Merge branch 'print_matmul_perf_flops' of...

Merge branch 'print_matmul_perf_flops' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into branch_for_ort2

src/targets/gpu/target.cpp

@@ -44,7 +44,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_SCHEDULE_PASS)
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_POINTWISE_FUSION)
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_POINTWISE_FUSION)
 
 struct id_pass
 {
@@ -93,14 +93,14 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         eliminate_common_subexpression{},
         dead_code_elimination{},
+        auto_contiguous{},
         simplify_algebra{},
         simplify_reshapes{},
         simplify_algebra{},
-        auto_contiguous{},
         simplify_reshapes{},
         propagate_constant{},
         dead_code_elimination{},
-        enable_pass(enabled(MIGRAPHX_ENABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
+        enable_pass(not enabled(MIGRAPHX_DISABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
         dead_code_elimination{},
         mlir_conv{&ctx},
         lowering{&ctx, options.offload_copy},

test/api/CMakeLists.txt

@@ -10,6 +10,7 @@ function(add_api_test TEST_NAME TEST_SRC TEST_DIR)
     add_dependencies(check ${NAME})
 endfunction()
 
+add_api_test(assign test_assign.cpp ${TEST_ONNX_DIR})
 add_api_test(compile_options test_compile_options.cpp ${TEST_ONNX_DIR})
 add_api_test(ref test_cpu.cpp ${TEST_ONNX_DIR})
 add_api_test(save_load test_save_load.cpp ${TEST_ONNX_DIR})

test/api/test_assign.cpp

+#include <migraphx/migraphx.h>
+#include <migraphx/migraphx.hpp>
+#include "test.hpp"
+
+TEST_CASE(shape_assign)
+{
+    auto s1_cpp = migraphx::shape{migraphx_shape_float_type, {1, 3}};
+    std::vector<size_t> lens{2, 3};
+
+    // handle ptr is const, workaround to construct shape using C API
+    migraphx_shape_t s2;
+    migraphx_shape_create(&s2, migraphx_shape_float_type, lens.data(), lens.size());
+    auto s2_cpp = migraphx::shape(s2, migraphx::own{});
+    CHECK(bool{s1_cpp != s2_cpp});
+    // use C++ API for assignment
+    s1_cpp.assign_to_handle(s2);
+    CHECK(bool{s1_cpp == s2_cpp});
+
+    auto s3_cpp = migraphx::shape{migraphx_shape_float_type, lens};
+    // use C API for assignment
+    migraphx_shape_assign_to(s2, s3_cpp.get_handle_ptr());
+    CHECK(bool{s2_cpp == s3_cpp});
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/auto_contiguous_test.cpp

@@ -135,4 +135,52 @@ TEST_CASE(two_transpose_gather)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(standard_reshape)
+{
+    migraphx::module m1;
+    {
+        auto data = m1.add_parameter("2x2", {migraphx::shape::float_type, {2, 3, 4, 5}});
+        auto add  = m1.add_instruction(migraphx::make_op("add"), data, data);
+        auto r = m1.add_instruction(migraphx::make_op("reshape", {{"dims", {2, 1, 12, 5}}}), add);
+        m1.add_return({r});
+    }
+    run_pass(m1);
+
+    migraphx::module m2;
+    {
+        auto data = m2.add_parameter("2x2", {migraphx::shape::float_type, {2, 3, 4, 5}});
+        auto add  = m2.add_instruction(migraphx::make_op("add"), data, data);
+        auto ca   = m2.add_instruction(migraphx::make_op("contiguous"), add);
+        auto r    = m2.add_instruction(migraphx::make_op("reshape", {{"dims", {2, 1, 12, 5}}}), ca);
+        m2.add_return({r});
+    }
+
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(dead_instruction)
+{
+    migraphx::module m1;
+    {
+        auto data = m1.add_parameter("2x2", {migraphx::shape::float_type, {2, 3, 4, 5}});
+        m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1, 3}}}), data);
+        auto r = m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1, 3}}}),
+                                    data);
+        m1.add_return({r});
+    }
+    run_pass(m1);
+
+    migraphx::module m2;
+    {
+        auto data = m2.add_parameter("2x2", {migraphx::shape::float_type, {2, 3, 4, 5}});
+        m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1, 3}}}), data);
+        auto r = m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1, 3}}}),
+                                    data);
+        auto cr = m2.add_instruction(migraphx::make_op("contiguous"), r);
+        m2.add_return({cr});
+    }
+
+    EXPECT(m1 == m2);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/eliminate_common_subexpression_test.cpp

