Merge branch 'develop' into blas_tuning

src/targets/gpu/target.cpp

@@ -48,6 +48,7 @@
 #include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/schedule.hpp>
+#include <migraphx/simplify_dyn_ops.hpp>
 #include <migraphx/simplify_qdq.hpp>
 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/split_single_dyn_dim.hpp>
@@ -109,6 +110,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
     {
         split_single_dyn_dim{},
         dead_code_elimination{},
+        simplify_dyn_ops{},
+        dead_code_elimination{},
         normalize_ops{},
         dead_code_elimination{},
         simplify_qdq{},

src/targets/gpu/time_op.cpp

@@ -41,8 +41,7 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
 }
 
 using milliseconds = std::chrono::duration<double, std::milli>;
-std::pair<double, double>
-time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
+double time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
 {
 
     // TODO: Use std::ref
@@ -51,21 +50,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
     auto output           = op.compute_shape(inputs);
     op.finalize(ctx, output, inputs);
     auto args = generate_arguments(inputs);
-    auto run  = [&] {
-        op.compute(ctx, output, args);
-        ctx.finish();
-    };
-    gctx.enable_perf_measurement();
+    auto start = context::create_event_for_timing();
+    auto stop  = context::create_event_for_timing();
+    auto run   = [&] { op.compute(ctx, output, args); };
     run();
-    double host_time   = 0.0;
-    double device_time = 0.0;
+    gctx.get_stream().record(start.get());
     for(auto i : range(n))
     {
         (void)i;
-        host_time += time<milliseconds>(run);
-        device_time += gctx.get_elapsed_ms();
+        run();
     }
-    return std::make_pair(host_time / n, device_time / n);
+    gctx.get_stream().record(stop.get());
+    gctx.finish();
+    return context::get_elapsed_ms(start.get(), stop.get()) / n;
 }
 
 } // namespace gpu

src/targets/ref/lowering.cpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 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

src/tf/parse_reshape.cpp

@@ -45,8 +45,7 @@ struct parse_reshape : op_parser<parse_reshape>
         auto s = args[1]->eval();
         std::vector<int64_t> dims;
         s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
-        return info.add_instruction(make_op("reshape", {{"dims", dims}}),
-                                    info.make_contiguous(args[0]));
+        return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
     }
 };
 

src/verify_args.cpp

@@ -28,19 +28,20 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 bool verify_args(const std::string& name,
-                 const argument& ref_arg,
                  const argument& target_arg,
-                 double tolerance)
+                 const verify::expected<argument>& ref_arg,
+                 verify::tolerance tols)
 {
     bool passed = true;
-    visit_all(ref_arg, target_arg)([&](auto ref, auto target) {
-        double error;
-        passed = verify::verify_range(ref, target, tolerance, &error);
+    visit_all(ref_arg.data(), target_arg)([&](auto ref, auto target) {
+        double rms_error;
+        passed =
+            verify::verify_range_with_tolerance(target, verify::expected{ref}, tols, &rms_error);
         if(not passed)
         {
             // TODO: Check for nans
             std::cout << "FAILED: " << name << std::endl;
-            std::cout << "error: " << error << std::endl;
+            std::cout << "RMS Error: " << rms_error << std::endl;
             if(ref.size() < 32)
                 std::cout << "ref:" << ref << std::endl;
             if(target.size() < 32)
@@ -78,16 +79,6 @@ bool verify_args(const std::string& name,
             if(verify::range_zero(target))
                 std::cout << "Target data is all zeros" << std::endl;
 
-            // auto mxdiff = max_diff(ref, target);
-            // std::cout << "Max diff: " << mxdiff << std::endl;
-
-            // auto idx = mismatch_idx(ref, target, float_equal);
-            // if(idx < verify::range_distance(ref))
-            // {
-            //     std::cout << "Mismatch at " << idx << ": " << ref[idx] << " != " << target[idx]
-            //               << std::endl;
-            // }
-
             auto ref_nan_idx = find_idx(ref, verify::not_finite);
             if(ref_nan_idx >= 0)
                 std::cout << "Non finite number found in ref at " << ref_nan_idx << ": "
@@ -97,11 +88,22 @@ bool verify_args(const std::string& name,
             if(target_nan_idx >= 0)
                 std::cout << "Non finite number found in target at " << target_nan_idx << ": "
                           << target[target_nan_idx] << std::endl;
-            // std::cout << std::endl;
+            std::cout << "MIGraphX verification passed successfully." << std::endl;
         }
     });
     return passed;
 }
 
