Fix a critical bug in softmax_with_cross_entropy_op backward. (#9120)

* Fix a critical bug in softmax_with_cross_entropy_op, which will lead to the wrong gradients.

* Enhance unit testing.

paddle/fluid/operators/softmax_with_cross_entropy_op.cu

@@ -23,21 +23,21 @@ using Tensor = framework::Tensor;
 
 namespace {
 template <typename T>
-__global__ void CrossEntropyGrad(T* logit_grad, const T* loss_grad,
-                                 const int64_t* labels, const int batch_size,
-                                 const int class_num) {
-  int tid = blockIdx.x * blockDim.x + threadIdx.x;
-  int sample_idx = tid / class_num;
-
-  if (tid < batch_size) {
-    PADDLE_ASSERT(labels[sample_idx] >= 0 && labels[sample_idx] < class_num);
-    logit_grad[tid * class_num + labels[tid]] -= static_cast<T>(1.);
+__global__ void CrossEntropyGrad(T* logit_grad, const int64_t* labels,
+                                 const int batch_size, const int class_num) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < batch_size;
+       i += blockDim.x * gridDim.x) {
+    int idx = i * class_num + labels[i];
+    logit_grad[idx] -= static_cast<T>(1.);
   }
+}
 
-  __syncthreads();
-
-  if (tid < batch_size * class_num) {
-    logit_grad[tid] *= loss_grad[sample_idx];
+template <typename T>
+__global__ void Scale(T* logit_grad, const T* loss_grad, const int num,
+                      const int class_num) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
+       i += blockDim.x * gridDim.x) {
+    logit_grad[i] *= loss_grad[i / class_num];
   }
 }
 
@@ -94,22 +94,22 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
     const int batch_size = logit_grad->dims()[0];
     const int class_num = logit_grad->dims()[1];
     int block = 512;
-    int grid = (batch_size * class_num + block - 1) / block;
+    auto stream = context.cuda_device_context().stream();
 
     if (context.Attr<bool>("soft_label")) {
+      int grid = (batch_size * class_num + block - 1) / block;
       const T* label_data = labels->data<T>();
-      SoftCrossEntropyGradientKernel<
-          T><<<grid, block, 0,
-               context.template device_context<platform::CUDADeviceContext>()
-                   .stream()>>>(logit_grad_data, loss_grad_data, label_data,
-                                batch_size, class_num);
+      SoftCrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
+          logit_grad_data, loss_grad_data, label_data, batch_size, class_num);
     } else {
+      int grid = (batch_size + block - 1) / block;
       const int64_t* label_data = labels->data<int64_t>();
-      CrossEntropyGrad<
-          T><<<grid, block, 0,
-               context.template device_context<platform::CUDADeviceContext>()
-                   .stream()>>>(logit_grad_data, loss_grad_data, label_data,
-                                batch_size, class_num);
+      CrossEntropyGrad<T><<<grid, block, 0, stream>>>(
+          logit_grad_data, label_data, batch_size, class_num);
+      int num = batch_size * class_num;
+      grid = (num + block - 1) / block;
+      Scale<T><<<grid, block, 0, stream>>>(logit_grad_data, loss_grad_data, num,
+                                           class_num);
     }
   }
 };

python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py

@@ -26,7 +26,7 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
 
     def setUp(self):
         self.op_type = "softmax_with_cross_entropy"
-        batch_size = 2
+        batch_size = 41
         class_num = 37
 
         logits = np.random.uniform(0.1, 1.0,
@@ -59,7 +59,7 @@ class TestSoftmaxWithCrossEntropyOp2(OpTest):
 
     def setUp(self):
         self.op_type = "softmax_with_cross_entropy"
-        batch_size = 2
+        batch_size = 41
         class_num = 37
 
         logits = np.random.uniform(0.1, 1.0,