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

src/targets/gpu/fuse_ops.cpp

@@ -553,11 +553,13 @@ struct find_gemm_pointwise
 {
     auto matcher() const
     {
-        return precompile_name("pointwise")(
+        auto gemm_op   = match::name("gpu::gemm")(match::nargs(3), match::used_once()).bind("gemm");
+        auto binary_op = match::all_of(
             match::nargs(3),
             match::either_arg(0, 1)(
-                match::any_of(match::standard_shape(), match::is_constant()).bind("c"),
-                match::name("gpu::gemm")(match::nargs(3), match::used_once()).bind("gemm")));
+                match::any_of(match::standard_shape(), match::is_constant()).bind("c"), gemm_op));
+        auto unary_op = match::all_of(match::nargs(2), match::arg(0)(gemm_op));
+        return precompile_name("pointwise")(match::any_of(binary_op, unary_op));
     }
 
     // TODO: Move to matcher.hpp
@@ -589,61 +591,84 @@ struct find_gemm_pointwise
         return match::name("@return")(match::args(match::any_of(add, mul_add, add_mul)));
     }
 
+    static auto match_mul(const std::string& input)
+    {
+        auto mul = match_mul_const(match_param(input), "alpha");
+        return match::name("@return")(match::args(mul));
+    }
+
     static float get_float(instruction_ref ins) { return ins->get_literal().at<float>(); }
 
     template <class Gemm>
     static bool update_gemm(Gemm& gemm, module_ref pm, unsigned input)
     {
         auto names = pm->get_parameter_names();
-        if(names.size() != 2)
-            return false;
         std::sort(names.begin(), names.end());
-        unsigned output = input == 0 ? 1 : 0;
-        auto mr         = match::match_instruction(
-            *pm, std::prev(pm->end()), match_add(names[input], names[output]));
-        if(mr.result == pm->end())
-            return false;
-        if(contains(mr.instructions, "alpha_mul"))
+        if(names.size() == 1)
+        {
+            auto mr = match::match_instruction(*pm, std::prev(pm->end()), match_mul(names[input]));
+            if(mr.result == pm->end())
+                return false;
             gemm.alpha *= get_float(mr.instructions["alpha"]);
-        else if(contains(mr.instructions, "beta_mul"))
-            gemm.beta *= get_float(mr.instructions["beta"]);
-        else if(contains(mr.instructions, "gamma_mul"))
+            return true;
+        }
+        else if(names.size() == 2)
         {
-            gemm.alpha *= get_float(mr.instructions["gamma"]);
-            gemm.beta *= get_float(mr.instructions["gamma"]);
+            unsigned output = input == 0 ? 1 : 0;
+            auto mr         = match::match_instruction(
+                *pm, std::prev(pm->end()), match_add(names[input], names[output]));
+            if(mr.result == pm->end())
+                return false;
+            if(contains(mr.instructions, "alpha_mul"))
+                gemm.alpha *= get_float(mr.instructions["alpha"]);
+            else if(contains(mr.instructions, "beta_mul"))
+                gemm.beta *= get_float(mr.instructions["beta"]);
+            else if(contains(mr.instructions, "gamma_mul"))
+            {
+                gemm.alpha *= get_float(mr.instructions["gamma"]);
+                gemm.beta *= get_float(mr.instructions["gamma"]);
+            }
+            return true;
+        }
+        else
+        {
+            return false;
         }
-        return true;
     }
 
     void apply(module& m, const match::matcher_result& r) const
     {
         auto ins      = r.result;
         auto gemm_ins = r.instructions["gemm"];
-        auto c_ins    = r.instructions["c"];
 
         auto gemm = any_cast<rocblas_gemm<op::dot>>(gemm_ins->get_operator());
 
         // Already fused gemm
         if(not float_equal(gemm.beta, 0))
             return;
-        gemm.beta = 1;
+        if(ins->inputs().size() == 3)
+            gemm.beta = 1;
 
         if(not update_gemm(
                gemm, ins->module_inputs().front(), ins->inputs().front() == gemm_ins ? 0 : 1))
             return;
 
