D3D11: Reduce overhead of clearing unused textures.

src/common/mathutil.h

@@ -524,6 +524,17 @@ struct Range
             return start < other.end;
         }
     }
+
+    void extend(T value)
+    {
+        start = value > start ? value : start;
+        end = value < end ? value : end;
+    }
+
+    bool empty() const
+    {
+        return end <= start;
+    }
 };
 
 typedef Range<int> RangeI;

src/libANGLE/renderer/d3d/RendererD3D.cpp

@@ -444,14 +444,7 @@ gl::Error RendererD3D::applyTextures(const gl::Data &data, gl::SamplerType shade
     // Set all the remaining textures to NULL
     size_t samplerCount = (shaderType == gl::SAMPLER_PIXEL) ? data.caps->maxTextureImageUnits
                                                             : data.caps->maxVertexTextureImageUnits;
-    for (size_t samplerIndex = samplerRange; samplerIndex < samplerCount; samplerIndex++)
-    {
-        gl::Error error = setTexture(shaderType, samplerIndex, NULL);
-        if (error.isError())
-        {
-            return error;
-        }
-    }
+    clearTextures(shaderType, samplerRange, samplerCount);
 
     return gl::Error(GL_NO_ERROR);
 }

src/libANGLE/renderer/d3d/RendererD3D.h

@@ -190,6 +190,9 @@ class RendererD3D : public Renderer, public BufferFactoryD3D
     void pushGroupMarker(GLsizei length, const char *marker) override;
     void popGroupMarker() override;
 
+    // In D3D11, faster than calling setTexture a jillion times
+    virtual gl::Error clearTextures(gl::SamplerType samplerType, size_t rangeStart, size_t rangeEnd) = 0;
+
   protected:
     virtual gl::Error drawArrays(const gl::Data &data, GLenum mode, GLsizei count, GLsizei instances, bool usesPointSize) = 0;
     virtual gl::Error drawElements(GLenum mode, GLsizei count, GLenum type, const GLvoid *indices,

src/libANGLE/renderer/d3d/d3d11/Renderer11.cpp

@@ -188,6 +188,45 @@ ANGLEFeatureLevel GetANGLEFeatureLevel(D3D_FEATURE_LEVEL d3dFeatureLevel)
 
 }
 
+void Renderer11::SRVCache::update(size_t resourceIndex, ID3D11ShaderResourceView *srv)
+{
+    ASSERT(resourceIndex < mCurrentSRVs.size());
+    SRVRecord *record = &mCurrentSRVs[resourceIndex];
+
+    record->srv = reinterpret_cast<uintptr_t>(srv);
+    if (srv)
+    {
+        record->resource = reinterpret_cast<uintptr_t>(GetViewResource(srv));
+        srv->GetDesc(&record->desc);
+        mHighestUsedSRV = std::max(resourceIndex + 1, mHighestUsedSRV);
+    }
+    else
+    {
+        record->resource = 0;
+
+        if (resourceIndex + 1 == mHighestUsedSRV)
+        {
+            do
+            {
+                --mHighestUsedSRV;
+            }
+            while (mHighestUsedSRV > 0 &&
+                   mCurrentSRVs[mHighestUsedSRV].srv == 0);
+        }
+    }
+}
+
+void Renderer11::SRVCache::clear()
+{
+    if (mCurrentSRVs.empty())
+    {
+        return;
+    }
+
+    memset(&mCurrentSRVs[0], 0, sizeof(SRVRecord) * mCurrentSRVs.size());
+    mHighestUsedSRV = 0;
+}
+
 Renderer11::Renderer11(egl::Display *display)
     : RendererD3D(display),
       mStateCache(this),
@@ -586,8 +625,8 @@ void Renderer11::initializeDevice()
     mForceSetPixelSamplerStates.resize(rendererCaps.maxTextureImageUnits);
     mCurPixelSamplerStates.resize(rendererCaps.maxTextureImageUnits);
 
