Merge remote-tracking branch 'origin/develop' into ck-proto

test/get_target_assignments.cpp → test/fpga/get_target_assignments.cpp

@@ -26,8 +26,9 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/register_target.hpp>
-#include <migraphx/ref/target.hpp>
+#include <migraphx/fpga/target.hpp>
 #include <migraphx/target_assignments.hpp>
+#include <migraphx/iterator_for.hpp>
 
 migraphx::program create_program()
 {
@@ -37,8 +38,8 @@ migraphx::program create_program()
     auto x    = mm->add_parameter("x", s);
     auto y    = mm->add_parameter("y", s);
     auto z    = mm->add_parameter("z", s);
-    auto diff = mm->add_instruction(migraphx::make_op("div"), x, y);
-    mm->add_instruction(migraphx::make_op("div"), diff, z);
+    auto diff = mm->add_instruction(migraphx::make_op("add"), x, y);
+    mm->add_instruction(migraphx::make_op("add"), diff, z);
     return p;
 }
 
@@ -47,14 +48,16 @@ TEST_CASE(is_supported)
     auto p       = create_program();
     auto targets = migraphx::get_targets();
     EXPECT(!targets.empty());
-    auto first_target = targets[0];
-    auto t            = migraphx::make_target(first_target);
+    auto t = migraphx::make_target("fpga");
 
     const auto assignments = p.get_target_assignments({t});
-    for(const auto& [ins, target] : assignments)
+    const auto* mod        = p.get_main_module();
+    EXPECT(mod->size() == assignments.size());
+
+    for(const auto ins : iterator_for(*mod))
     {
-        (void)ins;
-        EXPECT(target == first_target);
+        const auto& target = assignments.at(ins);
+        EXPECT(target == "fpga");
     }
 }
 

test/include/test.hpp

@@ -108,15 +108,7 @@ struct function
 };
 
 template <class Stream, class Iterator>
-inline Stream& stream_range(Stream& s, Iterator start, Iterator last)
-{
-    if(start != last)
-    {
-        s << *start;
-        std::for_each(std::next(start), last, [&](auto&& x) { s << ", " << x; });
-    }
-    return s;
-}
+Stream& stream_range(Stream& s, Iterator start, Iterator last);
 
 template <class Stream>
 inline Stream& operator<<(Stream& s, std::nullptr_t)
@@ -136,6 +128,17 @@ inline auto operator<<(Stream& s, const Range& v) -> decltype(stream_range(s, v.
     return s;
 }
 
+template <class Stream, class Iterator>
+inline Stream& stream_range(Stream& s, Iterator start, Iterator last)
+{
+    if(start != last)
+    {
+        s << *start;
+        std::for_each(std::next(start), last, [&](auto&& x) { s << ", " << x; });
+    }
+    return s;
+}
+
 template <class T>
 const T& get_value(const T& x)
 {

test/onnx/gen_onnx.py

@@ -3589,7 +3589,7 @@ def nms_test():
     st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
                                        [1])
     out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
-                                        [6, 3])
+                                        [None, 3])
 
     node = onnx.helper.make_node('NonMaxSuppression',
                                  inputs=[
@@ -3603,6 +3603,108 @@ def nms_test():
     return ([node], [b, s, mo, iou, st], [out])
 
 
+@onnx_test
+def nms_use_dyn_output_false_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT, [1, 1, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'],
+                                 use_dyn_output=0)
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_batch_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [None, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [None, 1, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'],
+                                 center_point_box=1,
+                                 use_dyn_output=1)
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_boxes_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, None, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [1, 1, None])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'])
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_classes_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [1, None, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'])
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
 @onnx_test
 def not_test():
     x = helper.make_tensor_value_info('0', TensorProto.INT32, [4])

test/onnx/onnx_test.cpp

@@ -954,13 +954,11 @@ TEST_CASE(conv_dynamic_img_same_upper)
 
 TEST_CASE(conv_dynamic_kernel_same_lower)
 {
-    std::cout << "here1\n";
     migraphx::program p;
     auto* mm = p.get_main_module();
     auto l0  = mm->add_parameter("0", {migraphx::shape::float_type, {1, 3, 5, 5}});
     auto l1  = mm->add_parameter(
         "1", {migraphx::shape::float_type, {{1, 1, 0}, {3, 3, 0}, {2, 4, 0}, {2, 4, 0}}});
-    std::cout << "here2\n";
     auto c0 = mm->add_instruction(
         migraphx::make_op("convolution",
                           {{"padding", {0, 0}},
@@ -970,12 +968,10 @@ TEST_CASE(conv_dynamic_kernel_same_lower)
                            {"use_dynamic_same_auto_pad", true}}),
         l0,
         l1);
-    std::cout << "here3\n";
     mm->add_return({c0});
 
