Caching the result of readPixelsSupported

The call was calling GR_GL_GetIntegerv 2 times for each readPixels
and thus was causing a loss of performance

(resubmit of issue 344793008)

Benchmark url: http://packages.gkny.fr/tst/index.html

BUG=skia:2681

Committed: https://skia.googlesource.com/skia/+/753a2964afe5661ce9b2a8ca77ca9d0aabd3173c

Committed: https://skia.googlesource.com/skia/+/8339371f1ec3c57a0741932fd96bff32c53d4e54

Review URL: https://codereview.chromium.org/364193004

4 files changed, 159 insertions, 12 deletions

diff --git a/src/core/SkTHashCache.h b/src/core/SkTHashCache.h
new file mode 100644
index 0000000000..cfee9722f2
--- /dev/null
+++ b/src/core/SkTHashCache.h
@@ -0,0 +1,77 @@
+/*
+ * Copyright 2014 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkTHASHCACHE_DEFINED
+#define	SkTHASHCACHE_DEFINED
+
+#include "SkTypes.h"
+#include "SkTDynamicHash.h"
+
+template <typename T,
+typename Key,
+typename Traits = T,
+int kGrowPercent = 75 >
+class SkTHashCache : public SkNoncopyable {
+public:
+
+    SkTHashCache() {
+        this->reset();
+    }
+
+    ~SkTHashCache() {
+        this->clear();
+    }
+
+    T* find(const Key& key) const {
+        return fDict->find(key);
+    }
+
+    /**
+     *  If element already in cache, return immediately the cached value
+     */
+    T& add(const T& add) {
+        Key key = Traits::GetKey(add);
+        if (T* val = this->find(key)) {
+            return *val;
+        }
+
+        T* element = SkNEW_ARGS(T, (add));
+
+        fDict->add(element);
+
+        return *element;
+    }
+
+    int size() const {
+        return fDict->count();
+    }
+
+    void reset() {
+        this->clear();
+
+        fDict.reset(SkNEW(DictType));
+    }
+
+private:
+    typedef SkTDynamicHash<T, Key, Traits, kGrowPercent> DictType;
+
+    void clear() {
+        if (fDict.get()) {
+            typename DictType::Iter it(fDict.get());
+
+            while (!it.done()) {
+                SkDELETE(&(*it));
+                ++it;
+            }
+        }
+    }
+
+    SkAutoTDelete<DictType> fDict;
+};
+
+#endif	/* SkHASHCACHE_DEFINED */
+
diff --git a/src/gpu/gl/GrGLCaps.cpp b/src/gpu/gl/GrGLCaps.cpp
index dc4ef4c160..2fe38479de 100644
--- a/src/gpu/gl/GrGLCaps.cpp
+++ b/src/gpu/gl/GrGLCaps.cpp
@@ -50,6 +50,8 @@ void GrGLCaps::reset() {
     fFBFetchSupport = false;
     fFBFetchColorName = NULL;
     fFBFetchExtensionString = NULL;
+
+    fReadPixelsSupportedCache.reset();
 }
 
 GrGLCaps::GrGLCaps(const GrGLCaps& caps) : GrDrawTargetCaps() {
@@ -582,7 +584,7 @@ void GrGLCaps::initConfigTexturableTable(const GrGLContextInfo& ctxInfo, const G
     }
 
     // Check for ASTC
-    fConfigTextureSupport[kASTC_12x12_GrPixelConfig] = 
+    fConfigTextureSupport[kASTC_12x12_GrPixelConfig] =
         ctxInfo.hasExtension("GL_KHR_texture_compression_astc_hdr") ||
         ctxInfo.hasExtension("GL_KHR_texture_compression_astc_ldr") ||
         ctxInfo.hasExtension("GL_OES_texture_compression_astc");
@@ -601,9 +603,9 @@ void GrGLCaps::initConfigTexturableTable(const GrGLContextInfo& ctxInfo, const G
     fConfigTextureSupport[kRGBA_float_GrPixelConfig] = hasFPTextures;
 }
 
-bool GrGLCaps::readPixelsSupported(const GrGLInterface* intf,
-                                   GrGLenum format,
-                                   GrGLenum type) const {
+bool GrGLCaps::doReadPixelsSupported(const GrGLInterface* intf,
+                                     GrGLenum format,
+                                     GrGLenum type) const {
     if (GR_GL_RGBA == format && GR_GL_UNSIGNED_BYTE == type) {
         // ES 2 guarantees this format is supported
         return true;
@@ -630,6 +632,26 @@ bool GrGLCaps::readPixelsSupported(const GrGLInterface* intf,
     return (GrGLenum)otherFormat == format && (GrGLenum)otherType == type;
 }
 
+bool GrGLCaps::readPixelsSupported(const GrGLInterface* intf,
+                                   GrGLenum format,
+                                   GrGLenum type,
+                                   GrGLenum currFboFormat) const {
+
+    ReadPixelsSupportedFormats::Key key = {format, type, currFboFormat};
+
+    ReadPixelsSupportedFormats* cachedValue = fReadPixelsSupportedCache.find(key);
+
+    if (NULL == cachedValue) {
+        bool value = doReadPixelsSupported(intf, format, type);
+        ReadPixelsSupportedFormats newValue(key, value);
+        fReadPixelsSupportedCache.add(newValue);
+
+        return newValue.value();
+    }
+
+    return cachedValue->value();
+}
+
 void GrGLCaps::initFSAASupport(const GrGLContextInfo& ctxInfo, const GrGLInterface* gli) {
 
     fMSFBOType = kNone_MSFBOType;
diff --git a/src/gpu/gl/GrGLCaps.h b/src/gpu/gl/GrGLCaps.h
index 887e2e9e2a..eb56798fda 100644
--- a/src/gpu/gl/GrGLCaps.h
+++ b/src/gpu/gl/GrGLCaps.h
@@ -11,8 +11,9 @@
 