+bool verify_args_with_tolerance(const std::string& name,
+                                const argument& target_arg,
+                                const verify::expected<argument>& ref_arg,
+                                std::size_t tolerance)
+{
+    double rms_tol = 0.001;
+    target_arg.visit([&](auto ta) { rms_tol = verify::get_rms_tol(ta, tolerance); });
+    verify::tolerance tols{rms_tol};
+    return verify_args(name, target_arg, ref_arg, tols);
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

test/CMakeLists.txt

@@ -25,92 +25,19 @@
 cmake_policy(SET CMP0057 NEW)
 
 find_package(Threads REQUIRED)
-include(ProcessorCount)
-ProcessorCount(N)
-set(CTEST_PARALLEL_LEVEL ${N} CACHE STRING "CTest parallel level")
-add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -j ${CTEST_PARALLEL_LEVEL} -C ${CMAKE_CFG_INTDIR} --timeout 5000)
-add_custom_target(tests)
-
-find_program(MIGRAPHX_GDB gdb)
-
-if(MIGRAPHX_GDB)
-    set(MIGRAPHX_TEST_GDB On CACHE BOOL "")
-else()
-    set(MIGRAPHX_TEST_GDB Off CACHE BOOL "")
-endif()
-
-set(SKIP_TESTS)
-
-function(add_test_command NAME EXE)
-    if(NAME IN_LIST SKIP_TESTS)
-        add_test(NAME ${NAME} COMMAND echo skipped)
-        set_tests_properties(${NAME} PROPERTIES DISABLED On)
-    elseif(WIN32)
-        set(WINPATH)
-
-        foreach(PATH ${CMAKE_FIND_ROOT_PATH})
-            list(APPEND WINPATH ${PATH}/bin)
-        endforeach()
-
-        file(GENERATE OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/test_${NAME}.cmd"
-            CONTENT "set PATH=${WINPATH};%PATH%
-                    %1 ${ARGN}")
-        add_test(NAME ${NAME} COMMAND ${WINE_CMD} cmd /c "${CMAKE_CURRENT_BINARY_DIR}/test_${NAME}.cmd" $<TARGET_FILE:${EXE}>)
-    else()
-        if(MIGRAPHX_TEST_GDB)
-            # add_test(NAME ${NAME} COMMAND ${MIGRAPHX_GDB}
-            # --batch
-            # --return-child-result
-            # -ex "set disable-randomization off"
-            # -ex run
-            # -ex backtrace
-            # --args $<TARGET_FILE:${EXE}> ${ARGN})
-            set(TEST_DIR ${CMAKE_CURRENT_BINARY_DIR}/gdb/test_${NAME})
-            file(MAKE_DIRECTORY ${TEST_DIR})
-
-            if(NOT EXISTS ${TEST_DIR})
-                message(FATAL_ERROR "Failed to create test directory: ${TEST_DIR}")
-            endif()
-
-            file(GENERATE OUTPUT "${TEST_DIR}/run.cmake"
-                CONTENT "
-                # Remove previous core dump
-                file(REMOVE ${TEST_DIR}/core)
-                execute_process(COMMAND $<TARGET_FILE:${EXE}> ${ARGN} WORKING_DIRECTORY ${TEST_DIR} RESULT_VARIABLE RESULT)
-                if(NOT RESULT EQUAL 0)
-                    # TODO: check for core files based on pid when setting /proc/sys/kernel/core_uses_pid
-                    if(EXISTS ${TEST_DIR}/core)
-                        set(\$ENV{UBSAN_OPTIONS} print_stacktrace=1)
-                        set(\$ENV{ASAN_OPTIONS} print_stacktrace=1)
-                        execute_process(COMMAND ${MIGRAPHX_GDB} $<TARGET_FILE:${EXE}> ${TEST_DIR}/core -batch -ex bt)
-                    endif()
-                    message(FATAL_ERROR \"Test failed\")
-                endif()
-            ")
-            add_test(NAME ${NAME} COMMAND ${CMAKE_COMMAND} -P "${TEST_DIR}/run.cmake")
-        else()
-            add_test(NAME ${NAME} COMMAND ${EXE} ${ARGN})
-        endif()
-    endif()
+rocm_test_link_libraries(Threads::Threads migraphx migraphx_ref migraphx_onnx migraphx_tf)
+rocm_test_include_directories(include)
 
-    set_tests_properties(${NAME} PROPERTIES FAIL_REGULAR_EXPRESSION "FAILED")
-endfunction()
-
-function(add_test_executable TEST_NAME)
-    add_executable(${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
-    set(TEST_COMMAND ${TEST_NAME})
-    add_test_command(${TEST_NAME} ${TEST_COMMAND})
-    add_dependencies(tests ${TEST_NAME})
-    add_dependencies(check ${TEST_NAME})
-    target_link_libraries(${TEST_NAME} Threads::Threads migraphx migraphx_onnx migraphx_ref)
-    target_include_directories(${TEST_NAME} PUBLIC include)
-endfunction(add_test_executable)
+set(MIGRAPHX_DISABLE_LARGE_BUFFER_TESTS Off CACHE BOOL "")
+if(MIGRAPHX_DISABLE_LARGE_BUFFER_TESTS)
+    add_compile_definitions(MIGRAPHX_DISABLE_LARGE_BUFFER_TESTS)
+endif()
 
 file(GLOB TESTS CONFIGURE_DEPENDS *.cpp)
 
 foreach(TEST ${TESTS})
     get_filename_component(BASE_NAME ${TEST} NAME_WE)
-    add_test_executable(test_${BASE_NAME} ${TEST})
+    rocm_add_test_executable(test_${BASE_NAME} ${TEST})
     rocm_clang_tidy_check(test_${BASE_NAME})
 endforeach()
 
