gpulib: implement rgb888_to_rgb565

and use it for menu

4 files changed, 37 insertions, 2 deletions

diff --git a/frontend/menu.c b/frontend/menu.c
index 3ebe9f5..3152cb2 100644
--- a/frontend/menu.c
+++ b/frontend/menu.c
@@ -590,7 +590,7 @@ static void draw_savestate_bg(int slot)
 
 	x = gpu->ulControl[5] & 0x3ff;
 	y = (gpu->ulControl[5] >> 10) & 0x1ff;
-	s = (u16 *)gpu->psxVRam + y * 1024 + (x & ~1);
+	s = (u16 *)gpu->psxVRam + y * 1024 + x;
 	w = psx_widths[(gpu->ulStatus >> 16) & 7];
 	tmp = gpu->ulControl[7];
 	h = ((tmp >> 10) & 0x3ff) - (tmp & 0x3ff);
@@ -2242,7 +2242,7 @@ static void menu_leave_emu(void)
 		}
 		else {
 			for (; h > 0; h--, d += g_menuscreen_w, s += last_psx_w * 3) {
-				bgr888_to_rgb565(d, s, w * 3);
+				rgb888_to_rgb565(d, s, w * 3);
 				menu_darken_bg(d, d, w, 0);
 			}
 		}
diff --git a/plugins/gpulib/cspace.c b/plugins/gpulib/cspace.c
index 408211f..1d5718c 100644
--- a/plugins/gpulib/cspace.c
+++ b/plugins/gpulib/cspace.c
@@ -39,5 +39,6 @@ void bgr888_to_rgb565(void *dst_, const void *src_, int bytes)
 }
 
 // TODO?
+void rgb888_to_rgb565(void *dst, const void *src, int bytes) {}
 void bgr888_to_rgb888(void *dst, const void *src, int bytes) {}
 
diff --git a/plugins/gpulib/cspace.h b/plugins/gpulib/cspace.h
index 644143b..8c9bcfa 100644
--- a/plugins/gpulib/cspace.h
+++ b/plugins/gpulib/cspace.h
@@ -6,6 +6,7 @@ extern "C"
 void bgr555_to_rgb565(void *dst, const void *src, int bytes);
 void bgr888_to_rgb888(void *dst, const void *src, int bytes);
 void bgr888_to_rgb565(void *dst, const void *src, int bytes);
+void rgb888_to_rgb565(void *dst, const void *src, int bytes);
 
 #ifdef __cplusplus
 }
diff --git a/plugins/gpulib/cspace_neon.s b/plugins/gpulib/cspace_neon.s
index abc0381..b458f06 100644
--- a/plugins/gpulib/cspace_neon.s
+++ b/plugins/gpulib/cspace_neon.s
@@ -129,4 +129,37 @@ bgr888_to_rgb565:
     bx          lr
 
 
+.global rgb888_to_rgb565
+rgb888_to_rgb565:
+    pld         [r1]
+    @ r2 /= 48
+    mov         r2, r2, lsr #4
+    movw        r3, #0x5556
+    movt        r3, #0x5555
+    umull       r12,r2, r3, r2
+
+    mov         r3, #0x07e0
+    vdup.16     q15, r3
+0:
+    pld         [r1, #48*3]
+    vld3.8      {d1-d3}, [r1, :64]!
+    vld3.8      {d5-d7}, [r1, :64]!
+
+    vshll.u8    q8, d2, #3      @ g
+    vshll.u8    q9, d6, #3
+    vshr.u8     d2, d1, #3      @ b
+    vshr.u8     d6, d5, #3
+    vzip.8      d2, d3          @ rb
+    vzip.8      d6, d7
+    vbit        q1, q8, q15
+    vbit        q3, q9, q15
+
+    vstmia      r0!, {d2,d3}
+    vstmia      r0!, {d6,d7}
+    subs        r2, r2, #1
+    bne         0b
+
+    bx          lr
+
+
 @ vim:filetype=armasm