Merge branch 'develop' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into dyn_test_runner

.github/workflows/gh_project.yaml → .github/workflows/add-to-project.yaml

-name: Update Github Project
+name: Add items to GH project
 
 on:
   pull_request:
@@ -16,4 +16,4 @@ jobs:
       - uses: actions/add-to-project@v0.4.0
         with:
           project-url: https://github.com/orgs/ROCmSoftwarePlatform/projects/20
-          github-token: $${{ ${{ github.token }} }}
+          github-token: ${{ secrets.TEST_PR_WORKFLOW }}

src/onnx/parse_matmul.cpp

@@ -43,55 +43,79 @@ struct parse_matmul : op_parser<parse_matmul>
                           const onnx_parser::node_info& info,
                           std::vector<instruction_ref> args) const
     {
-        auto l0      = args[0];
-        auto l1      = args[1];
-        auto l0_lens = l0->get_shape().lens();
-        auto l1_lens = l1->get_shape().lens();
+        auto a0 = args[0];
+        auto a1 = args[1];
+        auto s0 = a0->get_shape();
+        auto s1 = a1->get_shape();
 
-        // args[0] is a vector, prepend 1 to the shape
+        instruction_ref dot_res;
         bool is_a_prepended = false;
-        if(l0_lens.size() == 1)
+        bool is_b_appended  = false;
+        if(s0.ndim() == 1)
         {
             is_a_prepended = true;
-            l0_lens.insert(l0_lens.begin(), 1);
-            l0 = info.add_instruction(make_op("unsqueeze", {{"axes", {0}}}), args[0]);
+            a0             = info.add_instruction(make_op("unsqueeze", {{"axes", {0}}}), args[0]);
         }
-
-        bool is_b_appended = false;
-        if(l1_lens.size() == 1)
+        if(s1.ndim() == 1)
         {
             is_b_appended = true;
-            l1_lens.push_back(1);
-            l1 = info.add_instruction(make_op("unsqueeze", {{"axes", {1}}}), args[1]);
+            a1            = info.add_instruction(make_op("unsqueeze", {{"axes", {1}}}), args[1]);
         }
 
-        instruction_ref bl0 = l0;
-        instruction_ref bl1 = l1;
-        if(not std::equal(
-               l0_lens.rbegin() + 2, l0_lens.rend(), l1_lens.rbegin() + 2, l1_lens.rend()))
+        if(s0.dynamic() or s1.dynamic())
         {
-            auto l0_it = l0_lens.begin() + l0_lens.size() - 2;
-            std::vector<std::size_t> l0_broadcasted_lens(l0_lens.begin(), l0_it);
-            auto l1_it = l1_lens.begin() + l1_lens.size() - 2;
-            std::vector<std::size_t> l1_broadcasted_lens(l1_lens.begin(), l1_it);
-            auto output_lens = compute_broadcasted_lens(l0_broadcasted_lens, l1_broadcasted_lens);
-            l0_broadcasted_lens = output_lens;
-            l0_broadcasted_lens.insert(l0_broadcasted_lens.end(), l0_it, l0_lens.end());
-            l1_broadcasted_lens = output_lens;
-            l1_broadcasted_lens.insert(l1_broadcasted_lens.end(), l1_it, l1_lens.end());
-            if(l0_lens != l0_broadcasted_lens)
+            if(opd.op_name == "quant_dot")
+            {
+                MIGRAPHX_THROW("PARSE_MATMUL: dynamic MatMulInteger not supported");
+            }
+            auto s0_dds = a0->get_shape().to_dynamic().dyn_dims();
+            auto s1_dds = a1->get_shape().to_dynamic().dyn_dims();
+
+            // TODO: handling this case requires a new multibroadcast mode
+            if(not std::equal(
+                   s0_dds.rbegin() + 2, s0_dds.rend(), s1_dds.rbegin() + 2, s1_dds.rend()))
             {
-                bl0 = info.add_instruction(
-                    make_op("multibroadcast", {{"out_lens", l0_broadcasted_lens}}), l0);
+                MIGRAPHX_THROW("PARSE_MATMUL: dynamic shape broadcasting not supported");
             }
-            if(l1_lens != l1_broadcasted_lens)
+
+            dot_res = info.add_instruction(make_op(opd.op_name), a0, a1);
+        }
+        else
+        {
+            auto s0_lens        = a0->get_shape().lens();
+            auto s1_lens        = a1->get_shape().lens();
+            instruction_ref ba0 = a0;
+            instruction_ref ba1 = a1;
+            // try broadcasting if dimensions other than last two do not match
+            if(not std::equal(
+                   s0_lens.rbegin() + 2, s0_lens.rend(), s1_lens.rbegin() + 2, s1_lens.rend()))
             {
-                bl1 = info.add_instruction(
-                    make_op("multibroadcast", {{"out_lens", l1_broadcasted_lens}}), l1);
+                auto l0_it = s0_lens.begin() + s0_lens.size() - 2;
+                std::vector<std::size_t> l0_broadcasted_lens(s0_lens.begin(), l0_it);
+                auto l1_it = s1_lens.begin() + s1_lens.size() - 2;
+                std::vector<std::size_t> l1_broadcasted_lens(s1_lens.begin(), l1_it);
+                auto output_lens =
+                    compute_broadcasted_lens(l0_broadcasted_lens, l1_broadcasted_lens);
+                l0_broadcasted_lens = output_lens;
+                l0_broadcasted_lens.insert(l0_broadcasted_lens.end(), l0_it, s0_lens.end());
+                l1_broadcasted_lens = output_lens;
+                l1_broadcasted_lens.insert(l1_broadcasted_lens.end(), l1_it, s1_lens.end());
+                if(s0_lens != l0_broadcasted_lens)
+                {
+                    ba0 = info.add_instruction(
+                        make_op("multibroadcast", {{"out_lens", l0_broadcasted_lens}}), a0);
+                }
+                if(s1_lens != l1_broadcasted_lens)
+                {
+                    ba1 = info.add_instruction(
+                        make_op("multibroadcast", {{"out_lens", l1_broadcasted_lens}}), a1);
+                }
             }
+            dot_res = info.add_instruction(make_op(opd.op_name), ba0, ba1);
         }
-        instruction_ref dot_res = info.add_instruction(make_op(opd.op_name), bl0, bl1);
-        int64_t num_axis        = static_cast<int64_t>(dot_res->get_shape().lens().size());
+
+        // squeeze the appended or prepended dimensions
+        int64_t num_axis = dot_res->get_shape().ndim();
         if(is_a_prepended)
         {
             dot_res = info.add_instruction(make_op("squeeze", {{"axes", {num_axis - 2}}}), dot_res);

