[mdc.md] Plan for Manifold Dual Contouring (MDC) in OpenCL with detailed steps for input, processing, and output

Author: 3dJan

gladius/thegreatplan/mdc.md

@@ -0,0 +1,45 @@
+Implement Manifold Dual Contouring (MDC) in OpenCL with the following requirements:
+
+1. **Input**:
+   - A 3D grid of cells representing an implicit surface (Signed Distance Field).
+   - Hermite data: intersection points and normals for edges crossing the surface.
+
+2. **Steps**:
+   - **Kernel 1 (QEF Solve)**:
+     - For each cell, gather Hermite samples.
+     - Compute Quadratic Error Function (QEF) to find optimal vertex position.
+     - Detect ambiguous topology (e.g., more than 4 samples or complex sign changes).
+     - Output:
+       - `baseVertices[cellId]` (float4)
+       - `splitFlags[cellId]` (int: number of extra vertices needed)
+
+   - **Prefix Sum (Blelloch Scan)**:
+     - Implement an exclusive prefix sum on `splitFlags` to compute `vertexOffsets`.
+     - This determines where each thread writes its duplicate vertices in the global buffer.
+
+   - **Kernel 2 (Vertex Emission)**:
+     - Write primary vertex to `outVertices[cellId]`.
+     - If `splitFlags[cellId] > 0`, emit duplicates at positions:
+       `outVertices[vertexOffsets[cellId] + i]` for `i` in [0, splitFlags[cellId]-1].
+     - Apply small offsets to duplicates to maintain manifoldness.
+
+   - **Kernel 3 (Face Emission)**:
+     - Build faces referencing correct vertex IDs (including duplicates).
+     - Ensure manifold connectivity (each edge belongs to exactly two faces).
+
+3. **Blelloch Scan Details**:
+   - Implement upsweep and downsweep phases for prefix sum in OpenCL.
+   - Use local memory for block-level scan and global memory for inter-block aggregation.
+
+4. **Output**:
+   - A watertight, manifold mesh with vertices and faces stored in global buffers.
+
+5. **Performance Considerations**:
+   - Use work-groups for scan operations.
+   - Avoid race conditions with atomic operations or prefix sums.
+   - Optimize QEF solver using small matrix inversion or SVD.
+
+Provide:
+- Full OpenCL kernels for QEF pass, Blelloch scan, vertex emission, and face emission.
+- Host-side code to orchestrate kernel launches and buffer allocations.
+- Comments explaining each step.