fix the bug of yolo_box which can't run on nano and tx2 (#33422) (#33442)

paddle/fluid/operators/detection/yolo_box_op.cu

@@ -111,7 +111,14 @@ class YoloBoxOpCUDAKernel : public framework::OpKernel<T> {
     platform::GpuLaunchConfig config =
         platform::GetGpuLaunchConfig1D(ctx.cuda_device_context(), n * box_num);
 
-    KeYoloBoxFw<T><<<config.block_per_grid, config.thread_per_block, 0,
+    dim3 thread_num = config.thread_per_block;
+#ifdef WITH_NV_JETSON
+    if (config.compute_capability == 53 || config.compute_capability == 62) {
+      thread_num = 512;
+    }
+#endif
+
+    KeYoloBoxFw<T><<<config.block_per_grid, thread_num, 0,
                      ctx.cuda_device_context().stream()>>>(
         input_data, imgsize_data, boxes_data, scores_data, conf_thresh,
         anchors_data, n, h, w, an_num, class_num, box_num, input_size_h,

paddle/fluid/platform/gpu_launch_config.h

@@ -37,6 +37,7 @@ struct GpuLaunchConfig {
   dim3 theory_thread_count = dim3(1, 1, 1);
   dim3 thread_per_block = dim3(1, 1, 1);
   dim3 block_per_grid = dim3(1, 1, 1);
+  int compute_capability = 0;
 };
 
 inline GpuLaunchConfig GetGpuLaunchConfig1D(
@@ -67,11 +68,14 @@ inline GpuLaunchConfig GetGpuLaunchConfig1D(
       std::min(max_threads, context.GetMaxThreadsPerBlock());
   const int block_count =
       std::min(DivUp(physical_thread_count, thread_per_block), sm);
+  // Get compute_capability
+  const int capability = context.GetComputeCapability();
 
   GpuLaunchConfig config;
   config.theory_thread_count.x = theory_thread_count;
   config.thread_per_block.x = thread_per_block;
   config.block_per_grid.x = block_count;
+  config.compute_capability = capability;
   return config;
 }