Merge

src/analyze_streams.cpp

@@ -14,31 +14,31 @@ bool happens_before(const std::vector<std::size_t>& e1, const std::vector<std::s
            not std::equal(e1.begin(), e1.end(), e2.begin(), e2.end(), std::greater_equal<>{});
 }
 
-std::vector<stream_race> analyze_streams(const module& p, const stream_model& m)
+std::vector<stream_race> analyze_streams(const module& m, const stream_model& strmm)
 {
     using vector_clock = std::vector<std::size_t>;
     std::vector<stream_race> races;
-    auto nstream = m.get_nstream();
+    auto nstream = strmm.get_nstream();
     std::vector<vector_clock> vclock(nstream, vector_clock(nstream));
     std::unordered_map<instruction_ref, vector_clock> timestamp;
     std::unordered_map<std::size_t, vector_clock> events;
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
-        if(not m.has_stream(ins))
+        if(not strmm.has_stream(ins))
             continue;
-        std::size_t s = m.get_stream(ins);
+        std::size_t s = strmm.get_stream(ins);
         assert(s < nstream);
         assert(vclock.size() == nstream);
         assert(vclock[s].size() == nstream);
-        if(m.is_record(ins))
+        if(strmm.is_record(ins))
         {
             vclock[s][s]++;
-            auto event    = m.get_event_id(ins);
+            auto event    = strmm.get_event_id(ins);
             events[event] = vclock[s];
         }
-        else if(m.is_wait(ins))
+        else if(strmm.is_wait(ins))
         {
-            auto event = m.get_event_id(ins);
+            auto event = strmm.get_event_id(ins);
             if(not contains(events, event))
                 MIGRAPHX_THROW("Event is waited on before being recorded: " +
                                std::to_string(event));
@@ -57,21 +57,21 @@ std::vector<stream_race> analyze_streams(const module& p, const stream_model& m)
         }
         timestamp[ins] = vclock[s];
     }
-    for(auto ins : iterator_for(p))
+    for(auto ins : iterator_for(m))
     {
-        if(not m.has_stream(ins))
+        if(not strmm.has_stream(ins))
             continue;
         if(ins->inputs().empty())
             continue;
-        std::size_t s = m.get_stream(ins);
+        std::size_t s = strmm.get_stream(ins);
         // Find inputs from different streams
         std::vector<instruction_ref> inputs;
         fix([&](auto self, auto start) {
             for(auto input : start->inputs())
             {
-                if(not m.has_stream(input))
+                if(not strmm.has_stream(input))
                     self(input);
-                else if(m.get_stream(input) != s)
+                else if(strmm.get_stream(input) != s)
                     inputs.push_back(input);
             }
         })(ins);

src/api/CMakeLists.txt

@@ -3,7 +3,7 @@ add_library(migraphx_c
     api.cpp
 )
 set_target_properties(migraphx_c PROPERTIES EXPORT_NAME c)
-rocm_set_soversion(migraphx_c 2.0)
+rocm_set_soversion(migraphx_c 3.0)
 
 rocm_clang_tidy_check(migraphx_c)
 target_link_libraries(migraphx_c PRIVATE migraphx migraphx_tf migraphx_onnx migraphx_all_targets)

src/api/include/migraphx/migraphx.h

@@ -25,7 +25,8 @@ extern "C" {
 #endif
 
 // return code, more to be added later
-typedef enum {
+typedef enum
+{
     migraphx_status_success        = 0,
     migraphx_status_bad_param      = 1,
     migraphx_status_unknown_target = 3,
@@ -35,32 +36,13 @@ typedef enum {
 
 #define MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES(x, t) migraphx_shape_##x,
 /// An enum to represent the different data type inputs
-typedef enum {
+typedef enum
+{
     migraphx_shape_tuple_type,
     MIGRAPHX_SHAPE_VISIT_TYPES(MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES)
 } migraphx_shape_datatype_t;
 #undef MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES
 
