Merge branch 'trace-eliminate-contiguous' of...

Merge branch 'trace-eliminate-contiguous' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into nhwc_workaround

src/targets/cpu/include/migraphx/cpu/dnnl.hpp

@@ -400,7 +400,11 @@ struct dnnl_extend_op : dnnl_op<Derived, Primitive>
     }
 
     // dnnl has some issues with non-packed inputs
-    void required(const check_shapes& cs) const { cs.packed_or_broadcasted(); }
+    template <class T>
+    void required(const check_shapes<T>& cs) const
+    {
+        cs.packed_or_broadcasted();
+    }
 
     std::string name() const { return "dnnl::" + op.name(); }
     shape compute_shape(std::vector<shape> inputs) const

src/targets/cpu/target.cpp

@@ -61,7 +61,7 @@ namespace cpu {
 
 std::string target::name() const { return "cpu"; }
 
-// cppcheck-suppress constParameter
+// cppcheck-suppress constParameterReference
 std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_options&) const
 {
     auto& ctx = any_cast<context>(gctx);

src/targets/gpu/compile_gen.cpp

@@ -331,7 +331,7 @@ static std::vector<std::string> get_op_names(const module& m)
     {
         if(starts_with(ins.name(), "@"))
             continue;
-        if(ins.name() == "multibroadcast")
+        if(contains({"multibroadcast", "contiguous"}, ins.name()))
             continue;
         if(ins.name() == "pointwise")
         {

src/targets/gpu/device/include/migraphx/gpu/device/launch.hpp

@@ -41,7 +41,7 @@ struct index
 
     __device__ index_int nglobal() const { return blockDim.x * gridDim.x; } // NOLINT
 
-    __device__ index_int nlocal() const { return blockDim.x; } // NOLINT
+    __device__ index_int nlocal() const { return blockDim.x; }              // NOLINT
 
     template <class F>
     __device__ void global_stride(index_int n, F f) const
@@ -81,6 +81,12 @@ inline auto launch(hipStream_t stream, index_int global, index_int local)
         dim3 nthreads(local);
         // cppcheck-suppress UseDeviceLaunch
         hipLaunchKernelGGL((launcher<f_type>), nblocks, nthreads, 0, stream, f);
+        hipError_t kernel_launch_status = hipGetLastError();
+        if(kernel_launch_status != hipSuccess)
+        {
+            MIGRAPHX_THROW("MIGraphX device kernel failed to launch with error: " +
+                           std::string(hipGetErrorString(kernel_launch_status)));
+        }
     };
 }
 

src/targets/gpu/device/include/migraphx/gpu/device/nary.hpp

@@ -124,7 +124,7 @@ void nary_broadcast_vec_impl(
                     buffer[i] = binput.data()[i];
                 }
                 __syncthreads();
-                auto* bp = as_pointer(buffer);
+                const auto* bp = as_pointer(buffer);
                 // Process the data
                 for(size_t i = idx.global; i < nelements; i += nglobal)
                 {
@@ -219,7 +219,7 @@ void nary_double_broadcast_vec_impl(
                     buffer[i + bdim_vec_len] = binput2.data()[i];
                 }
                 __syncthreads();
-                auto* bp = as_pointer(buffer);
+                const auto* bp = as_pointer(buffer);
                 // Process the data
                 for(size_t i = idx.global; i < nelements; i += nglobal)
                 {

src/targets/gpu/device/topk.cpp

@@ -72,12 +72,12 @@ struct hip_heap_vector
             index_int l    = 2 * index + 1;
             index_int r    = 2 * index + 2;
 
-            if(l < n && compare(data[data_index(l)], data[data_index(index)]))
+            if(l < n and compare(data[data_index(l)], data[data_index(index)]))
             {
                 index = l;
             }
 