@@ -120,7 +47,7 @@ if(MIGRAPHX_ENABLE_GPU)
 
     foreach(TEST ${GPU_TESTS})
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
-        add_test_executable(test_gpu_${BASE_NAME} ${TEST})
+        rocm_add_test_executable(test_gpu_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_gpu_${BASE_NAME})
         set_tests_properties(test_gpu_${BASE_NAME} PROPERTIES
             COST 10
@@ -139,7 +66,7 @@ if(MIGRAPHX_ENABLE_FPGA)
 
     foreach(TEST ${FPGA_TESTS})
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
-        add_test_executable(test_fpga_${BASE_NAME} ${TEST})
+        rocm_add_test_executable(test_fpga_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_fpga_${BASE_NAME})
         set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES
             COST 10
@@ -161,22 +88,21 @@ foreach(ONNX_TEST ${ONNX_TESTS})
     target_link_libraries(${TEST_NAME} migraphx_onnx migraphx_ref)
     target_include_directories(${TEST_NAME} PUBLIC include)
     add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> WORKING_DIRECTORY ${TEST_ONNX_DIR})
-    add_dependencies(tests ${TEST_NAME})
-    add_dependencies(check ${TEST_NAME})
+    rocm_mark_as_test(${TEST_NAME})
 endforeach()
 
 # tf test
 set(TEST_TF_DIR ${CMAKE_CURRENT_SOURCE_DIR}/tf)
 add_executable(test_tf tf/tf_test.cpp)
+rocm_mark_as_test(test_tf)
 rocm_clang_tidy_check(test_tf)
 target_link_libraries(test_tf migraphx_tf)
 target_include_directories(test_tf PUBLIC include)
 add_test(NAME test_tf COMMAND $<TARGET_FILE:test_tf> WORKING_DIRECTORY ${TEST_TF_DIR})
-add_dependencies(tests test_tf)
-add_dependencies(check test_tf)
 
 add_subdirectory(api)
 add_subdirectory(verify)
+add_subdirectory(ref)
 
 if(MIGRAPHX_ENABLE_PYTHON)
     add_subdirectory(py)
@@ -195,8 +121,7 @@ if(MIGRAPHX_ENABLE_GPU AND MIGRAPHX_ENABLE_CPU AND MIGRAPHX_ENABLE_FPGA)
         target_link_libraries(${TEST_NAME} migraphx migraphx_onnx migraphx_tf migraphx_all_targets)
         target_include_directories(${TEST_NAME} PUBLIC include)
         add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> WORKING_DIRECTORY ${TEST_MULTI_TARGET_DIR})
-        add_dependencies(tests ${TEST_NAME})
-        add_dependencies(check ${TEST_NAME})
+        rocm_mark_as_test(${TEST_NAME})
     endforeach()
 endif()
 
@@ -213,7 +138,7 @@ int main() {}\n"
 #endif
 \n"
     )
-    add_test_executable(${NAME}
+    rocm_add_test_executable(${NAME}
         ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
         ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
     )
@@ -235,13 +160,13 @@ test_headers(migraphx ${CMAKE_SOURCE_DIR}/src/include/migraphx/*.hpp)
 test_headers(migraphx/ref ${CMAKE_SOURCE_DIR}/src/targets/ref/include/migraphx/ref/*.hpp)
 
 if(MIGRAPHX_ENABLE_GPU)
-    test_headers(migraphx/gpu ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp)
+    test_headers(migraphx/gpu HEADERS ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp DEPENDS migraphx_gpu)
 endif()
 if(MIGRAPHX_ENABLE_CPU)
-    test_headers(migraphx/cpu ${CMAKE_SOURCE_DIR}/src/targets/cpu/include/migraphx/cpu/*.hpp)
+    test_headers(migraphx/cpu HEADERS ${CMAKE_SOURCE_DIR}/src/targets/cpu/include/migraphx/cpu/*.hpp migraphx_cpu)
 endif()
 if(MIGRAPHX_ENABLE_FPGA)
-    test_headers(migraphx/fpga ${CMAKE_SOURCE_DIR}/src/targets/fpga/include/migraphx/fpga/*.hpp)
+    test_headers(migraphx/fpga HEADERS ${CMAKE_SOURCE_DIR}/src/targets/fpga/include/migraphx/fpga/*.hpp migraphx_fpga)
 endif()
 
 

test/auto_contiguous_test.cpp

@@ -158,6 +158,31 @@ TEST_CASE(two_transpose_gather)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(standard_reshape_lazy)
+{
+    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_lazy", {{"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_lazy", {{"dims", {2, 1, 12, 5}}}), ca);
+        m2.add_return({r});
+    }
+
+    EXPECT(m1 == m2);
+}
+
 TEST_CASE(standard_reshape)
 {
     migraphx::module m1;
@@ -173,8 +198,7 @@ TEST_CASE(standard_reshape)
     {
         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);
+        auto r = m2.add_instruction(migraphx::make_op("reshape", {{"dims", {2, 1, 12, 5}}}), add);
         m2.add_return({r});
     }
 

test/check_shapes_test.cpp

@@ -31,24 +31,39 @@
 
 using migraphx::shape;
 
