Improve Windows build support.

- Use multithreading (works with a sufficiently recent version of
  the Microsoft developer tools / C SDK).
- Detect CPU architecture and use AVX2 or NEON when appropriate.
- Automatically detect python installation path.

2 files changed, 67 insertions, 35 deletions

diff --git a/src/utils/prefetch.h b/src/utils/prefetch.h
index 89b14e8..a2c41ef 100644
--- a/src/utils/prefetch.h
+++ b/src/utils/prefetch.h
@@ -1,15 +1,7 @@
-#if defined(AVX2)
-
-#define prefetch(a, i) _mm_prefetch(a+i, _MM_HINT_T0)
-
-#else
 #if defined(__GNUC__) || defined(__clang__)
-
-#define prefetch(a, i) __builtin_prefetch(a+i, 0, 0)
-
+    #define prefetch(a, i) __builtin_prefetch(a+i, 0, 0)
+#elif defined(AVX2)
+    #define prefetch(a, i) _mm_prefetch((char *)(a+i), _MM_HINT_T0)
 #else
-
-#define prefetch(a, i) (void)i
-
-#endif
+    #define prefetch(a, i) (void)i
 #endif
diff --git a/src/utils/wrapthread.h b/src/utils/wrapthread.h
index 28654fe..4ac72a6 100644
--- a/src/utils/wrapthread.h
+++ b/src/utils/wrapthread.h
@@ -1,37 +1,77 @@
 #if THREADS == 1
+        #define WRAPTHREAD_NOTHREADS 1
+#elif defined(__unix__)
+        #define WRAPTHREAD_PTHREADS 1
+#elif defined(__has_include)
+        #if __has_include(<pthreads.h>)
+                #define WRAPTHREAD_PTHREADS 1
+        #elif __has_include(<threads.h>)
+                #define WRAPTHREAD_C11THREADS 1
+        #endif
+#else
+        #define WRAPTHREAD_NOTHREADS 1
+#endif
+
+#if WRAPTHREAD_PTHREADS
+        #include <pthread.h>
+
+        #define wrapthread_atomic _Atomic
+
+        #define wrapthread_return_t void *
+        #define wrapthread_return_val NULL
+
+        #define wrapthread_define_var_thread_t(x) pthread_t x
+        #define wrapthread_define_var_mutex_t(x) pthread_mutex_t x
+        #define wrapthread_define_struct_thread_t(x) pthread_t x
+        #define wrapthread_define_struct_mutex_t(x) pthread_mutex_t x
+
+        #define wrapthread_define_if_threads(T, x) T x
+
+        #define wrapthread_create(a, f, arg) pthread_create(a, NULL, f, arg)
+        #define wrapthread_join(a) pthread_join(a, NULL)
+        #define wrapthread_mutex_init(a) pthread_mutex_init(a, NULL)
+        #define wrapthread_mutex_lock(a) pthread_mutex_lock(a)
+        #define wrapthread_mutex_unlock(a) pthread_mutex_unlock(a)
+
+#elif WRAPTHREAD_C11THREADS
+        #include <threads.h>
+
+        #define wrapthread_atomic _Atomic
 
-#define wrapthread_atomic
+        #define wrapthread_return_t int
+        #define wrapthread_return_val 0
 
-#define wrapthread_define_var_thread_t(x) unused char x
-#define wrapthread_define_var_mutex_t(x) unused char x
-#define wrapthread_define_struct_thread_t(x) char x
-#define wrapthread_define_struct_mutex_t(x) char x
+        #define wrapthread_define_var_thread_t(x) thrd_t x
+        #define wrapthread_define_var_mutex_t(x) mtx_t x
+        #define wrapthread_define_struct_thread_t(x) thrd_t x
+        #define wrapthread_define_struct_mutex_t(x) mtx_t x
 
-#define wrapthread_define_if_threads(T, x) T x; (void)(x)
+        #define wrapthread_define_if_threads(T, x) T x
 
-#define wrapthread_create(a, b, c, d) c(d)
-#define wrapthread_join(a, b)
-#define wrapthread_mutex_init(a, b)
-#define wrapthread_mutex_lock(a)
-#define wrapthread_mutex_unlock(a)
+        #define wrapthread_create(a, f, arg) thrd_create(a, f, arg)
+        #define wrapthread_join(a) thrd_join(a, NULL)
+        #define wrapthread_mutex_init(a) mtx_init(a, mtx_plain)
+        #define wrapthread_mutex_lock(a) mtx_lock(a)
+        #define wrapthread_mutex_unlock(a) mtx_unlock(a)
 
 #else
 
-#include <pthread.h>
+        #define wrapthread_atomic
 
-#define wrapthread_atomic _Atomic
+        #define wrapthread_return_t int
+        #define wrapthread_return_val 0
 
-#define wrapthread_define_var_thread_t(x) pthread_t x
-#define wrapthread_define_var_mutex_t(x) pthread_mutex_t x
-#define wrapthread_define_struct_thread_t(x) pthread_t x
-#define wrapthread_define_struct_mutex_t(x) pthread_mutex_t x
+        #define wrapthread_define_var_thread_t(x) unused char x
+        #define wrapthread_define_var_mutex_t(x) unused char x
+        #define wrapthread_define_struct_thread_t(x) char x
+        #define wrapthread_define_struct_mutex_t(x) char x
 
-#define wrapthread_define_if_threads(T, x) T x
+        #define wrapthread_define_if_threads(T, x) T x; (void)(x)
 
-#define wrapthread_create(a, b, c, d) pthread_create(a, b, c, d)
-#define wrapthread_join(a, b) pthread_join(a, b)
-#define wrapthread_mutex_init(a, b) pthread_mutex_init(a, b)
-#define wrapthread_mutex_lock(a) pthread_mutex_lock(a)
-#define wrapthread_mutex_unlock(a) pthread_mutex_unlock(a)
+        #define wrapthread_create(t, f, arg) f(arg)
+        #define wrapthread_join(a)
+        #define wrapthread_mutex_init(a)
+        #define wrapthread_mutex_lock(a)
+        #define wrapthread_mutex_unlock(a)
 
 #endif