from_diagonal(&scale);
        let m = rotation_matrix * s;
        let cov3d = m * m.transpose();

        // Project covariance to 2D (using Jacobian of projection)
        let focal = camera.focal;
        let z = pos_cam.z.abs();
        let z2 = z * z;

        // Jacobian of perspective projection
        let j = na::Matrix2x3::new(
            focal.x / z,
            0.0,
            focal.x * pos_cam.x / z2,
            0.0,
            focal.y / z,
            focal.y * pos_cam.y / z2,
        );

        // View matrix rotation part
        let w = camera.view.fixed_view::<3, 3>(0, 0);

        // Transform covariance: cov2d = J * W * cov3d * W^T * J^T
        let t = w * cov3d;
        let cov2d_full = j * t * w.transpose() * j.transpose();

        // Add small constant for numerical stability
        let cov2d = na::Matrix2::new(
            cov2d_full[(0, 0)] + 0.3,
            cov2d_full[(0, 1)],
            cov2d_full[(1, 0)],
            cov2d_full[(1, 1)] + 0.3,
        );

        // Compute conic (inverse of covariance)
        let det = cov2d[(0, 0)] * cov2d[(1, 1)] - cov2d[(0, 1)] * cov2d[(1, 0)];
        if det.abs() < 1e-10 {
            return Ok(None); // Singular covariance
        }

        let conic = na::Vector3::new(
            cov2d[(1, 1)] / det,
            -cov2d[(0, 1)] / det,
            cov2d[(0, 0)] / det,
        );

        // Evaluate Spherical Harmonics for color
        let view_dir = (pos - camera.position).normalize();
        let color = self.eval_sh(model, idx, view_dir)?;

        // Get opacity (apply sigmoid)
        let opacity = sigmoid(gaussian.opacity);

        Ok(Some(ProjectedGaussian {
            mean2d,
            conic,
            color,
            opacity,
            depth,
        }))
    }

    /// Evaluate Spherical Harmonics for a Gaussian.
    fn eval_sh(
        &self,
        model: &GaussianModel,
        idx: usize,
        dir: na::Vector3<f32>,
    ) -> Result<na::Vector3<f32>, RenderError> {
        // `GaussianModel::sh_degree` is an unvalidated plain `u32`. Silently
        // clamping it to 3 while indexing with the *clamped* value would
        // compute a stride (`sh_coeffs_per_gaussian`) that is smaller than
        // the data's actual per-Gaussian layout whenever `sh_degree > 3`,
        // so every Gaussian after the first would read coefficients
        // belonging to a different Gaussian. Reject it outright instead.
        if model.sh_degree > 3 {
            return Err(RenderError::Rasterize(format!(
                "unsupported SH degree {}: only degrees 0-3 are implemented",
                model.sh_degree
            )));
        }
        let sh_degree = model.sh_degree;
        let sh_coeffs_per_gaussian = ((sh_degree + 1) * (sh_degree + 1) * 3) as usize;
        let sh_start = idx * sh_coeffs_per_gaussian;

        if sh_start + sh_coeffs_per_gaussian > model.sh_coeffs.len() {
            return Err(RenderError::Rasterize(format!(
                "SH coefficient index out of bounds: {} + {} > {}",
                sh_start,
                sh_coeffs_per_gaussian,
                model.sh_coeffs.len()
            )));
        }

        let sh = &model.sh_coeffs[sh_start..sh_start + sh_coeffs_per_gaussian];

        let result = match sh_degree {
            0 => eval_sh_degree0(sh),
            1 => eval_sh_degree1(dir, sh),
            2 => eval_sh_degree2(dir, sh),
            3 => eval_sh_degree3(dir, sh),
            _ => {
                return Err(RenderError::Rasterize(format!(
                    "Invalid SH degree: {}",
                    sh_degree
                )))
            }
        };