 #include "GrDrawTargetCaps.h"
 #include "GrGLStencilBuffer.h"
+#include "SkChecksum.h"
+#include "SkTHashCache.h"
 #include "SkTArray.h"
-#include "SkTDArray.h"
 
 class GrGLContextInfo;
 
@@ -252,7 +253,8 @@ public:
     /// Does ReadPixels support the provided format/type combo?
     bool readPixelsSupported(const GrGLInterface* intf,
                              GrGLenum format,
-                             GrGLenum type) const;
+                             GrGLenum type,
+                             GrGLenum currFboFormat) const;
 
     bool isCoreProfile() const { return fIsCoreProfile; }
 
@@ -323,6 +325,10 @@ private:
     void initConfigRenderableTable(const GrGLContextInfo&);
     void initConfigTexturableTable(const GrGLContextInfo&, const GrGLInterface*);
 
+    bool doReadPixelsSupported(const GrGLInterface* intf,
+                                   GrGLenum format,
+                                   GrGLenum type) const;
+
     // tracks configs that have been verified to pass the FBO completeness when
     // used as a color attachment
     VerifiedColorConfigs fVerifiedColorConfigs;
@@ -366,6 +372,46 @@ private:
     const char* fFBFetchColorName;
     const char* fFBFetchExtensionString;
 
+    class ReadPixelsSupportedFormats {
+    public:
+        struct Key {
+            GrGLenum fFormat;
+            GrGLenum fType;
+            GrGLenum fFboFormat;
+
+            bool operator==(const Key& rhs) const {
+                return fFormat == rhs.fFormat
+                        && fType == rhs.fType
+                        && fFboFormat == rhs.fFboFormat;
+            }
+
+            uint32_t getHash() const {
+                return SkChecksum::Murmur3(reinterpret_cast<const uint32_t*>(this), sizeof(*this));
+            }
+        };
+
+        ReadPixelsSupportedFormats(Key key, bool value) : fKey(key), fValue(value) {
+        }
+
+        static const Key& GetKey(const ReadPixelsSupportedFormats& element) {
+            return element.fKey;
+        }
+
+        static uint32_t Hash(const Key& key) {
+            return key.getHash();
+        }
+
+        bool value() const {
+            return fValue;
+        }
+    private:
+        Key fKey;
+        bool fValue;
+    };
+
+    mutable SkTHashCache<ReadPixelsSupportedFormats,
+                         ReadPixelsSupportedFormats::Key> fReadPixelsSupportedCache;
+
     typedef GrDrawTargetCaps INHERITED;
 };
 
diff --git a/src/gpu/gl/GrGpuGL.cpp b/src/gpu/gl/GrGpuGL.cpp
index 8be1118d64..a4665f9b3e 100644
--- a/src/gpu/gl/GrGpuGL.cpp
+++ b/src/gpu/gl/GrGpuGL.cpp
@@ -178,8 +178,6 @@ void GrGpuGL::contextAbandoned() {
 }
 
 ///////////////////////////////////////////////////////////////////////////////
-
-
 GrPixelConfig GrGpuGL::preferredReadPixelsConfig(GrPixelConfig readConfig,
                                                  GrPixelConfig surfaceConfig) const {
     if (GR_GL_RGBA_8888_PIXEL_OPS_SLOW && kRGBA_8888_GrPixelConfig == readConfig) {
@@ -190,9 +188,13 @@ GrPixelConfig GrGpuGL::preferredReadPixelsConfig(GrPixelConfig readConfig,
         // Mesa 3D takes a slow path on when reading back  BGRA from an RGBA surface and vice-versa.
         // Perhaps this should be guarded by some compiletime or runtime check.
         return surfaceConfig;
-    } else if (readConfig == kBGRA_8888_GrPixelConfig &&
-               !this->glCaps().readPixelsSupported(this->glInterface(),
-                                                   GR_GL_BGRA, GR_GL_UNSIGNED_BYTE)) {
+    } else if (readConfig == kBGRA_8888_GrPixelConfig
+            && !this->glCaps().readPixelsSupported(
+                this->glInterface(),
+                GR_GL_BGRA,
+                GR_GL_UNSIGNED_BYTE,
+                surfaceConfig
+            )) {
         return kRGBA_8888_GrPixelConfig;
     } else {
         return readConfig;
@@ -707,7 +709,7 @@ bool GrGpuGL::uploadTexData(const GrGLTexture::Desc& desc,
 }
 
 // TODO: This function is using a lot of wonky semantics like, if width == -1
-// then set width = desc.fWdith ... blah. A better way to do it might be to 
+// then set width = desc.fWdith ... blah. A better way to do it might be to
 // create a CompressedTexData struct that takes a desc/ptr and figures out
 // the proper upload semantics. Then users can construct this function how they
 // see fit if they want to go against the "standard" way to do it.