-/// Options to be passed when compiling
-typedef struct
-{
-    /// For targets with offloaded memory(such as the gpu), this will insert
-    /// instructions during compilation to copy the input parameters to the
-    /// offloaded memory and to copy the final result from the offloaded
-    /// memory back to main memory.
-    bool offload_copy;
-    /// Optimize math functions to use faster approximate versions. There may
-    /// be slight accuracy degredation when enabled.
-    bool fast_math;
-} migraphx_compile_options;
-
-/// Options for saving and loading files
-typedef struct
-{
-    /// Format to be used for file. It can either be json or msgpack
-    const char* format;
-} migraphx_file_options;
-
 typedef struct migraphx_shape* migraphx_shape_t;
 typedef const struct migraphx_shape* const_migraphx_shape_t;
 
@@ -82,6 +64,15 @@ typedef const struct migraphx_arguments* const_migraphx_arguments_t;
 typedef struct migraphx_shapes* migraphx_shapes_t;
 typedef const struct migraphx_shapes* const_migraphx_shapes_t;
 
+typedef struct migraphx_instruction* migraphx_instruction_t;
+typedef const struct migraphx_instruction* const_migraphx_instruction_t;
+
+typedef struct migraphx_instructions* migraphx_instructions_t;
+typedef const struct migraphx_instructions* const_migraphx_instructions_t;
+
+typedef struct migraphx_modules* migraphx_modules_t;
+typedef const struct migraphx_modules* const_migraphx_modules_t;
+
 typedef struct migraphx_module* migraphx_module_t;
 typedef const struct migraphx_module* const_migraphx_module_t;
 
@@ -94,6 +85,12 @@ typedef const struct migraphx_operation* const_migraphx_operation_t;
 typedef struct migraphx_onnx_options* migraphx_onnx_options_t;
 typedef const struct migraphx_onnx_options* const_migraphx_onnx_options_t;
 
+typedef struct migraphx_file_options* migraphx_file_options_t;
+typedef const struct migraphx_file_options* const_migraphx_file_options_t;
+
+typedef struct migraphx_compile_options* migraphx_compile_options_t;
+typedef const struct migraphx_compile_options* const_migraphx_compile_options_t;
+
 typedef struct migraphx_tf_options* migraphx_tf_options_t;
 typedef const struct migraphx_tf_options* const_migraphx_tf_options_t;
 
@@ -103,8 +100,24 @@ typedef const struct migraphx_quantize_op_names* const_migraphx_quantize_op_name
 typedef struct migraphx_quantize_int8_options* migraphx_quantize_int8_options_t;
 typedef const struct migraphx_quantize_int8_options* const_migraphx_quantize_int8_options_t;
 
+typedef struct migraphx_context* migraphx_context_t;
+typedef const struct migraphx_context* const_migraphx_context_t;
+
+typedef struct migraphx_experimental_custom_op* migraphx_experimental_custom_op_t;
+typedef const struct migraphx_experimental_custom_op* const_migraphx_experimental_custom_op_t;
+
+typedef migraphx_status (*migraphx_experimental_custom_op_compute_shape)(migraphx_shape_t out,
+                                                                         void* obj,
+                                                                         migraphx_shapes_t inputs);
+
+typedef migraphx_status (*migraphx_experimental_custom_op_copy)(void** out, void* input);
+
+typedef migraphx_status (*migraphx_experimental_custom_op_delete)(void* input);
+
 migraphx_status migraphx_shape_destroy(migraphx_shape_t shape);
 
+migraphx_status migraphx_shape_assign_to(migraphx_shape_t output, const_migraphx_shape_t input);
+
 migraphx_status migraphx_shape_create(migraphx_shape_t* shape,
                                       migraphx_shape_datatype_t type,
                                       size_t* lengths,
@@ -135,6 +148,9 @@ migraphx_shape_equal(bool* out, const_migraphx_shape_t shape, const_migraphx_sha
 
 migraphx_status migraphx_argument_destroy(migraphx_argument_t argument);
 
