Export and wrap test buffer

Samples/5_Domain_Specific/simpleVulkanMMAP/MonteCarloPi.cu

@@ -201,7 +201,7 @@ void MonteCarloPiSimulation::getIdealExecutionConfiguration() {
 void MonteCarloPiSimulation::setupSimulationAllocations() {
   CUdeviceptr d_ptr = 0U;
   size_t granularity = 0;
-  CUmemGenericAllocationHandle cudaPositionHandle, cudaInCircleHandle;
+  CUmemGenericAllocationHandle cudaPositionHandle, cudaInCircleHandle, cudaBDAHandle;
 
   CUmemAllocationProp allocProp = {};
   allocProp.type = CU_MEM_ALLOCATION_TYPE_PINNED;
@@ -222,11 +222,13 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
 
   size_t xyPositionVecSize = m_numPoints * sizeof(*m_xyVector);
   size_t inCircleVecSize = m_numPoints * sizeof(*m_pointsInsideCircle);
+  size_t bdaVecSize = sizeof(float) * 2;
 
   size_t xyPositionSize =
       ROUND_UP_TO_GRANULARITY(xyPositionVecSize, granularity);
   size_t inCircleSize = ROUND_UP_TO_GRANULARITY(inCircleVecSize, granularity);
-  m_totalAllocationSize = (xyPositionSize + inCircleSize);
+  size_t bdaSize = ROUND_UP_TO_GRANULARITY(bdaVecSize, granularity);
+  m_totalAllocationSize = (xyPositionSize + inCircleSize + bdaSize);
 
   // Reserve the required contiguous VA space for the allocations
   checkCudaErrors(
@@ -240,6 +242,8 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
       cuMemCreate(&cudaPositionHandle, xyPositionSize, &allocProp, 0));
   checkCudaErrors(
       cuMemCreate(&cudaInCircleHandle, inCircleSize, &allocProp, 0));
+  checkCudaErrors(
+      cuMemCreate(&cudaBDAHandle, bdaSize, &allocProp, 0));
 
   // Export the allocation to a platform-specific handle. The type of handle
   // requested here must match the requestedHandleTypes field in the prop
@@ -250,9 +254,12 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
   checkCudaErrors(
       cuMemExportToShareableHandle((void *)&m_inCircleShareableHandle,
                                    cudaInCircleHandle, ipcHandleTypeFlag, 0));
+  checkCudaErrors(cuMemExportToShareableHandle(
+      (void *)&m_bdaShareableHandle, cudaBDAHandle, ipcHandleTypeFlag, 0));
 
   CUdeviceptr va_position = d_ptr;
   CUdeviceptr va_InCircle = va_position + xyPositionSize;
+  CUdeviceptr va_BDA = va_InCircle + inCircleSize;
   m_pointsInsideCircle = (float *)va_InCircle;
   m_xyVector = (vec2 *)va_position;
 
@@ -261,6 +268,8 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
       cuMemMap(va_position, xyPositionSize, 0, cudaPositionHandle, 0));
   checkCudaErrors(
       cuMemMap(va_InCircle, inCircleSize, 0, cudaInCircleHandle, 0));
+  checkCudaErrors(
+      cuMemMap(va_BDA, bdaSize, 0, cudaBDAHandle, 0));
 
   // Release the handles for the allocation. Since the allocation is currently
   // mapped to a VA range with a previous call to cuMemMap the actual freeing of
@@ -268,6 +277,7 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
   // allocation will be kept live until it is unmapped.
   checkCudaErrors(cuMemRelease(cudaPositionHandle));
   checkCudaErrors(cuMemRelease(cudaInCircleHandle));
+  checkCudaErrors(cuMemRelease(cudaBDAHandle));
 
   CUmemAccessDesc accessDescriptor = {};
   accessDescriptor.location.id = m_cudaDevice;
@@ -278,6 +288,10 @@ void MonteCarloPiSimulation::setupSimulationAllocations() {
   // Read-Write access to the range.
   checkCudaErrors(
       cuMemSetAccess(d_ptr, m_totalAllocationSize, &accessDescriptor, 1));
+
+  // fill the BDA buffer with something
+  const float bdaValues[2] = { 42.0f, 17.0f };
+  cuMemcpyHtoD(va_BDA, &bdaValues[0], sizeof(float) * 2);
 }
 
 void MonteCarloPiSimulation::cleanupSimulationAllocations() {
@@ -290,6 +304,7 @@ void MonteCarloPiSimulation::cleanupSimulationAllocations() {
 
     checkIpcErrors(ipcCloseShareableHandle(m_posShareableHandle));
     checkIpcErrors(ipcCloseShareableHandle(m_inCircleShareableHandle));
+    checkIpcErrors(ipcCloseShareableHandle(m_bdaShareableHandle));
 