-        // const-fold input if not standard shape since rocblas can't handle it
-        if(not c_ins->get_shape().standard())
-        {
-            auto c = make_op("contiguous");
-            auto l = c.compute(c.compute_shape({c_ins->get_shape()}), {c_ins->eval()});
-            c_ins  = m.add_literal(l.get_shape(), l.data());
-        }
-
         auto inputs = gemm_ins->inputs();
         inputs.pop_back();
 
-        inputs.push_back(c_ins);
+        if(ins->inputs().size() == 3)
+        {
+            auto c_ins = r.instructions["c"];
+            // const-fold input if not standard shape since rocblas can't handle it
+            if(not c_ins->get_shape().standard())
+            {
+                auto c = make_op("contiguous");
+                auto l = c.compute(c.compute_shape({c_ins->get_shape()}), {c_ins->eval()});
+                c_ins  = m.add_literal(l.get_shape(), l.data());
+            }
+            inputs.push_back(c_ins);
+        }
+
         inputs.push_back(ins->inputs().back());
 
         m.replace_instruction(ins, gemm, inputs);

src/targets/gpu/include/migraphx/gpu/context.hpp

@@ -30,6 +30,7 @@
 #include <migraphx/gpu/hip.hpp>
 #include <migraphx/env.hpp>
 #include <migraphx/config.hpp>
+#include <migraphx/gpu/device_name.hpp>
 #include <unordered_map>
 #include <memory>
 
@@ -215,6 +216,10 @@ struct context
         return *current_device;
     }
 
+    bool get_exhaustive_tune_flag() const { return exhaustive_tune; }
+
+    void set_exhaustive_tune_flag(bool t) { exhaustive_tune = t; }
+
     hip_device::stream& get_stream() { return get_current_device().get_stream(); }
     hip_device::stream& get_stream(std::size_t n) { return get_current_device().get_stream(n); }
 
@@ -273,7 +278,8 @@ struct context
         auto v_streams        = v.at("streams");
         std::size_t n_streams = v_streams.without_key().to<std::size_t>();
 
-        this->current_device = std::make_shared<hip_device>(0, n_streams);
+        auto device          = get_device_id();
+        this->current_device = std::make_shared<hip_device>(device, n_streams);
     }
 
     void wait_for(any_ptr queue)
@@ -336,7 +342,8 @@ struct context
     // TODO: Make this a vector to support multiple devices
     std::shared_ptr<hip_device> current_device;
     std::vector<shared<hip_event_ptr>> events;
-    bool measure_perf = false;
+    bool exhaustive_tune = false;
+    bool measure_perf    = false;
     // for event perf timing
     shared<hip_event_ptr> start_event = nullptr;
     shared<hip_event_ptr> stop_event  = nullptr;

src/targets/gpu/include/migraphx/gpu/convolution.hpp

@@ -175,8 +175,9 @@ struct miopen_convolution
 
             auto* miopen_stream_handle = ctx.get_stream().get_miopen();
 
-            solution_ptr = find_solution(miopen_stream_handle, conv_problem.get());
-            auto status  = miopenGetSolutionWorkspaceSize(solution_ptr.get(), &workspace_size);
+            solution_ptr = find_solution(
+                miopen_stream_handle, conv_problem.get(), ctx.get_exhaustive_tune_flag());
+            auto status = miopenGetSolutionWorkspaceSize(solution_ptr.get(), &workspace_size);
             if(status != miopenStatusSuccess)
                 MIGRAPHX_THROW("MIOpen" + op.name() + " : failed to get solution's workspace size");
 
@@ -233,7 +234,7 @@ struct miopen_convolution
                                                        &perf,
                                                        workspace.implicit(),
                                                        workspace_size,
-                                                       false);
+                                                       ctx.get_exhaustive_tune_flag());
         if(status != miopenStatusSuccess)
             MIGRAPHX_THROW("MIOpen " + op.name() + " : find convolution failed");
         algo = perf.fwd_algo;

src/targets/gpu/include/migraphx/gpu/miopen.hpp

@@ -75,12 +75,19 @@ using miopen_find_options = MIGRAPHX_MANAGE_PTR(miopenFindOptions_t, miopenDestr
 using miopen_problem      = MIGRAPHX_MANAGE_PTR(miopenProblem_t, miopenDestroyProblem);
 using miopen_solution     = MIGRAPHX_MANAGE_PTR(miopenSolution_t, miopenDestroySolution);
 