+migraphx_status migraphx_argument_assign_to(migraphx_argument_t output,
+                                            const_migraphx_argument_t input);
+
 migraphx_status
 migraphx_argument_create(migraphx_argument_t* argument, const_migraphx_shape_t shape, void* buffer);
 
@@ -151,11 +167,17 @@ migraphx_argument_generate(migraphx_argument_t* out, const_migraphx_shape_t s, s
 
 migraphx_status migraphx_target_destroy(migraphx_target_t target);
 
+migraphx_status migraphx_target_assign_to(migraphx_target_t output, const_migraphx_target_t input);
+
 migraphx_status migraphx_target_create(migraphx_target_t* target, const char* name);
 
 migraphx_status migraphx_program_parameter_shapes_destroy(
     migraphx_program_parameter_shapes_t program_parameter_shapes);
 
+migraphx_status
+migraphx_program_parameter_shapes_assign_to(migraphx_program_parameter_shapes_t output,
+                                            const_migraphx_program_parameter_shapes_t input);
+
 migraphx_status migraphx_program_parameter_shapes_size(
     size_t* out, migraphx_program_parameter_shapes_t program_parameter_shapes);
 
@@ -170,6 +192,9 @@ migraphx_status migraphx_program_parameter_shapes_names(
 migraphx_status
 migraphx_program_parameters_destroy(migraphx_program_parameters_t program_parameters);
 
+migraphx_status migraphx_program_parameters_assign_to(migraphx_program_parameters_t output,
+                                                      const_migraphx_program_parameters_t input);
+
 migraphx_status
 migraphx_program_parameters_create(migraphx_program_parameters_t* program_parameters);
 
@@ -179,6 +204,9 @@ migraphx_status migraphx_program_parameters_add(migraphx_program_parameters_t pr
 
 migraphx_status migraphx_arguments_destroy(migraphx_arguments_t arguments);
 
+migraphx_status migraphx_arguments_assign_to(migraphx_arguments_t output,
+                                             const_migraphx_arguments_t input);
+
 migraphx_status migraphx_arguments_size(size_t* out, migraphx_arguments_t arguments);
 
 migraphx_status
@@ -186,21 +214,81 @@ migraphx_arguments_get(const_migraphx_argument_t* out, migraphx_arguments_t argu
 
 migraphx_status migraphx_shapes_destroy(migraphx_shapes_t shapes);
 
+migraphx_status migraphx_shapes_assign_to(migraphx_shapes_t output, const_migraphx_shapes_t input);
+
 migraphx_status migraphx_shapes_size(size_t* out, migraphx_shapes_t shapes);
 
 migraphx_status
 migraphx_shapes_get(const_migraphx_shape_t* out, migraphx_shapes_t shapes, size_t idx);
 
+migraphx_status migraphx_instruction_destroy(migraphx_instruction_t instruction);
+
+migraphx_status migraphx_instruction_assign_to(migraphx_instruction_t output,
+                                               const_migraphx_instruction_t input);
+
+migraphx_status migraphx_instructions_destroy(migraphx_instructions_t instructions);
+
+migraphx_status migraphx_instructions_assign_to(migraphx_instructions_t output,
+                                                const_migraphx_instructions_t input);
+
+migraphx_status migraphx_instructions_create(migraphx_instructions_t* instructions,
+                                             const_migraphx_instruction_t* ptr,
+                                             size_t size);
+
+migraphx_status migraphx_modules_destroy(migraphx_modules_t modules);
+
+migraphx_status migraphx_modules_assign_to(migraphx_modules_t output,
+                                           const_migraphx_modules_t input);
+
+migraphx_status
+migraphx_modules_create(migraphx_modules_t* modules, migraphx_module_t* ptr, size_t size);
+
+migraphx_status migraphx_module_create(migraphx_module_t* module, char* name);
+
 migraphx_status migraphx_module_print(const_migraphx_module_t module);
 