     migraphx::onnx_options options;
     options.default_dyn_dim_value = {2, 4, 0};
-    std::cout << "here\n";
     auto prog = migraphx::parse_onnx("conv_dynamic_kernel_same_lower_test.onnx", options);
     EXPECT(p == prog);
 }
@@ -3382,13 +3378,127 @@ TEST_CASE(nms_test)
     auto st = mm->add_parameter("score_threshold", sst);
 
     auto ret = mm->add_instruction(
-        migraphx::make_op("nonmaxsuppression", {{"center_point_box", 1}}), b, s, mo, iou, st);
+        migraphx::make_op("nonmaxsuppression", {{"center_point_box", true}}), b, s, mo, iou, st);
     mm->add_return({ret});
 
     auto prog = migraphx::parse_onnx("nms_test.onnx");
     EXPECT(p == prog);
 }
 
+TEST_CASE(nms_dynamic_batch_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape sb{migraphx::shape::float_type, {{1, 10, 0}, {6, 6, 0}, {4, 4, 0}}};
+    auto b = mm->add_parameter("boxes", sb);
+    migraphx::shape ss{migraphx::shape::float_type, {{1, 10, 0}, {1, 1, 0}, {6, 6, 0}}};
+    auto s = mm->add_parameter("scores", ss);
+    migraphx::shape smo{migraphx::shape::int64_type, {1}};
+    auto mo = mm->add_parameter("max_output_boxes_per_class", smo);
+    migraphx::shape siou{migraphx::shape::float_type, {1}};
+    auto iou = mm->add_parameter("iou_threshold", siou);
+    migraphx::shape sst{migraphx::shape::float_type, {1}};
+    auto st  = mm->add_parameter("score_threshold", sst);
+    auto ret = mm->add_instruction(
+        migraphx::make_op("nonmaxsuppression",
+                          {{"center_point_box", true}, {"use_dyn_output", true}}),
+        b,
+        s,
+        mo,
+        iou,
+        st);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 10, 0};
+    options.use_dyn_output        = true;
+
+    auto prog = migraphx::parse_onnx("nms_dynamic_batch_test.onnx", options);
+    EXPECT(p == prog);
+}
+
+TEST_CASE(nms_dynamic_boxes_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape sb{migraphx::shape::float_type, {{1, 1, 0}, {6, 20, 0}, {4, 4, 0}}};
+    auto b = mm->add_parameter("boxes", sb);
+    migraphx::shape ss{migraphx::shape::float_type, {{1, 1, 0}, {1, 1, 0}, {6, 20, 0}}};
+    auto s = mm->add_parameter("scores", ss);
+    migraphx::shape smo{migraphx::shape::int64_type, {1}};
+    auto mo = mm->add_parameter("max_output_boxes_per_class", smo);
+    migraphx::shape siou{migraphx::shape::float_type, {1}};
+    auto iou = mm->add_parameter("iou_threshold", siou);
+    migraphx::shape sst{migraphx::shape::float_type, {1}};
+    auto st  = mm->add_parameter("score_threshold", sst);
+    auto ret = mm->add_instruction(
+        migraphx::make_op("nonmaxsuppression", {{"use_dyn_output", true}}), b, s, mo, iou, st);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {6, 20, 0};
+    options.use_dyn_output        = true;
+
+    auto prog = migraphx::parse_onnx("nms_dynamic_boxes_test.onnx", options);
+    EXPECT(p == prog);
+}
+
+TEST_CASE(nms_dynamic_classes_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape sb{migraphx::shape::float_type, {1, 6, 4}};
+    auto b = mm->add_parameter("boxes", sb);
+    migraphx::shape ss{migraphx::shape::float_type, {{1, 1, 0}, {1, 10, 0}, {6, 6, 0}}};
+    auto s = mm->add_parameter("scores", ss);
+    migraphx::shape smo{migraphx::shape::int64_type, {1}};
+    auto mo = mm->add_parameter("max_output_boxes_per_class", smo);
+    migraphx::shape siou{migraphx::shape::float_type, {1}};
+    auto iou = mm->add_parameter("iou_threshold", siou);
+    migraphx::shape sst{migraphx::shape::float_type, {1}};
+    auto st  = mm->add_parameter("score_threshold", sst);
+    auto ret = mm->add_instruction(
+        migraphx::make_op("nonmaxsuppression", {{"use_dyn_output", true}}), b, s, mo, iou, st);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 10, 0};
+    options.use_dyn_output        = true;
+
+    auto prog = migraphx::parse_onnx("nms_dynamic_classes_test.onnx", options);
+    EXPECT(p == prog);
+}
+
+TEST_CASE(nms_overwrite_use_dyn_output_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape sb{migraphx::shape::float_type, {1, 6, 4}};
+    auto b = mm->add_parameter("boxes", sb);
+
+    migraphx::shape ss{migraphx::shape::float_type, {1, 1, 6}};
+    auto s = mm->add_parameter("scores", ss);
+
+    migraphx::shape smo{migraphx::shape::int64_type, {1}};
+    auto mo = mm->add_parameter("max_output_boxes_per_class", smo);
+
+    migraphx::shape siou{migraphx::shape::float_type, {1}};
+    auto iou = mm->add_parameter("iou_threshold", siou);
+
+    migraphx::shape sst{migraphx::shape::float_type, {1}};
+    auto st = mm->add_parameter("score_threshold", sst);
+
+    auto ret = mm->add_instruction(
+        migraphx::make_op("nonmaxsuppression", {{"use_dyn_output", true}}), b, s, mo, iou, st);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.use_dyn_output = true;
+
+    auto prog = migraphx::parse_onnx("nms_use_dyn_output_false_test.onnx", options);
+    EXPECT(p == prog);
+}
+
 TEST_CASE(nonzero_dynamic_test)
 {
     migraphx::program p;

