[MLU] fix bn_grad and hard_sigmoid_grad error (#44919)

paddle/fluid/operators/activation_op_mlu.cc

@@ -370,7 +370,7 @@ class HardSigmoidGradMLUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
-    auto* out = ctx.Input<Tensor>("Out");
+    auto* x = ctx.Input<Tensor>("X");
     auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     float slope = ctx.Attr<float>("slope");
     float offset = ctx.Attr<float>("offset");
@@ -381,7 +381,7 @@ class HardSigmoidGradMLUKernel : public framework::OpKernel<T> {
                                    1.0f /*sliced_dim useless*/,
                                    slope,
                                    offset);
-    MLUCnnlTensorDesc out_desc(*out);
+    MLUCnnlTensorDesc x_desc(*x);
     MLUCnnlTensorDesc dout_desc(*dout);
     MLUCnnlTensorDesc dx_desc(*dx);
     MLUCnnl::ActiveGrad(ctx,
@@ -392,8 +392,8 @@ class HardSigmoidGradMLUKernel : public framework::OpKernel<T> {
                         nullptr,
                         dout_desc.get(),
                         GetBasePtr(dout),
-                        out_desc.get(),
-                        GetBasePtr(out),
+                        x_desc.get(),
+                        GetBasePtr(x),
                         dx_desc.get(),
                         GetBasePtr(dx));
   }

paddle/fluid/operators/batch_norm_op_mlu.cc

@@ -273,7 +273,7 @@ class MLUBatchNormGradOpKernel : public framework::OpKernel<T> {
       const auto *running_mean = ctx.Input<Tensor>("Mean");
       const auto *running_variance = ctx.Input<Tensor>("Variance");
       MLUCnnl::FusedBatchNormGrad(ctx,
-                                  true /*is_training*/,
+                                  false /*is_training*/,
                                   transformed_desc.get(),
                                   GetBasePtr(&transformed_d_y),
                                   transformed_desc.get(),

paddle/fluid/operators/conv_transpose_op_mlu.cc

@@ -271,26 +271,18 @@ class Conv2DTransposeGradMLUKernel : public framework::OpKernel<T> {
           data_layout_mlu,
           ToCnnlDataType(input_grad_tensor.dtype()));
 
-      cnnlDataType_t tensor_dtype = ToCnnlDataType<T>();
-      cnnlDataType_t dt_onchip = ToCnnlDataType<T>();
-      MLUCnnl::Conv2D(ctx,
-                      conv_desc.get(),
-                      tensor_dtype,
-                      dt_onchip,
-                      nullptr /* input_position */,
-                      nullptr /* input_scale */,
-                      nullptr /* input_offset */,
-                      nullptr /* filter_position */,
-                      nullptr /* filter_scale */,
-                      nullptr /* filter_offset */,
-                      output_grad_desc.get(),
-                      GetBasePtr(&output_grad_tensor),
-                      trans_filter_desc.get(),
-                      GetBasePtr(&trans_filter),
-                      nullptr /* bias_desc*/,
-                      nullptr /* bias */,
-                      input_grad_desc.get(),
-                      GetBasePtr(&input_grad_tensor));
+      MLUCnnl::ConvolutionForward(ctx,
+                                  conv_desc.get(),
+                                  nullptr /*alpha*/,
+                                  nullptr /*beta*/,
+                                  nullptr /*bias_desc*/,
+                                  nullptr /*bias_ptr*/,
+                                  output_grad_desc.get(),
+                                  GetBasePtr(&output_grad_tensor),
+                                  trans_filter_desc.get(),
+                                  GetBasePtr(&trans_filter),
+                                  input_grad_desc.get(),
+                                  GetBasePtr(&input_grad_tensor));
       if (!channel_last) {
         // transpose output from NHWC to NCHW
         const std::vector<int> perm_to_nchw = {0, 3, 1, 2};

paddle/phi/kernels/funcs/activation_functor.h

@@ -1604,7 +1604,11 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
   }
 
   static constexpr ActBwdOpFwdDeps FwdDeps() {
+#ifdef PADDLE_WITH_MLU
+    return ActBwdOpFwdDeps::kDepX;
+#else
     return ActBwdOpFwdDeps::kDepOut;
+#endif
   }
 };