clean up intel labeled TODOs (#19476)

test=develop

paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc

@@ -159,8 +159,8 @@ class ConcatPrimitiveFactory {
   std::vector<memory::primitive_desc> srcs_pd;
   std::vector<memory> srcs;
   std::vector<primitive::at> inputs;
-  boost::optional<memory> dst_mem;  // TODO(mgallus): change to std::optional
-};                                  // upon introduction of C++17 to paddle
+  boost::optional<memory> dst_mem;
+};
 
 template <typename T>
 class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {

paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc

@@ -157,7 +157,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     int groups = ctx.Attr<int>("groups");
     bool is_conv3d = strides.size() == 3U;
 
-    // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
         is_conv3d
             ? dilations.size() == 3 && dilations[0] == 1 && dilations[1] == 1 &&
@@ -212,8 +211,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
         weights_tz, platform::MKLDNNGetDataType<T>(), weights_format);
-    std::vector<int> bias_tz;  // TODO(mgallus): avoid empty vector creation.
-                               // Currently used whenever bias is != nullptr.
+    std::vector<int> bias_tz;
     auto dst_md = platform::MKLDNNMemDesc(
         dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
@@ -364,7 +362,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                    "residual fusion does not support force output with fp32");
 
     bool is_conv3d = strides.size() == 3U;
-    // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
         is_conv3d
             ? dilations.size() == 3 && dilations[0] == 1 && dilations[1] == 1 &&

paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc

@@ -69,7 +69,6 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
     int groups = ctx.Attr<int>("groups");
 
-    // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
         dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
         "dilation in convolution is not implemented yet");
@@ -150,8 +149,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
         weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
-    std::vector<int> bias_tz;  // TODO(mgallus): avoid empty vector creation.
-                               // Currently used whenever bias is != nullptr.
+    std::vector<int> bias_tz;
     auto dst_md = platform::MKLDNNMemDesc(
         dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 

paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc

@@ -40,8 +40,6 @@ class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
       data[i] = dist(engine);
     }
 
-    // The format of output is set as the mkldnn's format
-    // TODO(@mozga-intel) The format of matrix sets inside the another layers.
     tensor->set_layout(DataLayout::kMKLDNN);
     tensor->set_format(mkldnn::memory::format::oihw);
   }

paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc

@@ -124,7 +124,5 @@ class QuantOpKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 namespace ops = paddle::operators;
 
-// TODO(Xiaoli) Support FP32->S8 quantization.
-
 REGISTER_OP_KERNEL(quantize, MKLDNN, ::paddle::platform::CPUPlace,
                    ops::QuantOpKernel<float>);

paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc

@@ -47,9 +47,7 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
     mkldnn::memory::data_type src_dt =
         paddle::framework::ToMKLDNNDataType(input->type());
-    mkldnn::memory::data_type dst_dt = src_dt;  // TODO(Xiaoli) support
-                                                // requantize from different
-                                                // data type (e.g., s8 to u8)
+    mkldnn::memory::data_type dst_dt = src_dt;
     mkldnn::memory::format src_fmt = memory::format::nhwc;
     mkldnn::memory::format dst_fmt = memory::format::nhwc;
 

paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc

@@ -139,7 +139,6 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       output->set_layout(DataLayout::kMKLDNN);
       output->set_format(output_format);
     } else {  // Fallback to naive version
-      // TODO(@mozga-intel) Add MKLDNN SelectedRows & LoDTensorArray support
       SumKernel<CPUDeviceContext, T> reference_kernel;
       reference_kernel.Compute(ctx);
     }

paddle/fluid/operators/ngraph/ops/reshape_op.h

@@ -57,8 +57,7 @@ static void BuildReshapeNode(
   std::shared_ptr<ngraph::Node> input =
       platform::GetInputNode(op, "X", ngb_node_map);
   auto input_shape = input->get_shape();
-  // TODO(mozga-intel) The vector of shape is not supported yet, that's
-  // asDispensable() operator"
+
   std::shared_ptr<ngraph::Node> shape =
       platform::GetInputNode(op, "Shape", ngb_node_map);
 

python/paddle/fluid/contrib/slim/quantization/mkldnn_post_training_strategy.py

@@ -83,8 +83,6 @@ class MKLDNNPostTrainingQuantStrategy(Strategy):
         if six.PY3:
             data = warmup_reader.__next__()
 
-        # TODO (Intel) Remove limits that MKLDNNPostTrainingQuantStrategy
-        # only support image classification
         num_images = len(data)
         image_data = [img.tolist() for (img, _) in data]
         image_data = np.array(image_data).astype("float32").reshape(

python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py

@@ -152,73 +152,6 @@ class TestElementwiseMulMKLDNNOp_Integrated_With_Convs(ElementwiseMulOp):
         pass
 
 
-# TODO(LeoZhao-Intel): re-enable this case
-# https://github.com/PaddlePaddle/Paddle/issues/16764
-@unittest.skip("Not supported well on avx2.")
-class TestElementwiseMulMKLDNNOp_BroadcastNCHW16c(ElementwiseMulOp):
-    def init_input_output(self):
-        x = np.random.rand(1, 16, 2, 2).astype(self.dtype)
-        self.x = x.transpose(0, 2, 3, 1).reshape(1, 16, 2, 2)
-        self.y = np.random.rand(1, 16).astype(self.dtype)
-
-        self.out = x * self.y.reshape(1, 16, 1, 1)
-        self.out = self.out.transpose(0, 2, 3, 1).reshape(1, 16, 2, 2)
-
-    def setUp(self):
-        super(TestElementwiseMulMKLDNNOp_BroadcastNCHW16c, self).setUp()
-        self.attrs["x_data_format"] = "nchw16c"
-        self.attrs["y_data_format"] = "nc"
-        self._cpu_only = True
-
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-    def init_axis(self):
-        self.axis = 0
-
-    def test_check_grad_normal(self):
-        pass
-
-    def test_check_grad_ingore_x(self):
-        pass
-
-    def test_check_grad_ingore_y(self):
-        pass
-
-
-@unittest.skip(
-    "Not implemented yet.")  # TODO(mgallus): enable when implemented.
-class TestElementwiseMulMKLDNNOp_BroadcastNCHW8c(ElementwiseMulOp):
-    def init_input_output(self):
-        x = np.random.rand(1, 8, 2, 2).astype(self.dtype)
-        self.x = x.transpose(0, 2, 3, 1).reshape(1, 8, 2, 2)
-        self.y = np.random.rand(1, 8).astype(self.dtype)
-
-        self.out = x * self.y.reshape(1, 8, 1, 1)
-        self.out = self.out.transpose(0, 2, 3, 1).reshape(1, 8, 2, 2)
-
-    def setUp(self):
-        super(TestElementwiseMulMKLDNNOp_BroadcastNCHW8c, self).setUp()
-        self.attrs["x_data_format"] = "nchw8c"
-        self.attrs["y_data_format"] = "nc"
-        self._cpu_only = True
-
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-    def init_axis(self):
-        self.axis = 0
-
-    def test_check_grad_normal(self):
-        pass
-
-    def test_check_grad_ingore_x(self):
-        pass
-
-    def test_check_grad_ingore_y(self):
-        pass
-
-
 class TestElementwiseMulMKLDNNOp_FallbackNCHW(ElementwiseMulOp):
     def init_input_output(self):
         self.x = np.random.rand(1, 16, 2, 2).astype(self.dtype)