merge develop branch

src/include/migraphx/op/flatten.hpp

@@ -39,7 +39,7 @@ struct flatten
     std::string name() const { return "flatten"; }
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this}.has(1).standard();
         auto&& lens = inputs.front().lens();
         auto x =
             std::accumulate(lens.begin(), lens.begin() + axis, std::size_t{1}, std::multiplies<>{});

src/onnx/parse_generic_op.cpp

@@ -49,7 +49,7 @@ struct parse_generic_op : op_parser<parse_generic_op>
 
     bool needs_contiguous(const std::string& op_name) const
     {
-        return contains({"gather", "scatter"}, op_name);
+        return contains({"flatten", "gather", "scatter"}, op_name);
     }
 
     instruction_ref parse(const op_desc& opd,

test/eliminate_contiguous_test.cpp

@@ -131,4 +131,32 @@ TEST_CASE(non_standard_return_input)
     EXPECT(std::distance(m.begin(), m.end()) == count);
 }
 
+TEST_CASE(non_standard_flatten_op)
+{
+    migraphx::module m;
+
+    auto l = m.add_parameter("x", {migraphx::shape::float_type, {2, 6, 6, 6}});
+    auto t = m.add_instruction(
+        migraphx::make_op("slice", {{"axes", {2, 3}}, {"starts", {1, 1}}, {"ends", {6, 6}}}), l);
+    auto c = m.add_instruction(migraphx::make_op("contiguous"), t);
+    m.add_instruction(migraphx::make_op("flatten"), c);
+    auto count = std::distance(m.begin(), m.end());
+    run_pass(m);
+    EXPECT(std::distance(m.begin(), m.end()) == count);
+}
+
+TEST_CASE(standard_flatten_op)
+{
+    migraphx::module m;
+
+    auto l = m.add_parameter("x", {migraphx::shape::float_type, {2, 6, 6, 6}});
+    auto t = m.add_instruction(
+        migraphx::make_op("slice", {{"axes", {0, 1}}, {"starts", {1, 1}}, {"ends", {6, 6}}}), l);
+    auto c = m.add_instruction(migraphx::make_op("contiguous"), t);
+    m.add_instruction(migraphx::make_op("flatten"), c);
+    auto count = std::distance(m.begin(), m.end());
+    run_pass(m);
+    EXPECT(std::distance(m.begin(), m.end()) == (count - 1));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/onnx/gen_onnx.py

@@ -1232,98 +1232,6 @@ def equal_bool_test():
     return ([node1, node2], [x1, x2], [y])
 
 
-@onnx_test
-def greater_test():
-    ax1 = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
-    x1 = helper.make_tensor("x1",
-                            data_type=TensorProto.FLOAT,
-                            dims=(2, 3),
-                            vals=ax1.astype(np.float32))
-
-    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [2, 3])
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
-
-    node = onnx.helper.make_node(
-        'Greater',
-        inputs=['x1', 'x2'],
-        outputs=['y'],
-    )
-
-    return ([node], [x2], [y], [x1])
-
-
-@onnx_test
-def greater_bool_test():
-
-    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [2, 3])
-    x2 = helper.make_tensor_value_info('x2', TensorProto.BOOL, [2, 3])
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
-
-    node1 = onnx.helper.make_node('Cast', inputs=['x1'], outputs=['bx1'], to=9)
-
-    node2 = onnx.helper.make_node(
-        'Greater',
-        inputs=['bx1', 'x2'],
-        outputs=['y'],
-    )
-
-    return ([node1, node2], [x1, x2], [y])
-
-
-@onnx_test
-def less_test():
-    ax1 = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
-    x1 = helper.make_tensor("x1",
-                            data_type=TensorProto.FLOAT,
-                            dims=(2, 3),
-                            vals=ax1.astype(np.float32))
-
-    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [2, 3])
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
-
-    node = onnx.helper.make_node(
-        'Less',
-        inputs=['x1', 'x2'],
-        outputs=['y'],
-    )
-
-    return ([node], [x2], [y], [x1])
-
-
-@onnx_test
-def less_bool_test():
-
-    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [2, 3])
-    x2 = helper.make_tensor_value_info('x2', TensorProto.BOOL, [2, 3])
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
-
-    node1 = onnx.helper.make_node('Cast', inputs=['x1'], outputs=['bx1'], to=9)
-
-    node2 = onnx.helper.make_node(
-        'Less',
-        inputs=['bx1', 'x2'],
-        outputs=['y'],
-    )
-
-    return ([node1, node2], [x1, x2], [y])
-
-
-@onnx_test
-def lessorequal_test():
-
-    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [3])
-    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [3])
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3])
-
-    node = onnx.helper.make_node(
-        'LessOrEqual',
-        inputs=['x1', 'x2'],
-        outputs=['y'],
-    )
-
-    return ([node], [x1, x2], [y])
-
-
 @onnx_test
 def erf_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10, 15])
@@ -1391,6 +1299,29 @@ def flatten_test():
     return ([node, node2], [x], [y, y2])
 
 
+@onnx_test
+def flatten_nonstd_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [2, 3, 5, 4])
+    y = helper.make_tensor_value_info('2', TensorProto.FLOAT, [6, 20])
+    y2 = helper.make_tensor_value_info('3', TensorProto.FLOAT, [2, 60])
+
+    trans = helper.make_node(
+        'Transpose',
+        inputs=['0'],
+        outputs=['tx'],
+        perm=[0, 1, 3, 2],
+    )
+
+    node = onnx.helper.make_node('Flatten',
+                                 inputs=['tx'],
+                                 axis=2,
+                                 outputs=['2'])
+
+    node2 = onnx.helper.make_node('Flatten', inputs=['tx'], outputs=['3'])
+
+    return ([trans, node, node2], [x], [y, y2])
+
+
 @onnx_test
 def floor_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10])
@@ -1534,6 +1465,44 @@ def globalmaxpool_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def greater_test():
+    ax1 = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
+    x1 = helper.make_tensor("x1",
+                            data_type=TensorProto.FLOAT,
+                            dims=(2, 3),
+                            vals=ax1.astype(np.float32))
+
+    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [2, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
+
+    node = onnx.helper.make_node(
+        'Greater',
+        inputs=['x1', 'x2'],
+        outputs=['y'],
+    )
+
+    return ([node], [x2], [y], [x1])
+
+
+@onnx_test
+def greater_bool_test():
+
+    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [2, 3])
+    x2 = helper.make_tensor_value_info('x2', TensorProto.BOOL, [2, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
+
+    node1 = onnx.helper.make_node('Cast', inputs=['x1'], outputs=['bx1'], to=9)
+
+    node2 = onnx.helper.make_node(
+        'Greater',
+        inputs=['bx1', 'x2'],
+        outputs=['y'],
+    )
+
+    return ([node1, node2], [x1, x2], [y])
+
+
 @onnx_test
 def group_conv_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 4, 16, 16])
@@ -2231,6 +2200,60 @@ def leaky_relu_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def less_test():
+    ax1 = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
+    x1 = helper.make_tensor("x1",
+                            data_type=TensorProto.FLOAT,
+                            dims=(2, 3),
+                            vals=ax1.astype(np.float32))
+
+    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [2, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
+
+    node = onnx.helper.make_node(
+        'Less',
+        inputs=['x1', 'x2'],
+        outputs=['y'],
+    )
+
+    return ([node], [x2], [y], [x1])
+
+
+@onnx_test
+def less_bool_test():
+
+    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [2, 3])
+    x2 = helper.make_tensor_value_info('x2', TensorProto.BOOL, [2, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3])
+
+    node1 = onnx.helper.make_node('Cast', inputs=['x1'], outputs=['bx1'], to=9)
+
+    node2 = onnx.helper.make_node(
+        'Less',
+        inputs=['bx1', 'x2'],
+        outputs=['y'],
+    )
+
+    return ([node1, node2], [x1, x2], [y])
+
+
+@onnx_test
+def lessorequal_test():
+
+    x1 = helper.make_tensor_value_info('x1', TensorProto.FLOAT, [3])
+    x2 = helper.make_tensor_value_info('x2', TensorProto.FLOAT, [3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3])
+
+    node = onnx.helper.make_node(
+        'LessOrEqual',
+        inputs=['x1', 'x2'],
+        outputs=['y'],
+    )
+
+    return ([node], [x1, x2], [y])
+
+
 @onnx_test
 def log_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10])

test/onnx/onnx_test.cpp

@@ -1183,6 +1183,21 @@ TEST_CASE(flatten_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(flatten_nonstd_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 5, 4}});
+    auto l1  = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), l0);
+    auto l2  = mm->add_instruction(migraphx::make_op("contiguous"), l1);
+    mm->add_instruction(migraphx::make_op("flatten", {{"axis", 2}}), l2);
+    auto l3 = mm->add_instruction(migraphx::make_op("contiguous"), l1);
+    mm->add_instruction(migraphx::make_op("flatten", {{"axis", 1}}), l3);
+    auto prog = optimize_onnx("flatten_nonstd_test.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(floor_test)
 {
     migraphx::program p;