@@ -6,6 +6,12 @@
 
 #include <test.hpp>
 
+void run_pass(migraphx::program& p)
+{
+    migraphx::run_passes(
+        p, {migraphx::eliminate_common_subexpression{}, migraphx::dead_code_elimination{}});
+}
+
 void run_pass(migraphx::module& m)
 {
     migraphx::run_passes(
@@ -142,4 +148,59 @@ TEST_CASE(cse_test_literal)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(cse_test_submodule)
+{
+    migraphx::shape si{migraphx::shape::int64_type};
+    migraphx::shape s{migraphx::shape::int64_type, {1}};
+    migraphx::shape sc{migraphx::shape::bool_type};
+
+    auto create_program = [&](bool remove_literal = false) {
+        migraphx::program p;
+        std::vector<bool> vc    = {true};
+        std::vector<int64_t> vd = {3};
+        auto* mm                = p.get_main_module();
+
+        auto in_cond = mm->add_parameter("ccond", sc);
+        auto in_val  = mm->add_parameter("val", s);
+        auto b0      = mm->add_literal(migraphx::literal(sc, vc));
+        auto b1      = b0;
+        if(not(remove_literal))
+            b1 = mm->add_literal(migraphx::literal(sc, vc));
+
+        auto* body1 = p.create_module("loop_module1");
+        body1->add_parameter("#loop_module_in_1", sc);
+        auto in_v1 = body1->add_parameter("#loop_module_in_2", s);
+        auto l1    = body1->add_literal(migraphx::literal(si, vd));
+        auto ad1   = body1->add_instruction(migraphx::make_op("add"), l1, l1);
+        auto val1  = body1->add_instruction(migraphx::make_op("add"), in_v1, ad1);
+        auto cond1 = body1->add_instruction(
+            migraphx::make_op("convert", {{"target_type", migraphx::shape::bool_type}}), b0);
+        auto cond2 = body1->add_instruction(
+            migraphx::make_op("convert", {{"target_type", migraphx::shape::bool_type}}), b1);
+        body1->add_return({cond1, cond2, val1, val1});
+
+        auto* body2 = p.create_module("loop_module2");
+        body2->add_parameter("#loop_module_in_1", sc);
+        auto in_v2 = body2->add_parameter("#loop_module_in_2", s);
+        auto l2    = body2->add_literal(migraphx::literal(si, vd));
+        auto ad2   = body2->add_instruction(migraphx::make_op("add"), l2, l2);
+        auto val2  = body2->add_instruction(migraphx::make_op("add"), in_v2, ad2);
+        auto cond3 = body2->add_instruction(
+            migraphx::make_op("convert", {{"target_type", migraphx::shape::bool_type}}), b1);
+        body2->add_return({cond3, val2, val2});
+
+        auto loop1 = mm->add_instruction(
+            migraphx::make_op("loop", {{"max_iterations", 1}}), {in_cond, in_val}, {body1});
+        auto loop2 = mm->add_instruction(
+            migraphx::make_op("loop", {{"max_iterations", 1}}), {in_cond, in_val}, {body2});
+
+        mm->add_return({loop1, loop2});
+
+        return p;
+    };
+    auto p = create_program();
+    run_pass(p);
+    EXPECT(p == create_program(true));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/fast_math.cpp