-            if(r < n && compare(data[data_index(r)], data[data_index(index)]))
+            if(r < n and compare(data[data_index(r)], data[data_index(index)]))
             {
                 index = r;
                 if(compare(data[data_index(l)], data[data_index(r)]))

src/targets/gpu/device_name.cpp

@@ -31,20 +31,6 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-template <class HipDeviceProp>
-std::string get_arch_name(rank<0>, const HipDeviceProp& props)
-{
-    return "gfx" + std::to_string(props.gcnArch);
-}
-
-template <class HipDeviceProp>
-auto get_arch_name(rank<1>, const HipDeviceProp& props) -> decltype(std::string(props.gcnArchName))
-{
-    return std::string(props.gcnArchName);
-}
-
-std::string get_arch_name(const hipDeviceProp_t& props) { return get_arch_name(rank<1>{}, props); }
-
 int get_device_id()
 {
     int device;
@@ -60,7 +46,7 @@ std::string get_device_name()
     auto status = hipGetDeviceProperties(&props, get_device_id());
     if(status != hipSuccess)
         MIGRAPHX_THROW("Failed to get device properties");
-    return get_arch_name(props);
+    return props.gcnArchName;
 }
 
 } // namespace gpu

src/targets/gpu/fuse_mlir.cpp

@@ -86,7 +86,7 @@ struct mlir_op
         size_t param_cnt               = 0;
         std::vector<std::string> names = mod->get_parameter_names();
         std::sort(names.begin(), names.end());
-        for(std::string param_name : names)
+        for(const std::string& param_name : names)
         {
             ins_shapes[mod->get_parameter(param_name)] = inputs[param_cnt++];
         }
@@ -211,42 +211,47 @@ struct find_mlir_op
             return false;
         }
         const std::initializer_list<std::string> any_type_ops = {"@literal", "@param", "@return"};
-        const std::initializer_list<std::string> no_bool_ops  = {"convolution",
-                                                                "quant_convolution",
-                                                                "dot",
-                                                                "quant_dot",
-                                                                "add",
-                                                                "clip",
-                                                                "relu",
-                                                                "sub",
-                                                                "mul",
-                                                                "div",
-                                                                "pow",
-                                                                "where",
-                                                                "quantizelinear",
-                                                                "dequantizelinear",
-                                                                "abs",
-                                                                "neg"};
-        const std::initializer_list<std::string> fp_only_ops  = {"ceil",
-                                                                "erf",
-                                                                "exp",
-                                                                "floor",
-                                                                "log",
-                                                                "recip",
-                                                                "rsqrt",
-                                                                "sigmoid"
-                                                                "softmax",
-                                                                "tanh"};
+        const std::initializer_list<std::string> no_bool_ops  = {
+            "convolution",
+            "quant_convolution",
+            "dot",
+            "quant_dot",
+            "add",
+            "clip",
+            "relu",
+            "sub",
+            "mul",
+            "div",
+            "pow",
+            "where",
+            "quantizelinear",
+            "dequantizelinear",
+            "abs",
+            "neg",
+        };
+        const std::initializer_list<std::string> fp_only_ops = {
+            "ceil",
+            "erf",
+            "exp",
+            "floor",
+            "log",
+            "recip",
+            "rsqrt",
+            // There are bugs in MLIR right now for models using sigmoid so disable it for now
+            // "sigmoid",
+            "softmax",
+            "tanh",
+        };
         bool is_float = contains({type_t::float_type, type_t::half_type}, result_type);
         if(contains(any_type_ops, name))
             return true;
-        if(result_type != type_t::bool_type && contains(no_bool_ops, name))
+        if(result_type != type_t::bool_type and contains(no_bool_ops, name))
             return true;
-        if(is_float && contains(fp_only_ops, name))
+        if(is_float and contains(fp_only_ops, name))
             return true;
         // Only conversions between floating types are known to be unambigiously
         // supported.
-        if(is_float && name == "convert")
+        if(is_float and name == "convert")
         {
             return std::all_of(i.inputs().begin(), i.inputs().end(), [](const auto& arg) {
                 return contains({type_t::float_type, type_t::half_type}, arg->get_shape().type());