-inline miopen_solution find_solution(miopenHandle_t handle, miopenProblem_t problem)
+inline miopen_solution
+find_solution(miopenHandle_t handle, miopenProblem_t problem, bool tune = false)
 {
     miopenSolution_t solution;
-    size_t found = 0;
-    auto status  = miopenFindSolutions(handle, problem, nullptr, &solution, &found, 1);
-    auto result  = miopen_solution{solution};
+    size_t found           = 0;
+    miopen_find_options fo = nullptr;
+    if(tune)
+    {
+        fo = make_obj<miopen_find_options>(&miopenCreateFindOptions);
+        miopenSetFindOptionTuning(fo.get(), 1);
+    }
+    auto status = miopenFindSolutions(handle, problem, fo.get(), &solution, &found, 1);
+    auto result = miopen_solution{solution};
     if(status != miopenStatusSuccess or found == 0)
         MIGRAPHX_THROW("MIOpen miopenFindSolutions failed");
     return result;

src/targets/gpu/include/migraphx/gpu/prefuse_ops.hpp

@@ -30,14 +30,14 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-struct module;
+struct module_pass_manager;
 
 namespace gpu {
 
 struct prefuse_ops
 {
     std::string name() const { return "gpu::prefuse_ops"; }
-    void apply(module& m) const;
+    void apply(module_pass_manager& mpm) const;
 };
 
 } // namespace gpu

src/targets/gpu/kernels/include/migraphx/kernels/array.hpp

@@ -105,7 +105,7 @@ constexpr auto array_for_each(T& x, Ts&... xs)
             }
             else
             {
-                using vec_type = std::remove_reference_t<decltype(array2vec(x))>;
+                using vec_type = remove_reference_t<decltype(array2vec(x))>;
                 f(array2vec(x), __builtin_convertvector(array2vec(xs), vec_type)...);
             }
         }

src/targets/gpu/kernels/include/migraphx/kernels/dpp.hpp

@@ -72,7 +72,7 @@ __device__ T dpp_mov(T& x)
     }
     return output.data;
 }
-#endif
+#endif // MIGRAPHX_HAS_DPP
 
 } // namespace migraphx
 #endif // MIGRAPHX_GUARD_KERNELS_DPP_HPP

src/targets/gpu/kernels/include/migraphx/kernels/gathernd.hpp

@@ -26,7 +26,7 @@
 
 #include <migraphx/kernels/index.hpp>
 #include <migraphx/kernels/algorithm.hpp>
-
+#include <migraphx/kernels/ops.hpp>
 namespace migraphx {
 
 template <class T>
@@ -53,23 +53,17 @@ __device__ void gathernd(const T& data_t, const U& indices_t, const V& output_t,
     auto indices_shape_lens = indices_shape.lens;
     auto data_shape_lens    = data_shape.lens;
     auto num_slice_dims     = indices_shape_lens.back();
-    std::size_t num_slices  = accumulate(indices_shape_lens.begin(),
-                                        indices_shape_lens.end() - 1,
-                                        1,
-                                        std::multiplies<std::size_t>());
-    std::size_t slice_size  = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
+    std::size_t num_slices =
+        accumulate(indices_shape_lens.begin(), indices_shape_lens.end() - 1, 1, op::product{});
+    std::size_t slice_size = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
                                         data_shape_lens.end(),
                                         1,
-                                        std::multiplies<std::size_t>());
-    const std::size_t num_batches       = accumulate(data_shape_lens.begin(),
-                                               data_shape_lens.begin() + batch_dims,
-                                               1,
-                                               std::multiplies<std::size_t>());
-    const std::size_t data_batch_stride = accumulate(data_shape_lens.begin() + batch_dims,
-                                                     data_shape_lens.end(),
-                                                     1,
-                                                     std::multiplies<std::size_t>());
-    const auto num_slices_per_batch     = num_slices / num_batches;
+                                        op::product{});
+    const std::size_t num_batches =
+        accumulate(data_shape_lens.begin(), data_shape_lens.begin() + batch_dims, 1, op::product{});
+    const std::size_t data_batch_stride =
+        accumulate(data_shape_lens.begin() + batch_dims, data_shape_lens.end(), 1, op::product{});
+    const auto num_slices_per_batch = num_slices / num_batches;
 
