[WIP] port forest tutorial to new host/device memory support

RenderKit · Nov 27, 2024 · 96782b9 · 96782b9
1 parent 190b4aa
commit 96782b9
Show file tree

Hide file tree

Showing 9 changed files with 160 additions and 109 deletions.
diff --git a/common/cmake/dpcpp.cmake b/common/cmake/dpcpp.cmake
@@ -52,14 +52,7 @@ IF (EMBREE_SYCL_SUPPORT)
     SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} /debug:none")    # FIXME: debug information generation takes forever in SYCL
     SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} /DNDEBUG")    # FIXME: debug information generation takes forever in SYCL
   ELSE()
-    message("CMAKE_BUILD_TYPE ${CMAKE_BUILD_TYPE}")
-    if((CMAKE_BUILD_TYPE STREQUAL "Debug"))
-      message(STATUS "generate debug information")
-      SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} -g")            # FIXME: debug information generation takes forever in SYCL
-    else()
-      message(STATUS "generate NO debug information")
-      SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} -g0")            # FIXME: debug information generation takes forever in SYCL
-    endif()
+    SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} -g0")            # FIXME: debug information generation takes forever in SYCL
     SET(CMAKE_CXX_FLAGS_SYCL "${CMAKE_CXX_FLAGS_SYCL} -UDEBUG -DNDEBUG") # FIXME: assertion still not working in SYCL
   ENDIF()
 

diff --git a/common/sys/alloc.cpp b/common/sys/alloc.cpp
@@ -89,10 +89,14 @@ namespace embree
     total_allocations++;    
 
     void* ptr = nullptr;
-    if (mode == EMBREE_USM_SHARED_DEVICE_READ_ONLY)
+    if (mode == EMBREE_USM_SHARED_DEVICE_READ_ONLY) {
+      std::cout << "XXX shared memory allocation!" << std::endl;
       ptr = sycl::aligned_alloc_shared(align,size,*device,*context,sycl::ext::oneapi::property::usm::device_read_only());
-    else
+    }
+    else {
+      std::cout << "XXX shared memory allocation!" << std::endl;
       ptr = sycl::aligned_alloc_shared(align,size,*device,*context);
+    }
 
     if (size != 0 && ptr == nullptr)
       throw std::bad_alloc();
@@ -114,9 +118,15 @@ namespace embree
     void* ptr = nullptr;
     if (type == EmbreeMemoryType::SHARED) {
       if (mode == EMBREE_USM_SHARED_DEVICE_READ_ONLY)
+      {
+        std::cout << "XXX shared memory allocation!" << std::endl;
         ptr = sycl::aligned_alloc_shared(align,size,*device,*context,sycl::ext::oneapi::property::usm::device_read_only());
+      }
       else
+      {
+        std::cout << "XXX shared memory allocation!" << std::endl;
         ptr = sycl::aligned_alloc_shared(align,size,*device,*context);
+      }
     }
     else if (type == EmbreeMemoryType::HOST) {
       ptr = sycl::aligned_alloc_host(align,size,*context);

diff --git a/include/embree4/rtcore_buffer.h b/include/embree4/rtcore_buffer.h
@@ -57,6 +57,11 @@ RTC_API void rtcCommitBufferWithQueue(RTCBuffer buffer, sycl::queue queue);
 /* Returns a pointer to the buffer data. */
 RTC_API void* rtcGetBufferData(RTCBuffer buffer);
 
+/* Returns a pointer to the buffer data on the device. Returns the same pointer as
+  rtcGetBufferData if the device is no SYCL device or if Embree is executed on a
+  system with unified memory (e.g., iGPUs). */
+RTC_API void* rtcGetBufferDataDevice(RTCBuffer buffer);
+
 /* Retains the buffer (increments the reference count). */
 RTC_API void rtcRetainBuffer(RTCBuffer buffer);
 

diff --git a/kernels/common/buffer.h b/kernels/common/buffer.h
@@ -104,6 +104,21 @@ namespace embree
       return ptr;
     }
 
+    /*! gets buffer pointer */
+    void* dataDevice()
+    {
+      /* report error if buffer is not existing */
+      if (!device)
+        throw_RTCError(RTC_ERROR_INVALID_ARGUMENT, "invalid buffer specified");
+
+      /* return buffer */
+#if defined(EMBREE_SYCL_SUPPORT)
+      return dptr;
+#else
+      return ptr;
+#endif
+    }
+
     /*! returns pointer to first element */
     __forceinline char* getPtr() const {
       return ptr;

diff --git a/kernels/common/rtcore.cpp b/kernels/common/rtcore.cpp
@@ -242,6 +242,18 @@ RTC_NAMESPACE_BEGIN;
     return nullptr;
   }
 
