From a327967e78393018a9f2a7edb38bc3af657e597a Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Tue, 7 Dec 2010 13:00:46 +0200
Subject: [PATCH] simple neon BGR555 to RGB565 converter

---
 Makefile                   |  4 ++++
 frontend/arm_utils.s       | 48 ++++++++++++++++++++++++++++++++++++++
 plugins/dfxvideo/draw_fb.c |  9 ++++++-
 3 files changed, 60 insertions(+), 1 deletion(-)
 create mode 100644 frontend/arm_utils.s

diff --git a/Makefile b/Makefile
index bf3eee26..d0b4bc76 100644
--- a/Makefile
+++ b/Makefile
@@ -7,6 +7,7 @@ CFLAGS += -ggdb -Ifrontend
 LDFLAGS += -lz -lpthread -ldl
 ifdef CROSS_COMPILE
 CFLAGS += -mcpu=cortex-a8 -mtune=cortex-a8 -mfloat-abi=softfp -ffast-math
+ASFLAGS += -mcpu=cortex-a8 -mfpu=neon
 endif
 ifndef DEBUG
 CFLAGS += -O2
@@ -55,6 +56,9 @@ OBJS += gui/Config.o gui/Plugin.o
 
 OBJS += frontend/main.o frontend/plugin.o frontend/plugin_lib.o
 OBJS += frontend/linux/fbdev.o
+ifdef CROSS_COMPILE
+OBJS += frontend/arm_utils.o
+endif
 
 $(TARGET): $(OBJS)
 	$(CC) -o $@ $^ $(LDFLAGS) -Wl,-Map=$@.map
diff --git a/frontend/arm_utils.s b/frontend/arm_utils.s
new file mode 100644
index 00000000..edaafb86
--- /dev/null
+++ b/frontend/arm_utils.s
@@ -0,0 +1,48 @@
+/*
+ * (C) Gražvydas "notaz" Ignotas, 2010
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+.text
+.align 2
+
+.global bgr555_to_rgb565
+bgr555_to_rgb565:
+    mov         r3, #0x03e0
+    vdup.16     q15, r3
+    mov         r2, r2, lsr #6
+0:
+    vldmia      r1!, {q0-q3}
+    vshr.u16    q4, q0, #10
+    vshr.u16    q5, q1, #10
+    vshr.u16    q6, q2, #10
+    vshr.u16    q7, q3, #10
+    vshl.u16    q8, q0, #11
+    vshl.u16    q9, q1, #11
+    vshl.u16    q10, q2, #11
+    vshl.u16    q11, q3, #11
+    vand        q0, q0, q15
+    vand        q1, q1, q15
+    vand        q2, q2, q15
+    vand        q3, q3, q15
+    vshl.u16    q0, q0, #1
+    vshl.u16    q1, q1, #1
+    vshl.u16    q2, q2, #1
+    vshl.u16    q3, q3, #1
+    vorr        q0, q0, q4
+    vorr        q1, q1, q5
+    vorr        q2, q2, q6
+    vorr        q3, q3, q7
+    vorr        q0, q0, q8
+    vorr        q1, q1, q9
+    vorr        q2, q2, q10
+    vorr        q3, q3, q11
+    vstmia      r0!, {q0-q3}
+    subs        r2, r2, #1
+    bne         0b
+
+    bx          lr
+
+@ vim:filetype=armasm
diff --git a/plugins/dfxvideo/draw_fb.c b/plugins/dfxvideo/draw_fb.c
index ce3f5154..e2510710 100644
--- a/plugins/dfxvideo/draw_fb.c
+++ b/plugins/dfxvideo/draw_fb.c
@@ -36,9 +36,13 @@ PSXPoint_t     ptCursorPoint[8];
 unsigned short usCursorActive = 0;
 char *         pCaptionText;
 
+#ifndef __arm__
+#define bgr555_to_rgb565 memcpy
+#endif
 
 static void blit(void)
 {
+ extern void bgr555_to_rgb565(void *dst, void *src, int bytes);
  int x = PSXDisplay.DisplayPosition.x;
  int y = PSXDisplay.DisplayPosition.y;
  int w = PreviousPSXDisplay.Range.x1;
@@ -46,6 +50,9 @@ static void blit(void)
  int pitch = PreviousPSXDisplay.DisplayMode.x * 2;
  unsigned char *dest = pl_fbdev_buf;
 
+ if (w <= 0)
+   return;
+
  // TODO: clear border if centering
 
  // account for centering
@@ -57,7 +64,7 @@ static void blit(void)
    unsigned short *srcs = psxVuw + y * 1024 + x;
    for (; h-- > 0; dest += pitch, srcs += 1024)
    {
-     memcpy(dest, srcs, w * 2);
+     bgr555_to_rgb565(dest, srcs, w * 2);
    }
  }
 }
-- 
2.39.2