     ind.global_stride(output_shape.elements(), [&](auto i) {
         const auto* indices_ptr     = indices_t.data();
@@ -83,15 +77,15 @@ __device__ void gathernd(const T& data_t, const U& indices_t, const V& output_t,
             int64_t index                   = slice_indices[idx];
             const std::size_t input_dim_idx = batch_dims + idx;
             const auto input_dim            = data_shape_lens[input_dim_idx];
-            assert(index >= -static_cast<int64_t>(input_dim) and
-                   index < static_cast<int64_t>(input_dim));
+            MIGRAPHX_ASSERT(index >= -static_cast<int64_t>(input_dim) and
+                            index < static_cast<int64_t>(input_dim));
             if(index < 0)
                 index += input_dim;
             std::size_t size_from_slice_dims =
                 accumulate(data_shape_lens.begin() + batch_dims + idx + 1,
                            data_shape_lens.begin() + batch_dims + num_slice_dims,
                            slice_size,
-                           std::multiplies<std::size_t>());
+                           op::product{});
             relative_slice_offset += index * size_from_slice_dims;
         }
 

src/targets/gpu/kernels/include/migraphx/kernels/hip.hpp

@@ -24,11 +24,14 @@
 #ifndef MIGRAPHX_GUARD_KERNELS_HIP_HPP
 #define MIGRAPHX_GUARD_KERNELS_HIP_HPP
 
-// Workaround macro redefinition issue with clang tidy
-#if defined(__HIP_PLATFORM_HCC__) && defined(MIGRAPHX_USE_CLANG_TIDY)
-#undef __HIP_PLATFORM_HCC__ // NOLINT
-#endif
-
+#ifndef MIGRAPHX_USE_HIPRTC
 #include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <hip/math_functions.h>
+#include <hip/hip_math_constants.h>
+#elif defined(MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS)
+#include <hip/hip_common.h>
+#include <hip/hip_math_constants.h>
+#endif
 
 #endif // MIGRAPHX_GUARD_KERNELS_HIP_HPP

src/targets/gpu/kernels/include/migraphx/kernels/math.hpp

@@ -28,8 +28,7 @@
 #include <migraphx/kernels/vec.hpp>
 #include <migraphx/kernels/functional.hpp>
 #include <migraphx/kernels/type_traits.hpp>
-#include <hip/hip_fp16.h>
-#include <hip/math_functions.h>
+#include <migraphx/kernels/hip.hpp>
 
 namespace migraphx {
 
@@ -222,7 +221,7 @@ constexpr auto min(const T& a, const U& b)
 template <class T, MIGRAPHX_REQUIRES(is_same<vec_type<T>, half>{})>
 constexpr T sin(T x)
 {
-    constexpr const T shift = M_PI_2;
+    constexpr const T shift = HIP_PIO2_F;
     return migraphx::cos(shift - x);
 }
 

src/targets/gpu/kernels/include/migraphx/kernels/shape.hpp

@@ -76,14 +76,6 @@ struct shape
 
     constexpr index_int index(index_array x) const { return x.dot(strides); }
 
-    constexpr index_int index(std::initializer_list<index_int> x) const
-    {
-        index_int idx = 0;
-        for(index_int i = 0; i < x.size(); i++)
-            idx += *(x.begin() + i) * strides[i];
-        return idx;
-    }
-
     constexpr index_int index(index_int i) const
     {
         if(this->standard())