test/op_shape_test.cpp

@@ -1135,6 +1135,149 @@ TEST_CASE(multinomial)
     throws_shape(migraphx::make_op("multinomial", {{"dtype", dtype}}), s, s);
 }
 
+TEST_CASE(nms_shape)
+{
+    // use_dyn_output == false
+    migraphx::shape boxes_s{migraphx::shape::float_type, {1, 6, 4}};
+    migraphx::shape scores_s{migraphx::shape::float_type, {1, 1, 6}};
+    migraphx::shape max_out_s{migraphx::shape::int64_type, {1}};
+    migraphx::shape iou_thres_s{migraphx::shape::float_type, {1}};
+    migraphx::shape score_thres_s{migraphx::shape::float_type, {1}};
+    migraphx::shape output_s{migraphx::shape::int64_type, {6, 3}};
+    expect_shape(output_s,
+                 migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", false}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // use_dyn_output == true
+    output_s = {migraphx::shape::int64_type, {{0, 6, 0}, {3, 3, 0}}};
+    expect_shape(output_s,
+                 migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic batches
+    boxes_s  = {migraphx::shape::float_type, {{1, 3, 0}, {6, 6, 0}, {4, 4, 0}}};
+    scores_s = {migraphx::shape::float_type, {{1, 3, 0}, {1, 1, 0}, {6, 6, 0}}};
+    output_s = {migraphx::shape::int64_type, {{0, 18, 0}, {3, 3, 0}}};
+    expect_shape(output_s,
+                 migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic num boxes
+    boxes_s  = {migraphx::shape::float_type, {{1, 1, 0}, {6, 20, 0}, {4, 4, 0}}};
+    scores_s = {migraphx::shape::float_type, {{1, 1, 0}, {1, 1, 0}, {6, 20, 0}}};
+    output_s = {migraphx::shape::int64_type, {{0, 20, 0}, {3, 3, 0}}};
+    expect_shape(output_s,
+                 migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // use_dyn_output false with dynamic input shape
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", false}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic classes
+    boxes_s  = {migraphx::shape::float_type, {{1, 1, 0}, {6, 6, 0}, {4, 4, 0}}};
+    scores_s = {migraphx::shape::float_type, {{1, 1, 0}, {1, 3, 0}, {6, 6, 0}}};
+    output_s = {migraphx::shape::int64_type, {{0, 6, 0}, {3, 3, 0}}};
+    expect_shape(output_s,
+                 migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // fixed mismatch batches
+    boxes_s  = {migraphx::shape::float_type, {2, 6, 4}};
+    scores_s = {migraphx::shape::float_type, {1, 1, 6}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // fixed mismatch num boxes
+    boxes_s  = {migraphx::shape::float_type, {1, 6, 4}};
+    scores_s = {migraphx::shape::float_type, {1, 1, 4}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic mismatch batches
+    boxes_s  = {migraphx::shape::float_type, {{1, 4, 0}, {6, 6, 0}, {4, 4, 0}}};
+    scores_s = {migraphx::shape::float_type, {{2, 8, 0}, {1, 1, 0}, {6, 6, 0}}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic mismatch num boxes
+    boxes_s  = {migraphx::shape::float_type, {{1, 1, 0}, {6, 8, 0}, {4, 4, 0}}};
+    scores_s = {migraphx::shape::float_type, {{1, 1, 0}, {1, 1, 0}, {3, 9, 0}}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+
+    // dynamic number of classes, fixed boxes_s, mismatch batches
+    boxes_s  = {migraphx::shape::float_type, {1, 6, 4}};
+    scores_s = {migraphx::shape::float_type, {{1, 3, 0}, {1, 3, 0}, {6, 6, 0}}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+    // dynamic number of classes, fixed boxes_s, mismatch num boxes
+    boxes_s  = {migraphx::shape::float_type, {1, 6, 4}};
+    scores_s = {migraphx::shape::float_type, {{1, 1, 0}, {1, 3, 0}, {4, 8, 0}}};
+    throws_shape(migraphx::make_op("nonmaxsuppression",
+                                   {{"center_point_box", true}, {"use_dyn_output", true}}),
+                 boxes_s,
+                 scores_s,
+                 max_out_s,
+                 iou_thres_s,
+                 score_thres_s);
+}
+
 TEST_CASE(pooling_shape)
 {
     migraphx::shape output{migraphx::shape::float_type, {4, 3, 1, 1}};