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

@@ -46,13 +46,7 @@ using hip_event_ptr = MIGRAPHX_MANAGE_PTR(hipEvent_t, hipEventDestroy);
 
 struct hip_device
 {
-    hip_device()
-    {
-        device_props.gcnArchName[0]      = '\0';
-        device_props.gcnArch             = 0;
-        device_props.multiProcessorCount = 0;
-        add_stream();
-    }
+    hip_device() : device_props{} { add_stream(); }
 
     hip_device(std::size_t id, std::size_t n) : device_id(id)
     {
@@ -171,7 +165,7 @@ struct hip_device
 
     std::size_t stream_id() const { return current_stream; }
 
-    std::string get_device_name() const { return get_arch_name(device_props); }
+    std::string get_device_name() const { return device_props.gcnArchName; }
 
     std::string get_gfx_name() const { return trim(split_string(get_device_name(), ':').front()); }
 

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

@@ -33,8 +33,6 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPHX_GPU_EXPORT std::string get_arch_name(const hipDeviceProp_t& props);
-
 MIGRAPHX_GPU_EXPORT std::string get_device_name();
 
 MIGRAPHX_GPU_EXPORT int get_device_id();

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

@@ -92,7 +92,7 @@ struct hip_sync_stream
         return inputs.front();
     }
 
-    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
+    argument compute(const context& ctx, const shape&, const std::vector<argument>& args) const
     {
         gpu_sync(ctx);
         if(args.empty())

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

@@ -37,7 +37,7 @@ struct module;
 namespace gpu {
 
 MIGRAPHX_GPU_EXPORT std::string dump_mlir(const module& m);
-MIGRAPHX_GPU_EXPORT code_object_op compile_mlir(const context& ctx,
+MIGRAPHX_GPU_EXPORT code_object_op compile_mlir(const context& migraphx_ctx,
                                                 module m,
                                                 const std::vector<instruction_ref>& inputs,
                                                 const value& solution);
@@ -47,7 +47,8 @@ MIGRAPHX_GPU_EXPORT instruction_ref insert_mlir(module& m,
                                                 code_object_op co,
                                                 const std::vector<instruction_ref>& inputs);
 
-MIGRAPHX_GPU_EXPORT tuning_config get_tuning_config_mlir(module m,
+MIGRAPHX_GPU_EXPORT tuning_config get_tuning_config_mlir(const context& migraphx_ctx,
+                                                         module m,
                                                          const std::vector<shape>& inputs);
 
 } // namespace gpu

src/targets/gpu/jit/ck_gemm.cpp

@@ -300,7 +300,8 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
         const auto& b_shape = inputs[1];
         const auto& c_shape = inputs.back();
 
-        auto rank = a_shape.lens().size();
+        // cppcheck-suppress unreadVariable
+        auto rank = a_shape.ndim();
 
         auto batch_count = get_batch_count(c_shape);
         auto m           = c_shape.lens()[rank - 2];

src/targets/gpu/jit/mlir.cpp

@@ -37,7 +37,7 @@ struct mlir_compiler : compiler<mlir_compiler>
     operation compile_op(context&, const std::vector<shape>&, const value&) const { return {}; }
 
     compiler_replace
-    compile(context& ctx, instruction_ref ins, const operation&, const value& solution) const
+    compile(const context& ctx, instruction_ref ins, const operation&, const value& solution) const
     {
         auto* smod = ins->module_inputs().front();
         assert(smod->get_parameter_names().size() == ins->inputs().size() - 1);
@@ -52,14 +52,16 @@ struct mlir_compiler : compiler<mlir_compiler>
                 }};
     }
 
-    optional<tuning_config>
-    get_tuning_config(context&, instruction_ref ins, const operation&, bool exhaustive) const
+    optional<tuning_config> get_tuning_config(const context& ctx,
+                                              instruction_ref ins,
+                                              const operation&,
+                                              bool exhaustive) const
     {
         if(not exhaustive)
             return nullopt;
         auto shapes = to_shapes(ins->inputs());
         auto* smod  = ins->module_inputs().front();
-        return get_tuning_config_mlir(*smod, shapes);
+        return get_tuning_config_mlir(ctx, *smod, shapes);
     }
 };
 