src/targets/gpu/kernels/include/migraphx/kernels/types.hpp

@@ -28,8 +28,45 @@
 
 namespace migraphx {
 
-using index_int = std::uint32_t;
-using diff_int  = std::int32_t;
+#if defined(MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS) and defined(MIGRAPHX_USE_HIPRTC)
+using int8_t   = signed char;
+using uint8_t  = unsigned char;
+using int16_t  = signed short;
+using uint16_t = unsigned short;
+using int32_t  = signed int;
+using uint32_t = unsigned int;
+using int64_t  = signed long long;
+using uint64_t = unsigned long long;
+#elif defined(MIGRAPHX_USE_HIPRTC)
+using int8_t   = __hip_int8_t;
+using uint8_t  = __hip_uint8_t;
+using int16_t  = __hip_int16_t;
+using uint16_t = __hip_uint16_t;
+using int32_t  = __hip_int32_t;
+using uint32_t = __hip_uint32_t;
+using int64_t  = __hip_int64_t;
+using uint64_t = __hip_uint64_t;
+#else
+using int8_t   = std::int8_t;
+using uint8_t  = std::uint8_t;
+using int16_t  = std::int16_t;
+using uint16_t = std::uint16_t;
+using int32_t  = std::int32_t;
+using uint32_t = std::uint32_t;
+using int64_t  = std::int64_t;
+using uint64_t = std::uint64_t;
+#endif // MIGRAPHX_USE_HIPRTC
+using index_int = uint32_t;
+using diff_int  = int32_t;
+
+static_assert(sizeof(int8_t) == 1, "int8_t must be 1 bytes");
+static_assert(sizeof(uint8_t) == 1, "uint8_t must be 1 bytes");
+static_assert(sizeof(int16_t) == 2, "int16_t must be 2 bytes");
+static_assert(sizeof(uint16_t) == 2, "uint16_t must be 2 bytes");
+static_assert(sizeof(int32_t) == 4, "int32_t must be 4 bytes");
+static_assert(sizeof(uint32_t) == 4, "uint32_t must be 4 bytes");
+static_assert(sizeof(int64_t) == 8, "int64_t must be 8 bytes");
+static_assert(sizeof(uint64_t) == 8, "uint64_t must be 8 bytes");
 
 #define MIGRAPHX_DEVICE_CONSTEXPR constexpr __device__ __host__ // NOLINT
 

src/targets/gpu/lowering.cpp

@@ -83,8 +83,7 @@ struct miopen_apply
         auto& ctx      = get_context();
         int8_x4_format = get_int8_x4_format(ctx);
         compute_fp32   = get_compute_fp32_flag();
-
-        offload_copy = (mod->name() == "main") ? pass->offload_copy : false;
+        offload_copy   = (mod->name() == "main") ? pass->offload_copy : false;
 
         add_generic_op("contiguous");
 

src/targets/gpu/prefuse_ops.cpp

@@ -26,6 +26,8 @@
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/register_op.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/dead_code_elimination.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -90,7 +92,9 @@ struct find_layernorm
     {
         auto ins   = r.result;
         auto x_ins = r.instructions["x"];
-        auto eps   = r.instructions["eps"]->eval().at<float>();
+        float eps  = 0;
+        if(contains(r.instructions, "eps"))
+            eps = r.instructions["eps"]->eval().at<float>();
 
         m.replace_instruction(ins, layernorm{eps}, x_ins);
     }
@@ -100,24 +104,26 @@ struct find_add_layernorm
 {
     auto matcher() const
     {
-        return match::layernorm()(
-            match::var("x")(match::name("add")(match::used_once()).bind("add")));
+        return match::name("gpu::prelayernorm")(
+            match::args(match::name("add")(match::used_once()).bind("add")));
     }
 
     void apply(module& m, const match::matcher_result& r) const
     {
         auto ins     = r.result;
         auto add_ins = r.instructions["add"];
-        auto eps     = r.instructions["eps"]->eval().at<float>();
+        auto op      = any_cast<layernorm>(ins->get_operator());
 
-        m.replace_instruction(ins, add_layernorm{eps}, add_ins->inputs());
+        m.replace_instruction(ins, add_layernorm{op.epsilon}, add_ins->inputs());
     }
 };
 } // namespace
 
