Add specialized sRGB blitter for SkOpts

bench/SkBlend_optsBench.cpp

Index: bench/SkBlend_optsBench.cpp
diff --git a/bench/SkBlend_optsBench.cpp b/bench/SkBlend_optsBench.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..9d20ea332805b4e61fa84c71bdc624dcc7c8b15b
--- /dev/null
+++ b/bench/SkBlend_optsBench.cpp
@@ -0,0 +1,165 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include <tuple>
+
+#include "Benchmark.h"
+#include "Resources.h"
+#include "SkCpu.h"
+#include "SkImage.h"
+#include "SkImage_Base.h"
+#include "SkNx.h"
+#include "SkOpts.h"
+#include "SkString.h"
+
+#define INNER_LOOPS 10
+
+namespace sk_default {
+extern void brute_force_srcover_srgb_srgb(
+    uint32_t* dst, const uint32_t* const srcStart, int ndst, const int nsrc);
+}
+
+class SrcOverVSkOptsBruteForce {
+public:
+    static SkString Name() { return SkString{"VSkOptsBruteForce"}; }
+    static bool WorksOnCpu() { return true; }
+    static void BlendN(uint32_t* dst, int count, const uint32_t* src) {
+        sk_default::brute_force_srcover_srgb_srgb(dst, src, count, count);
+    }
+};
+
+namespace sk_default {
+extern void trivial_srcover_srgb_srgb(
+    uint32_t* dst, const uint32_t* const srcStart, int ndst, const int nsrc);
+}
+
+class SrcOverVSkOptsTrivial {
+public:
+    static SkString Name() { return SkString{"VSkOptsTrivial"}; }
+    static bool WorksOnCpu() { return true; }
+    static void BlendN(uint32_t* dst, int count, const uint32_t* src) {
+        sk_default::trivial_srcover_srgb_srgb(dst, src, count, count);
+    }
+};
+
+namespace sk_default {
+extern void best_non_simd_srcover_srgb_srgb(
+    uint32_t* dst, const uint32_t* const srcStart, int ndst, const int nsrc);
+}
+
+class SrcOverVSkOptsNonSimdCore {
+public:
+    static SkString Name() { return SkString{"VSkOptsNonSimdCore"}; }
+    static bool WorksOnCpu() { return true; }
+    static void BlendN(uint32_t* dst, int count, const uint32_t* src) {
+        sk_default::best_non_simd_srcover_srgb_srgb(dst, src, count, count);
+    }
+};
+
+namespace sk_default {
+extern void srcover_srgb_srgb(
+    uint32_t* dst, const uint32_t* const srcStart, int ndst, const int nsrc);
+}
+
+class SrcOverVSkOptsDefault {
+public:
+    static SkString Name() { return SkString{"VSkOptsDefault"}; }
+    static bool WorksOnCpu() { return true; }
+    static void BlendN(uint32_t* dst, int count, const uint32_t* src) {
+        sk_default::srcover_srgb_srgb(dst, src, count, count);
+    }
+};
+
+namespace sk_sse41 {
+   extern void srcover_srgb_srgb(
+       uint32_t* dst, const uint32_t* const srcStart, int ndst, const int nsrc);
+}
+
+class SrcOverVSkOptsSSE41 {
+public:
+    static SkString Name() { return SkString{"VSkOptsSSE41"}; }
+    static bool WorksOnCpu() { return SkCpu::Supports(SkCpu::SSE41); }
+    static void BlendN(uint32_t* dst, int count, const uint32_t* src) {
+        sk_sse41::srcover_srgb_srgb(dst, src, count, count);
+    }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <typename Blender>
+class LinearSrcOverBench : public Benchmark {
+public:
+    LinearSrcOverBench(const char* fileName) {
+        fName = "LinearSrcOver";
+        fName.append(fileName);
+        fName.append(Blender::Name());
+
+        sk_sp<SkImage> image = GetResourceAsImage(fileName);
+        SkBitmap bm;
+        if (!as_IB(image)->getROPixels(&bm)) {
+            SkFAIL("Could not read resource");
+        }
+        bm.peekPixels(&fPixmap);
+        fDst.reset(fPixmap.rowBytesAsPixels());
+    }
+
+protected:
+    bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
+    const char* onGetName() override { return fName.c_str(); }
+    void onDraw(int loops, SkCanvas*) override {
+        if (!Blender::WorksOnCpu()) return;

[f(malita), 2016/05/06 17:43:56]

mtklein is probably the right person for this question, but what's the effect of
returning early from onDraw()?  Is bench going to attempt to time a no-op loop?

I would think isSuitableFor() is a better place to veto the bench.

[herb_g, 2016/05/06 20:57:45]

Done.

Thanks for the insight.

+        SkPixmap& pixels = fPixmap;

[f(malita), 2016/05/06 17:43:56]

Nit: why alias fPixmap?

[herb_g, 2016/05/06 20:57:45]

Done.

+        SkASSERT(pixels.colorType() == kRGBA_8888_SkColorType);
+
+        const int width = pixels.rowBytesAsPixels();
+        fCount = width;

[f(malita), 2016/05/06 17:43:56]

Nit: looks like this could live in the ctor?

[herb_g, 2016/05/06 20:57:45]

Done.

+
+        for (int i = 0; i < loops * INNER_LOOPS; ++i) {
+            const uint32_t* src = pixels.addr32();
+            for (int y = 0; y < pixels.height(); y++) {
+                Blender::BlendN(fDst.get(), width, src);
+                src += width;
+            }
+        }
+    }
+
+    void onPostDraw(SkCanvas*) override {
+        // Make sure the compiler does not optimize away the operation.
+        volatile uint32_t v = 0;
+        for (int i = 0; i < fCount; i++) {
+            v ^= fDst[i];
+        }
+    }
+
+private:
+    int fCount;
+    SkAutoTArray<uint32_t> fDst;
+    SkString fName;
+    SkPixmap fPixmap;
+
+    typedef Benchmark INHERITED;
+};
+
+#if defined(SK_CPU_X86) && !defined(SK_BUILD_FOR_IOS)
+#define BENCHES(fileName) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsBruteForce>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsTrivial>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsNonSimdCore>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsDefault>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsSSE41>(fileName); )
+#else
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsBruteForce>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsTrivial>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsNonSimdCore>(fileName); ) \
+DEF_BENCH( return new LinearSrcOverBench<SrcOverVSkOptsDefault>(fileName); )
+#endif
+
+BENCHES("yellow_rose.png")
+BENCHES("baby_tux.png")
+BENCHES("plane.png")
+BENCHES("mandrill_512.png")
+BENCHES("iconstrip.png")