-bool create_shapes(bool dynamic_allowed)
+void create_shapes(bool dynamic_allowed)
 {
-    try
-    {
-        shape a{shape::int64_type, {3}};
-        shape b{shape::float_type, {{3, 6}, {4, 4}}};
-        auto op = migraphx::make_op("add");
-        migraphx::check_shapes{{a, b}, op, dynamic_allowed}.has(2);
-        return true;
-    }
-    catch(...)
-    {
-        return false;
-    }
+    shape a{shape::int64_type, {3}};
+    shape b{shape::float_type, {{3, 6}, {4, 4}}};
+    migraphx::check_shapes{{a, b}, "", dynamic_allowed}.has(2);
 }
 
-TEST_CASE(allow_dynamic_shape) { EXPECT(create_shapes(true)); }
+TEST_CASE(allow_dynamic_shape)
+{
+    EXPECT(not test::throws([] { create_shapes(true); }));
+}
+
+TEST_CASE(fail_dynamic_shape)
+{
+    EXPECT(test::throws([] { create_shapes(false); }));
+}
 
-TEST_CASE(fail_dynamic_shape) { EXPECT(not create_shapes(false)); }
+TEST_CASE(same_layout_fail)
+{
+    EXPECT(test::throws([] {
+        shape a{shape::float_type, {2, 3}};
+        shape b{shape::float_type, {2, 3}, {1, 2}};
+        migraphx::check_shapes{{a, b}, ""}.same_layout();
+    }));
+}
+
+TEST_CASE(same_layout_pass)
+{
+    EXPECT(not test::throws([] {
+        shape a{shape::float_type, {2, 3}, {1, 2}};
+        shape b{shape::float_type, {2, 3}, {1, 2}};
+        migraphx::check_shapes{{a, b}, ""}.same_layout();
+    }));
+}
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/common_dims.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 <migraphx/common_dims.hpp>
+#include <test.hpp>
+
+using axes_map = std::vector<std::vector<std::size_t>>;
+
+TEST_CASE(common_d1_less)
+{
+    auto cd = migraphx::common_dims::compute({2, 32, 40, 8}, {2, 1280, 8});
+    EXPECT(cd.dims == std::vector<std::size_t>{2, 32, 40, 8});
+    EXPECT(cd.axes_map1 == axes_map{{0}, {1}, {2}, {3}});
+    EXPECT(cd.axes_map2 == axes_map{{0}, {1, 2}, {3}});
+}
+
+TEST_CASE(common1)
+{
+    auto cd = migraphx::common_dims::compute({2, 32, 2560}, {2, 1280, 8, 8});
+    EXPECT(cd.dims == std::vector<std::size_t>{2, 32, 40, 8, 8});
+    EXPECT(cd.axes_map1 == axes_map{{0}, {1}, {2, 3, 4}});
+    EXPECT(cd.axes_map2 == axes_map{{0}, {1, 2}, {3}, {4}});
+}
+
+TEST_CASE(common2)
+{
+    auto cd = migraphx::common_dims::compute({2, 1280, 8, 8}, {2, 32, 2560});
+    EXPECT(cd.dims == std::vector<std::size_t>{2, 32, 40, 8, 8});
+    EXPECT(cd.axes_map1 == axes_map{{0}, {1, 2}, {3}, {4}});
+    EXPECT(cd.axes_map2 == axes_map{{0}, {1}, {2, 3, 4}});
+}
+
+TEST_CASE(common_error1)
+{
+    auto cd = migraphx::common_dims::compute({6, 35}, {3, 7, 2, 5});
+    EXPECT(cd.dims.empty());
+}
+
+TEST_CASE(common_error2)
+{
+    auto cd = migraphx::common_dims::compute({3, 7, 2, 5}, {6, 35});
+    EXPECT(cd.dims.empty());
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/eliminate_contiguous_test.cpp

@@ -196,15 +196,47 @@ TEST_CASE(contiguous_pointwise)
             migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", {2, 3, 8, 8}}}), y);
         auto yc  = mm->add_instruction(migraphx::make_op("contiguous"), yb);
         auto add = add_pointwise(p, "main:pointwise0", {x, yc}, single_pointwise("add"));
-        mm->add_instruction(pass_op{}, add);
+        auto cadd = mm->add_instruction(migraphx::make_op("contiguous"), add);
+        mm->add_instruction(pass_op{}, cadd);
     }
     auto count = std::distance(mm->begin(), mm->end());
     run_pass(*mm);
-    EXPECT(std::distance(mm->begin(), mm->end()) == (count - 1));
+    EXPECT(std::distance(mm->begin(), mm->end()) == (count - 2));
     EXPECT(std::none_of(
         mm->begin(), mm->end(), [](auto&& ins) { return ins.name() == "contiguous"; }));
 }
 
