Optimized premultiplying swizzles for NEON

src/core/SkOpts.cpp

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkOnce.h"
 #include "SkOpts.h"
 
 #define SK_OPTS_NS sk_default
 #include "SkBlitMask_opts.h"
 #include "SkBlitRow_opts.h"
 #include "SkBlurImageFilter_opts.h"
 #include "SkColorCubeFilter_opts.h"
 #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"
 
 #if defined(SK_CPU_X86) && !defined(SK_BUILD_FOR_IOS)
     #if defined(SK_BUILD_FOR_WIN32)
         #include <intrin.h>
         static void cpuid (uint32_t abcd[4]) { __cpuid  ((int*)abcd, 1);    }
         static void cpuid7(uint32_t abcd[4]) { __cpuidex((int*)abcd, 7, 0); }
         static uint64_t xgetbv(uint32_t xcr) { return _xgetbv(xcr); }
@@ skipping 13 matching lines @@
             return (uint64_t)(edx) << 32 | eax;
         }
     #endif
 #elif !defined(SK_ARM_HAS_NEON)      && \
        defined(SK_CPU_ARM32)         && \
        defined(SK_BUILD_FOR_ANDROID) && \
       !defined(SK_BUILD_FOR_ANDROID_FRAMEWORK)
     #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
     // CPU-specialized, e.g. a typical x86-64 machine might start with SSE2 defaults.
     // They'll still get a chance to be replaced with even better ones, e.g. using SSE4.1.
     decltype(rsqrt)                     rsqrt = sk_default::rsqrt;
     decltype(memset16)               memset16 = sk_default::memset16;
     decltype(memset32)               memset32 = sk_default::memset32;
     decltype(create_xfermode) create_xfermode = sk_default::create_xfermode;
     decltype(color_cube_filter_span) color_cube_filter_span = sk_default::color_cube_filter_span;
@@ skipping 11 matching lines @@
     decltype(fill_block_dimensions) fill_block_dimensions = sk_default::fill_block_dimensions;
 
     decltype(blit_mask_d32_a8) blit_mask_d32_a8 = sk_default::blit_mask_d32_a8;
 
     decltype(blit_row_color32) blit_row_color32 = sk_default::blit_row_color32;
 
     decltype(matrix_translate)       matrix_translate       = sk_default::matrix_translate;
     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();
     void Init_sse41();
     void Init_sse42() {}
     void Init_avx();
     void Init_avx2() {}
     void Init_neon();
 
     static void init() {
@@ skipping 26 matching lines @@
 
     SK_DECLARE_STATIC_ONCE(gInitOnce);
     void Init() { SkOnce(&gInitOnce, init); }
 
 #if SK_ALLOW_STATIC_GLOBAL_INITIALIZERS
     static struct AutoInit {
         AutoInit() { Init(); }
     } gAutoInit;
 #endif
 }