src/simplify_algebra.cpp

@@ -1065,11 +1065,23 @@ struct find_split_reshape
             return;
         }
 
+        // Only want to apply this optimization if each split output is followed by
+        // a contiguous op and a reshape
+        if(std::any_of(split_outputs.begin(), split_outputs.end(), [](auto i) {
+               if(i->outputs().size() == 1)
+               {
+                   auto cont = i->outputs().front();
+                   return cont->outputs().size() != 1;
+               }
+               return false;
+           }))
+        {
+            return;
+        }
+
         std::vector<instruction_ref> vec_rsp(split_outputs.size());
         std::transform(split_outputs.begin(), split_outputs.end(), vec_rsp.begin(), [](auto i) {
-            assert(i->outputs().size() == 1);
             auto cont = i->outputs().front();
-            assert(cont->outputs().size() == 1);
             return cont->outputs().front();
         });
 

src/simplify_reshapes.cpp

@@ -763,16 +763,23 @@ struct find_transpose_slice
         // Compute axis before transpose to use for unsqueeze
         auto perm    = ins->get_operator().to_value()["permutation"].to_vector<int64_t>();
         auto preaxis = std::find(perm.begin(), perm.end(), axis) - perm.begin();
-        // Make unsqeeze
+        // Make unsqueeze
+        std::vector<int64_t> steps(sdistance.size());
+        std::transform(
+            slice.axes.begin(),
+            slice.axes.end(),
+            sdistance.begin(),
+            steps.begin(),
+            [&](const auto ax, const auto sdis) { return ins->get_shape().lens().at(ax) / sdis; });
         auto unsqueeze = m.insert_instruction(
-            ins, make_op("unsqueeze", {{"axes", {preaxis}}, {"steps", sdistance}}), ins->inputs());
+            ins, make_op("unsqueeze", {{"axes", {preaxis}}, {"steps", steps}}), ins->inputs());
         // Make transpose
         std::transform(perm.begin(), perm.end(), perm.begin(), [&](auto i) {
-            if(i > preaxis)
+            if(i >= preaxis)
                 return i + 1;
             return i;
         });
-        perm.insert(perm.begin(), preaxis + 1);
+        perm.insert(perm.begin(), preaxis);
         auto transpose =
             m.insert_instruction(ins, make_op("transpose", {{"permutation", perm}}), unsqueeze);
         // Slice and squeeze

test/onnx/gen_onnx.py

@@ -3563,6 +3563,81 @@ def matmul_vv_test():
     return ([node], [m1, m2], [y])
 
