diff --git a/include/fmpi_core.h b/include/fmpi_core.h
index 40e985c212ad72eb5688b53b4123d4f7493e3d61..10664ddaff56a8d9ac93f510cf7dfda8b8ff4ab9 100644
--- a/include/fmpi_core.h
+++ b/include/fmpi_core.h
@@ -150,12 +150,12 @@ do { \
FMPI_REGISTER_SYNC_TASK()
------------------------------------------------------------------------------*/
#define FMPI_REGISTER_SYNC_TASK(ctx, func, stencil, ...) \
- FMPI_TASK_REGISTER_IMPL(func, sync, ctx, stencil, __VA_ARGS__)
+ FMPI_TASK_REGISTER_SYNC_IMPL(func, ctx, stencil, __VA_ARGS__)
/*------------------------------------------------------------------------------
FMPI_REGISTER_ASYNC_TASK()
------------------------------------------------------------------------------*/
#define FMPI_REGISTER_ASYNC_TASK(ctx, func, stencil, ...) \
- FMPI_TASK_REGISTER_IMPL(func, async, ctx, stencil, __VA_ARGS__)
+ FMPI_TASK_REGISTER_ASYNC_IMPL(func, ctx, stencil, __VA_ARGS__)
/*==============================================================================
GUARD
==============================================================================*/
diff --git a/include/fmpi_task.h b/include/fmpi_task.h
index 39e26cad537fb3f28fadee7ef3896c3425f0dae5..9967f7db60d3bc2612df4816a026419392ee9ac6 100644
--- a/include/fmpi_task.h
+++ b/include/fmpi_task.h
@@ -116,7 +116,7 @@ typedef struct fmpi_task {
PUBLIC FUNCTION
==============================================================================*/
/*------------------------------------------------------------------------------
- fmpi_task_register_sync()
+ fmpi_task_register()
------------------------------------------------------------------------------*/
/**
* TODO
@@ -136,16 +136,10 @@ typedef struct fmpi_task {
* TODO
* }
*/
-struct fmpi_task fmpi_task_register_sync(
+struct fmpi_task fmpi_task_register(
struct fmpi_ctx * ctx, fmpi_task_func func, const char * name,
- struct fmpi_stencil stencil, const struct fmpi_task_args * args
-);
-/*------------------------------------------------------------------------------
- fmpi_task_register_async()
-------------------------------------------------------------------------------*/
-struct fmpi_task fmpi_task_register_async(
- struct fmpi_ctx * ctx, fmpi_task_func func, const char * name,
- struct fmpi_stencil stencil, const struct fmpi_task_args * args
+ enum fmpi_task_type type, struct fmpi_stencil stencil,
+ const struct fmpi_task_args * args
);
/*------------------------------------------------------------------------------
fmpi_task_run_sync()
diff --git a/include/internal/generic/fmpi_task_generic.h b/include/internal/generic/fmpi_task_generic.h
index f1624e7dc1155b6f3b2789b2ed6aa96118566cb9..350abd0e6db958d73fcbdf5480892b0bf9e2e6eb 100644
--- a/include/internal/generic/fmpi_task_generic.h
+++ b/include/internal/generic/fmpi_task_generic.h
@@ -73,18 +73,22 @@ _Pragma("GCC diagnostic warning \"-Wincompatible-pointer-types\"")\
return futhark_entry_##FUNC(ctx, out, FMPI_PRIV_TASK_ARGS_##N(args_in)); \
}
-#define FMPI_TASK_REGISTER_IMPL(FUNC, TYPE, ctx, stencil, ...) \
+#define FMPI_TASK_REGISTER_SYNC_IMPL(FUNC, ctx, stencil, ...) \
FMPI_PRIV_TASK_CONCAT(FMPI_PRIV_TASK_REGISTER_, CPL_ARG_COUNT(__VA_ARGS__)) \
- (FUNC, TYPE, ctx, stencil ,__VA_ARGS__)
+ (FUNC, ctx, FMPI_TASK_TYPE_SYNC, stencil ,__VA_ARGS__)
-#define FMPI_PRIV_TASK_REGISTER_1(FUNC, TYPE, ctx, stencil, arg_out) \
- fmpi_task_register_##TYPE((ctx), FUNC##_0, #FUNC, (stencil), &(struct fmpi_task_args){ \
+#define FMPI_TASK_REGISTER_ASYNC_IMPL(FUNC, ctx, stencil, ...) \
+ FMPI_PRIV_TASK_CONCAT(FMPI_PRIV_TASK_REGISTER_, CPL_ARG_COUNT(__VA_ARGS__)) \
+ (FUNC, ctx, FMPI_TASK_TYPE_ASYNC, stencil ,__VA_ARGS__)
+
+#define FMPI_PRIV_TASK_REGISTER_1(FUNC, ctx, type, stencil, arg_out) \
+ fmpi_task_register((ctx), FUNC##_0, #FUNC, (type), (stencil), &(struct fmpi_task_args){ \
.out = (arg_out), \
.cnt = 0 \
})
-#define FMPI_PRIV_TASK_REGISTER_N(N, FUNC, TYPE, ctx, stencil, arg_out, ...) \
- fmpi_task_register_##TYPE((ctx), FUNC##_##N, #FUNC, (stencil), &(struct fmpi_task_args){ \
