refactor: use new Cuda handle to enable runtime loading of driver

CMakeLists.txt

@@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.20 FATAL_ERROR)
 project(
   ViennaRay
   LANGUAGES CXX
-  VERSION 4.0.0)
+  VERSION 4.1.0)
 
 # --------------------------------------------------------------------------------------------------------
 # Library switches
@@ -99,7 +99,7 @@ include("cmake/cpm.cmake")
 
 CPMAddPackage(
   NAME ViennaCore
-  VERSION 2.0.0
+  VERSION 2.1.0
   GIT_REPOSITORY "https://github.com/ViennaTools/ViennaCore"
   OPTIONS "VIENNACORE_USE_GPU ${VIENNARAY_USE_GPU}")
 

README.md

@@ -63,7 +63,7 @@ We recommend using [CPM.cmake](https://github.com/cpm-cmake/CPM.cmake) to consum
 * Installation with CPM
 
   ```cmake
-  CPMAddPackage("gh:viennatools/viennaray@4.0.0") # Use the latest release version
+  CPMAddPackage("gh:viennatools/viennaray@4.1.0") # Use the latest release version
   ```
 
 * With a local installation

gpu/examples/trenchTriangles.cpp

@@ -1,6 +1,5 @@
 #include "gpu/raygTraceTriangle.hpp"
 
-#include <cuda_runtime.h>
 #include <omp.h>
 
 // #define COUNT_RAYS

include/viennaray/gpu/raygDiskGeometry.hpp

@@ -170,20 +170,21 @@ struct DiskGeometry {
     CudaBuffer outputBuffer;
     outputBuffer.alloc(blasBufferSizes.outputSizeInBytes);
 
-    optixAccelBuild(context.optix, 0, &accelOptions, diskInput.data(),
-                    numBuildInputs, tempBuffer.dPointer(),
-                    tempBuffer.sizeInBytes, outputBuffer.dPointer(),
-                    outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelBuild(
+        context.optix, 0, &accelOptions, diskInput.data(), numBuildInputs,
+        tempBuffer.dPointer(), tempBuffer.sizeInBytes, outputBuffer.dPointer(),
+        outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1));
+    context.sync();
 
     // perform compaction
     uint64_t compactedSize;
     compactedSizeBuffer.download(&compactedSize, 1);
 
     asBuffer.alloc(compactedSize);
-    optixAccelCompact(context.optix, 0, asHandle, asBuffer.dPointer(),
-                      asBuffer.sizeInBytes, &asHandle);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelCompact(context.optix, 0, asHandle,
+                                  asBuffer.dPointer(), asBuffer.sizeInBytes,
+                                  &asHandle));
+    context.sync();
 
     // clean up
     outputBuffer.free(); // << the UNcompacted, temporary output buffer

include/viennaray/gpu/raygLineGeometry.hpp

@@ -146,20 +146,21 @@ struct LineGeometry {
     CudaBuffer outputBuffer;
     outputBuffer.alloc(blasBufferSizes.outputSizeInBytes);
 
-    optixAccelBuild(context.optix, 0, &accelOptions, lineInput.data(),
-                    numBuildInputs, tempBuffer.dPointer(),
-                    tempBuffer.sizeInBytes, outputBuffer.dPointer(),
-                    outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelBuild(
+        context.optix, 0, &accelOptions, lineInput.data(), numBuildInputs,
+        tempBuffer.dPointer(), tempBuffer.sizeInBytes, outputBuffer.dPointer(),
+        outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1));
+    context.sync();
 
     // perform compaction
     uint64_t compactedSize;
     compactedSizeBuffer.download(&compactedSize, 1);
 
     asBuffer.alloc(compactedSize);
-    optixAccelCompact(context.optix, 0, asHandle, asBuffer.dPointer(),
-                      asBuffer.sizeInBytes, &asHandle);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelCompact(context.optix, 0, asHandle,
+                                  asBuffer.dPointer(), asBuffer.sizeInBytes,
+                                  &asHandle));
+    context.sync();
 
