Avoid copying data if it comes from/to a shared buffer, even internally.

[simgrid.git] / include / smpi / smpi.h

diff --git a/include/smpi/smpi.h b/include/smpi/smpi.h
index d1ac611..bc9dcb7 100644 (file)
--- a/include/smpi/smpi.h
+++ b/include/smpi/smpi.h
@@ -8,6 +8,7 @@
 
 #include <simgrid/forward.h>
 #include <smpi/forward.hpp>
+#include <smpi/smpi_helpers_internal.h>
 #include <xbt/function_types.h>
 
 #include <stddef.h>
@@ -29,7 +30,7 @@
   type _XBT_CONCAT(P, name) args
 #endif
 
-SG_BEGIN_DECL()
+SG_BEGIN_DECL
 #define MPI_THREAD_SINGLE     0
 #define MPI_THREAD_FUNNELED   1
 #define MPI_THREAD_SERIALIZED 2
@@ -298,7 +299,7 @@ XBT_PUBLIC_DATA const MPI_Datatype MPI_INTEGER16;
 XBT_PUBLIC_DATA const MPI_Datatype MPI_COUNT;
 
 //defines for fortran compatibility
-#if defined(__alpha__) || defined(__sparc64__) || defined(__x86_64__) || defined(__ia64__)
+#if defined(__alpha__) || defined(__sparc64__) || defined(__x86_64__) || defined(__ia64__) || defined(__aarch64__)
 #define MPI_INTEGER MPI_INT
 #define MPI_2INTEGER MPI_2INT
 #define MPI_LOGICAL MPI_INT
@@ -1006,38 +1007,41 @@ XBT_PUBLIC void smpi_trace_set_call_location__(const char* file, int* line);
 
 #define SMPI_ITER_NAME1(line) _XBT_CONCAT(iter_count, line)
 #define SMPI_ITER_NAME(line) SMPI_ITER_NAME1(line)
-#define SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, global, iters, thres)\
-  int SMPI_ITER_NAME(__LINE__)=0;\
-  {loop_init;\
-  while(loop_end){\
-    SMPI_ITER_NAME(__LINE__)++;\
-    loop_iter;\
-  }}                                                                         \
-  for(loop_init; \
-      loop_end ? (smpi_sample_1(global, __FILE__, __LINE__, iters, thres), (smpi_sample_2(global, __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__)))) :\
-      smpi_sample_exit(global, __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__));\
-      smpi_sample_3(global, __FILE__, __LINE__),loop_iter)
-#define SMPI_SAMPLE_LOCAL(loop_init, loop_end, loop_iter, iters, thres) SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, 0, iters, thres)
-#define SMPI_SAMPLE_GLOBAL(loop_init, loop_end, loop_iter,iters, thres) SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, 1, iters, thres)
+#define SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, global, iters, thres)                                         \
+  int SMPI_ITER_NAME(__LINE__) = 0;                                                                                    \
+  {                                                                                                                    \
+    loop_init;                                                                                                         \
+    while (loop_end) {                                                                                                 \
+      SMPI_ITER_NAME(__LINE__)++;                                                                                      \
+      (loop_iter);                                                                                                     \
+    }                                                                                                                  \
+  }                                                                                                                    \
+  for (loop_init; (loop_end) ? (smpi_sample_1((global), __FILE__, __LINE__, (iters), (thres)),                         \
+                                (smpi_sample_2((global), __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__))))               \
+                             : smpi_sample_exit((global), __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__));               \
+       smpi_sample_3((global), __FILE__, __LINE__), (loop_iter))
+#define SMPI_SAMPLE_LOCAL(loop_init, loop_end, loop_iter, iters, thres)                                                \
+  SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 0, (iters), (thres))
+#define SMPI_SAMPLE_GLOBAL(loop_init, loop_end, loop_iter, iters, thres)                                               \
+  SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 1, (iters), (thres))
 #define SMPI_SAMPLE_DELAY(duration) for(smpi_execute(duration); 0; )
 #define SMPI_SAMPLE_FLOPS(flops) for(smpi_execute_flops(flops); 0; )
-
 XBT_PUBLIC void* smpi_shared_malloc(size_t size, const char* file, int line);
-#define SMPI_SHARED_MALLOC(size) smpi_shared_malloc(size, __FILE__, __LINE__)
+#define SMPI_SHARED_MALLOC(size) smpi_shared_malloc((size), __FILE__, __LINE__)
 XBT_PUBLIC void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks);
 #define SMPI_PARTIAL_SHARED_MALLOC(size, shared_block_offsets, nb_shared_blocks)                                       \
-  smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks)
+  smpi_shared_malloc_partial((size), (shared_block_offsets), (nb_shared_blocks))
+
 
-XBT_PUBLIC void smpi_shared_free(void* data);
 #define SMPI_SHARED_FREE(data) smpi_shared_free(data)
 
 XBT_PUBLIC int smpi_shared_known_call(const char* func, const char* input);
 XBT_PUBLIC void* smpi_shared_get_call(const char* func, const char* input);
 XBT_PUBLIC void* smpi_shared_set_call(const char* func, const char* input, void* data);
 #define SMPI_SHARED_CALL(func, input, ...)                                                                             \
-  (smpi_shared_known_call(_XBT_STRINGIFY(func), input)                                                                 \
-       ? smpi_shared_get_call(_XBT_STRINGIFY(func), input)                                                             \
-       : smpi_shared_set_call(_XBT_STRINGIFY(func), input, (func(__VA_ARGS__))))
+  (smpi_shared_known_call(_XBT_STRINGIFY(func), (input))                                                               \
+       ? smpi_shared_get_call(_XBT_STRINGIFY(func), (input))                                                           \
+       : smpi_shared_set_call(_XBT_STRINGIFY(func), (input), ((func)(__VA_ARGS__))))
 
 /* Fortran specific stuff */
 
@@ -1053,11 +1057,11 @@ XBT_PUBLIC void SMPI_app_instance_register(const char* name, xbt_main_func_t cod
 XBT_PUBLIC void SMPI_init();
 XBT_PUBLIC void SMPI_finalize();
 
-SG_END_DECL()
+SG_END_DECL
 
 /* C++ declarations for shared_malloc */
 #ifdef __cplusplus
-XBT_PUBLIC int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>>& private_blocks, size_t* offset);
+XBT_PUBLIC int smpi_is_shared(const void* ptr, std::vector<std::pair<size_t, size_t>>& private_blocks, size_t* offset);
 
 std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>>& vec,
                                                                       size_t offset, size_t buff_size);