+  RTC_API void* rtcGetBufferDataDevice(RTCBuffer hbuffer)
+  {
+    Buffer* buffer = (Buffer*)hbuffer;
+    RTC_CATCH_BEGIN;
+    RTC_TRACE(rtcGetBufferDataDevice);
+    RTC_VERIFY_HANDLE(hbuffer);
+    RTC_ENTER_DEVICE(hbuffer);
+    return buffer->dataDevice();
+    RTC_CATCH_END2(buffer);
+    return nullptr;
+  }
+
   RTC_API void* rtcGetBufferData(RTCBuffer hbuffer)
   {
     Buffer* buffer = (Buffer*)hbuffer;

diff --git a/kernels/sycl/rthwif_embree_builder.cpp b/kernels/sycl/rthwif_embree_builder.cpp
@@ -673,7 +673,7 @@ namespace embree
       throw_RTCError(RTC_ERROR_INVALID_ARGUMENT, "AccelBuffer constructor called with non-GPU device");
     }
 
-    unifiedMemory = true; //gpu_device->has_unified_memory();
+    unifiedMemory = gpu_device->has_unified_memory();
 
     if (unifiedMemory)
     {

diff --git a/tutorials/CMakeLists.txt b/tutorials/CMakeLists.txt
@@ -79,40 +79,40 @@ IF (EMBREE_TUTORIALS_GLFW)
 ENDIF()
 
 ADD_SUBDIRECTORY(common)
-ADD_SUBDIRECTORY(embree_info)
-ADD_SUBDIRECTORY(minimal)
-ADD_SUBDIRECTORY(verify)
-ADD_SUBDIRECTORY(triangle_geometry)
-ADD_SUBDIRECTORY(dynamic_scene)
-ADD_SUBDIRECTORY(voronoi)
-ADD_SUBDIRECTORY(closest_point)
-ADD_SUBDIRECTORY(user_geometry)
-ADD_SUBDIRECTORY(viewer)
-ADD_SUBDIRECTORY(instanced_geometry)
-ADD_SUBDIRECTORY(multi_instanced_geometry)
-ADD_SUBDIRECTORY(intersection_filter)
-ADD_SUBDIRECTORY(pathtracer)
-ADD_SUBDIRECTORY(hair_geometry)
-ADD_SUBDIRECTORY(subdivision_geometry)
-ADD_SUBDIRECTORY(displacement_geometry)
-ADD_SUBDIRECTORY(grid_geometry)
-ADD_SUBDIRECTORY(bvh_builder)
-ADD_SUBDIRECTORY(lazy_geometry)
-ADD_SUBDIRECTORY(bvh_access)
-ADD_SUBDIRECTORY(quaternion_motion_blur)
-ADD_SUBDIRECTORY(motion_blur_geometry)
-ADD_SUBDIRECTORY(interpolation)
-ADD_SUBDIRECTORY(curve_geometry)
-ADD_SUBDIRECTORY(point_geometry)
-ADD_SUBDIRECTORY(buildbench)
-ADD_SUBDIRECTORY(convert)
-ADD_SUBDIRECTORY(collide)
-ADD_SUBDIRECTORY(next_hit)
-ADD_SUBDIRECTORY(multiscene_geometry)
-ADD_SUBDIRECTORY(ray_mask)
+#ADD_SUBDIRECTORY(embree_info)
+#ADD_SUBDIRECTORY(minimal)
+#ADD_SUBDIRECTORY(verify)
+#ADD_SUBDIRECTORY(triangle_geometry)
+#ADD_SUBDIRECTORY(dynamic_scene)
+#ADD_SUBDIRECTORY(voronoi)
+#ADD_SUBDIRECTORY(closest_point)
+#ADD_SUBDIRECTORY(user_geometry)
+#ADD_SUBDIRECTORY(viewer)
+#ADD_SUBDIRECTORY(instanced_geometry)
+#ADD_SUBDIRECTORY(multi_instanced_geometry)
+#ADD_SUBDIRECTORY(intersection_filter)
+#ADD_SUBDIRECTORY(pathtracer)
+#ADD_SUBDIRECTORY(hair_geometry)
+#ADD_SUBDIRECTORY(subdivision_geometry)
+#ADD_SUBDIRECTORY(displacement_geometry)
+#ADD_SUBDIRECTORY(grid_geometry)
+#ADD_SUBDIRECTORY(bvh_builder)
+#ADD_SUBDIRECTORY(lazy_geometry)
+#ADD_SUBDIRECTORY(bvh_access)
+#ADD_SUBDIRECTORY(quaternion_motion_blur)
+#ADD_SUBDIRECTORY(motion_blur_geometry)
+#ADD_SUBDIRECTORY(interpolation)
+#ADD_SUBDIRECTORY(curve_geometry)
+#ADD_SUBDIRECTORY(point_geometry)
+#ADD_SUBDIRECTORY(buildbench)
+#ADD_SUBDIRECTORY(convert)
+#ADD_SUBDIRECTORY(collide)
+#ADD_SUBDIRECTORY(next_hit)
+#ADD_SUBDIRECTORY(multiscene_geometry)
+#ADD_SUBDIRECTORY(ray_mask)
 ADD_SUBDIRECTORY(forest)
-ADD_SUBDIRECTORY(debug_device_memory)
-ADD_SUBDIRECTORY(embree_tests)
+#ADD_SUBDIRECTORY(debug_device_memory)
+#ADD_SUBDIRECTORY(embree_tests)
 
 
 ENDIF()