rocr: SVMPrefetch to a particular numa node (#1063)

In order for hipMemPrefetchAysnc_v2() api to work, we need rocr to
migrates the ranges of pages requested to the particular NUMA node in
question, via move_pages().

Signed-off-by: Sunday Clement <Sunday.Clement@amd.com>

projects/rocr-runtime/CMakeLists.txt

@@ -60,6 +60,13 @@ if (NOT DEFINED BUILD_ROCR)
   set(BUILD_ROCR ON)
 endif()
 
+find_package(NUMA)
+if(NUMA_FOUND)
+  set(NUMA "${NUMA_LIBRARIES}")
+else()
+  find_library(NUMA NAMES numa REQUIRED)
+endif()
+
 function(add_rocm_subdir subdir subdir_assigns)
     message("add_rocm_subdir() -- " ${subdir})
     # message("  subdir_assigns before:" ${subdir_assigns} "EOM")
@@ -238,7 +245,7 @@ set(CPACK_DEBIAN_DEV_PACKAGE_DEPENDS "libdrm-amdgpu-dev | libdrm-dev, rocm-core,
 
 set(CPACK_DEBIAN_DEV_PACKAGE_RECOMMENDS "libdrm-amdgpu-dev")
 
-set(CPACK_DEBIAN_BINARY_PACKAGE_DEPENDS "libdrm-amdgpu-amdgpu1 | libdrm-amdgpu1, libnuma1, libelf1")
+set(CPACK_DEBIAN_BINARY_PACKAGE_DEPENDS "libdrm-amdgpu-amdgpu1 | libdrm-amdgpu1, libnuma1, numactl, libelf1")
 set(CPACK_DEBIAN_ASAN_PACKAGE_DEPENDS "libdrm-amdgpu-dev | libdrm-dev, rocm-core-asan, libdrm-amdgpu-amdgpu1 | libdrm-amdgpu1, libnuma1, libelf1")
 set(CPACK_DEBIAN_ASAN_PACKAGE_RECOMMENDS "libdrm-amdgpu-dev")
 

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp

@@ -56,6 +56,7 @@
 #include <sys/mman.h>
 #include <sys/socket.h>
 #include <sys/un.h>
+#include <numaif.h>      // for move_pages
 #else
 #define debug_warning(__VA_ARGS__)
 #endif
@@ -3354,12 +3355,29 @@ hsa_status_t Runtime::SvmPrefetch(void* ptr, size_t size, hsa_agent_t agent,
           op->dep_signals[op->remaining_deps], HSA_SIGNAL_CONDITION_EQ, 0, signal_handler, arg);
       return false;
     }
+    
+    Agent* dest = Runtime::runtime_singleton_->GetSVMPrefetchAgent(op->base, op->size);
 
-    HSA_SVM_ATTRIBUTE attrib;
-    attrib.type = HSA_SVM_ATTR_PREFETCH_LOC;
-    attrib.value = op->node_id;
-    HSAKMT_STATUS error = HSAKMT_CALL(hsaKmtSVMSetAttr(op->base, op->size, 1, &attrib));
-    assert(error == HSAKMT_STATUS_SUCCESS && "KFD Prefetch failed.");
+    if (dest == nullptr || dest->device_type() == Agent::kAmdGpuDevice) {
+      // Prefetch location is not valid for move_pages usecase.
+      HSA_SVM_ATTRIBUTE attrib;
+      attrib.type = HSA_SVM_ATTR_PREFETCH_LOC;
+      attrib.value = op->node_id;
+      HSAKMT_STATUS error = HSAKMT_CALL(hsaKmtSVMSetAttr(op->base, op->size, 1, &attrib));
+      assert(error == HSAKMT_STATUS_SUCCESS && "KFD Prefetch failed.");
+    } else {
+      // Migrate pages to the requested CPU NUMA node
+      void* base_ptr = op->base;
+      size_t num_pages = op->size / 4096;
+      std::vector<void*> pages(num_pages);
+      for (size_t i = 0; i < num_pages; ++i)
+          pages[i] = static_cast<uint8_t*>(base_ptr) + i * 4096;
+      std::vector<int> nodes(num_pages, op->node_id);
+      std::vector<int> status(num_pages, -1);
+
+      int ret = move_pages(0, num_pages, pages.data(), nodes.data(), status.data(), 0);
+      assert(ret == 0 && "move_pages failed");
+    }
 
     removePrefetchRanges(op);