-void prefuse_ops::apply(module& m) const
+void prefuse_ops::apply(module_pass_manager& mpm) const
 {
-    match::find_matches(m, find_add_layernorm{}, find_layernorm{});
+    match::find_matches(mpm.get_module(), find_layernorm{});
+    mpm.run_pass(dead_code_elimination{});
+    match::find_matches(mpm.get_module(), find_add_layernorm{});
 }
 
 } // namespace gpu

src/targets/gpu/target.cpp

@@ -38,6 +38,7 @@
 #include <migraphx/layout_nhwc.hpp>
 #include <migraphx/memory_coloring.hpp>
 #include <migraphx/normalize_ops.hpp>
+#include <migraphx/optimize_module.hpp>
 #include <migraphx/preallocate_param.hpp>
 #include <migraphx/propagate_constant.hpp>
 #include <migraphx/register_target.hpp>
@@ -90,6 +91,7 @@ pass enable_pass(bool enabled, pass p)
 std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_options& options) const
 {
     auto& ctx = any_cast<context>(gctx);
+    ctx.set_exhaustive_tune_flag(options.exhaustive_tune);
     std::set<shape::type_t> unsupported_types(shape::types().begin(), shape::types().end());
     unsupported_types.erase(shape::type_t::float_type);
     unsupported_types.erase(shape::type_t::half_type);
@@ -118,21 +120,13 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         rewrite_pooling{},
         dead_code_elimination{},
         rewrite_gelu{},
-        dead_code_elimination{},
-        eliminate_common_subexpression{},
-        dead_code_elimination{},
-        simplify_algebra{},
-        simplify_reshapes{},
+        optimize_module{},
         enable_pass(enabled(MIGRAPHX_ENABLE_NHWC{}), layout_nhwc{}),
         dead_code_elimination{},
-        simplify_reshapes{},
-        simplify_algebra{},
         prefuse_ops{},
         dead_code_elimination{},
         auto_contiguous{},
-        simplify_reshapes{},
-        propagate_constant{},
-        dead_code_elimination{},
+        optimize_module{},
         enable_pass(not enabled(MIGRAPHX_DISABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
         dead_code_elimination{},
         fuse_mlir{&ctx},

test/CMakeLists.txt

-#####################################################################################
+# ####################################################################################
 # The MIT License (MIT)
 #
 # Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
@@ -20,7 +20,7 @@
 # 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.
-#####################################################################################
+# ####################################################################################
 
 cmake_policy(SET CMP0057 NEW)
 
@@ -49,27 +49,31 @@ function(add_test_command NAME EXE)
         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})
+            # 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})
+
+            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
@@ -90,22 +94,27 @@ function(add_test_command NAME EXE)
             add_test(NAME ${NAME} COMMAND ${EXE} ${ARGN})
         endif()
     endif()
+
     set_tests_properties(${NAME} PROPERTIES FAIL_REGULAR_EXPRESSION "FAILED")
 endfunction()
 
 function(add_test_executable TEST_NAME)
-    add_executable (${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
+    add_executable(${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
     target_link_libraries(${TEST_NAME} ${CMAKE_THREAD_LIBS_INIT})
+
     # Cmake does not add flags correctly for gcc
-    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU") 
+    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU")
         set_target_properties(${TEST_NAME} PROPERTIES COMPILE_FLAGS -pthread LINK_FLAGS -pthread)
     endif()
+
     separate_arguments(MIOPEN_TEST_FLAGS_ARGS UNIX_COMMAND ${MIOPEN_TEST_FLAGS})
+
     if(MIOPEN_TEST_ALL)
         set(TEST_COMMAND ${TEST_NAME} ${MIOPEN_TEST_FLOAT_ARG} --all ${MIOPEN_TEST_FLAGS_ARGS})
     else()
         set(TEST_COMMAND ${TEST_NAME} ${MIOPEN_TEST_FLOAT_ARG} ${MIOPEN_TEST_FLAGS_ARGS})
     endif()
+
     add_test_command(${TEST_NAME} ${TEST_COMMAND})
     add_dependencies(tests ${TEST_NAME})
     add_dependencies(check ${TEST_NAME})
@@ -129,11 +138,11 @@ if(MIGRAPHX_ENABLE_GPU)
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
         add_test_executable(test_gpu_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_gpu_${BASE_NAME})
-        set_tests_properties(test_gpu_${BASE_NAME} PROPERTIES 
-            COST 10 
+        set_tests_properties(test_gpu_${BASE_NAME} PROPERTIES
+            COST 10
             RESOURCE_LOCK gpu
         )
-        target_link_libraries(test_gpu_${BASE_NAME} migraphx_gpu)
+        target_link_libraries(test_gpu_${BASE_NAME} migraphx_gpu migraphx_kernels)
     endforeach()
 endif()
 