+TEST_CASE(contiguous_nhwc_pointwise)
+{
+    auto s =
+        migraphx::shape::from_permutation(migraphx::shape::float_type, {2, 3, 8, 8}, {0, 2, 3, 1});
+    migraphx::program p1;
+    {
+        auto* mm = p1.get_main_module();
+        auto x   = mm->add_parameter("x", s);
+        auto y   = mm->add_parameter("y", migraphx::shape{migraphx::shape::float_type, {3}});
+        auto yb  = mm->add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", {2, 3, 8, 8}}}), y);
+        auto yc   = mm->add_instruction(migraphx::make_op("contiguous"), yb);
+        auto add  = add_pointwise(p1, "main:pointwise0", {x, yc}, single_pointwise("add"));
+        auto cadd = mm->add_instruction(migraphx::make_op("contiguous"), add);
+        mm->add_instruction(pass_op{}, cadd);
+    }
+    run_pass(*p1.get_main_module());
+    migraphx::program p2;
+    {
+        auto* mm = p2.get_main_module();
+        auto x   = mm->add_parameter("x", s);
+        auto y   = mm->add_parameter("y", migraphx::shape{migraphx::shape::float_type, {3}});
+        auto yb  = mm->add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", {2, 3, 8, 8}}}), y);
+        auto add  = add_pointwise(p2, "main:pointwise0", {x, yb}, single_pointwise("add"));
+        auto cadd = mm->add_instruction(migraphx::make_op("contiguous"), add);
+        mm->add_instruction(pass_op{}, cadd);
+    }
+    EXPECT(p1 == p2);
+}
+
 TEST_CASE(slice_contiguous)
 {
     migraphx::module m;

test/eliminate_pad_test.cpp

@@ -27,7 +27,7 @@
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/instruction.hpp>
 #include <basic_ops.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/common.hpp>
 #include <migraphx/make_op.hpp>
 
 #include <test.hpp>
@@ -58,9 +58,8 @@ create_conv(migraphx::instruction_ref& l_img,
     migraphx::shape s_weights{migraphx::shape::int32_type, {4, channels, 3, 3}};
     std::vector<int32_t> weights(4 * channels * 3 * 3);
     auto l_weights = m.add_literal(migraphx::literal{s_weights, weights});
-    migraphx::op::convolution op;
-    op.padding_mode = padding_mode;
-    return m.add_instruction(op, l_img, l_weights);
+    return m.add_instruction(
+        migraphx::make_op("convolution", {{"padding_mode", padding_mode}}), l_img, l_weights);
 }
 
 TEST_CASE(rewrite_pad)

test/float_equal.cpp

@@ -112,7 +112,10 @@ TEST_CASE_REGISTER(test_limits<double, int>);
 TEST_CASE_REGISTER(test_limits<double, migraphx::half>);
 TEST_CASE_REGISTER(test_limits<float, int>);
 TEST_CASE_REGISTER(test_limits<int, migraphx::half>);
+#ifndef _WIN32
+// On Windows, types int and long have the same min and max values.
 TEST_CASE_REGISTER(test_limits<long, int>);
+#endif
 TEST_CASE_REGISTER(test_limits<long, char>);
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/fuse_pointwise.cpp

@@ -21,8 +21,9 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/fuse_pointwise.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/program.hpp>
@@ -361,4 +362,154 @@ TEST_CASE(no_input)
     EXPECT(p == p2);
 }
 