test/simplify_reshapes_test.cpp

@@ -39,6 +39,15 @@ void run_pass(migraphx::module& m)
     migraphx::run_passes(m, {migraphx::simplify_reshapes{}, migraphx::dead_code_elimination{}});
 }
 
+inline std::vector<std::vector<std::size_t>> to_lens(const std::vector<migraphx::shape>& shapes)
+{
+    std::vector<std::vector<std::size_t>> result;
+    std::transform(shapes.begin(), shapes.end(), std::back_inserter(result), [&](const auto& s) {
+        return s.lens();
+    });
+    return result;
+}
+
 TEST_CASE(double_contig)
 {
     migraphx::program p;
@@ -1275,4 +1284,82 @@ TEST_CASE(transpose_slice_single_transpose)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(transpose_slice_non_packed_axis)
+{
+    migraphx::module m1;
+    {
+        auto x = m1.add_parameter("x", {migraphx::shape::float_type, {2, 384, 36, 64}});
+        auto transpose =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto sqrt = m1.add_instruction(migraphx::make_op("sqrt"), slice);
+        m1.add_return({sqrt});
+    }
+    auto output_shapes = m1.get_output_shapes();
+    run_pass(m1);
+    EXPECT(m1.get_output_shapes() == output_shapes);
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {2, 384, 36, 64}});
+        auto unsqueeze =
+            m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {12}}}), x);
+        auto transpose = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {3, 0, 2, 1, 4}}}), unsqueeze);
+        auto slice = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), transpose);
+        auto squeeze = m2.add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), slice);
+        auto sqrt    = m2.add_instruction(migraphx::make_op("sqrt"), squeeze);
+        m2.add_return({sqrt});
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice_non_packed_multi_axis)
+{
+    migraphx::module m1;
+    {
+        auto x = m1.add_parameter("x", {migraphx::shape::float_type, {2, 384, 36, 64}});
+        auto transpose =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        m1.add_return({slice1, transpose2, slice3});
+    }
+    auto output_shapes = m1.get_output_shapes();
+    run_pass(m1);
+    EXPECT(to_lens(m1.get_output_shapes()) == to_lens(output_shapes));
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {2, 384, 36, 64}});
+        auto unsqueeze =
+            m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {12}}}), x);
+        auto transpose = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {3, 0, 2, 1, 4}}}), unsqueeze);
+        auto slice1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), transpose);
+        auto squeeze1 = m2.add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), slice1);
+        auto slice2   = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {1}}, {"ends", {2}}}), transpose);
+        auto squeeze2   = m2.add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), slice2);
+        auto transpose2 = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), squeeze2);
+        auto slice3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {2}}, {"ends", {3}}}), transpose);
+        auto squeeze3 = m2.add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), slice3);
+        m2.add_return({squeeze1, transpose2, squeeze3});
+    }
+    EXPECT(m1.sort() == m2.sort());
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/verify/test_layernorm.cpp

@@ -68,7 +68,7 @@ struct test_layernorm : verify_program<test_layernorm>
     {
         migraphx::program p;
         auto* mm                 = p.get_main_module();
-        std::vector<size_t> dims = {1, 1, 5};
+        std::vector<size_t> dims = {1, 2, 5};
         auto x = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, dims});
         add_layernorm(*mm, x, dims);
         return p;

test/verify/test_softmax_large1.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/common.hpp>
+
+struct test_softmax_large1 : verify_program<test_softmax_large1>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm   = p.get_main_module();
+        auto x     = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {2, 4}});
+        auto large = mm->add_literal({migraphx::shape{migraphx::shape::float_type}, {10000}});
+        auto add   = migraphx::add_common_op(*mm, migraphx::make_op("add"), {x, large});
+        mm->add_instruction(migraphx::make_op("softmax", {{"axis", -1}}), add);
+        return p;
+    }
+};