@@ -145,8 +154,8 @@ if(MIGRAPHX_ENABLE_FPGA)
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
         add_test_executable(test_fpga_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_fpga_${BASE_NAME})
-        set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES 
-            COST 10 
+        set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES
+            COST 10
             RESOURCE_LOCK fpga
         )
         target_link_libraries(test_fpga_${BASE_NAME} migraphx_fpga)
@@ -155,7 +164,8 @@ endif()
 
 # Onnx test
 set(TEST_ONNX_DIR ${CMAKE_CURRENT_SOURCE_DIR}/onnx)
-file (GLOB ONNX_TESTS ${TEST_ONNX_DIR}/*.cpp)
+file(GLOB ONNX_TESTS ${TEST_ONNX_DIR}/*.cpp)
+
 foreach(ONNX_TEST ${ONNX_TESTS})
     get_filename_component(BASE_NAME ${ONNX_TEST} NAME_WE)
     set(TEST_NAME test_${BASE_NAME})
@@ -163,7 +173,7 @@ foreach(ONNX_TEST ${ONNX_TESTS})
     rocm_clang_tidy_check(${TEST_NAME})
     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_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> WORKING_DIRECTORY ${TEST_ONNX_DIR})
     add_dependencies(tests ${TEST_NAME})
     add_dependencies(check ${TEST_NAME})
 endforeach()
@@ -174,26 +184,26 @@ add_executable(test_tf tf/tf_test.cpp)
 rocm_clang_tidy_check(test_tf)
 target_link_libraries(test_tf migraphx_tf migraphx_ref)
 target_include_directories(test_tf PUBLIC include)
-add_test(NAME test_tf COMMAND $<TARGET_FILE:test_tf> WORKING_DIRECTORY ${TEST_TF_DIR}) 
+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)
+
 if(MIGRAPHX_ENABLE_PYTHON)
-add_subdirectory(py)
+    add_subdirectory(py)
 endif()
 
 function(test_header NAME HEADER)
-
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp 
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
         "#include <${HEADER}>\nint main() {}\n"
     )
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp 
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
         "#include <${HEADER}>\n"
     )
     add_test_executable(${NAME}
-        ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp 
+        ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
         ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
     )
 endfunction()
@@ -206,6 +216,7 @@ function(test_headers PREFIX)
         string(MAKE_C_IDENTIFIER ${HEADER_REL} TEST_NAME)
         get_filename_component(BASE_NAME ${HEADER} NAME_WE)
         test_header(header_${TEST_NAME} ${PREFIX}/${BASE_NAME}.hpp)
+
         if(MIGRAPHX_ENABLE_GPU)
             target_link_libraries(header_${TEST_NAME} migraphx_gpu)
         endif()
@@ -214,6 +225,7 @@ endfunction()
 