+migraphx_status migraphx_module_add_instruction(migraphx_instruction_t* out,
+                                                migraphx_module_t module,
+                                                migraphx_operation_t op,
+                                                migraphx_instructions_t args);
+
+migraphx_status migraphx_module_add_instruction_with_mod_args(migraphx_instruction_t* out,
+                                                              migraphx_module_t module,
+                                                              migraphx_operation_t op,
+                                                              migraphx_instructions_t args,
+                                                              migraphx_modules_t module_refs);
+
+migraphx_status migraphx_module_add_literal(migraphx_instruction_t* out,
+                                            migraphx_module_t module,
+                                            const_migraphx_shape_t shape,
+                                            const char* buffer);
+
+migraphx_status migraphx_module_add_parameter(migraphx_instruction_t* out,
+                                              migraphx_module_t module,
+                                              const char* name,
+                                              const_migraphx_shape_t shape);
+
+migraphx_status migraphx_module_add_return(migraphx_instruction_t* out,
+                                           migraphx_module_t module,
+                                           migraphx_instructions_t args);
+
 migraphx_status migraphx_program_destroy(migraphx_program_t program);
 
+migraphx_status migraphx_program_assign_to(migraphx_program_t output,
+                                           const_migraphx_program_t input);
+
+migraphx_status migraphx_program_create(migraphx_program_t* program);
+
 migraphx_status migraphx_program_get_main_module(migraphx_module_t* out,
                                                  migraphx_program_t program);
 
+migraphx_status migraphx_program_create_module(migraphx_module_t* out,
+                                               migraphx_program_t program,
+                                               const char* name);
+
 migraphx_status migraphx_program_compile(migraphx_program_t program,
                                          migraphx_target_t target,
-                                         migraphx_compile_options* options);
+                                         migraphx_compile_options_t options);
 
 migraphx_status migraphx_program_get_parameter_shapes(migraphx_program_parameter_shapes_t* out,
                                                       migraphx_program_t program);
@@ -219,22 +307,32 @@ migraphx_status migraphx_program_run(migraphx_arguments_t* out,
 migraphx_status
 migraphx_program_equal(bool* out, const_migraphx_program_t program, const_migraphx_program_t x);
 
+migraphx_status migraphx_program_experimental_get_context(migraphx_context_t* out,
+                                                          const_migraphx_program_t program);
+
 migraphx_status migraphx_operation_destroy(migraphx_operation_t operation);
 
+migraphx_status migraphx_operation_assign_to(migraphx_operation_t output,
+                                             const_migraphx_operation_t input);
+
 migraphx_status migraphx_operation_create(migraphx_operation_t* operation,
                                           const char* name,
-                                          const char* attributes);
+                                          const char* attributes,
+                                          ...);
 
 migraphx_status migraphx_operation_name(char* out, size_t out_size, migraphx_operation_t operation);
 
 migraphx_status
-migraphx_load(migraphx_program_t* out, const char* name, migraphx_file_options* options);
+migraphx_load(migraphx_program_t* out, const char* name, migraphx_file_options_t options);
 
 migraphx_status
-migraphx_save(migraphx_program_t p, const char* name, migraphx_file_options* options);
+migraphx_save(migraphx_program_t p, const char* name, migraphx_file_options_t options);
 
 migraphx_status migraphx_onnx_options_destroy(migraphx_onnx_options_t onnx_options);
 
