tune the preloads a bit

diff --git a/src/video/SDL_blit_A.c b/src/video/SDL_blit_A.c
index b013ed3..504451b 100644 (file)
--- a/src/video/SDL_blit_A.c
+++ b/src/video/SDL_blit_A.c
@@ -77,6 +77,7 @@ static void name(SDL_BlitInfo *info) \
        int dststride = width * dstBpp + info->d_skip; \
 \
        while ( height-- ) { \
        int dststride = width * dstBpp + info->d_skip; \
 \
        while ( height-- ) { \

+           __builtin_prefetch(dst + dststride); \

            neon_name(dst, src, width); \
            src += srcstride; \
            dst += dststride; \
            neon_name(dst, src, width); \
            src += srcstride; \
            dst += dststride; \

diff --git a/src/video/SDL_blit_neon.S b/src/video/SDL_blit_neon.S
index 979bb2a..dcbfd74 100644 (file)
--- a/src/video/SDL_blit_neon.S
+++ b/src/video/SDL_blit_neon.S
@@ -63,13 +63,13 @@
     vdup.i16   q12, r12
 
 0:
     vdup.i16   q12, r12
 
 0:

-    pld        [r1, #64*2]

     cmp        r2, #8
     cmp        r2, #8

-    pld        [r0, #64*2]

     blt        3f
 1:
     vld4.8     {d4-d7}, [r1]!
     blt        3f
 1:
     vld4.8     {d4-d7}, [r1]!

+    pld        [r1, #64*2]

     vld4.8     {d0-d3}, [r0]
     vld4.8     {d0-d3}, [r0]

+    pld        [r0, #64+32]

 2:
 .if \bgr2rgb
     vswp       d4, d6          @ BGR->RGB
 2:
 .if \bgr2rgb
     vswp       d4, d6          @ BGR->RGB


@@ -164,13 +164,13 @@ do_argb_finish:
 .endif
     vdup.i16   q12, r12
 0:
 .endif
     vdup.i16   q12, r12
 0:

-    pld        [r1, #64*2]

     cmp        r2, #8
     cmp        r2, #8

-    pld        [r0, #64*2]

     blt        3f
 1:
     vld4.8     {d4-d7}, [r1]!
     blt        3f
 1:
     vld4.8     {d4-d7}, [r1]!

+    pld        [r1, #64*2]

     vld2.8     {d1-d2}, [r0]
     vld2.8     {d1-d2}, [r0]

+    pld        [r0, #64+32]

 .if \bgr2rgb
     vswp       d4, d6          @ BGR->RGB
 .endif
 .if \bgr2rgb
     vswp       d4, d6          @ BGR->RGB
 .endif