        Ok(result)
    }

    /// Blend Gaussians for a single pixel (front-to-back alpha compositing).
    ///
    /// Returns `([r, g, b, alpha], depth)`. `depth` is the alpha-weighted
    /// accumulated depth, mirroring `shaders/rasterize_fwd.wgsl` exactly:
    /// the same per-Gaussian `weight = T * alpha` used for color is also
    /// applied to `g.depth` inside the same loop
    /// (`depth_acc += weight * wg_depth[b]` on the GPU side), with no
    /// background contribution added to depth afterward — the GPU writes
    /// `out_depth[pixel_idx] = depth_acc` directly, unlike color which gets
    /// `color += T * background`.
    fn blend_pixel(&self, px: f32, py: f32, gaussians: &[ProjectedGaussian]) -> ([f32; 4], f32) {
        let pixel_f = na::Vector2::new(px + 0.5, py + 0.5);

        let mut t = 1.0f32; // transmittance
        let mut color = na::Vector3::zeros();
        let mut depth_acc = 0.0f32;

        for g in gaussians {
            if t < 1.0 / 255.0 {
                break; // Early termination
            }

            // Evaluate 2D Gaussian
            let d = pixel_f - g.mean2d;
            let power = -0.5
                * (g.conic.x * d.x * d.x + 2.0 * g.conic.y * d.x * d.y + g.conic.z * d.y * d.y);

            if !(-4.0..=0.0).contains(&power) {
                continue; // Outside valid range
            }

            let alpha_raw = g.opacity * power.exp();
            let alpha = alpha_raw.min(0.99);

            if alpha < 1.0 / 255.0 {
                continue; // Too transparent
            }

            let weight = t * alpha;
            color += weight * g.color;
            depth_acc += weight * g.depth;
            t *= 1.0 - alpha;
        }

        // Add background (color only — depth has no background term, as on
        // the GPU).
        let bg = na::Vector3::new(
            self.config.background[0],
            self.config.background[1],
            self.config.background[2],
        );
        color += t * bg;

        ([color.x, color.y, color.z, 1.0 - t], depth_acc)
    }
}

/// Sigmoid function.
fn sigmoid(x: f32) -> f32 {
    1.0 / (1.0 + (-x).exp())
}

/// Evaluate SH degree 0 (DC term only).
fn eval_sh_degree0(sh: &[f32]) -> na::Vector3<f32> {
    let dc = na::Vector3::new(sh[0], sh[1], sh[2]);
    (SH_C0 * dc + na::Vector3::new(0.5, 0.5, 0.5)).sup(&na::Vector3::zeros())
}

/// Evaluate SH degree 1 (DC + linear).
fn eval_sh_degree1(dir: na::Vector3<f32>, sh: &[f32]) -> na::Vector3<f32> {
    let x = dir.x;
    let y = dir.y;
    let z = dir.z;

    // DC term
    let mut result = SH_C0 * na::Vector3::new(sh[0], sh[1], sh[2]);

    // Linear terms: coefficients [3..12]
    // Y_1^{-1} = -y, Y_1^0 = z, Y_1^1 = -x
    result += SH_C1 * (-y) * na::Vector3::new(sh[3], sh[4], sh[5]);
    result += SH_C1 * z * na::Vector3::new(sh[6], sh[7], sh[8]);
    result += SH_C1 * (-x) * na::Vector3::new(sh[9], sh[10], sh[11]);

    (result + na::Vector3::new(0.5, 0.5, 0.5)).sup(&na::Vector3::zeros())
}

/// Evaluate SH degree 2 (DC + linear + quadratic).
///
/// All terms are accumulated without intermediate clamping to match the GPU
/// shader, which evaluates every SH band in a single pass and only clamps the
/// final result.
fn eval_sh_degree2(dir: na::Vector3<f32>, sh: &[f32]) -> na::Vector3<f32> {
    let x = dir.x;
    let y = dir.y;
    let z = dir.z;

    // DC term (degree 0)
    let mut result = SH_C0 * na::Vector3::new(sh[0], sh[1], sh[2]);

    // Linear terms (degree 1): same as in eval_sh_degree1 but without clamping
    result += SH_C1 * (-y) * na::Vector3::new(sh[3], sh[4], sh[5]);
    result += SH_C1 * z * na::Vector3::new(sh[6], sh[7], sh[8]);
    result += SH_C1 * (-x) * na::Vector3::new(sh[9], sh[10], sh[11]);

    // Quadratic terms (degree 2)
    let xx = x * x;
    let yy = y * y;
    let zz = z * z;
    let xy = x * y;
    let yz = y * z;
    let xz = x * z;

    result += SH_C2_0 * xy * na::Vector3::new(sh[12], sh[13], sh[14]);
    result += SH_C2_1 * yz * na::Vector3::new(sh[15], sh[16], sh[17]);
    result += SH_C2_2 * (2.0 * zz - xx - yy) * na::Vector3::new(sh[18], sh[19], sh[20]);
    result += SH_C2_3 * xz * na::Vector3::new(sh[21], sh[22], sh[23]);
    result += SH_C2_4 * (xx - yy) * na::Vector3::new(sh[24], sh[25], sh[26]);