+migraphx_status migraphx_onnx_options_assign_to(migraphx_onnx_options_t output,
+                                                const_migraphx_onnx_options_t input);
+
 migraphx_status migraphx_onnx_options_create(migraphx_onnx_options_t* onnx_options);
 
 migraphx_status migraphx_onnx_options_set_input_parameter_shape(
@@ -243,6 +341,33 @@ migraphx_status migraphx_onnx_options_set_input_parameter_shape(
 migraphx_status migraphx_onnx_options_set_default_dim_value(migraphx_onnx_options_t onnx_options,
                                                             size_t value);
 
+migraphx_status
+migraphx_onnx_options_set_default_loop_iterations(migraphx_onnx_options_t onnx_options,
+                                                  int64_t value);
+
+migraphx_status migraphx_file_options_destroy(migraphx_file_options_t file_options);
+
+migraphx_status migraphx_file_options_assign_to(migraphx_file_options_t output,
+                                                const_migraphx_file_options_t input);
+
+migraphx_status migraphx_file_options_create(migraphx_file_options_t* file_options);
+
+migraphx_status migraphx_file_options_set_file_format(migraphx_file_options_t file_options,
+                                                      const char* format);
+
+migraphx_status migraphx_compile_options_destroy(migraphx_compile_options_t compile_options);
+
+migraphx_status migraphx_compile_options_assign_to(migraphx_compile_options_t output,
+                                                   const_migraphx_compile_options_t input);
+
+migraphx_status migraphx_compile_options_create(migraphx_compile_options_t* compile_options);
+
+migraphx_status
+migraphx_compile_options_set_offload_copy(migraphx_compile_options_t compile_options, bool value);
+
+migraphx_status migraphx_compile_options_set_fast_math(migraphx_compile_options_t compile_options,
+                                                       bool value);
+
 migraphx_status
 migraphx_parse_onnx(migraphx_program_t* out, const char* name, migraphx_onnx_options_t options);
 
@@ -253,6 +378,9 @@ migraphx_status migraphx_parse_onnx_buffer(migraphx_program_t* out,
 
 migraphx_status migraphx_tf_options_destroy(migraphx_tf_options_t tf_options);
 
+migraphx_status migraphx_tf_options_assign_to(migraphx_tf_options_t output,
+                                              const_migraphx_tf_options_t input);
+
 migraphx_status migraphx_tf_options_create(migraphx_tf_options_t* tf_options);
 
 migraphx_status migraphx_tf_options_set_nhwc(migraphx_tf_options_t tf_options, bool is_nhwc);
@@ -274,6 +402,9 @@ migraphx_parse_tf(migraphx_program_t* out, const char* name, migraphx_tf_options
 
 migraphx_status migraphx_quantize_op_names_destroy(migraphx_quantize_op_names_t quantize_op_names);
 
+migraphx_status migraphx_quantize_op_names_assign_to(migraphx_quantize_op_names_t output,
+                                                     const_migraphx_quantize_op_names_t input);
+
 migraphx_status migraphx_quantize_op_names_create(migraphx_quantize_op_names_t* quantize_op_names);
 
 migraphx_status migraphx_quantize_op_names_add(migraphx_quantize_op_names_t quantize_op_names,
@@ -287,6 +418,10 @@ migraphx_status migraphx_quantize_fp16(migraphx_program_t prog);
 migraphx_status
 migraphx_quantize_int8_options_destroy(migraphx_quantize_int8_options_t quantize_int8_options);
 
+migraphx_status
+migraphx_quantize_int8_options_assign_to(migraphx_quantize_int8_options_t output,
+                                         const_migraphx_quantize_int8_options_t input);
+
 migraphx_status
 migraphx_quantize_int8_options_create(migraphx_quantize_int8_options_t* quantize_int8_options);
 
@@ -301,6 +436,30 @@ migraphx_status migraphx_quantize_int8(migraphx_program_t prog,
                                        migraphx_target_t target,
                                        migraphx_quantize_int8_options_t options);
 