src/targets/gpu/mlir.cpp

@@ -36,7 +36,10 @@
 #include <mutex>
 #if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 3
 #warning "Incompatible version of rocMLIR library used, disabling"
+// Only undefine when not using cppcheck
+#ifndef CPPCHECK
 #undef MIGRAPHX_MLIR
+#endif
 #else
 #include <mlir-c/RegisterRocMLIR.h>
 #endif
@@ -50,6 +53,7 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/gpu/code_object_op.hpp>
 #include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/compile_gen.hpp>
 #include <migraphx/gpu/device_name.hpp>
 #include <migraphx/gpu/perfdb.hpp>
 #include <migraphx/gpu/tuning_config.hpp>
@@ -172,12 +176,6 @@ std::string mlir_print(F f, T x)
     return ss.str();
 }
 
-bool has_xdlops(const std::string& target_arch)
-{
-    const auto device_name = trim(split_string(target_arch, ':').front());
-    return (starts_with(device_name, "gfx9") and device_name >= "gfx908");
-}
-
 struct mlir_program
 {
     mlir_program()
@@ -512,7 +510,8 @@ struct mlir_program
         ops.add_attributes({{"function_type", make_function_type(inputs, outputs)},
                             {"sym_name", sym_name},
                             {"kernel", std::string("mixr")},
-                            {"arch", target_arch}});
+                            {"arch", target_arch},
+                            {"num_cu", num_cu}});
         ops.add_region(std::move(region));
         insert(body, std::move(ops));
 
@@ -559,14 +558,7 @@ struct mlir_program
 
     static std::string get_symbol_name(const module& m)
     {
-        for(auto ins : iterator_for(m))
-        {
-            if(ins->name() == "convolution" or ins->name() == "dot")
-            {
-                return "mlir_" + ins->name();
-            }
-        }
-        return "main";
+        return "mlir_" + gen::generate_name_from_ops(m);
     }
 
     void parse(const module& m)
@@ -602,9 +594,6 @@ struct mlir_program
             {
                 pp =
                     problem_params{ins->get_operator(), to_shapes(ins->inputs()), ins->get_shape()};
-                // check if HW supports xdlops
-                if(has_xdlops(target_arch))
-                    ops.add_attributes({{"xdlopsV2", true}});
             }
 
             std::vector<MlirValue> inputs;
@@ -653,7 +642,12 @@ struct mlir_program
         return op;
     }
 
-    void find_target() { target_arch = get_device_name(); }
+    void set_gpu_properties(const context& migraphx_ctx)
+    {
+        const auto& device = migraphx_ctx.get_current_device();
+        target_arch  = device.get_device_name();
+        num_cu       = device.get_cu_count();
+    }
 
     std::pair<std::size_t, std::size_t> get_launch_params() const
     {
@@ -667,7 +661,7 @@ struct mlir_program
 
     value::binary get_binary() const
     {
-        int size = 0;
+        size_t size = 0;
         mlirGetBinary(mmodule.get(), &size, nullptr);
         value::binary result(size);
         if(mlirGetBinary(mmodule.get(), &size, reinterpret_cast<char*>(result.data())))
@@ -675,30 +669,41 @@ struct mlir_program
         MIGRAPHX_THROW("Failed to compile mlir program");
     }
 
-    void set_tuning(const value& v)
+    void set_tuning(const value& v) MIGRAPHX_TIDY_CONST
     {
-        auto str = v.to<std::string>();
-        // We need to make a copy of the buffer since mlirRockTuningSetFromStr may modify the string
-        std::vector<char> buffer(str.begin(), str.end());
-        buffer.push_back(0);
-        if(not mlirRockTuningSetFromStr(mmodule.get(), buffer.data()))
-            MIGRAPHX_THROW("Failed setting tuning key: " + str);
+        const auto* str = v.if_string();
+        if(str == nullptr)
+            MIGRAPHX_THROW("mlir tuning solutions must be strings");
+        if(not mlirRockTuningSetFromStr(mmodule.get(), make_mlir_string_ref(*str)))
+            MIGRAPHX_THROW("Failed setting tuning key: " + *str);
     }
 
