Fix ignore index of c_softmax_with_cross_entropy_op. (#52835)

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* Remove climits.

* Fix bug of c_softmax_with_cross_entropy_op. Support ignore_index is
negative number.

paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu

@@ -40,6 +40,7 @@ template <typename T, typename IndexT>
 __global__ void MaskLabelByIndex(T* predicted_logits,
                                  const T* logit,
                                  const IndexT* label,
+                                 const IndexT ignore_index,
                                  const int start_index,
                                  const int end_index,
                                  const int64_t N,
@@ -47,13 +48,15 @@ __global__ void MaskLabelByIndex(T* predicted_logits,
                                  const int nranks) {
   CUDA_KERNEL_LOOP(i, N) {
     auto real_label = label[i];
-    PADDLE_ENFORCE((real_label < D * nranks) && (real_label >= 0),
+    PADDLE_ENFORCE(((real_label < D * nranks) && (real_label >= 0)) ||
+                       (real_label == ignore_index),
                    "The index is out of bounds, "
                    "please check whether the value of label and "
                    "input meet the class number. It should "
-                   "be less than [%d], but received [%d]",
-                   D * nranks,
-                   real_label);
+                   "be less than [%ld] or equal to [%ld], but received [%ld]",
+                   static_cast<int64_t>(D * nranks),
+                   static_cast<int64_t>(ignore_index),
+                   static_cast<int64_t>(real_label));
 
     if (real_label >= start_index && real_label < end_index) {
       predicted_logits[i] = logit[i * D + real_label - start_index];
@@ -204,10 +207,11 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
     const auto& label_type = framework::TransToProtoVarType(labels->dtype());
 
     if (label_type == framework::proto::VarType::INT32) {
-      MaskLabelByIndex<T, int32_t>
-          <<<blocks, threads, 0, dev_ctx.stream()>>>(predicted_logits.data<T>(),
+      MaskLabelByIndex<T, int32_t><<<blocks, threads, 0, dev_ctx.stream()>>>(
+          predicted_logits.data<T>(),
           softmax_2d.data<T>(),
           labels->data<int32_t>(),
+          static_cast<int32_t>(ignore_index),
           start_index,
           end_index,
           N,
@@ -218,6 +222,7 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
           <<<blocks, threads, 0, dev_ctx.stream()>>>(predicted_logits.data<T>(),
                                                      softmax_2d.data<T>(),
                                                      labels->data<int64_t>(),
+                                                     ignore_index,
                                                      start_index,
                                                      end_index,
                                                      N,
@@ -362,20 +367,22 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
     const auto& label_type = framework::TransToProtoVarType(labels->dtype());
 
     if (label_type == framework::proto::VarType::INT32) {
-      MaskLabelByIndex<T, int32_t>
-          <<<blocks, threads, 0, dev_ctx.stream()>>>(predicted_logits.data<T>(),
+      MaskLabelByIndex<T, int32_t><<<blocks, threads, 0, dev_ctx.stream()>>>(
+          predicted_logits.data<T>(),
           softmax_2d.data<T>(),
           labels->data<int32_t>(),
+          static_cast<int32_t>(ignore_index),
           start_index,
           end_index,
           N,
           D,
           nranks);
     } else if (label_type == framework::proto::VarType::INT64) {
-      MaskLabelByIndex<T, int64_t>
-          <<<blocks, threads, 0, dev_ctx.stream()>>>(predicted_logits.data<T>(),
+      MaskLabelByIndex<T, int64_t><<<blocks, threads, 0, dev_ctx.stream()>>>(
+          predicted_logits.data<T>(),
           softmax_2d.data<T>(),
           labels->data<int64_t>(),
+          static_cast<int32_t>(ignore_index),
           start_index,
           end_index,
           N,