+migraphx_status migraphx_context_finish(const_migraphx_context_t context);
+
+migraphx_status migraphx_context_get_queue(void** out, migraphx_context_t context);
+
+migraphx_status
+migraphx_experimental_custom_op_destroy(migraphx_experimental_custom_op_t experimental_custom_op);
+
+migraphx_status
+migraphx_experimental_custom_op_assign_to(migraphx_experimental_custom_op_t output,
+                                          const_migraphx_experimental_custom_op_t input);
+
+migraphx_status
+migraphx_experimental_custom_op_create(migraphx_experimental_custom_op_t* experimental_custom_op,
+                                       void* obj,
+                                       migraphx_experimental_custom_op_copy c,
+                                       migraphx_experimental_custom_op_delete d,
+                                       const char* name);
+
+migraphx_status migraphx_experimental_custom_op_set_compute_shape(
+    migraphx_experimental_custom_op_t obj, migraphx_experimental_custom_op_compute_shape input);
+
+migraphx_status
+migraphx_experimental_custom_op_register(migraphx_experimental_custom_op_t experimental_custom_op);
+
 #ifdef __cplusplus
 }
 #endif

src/api/migraphx.py

@@ -178,14 +178,58 @@ def shapes(h):
              returns='const migraphx::shape&')
 
 
+@api.handle('migraphx_instruction', 'migraphx::instruction_ref')
+def instruction(h):
+    pass
+
+
+@api.handle('migraphx_instructions', 'std::vector<migraphx::instruction_ref>')
+def instructions(h):
+    h.constructor(
+        'create',
+        api.params(ptr='const_migraphx_instruction_t*', size='size_t'),
+        fname='migraphx::to_obj_vector<const_migraphx_instruction_t>')
+
+
+@api.handle('migraphx_modules', 'std::vector<migraphx::module*>')
+def modules(h):
+    h.constructor('create',
+                  api.params(ptr='migraphx_module_t*', size='size_t'),
+                  fname='migraphx::to_objptr_vector<migraphx::module*>')
+
+
 @auto_handle(ref=True)
 def module(h):
+    h.constructor('create', api.params(name='std::string'))
     h.method('print', invoke='migraphx::print_module($@)', const=True)
+    h.method('add_instruction',
+             api.params(op='migraphx::operation',
+                        args='std::vector<migraphx::instruction_ref>'),
+             returns='migraphx::instruction_ref')
+    h.method('add_instruction_with_mod_args',
+             api.params(op='migraphx::operation',
+                        args='std::vector<migraphx::instruction_ref>',
+                        module_refs='std::vector<migraphx::module*>'),
+             fname='add_instruction',
+             returns='migraphx::instruction_ref')
+    h.method('add_literal',
+             api.params(shape='const migraphx::shape&', buffer='const char*'),
+             returns='migraphx::instruction_ref')
+    h.method('add_parameter',
+             api.params(name='const char*', shape='const migraphx::shape&'),
+             returns='migraphx::instruction_ref')
+    h.method('add_return',
+             api.params(args='std::vector<migraphx::instruction_ref>'),
+             returns='migraphx::instruction_ref')
 
 
 @auto_handle()
 def program(h):
+    h.constructor('create')
     h.method('get_main_module', returns='migraphx::module*')
+    h.method('create_module',
+             api.params(name='const char*'),
+             returns='migraphx::module*')
     h.method(
         'compile',
         api.params(target='migraphx::target',
@@ -207,12 +251,18 @@ def program(h):
              invoke='migraphx::equal($@)',
              returns='bool',
              const=True)
+    h.method('experimental_get_context',
+             invoke='migraphx::get_context($@)',
+             const=True,
+             returns='migraphx::context')
 
 
 @auto_handle()
 def operation(h):
     h.constructor('create',
-                  api.params(name='const char*', attributes='const char*'),
+                  api.params(name='const char*',
+                             attributes='const char*',
+                             vlist='...'),
                   fname='migraphx::create_op')
     h.method('name', returns='std::string')
 