     tuning_config get_tuning_config() MIGRAPHX_TIDY_CONST
     {
         tuning_config tc;
         run_high_level_pipeline();
-        mlir_tuning_space params{mlirRockTuningSpaceCreate(mmodule.get())};
-        for(auto i : range(mlirRockTuningGetNumParamsFull(params.get())))
+        mlir_tuning_space params{
+            mlirRockTuningSpaceCreate(mmodule.get(), RocmlirTuningParamSetKindFull)};
+        for(auto i : range(mlirRockTuningGetNumParams(params.get())))
         {
             mlir_tuning_param param{mlirRockTuningParamCreate()};
             if(not mlirRockTuningParamGet(params.get(), i, param.get()))
                 MIGRAPHX_THROW("Incorrect mlir tuning parameter: " + std::to_string(i));
-            tc.solutions.push_back(std::string{mlirRockTuningGetParamStr(param.get())});
+            std::array<char, ROCMLIR_TUNING_KEY_BUFSZ> perf_key;
+            size_t perf_key_bytes =
+                mlirRockTuningParamToString(param.get(), perf_key.data(), perf_key.size());
+            if(perf_key_bytes > perf_key.size())
+                MIGRAPHX_THROW("Tuning perf key was " + std::to_string(perf_key_bytes) +
+                               " bytes and thus too long");
+            tc.solutions.emplace_back(perf_key.begin(), perf_key.begin() + perf_key_bytes);
         }
-        mlir_tuning_table tuning_table{mlirRockTuningTableCreate()};
-        tc.problem = std::string{mlirRockTuningGetKey(tuning_table.get(), mmodule.get())};
+        std::array<char, ROCMLIR_TUNING_KEY_BUFSZ> tuning_key;
+        size_t tuning_key_bytes =
+            mlirRockTuningGetKey(mmodule.get(), tuning_key.data(), tuning_key.size());
+        if(tuning_key_bytes > tuning_key.size())
+            MIGRAPHX_THROW("Tuning table key was " + std::to_string(tuning_key_bytes) +
+                           " bytes and thus too long");
+        tc.problem = std::string(tuning_key.begin(), tuning_key.begin() + tuning_key_bytes);
         return tc;
     }
 
@@ -706,10 +711,10 @@ struct mlir_program
 
     // This function appends to tuning cfg file that could be
     // used with rocMLIR tuning scripts.
-    void dump_tuning_cfg(const char* prob_config) const
+    void dump_tuning_cfg(const std::string& prob_config) const
     {
         std::string tuning_cfg_path = string_value_of(MIGRAPHX_MLIR_TUNING_CFG{});
-        if(!tuning_cfg_path.empty())
+        if(not tuning_cfg_path.empty())
         {
             std::vector<std::string> tokens = split_string(prob_config, '\t');
             std::string prob                = tokens[1];
@@ -726,51 +731,66 @@ struct mlir_program
         }
     }
 
-    static mlir_tuning_table create_tuning_table()
+    static std::pair<mlir_tuning_table, bool> load_tuning_table()
     {
         mlir_tuning_table tuning_table{mlirRockTuningTableCreate()};
+        bool found_table           = false;
         std::string tuning_db_path = string_value_of(MIGRAPHX_MLIR_TUNING_DB{});
-        if(!tuning_db_path.empty())
+        if(not tuning_db_path.empty())
         {
             std::ifstream tuning_db_tsv(tuning_db_path);
             if(tuning_db_tsv)
             {
+                found_table = true;
                 std::string line;
                 while(std::getline(tuning_db_tsv, line))
                 {
                     std::vector<std::string> tokens = split_string(line, '\t');
                     std::string arch                = tokens[0];
-                    std::string prob                = tokens[1];
-                    std::string perf                = tokens[2];
-                    std::string key                 = arch.append("\t").append(prob);
-                    mlirRockTuningUpdateTable(tuning_table.get(), key.c_str(), perf.c_str(), 1.0);
+                    std::string num_cu              = tokens[1];
+                    std::string prob                = tokens[2];
+                    std::string perf                = tokens[3];
+                    std::string key = arch.append("\t").append(num_cu).append("\t").append(prob);
+                    mlirRockTuningUpdateTable(tuning_table.get(),
+                                              make_mlir_string_ref(key),
+                                              make_mlir_string_ref(perf),
+                                              1.0);
                 }
             }
         }
         else
         {
+            found_table = false;
             std::cerr
                 << "WARNING: MLIR tuning db not found. Please set MIGRAPHX_MLIR_TUNING_DB for "
                    "optimal performance."
                 << std::endl;
         }
-        return tuning_table;
+        return std::make_pair(std::move(tuning_table), found_table);
     }
 
     bool get_module_tuned() const
     {
-        static mlir_tuning_table tuning_table = create_tuning_table();
-        // The tuning table as currently implemented is currently not
-        // thread safe. This will be fixed in the future. For now,
-        // stick a mutex around all tuning table interaction.
-        static std::mutex lock;
-        std::lock_guard<std::mutex> guard(lock);
-        if(!mlirRockTuningSetFromTable(tuning_table.get(), mmodule.get()))
+        static std::pair<mlir_tuning_table, bool> tuning_table = load_tuning_table();
+        if(not mlirRockTuningSetFromTable(tuning_table.first.get(), mmodule.get()))
         {
-            const char* prob_config = mlirRockTuningGetKey(tuning_table.get(), mmodule.get());
-            std::stringstream key(prob_config);
-            std::cerr << "fails to set param on" << prob_config << std::endl;
-            dump_tuning_cfg(prob_config);
+            std::array<char, ROCMLIR_TUNING_KEY_BUFSZ> prob_config;
+            size_t prob_config_bytes =
+                mlirRockTuningGetKey(mmodule.get(), prob_config.data(), prob_config.size());
+            if(prob_config_bytes >= prob_config.size())
+            {
+                std::cerr << "MLIR tuning key overflowed buffer, needed " << prob_config_bytes
+                          << " bytes" << std::endl;
+                return false;
+            }
+            std::string prob_config_str(prob_config.begin(),
+                                        prob_config.begin() + prob_config_bytes);
+            if(tuning_table.second)
+            {
+                std::cerr << "NOTE: MLIR tuning table did not include a key for " << prob_config_str
+                          << std::endl;
+            }
+            dump_tuning_cfg(prob_config_str);
             return false;
         }
         return true;
@@ -781,7 +801,8 @@ struct mlir_program
     mlir_module mmodule;
     problem_params pp;
     std::deque<std::string> strings{};
-    std::string target_arch;
+    std::string target_arch = "";
+    std::size_t num_cu      = 0;
     std::string sym_name;
 };
 
@@ -838,7 +859,7 @@ void adjust_param_shapes(module& m, const std::vector<shape>& inputs)
     }
 }
 
-code_object_op compile_mlir(const context&,
+code_object_op compile_mlir(const context& migraphx_ctx,
                             module m,
                             const std::vector<instruction_ref>& inputs,
                             const value& solution)
@@ -850,7 +871,7 @@ code_object_op compile_mlir(const context&,
         std::cout << m << std::endl;
 
     mlir_program mp;
-    mp.find_target();
+    mp.set_gpu_properties(migraphx_ctx);
     mp.parse(m);
     auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
     if(trace)
@@ -877,12 +898,13 @@ instruction_ref insert_mlir(module& m,
     return m.insert_instruction(ins, co, refs);
 }
 
-tuning_config get_tuning_config_mlir(module m, const std::vector<shape>& inputs)
+tuning_config
+get_tuning_config_mlir(const context& migraphx_ctx, module m, const std::vector<shape>& inputs)
 {
     adjust_param_shapes(m, inputs);
 
     mlir_program mp;
-    mp.find_target();
+    mp.set_gpu_properties(migraphx_ctx);
     mp.parse(m);
     return mp.get_tuning_config();
 }
@@ -909,10 +931,14 @@ instruction_ref
 insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<instruction_ref>&)
 {
     use(co);
+    use(m);
     return m.end();
 }
 