     // clean up
     outputBuffer.free(); // << the UNcompacted, temporary output buffer

include/viennaray/gpu/raygTrace.hpp

@@ -229,11 +229,9 @@ template <class T, int D> class Trace {
       VIENNACORE_LOG_WARNING(
           "Passed cell data does not match number of elements.");
     }
-    cellDataBuffer_ = passedCellDataBuffer;
-#ifndef NDEBUG
-    // In debug mode, we set the buffer as reference to avoid accidental frees
-    cellDataBuffer_.isRef = true;
-#endif
+    cellDataBuffer_ = passedCellDataBuffer; // note: this is a reference, and
+                                            // does not own the memory, so no
+                                            // free() in destructor
     numCellData_ = numData;
   }
 
@@ -397,7 +395,7 @@ template <class T, int D> class Trace {
     }
     directCallablePGs_.clear();
     for (auto &s : streams_) {
-      CUDA_CHECK(cuStreamDestroy(s));
+      context_->ch.cuStreamDestroy_(s);
     }
   }
 
@@ -429,7 +427,7 @@ template <class T, int D> class Trace {
     // each particle gets its own stream
     streams_.resize(particles_.size());
     for (size_t i = 0; i < particles_.size(); i++) {
-      CUDA_CHECK(cuStreamCreate(&streams_[i], CU_STREAM_DEFAULT));
+      context_->ch.cuStreamCreate_(&streams_[i], CU_STREAM_DEFAULT);
     }
 
     return numFluxes_;
@@ -469,7 +467,7 @@ template <class T, int D> class Trace {
       return;
 
     for (auto &s : streams_) {
-      CUDA_CHECK(cuStreamSynchronize(s));
+      context_->ch.cuStreamSynchronize_(s);
     }
     isSynced_ = true;
   }

include/viennaray/gpu/raygTriangleGeometry.hpp

@@ -70,6 +70,8 @@ struct TriangleGeometry {
     // device pointers
     CUdeviceptr d_geoVertices = geometryVertexBuffer.dPointer();
     CUdeviceptr d_geoIndices = geometryIndexBuffer.dPointer();
+    assert((d_geoVertices != 0) && (d_geoIndices != 0) &&
+           "Device pointers for geometry vertices and indices cannot be null!");
 
     triangleInput[0].triangleArray.vertexFormat = OPTIX_VERTEX_FORMAT_FLOAT3;
     triangleInput[0].triangleArray.vertexStrideInBytes = sizeof(Vec3Df);
@@ -160,20 +162,21 @@ struct TriangleGeometry {
     CudaBuffer outputBuffer;
     outputBuffer.alloc(blasBufferSizes.outputSizeInBytes);
 
-    optixAccelBuild(context.optix, 0, &accelOptions, triangleInput.data(),
-                    numBuildInputs, tempBuffer.dPointer(),
-                    tempBuffer.sizeInBytes, outputBuffer.dPointer(),
-                    outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelBuild(
+        context.optix, 0, &accelOptions, triangleInput.data(), numBuildInputs,
+        tempBuffer.dPointer(), tempBuffer.sizeInBytes, outputBuffer.dPointer(),
+        outputBuffer.sizeInBytes, &asHandle, &emitDesc, 1));
+    context.sync();
 
     // perform compaction
     uint64_t compactedSize;
     compactedSizeBuffer.download(&compactedSize, 1);
 
     asBuffer.alloc(compactedSize);
-    optixAccelCompact(context.optix, 0, asHandle, asBuffer.dPointer(),
-                      asBuffer.sizeInBytes, &asHandle);
-    cuCtxSynchronize();
+    OPTIX_CHECK(optixAccelCompact(context.optix, 0, asHandle,
+                                  asBuffer.dPointer(), asBuffer.sizeInBytes,
+                                  &asHandle));
+    context.sync();
 
     // clean up
     outputBuffer.free(); // << the UNcompacted, temporary output buffer