neon code for bgra to rgba conversion for copyFTBitmap

src/opts/SkFontHost_FreeType_common_opts_neon.cpp

Index: src/opts/SkFontHost_FreeType_common_opts_neon.cpp
diff --git a/src/opts/SkFontHost_FreeType_common_opts_neon.cpp b/src/opts/SkFontHost_FreeType_common_opts_neon.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..22a9fdfc0ad316a3cb928ecdb3b6e630bf7a9468
--- /dev/null
+++ b/src/opts/SkFontHost_FreeType_common_opts_neon.cpp
@@ -0,0 +1,60 @@
+#include <arm_neon.h>
+
+void BGRA2RGBA_Neon(const uint8_t* src, uint8_t* dst, size_t height, const size_t width, const size_t dstRowBytes, const int srcPitch)

[reed1, 2015/01/29 15:47:19]

1. col max is 100
2. why is dst..rowBytes and src..Pitch? Skia convention is to use rowBytes

[frederic.ma, 2015/01/29 18:33:09]

1. Col max fixed
2. dst..rowBytes and src..Pitch renamed.

+{
+    asm volatile (
+
+        "cmp    %[height], #0                                           \n\t"   // if (height == 0) exit
+        "beq    End                                                     \n\t"
+
+        "RowLoop:"
+
+            "mov    r1, %[width]                                        \n\t"   // reset r1 = width for each new line
+            "mov    r2, %[src]                                          \n\t"
+            "mov    r3, %[dst]                                          \n\t"
+
+            "ColumnBatchLoop:"
+
+                "cmp    r1, #8                                          \n\t"   // while (r1 >= 8)
+                "blt    PixelLoopStart                                  \n\t"
+
+                // load src, 8 pixels at a time, each pixel taking 32bit value
+                "vld4.8 {d0, d1, d2, d3}, [r2]!                         \n\t"   //d0: B0B1B2B3...B7 d1: G0G1G2G3...G7 d2: R0R1R2R3...G7 d3: A0A1A2A3...A7
+                "vswp d0, d2                                            \n\t"   //swap d0 and d2 so we have RGBA format
+                "vst4.8 {d0, d1, d2, d3}, [r3]!                         \n\t"   //write back to dst
+
+                "sub    r1, r1, #8                                      \n\t"   //r1 -= 8 because we are processing 8 pixels at a time
+                "b      ColumnBatchLoop                                 \n\t"
+
+                "PixelLoopStart:                                        \n\t"
+
+                    "cmp    r1, #0                                      \n\t"   // while (r1 > 0)
+                    "beq    PixelLoopEnd                                \n\t"
+
+                "PixelLoop:"
+
+                    "vld4.8  {d0[0], d1[0], d2[0], d3[0]}, [r2]!        \n\t"   //load 1 32bits pixel at a time and swap again
+                    "vswp d0, d2                                        \n\t"
+                    "vst4.8  {d0[0], d1[0], d2[0], d3[0]}, [r3]!        \n\t"
+
+                    "subs    r1, r1, #1                                 \n\t"   // pixel counter decrement
+                    "bne     PixelLoop                                  \n\t"
+
+                "PixelLoopEnd:                                          \n\t"
+
+            "add    %[src], %[src], %[srcPitch]                         \n\t"
+            "add    %[dst], %[dst], %[dstRowBytes]                      \n\t"
+
+            "subs   %[height], %[height], #1                            \n\t"   // while (height)
+            "bne    RowLoop                                             \n\t"   // process next line
+
+        "End:                                                           \n\t"
+
+            : [height] "+r" (height), [dst] "+r" (dst), [src] "+r" (src)
+            : [width] "r" (width), [srcPitch] "r" (srcPitch), [dstRowBytes] "r" (dstRowBytes)
+            : "cc", "memory",
+              "r1", "r2", "r3",
+              "d0", "d1", "d2", "d3");
+}
+
+