 
+@onnx_test()
+def matmul_dyn_mm_test():
+    m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None, 7])
+    m2 = helper.make_tensor_value_info('2', TensorProto.FLOAT, [7, None])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None, None])
+
+    node = onnx.helper.make_node(
+        'MatMul',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
+@onnx_test()
+def matmul_dyn_mv_test():
+    m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None, 7])
+    m2 = helper.make_tensor_value_info('2', TensorProto.FLOAT, [7])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None, 1])
+
+    node = onnx.helper.make_node(
+        'MatMul',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
+@onnx_test()
+def matmul_dyn_vm_test():
+    m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [7])
+    m2 = helper.make_tensor_value_info('2', TensorProto.FLOAT, [7, None])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, None])
+
+    node = onnx.helper.make_node(
+        'MatMul',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
+@onnx_test()
+def matmul_dyn_vv_test():
+    m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None])
+    m2 = helper.make_tensor_value_info('2', TensorProto.FLOAT, [None])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1])
+
+    node = onnx.helper.make_node(
+        'MatMul',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
+@onnx_test()
+def matmul_dyn_broadcast_error():
+    m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [7])
+    m2 = helper.make_tensor_value_info('2', TensorProto.FLOAT, [5, 7, None])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [5, None])
+
+    node = onnx.helper.make_node(
+        'MatMul',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
 @onnx_test()
 def matmulinteger_test():
     m1 = helper.make_tensor_value_info('1', TensorProto.INT8, [3, 6, 16])
@@ -3578,6 +3653,21 @@ def matmulinteger_test():
     return ([node], [m1, m2], [y])
 
 
+@onnx_test()
+def matmulinteger_dyn_error():
+    m1 = helper.make_tensor_value_info('1', TensorProto.INT8, [None, 6, 16])
+    m2 = helper.make_tensor_value_info('2', TensorProto.INT8, [None, 16, 8])
+    y = helper.make_tensor_value_info('y', TensorProto.INT32, [None, 6, 8])
+
+    node = onnx.helper.make_node(
+        'MatMulInteger',
+        inputs=['1', '2'],
+        outputs=['y'],
+    )
+
+    return ([node], [m1, m2], [y])
+
+
 @onnx_test()
 def max_test():
     a = helper.make_tensor_value_info('0', TensorProto.FLOAT, [3])

test/onnx/onnx_test.cpp

@@ -3432,6 +3432,92 @@ TEST_CASE(matmul_vv_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(matmul_dyn_mm_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter(
+        "1", migraphx::shape{migraphx::shape::float_type, {{4, 8, 6}, {7, 7, 0}}});
+    auto l1 = mm->add_parameter(
+        "2", migraphx::shape{migraphx::shape::float_type, {{7, 7, 0}, {1, 5, 3}}});
+    auto ret = migraphx::add_apply_alpha_beta(*mm, {l0, l1}, migraphx::make_op("dot"), 1.0f, 0.0f);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.map_dyn_input_dims["1"] = {{4, 8, 6}, {7, 7, 0}};
+    options.map_dyn_input_dims["2"] = {{7, 7, 0}, {1, 5, 3}};
+    auto prog                       = parse_onnx("matmul_dyn_mm_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(matmul_dyn_mv_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter(
+        "1", migraphx::shape{migraphx::shape::float_type, {{4, 8, 6}, {7, 7, 0}}});
+    auto l1  = mm->add_parameter("2", migraphx::shape{migraphx::shape::float_type, {7}});
+    auto sl1 = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), l1);
+    auto res = migraphx::add_apply_alpha_beta(*mm, {l0, sl1}, migraphx::make_op("dot"), 1.0f, 0.0f);
+    auto ret = mm->add_instruction(migraphx::make_op("squeeze", {{"axes", {1}}}), res);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.map_dyn_input_dims["1"] = {{4, 8, 6}, {7, 7, 0}};
+    auto prog                       = parse_onnx("matmul_dyn_mv_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(matmul_dyn_vm_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {7}});
+    auto l1  = mm->add_parameter(
+        "2", migraphx::shape{migraphx::shape::float_type, {{7, 7, 0}, {4, 10, 8}}});
+    auto sl0 = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0}}}), l0);
+    auto res = migraphx::add_apply_alpha_beta(*mm, {sl0, l1}, migraphx::make_op("dot"), 1.0f, 0.0f);
+    auto ret = mm->add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), res);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.map_dyn_input_dims["2"] = {{7, 7, 0}, {4, 10, 8}};
+    auto prog                       = parse_onnx("matmul_dyn_vm_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(matmul_dyn_vv_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape::dynamic_dimension dd{5, 8, 7};
+    auto l0  = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {dd}});
+    auto l1  = mm->add_parameter("2", migraphx::shape{migraphx::shape::float_type, {dd}});
+    auto sl0 = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0}}}), l0);
+    auto sl1 = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), l1);
+    auto res =
+        migraphx::add_apply_alpha_beta(*mm, {sl0, sl1}, migraphx::make_op("dot"), 1.0f, 0.0f);
+    auto sr0 = mm->add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), res);
+    auto ret = mm->add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), sr0);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = dd;
+    auto prog                     = parse_onnx("matmul_dyn_vv_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(matmul_dyn_broadcast_error)
+{
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    EXPECT(test::throws([&] { migraphx::parse_onnx("matmul_dyn_broadcast_error.onnx", options); }));
+}
+
 TEST_CASE(matmulinteger_test)
 {
     migraphx::program p;
@@ -3445,6 +3531,13 @@ TEST_CASE(matmulinteger_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(matmulinteger_dyn_error)
+{
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    EXPECT(test::throws([&] { migraphx::parse_onnx("matmulinteger_dyn_error.onnx", options); }));
+}
+
 TEST_CASE(max_test)
 {
     migraphx::program p;

