[Relay][AlterOpLayout] NHWC to NCHWc pad operator. (#4103)

* [Relay][AlterOpLayout] NHWC to NCHWc pad operator.

* Fixing culprit.

* Flaky test 1.

* Flaky test 2.

src/relay/op/nn/pad.cc

@@ -36,6 +36,82 @@ namespace relay {
 // relay.nn.pad
 TVM_REGISTER_NODE_TYPE(PadAttrs);
 
+Array<Array<Layout> > PadInferCorrectLayout(
+    const Attrs& attrs,
+    const Array<Layout>& new_in_layouts,
+    const Array<Layout>& old_in_layouts,
+    const Array<Array<IndexExpr>> &old_in_shapes) {
+  // NOTE: Discard "const" qualifier here.
+  PadAttrs *params = const_cast<PadAttrs*>(attrs.as<PadAttrs>());
+
+  Layout ret;
+  // If new_in_layouts are defined, this code tries to modify the layout.
+  bool is_layout_modified = new_in_layouts.defined();
+  if (new_in_layouts.defined()) {
+    // Create a map of axis to param_width. For the new layout, a new param_width is generated using
+    // the map. The new layout is rejected, if the padding is happening along the axis which was
+    // split.
+
+    // 1) Create a map from axis to param_width using old layout.
+    std::map<std::string, tvm::Array<tvm::Expr>> axis_pad_width;
+    int index_counter = 0;
+    CHECK_EQ(new_in_layouts.size(), 1);
+    CHECK_EQ(old_in_layouts.size(), 1);
+    for (auto iter_var : old_in_layouts[0]->axes) {
+      const auto& old_layout_axis = LayoutAxis::Get(iter_var);
+      axis_pad_width.emplace(old_layout_axis.name(), params->pad_width[index_counter]);
+      index_counter++;
+    }
+
+    // 2) Create new pad width by walking over the new layout and using the map.
+    tvm::Array<tvm::Array<tvm::Expr>> new_pad_width;
+    for (auto iter_var : new_in_layouts[0]->axes) {
+      const auto& new_layout_axis = LayoutAxis::Get(iter_var);
+      auto axis_name = new_layout_axis.name();
+      if (axis_pad_width.count(axis_name) != 0 && new_layout_axis.IsPrimal()) {
+        // This is primal axis. So, directly use the original pad_width.
+        new_pad_width.push_back(axis_pad_width.at(axis_name));
+      } else {
+        // This is the axis that got split. So, check that pad_width was [0, 0] originally.
+        const auto& dual_axis = new_layout_axis.ToPrimal();
+        auto dual_axis_name = dual_axis.name();
+        CHECK(axis_pad_width.count(dual_axis_name))
+            << "Missing axis " << dual_axis << " in " << old_in_layouts[0].name();
+        new_pad_width.push_back(axis_pad_width.at(dual_axis_name));
+
+        // If any pad_width element is not zero, do not change the layout.
+        for (auto width : axis_pad_width.at(dual_axis_name)) {
+          if (auto* width_imm = width.as<IntImm>()) {
+            if (width_imm->value != 0) {
+              is_layout_modified = false;
+            }
+          } else {
+            is_layout_modified = false;
+          }
+        }
+      }
+    }
+
+    // If the above conditions satisfied, we can set the newly created pad_width and use the new
+    // layout.
+    if (is_layout_modified) {
+      ret = new_in_layouts[0];
+      params->pad_width = new_pad_width;
+    }
+  }
+
+  if (!is_layout_modified) {
+    if (old_in_layouts.defined()) {
+      CHECK_EQ(old_in_layouts.size(), 1);
+      ret = old_in_layouts[0];
+    } else {
+      ret = Layout::Undef();
+    }
+  }
+
+  return Array<Array<Layout> >{{ret}, {ret}};
+}
+
 bool PadRel(const Array<Type>& types,
             int num_inputs,
             const Attrs& attrs,
@@ -133,6 +209,7 @@ RELAY_REGISTER_OP("nn.pad")
 .add_argument("data", "Tensor", "The input tensor.")
 .set_support_level(2)
 .add_type_rel("Pad", PadRel)
+.set_attr<FInferCorrectLayout>("FInferCorrectLayout", PadInferCorrectLayout)
 .set_attr<TOpPattern>("TOpPattern", kInjective)
 .set_attr<FTVMCompute>("FTVMCompute", PadCompute);
 

tests/python/frontend/coreml/test_forward.py

@@ -47,8 +47,10 @@ def run_model_checkonly(model_file, model_name='', input_name='image'):
     model = cm.models.MLModel(model_file)
     x = model_zoo.get_cat_image()
     shape_dict = {input_name : x.shape}
-    mod, params = relay.frontend.from_coreml(model, shape_dict)
+    # Some Relay passes change operators on the fly. Ensuring that we generate
+    # new graph for each target.
     for target, ctx in ctx_list():
+        mod, params = relay.frontend.from_coreml(model, shape_dict)
         tvm_output = get_tvm_output(mod["main"], x, params, target, ctx)
         print(target, ctx, model_name, 'prediction id: ', np.argmax(tvm_output.flat))
 

tests/python/relay/test_pass_alter_op_layout.py

@@ -641,6 +641,120 @@ def test_alter_layout_prelu():
     assert(analysis.alpha_equal(a, b))
 