 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 ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp)
 endif()

test/api/test_gpu.cpp

@@ -35,6 +35,7 @@ TEST_CASE(load_and_run)
     auto shapes_before = p.get_output_shapes();
     migraphx::compile_options options;
     options.set_offload_copy();
+    options.set_exhaustive_tune_flag();
     p.compile(migraphx::target("gpu"), options);
     auto shapes_after = p.get_output_shapes();
     CHECK(shapes_before.size() == 1);

test/api/test_save_load.cpp

@@ -30,7 +30,6 @@ TEST_CASE(load_save_default)
     std::string filename = "migraphx_api_load_save.mxr";
     auto p1              = migraphx::parse_onnx("conv_relu_maxpool_test.onnx");
     auto s1              = p1.get_output_shapes();
-
     migraphx::save(p1, filename.c_str());
     auto p2 = migraphx::load(filename.c_str());
     auto s2 = p2.get_output_shapes();

test/gpu/jit.cpp

@@ -35,13 +35,14 @@
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compiler.hpp>
+#include <migraphx_kernels.hpp>
 
 // NOLINTNEXTLINE
 const std::string write_2s = R"__migraphx__(
 #include <hip/hip_runtime.h>
 
 extern "C" {
-__global__ void write(int8_t* data) 
+__global__ void write(char* data) 
 {
     int num = threadIdx.x + blockDim.x * blockIdx.x;
     data[num] = 2;
@@ -58,7 +59,7 @@ const std::string add_2s_binary = R"__migraphx__(
 #include <hip/hip_runtime.h>
 
 extern "C" {
-__global__ void add_2(std::int8_t* x, std::int8_t* y) 
+__global__ void add_2(char* x, char* y) 
 {
     int num = threadIdx.x + blockDim.x * blockIdx.x;
     y[num] = x[num] + 2;
@@ -137,7 +138,8 @@ int main() {}
 const std::string math_template = R"__migraphx__(
 #include <migraphx/kernels/pointwise.hpp>
 #include <migraphx/kernels/math.hpp>
-
+#include <migraphx/kernels/types.hpp>
+using namespace migraphx;
 extern "C" {
 __global__ void kernel(${type}* p) 
 {

test/memory_coloring_test.cpp

@@ -691,7 +691,7 @@ TEST_CASE(test38)
     auto p83    = m.add_instruction(pass_op{}, p78, p77);
     m.add_instruction(pass_op{}, output, p83, p63);
     run_pass(m);
-    CHECK(m.get_parameter_shape("scratch").bytes() == 7225344); // Optimal solution is 6422528
+    CHECK(m.get_parameter_shape("scratch").bytes() == 6422528);
     CHECK(no_allocate(m));
 }
 
@@ -729,7 +729,7 @@ TEST_CASE(test39)
         run_pass(*smod);
     }
 
-    CHECK(mm->get_parameter_shape("scratch").bytes() == 4);
+    CHECK(mm->get_parameter_shape("scratch").bytes() == 1);
     CHECK(then_mod->get_parameter_shape("scratch").bytes() == 24);
     CHECK(else_mod->get_parameter_shape("scratch").bytes() == 24);
     CHECK(no_allocate(*mm));
@@ -3374,7 +3374,7 @@ TEST_CASE(rnn_dom)
     m.add_instruction(pass_op{}, moutput, mx250, mx249, mx248);
 
     run_pass(m);
-    CHECK(m.get_parameter_shape("scratch").bytes() == 1600);
+    CHECK(m.get_parameter_shape("scratch").bytes() == 1824); // Optimal is 1600
     CHECK(no_allocate(m));
     CHECK(is_disjoint({mx0, mx8}));
     CHECK(is_disjoint({mx0, mx8}));
@@ -3790,4 +3790,23 @@ TEST_CASE(literal_test)
     CHECK(lit == result);
 }
 
+TEST_CASE(test_tuple)
+{
+    migraphx::module m;
+
+    auto s1 = migraphx::shape{migraphx::shape::float_type, {8}};
+    auto s2 = migraphx::shape{migraphx::shape::half_type, {10}};
+
+    auto s = migraphx::shape{{s1, s2}};
+
+    auto a1 = add_alloc(m, s);
+    auto m1 = m.add_instruction(pass_op{}, a1);
+    auto a2 = add_alloc(m, {migraphx::shape::float_type, {4}});
+    m.add_instruction(pass_op{}, a2, m1);
+    run_pass(m);
+    CHECK(m.get_parameter_shape("scratch").bytes() == 68);
+    CHECK(no_allocate(m));
+    CHECK(is_disjoint({a1, a2}));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }