Add cont before reshape in parsing resize (#792)

* add a contiguous check before the reshape in parsing resize

* clang format

* add a unit test to the fix in parse_resize

* clang format

src/onnx/parse_resize.cpp

@@ -189,8 +189,9 @@ struct parse_resize : op_parser<parse_resize>
         // reshape input to one-dimension
         std::vector<int64_t> rsp_lens = {static_cast<int64_t>(in_s.elements())};
         shape ind_s{shape::int32_type, out_lens};
-        auto rsp     = info.add_instruction(make_op("reshape", {{"dims", rsp_lens}}), args[0]);
-        auto ins_ind = info.add_literal(literal(ind_s, ind));
+        auto arg_cont = info.make_contiguous(args[0]);
+        auto rsp      = info.add_instruction(make_op("reshape", {{"dims", rsp_lens}}), arg_cont);
+        auto ins_ind  = info.add_literal(literal(ind_s, ind));
         return info.add_instruction(make_op("gather", {{"axis", 0}}), rsp, ins_ind);
     }
 };

test/onnx/gen_onnx.py

@@ -3069,6 +3069,32 @@ def resize_downsample_c_test():
     return ([node], [X], [Y], [scale_tensor])
 
 
+@onnx_test
+def resize_nonstd_input_test():
+    scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32)
+    scale_tensor = helper.make_tensor(name='scales',
+                                      data_type=TensorProto.FLOAT,
+                                      dims=scales.shape,
+                                      vals=scales.flatten().astype(np.float32))
+
+    X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [1, 1, 4, 2])
+    Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1, 1, 1, 2])
+
+    trn = onnx.helper.make_node('Transpose',
+                                inputs=['X'],
+                                outputs=['TX'],
+                                perm=[0, 1, 3, 2])
+
+    node = onnx.helper.make_node('Resize',
+                                 inputs=['TX', '', 'scales'],
+                                 outputs=['Y'],
+                                 coordinate_transformation_mode='asymmetric',
+                                 mode='nearest',
+                                 nearest_mode='ceil')
+
+    return ([trn, node], [X], [Y], [scale_tensor])
+
+
 @onnx_test
 def resize_outsize_test():
     out_lens = np.array([1, 1, 4, 6], dtype=np.int64)

test/onnx/onnx_test.cpp

@@ -2690,6 +2690,35 @@ TEST_CASE(resize_outsize_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(resize_nonstd_input_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    std::vector<float> ds = {1.0f, 1.0f, 0.6f, 0.6f};
+    migraphx::shape ss{migraphx::shape::float_type, {4}};
+    mm->add_literal(migraphx::literal{ss, ds});
+
+    migraphx::shape sx{migraphx::shape::float_type, {1, 1, 4, 2}};
+    auto inx = mm->add_parameter("X", sx);
+
+    migraphx::shape si{migraphx::shape::int32_type, {1, 1, 1, 2}};
+    std::vector<int> ind = {0, 4};
+    auto li              = mm->add_literal(migraphx::literal(si, ind));
+
+    auto tx = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), inx);
+    mm->add_instruction(migraphx::make_op("undefined"));
+    auto tx_cont = mm->add_instruction(migraphx::make_op("contiguous"), tx);
+
+    auto lrsp = mm->add_instruction(migraphx::make_op("reshape", {{"dims", {8}}}), tx_cont);
+    auto r    = mm->add_instruction(migraphx::make_op("gather", {{"axis", 0}}), lrsp, li);
+    mm->add_return({r});
+
+    auto prog = migraphx::parse_onnx("resize_nonstd_input_test.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(resize_upsample_pc_test)
 {
     migraphx::program p;