     // Free the virtual address region.
     checkCudaErrors(

Samples/5_Domain_Specific/simpleVulkanMMAP/MonteCarloPi.h

@@ -61,7 +61,7 @@ class MonteCarloPiSimulation {
   // Shareable Handles(a file descriptor on Linux and NT Handle on Windows),
   // used for sharing cuda
   // allocated memory with Vulkan
-  ShareableHandle m_posShareableHandle, m_inCircleShareableHandle;
+  ShareableHandle m_posShareableHandle, m_inCircleShareableHandle, m_bdaShareableHandle;
 
   // Cuda Device corresponding to the Vulkan Physical device
   int m_cudaDevice;
@@ -90,6 +90,7 @@ class MonteCarloPiSimulation {
   ShareableHandle &getInCircleShareableHandle() {
     return m_inCircleShareableHandle;
   }
+  ShareableHandle &getBDAShareableHandle() { return m_bdaShareableHandle; }
 };
 
 #endif  // __PISIM_H__

Samples/5_Domain_Specific/simpleVulkanMMAP/main.cpp

@@ -59,8 +59,8 @@ std::string execution_path;
 class VulkanCudaPi : public VulkanBaseApp {
   typedef struct UniformBufferObject_st { float frame; } UniformBufferObject;
 
-  VkBuffer m_inCircleBuffer, m_xyPositionBuffer;
-  VkDeviceMemory m_inCircleMemory, m_xyPositionMemory;
+  VkBuffer m_inCircleBuffer, m_xyPositionBuffer, m_bdaBuffer;
+  VkDeviceMemory m_inCircleMemory, m_xyPositionMemory, m_bdaMemory;
   VkSemaphore m_vkWaitSemaphore, m_vkSignalSemaphore;
   MonteCarloPiSimulation m_sim;
   UniformBufferObject m_ubo;
@@ -75,8 +75,10 @@ class VulkanCudaPi : public VulkanBaseApp {
       : VulkanBaseApp("simpleVulkanMMAP", ENABLE_VALIDATION),
         m_inCircleBuffer(VK_NULL_HANDLE),
         m_xyPositionBuffer(VK_NULL_HANDLE),
+        m_bdaBuffer(VK_NULL_HANDLE),
         m_inCircleMemory(VK_NULL_HANDLE),
         m_xyPositionMemory(VK_NULL_HANDLE),
+        m_bdaMemory(VK_NULL_HANDLE),
         m_sim(num_points),
         m_ubo(),
         m_stream(0),
@@ -123,6 +125,12 @@ class VulkanCudaPi : public VulkanBaseApp {
     if (m_inCircleMemory != VK_NULL_HANDLE) {
       vkFreeMemory(m_device, m_inCircleMemory, nullptr);
     }
+    if (m_bdaBuffer != VK_NULL_HANDLE) {
+      vkDestroyBuffer(m_device, m_bdaBuffer, nullptr);
+    }
+    if (m_bdaMemory != VK_NULL_HANDLE) {
+      vkFreeMemory(m_device, m_bdaMemory, nullptr);
+    }
   }
 
   void fillRenderingCommandBuffer(VkCommandBuffer& commandBuffer) {
@@ -226,14 +234,21 @@ class VulkanCudaPi : public VulkanBaseApp {
         VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, m_inCircleBuffer,
         m_inCircleMemory);
 
+    importExternalBuffer(
+        (void*)(uintptr_t)m_sim.getBDAShareableHandle(),
+        getDefaultMemHandleType(), sizeof(float) * 2,
+        VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
+        VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, m_bdaBuffer,
+        m_bdaMemory);
+
     // (SE) get function ptr
     auto* vkGetBufferDeviceAddressKHR = (PFN_vkGetBufferDeviceAddressKHR)vkGetDeviceProcAddr(m_device, "vkGetBufferDeviceAddressKHR");
     std::cout << "DEBUG: vkGetBufferDeviceAddressKHR = " << (void*)vkGetBufferDeviceAddressKHR << std::endl;
   
-    // get BDA for the circle buffer
+    // get BDA
     VkBufferDeviceAddressInfoKHR bda_info{VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO_KHR,
                                           nullptr,
-                                          m_inCircleBuffer};
+                                          m_bdaBuffer};
     auto bda = vkGetBufferDeviceAddressKHR(m_device, &bda_info);
     std::cout << "DEBUG: BDA = " << (void*)bda << std::endl;