+TEST_CASE(add_reshape_add)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {3, 10, 16}};
+    migraphx::shape s2{migraphx::shape::float_type, {3, 40, 2, 2}};
+    migraphx::shape s3{migraphx::shape::float_type, {3, 10, 4, 2, 2}};
+    migraphx::program p1;
+    {
+        auto* mm  = p1.get_main_module();
+        auto x    = mm->add_parameter("x", s1);
+        auto y    = mm->add_parameter("y", s1);
+        auto z    = mm->add_parameter("z", s2);
+        auto add1 = mm->add_instruction(migraphx::make_op("add"), x, y);
+        auto reshape =
+            mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), add1);
+        auto add2 = mm->add_instruction(migraphx::make_op("add"), reshape, z);
+        mm->add_return({add2});
+    }
+    run_pass(p1);
+    migraphx::program p2;
+    {
+        auto* mm = p2.get_main_module();
+        auto x   = mm->add_parameter("x", s1);
+        auto y   = mm->add_parameter("y", s1);
+        auto z   = mm->add_parameter("z", s2);
+        auto x2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), x);
+        auto y2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), y);
+        auto z2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), z);
+        auto fadd =
+            add_pointwise(p2, "main:pointwise0", {x2, y2, z2}, [=](auto* pm, const auto& inputs) {
+                auto add1 = pm->add_instruction(migraphx::make_op("add"), inputs[0], inputs[1]);
+                return pm->add_instruction(migraphx::make_op("add"), add1, inputs[2]);
+            });
+        auto reshape =
+            mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), fadd);
+        mm->add_return({reshape});
+    }
+    EXPECT(p1.sort() == p2.sort());
+}
+
+TEST_CASE(add_reshape_add_nonstandard)
+{
+    migraphx::shape s1 =
+        migraphx::shape::from_permutation(migraphx::shape::float_type, {3, 10, 16}, {2, 0, 1});
+    migraphx::shape s2{migraphx::shape::float_type, {3, 40, 2, 2}};
+    migraphx::shape s3{migraphx::shape::float_type, {3, 10, 4, 2, 2}};
+    migraphx::program p1;
+    {
+        auto* mm     = p1.get_main_module();
+        auto x       = mm->add_parameter("x", s1);
+        auto y       = mm->add_parameter("y", s1);
+        auto z       = mm->add_parameter("z", s2);
+        auto add1    = mm->add_instruction(migraphx::make_op("add"), x, y);
+        auto c       = mm->add_instruction(migraphx::make_op("contiguous"), add1);
+        auto reshape = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), c);
+        auto add2    = mm->add_instruction(migraphx::make_op("add"), reshape, z);
+        mm->add_return({add2});
+    }
+    run_pass(p1);
+    migraphx::program p2;
+    {
+        auto* mm = p2.get_main_module();
+        auto x   = mm->add_parameter("x", s1);
+        auto y   = mm->add_parameter("y", s1);
+        auto z   = mm->add_parameter("z", s2);
+        auto cx  = mm->add_instruction(migraphx::make_op("contiguous"), x);
+        auto cy  = mm->add_instruction(migraphx::make_op("contiguous"), y);
+        auto x2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), cx);
+        auto y2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), cy);
+        auto z2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s3.lens()}}), z);
+        auto fadd =
+            add_pointwise(p2, "main:pointwise0", {x2, y2, z2}, [=](auto* pm, const auto& inputs) {
+                auto add1 = pm->add_instruction(migraphx::make_op("add"), inputs[0], inputs[1]);
+                return pm->add_instruction(migraphx::make_op("add"), add1, inputs[2]);
+            });
+        auto reshape =
+            mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), fadd);
+        mm->add_return({reshape});
+    }
+    EXPECT(p1.sort() == p2.sort());
+}
+
+TEST_CASE(add_unsqueeze_add_nonstandard)
+{
+    migraphx::shape s1 =
+        migraphx::shape::from_permutation(migraphx::shape::float_type, {3, 10, 16}, {2, 0, 1});
+    migraphx::shape s2{migraphx::shape::float_type, {3, 10, 1, 16}};
+    migraphx::program p1;
+    {
+        auto* mm       = p1.get_main_module();
+        auto x         = mm->add_parameter("x", s1);
+        auto y         = mm->add_parameter("y", s1);
+        auto z         = mm->add_parameter("z", s2);
+        auto add1      = mm->add_instruction(migraphx::make_op("add"), x, y);
+        auto unsqueeze = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}}), add1);
+        auto add2      = mm->add_instruction(migraphx::make_op("add"), unsqueeze, z);
+        mm->add_return({add2});
+    }
+    run_pass(p1);
+    migraphx::program p2;
+    {
+        auto* mm = p2.get_main_module();
+        auto x   = mm->add_parameter("x", s1);
+        auto y   = mm->add_parameter("y", s1);
+        auto z   = mm->add_parameter("z", s2);
+        auto cx  = mm->add_instruction(migraphx::make_op("contiguous"), x);
+        auto cy  = mm->add_instruction(migraphx::make_op("contiguous"), y);
+        auto x2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), cx);
+        auto y2  = mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), cy);
+        auto fadd =
+            add_pointwise(p2, "main:pointwise0", {x2, y2, z}, [=](auto* pm, const auto& inputs) {
+                auto add1 = pm->add_instruction(migraphx::make_op("add"), inputs[0], inputs[1]);
+                return pm->add_instruction(migraphx::make_op("add"), add1, inputs[2]);
+            });
+        mm->add_return({fadd});
+    }
+    EXPECT(p1.sort() == p2.sort());
+}
+
+TEST_CASE(add_reshape_add_error)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {6, 35}};
+    migraphx::shape s2{migraphx::shape::float_type, {3, 7, 2, 5}};
+    migraphx::program p1;
+    {
+        auto* mm  = p1.get_main_module();
+        auto x    = mm->add_parameter("x", s1);
+        auto y    = mm->add_parameter("y", s1);
+        auto z    = mm->add_parameter("z", s2);
+        auto add1 = mm->add_instruction(migraphx::make_op("add"), x, y);
+        auto reshape =
+            mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), add1);
+        auto add2 = mm->add_instruction(migraphx::make_op("add"), reshape, z);
+        mm->add_return({add2});
+    }
+    run_pass(p1);
+    migraphx::program p2;
+    {
+        auto* mm   = p2.get_main_module();
+        auto x     = mm->add_parameter("x", s1);
+        auto y     = mm->add_parameter("y", s1);
+        auto z     = mm->add_parameter("z", s2);
+        auto fadd1 = add_pointwise(p2, "main:pointwise0", {x, y}, single_pointwise("add"));
+        auto reshape =
+            mm->add_instruction(migraphx::make_op("reshape", {{"dims", s2.lens()}}), fadd1);
+        auto fadd2 = add_pointwise(p2, "main:pointwise1", {reshape, z}, single_pointwise("add"));
+        mm->add_return({fadd2});
+    }
+    EXPECT(p1.sort() == p2.sort());
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/codegen_literal.cpp

@@ -80,7 +80,7 @@ TEST_CASE(mul_literal_round_test)
     migraphx::target gpu_t = migraphx::make_target("gpu");
     run_prog(p, gpu_t, m, gpu_result);
 
-    EXPECT(migraphx::verify::verify_range(ref_result, gpu_result));
+    EXPECT(migraphx::verify::verify_rms_range(gpu_result, ref_result));
 }
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/fuse_mlir.cpp

@@ -34,7 +34,8 @@
 
 void run_pass(migraphx::program& p)
 {
-    migraphx::run_passes(p, {migraphx::gpu::fuse_mlir{}, migraphx::dead_code_elimination{}});
+    migraphx::run_passes(
+        p, {migraphx::gpu::fuse_mlir{.enable_extra = true}, migraphx::dead_code_elimination{}});
 }
 
 template <class F>
@@ -151,7 +152,6 @@ TEST_CASE(int_quant_dot_tanh_fails)
 
 int main(int argc, const char* argv[])
 {
-    if(migraphx::gpu::mlir_enabled())
-        test::run(argc, argv);
+    test::run(argc, argv);
     return 0;
 }

test/gpu/fuse_ops.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 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 "make_precompile_op.hpp"
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/gpu/fuse_ops.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/program.hpp>
+#include <basic_ops.hpp>
+#include <migraphx/make_op.hpp>
+#include <test.hpp>
+#include <pointwise.hpp>
+
+void run_pass(migraphx::program& p)
+{
+    migraphx::run_passes(p, {migraphx::gpu::fuse_ops{}, migraphx::dead_code_elimination{}});
+}
+
+TEST_CASE(layernorm_pointwise)
+{
+    migraphx::shape s{migraphx::shape::float_type, {2, 3, 4}};
+    auto create_program = [=](bool first_arg_layernorm) {
+        migraphx::program p;
+        auto* mm       = p.get_main_module();
+        auto x         = mm->add_parameter("x", s);
+        auto y         = mm->add_parameter("y", s);
+        auto z         = mm->add_parameter("z", s);
+        auto alloc     = migraphx::make_op("allocate", {{"shape", to_value(s)}});
+        auto alloc_ins = mm->add_instruction(alloc);
+        auto* pw_add1 =
+            create_pointwise_module(p, "main:pointwise0", {x, y}, single_pointwise("add"));
+        auto add1 =
+            mm->add_instruction(make_precompile_op("pointwise"), {x, y, alloc_ins}, {pw_add1});
+        auto alloc_ins2 = mm->add_instruction(alloc);
+        auto layernorm_ins =
+            mm->add_instruction(make_precompile_op("gpu::prelayernorm"), add1, alloc_ins2);
+        std::vector<migraphx::instruction_ref> pw_inputs = {layernorm_ins, z};
+        if(not first_arg_layernorm)
+        {
+            pw_inputs = {z, layernorm_ins};
+        }
+        auto* pw_add2 =
+            create_pointwise_module(p, "main:pointwise1", pw_inputs, single_pointwise("add"));
+        auto alloc_ins3 = mm->add_instruction(alloc);
+        pw_inputs.push_back(alloc_ins3);
+        auto add2 = mm->add_instruction(make_precompile_op("pointwise"), pw_inputs, {pw_add2});
+        mm->add_return({add2});
+        return p;
+    };
+
+    auto create_fused_program = [=]() {
+        migraphx::program p;
+        auto* mm       = p.get_main_module();
+        auto x         = mm->add_parameter("x", s);
+        auto y         = mm->add_parameter("y", s);
+        auto z         = mm->add_parameter("z", s);
+        auto alloc     = migraphx::make_op("allocate", {{"shape", to_value(s)}});
+        auto alloc_ins = mm->add_instruction(alloc);
+        auto* pw_add1 =
+            create_pointwise_module(p, "main:pointwise0", {x, y}, single_pointwise("add"));
+        auto add1 =
+            mm->add_instruction(make_precompile_op("pointwise"), {x, y, alloc_ins}, {pw_add1});
+        auto alloc_ins2 = mm->add_instruction(alloc);
+        auto* pw_add2 =
+            create_pointwise_module(p, "main:pointwise1", {x, z}, single_pointwise("add"));
+        auto layernorm_ins = mm->add_instruction(
+            make_precompile_op("gpu::prelayernorm"), {add1, z, alloc_ins2}, {pw_add2});
+        mm->add_return({layernorm_ins});
+        return p;
+    };
+
+    {
+        migraphx::program p1 = create_program(true);
+        run_pass(p1);
+        migraphx::program p2 = create_fused_program();
+        EXPECT(p1 == p2);
+    }
+    {
+        migraphx::program p1 = create_program(false);
+        run_pass(p1);
+        migraphx::program p2 = create_fused_program();
+        EXPECT(p1 == p2);
+    }
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/jit.cpp

@@ -155,7 +155,7 @@ int main() {}
 
 migraphx::src_file make_src_file(const std::string& name, const std::string& content)
 {
-    return {name, std::make_pair(content.data(), content.data() + content.size())};
+    return {name, content};
 }
 
 TEST_CASE(simple_compile_hip)
@@ -218,6 +218,15 @@ TEST_CASE(compile_warnings)
 #endif
 }
 
+TEST_CASE(has_flags)
+{
+    EXPECT(migraphx::gpu::hip_has_flags({"--std=c++17"}));
+    EXPECT(not migraphx::gpu::hip_has_flags({"--non-existent-flag-to-test-in-migraphx"}));
+    EXPECT(migraphx::gpu::hip_has_flags({"-Wunused-parameter"}));
+    EXPECT(not migraphx::gpu::hip_has_flags(
+        {"-Wnon-existent-warnings-flag-to-test-in-migraphx", "-Werror"}));
+}
+
 TEST_CASE(code_object_hip)
 {
     auto binaries = migraphx::gpu::compile_hip_src(

test/gpu/manage_host_buffer.cpp

@@ -53,7 +53,6 @@ TEST_CASE(host_same_buffer_copy)
     migraphx::parameter_map pp;
     std::vector<float> a_vec(ss.elements(), -1);
     std::vector<float> b_vec(ss.elements(), 2);
-    std::vector<float> c_vec(ss.elements(), 0);
     pp["a"] = migraphx::argument(ss, a_vec.data());
     pp["b"] = migraphx::argument(ss, b_vec.data());
     std::vector<float> gpu_result;
@@ -64,7 +63,8 @@ TEST_CASE(host_same_buffer_copy)
     auto result = p.eval(pp).back();
     std::vector<float> results_vector(ss.elements(), -1);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
-    EXPECT(migraphx::verify::verify_range(c_vec, results_vector));
+    std::vector<float> gold_vec(ss.elements(), 0);
+    EXPECT(migraphx::verify::verify_rms_range(results_vector, gold_vec));
 }
 
 TEST_CASE(arguments_lifetime)

test/gpu/mlir.cpp

@@ -133,7 +133,8 @@ bool verify_mlir(const migraphx::module& mmlir)
     auto inputs = generate_params(ref);
 
     auto mlir = create_program_from_mlir(mmlir);
-    return migraphx::verify_args("mlir", run_ref(ref, inputs), run_gpu(mlir, inputs));
+    return migraphx::verify_args_with_tolerance(
+        "mlir", run_gpu(mlir, inputs), migraphx::verify::expected{run_ref(ref, inputs)});
 }
 
 TEST_CASE(conv)

test/gpu/quantization.cpp

@@ -24,7 +24,7 @@
 #include <iostream>
 #include <vector>
 #include <migraphx/gpu/fuse_mlir.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/make_op.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/quantization.hpp>
 #include <migraphx/generate.hpp>
@@ -40,7 +40,6 @@
 TEST_CASE(gpu_target_copy)
 {
     migraphx::target gpu_t = migraphx::make_target("gpu");
-    migraphx::target ref_t = migraphx::make_target("ref");
     migraphx::shape s{migraphx::shape::int8_type, {2, 3, 4, 5}};
 
     auto ref_arg_orig  = migraphx::generate_argument(s, 0x123456L);
@@ -52,7 +51,7 @@ TEST_CASE(gpu_target_copy)
     std::vector<int8_t> val_final;
     ref_arg_final.visit([&](auto v) { val_final.assign(v.begin(), v.end()); });
 
-    EXPECT(migraphx::verify::verify_range(val_orig, val_final));
+    EXPECT(migraphx::verify::verify_rms_range(val_orig, val_final));
 }
 
 TEST_CASE(int8_quantization)
@@ -90,7 +89,7 @@ TEST_CASE(int8_quantization)
         migraphx::shape sc{migraphx::shape::float_type, {5, 8}};
         auto pa = mm->add_parameter("a", sa);
         auto pb = mm->add_parameter("b", sb);
-        mm->add_instruction(migraphx::op::dot{}, pa, pb);
+        mm->add_instruction(migraphx::make_op("dot"), pa, pb);
 
         return p;
     };
@@ -118,9 +117,12 @@ TEST_CASE(int8_quantization)
         // the regular pipeline uses the rewrite_quantization in the much
         // earlier stage.
         if(migraphx::gpu::mlir_enabled())
-            EXPECT(migraphx::verify::verify_range(ref_result, gpu_result, 1e5));
+            EXPECT(migraphx::verify::verify_range_with_tolerance(
+                gpu_result,
+                migraphx::verify::expected{ref_result},
+                migraphx::verify::tolerance{0.01}));
         else
-            EXPECT(migraphx::verify::verify_range(ref_result, gpu_result));
+            EXPECT(migraphx::verify::verify_rms_range(gpu_result, ref_result));
     }
 }