--- a/sdl/sdl.cpp
+++ b/sdl/sdl.cpp
@@ -2216,6 +2217,134 @@
 	}
 }
 
+#define THREADS_STRETCH_CNT (3)
+struct {
+	struct filter_data *in;
+	struct filter_data *out;
+	struct {
+		unsigned int src_y_start;
+		unsigned int src_y_end;
+		unsigned int dst_off;
+		SDL_sem *sem_in;
+		SDL_sem *sem_out;
+	} thread[THREADS_STRETCH_CNT];
+	volatile int exit;
+} threads_stretch_data;
+
+
+static int filter_stretch_thread(void *num) {
+	int thread_num = *((int*) &num);
+	while (!threads_stretch_data.exit) {
+		SDL_SemWait(threads_stretch_data.thread[thread_num].sem_in);
+		
+		if (threads_stretch_data.exit) {
+			return 0;
+		}
+
+		struct filter_data *in = threads_stretch_data.in;
+		struct filter_data *out = threads_stretch_data.out;
+		struct filter_stretch_data *data =
+			(struct filter_stretch_data *)out->data;
+		uint8_t *h_table = data->h_table;
+		uint8_t *v_table = data->v_table;
+		uint32_t *dst = (uint32_t*) (out->buf.u8 + threads_stretch_data.thread[thread_num].dst_off);
+		unsigned int dst_pitch = out->pitch;
+		unsigned int src_pitch = in->pitch;
+		unsigned int src_y_start = threads_stretch_data.thread[thread_num].src_y_start;
+		unsigned int src_y_end = threads_stretch_data.thread[thread_num].src_y_end;
+		unsigned int src_w = in->width;
+		unsigned int dst_w = out->width;
+		uint32_t *src = (uint32_t*) (in->buf.u8 + src_pitch * src_y_start);
+
+		dst_pitch /= sizeof(uint32_t);
+		src_pitch /= sizeof(uint32_t);
+		for (unsigned int src_y = src_y_start; src_y != src_y_end; src_y++) {
+			uint8_t v_repeat = v_table[src_y];
+			unsigned int src_x;
+			unsigned int dst_x;
+
+			if (!v_repeat) {
+				src += src_pitch;
+				continue;
+			}
+			for (src_x = 0, dst_x = 0; (src_x != src_w); ++src_x) {
+				uint8_t h_repeat = h_table[src_x];
+
+				while (h_repeat--)
+					dst[dst_x++] = src[src_x];
+			}
+
+			dst += dst_pitch;
+			while (--v_repeat) {
+				memcpy(dst, (dst - dst_pitch), (dst_w * sizeof(*dst)));
+				dst += dst_pitch;
+			}
+			src += src_pitch;
+		}
+
+		SDL_SemPost(threads_stretch_data.thread[thread_num].sem_out);
+	}
+}
+
+static void threads_stretch_data_init() {
+	threads_stretch_data.in = NULL;
+	threads_stretch_data.out = NULL;
+	for (int i = 0; i < THREADS_STRETCH_CNT; i++) {
+		threads_stretch_data.thread[i].sem_in = SDL_CreateSemaphore(0);
+		threads_stretch_data.thread[i].sem_out = SDL_CreateSemaphore(0);
+		SDL_CreateThread(filter_stretch_thread, (void*) i);
+	}
+	threads_stretch_data.exit = 0;
+}
+
+static void filter_stretch_4_master(const struct filter_data *in,
+                                    struct filter_data *out)
+{
+	struct filter_stretch_data *data =
+		(struct filter_stretch_data *)out->data;
+	uint8_t *h_table = data->h_table;
+	uint8_t *v_table = data->v_table;
+	uint32_t *dst = (uint32_t*) out->buf.u8;
+	unsigned int dst_pitch = out->pitch;
+	unsigned int dst_w = out->width;
+	uint32_t *src = (uint32_t*) in->buf.u8;
+	unsigned int src_pitch = in->pitch;
+	unsigned int src_w = in->width;
+	unsigned int src_h = in->height;
+	unsigned int src_y;
+
+	threads_stretch_data.in = const_cast<struct filter_data*>(in);
+	threads_stretch_data.out = out;
+
+	threads_stretch_data.thread[0].src_y_start = 0;
+	threads_stretch_data.thread[0].src_y_end = src_h / 3;
+	threads_stretch_data.thread[0].dst_off = 0;
+	threads_stretch_data.thread[1].src_y_start = src_h / 3;
+	threads_stretch_data.thread[1].src_y_end = (src_h * 2) / 3;
+	threads_stretch_data.thread[2].src_y_start = (src_h * 2) / 3;
+	threads_stretch_data.thread[2].src_y_end = src_h;
+
+	uint32_t dst_off = 0;
+	for (src_y = 0; src_y != src_h / 3; src_y++) {
+		uint8_t v_repeat = v_table[src_y];
+		dst_off += v_repeat * dst_pitch;
+	}
+	threads_stretch_data.thread[1].dst_off = dst_off;
+
+	for (; src_y != (src_h * 2) / 3; src_y++) {
+		uint8_t v_repeat = v_table[src_y];
+		dst_off += v_repeat * dst_pitch;
+	}
+	threads_stretch_data.thread[2].dst_off = dst_off;
+	for (int i = 0; i < THREADS_STRETCH_CNT; i++) {
+		SDL_SemPost(threads_stretch_data.thread[i].sem_in);
+	}
+
+	for (int i = 0; i < THREADS_STRETCH_CNT; i++) {
+		SDL_SemWait(threads_stretch_data.thread[i].sem_out);
+	}
+}
+
 /**
  * This filter stretches the input buffer to fill the entire output.
  * @param in Input buffer data.
@@ -2231,7 +2360,7 @@
 		{ 1, filter_stretch_X<uint8_t> },
 		{ 2, filter_stretch_X<uint16_t> },
 		{ 3, filter_stretch_3 },
-		{ 4, filter_stretch_X<uint32_t> },
+		{ 4, filter_stretch_4_master },
 	};
 	struct filter_stretch_data *data;
 	unsigned int dst_w;
@@ -3872,6 +4002,7 @@
 			break;
 		}
 	}
+	threads_stretch_data_init();
 	return 1;
 fail:
 	fprintf(stderr, "sdl: can't initialize graphics.\n");
@@ -6906,6 +7037,11 @@
 
 #ifdef WITH_THREADS
 	screen_update_thread_stop();
+	threads_stretch_data.exit = 1;
+	for (int i = 0; i < THREADS_STRETCH_CNT; i++) {
+		SDL_SemPost(threads_stretch_data.thread[i].sem_in);
+		SDL_SemPost(threads_stretch_data.thread[i].sem_out);
+	}
 #endif
 	if (mdscr.data) {
 		free((void*)mdscr.data);