-    mCurVertexSRVs.resize(rendererCaps.maxVertexTextureImageUnits);
-    mCurPixelSRVs.resize(rendererCaps.maxTextureImageUnits);
+    mCurVertexSRVs.initialize(rendererCaps.maxVertexTextureImageUnits);
+    mCurPixelSRVs.initialize(rendererCaps.maxTextureImageUnits);
 
     markAllStateDirty();
 
@@ -605,6 +644,9 @@ void Renderer11::initializeDevice()
         angleFeatureLevel = ANGLE_FEATURE_LEVEL_11_1;
     }
 
+    // Initialize cached NULL SRV block
+    mNullSRVs.resize(getRendererCaps().maxTextureImageUnits, nullptr);
+
     ANGLE_HISTOGRAM_ENUMERATION("GPU.ANGLE.D3D11FeatureLevel",
                                 angleFeatureLevel,
                                 NUM_ANGLE_FEATURE_LEVELS);
@@ -2184,8 +2226,8 @@ void Renderer11::markAllStateDirty()
     // We reset the current SRV data because it might not be in sync with D3D's state
     // anymore. For example when a currently used SRV is used as an RTV, D3D silently
     // remove it from its state.
-    memset(mCurVertexSRVs.data(), 0, sizeof(SRVRecord) * mCurVertexSRVs.size());
-    memset(mCurPixelSRVs.data(), 0, sizeof(SRVRecord) * mCurPixelSRVs.size());
+    mCurVertexSRVs.clear();
+    mCurPixelSRVs.clear();
 
     ASSERT(mForceSetVertexSamplerStates.size() == mCurVertexSRVs.size());
     for (size_t vsamplerId = 0; vsamplerId < mForceSetVertexSamplerStates.size(); ++vsamplerId)
@@ -3675,7 +3717,7 @@ void Renderer11::setShaderResource(gl::SamplerType shaderType, UINT resourceSlot
     auto &currentSRVs = (shaderType == gl::SAMPLER_VERTEX ? mCurVertexSRVs : mCurPixelSRVs);
 
     ASSERT(static_cast<size_t>(resourceSlot) < currentSRVs.size());
-    auto &record = currentSRVs[resourceSlot];
+    const SRVRecord &record = currentSRVs[resourceSlot];
 
     if (record.srv != reinterpret_cast<uintptr_t>(srv))
     {
@@ -3688,16 +3730,7 @@ void Renderer11::setShaderResource(gl::SamplerType shaderType, UINT resourceSlot
             mDeviceContext->PSSetShaderResources(resourceSlot, 1, &srv);
         }
 
-        record.srv = reinterpret_cast<uintptr_t>(srv);
-        if (srv)
-        {
-            record.resource = reinterpret_cast<uintptr_t>(GetViewResource(srv));
-            srv->GetDesc(&record.desc);
-        }
-        else
-        {
-            record.resource = 0;
-        }
+        currentSRVs.update(resourceSlot, srv);
     }
 }
 
@@ -3706,4 +3739,38 @@ void Renderer11::createAnnotator()
     mAnnotator = new DebugAnnotator11();
 }
 
+gl::Error Renderer11::clearTextures(gl::SamplerType samplerType, size_t rangeStart, size_t rangeEnd)
+{
+    if (rangeStart == rangeEnd)
+    {
+        return gl::Error(GL_NO_ERROR);
+    }
+
+    auto &currentSRVs = (samplerType == gl::SAMPLER_VERTEX ? mCurVertexSRVs : mCurPixelSRVs);
+
+    gl::Range<size_t> clearRange(rangeStart, rangeStart);
+    clearRange.extend(std::min(rangeEnd, currentSRVs.highestUsed()));
+
+    if (clearRange.empty())
+    {
+        return gl::Error(GL_NO_ERROR);
+    }
+
+    if (samplerType == gl::SAMPLER_VERTEX)
+    {
+        mDeviceContext->VSSetShaderResources(rangeStart, rangeEnd - rangeStart, &mNullSRVs[0]);
+    }
+    else
+    {
+        mDeviceContext->PSSetShaderResources(rangeStart, rangeEnd - rangeStart, &mNullSRVs[0]);
+    }
+
+    for (size_t samplerIndex = rangeStart; samplerIndex < rangeEnd; ++samplerIndex)
+    {
+        currentSRVs.update(samplerIndex, nullptr);
+    }
+
+    return gl::Error(GL_NO_ERROR);
+}
+
 }