-#include <test.hpp>
-#include <migraphx/quantization.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/op/add.hpp>
-#include <migraphx/op/mul.hpp>
-#include <migraphx/op/multibroadcast.hpp>
-#include <migraphx/op/pow.hpp>
-#include <migraphx/op/tanh.hpp>
-#include <migraphx/gpu/target.hpp>
-#include <migraphx/instruction.hpp>
-
-migraphx::program create_gelu()
-{
-    migraphx::program p;
-    auto* mm                 = p.get_main_module();
-    std::vector<float> data0 = {0.044715};
-    std::vector<float> data1 = {0.797885};
-    std::vector<float> data2 = {3};
-    std::vector<float> data3 = {0.5};
-    migraphx::shape s0{migraphx::shape::float_type, {1}};
-
-    std::vector<size_t> x_dims{1, 1, 5};
-
-    auto x         = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, x_dims});
-    auto const_val = mm->add_literal(migraphx::literal{s0, data0});
-    auto sqrt_2_pi = mm->add_literal(migraphx::literal{s0, data1});
-    auto three_val = mm->add_literal(migraphx::literal{s0, data2});
-    auto half_val  = mm->add_literal(migraphx::literal{s0, data3});
-
-    auto mbcast_3         = mm->add_instruction(migraphx::op::multibroadcast{x_dims}, three_val);
-    auto pow_op           = mm->add_instruction(migraphx::op::pow{}, x, mbcast_3);
-    auto mbcast_const     = mm->add_instruction(migraphx::op::multibroadcast{x_dims}, const_val);
-    auto mul_const        = mm->add_instruction(migraphx::op::mul{}, mbcast_const, pow_op);
-    auto add_x            = mm->add_instruction(migraphx::op::add{}, x, mul_const);
-    auto mbcast_sqrt_2_pi = mm->add_instruction(migraphx::op::multibroadcast{x_dims}, sqrt_2_pi);
-    auto mul_add_x        = mm->add_instruction(migraphx::op::mul{}, mbcast_sqrt_2_pi, add_x);
-    auto tanh_op          = mm->add_instruction(migraphx::op::tanh{}, mul_add_x);
-    auto mbcast_half      = mm->add_instruction(migraphx::op::multibroadcast{x_dims}, half_val);
-    auto mul_half         = mm->add_instruction(migraphx::op::mul{}, mbcast_half, tanh_op);
-    auto add_mul_half     = mm->add_instruction(migraphx::op::add{}, mul_half, mbcast_half);
-    auto mul_x            = mm->add_instruction(migraphx::op::mul{}, x, add_mul_half);
-    mm->add_return({mul_x});
-
-    return p;
-}
-
-TEST_CASE(enable_fast_gelu)
-{
-    migraphx::program p = create_gelu();
-    p.compile(migraphx::gpu::target{});
-    CHECK(any_of(*p.get_main_module(), [&](auto&& i) { return i.name() == "gpu::gelu"; }));
-}
-
-TEST_CASE(disable_fast_gelu)
-{
-    migraphx::program p = create_gelu();
-    migraphx::compile_options options;
-    options.fast_math = false;
-    p.compile(migraphx::gpu::target{}, options);
-    CHECK(any_of(*p.get_main_module(), [&](auto&& i) { return i.name() == "gpu::gelu_new"; }));
-}
-
-int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/op_shape_test.cpp

@@ -1446,6 +1446,27 @@ TEST_CASE(test_squeeze_all)
     expect_shape(s2, migraphx::make_op("squeeze", {{"axes", {0}}}), s1);
 }
 
+TEST_CASE(test_squeeze_transpose)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 4, 1}, {4, 1, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 4}, {4, 1}};
+    expect_shape(s2, migraphx::make_op("squeeze", {{"axes", {2}}}), s1);
+}
+
+TEST_CASE(test_squeeze_multibroadcast)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 1, 4}, {0, 1, 1, 0}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 4}, {0, 1, 0}};
+    expect_shape(s2, migraphx::make_op("squeeze", {{"axes", {2}}}), s1);
+}
+
+TEST_CASE(test_squeeze_slice)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 1, 4}, {108, 36, 6, 1}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 4}, {108, 36, 1}};
+    expect_shape(s2, migraphx::make_op("squeeze", {{"axes", {2}}}), s1);
+}
+
 TEST_CASE(test_squeeze_negative_axis)
 {
     migraphx::shape s1{migraphx::shape::float_type, {4, 1, 3, 1, 3}};
@@ -1492,6 +1513,55 @@ TEST_CASE(test_unsqueeze_scalar_tensor2)
     throws_shape(migraphx::make_op("unsqueeze", {{"axes", {-2}}}), s);
 }
 
+TEST_CASE(test_unsqueeze_transpose)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 4, 3}, {12, 1, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 4, 1, 3}, {12, 1, 1, 4}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_multibroadcast)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}, {0, 1, 0}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {0, 1, 1, 0}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_slice)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}, {108, 36, 1}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {108, 36, 36, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_axis_zero)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {1, 2, 3, 4}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {0}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_axis_last)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 4, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {-1}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_multiple_axes_1)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {1, 2, 3, 4, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {0, -1}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_multiple_axes_2)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {1, 1, 2, 3, 4}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {0, 1}}}), s1);
+}
+
 TEST_CASE(transpose_shape)
 {
     migraphx::shape input{migraphx::shape::float_type, {2, 2}};