@@ -243,6 +293,30 @@ def onnx_options(h):
         api.params(value='size_t'),
         invoke='migraphx::set_default_dim_value($@)',
     )
+    h.method(
+        'set_default_loop_iterations',
+        api.params(value='int64_t'),
+        invoke='migraphx::set_default_loop_iterations($@)',
+    )
+
+
+@auto_handle()
+def file_options(h):
+    h.constructor('create')
+    h.method('set_file_format',
+             api.params(format='const char*'),
+             invoke='migraphx::set_file_format($@)')
+
+
+@auto_handle()
+def compile_options(h):
+    h.constructor('create')
+    h.method('set_offload_copy',
+             api.params(value='bool'),
+             invoke='migraphx::set_offload_copy($@)')
+    h.method('set_fast_math',
+             api.params(value='bool'),
+             invoke='migraphx::set_fast_math($@)')
 
 
 api.add_function('migraphx_parse_onnx',
@@ -327,3 +401,19 @@ api.add_function('migraphx_quantize_int8',
                             target='migraphx::target',
                             options='migraphx::quantize_int8_options'),
                  fname='migraphx::quantize_int8_wrap')
+
+
+@auto_handle(ref=True)
+def context(h):
+    h.method('finish', const=True)
+    h.method('get_queue', returns='void*', fname='get_queue().unsafe_get')
+
+
+@api.interface('migraphx_experimental_custom_op',
+               'migraphx::experimental_custom_op')
+def experimental_custom_op(h):
+    h.constructor('create', api.params(name='const char*'))
+    h.virtual('compute_shape',
+              api.params(inputs='std::vector<migraphx::shape>'),
+              returns='migraphx::shape')
+    h.method('register', invoke='migraphx::register_custom_op($@)')

src/apply_alpha_beta.cpp

+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/common.hpp>
+#include <migraphx/apply_alpha_beta.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+instruction_ref insert_apply_alpha_beta(module& m,
+                                        instruction_ref pos,
+                                        const std::vector<instruction_ref>& args,
+                                        const operation& op,
+                                        const literal& alpha,
+                                        const literal& beta)
+{
+    auto a          = args[0];
+    auto b          = args[1];
+    auto input_type = a->get_shape().type();
+    if(!float_equal(alpha.at<float>(0), 1.0))
+    {
+        auto alpha_literal = m.add_literal(alpha);
+        a                  = insert_common_op(m, pos, migraphx::make_op("mul"), {alpha_literal, a});
+        if(a->get_shape().type() != input_type)
+        {
+            a = m.insert_instruction(pos, make_op("convert", {{"target_type", input_type}}), a);
+        }
+    }
+    auto op_res = m.insert_instruction(pos, op, a, b);
+    if(args.size() == 3)
+    {
+        if(not float_equal(beta.at<float>(0), 0.0) && args[2]->get_shape().elements() > 0)
+        {
+            auto out_lens = op_res->get_shape().lens();
+            auto c        = args[2];
+            auto c_lens   = c->get_shape().lens();
+            input_type    = c->get_shape().type();
+            if(out_lens != c_lens)
+            {
+                c = m.insert_instruction(
+                    pos, migraphx::make_op("multibroadcast", {{"out_lens", out_lens}}), args[2]);
+            }
+            auto beta_literal = m.add_literal(beta);
+            auto beta_c = insert_common_op(m, pos, migraphx::make_op("mul"), {c, beta_literal});
+            if(beta_c->get_shape().type() != input_type)
+            {
+                beta_c = m.insert_instruction(
+                    pos, migraphx::make_op("convert", {{"target_type", input_type}}), beta_c);
+            }
+            return m.insert_instruction(pos, migraphx::make_op("add"), op_res, beta_c);
+        }
+    }
+    return op_res;
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/argument.cpp

