enable neighbor_list use uva (PaddlePaddle#220)

Co-authored-by: root <[email protected]>
danleifeng · Sep 12, 2023 · ce95a64 · ce95a64
1 parent afec065
commit ce95a64
Show file tree

Hide file tree

Showing 2 changed files with 31 additions and 4 deletions.
diff --git a/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.cu b/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.cu
@@ -21,6 +21,11 @@
 #ifdef PADDLE_WITH_HETERPS
 #include "paddle/fluid/framework/fleet/heter_ps/gpu_graph_utils.h"
 #include "paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table.h"
+#define ALIGN_INT64(LEN) (uint64_t((LEN) + 7) & uint64_t(~7))
+#define HBMPS_MAX_BUFF 1024 * 1024
+
+DECLARE_bool(enable_neighbor_list_use_uva);
+
 namespace paddle {
 namespace framework {
 /*
@@ -882,8 +887,14 @@ void GpuPsGraphTable::build_graph_on_single_gpu(const GpuPsCommGraph& g,
     gpu_graph_list_[offset].node_size = 0;
   }
   if (g.neighbor_size) {
-    cudaError_t cudaStatus = cudaMalloc(&gpu_graph_list_[offset].neighbor_list,
+    cudaError_t cudaStatus;
+    if (!FLAGS_enable_neighbor_list_use_uva) {
+      cudaStatus = cudaMalloc(&gpu_graph_list_[offset].neighbor_list,
+                                        g.neighbor_size * sizeof(uint64_t));
+    } else {
+      cudaStatus = cudaMallocManaged(&gpu_graph_list_[offset].neighbor_list,
                                         g.neighbor_size * sizeof(uint64_t));
+    }
     PADDLE_ENFORCE_EQ(cudaStatus,
                       cudaSuccess,
                       platform::errors::InvalidArgument(
@@ -945,9 +956,13 @@ void GpuPsGraphTable::build_graph_from_cpu(
       gpu_graph_list_[offset].node_size = 0;
     }
     if (cpu_graph_list[i].neighbor_size) {
-      CUDA_CHECK(
-          cudaMalloc(&gpu_graph_list_[offset].neighbor_list,
-                     cpu_graph_list[i].neighbor_size * sizeof(uint64_t)));
+      if (!FLAGS_enable_neighbor_list_use_uva) {
+        CUDA_CHECK(cudaMalloc(&gpu_graph_list_[offset].neighbor_list,
+                    cpu_graph_list[i].neighbor_size * sizeof(uint64_t)));
+      } else {
+        CUDA_CHECK(cudaMallocManaged(&gpu_graph_list_[offset].neighbor_list,
+                    cpu_graph_list[i].neighbor_size * sizeof(uint64_t)));
+      }
 
       CUDA_CHECK(cudaMemcpy(gpu_graph_list_[offset].neighbor_list,
                             cpu_graph_list[i].neighbor_list,

diff --git a/paddle/phi/core/flags.cc b/paddle/phi/core/flags.cc
@@ -852,6 +852,18 @@ PHI_DEFINE_EXPORTED_bool(graph_load_in_parallel,
                          false,
                          "It controls whether load graph node and edge with "
                          "mutli threads parallely.");
+
+/**
+ * Distributed related FLAG
+ * Name: FLAGS_enable_neighbor_list_use_uva
+ * Since Version: 2.2.0
+ * Value Range: bool, default=false
+ * Example:
+ * Note: Control whether store neighbor_list with UVA
+ */
+PHI_DEFINE_EXPORTED_bool(enable_neighbor_list_use_uva,
+                            false,
+                            "It controls whether store neighbor_list with UVA");
 
 /**
  * Distributed related FLAG