 
+def test_alter_layout_pad():
+    """ Check NCHW, NHWC and corner case for pad layout conversion"""
+    # Register alter op layout. "level" is used to override the previously registered functions.
+    @register_alter_op_layout("nn.conv2d", level=112)
+    def alter_conv2d(attrs, inputs, tinfos):
+        data, weight = inputs
+        new_attrs = dict(attrs)
+        new_attrs['data_layout'] = 'NCHW16c'
+        return relay.nn.conv2d(data, weight, **new_attrs)
+
+    # Check NCHW conversion.
+    def before_nchw():
+        x = relay.var("x", shape=(1, 64, 56, 56))
+        weight1 = relay.var('weight1')
+        y = relay.nn.conv2d(x, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1))
+        ret = relay.nn.pad(y, pad_width=((0, 0), (0, 0), (1, 1), (1, 1)))
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    def expected_nchw():
+        x = relay.var("x", shape=(1, 64, 56, 56))
+        weight1 = relay.var('weight1')
+        y = relay.layout_transform(x, "NCHW", "NCHW16c")
+        y = relay.nn.conv2d(y, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1),
+                            data_layout="NCHW16c")
+        ret = relay.nn.pad(y, pad_width=((0, 0), (0, 0), (1, 1), (1, 1), (0, 0)))
+        ret = relay.layout_transform(ret, "NCHW16c", "NCHW")
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    a = before_nchw()
+    a = run_opt_pass(a, transform.AlterOpLayout())
+
+    b = expected_nchw()
+    b = run_opt_pass(b, transform.InferType())
+
+    assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
+
+    # Check NHWC conversion.
+    def before_nhwc():
+        x = relay.var("x", shape=(1, 56, 56, 64))
+        weight1 = relay.var('weight1')
+        y = relay.nn.conv2d(x, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1),
+                            data_layout='NHWC')
+        ret = relay.nn.pad(y, pad_width=((0, 0), (1, 1), (1, 1), (0, 0)))
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    def expected_nhwc():
+        x = relay.var("x", shape=(1, 56, 56, 64))
+        weight1 = relay.var('weight1')
+        y = relay.layout_transform(x, "NHWC", "NCHW16c")
+        y = relay.nn.conv2d(y, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1),
+                            data_layout="NCHW16c")
+        ret = relay.nn.pad(y, pad_width=((0, 0), (0, 0), (1, 1), (1, 1), (0, 0)))
+        ret = relay.layout_transform(ret, "NCHW16c", "NHWC")
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    a = before_nhwc()
+    a = run_opt_pass(a, transform.AlterOpLayout())
+
+    b = expected_nhwc()
+    b = run_opt_pass(b, transform.InferType())
+
+    assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
+
+    # Check that conversion does not happen when padding along split axis..
+    def before():
+        x = relay.var("x", shape=(1, 64, 56, 56))
+        weight1 = relay.var('weight1')
+        y = relay.nn.conv2d(x, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1))
+        ret = relay.nn.pad(y, pad_width=((0, 0), (1, 1), (1, 1), (1, 1)))
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    def expected():
+        x = relay.var("x", shape=(1, 64, 56, 56))
+        weight1 = relay.var('weight1')
+        y = relay.layout_transform(x, "NCHW", "NCHW16c")
+        y = relay.nn.conv2d(y, weight1,
+                            channels=32,
+                            kernel_size=(3, 3),
+                            padding=(1, 1),
+                            data_layout="NCHW16c")
+        ret = relay.layout_transform(y, "NCHW16c", "NCHW")
+        ret = relay.nn.pad(ret, pad_width=((0, 0), (1, 1), (1, 1), (1, 1)))
+        y = relay.Function(analysis.free_vars(ret), ret)
+        return y
+
+    a = before()
+    a = run_opt_pass(a, transform.AlterOpLayout())
+
+    b = expected()
+    b = run_opt_pass(b, transform.InferType())
+
+    assert analysis.alpha_equal(a, b), "Actual = \n" + str(a)
+
+
 def test_alter_layout_pool():
     """ Check NCHW, NHWC pool layout conversion"""
     # Register alter op layout. "level" is used to override the previously registered functions.
@@ -815,5 +929,6 @@ if __name__ == "__main__":
     test_alter_layout_strided_slice()
     test_alter_layout_depthwise_conv2d()
     test_alter_layout_prelu()
+    test_alter_layout_pad()
     test_alter_layout_pool()
     test_alter_layout_sum()