test/simplify_algebra_test.cpp

@@ -2919,4 +2919,53 @@ TEST_CASE(reorder_slice_ins_deps)
     EXPECT(m == create_module());
 }
 
+TEST_CASE(dot_fusion_reshape)
+{
+    migraphx::module m1;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {2, 4096, 320}};
+        auto input = m1.add_parameter("input", s);
+
+        auto p0 = m1.add_literal(
+            migraphx::generate_literal({migraphx::shape::float_type, {2, 320, 320}}, 0));
+        auto p1 = m1.add_literal(
+            migraphx::generate_literal({migraphx::shape::float_type, {2, 320, 320}}, 1));
+
+        auto d0 = m1.add_instruction(migraphx::make_op("dot"), input, p0);
+        auto d1 = m1.add_instruction(migraphx::make_op("dot"), input, p1);
+
+        auto r0 =
+            m1.add_instruction(migraphx::make_op("reshape", {{"dims", {2, 4096, 8, 40}}}), d0);
+        m1.add_return({r0, d1});
+    };
+
+    migraphx::module m2;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {2, 4096, 320}};
+        auto input = m2.add_parameter("input", s);
+
+        auto p0 = m2.add_literal(
+            migraphx::generate_literal({migraphx::shape::float_type, {2, 320, 320}}, 0));
+        auto p1 = m2.add_literal(
+            migraphx::generate_literal({migraphx::shape::float_type, {2, 320, 320}}, 1));
+
+        auto c = m2.add_instruction(migraphx::make_op("concat", {{"axis", 2}}), p0, p1);
+        auto d = m2.add_instruction(migraphx::make_op("dot"), input, c);
+
+        auto s0 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {320}}}), d);
+        auto s1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {320}}, {"ends", {640}}}), d);
+
+        auto cont0 = m2.add_instruction(migraphx::make_op("contiguous"), s0);
+        auto r0 =
+            m2.add_instruction(migraphx::make_op("reshape", {{"dims", {2, 4096, 8, 40}}}), cont0);
+
+        m2.add_return({r0, s1});
+    };
+
+    run_pass(m1);
+    EXPECT(m1.sort() == m2.sort());
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/simplify_reshapes_test.cpp

@@ -1405,9 +1405,9 @@ TEST_CASE(transpose_slice_non_packed_axis)
     {
         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);
+            m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {3}}}), x);
         auto transpose = m2.add_instruction(
-            migraphx::make_op("transpose", {{"permutation", {3, 0, 2, 1, 4}}}), unsqueeze);
+            migraphx::make_op("transpose", {{"permutation", {2, 0, 3, 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);
@@ -1444,9 +1444,9 @@ TEST_CASE(transpose_slice_non_packed_multi_axis)
     {
         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);
+            m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {3}}}), x);
         auto transpose = m2.add_instruction(
-            migraphx::make_op("transpose", {{"permutation", {3, 0, 2, 1, 4}}}), unsqueeze);
+            migraphx::make_op("transpose", {{"permutation", {2, 0, 3, 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);

test/verify/test_trans_slice.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>
+
+struct test_trans_slice : verify_program<test_trans_slice>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+
+        auto x = mm->add_parameter("x", {migraphx::shape::float_type, {2, 384, 36, 64}});
+        auto transpose =
+            mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = mm->add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = mm->add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto transpose2 = mm->add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), slice2);
+        auto slice3 = mm->add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        mm->add_return({slice1, transpose2, slice3});
+
+        return p;
+    }
+};