+#define FMPI_PRIV_TASK_REGISTER_N(N, FUNC, ctx, type, stencil, arg_out, ...) \
+ fmpi_task_register((ctx), FUNC##_##N, #FUNC, (type), (stencil), &(struct fmpi_task_args){ \
.in = {__VA_ARGS__}, \
.out = (arg_out), \
.cnt = N \
diff --git a/src/fmpi_task.c b/src/fmpi_task.c
index f7b6503b7769d2f0585ce9e8cafb7f3dfd7b2f59..2845ae351d7a0d17c745b92c191123762e662107 100644
--- a/src/fmpi_task.c
+++ b/src/fmpi_task.c
@@ -40,21 +40,22 @@
PUBLIC FUNCTION DEFINITION
==============================================================================*/
/*------------------------------------------------------------------------------
- fmpi_task_register_sync()
+ fmpi_task_register()
------------------------------------------------------------------------------*/
-struct fmpi_task fmpi_task_register_sync(
+struct fmpi_task fmpi_task_register(
struct fmpi_ctx * const ctx, const fmpi_task_func func,
- const char * const name, const struct fmpi_stencil stencil,
- const struct fmpi_task_args * const args
+ const char * const name, const enum fmpi_task_type type,
+ const struct fmpi_stencil stencil, const struct fmpi_task_args * const args
){
assert(ctx != NULL);
assert(ctx->task_cnt < FMPI_TASK_MAX);
assert(name != NULL);
assert(args != NULL);
+
struct fmpi_task task = {
.func = func,
.name = name,
- .type = FMPI_TASK_TYPE_SYNC,
+ .type = type,
.args = *args,
.stencil = stencil
};
@@ -63,58 +64,33 @@ struct fmpi_task fmpi_task_register_sync(
const struct fmpi_data * const data = (stencil.type != FMPI_STENCIL_NONE)
? &task.domains[i].halo
: &task.domains[i].inner;
+ void * gpu_data = NULL;
+ const char * func_name = NULL;
+ if(task.type == FMPI_TASK_TYPE_SYNC) {
//! @todo Could fmpi_futhark_new_data_async() be called here instead?
- void * gpu_data = fmpi_futhark_new_data_sync(
- ctx->fut, data->raw, data->type.base, data->dim_cnt,
- data->dim_len[0], data->dim_len[1], data->dim_len[2]
- );
- if(gpu_data == NULL) {
- FMPI_RAISE_ERROR(ctx->err_handler, "FMPI",
- "fmpi_futhark_new_data_sync() failed!"
+ gpu_data = fmpi_futhark_new_data_sync(
+ ctx->fut, data->raw, data->type.base, data->dim_cnt,
+ data->dim_len[0], data->dim_len[1], data->dim_len[2]
);
+ func_name = "fmpi_futhark_new_data_sync";
+ } else {
+ gpu_data = fmpi_futhark_new_data_async(
+ ctx->fut, data->raw, data->type.base, data->dim_cnt,
+ data->dim_len[0], data->dim_len[1], data->dim_len[2]
+ );
+ func_name = "fmpi_futhark_new_data_async";
}
- task.args.in[i].gpu = gpu_data;
- }
- ctx->tasks[ctx->task_cnt++] = task;
- fmpi_futhark_sync(ctx->fut);
- return task;
-}
-/*------------------------------------------------------------------------------
- fmpi_task_register_async()
-------------------------------------------------------------------------------*/
-struct fmpi_task fmpi_task_register_async(
- struct fmpi_ctx * const ctx, const fmpi_task_func func,
- const char * const name, const struct fmpi_stencil stencil,
- const struct fmpi_task_args * const args
-){
- assert(ctx != NULL);
- assert(ctx->task_cnt < FMPI_TASK_MAX);
- assert(name != NULL);
- assert(args != NULL);
- struct fmpi_task task = {
- .func = func,
- .name = name,
- .type = FMPI_TASK_TYPE_ASYNC,
- .args = *args,
- .stencil = stencil
- };
- for(size_t i = 0; i < task.args.cnt; i++) {
- task.domains[i] = fmpi_new_domain(ctx, &args->in[i], stencil);
- const struct fmpi_data * const data = (stencil.type != FMPI_STENCIL_NONE)
- ? &task.domains[i].halo
- : &task.domains[i].inner;
- void * gpu_data = fmpi_futhark_new_data_async(
- ctx->fut, data->raw, data->type.base, data->dim_cnt,
- data->dim_len[0], data->dim_len[1], data->dim_len[2]
- );
if(gpu_data == NULL) {
FMPI_RAISE_ERROR(ctx->err_handler, "FMPI",
- "fmpi_futhark_new_data_async() failed!"
+ "%s() failed!", func_name
);
}
task.args.in[i].gpu = gpu_data;
}
ctx->tasks[ctx->task_cnt++] = task;
+ if(task.type == FMPI_TASK_TYPE_SYNC) {
+ fmpi_futhark_sync(ctx->fut);
+ }
return task;
}
/*------------------------------------------------------------------------------