test/print_graph_test.cpp

@@ -34,12 +34,12 @@ TEST_CASE(basic_graph_test)
     EXPECT(migraphx::contains(test, "\"main:@0\"[label=\"@literal\"]"));
     EXPECT(migraphx::contains(test, "\"y\"[label=\"@param:y\"]"));
     EXPECT(migraphx::contains(test, "\"x\"[label=\"@param:x\"]"));
-    EXPECT(migraphx::contains(test, "\"main:@1\"[label=\"sum\"]"));
-    EXPECT(migraphx::contains(test, "\"main:@2\"[label=\"sum\"]"));
-    EXPECT(migraphx::contains(test, "\"x\" -> \"main:@1\""));
-    EXPECT(migraphx::contains(test, "\"y\" -> \"main:@1\""));
-    EXPECT(migraphx::contains(test, "\"main:@1\" -> \"main:@2\""));
-    EXPECT(migraphx::contains(test, "\"main:@0\" -> \"main:@2\""));
+    EXPECT(migraphx::contains(test, "\"main:@3\"[label=\"sum\"]"));
+    EXPECT(migraphx::contains(test, "\"main:@4\"[label=\"sum\"]"));
+    EXPECT(migraphx::contains(test, "\"x\" -> \"main:@3\""));
+    EXPECT(migraphx::contains(test, "\"y\" -> \"main:@3\""));
+    EXPECT(migraphx::contains(test, "\"main:@3\" -> \"main:@4\""));
+    EXPECT(migraphx::contains(test, "\"main:@0\" -> \"main:@4\""));
     EXPECT(migraphx::contains(test, "[label=\"int64_type, {1}, {0}\"]"));
 }
 

test/ref_ops_nonstd_shape_test.cpp

@@ -3,21 +3,17 @@
 #include <migraphx/literal.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/ref/target.hpp>
+#include <migraphx/generate.hpp>
 #include <migraphx/verify.hpp>
 #include <migraphx/make_op.hpp>
-#include <migraphx/auto_contiguous.hpp>
 #include <migraphx/pass_manager.hpp>
 #include "test.hpp"
 
 TEST_CASE(argmax_test_nonstd_shape)
 {
     migraphx::program p;
-    auto* mm                = p.get_main_module();
-    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
-                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
-                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
-    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
-    auto dl = mm->add_literal(migraphx::literal{data_shape, data});
+    auto* mm = p.get_main_module();
+    auto dl = mm->add_literal(migraphx::generate_literal({migraphx::shape::float_type, {2, 3, 4}}));
     auto dl_trans =
         mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 2, 0}}}), dl);
     mm->add_instruction(migraphx::make_op("argmax", {{"axis", -3}}), dl_trans);
@@ -35,12 +31,8 @@ TEST_CASE(argmax_test_nonstd_shape)
 TEST_CASE(argmin_test_nonstd_shape)
 {
     migraphx::program p;
-    auto* mm                = p.get_main_module();
-    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
-                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
-                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
-    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
-    auto dl = mm->add_literal(migraphx::literal{data_shape, data});
+    auto* mm = p.get_main_module();
+    auto dl = mm->add_literal(migraphx::generate_literal({migraphx::shape::float_type, {2, 3, 4}}));
     auto dl_trans =
         mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 2, 0}}}), dl);
     mm->add_instruction(migraphx::make_op("argmin", {{"axis", -1}}), dl_trans);
@@ -55,4 +47,125 @@ TEST_CASE(argmin_test_nonstd_shape)
     EXPECT(migraphx::verify_range(result_vec, res_gold_vec));
 }
 
