New pthreadpool_parallelize_* API

1 files changed, 139 insertions, 27 deletions

diff --git a/include/pthreadpool.h b/include/pthreadpool.h
index a99105e..9f7f4b3 100644
--- a/include/pthreadpool.h
+++ b/include/pthreadpool.h
@@ -1,16 +1,23 @@
-#include <stddef.h>
+#ifndef PTHREADPOOL_H_
+#define PTHREADPOOL_H_
 
-#ifndef PTHREADPOOL_H
-#define PTHREADPOOL_H
+#include <stddef.h>
+#include <stdint.h>
 
 typedef struct pthreadpool* pthreadpool_t;
 
-typedef void (*pthreadpool_function_1d_t)(void*, size_t);
-typedef void (*pthreadpool_function_1d_tiled_t)(void*, size_t, size_t);
-typedef void (*pthreadpool_function_2d_t)(void*, size_t, size_t);
-typedef void (*pthreadpool_function_2d_tiled_t)(void*, size_t, size_t, size_t, size_t);
-typedef void (*pthreadpool_function_3d_tiled_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t);
-typedef void (*pthreadpool_function_4d_tiled_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_1d_t)(void*, size_t);
+typedef void (*pthreadpool_task_1d_tile_1d_t)(void*, size_t, size_t);
+typedef void (*pthreadpool_task_2d_t)(void*, size_t, size_t);
+typedef void (*pthreadpool_task_2d_tile_1d_t)(void*, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_2d_tile_2d_t)(void*, size_t, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_3d_tile_2d_t)(void*, size_t, size_t, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_4d_tile_2d_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_5d_tile_2d_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t, size_t);
+typedef void (*pthreadpool_task_6d_tile_2d_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t, size_t, size_t);
+
+
+#define PTHREADPOOL_FLAG_DISABLE_DENORMALS 0x00000001
 
 #ifdef __cplusplus
 extern "C" {
@@ -37,7 +44,6 @@ pthreadpool_t pthreadpool_create(size_t threads_count);
  */
 size_t pthreadpool_get_threads_count(pthreadpool_t threadpool);
 
-
 /**
  * Processes items in parallel using threads from a thread pool.
  *
@@ -53,27 +59,143 @@ size_t pthreadpool_get_threads_count(pthreadpool_t threadpool);
  * @param[in]  items       The number of items to process. The @a function
  *    will be called once for each item.
  */
-void pthreadpool_compute_1d(
+void pthreadpool_parallelize_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_1d_t function,
+	void* argument,
+	size_t range,
+	uint32_t flags);
+
+void pthreadpool_parallelize_1d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_1d_tile_1d_t function,
+	void* argument,
+	size_t range,
+	size_t tile,
+	uint32_t flags);
+
+void pthreadpool_parallelize_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	uint32_t flags);
+
+void pthreadpool_parallelize_2d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_2d_tile_1d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t tile_j,
+	uint32_t flags);
+
+void pthreadpool_parallelize_2d_tile_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_2d_tile_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t tile_i,
+	size_t tile_j,
+	uint32_t flags);
+
+void pthreadpool_parallelize_3d_tile_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_3d_tile_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t tile_j,
+	size_t tile_k,
+	uint32_t flags);
+
+void pthreadpool_parallelize_4d_tile_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_4d_tile_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t tile_k,
+	size_t tile_l,
+	uint32_t flags);
+
+void pthreadpool_parallelize_5d_tile_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_5d_tile_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t tile_l,
+	size_t tile_m,
+	uint32_t flags);
+
+void pthreadpool_parallelize_6d_tile_2d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_6d_tile_2d_t function,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t range_n,
+	size_t tile_m,
+	size_t tile_n,
+	uint32_t flags);
+
+/**
+ * Terminates threads in the thread pool and releases associated resources.
+ *
+ * @warning  Accessing the thread pool after a call to this function constitutes
+ *    undefined behaviour and may cause data corruption.
+ *
+ * @param[in,out]  threadpool  The thread pool to destroy.
+ */
+void pthreadpool_destroy(pthreadpool_t threadpool);
+
+/* Legacy API for compatibility with pre-existing users (e.g. NNPACK) */
+#if defined(__GNUC__)
+	#define PTHREADPOOL_DEPRECATED __attribute__((__deprecated__))
+#else
+	#define PTHREADPOOL_DEPRECATED
+#endif
+
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_1d_t)(void*, size_t);
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_1d_tiled_t)(void*, size_t, size_t);
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_2d_t)(void*, size_t, size_t);
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_2d_tiled_t)(void*, size_t, size_t, size_t, size_t);
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_3d_tiled_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t);
+typedef PTHREADPOOL_DEPRECATED void (*pthreadpool_function_4d_tiled_t)(void*, size_t, size_t, size_t, size_t, size_t, size_t, size_t, size_t);
+
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_1d(
 	pthreadpool_t threadpool,
 	pthreadpool_function_1d_t function,
 	void* argument,
 	size_t range);
 
-void pthreadpool_compute_1d_tiled(
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_1d_tiled(
 	pthreadpool_t threadpool,
 	pthreadpool_function_1d_tiled_t function,
 	void* argument,
 	size_t range,
 	size_t tile);
 
-void pthreadpool_compute_2d(
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_function_2d_t function,
 	void* argument,
 	size_t range_i,
 	size_t range_j);
 
-void pthreadpool_compute_2d_tiled(
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_2d_tiled(
 	pthreadpool_t threadpool,
 	pthreadpool_function_2d_tiled_t function,
 	void* argument,
@@ -82,7 +204,7 @@ void pthreadpool_compute_2d_tiled(
 	size_t tile_i,
 	size_t tile_j);
 
-void pthreadpool_compute_3d_tiled(
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_3d_tiled(
 	pthreadpool_t threadpool,
 	pthreadpool_function_3d_tiled_t function,
 	void* argument,
@@ -93,7 +215,7 @@ void pthreadpool_compute_3d_tiled(
 	size_t tile_j,
 	size_t tile_k);
 
-void pthreadpool_compute_4d_tiled(
+PTHREADPOOL_DEPRECATED void pthreadpool_compute_4d_tiled(
 	pthreadpool_t threadpool,
 	pthreadpool_function_4d_tiled_t function,
 	void* argument,
@@ -106,18 +228,8 @@ void pthreadpool_compute_4d_tiled(
 	size_t tile_k,
 	size_t tile_l);
 
-/**
- * Terminates threads in the thread pool and releases associated resources.
- *
- * @warning  Accessing the thread pool after a call to this function constitutes
- *    undefined behaviour and may cause data corruption.
- *
- * @param[in,out]  threadpool  The thread pool to destroy.
- */
-void pthreadpool_destroy(pthreadpool_t threadpool);
-
 #ifdef __cplusplus
 } /* extern "C" */
 #endif
 
-#endif /* PTHREADPOOL_H */
+#endif /* PTHREADPOOL_H_ */