Location

src/backend/cpu/
  cpu_backend.c       — RHI function table implementation

src/rasterizer/
  rasterizer.h        — Shared software rasterizer interface
  rasterizer.c        — Triangle rasterization, clipping, line drawing

The CPU backend is a thin RHI adapter that delegates all rasterization to the shared software rasterizer at src/rasterizer/.

Capabilities

The CPU backend is always available. It requires no GPU, no drivers, and no platform-specific code.

Internal Types

MopSwFramebuffer

typedef struct MopSwFramebuffer {
    int       width, height;
    uint8_t  *color;       /* RGBA8, size = w * h * 4 */
    float    *depth;       /* float,  size = w * h     */
    uint32_t *object_id;   /* uint32, size = w * h     */
} MopSwFramebuffer;

RHI Handle Mapping

Rasterization Pipeline

For each draw call:

1. Vertex fetch: Read MopVertex array from buffer
2. Triangle loop: For each 3 indices: a. MVP transform: Multiply position by mvp → clip space b. Normal transform: Multiply normal by model matrix (upper 3x3) → world space c. Frustum clipping: Sutherland-Hodgman against 6 planes d. Triangle fan: Clipped polygon → fan of sub-triangles e. For each sub-triangle:
   • Perspective divide: xyz / w → NDC [-1, 1]
   • Viewport transform: NDC → screen pixels (Y flipped for top-left origin)
   • Backface cull: Screen-space signed area test
   • Flat shading: Average face normal · directional light
   • Rasterize: Half-space edge functions (solid) or Bresenham (wireframe)
   • Per-pixel: Depth test → write color, depth, object_id

Clipping

Uses Sutherland-Hodgman algorithm. The triangle is clipped sequentially against 6 frustum planes in clip space:

+X: w + x ≥ 0
-X: w - x ≥ 0
+Y: w + y ≥ 0
-Y: w - y ≥ 0
+Z: w + z ≥ 0
-Z: w - z ≥ 0

Maximum output vertices after clipping a single triangle: 24 (each plane can add at most one vertex; 3 + 6 × 3 < 24 with some margin).

Lighting

Fixed directional light at (0.3, 1.0, 0.5) (world space). Ambient factor: 0.2. Diffuse factor: 0.8 * max(N·L, 0). Total: clamp(ambient + diffuse).

Edge Function Rasterization

For each pixel in the triangle's bounding box:

w0 = edge(v1, v2, pixel)
w1 = edge(v2, v0, pixel)
w2 = edge(v0, v1, pixel)

CW and CCW triangles are both handled: if the triangle's signed area is negative (CW after Y-flip), all edge values are negated before the ≥ 0 test. Barycentric weights are normalized by absolute area.