[OP] fix histogram op when input tensor is empty, test=develop (#33970)

paddle/fluid/operators/histogram_op.cu

@@ -81,6 +81,13 @@ class HistogramCUDAKernel : public framework::OpKernel<T> {
     const T* input_data = input->data<T>();
     const int input_numel = input->numel();
 
+    int64_t* out_data = output->mutable_data<int64_t>(context.GetPlace());
+    math::SetConstant<platform::CUDADeviceContext, int64_t>()(
+        context.template device_context<platform::CUDADeviceContext>(), output,
+        static_cast<int64_t>(0));
+
+    if (input_data == nullptr) return;
+
     T output_min = static_cast<T>(minval);
     T output_max = static_cast<T>(maxval);
 
@@ -126,11 +133,6 @@ class HistogramCUDAKernel : public framework::OpKernel<T> {
             "But received max is %d, min is %d",
             maxval, minval));
 
-    int64_t* out_data = output->mutable_data<int64_t>(context.GetPlace());
-    math::SetConstant<platform::CUDADeviceContext, int64_t>()(
-        context.template device_context<platform::CUDADeviceContext>(), output,
-        static_cast<int64_t>(0));
-
     auto stream =
         context.template device_context<platform::CUDADeviceContext>().stream();
     KernelHistogram<

paddle/fluid/operators/histogram_op.h

@@ -38,6 +38,13 @@ class HistogramKernel : public framework::OpKernel<T> {
     const T* input_data = input->data<T>();
     auto input_numel = input->numel();
 
+    int64_t* out_data = output->mutable_data<int64_t>(context.GetPlace());
+    math::SetConstant<DeviceContext, int64_t>()(
+        context.template device_context<DeviceContext>(), output,
+        static_cast<int64_t>(0));
+
+    if (input_data == nullptr) return;
+
     T output_min = static_cast<T>(minval);
     T output_max = static_cast<T>(maxval);
     if (output_min == output_max) {
@@ -63,11 +70,6 @@ class HistogramKernel : public framework::OpKernel<T> {
             "But received max is %d, min is %d",
             maxval, minval));
 
-    int64_t* out_data = output->mutable_data<int64_t>(context.GetPlace());
-    math::SetConstant<DeviceContext, int64_t>()(
-        context.template device_context<DeviceContext>(), output,
-        static_cast<int64_t>(0));
-
     for (int64_t i = 0; i < input_numel; i++) {
       if (input_data[i] >= output_min && input_data[i] <= output_max) {
         const int64_t bin = (int64_t)((input_data[i] - output_min) * nbins /