Optimized premultiplying swizzles for NEON

src/core/SkOpts.cpp

Index: src/core/SkOpts.cpp
diff --git a/src/core/SkOpts.cpp b/src/core/SkOpts.cpp
index ee88b231690b0e8d7f0cd895a9609f9bf680235b..e8408bef0caca130386c7a52af695d7bce8a24d4 100644
--- a/src/core/SkOpts.cpp
+++ b/src/core/SkOpts.cpp
@@ -16,6 +16,7 @@
 #include "SkFloatingPoint_opts.h"
 #include "SkMatrix_opts.h"
 #include "SkMorphologyImageFilter_opts.h"
+#include "SkSwizzler_opts.h"
 #include "SkTextureCompressor_opts.h"
 #include "SkUtils_opts.h"
 #include "SkXfermode_opts.h"
@@ -49,58 +50,6 @@
     #include <cpu-features.h>
 #endif
 
-namespace sk_default {
-
-// These variable names in these functions just pretend the input is BGRA.
-// They work fine with both RGBA and BGRA.
-
-static void premul_xxxa(uint32_t dst[], const uint32_t src[], int count) {
-    for (int i = 0; i < count; i++) {
-        uint8_t a = src[i] >> 24,
-                r = src[i] >> 16,
-                g = src[i] >>  8,
-                b = src[i] >>  0;
-        r = (r*a+127)/255;
-        g = (g*a+127)/255;
-        b = (b*a+127)/255;
-        dst[i] = (uint32_t)a << 24
-               | (uint32_t)r << 16
-               | (uint32_t)g <<  8
-               | (uint32_t)b <<  0;
-    }
-}
-
-static void swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) {
-    for (int i = 0; i < count; i++) {
-        uint8_t a = src[i] >> 24,
-                r = src[i] >> 16,
-                g = src[i] >>  8,
-                b = src[i] >>  0;
-        dst[i] = (uint32_t)a << 24
-               | (uint32_t)b << 16
-               | (uint32_t)g <<  8
-               | (uint32_t)r <<  0;
-    }
-}
-
-static void premul_swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) {
-    for (int i = 0; i < count; i++) {
-        uint8_t a = src[i] >> 24,
-                r = src[i] >> 16,
-                g = src[i] >>  8,
-                b = src[i] >>  0;
-        r = (r*a+127)/255;
-        g = (g*a+127)/255;
-        b = (b*a+127)/255;
-        dst[i] = (uint32_t)a << 24
-               | (uint32_t)b << 16
-               | (uint32_t)g <<  8
-               | (uint32_t)r <<  0;
-    }
-}
-
-}  // namespace sk_default
-
 namespace SkOpts {
     // Define default function pointer values here...
     // If our global compile options are set high enough, these defaults might even be
@@ -132,9 +81,9 @@ namespace SkOpts {
     decltype(matrix_scale_translate) matrix_scale_translate = sk_default::matrix_scale_translate;
     decltype(matrix_affine)          matrix_affine          = sk_default::matrix_affine;
 
-    decltype(       premul_xxxa)        premul_xxxa = sk_default::       premul_xxxa;
-    decltype(       swaprb_xxxa)        swaprb_xxxa = sk_default::       swaprb_xxxa;
-    decltype(premul_swaprb_xxxa) premul_swaprb_xxxa = sk_default::premul_swaprb_xxxa;
+    decltype(       premul_xxxa)        premul_xxxa = sk_default::premul_xxxa<false>;

[mtklein, 2016/01/13 17:04:36]

I'd prefer to add shims to SkSwizzler_opts.h to keep the names the same.  It's
easier to debug/profile if we see premul_swaprb_xxxa than premul_xxxa<1>.

static void       premul_xxxa(...) { premul_xxxa_should_swaprb<false>(...); }
static void premul_swaprb_xxxa(...) { premul_xxxa_should_swaprb<true>(...); }

 _should_swaprb() ought to inline into both, and I can't imagine seeing any
effect on speed even if it didn't.

[msarett, 2016/01/13 18:52:05]

Agreed on all counts!  Done.

+    decltype(       swaprb_xxxa)        swaprb_xxxa = sk_default::swaprb_xxxa;
+    decltype(premul_swaprb_xxxa) premul_swaprb_xxxa = sk_default::premul_xxxa<true>;
 
     // Each Init_foo() is defined in src/opts/SkOpts_foo.cpp.
     void Init_ssse3();