Merge pull request 'perf(mesh): dirty-range vertex upload for deforming-mesh Refit (#119)' (#138) from claude/issue-119 into master

tests/BLASBuildOptions/main.cpp

@@ -157,6 +157,42 @@ int main() {
     Check(tri.blasAddr == triAddrBefore,
           "Refit kept the same blasAddr (instances stay valid)");
 
+    // Dirty-range refit (issue #119): move only a sub-window of the vertices
+    // and pass it as [offset, count). The ranged upload patches just those
+    // vertices in place; the UPDATE must still keep the same AS handle /
+    // blasAddr, and the validation layer (checked at the end) is the real proof
+    // that the ranged host-write / flush / barrier — and the staged variant
+    // below — are spec-correct. The untouched vertices retain their last values
+    // in the device buffer, so the refit BLAS is over the full deformed cube.
+    {
+        // Full array of moved positions; we declare only the back half [4,8)
+        // as the dirty window, so only those 4 vertices are re-uploaded.
+        auto verts = CubeVerts(2.0f);
+        auto idx   = CubeIndices();
+        VkCommandBuffer cmd = BeginCmd();
+        tri.Refit(verts, idx, /*dirtyVertexOffset*/ 4, /*dirtyVertexCount*/ 4, cmd);
+        SubmitWait(cmd);
+    }
+    Check(tri.accelerationStructure == triHandleBefore,
+          "dirty-range Refit kept the same AS handle (in-place UPDATE, #119)");
+    Check(tri.blasAddr == triAddrBefore,
+          "dirty-range Refit kept the same blasAddr (#119)");
+
+    // A dirty-range refit on a mesh whose counts changed must still fall back
+    // to a full rebuild (the dirty window is irrelevant on that path) and
+    // succeed — same robustness as the full-span Refit fallback.
+    {
+        auto verts = CubeVerts(1.0f);
+        verts.push_back({2.0f, 2.0f, 2.0f});   // count change → topology change
+        auto idx = CubeIndices();
+        idx.insert(idx.end(), {0u, 1u, 8u});
+        VkCommandBuffer cmd = BeginCmd();
+        tri.Refit(verts, idx, /*dirtyVertexOffset*/ 8, /*dirtyVertexCount*/ 1, cmd);
+        SubmitWait(cmd);
+    }
+    Check(tri.blasAddr != 0 && tri.builtPrimitiveCount == 13,
+          "dirty-range Refit with a count change rebuilt the BLAS (13 prims, #119)");
+
     // ── Triangle BLAS: fast-build, no update → flags reflect the choice. ─
     Mesh triFast;
     {
@@ -340,6 +376,19 @@ int main() {
         Check(staged.accelerationStructure == stagedHandle,
               "staged-upload Refit kept the same AS handle (in-place UPDATE, #73)");
 
+        // Dirty-range refit on the staged path (#119): the ranged upload must
+        // take the stage-a-sub-slice + vkCmdCopyBuffer-at-offset path (the
+        // buffer is pure DEVICE_LOCAL), keep the AS handle, and stay
+        // validation-clean. Only the front half [0,4) is declared dirty.
+        {
+            auto verts3 = CubeVerts(2.0f);
+            VkCommandBuffer cmd = BeginCmd();
+            staged.Refit(verts3, idx, /*dirtyVertexOffset*/ 0, /*dirtyVertexCount*/ 4, cmd);
+            SubmitWait(cmd);
+        }
+        Check(staged.accelerationStructure == stagedHandle,
+              "staged-upload dirty-range Refit kept the same AS handle (#119)");
+
         Device::directWriteBudget = savedBudget;
     }