⚡ Bolt: [Remove Per-Draw Vector Allocations]

.jules/bolt.md

@@ -1,3 +1,7 @@
 ## 2025-03-10 - [PPM Framebuffer IO Optimization]
 **Learning:** `std::fwrite` has overhead for every call. Writing 3 bytes per pixel individually for an 800x600 image causes 480,000 library calls per frame. Buffering an entire row of pixels (or the whole image) and making one `fwrite` call per row (or per image) provides a massive performance boost (nearly 2x speedup for the offline demo application).
 **Action:** Always batch I/O operations. When writing image files or any large binary data, buffer the data in memory and write in large chunks rather than making many small `fwrite` calls.
+
+## 2025-03-17 - [Renderer Dynamic Vector Allocation Avoidance]
+**Learning:** `std::vector` allocations in `Renderer::drawMesh` per draw call cause large memory overhead when called thousands of times per frame. Using class members instead and just calling `.resize()` helps reuse their capacity.
+**Action:** Always allocate vectors as members in render loops and prefer `.resize()` instead of local instantiations when the vector sizes match vertex or index counts.

include/soft_render/render/renderer.hpp

@@ -55,6 +55,8 @@ class Renderer {
     pipeline::FragmentShader fs_;
     pipeline::Uniforms uniforms_;
     pipeline::SceneLighting lighting_;
+    std::vector<pipeline::ClipVertex> transformedBuffer_;
+    std::vector<pipeline::Triangle> triangleBuffer_;
 };
 
 }

src/render/renderer.cpp

@@ -58,16 +58,16 @@ void Renderer::drawMesh(const pipeline::Vertex* verts, int vCount,
                          const uint32_t* indices, int iCount,
                          const pipeline::Material& mat) {
     // Transform all vertices
-    std::vector<pipeline::ClipVertex> transformed(vCount);
-    vp_.processBatch(verts, transformed.data(), vCount, uniforms_);
+    transformedBuffer_.resize(vCount);
+    vp_.processBatch(verts, transformedBuffer_.data(), vCount, uniforms_);
 
     // Build triangle list
     int triCount = iCount / 3;
-    std::vector<pipeline::Triangle> tris(triCount);
+    triangleBuffer_.resize(triCount);
     for (int i = 0; i < triCount; ++i) {
-        tris[i].v[0] = transformed[indices[i*3 + 0]];
-        tris[i].v[1] = transformed[indices[i*3 + 1]];
-        tris[i].v[2] = transformed[indices[i*3 + 2]];
+        triangleBuffer_[i].v[0] = transformedBuffer_[indices[i*3 + 0]];
+        triangleBuffer_[i].v[1] = transformedBuffer_[indices[i*3 + 1]];
+        triangleBuffer_[i].v[2] = transformedBuffer_[indices[i*3 + 2]];
     }
 
     // Fragment shader
@@ -76,7 +76,7 @@ void Renderer::drawMesh(const pipeline::Vertex* verts, int vCount,
 
     // Rasterize
     pipeline::Rasterizer rast(fb_);
-    rast.rasterizeBatch(tris.data(), triCount, fragCb);
+    rast.rasterizeBatch(triangleBuffer_.data(), triCount, fragCb);
 }
 
 void Renderer::draw(const DrawCall& call) {