fix tf fusion regressions (#755)

* fix relu6

* add more transposes

* add tests

* formatting
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

src/tf/parse_binary_op.cpp

@@ -9,6 +9,7 @@ namespace tf {
 
 struct parse_binary_op : op_parser<parse_binary_op>
 {
+    bool transpose() const { return true; }
     std::vector<op_desc> operators() const
     {
         return {{"Add", "add"},

src/tf/parse_generic_op.cpp

@@ -9,6 +9,7 @@ namespace tf {
 
 struct parse_generic_op : op_parser<parse_generic_op>
 {
+    bool transpose() const { return true; }
     std::vector<op_desc> operators() const
     {
         return {{"All", "identity"},

src/tf/parse_relu6.cpp

@@ -10,6 +10,7 @@ namespace tf {
 
 struct parse_relu6 : op_parser<parse_relu6>
 {
+    bool transpose() const { return true; }
     std::vector<op_desc> operators() const { return {{"Relu6"}}; }
 
     instruction_ref parse(const op_desc& /*opd*/,

test/tf/gen_tf_pb.py

@@ -196,6 +196,23 @@ def conv_test(g1):
         tf.nn.conv2d(g1_input, g1_weights, [1, 1, 1, 1], "SAME", name='conv1')
 
 
+@tf_test
+def conv_add_test(g1):
+    with g1.as_default():
+        g1_input = tf.compat.v1.placeholder(tf.float32,
+                                            shape=(1, 16, 16, 3),
+                                            name='0')
+        g1_weights = tf.constant(value=1.0,
+                                 dtype=tf.float32,
+                                 shape=(3, 3, 3, 32),
+                                 name='1')
+        conv = tf.nn.conv2d(g1_input,
+                            g1_weights, [1, 1, 1, 1],
+                            "SAME",
+                            name='conv1')
+        tf.add(conv, conv, name='add1')
+
+
 @tf_test
 def conv_nchw_test(g1):
     with g1.as_default():
@@ -213,6 +230,40 @@ def conv_nchw_test(g1):
                      name='conv1')
 
 
+@tf_test
+def conv_relu_test(g1):
+    with g1.as_default():
+        g1_input = tf.compat.v1.placeholder(tf.float32,
+                                            shape=(1, 16, 16, 3),
+                                            name='0')
+        g1_weights = tf.constant(value=1.0,
+                                 dtype=tf.float32,
+                                 shape=(3, 3, 3, 32),
+                                 name='1')
+        conv = tf.nn.conv2d(g1_input,
+                            g1_weights, [1, 1, 1, 1],
+                            "SAME",
+                            name='conv1')
+        tf.nn.relu(conv, name='relu1')
+
+
+@tf_test
+def conv_relu6_test(g1):
+    with g1.as_default():
+        g1_input = tf.compat.v1.placeholder(tf.float32,
+                                            shape=(1, 16, 16, 3),
+                                            name='0')
+        g1_weights = tf.constant(value=1.0,
+                                 dtype=tf.float32,
+                                 shape=(3, 3, 3, 32),
+                                 name='1')
+        conv = tf.nn.conv2d(g1_input,
+                            g1_weights, [1, 1, 1, 1],
+                            "SAME",
+                            name='conv1')
+        tf.nn.relu6(conv, name='relu1')
+
+
 @tf_test
 def depthwiseconv_test(g1):
     with g1.as_default():
@@ -582,7 +633,10 @@ if __name__ == '__main__':
     concat_test()
     const_test()
     conv_test()
+    conv_add_test()
     conv_nchw_test()
+    conv_relu_test()
+    conv_relu6_test()
     depthwiseconv_test()
     expanddims_test()
     gather_test()

test/tf/tf_test.cpp

@@ -303,6 +303,17 @@ TEST_CASE(conv_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(conv_add_test)
+{
+    migraphx::program p = create_conv();
+    auto* mm            = p.get_main_module();
+    auto l0             = std::prev(mm->end());
+    mm->add_instruction(migraphx::make_op("add"), l0, l0);
+    auto prog = optimize_tf("conv_add_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(conv_nchw_test)
 {
     migraphx::program p = create_conv();
@@ -311,6 +322,35 @@ TEST_CASE(conv_nchw_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(conv_relu_test)
+{
+    migraphx::program p = create_conv();
+    auto* mm            = p.get_main_module();
+    auto l0             = std::prev(mm->end());
+    mm->add_instruction(migraphx::make_op("relu"), l0);
+    auto prog = optimize_tf("conv_relu_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(conv_relu6_test)
+{
+    migraphx::program p = create_conv();
+    auto* mm            = p.get_main_module();
+    std::vector<size_t> input_lens{1, 32, 16, 16};
+    auto l0      = std::prev(mm->end());
+    auto min_val = mm->add_literal(0.0f);
+    auto max_val = mm->add_literal(6.0f);
+    min_val      = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"output_lens", input_lens}}), min_val);
+    max_val = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"output_lens", input_lens}}), max_val);
+    mm->add_instruction(migraphx::make_op("clip"), l0, min_val, max_val);
+    auto prog = optimize_tf("conv_relu6_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(depthwiseconv_test)
 {
     migraphx::program p;