    // Apply clamp ONCE at the end (matches GPU behavior)
    (result + na::Vector3::new(0.5, 0.5, 0.5)).sup(&na::Vector3::zeros())
}

/// Evaluate SH degree 3 (DC + linear + quadratic + cubic).
///
/// All terms are accumulated without intermediate clamping to match the GPU
/// shader, which evaluates every SH band in a single pass and only clamps the
/// final result.
fn eval_sh_degree3(dir: na::Vector3<f32>, sh: &[f32]) -> na::Vector3<f32> {
    let x = dir.x;
    let y = dir.y;
    let z = dir.z;

    // DC term (degree 0)
    let mut result = SH_C0 * na::Vector3::new(sh[0], sh[1], sh[2]);

    // Linear terms (degree 1)
    result += SH_C1 * (-y) * na::Vector3::new(sh[3], sh[4], sh[5]);
    result += SH_C1 * z * na::Vector3::new(sh[6], sh[7], sh[8]);
    result += SH_C1 * (-x) * na::Vector3::new(sh[9], sh[10], sh[11]);

    // Quadratic terms (degree 2)
    let xx = x * x;
    let yy = y * y;
    let zz = z * z;
    let xy = x * y;
    let yz = y * z;
    let xz = x * z;

    result += SH_C2_0 * xy * na::Vector3::new(sh[12], sh[13], sh[14]);
    result += SH_C2_1 * yz * na::Vector3::new(sh[15], sh[16], sh[17]);
    result += SH_C2_2 * (2.0 * zz - xx - yy) * na::Vector3::new(sh[18], sh[19], sh[20]);
    result += SH_C2_3 * xz * na::Vector3::new(sh[21], sh[22], sh[23]);
    result += SH_C2_4 * (xx - yy) * na::Vector3::new(sh[24], sh[25], sh[26]);

    // Cubic terms (degree 3)
    result += SH_C3_0 * y * (3.0 * xx - yy) * na::Vector3::new(sh[27], sh[28], sh[29]);
    result += SH_C3_1 * x * y * z * na::Vector3::new(sh[30], sh[31], sh[32]);
    result += SH_C3_2 * y * (4.0 * zz - xx - yy) * na::Vector3::new(sh[33], sh[34], sh[35]);
    result +=
        SH_C3_3 * z * (2.0 * zz - 3.0 * xx - 3.0 * yy) * na::Vector3::new(sh[36], sh[37], sh[38]);
    result += SH_C3_4 * x * (4.0 * zz - xx - yy) * na::Vector3::new(sh[39], sh[40], sh[41]);
    result += SH_C3_5 * z * (xx - yy) * na::Vector3::new(sh[42], sh[43], sh[44]);
    result += SH_C3_6 * x * (xx - 3.0 * yy) * na::Vector3::new(sh[45], sh[46], sh[47]);

    // Apply clamp ONCE at the end (matches GPU behavior)
    (result + na::Vector3::new(0.5, 0.5, 0.5)).sup(&na::Vector3::zeros())
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_sigmoid() {
        assert!((sigmoid(0.0) - 0.5).abs() < 1e-6);
        assert!(sigmoid(10.0) > 0.99);
        assert!(sigmoid(-10.0) < 0.01);
    }

    #[test]
    fn test_sh_degree0() {
        let sh = vec![1.0, 0.5, 0.25];
        let result = eval_sh_degree0(&sh);

        // Expected: SH_C0 * [1.0, 0.5, 0.25] + 0.5
        let expected = na::Vector3::new(SH_C0 * 1.0 + 0.5, SH_C0 * 0.5 + 0.5, SH_C0 * 0.25 + 0.5);

        assert!((result - expected).norm() < 1e-5);
    }

    #[test]
    fn test_cpu_rasterizer_creation() {
        let config = RasterConfig::new().with_resolution(512, 512);
        let _rasterizer = CpuRasterizer::new(config);
    }

