fix to endcaps. fixing incorrect unmap of instance buffer when made a…

…s host buffer
gprt-org · Feb 16, 2025 · 1aec048 · 1aec048
1 parent 63b7975
commit 1aec048
Show file tree

Hide file tree

Showing 3 changed files with 13 additions and 9 deletions.
diff --git a/gprt/cmake/embed_devicecode.cmake b/gprt/cmake/embed_devicecode.cmake
@@ -46,7 +46,7 @@ function(embed_devicecode)
     -zero-initialize # zero-initialize all variables
     -Wno-39001 # for VK_EXT_mutable_descriptor_type, allows overlapping bindings
     -g
-    -O0
+    -O3
     -I ${GPRT_INCLUDE_DIR}
     -o ${CMAKE_CURRENT_BINARY_DIR}/${EMBED_DEVICECODE_OUTPUT_TARGET}.spv
     DEPENDS ${EMBED_DEVICECODE_SOURCES} ${EMBED_DEVICECODE_HEADERS} ${GPRT_INCLUDE_DIR}/gprt.slang ${GPRT_INCLUDE_DIR}/gprt.h

diff --git a/gprt/gprt.cpp b/gprt/gprt.cpp
@@ -4577,10 +4577,12 @@ Context::buildSBT(GPRTBuildSBTFlags flags) {
         VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
         // means we can use this buffer as a storage buffer resource
         VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
-    const VkMemoryPropertyFlags memoryUsageFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |   // mappable to host with
-                                                                                           // vkMapMemory
-                                                   VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;   // means "flush" and
-                                                                                           // "invalidate" not needed
+    const VkMemoryPropertyFlags memoryUsageFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;   // means most efficient for
+                                                                                        // device access
+    // const VkMemoryPropertyFlags memoryUsageFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |   // mappable to host with
+    //                                                                                        // vkMapMemory
+    //                                                VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;   // means "flush" and
+    //                                                                                        // "invalidate" not needed
 
     // std::cout<<"Todo, get some smarter memory allocation working..."
     // <<std::endl;
@@ -4724,7 +4726,9 @@ Context::buildSBT(GPRTBuildSBTFlags flags) {
           // SBT data... So, if we have a bunch of instances set by reference on
           // device, we need to eventually do something smarter here...
           // Todo, move this setup to the device...
-          instanceAccel->instancesBuffer->map();
+
+          bool previouslyMapped = (instanceAccel->instancesBuffer->mapped != nullptr);
+          if (!previouslyMapped) instanceAccel->instancesBuffer->map();
           gprt::Instance *instances = (gprt::Instance *) instanceAccel->instancesBuffer->mapped;
 
           for (uint32_t blasID = 0; blasID < instanceAccel->numInstances; ++blasID) {
@@ -4797,7 +4801,7 @@ Context::buildSBT(GPRTBuildSBTFlags flags) {
             }
           }
 
-          instanceAccel->instancesBuffer->unmap();
+          if (!previouslyMapped) instanceAccel->instancesBuffer->unmap();
         }
       }
       hitgroupTable->unmap();

diff --git a/gprt/gprt_host.h b/gprt/gprt_host.h
@@ -259,8 +259,8 @@ typedef enum {
 // } GPRTCellType;
 
 typedef enum {
-  GPRT_LSS_CHAINED_END_CAPS = 0,
-  GPRT_LSS_NO_END_CAPS = 1
+  GPRT_LSS_CHAINED_END_CAPS = 1,
+  GPRT_LSS_NO_END_CAPS = 2
 } GPRTLSSFlags;
 
 typedef enum {