arm: provide bgr888_to_rgb565 for v6

diff --git a/Makefile b/Makefile
index 7f95fe4..89e54ff 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -319,7 +319,7 @@ endif
 OBJS += frontend/cspace.o
 ifeq "$(HAVE_NEON_ASM)" "1"
 OBJS += frontend/cspace_neon.o
-frontend/cspace.o: CFLAGS += -DHAVE_bgr555_to_rgb565 -DHAVE_bgr888_to_x
+frontend/cspace.o: CFLAGS += -DHAVE_bgr555_to_rgb565
 else
 ifeq "$(ARCH)" "arm"
 OBJS += frontend/cspace_arm.o

diff --git a/frontend/cspace.c b/frontend/cspace.c
index bd5e7a3..2d5349e 100644 (file)
--- a/frontend/cspace.c
+++ b/frontend/cspace.c
@@ -10,6 +10,7 @@
 
 #include <stdint.h>
 #include "cspace.h"
+#include "compiler_features.h"
 
 /*
  * note: these are intended for testing and should be avoided
@@ -111,9 +112,7 @@ void bgr555_to_rgb565(void *dst_, const void *src_, int bytes)
 
 #endif
 
-#ifndef HAVE_bgr888_to_x
-
-void bgr888_to_rgb565(void *dst_, const void *src_, int bytes)
+void attr_weak bgr888_to_rgb565(void *dst_, const void *src_, int bytes)
 {
     const unsigned char *src = src_;
     unsigned int *dst = dst_;
@@ -140,8 +139,6 @@ void bgr888_to_rgb565(void *dst_, const void *src_, int bytes)
 void rgb888_to_rgb565(void *dst, const void *src, int bytes) {}
 void bgr888_to_rgb888(void *dst, const void *src, int bytes) {}
 
-#endif // __ARM_NEON__
-
 /* YUV stuff */
 static int yuv_ry[32], yuv_gy[32], yuv_by[32];
 static unsigned char yuv_u[32 * 2], yuv_v[32 * 2];

diff --git a/frontend/cspace_arm.S b/frontend/cspace_arm.S
index 177b085..3ef5083 100644 (file)
--- a/frontend/cspace_arm.S
+++ b/frontend/cspace_arm.S
@@ -68,3 +68,52 @@ FUNCTION(bgr555_to_rgb565): @ void *dst, const void *src, int bytes
     bgt      2b
 
     pop      {r4-r11,pc}
+
+
+#ifdef HAVE_ARMV6 /* v6-only due to potential misaligned reads */
+
+# r1b0g0r0 g2r2b1g1 b3g3r3b2
+FUNCTION(bgr888_to_rgb565):
+    pld      [r1]
+    push     {r4-r10,lr}
+
+    mov      r10, #0x001f          @ b mask
+    mov      r12, #0x07e0          @ g mask
+    mov      lr,  #0xf800          @ r mask
+
+0:
+    ldr      r3, [r1], #4          @ may be unaligned
+    ldr      r4, [r1], #4
+    ldr      r5, [r1], #4
+    pld      [r1, #32*1]
+    and      r6, r10,r3, lsr #16+3 @ b0
+    and      r7, r12,r3, lsr #5    @ g0
+    and      r8, lr, r3, lsl #8    @ r0
+    and      r9, lr, r3, lsr #16   @ r1
+    orr      r6, r6, r7
+    orr      r6, r6, r8            @ r0g0b0
+
+    and      r7, r12,r4, lsl #3    @ g1
+    and      r8, r10,r4, lsr #11   @ b1
+    orr      r9, r9, r7
+    orr      r9, r9, r8            @ r1g1b1
+    and      r7, lr, r4, lsr #8    @ r2
+    and      r8, r12,r4, lsr #21   @ g2
+    pkhbt    r9, r6, r9, lsl #16
+    str      r9, [r0], #4
+
+    and      r6, r10,r5, lsr #3    @ b2
+    orr      r7, r7, r8
+    orr      r6, r6, r7            @ r2g2b2
+    and      r7, lr, r5            @ r3
+    and      r8, r12,r5, lsr #13   @ g3
+    orr      r7, r7, r5, lsr #27   @ r3b3
+    orr      r7, r7, r8            @ r3g3b3
+    pkhbt    r7, r6, r7, lsl #16
+    str      r7, [r0], #4
+    subs     r2, r2, #12
+    bgt      0b
+
+    pop      {r4-r10,pc}
+
+#endif /* HAVE_ARMV6 */

diff --git a/include/compiler_features.h b/include/compiler_features.h
index 0ab8468..d698363 100644 (file)
--- a/include/compiler_features.h
+++ b/include/compiler_features.h
@@ -8,11 +8,13 @@
 #  define noinline       __attribute__((noinline,noclone))
 # endif
 # define attr_unused     __attribute__((unused))
+# define attr_weak       __attribute__((weak))
 #else
 # define likely(x)       (x)
 # define unlikely(x)     (x)
 # define noinline
 # define attr_unused
+# define attr_weak
 #endif
 
 #ifndef __has_builtin