src/libANGLE/renderer/d3d/d3d11/Renderer11.h

@@ -266,6 +266,7 @@ class Renderer11 : public RendererD3D
 
   protected:
     void createAnnotator() override;
+    gl::Error clearTextures(gl::SamplerType samplerType, size_t rangeStart, size_t rangeEnd) override;
 
   private:
     void generateCaps(gl::Caps *outCaps, gl::TextureCapsMap *outTextureCaps, gl::Extensions *outExtensions) const override;
@@ -320,8 +321,41 @@ class Renderer11 : public RendererD3D
         uintptr_t resource;
         D3D11_SHADER_RESOURCE_VIEW_DESC desc;
     };
-    std::vector<SRVRecord> mCurVertexSRVs;
-    std::vector<SRVRecord> mCurPixelSRVs;
+
+    // A cache of current SRVs that also tracks the highest 'used' (non-NULL) SRV
+    // We might want to investigate a more robust approach that is also fast when there's
+    // a large gap between used SRVs (e.g. if SRV 0 and 7 are non-NULL, this approach will
+    // waste time on SRVs 1-6.)
+    class SRVCache : angle::NonCopyable
+    {
+      public:
+        SRVCache()
+            : mHighestUsedSRV(0)
+        {
+        }
+
+        void initialize(size_t size)
+        {
+            mCurrentSRVs.resize(size);
+        }
+
+        size_t size() const { return mCurrentSRVs.size(); }
+        size_t highestUsed() const { return mHighestUsedSRV; }
+
+        const SRVRecord &operator[](size_t index) const { return mCurrentSRVs[index]; }
+        void clear();
+        void update(size_t resourceIndex, ID3D11ShaderResourceView *srv);
+
+      private:
+        std::vector<SRVRecord> mCurrentSRVs;
+        size_t mHighestUsedSRV;
+    };
+
+    SRVCache mCurVertexSRVs;
+    SRVCache mCurPixelSRVs;
+
+    // A block of NULL pointers, cached so we don't re-allocate every draw call
+    std::vector<ID3D11ShaderResourceView*> mNullSRVs;
 
     // Currently applied blend state
     bool mForceSetBlendState;

src/libANGLE/renderer/d3d/d3d9/Renderer9.cpp

@@ -2942,4 +2942,19 @@ void Renderer9::createAnnotator()
     mAnnotator = new DebugAnnotator9();
 }
 
+gl::Error Renderer9::clearTextures(gl::SamplerType samplerType, size_t rangeStart, size_t rangeEnd)
+{
+    // TODO(jmadill): faster way?
+    for (size_t samplerIndex = rangeStart; samplerIndex < rangeEnd; samplerIndex++)
+    {
+        gl::Error error = setTexture(samplerType, samplerIndex, nullptr);
+        if (error.isError())
+        {
+            return error;
+        }
+    }
+
+    return gl::Error(GL_NO_ERROR);
+}
+
 }

src/libANGLE/renderer/d3d/d3d9/Renderer9.h

@@ -235,6 +235,7 @@ class Renderer9 : public RendererD3D
 
   protected:
     void createAnnotator() override;
+    gl::Error clearTextures(gl::SamplerType samplerType, size_t rangeStart, size_t rangeEnd) override;
 
   private:
     void generateCaps(gl::Caps *outCaps, gl::TextureCapsMap *outTextureCaps, gl::Extensions *outExtensions) const override;