@@ -8,7 +8,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 argument::argument(const shape& s) : m_shape(s)
 {
     auto buffer = make_shared_array<char>(s.bytes());
-    m_data      = {[=]() mutable { return buffer.get(); }};
+    assign_buffer({[=]() mutable { return buffer.get(); }});
 }
 
 argument::argument(shape s, std::nullptr_t)
@@ -18,14 +18,17 @@ argument::argument(shape s, std::nullptr_t)
 
 argument::argument(const shape& s, const argument::data_t& d) : m_shape(s), m_data(d) {}
 
-argument argument::load(const shape& s, char* buffer)
+void argument::assign_buffer(std::function<char*()> d)
 {
+    const shape& s = m_shape;
     if(s.type() != shape::tuple_type)
-        return argument{s, buffer};
+    {
+        m_data = {std::move(d)};
+        return;
+    }
     // Collect all shapes
     std::unordered_map<std::size_t, shape> shapes;
     {
-        // cppcheck-suppress variableScope
         std::size_t i = 0;
         fix([&](auto self, auto ss) {
             if(ss.sub_shapes().empty())
@@ -56,21 +59,23 @@ argument argument::load(const shape& s, char* buffer)
     }
     assert(offset == s.bytes());
 
-    // cppcheck-suppress variableScope
     std::size_t i = 0;
-    return fix<argument>([&](auto self, auto ss) {
+    m_data        = fix<data_t>([&](auto self, auto ss) {
+        data_t result;
         if(ss.sub_shapes().empty())
         {
-            argument r{shapes[i], buffer + offsets[i]};
+            auto n = offsets[i];
+            result = {[d, n]() mutable { return d() + n; }};
             i++;
-            return r;
+            return result;
         }
-        std::vector<argument> subs;
+        std::vector<data_t> subs;
         std::transform(ss.sub_shapes().begin(),
                        ss.sub_shapes().end(),
                        std::back_inserter(subs),
                        [&](auto child) { return self(child); });
-        return argument{subs};
+        result.sub = subs;
+        return result;
     })(s);
 }
 
@@ -99,7 +104,11 @@ bool argument::empty() const { return not m_data.get and m_data.sub.empty(); }
 
 const shape& argument::get_shape() const { return this->m_shape; }
 
-argument argument::reshape(const shape& s) const { return {s, this->m_data}; }
+argument argument::reshape(const shape& s) const
+{
+    assert(s.element_space() <= this->get_shape().element_space());
+    return {s, this->m_data};
+}
 
 argument::data_t argument::data_t::share() const
 {
@@ -148,5 +157,13 @@ std::vector<argument> argument::get_sub_objects() const
     return result;
 }
 
+argument argument::element(std::size_t i) const
+{
+    assert(this->get_shape().sub_shapes().empty());
+    auto idx    = this->get_shape().index(i);
+    auto offset = this->get_shape().type_size() * idx;
+    return argument{shape{this->get_shape().type()}, this->data() + offset};
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/common.cpp

@@ -96,7 +96,7 @@ instruction_ref insert_common_op(module& m,
         if(input->get_shape().lens() != common.lens())
         {
             input = m.insert_instruction(
-                ins, make_op("multibroadcast", {{"output_lens", common.lens()}}), input);
+                ins, make_op("multibroadcast", {{"out_lens", common.lens()}}), input);
         }
         if(input->get_shape().type() != common.type())
         {

src/compile_src.cpp

@@ -28,13 +28,20 @@ std::vector<char> src_compiler::compile(const std::vector<src_file>& srcs) const
         {
             params += " " + src.path.filename().string();
             if(out.empty())
-                out = src.path.stem().string() + ".o";
+                out = src.path.stem().string() + out_ext;
         }
     }
 
     params += " -o " + out;
 
+    if(not launcher.empty())
+    {
+        td.execute(launcher, compiler + " " + params);
+    }
+    else
+    {
         td.execute(compiler, params);
+    }
 
     auto out_path = td.path / out;
     if(not fs::exists(out_path))