    #[test]
    fn test_blend_pixel_empty() {
        let config = RasterConfig::new().with_resolution(512, 512);
        let rasterizer = CpuRasterizer::new(config);

        let gaussians = vec![];
        let (pixel, depth) = rasterizer.blend_pixel(0.0, 0.0, &gaussians);

        // Should return background color (black by default) and zero
        // accumulated depth (no Gaussians contributed).
        assert_eq!(pixel, [0.0, 0.0, 0.0, 0.0]);
        assert_eq!(depth, 0.0);
    }

    #[test]
    fn test_blend_pixel_depth_matches_weighted_sum() {
        // Two exactly-overlapping, fully-opaque (post-clamp) Gaussians at
        // the pixel center: depth_acc must equal sum(weight_i * depth_i)
        // with weight_i = T_i * alpha_i, exactly mirroring
        // shaders/rasterize_fwd.wgsl's `depth_acc += weight * wg_depth[b]`.
        let config = RasterConfig::new().with_resolution(64, 64);
        let rasterizer = CpuRasterizer::new(config);

        let near = ProjectedGaussian {
            mean2d: na::Vector2::new(10.5, 10.5),
            conic: na::Vector3::new(10.0, 0.0, 10.0),
            color: na::Vector3::new(1.0, 0.0, 0.0),
            opacity: 1.0,
            depth: 1.0,
        };
        let far = ProjectedGaussian {
            mean2d: na::Vector2::new(10.5, 10.5),
            conic: na::Vector3::new(10.0, 0.0, 10.0),
            color: na::Vector3::new(0.0, 1.0, 0.0),
            opacity: 1.0,
            depth: 10.0,
        };

        let (_pixel, depth) = rasterizer.blend_pixel(10.0, 10.0, &[near, far]);

        // Both Gaussians sit dead-center (power = 0) with opacity clamped
        // to 0.99: weight_near = 1.0 * 0.99, weight_far = (1.0 - 0.99) * 0.99.
        let expected = 0.99 * 1.0 + (0.01 * 0.99) * 10.0;
        assert!(
            (depth - expected).abs() < 1e-4,
            "expected depth={expected}, got {depth}"
        );
        // The nearer Gaussian dominates because its weight (0.99) far
        // exceeds the farther one's residual-transmittance weight (0.0099).
        let near_contribution = 0.99 * 1.0;
        assert!(
            depth < near_contribution * 1.2,
            "nearer Gaussian should dominate accumulated depth: depth={depth}"
        );
    }

    #[test]
    fn test_blend_pixel_depth_zero_when_pixel_uncovered() {
        // A Gaussian far from the queried pixel contributes ~0 weight, so
        // depth should stay at (approximately) zero rather than picking up
        // a stray, uncontributed value.
        let config = RasterConfig::new().with_resolution(64, 64);
        let rasterizer = CpuRasterizer::new(config);

        let g = ProjectedGaussian {
            mean2d: na::Vector2::new(1000.5, 1000.5), // far outside the pixel
            conic: na::Vector3::new(10.0, 0.0, 10.0),
            color: na::Vector3::new(1.0, 0.0, 0.0),
            opacity: 1.0,
            depth: 42.0,
        };

        let (_pixel, depth) = rasterizer.blend_pixel(10.0, 10.0, &[g]);
        assert_eq!(depth, 0.0, "uncovered pixel should have zero depth");
    }

    #[test]
    fn test_eval_sh_rejects_degree_above_three() {
        // A model claiming sh_degree > 3 must error out rather than
        // silently clamp to 3 and read coefficients at the wrong stride.
        let config = RasterConfig::new().with_resolution(4, 4);
        let rasterizer = CpuRasterizer::new(config);
        let model = GaussianModel {
            gaussians: Vec::new(),
            // Give it *some* coefficients so the failure is unambiguously
            // about the degree check, not an out-of-bounds slice: degree-4
            // stride would be (4+1)^2*3 = 75.
            sh_coeffs: vec![0.0f32; 75],
            sh_degree: 4,
            face_indices: Vec::new(),
            barycentric: Vec::new(),
            local_offsets: Vec::new(),
            is_rigid: Vec::new(),
        };
        let dir = na::Vector3::new(0.0, 0.0, 1.0);
        let result = rasterizer.eval_sh(&model, 0, dir);
        assert!(result.is_err(), "sh_degree > 3 must be rejected");
    }
}