-tuning_config get_tuning_config_mlir(module, const std::vector<shape>&) { return {}; }
+tuning_config get_tuning_config_mlir(const context&, module, const std::vector<shape>&)
+{
+    return {};
+}
 // NOLINTEND(performance-unnecessary-value-param)
 
 #endif

src/targets/gpu/time_op.cpp

@@ -34,7 +34,7 @@ namespace gpu {
 std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsigned long seed = 0)
 {
     std::vector<argument> args;
-    std::transform(shapes.begin(), shapes.end(), std::back_inserter(args), [&](auto& s) {
+    std::transform(shapes.begin(), shapes.end(), std::back_inserter(args), [&](const auto& s) {
         return to_gpu(generate_argument(s, seed++));
     });
     return args;

src/tf/tf_parser.cpp

@@ -338,7 +338,7 @@ void tf_parser::parse_node(const std::string& name)
             std::string input_name = input;
             // if input has trailing `:0` index then remove it
             auto multi_out_idx = input.find(':');
-            if(multi_out_idx != std::string::npos && input.substr(multi_out_idx + 1) == "0")
+            if(multi_out_idx != std::string::npos and input.substr(multi_out_idx + 1) == "0")
             {
                 input_name = input.substr(0, multi_out_idx);
             }

src/value.cpp

@@ -285,7 +285,7 @@ bool value::contains(const std::string& pkey) const
 }
 std::size_t value::size() const
 {
-    auto* a = if_array_impl(x);
+    const auto* a = if_array_impl(x);
     if(a == nullptr)
         return 0;
     return a->size();

test/CMakeLists.txt

@@ -98,17 +98,11 @@ endfunction()
 
 function(add_test_executable TEST_NAME)
     add_executable(${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
-    target_link_libraries(${TEST_NAME} ${CMAKE_THREAD_LIBS_INIT})
-
-    # Cmake does not add flags correctly for gcc
-    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU")
-        set_target_properties(${TEST_NAME} PROPERTIES COMPILE_FLAGS -pthread LINK_FLAGS -pthread)
-    endif()
     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} migraphx migraphx_onnx migraphx_ref)
+    target_link_libraries(${TEST_NAME} Threads::Threads migraphx migraphx_onnx migraphx_ref)
     target_include_directories(${TEST_NAME} PUBLIC include)
 endfunction(add_test_executable)
 
@@ -208,11 +202,16 @@ endif()
 
 
 function(test_header NAME HEADER)
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
-        "#include <${HEADER}>\nint main() {}\n"
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp "
+#include <${HEADER}>
+int main() {}\n"
     )
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
-        "#include <${HEADER}>\n"
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp "
+#include <${HEADER}>
+#if defined(min) || defined(max) || defined(near) || defined(far)
+#error \"Do not include windows.h in header files\"
+#endif
+\n"
     )
     add_test_executable(${NAME}
         ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp

test/api/test_cpu.cpp

@@ -145,15 +145,15 @@ TEST_CASE(zero_parameter)
 
 TEST_CASE(set_scalar_parameter)
 {
-    auto p1 = migraphx::parse_onnx("add_bcast_test.onnx");
-    migraphx::shape s1(migraphx_shape_float_type, {3, 4});
+    auto p1 = migraphx::parse_onnx("implicit_add_bcast_test.onnx");
+    migraphx::shape s1(migraphx_shape_float_type, {3, 4, 1});
     auto param_shapes = p1.get_parameter_shapes();
     auto s1_orig      = param_shapes["1"];
     CHECK(bool{s1 == s1_orig});
 
     migraphx::onnx_options option;
     option.set_input_parameter_shape("1", {});
-    auto p2 = migraphx::parse_onnx("add_bcast_test.onnx", option);
+    auto p2 = migraphx::parse_onnx("implicit_add_bcast_test.onnx", option);
     migraphx::shape s_scalar(migraphx_shape_float_type);
     auto param_shapes_1 = p2.get_parameter_shapes();
     auto s_scalar_after = param_shapes_1["1"];