+TEST_CASE(squeeze_transpose_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0 =
+        mm->add_literal(migraphx::generate_literal({migraphx::shape::float_type, {4, 1, 3, 1, 3}}));
+    auto l0_trans =
+        mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 2, 3, 0, 4}}}), l0);
+    mm->add_instruction(migraphx::make_op("squeeze"), l0_trans);
+    auto p_uncompiled = p;
+    // contiguous is required to read the values in standard shaped order
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {3, 4, 3}});
+    EXPECT(result == expected_result);
+}
+
+TEST_CASE(squeeze_multibroadcast_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0 =
+        mm->add_literal(migraphx::generate_literal({migraphx::shape::float_type, {1, 3, 1, 3}}));
+    auto l0_brcst = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {4, 1, 3, 4, 3}}}), l0);
+    mm->add_instruction(migraphx::make_op("squeeze"), l0_brcst);
+    auto p_uncompiled   = p;
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {4, 3, 4, 3}});
+    EXPECT(result == expected_result);
+}
+
+TEST_CASE(squeeze_slice_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0 =
+        mm->add_literal(migraphx::generate_literal({migraphx::shape::float_type, {1, 3, 4, 3}}));
+    auto l0_slice = mm->add_instruction(
+        migraphx::make_op("slice", {{"axes", {2}}, {"starts", {2}}, {"ends", {3}}}), l0);
+    mm->add_instruction(migraphx::make_op("squeeze"), l0_slice);
+    auto p_uncompiled   = p;
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {3, 3}});
+    EXPECT(result == expected_result);
+}
+
+TEST_CASE(unsqueeze_transpose_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s1{migraphx::shape::float_type, {4, 3, 3}};
+    auto l0 = mm->add_literal(migraphx::generate_literal(s1));
+    auto l0_trans =
+        mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1}}}), l0);
+    mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}}), l0_trans);
+    auto p_uncompiled   = p;
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {3, 4, 1, 3}});
+    EXPECT(result == expected_result);
+}
+
+TEST_CASE(unsqueeze_multibroadcast_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s1{migraphx::shape::float_type, {4, 1, 3}};
+    auto l0 = mm->add_literal(migraphx::generate_literal(s1));
+    auto l0_brcst =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {4, 4, 3, 3}}}), l0);
+    mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}}), l0_brcst);
+    auto p_uncompiled   = p;
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {4, 4, 1, 3, 3}});
+    EXPECT(result == expected_result);
+}
+
+TEST_CASE(unsqueeze_slice_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4, 4}};
+    auto l0       = mm->add_literal(migraphx::generate_literal(s1));
+    auto l0_slice = mm->add_instruction(
+        migraphx::make_op("slice", {{"axes", {3}}, {"starts", {2}}, {"ends", {3}}}), l0);
+    mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), l0_slice);
+    auto p_uncompiled   = p;
+    auto* mm_uncompiled = p_uncompiled.get_main_module();
+    mm_uncompiled->add_instruction(migraphx::make_op("contiguous"),
+                                   std::prev(mm_uncompiled->end()));
+    p.compile(migraphx::ref::target{});
+    auto result          = p.eval({}).back();
+    auto expected_result = p_uncompiled.eval({}).back();
+    EXPECT(result.get_shape() == migraphx::shape{migraphx::shape::float_type, {2, 1, 3, 4, 1}});
+    EXPECT(result == expected_result);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/simplify_reshapes_test.cpp

 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/operators.hpp>
 #include <migraphx/instruction.hpp>

test/verify/test_sub_int.cpp

+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_sub_int : verify_program<test_sub_int>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape s{migraphx::shape::float_type, {3}};
+        auto x  = mm->add_parameter("x", {migraphx::shape::int16_type, {4, 5}});
+        auto y  = mm->add_parameter("y", {migraphx::shape::int16_type, {2, 3, 4, 5}});
+        auto xb = mm->add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {2, 3, 4, 5}}}), x);
+        auto diff = mm->add_instruction(migraphx::make_op("sub"), y, xb);
+        mm->add_return({diff});
+        return p;
+    }
+};

tools/api.py

@@ -15,7 +15,7 @@ c_api_body_preamble: List[str] = []
 cpp_header_preamble: List[str] = []
 
 
-def bad_param_error(msg):
+def bad_param_error(msg: str):
     return 'throw std::runtime_error("{}")'.format(msg)
 
 
@@ -89,7 +89,7 @@ class Type:
         else:
             return t.remove_const()
 
-    def const_compatible(self, t):
+    def const_compatible(self, t: 'Type'):
         if t.is_const():
             return self.add_const()
         return self
@@ -720,6 +720,7 @@ def add_handle(name: str,
                destroy: Optional[str] = None,
                ref: Optional[bool] = None) -> None:
     opaque_type = ctype + '_t'
+    const_opaque_type = 'const_' + opaque_type
 
     def handle_wrap(p):
         t = Type(opaque_type)
@@ -747,6 +748,9 @@ def add_handle(name: str,
         add_function(destroy or ctype + '_' + 'destroy',
                      params({name: opaque_type}),
                      fname='destroy')
+        add_function(ctype + '_' + 'assign_to',
+                     params(output=opaque_type, input=const_opaque_type),
+                     invoke='*output = *input')
     add_handle_preamble()
     c_header_preamble.append(handle_typedef.substitute(locals()))
     c_api_body_preamble.append(handle_definition.substitute(locals()))

tools/generate.sh

@@ -14,4 +14,6 @@ function api {
 }
 
 api $DIR/api/migraphx.h $SRC_DIR/api/include/migraphx/migraphx.h
+echo "Finished generating header migraphx.h"
 api $DIR/api/api.cpp $SRC_DIR/api/api.cpp
+echo "Finished generating source api.cpp "