From 081e4307b47d2c9e119b470d274886188a14391a Mon Sep 17 00:00:00 2001
From: Zhang Zheng <32410583+ZzSean@users.noreply.github.com>
Date: Fri, 18 Mar 2022 09:40:54 +0800
Subject: [PATCH] Optimize perf of softmax_with_cross_entropy_bwd (#40643)

* Optimize perf of softmax_with_cross_entropy_bwd

* fix

* fix
---
 .../softmax_with_cross_entropy_op.cu          | 31 ++++++++++++-------
 1 file changed, 19 insertions(+), 12 deletions(-)

diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu
index 19a395e723..41545a1ca2 100644
--- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu
+++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu
@@ -760,8 +760,9 @@ static void SoftmaxWithCrossEntropyHardLabel(
 */
 template <typename T, typename LabelT>
 __global__ void SoftmaxWithCrossEntropyGradHardLabel(
-    T* logits_grad, const T* loss_grad, const LabelT* labels, const int64_t n,
-    const int64_t dim, const int64_t d, const int ignore_index) {
+    T* logits_grad, const T* loss_grad, const T* softmax, const LabelT* labels,
+    const int64_t n, const int64_t dim, const int64_t d,
+    const int ignore_index) {
   int64_t idx = blockIdx.x * blockDim.x + threadIdx.x;
   int64_t idx_n = idx / (d * dim);
   int64_t idx_dim = (idx / d) % dim;
@@ -773,10 +774,9 @@ __global__ void SoftmaxWithCrossEntropyGradHardLabel(
     if (lbl == ignore_index) {
       logits_grad[idx] = static_cast<T>(0.0);
     } else if (lbl == idx_dim) {
-      logits_grad[idx] =
-          (logits_grad[idx] - static_cast<T>(1.0)) * loss_grad[ids];
+      logits_grad[idx] = (softmax[idx] - static_cast<T>(1.0)) * loss_grad[ids];
     } else {
-      logits_grad[idx] *= loss_grad[ids];
+      logits_grad[idx] = softmax[idx] * loss_grad[ids];
     }
   }
 }
@@ -1395,11 +1395,20 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
     Tensor* logit_grad =
         context.Output<Tensor>(framework::GradVarName("Logits"));
     const Tensor* softmax = context.Input<Tensor>("Softmax");
-    if (logit_grad != softmax) {
+    auto stream = context.cuda_device_context().stream();
+    auto ignore_index = context.Attr<int>("ignore_index");
+    auto use_softmax = context.Attr<bool>("use_softmax");
+
+    T* logit_grad_data = nullptr;
+    bool copy_flag = (logit_grad != softmax && (!use_softmax || soft_label));
+    if (copy_flag) {
       framework::TensorCopy(*softmax, context.GetPlace(),
                             context.device_context(), logit_grad);
+      logit_grad_data = logit_grad->template data<T>();
+    } else {
+      logit_grad_data =
+          logit_grad->template mutable_data<T>(context.GetPlace());
     }
-    T* logit_grad_data = logit_grad->template data<T>();
 
     const int rank = logit_grad->dims().size();
     const int axis = phi::funcs::CanonicalAxis(context.Attr<int>("axis"), rank);
@@ -1414,9 +1423,6 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
 #else
     int block = 512;
 #endif
-    auto stream = context.cuda_device_context().stream();
-    auto ignore_index = context.Attr<int>("ignore_index");
-    auto use_softmax = context.Attr<bool>("use_softmax");
 
     // do not with softmax op, and input is softmax
     if (!use_softmax) {
@@ -1451,11 +1457,12 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
       SoftCrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
           logit_grad_data, loss_grad_data, label_data, n, d, remain);
     } else {
+      const T* softmax_data = softmax->template data<T>();
       const auto* label_data = labels.template data<LabelT>();
       int grid = (n * d + block - 1) / block;
       SoftmaxWithCrossEntropyGradHardLabel<T><<<grid, block, 0, stream>>>(
-          logit_grad_data, loss_grad_data, label_data, n, d / remain, remain,
-          ignore_index);
+          logit_grad_data, loss_grad_data, softmax_data, label_data, n,
+          d / remain, remain, ignore_index);
     }
   }
 };
-- 
GitLab