Implement instanced rendering for simple shapes

src/gpu/GrInstancedRendering.h

Index: src/gpu/GrInstancedRendering.h
diff --git a/src/gpu/GrInstancedRendering.h b/src/gpu/GrInstancedRendering.h
new file mode 100644
index 0000000000000000000000000000000000000000..6d0a52bbc24cc33e539295fdda507e0e4238a2e7
--- /dev/null
+++ b/src/gpu/GrInstancedRendering.h
@@ -0,0 +1,208 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef GrInstancedRendering_DEFINED
+#define GrInstancedRendering_DEFINED
+
+#include "GrAllocator.h"
+#include "GrInstancedRenderingTypes.h"
+#include "batches/GrDrawBatch.h"
+
+class GrInstanceProcessor;
+class GrResourceProvider;
+
+/**
+ * This class serves as a centralized clearinghouse for instanced rendering. It accumulates data for

[bsalomon, 2016/04/25 13:22:38]

If we had a completely different instanced primitive type that didn't batch with
(d)rrect family of shapes would we use this same object? I'm wondering whether
this is really a InstancedRRectRenderer. I know we did it with path rendering
but having class names that are verbs strikes me as a little odd.

[Chris Dalton, 2016/04/25 17:01:19]

My vision for this class is yes! The basic geometry is so small and simple,
there's no reason we can't fill the vertex and index buffers with explicit
polygons, beziers, etc.

Right now the shaders only rasterize arcs and rectangles, but gl_SampleMask is
PERFECT for Loop-Blinn. We could (and should) easily add support for various
types of non-overlapping shapes.

Then a lot of simple-ish path rendering can be achieved by selecting the right
index range from a pre-generated library of simple shapes, and pushing a few
parameters to a texel buffer.

[bsalomon, 2016/04/27 14:10:32]

Ok, let's keep the name then

+ * instanced draws into one location, and creates special batches that pull from this data. The
+ * nature of instanced rendering allows these batches to combine well and render efficiently.
+ *
+ * During a flush, this class assembles the accumulated draw data into a single vertex and texel
+ * buffer, and its subclass draws the batches using backend-specific instanced rendering APIs.
+ *
+ * This class is responsible for the CPU side of instanced rendering. Shaders are implemented by
+ * GrInstanceProcessor.
+ */
+class GrInstancedRendering : public SkNoncopyable, protected GrInstancedRenderingTypes {
+public:
+    virtual ~GrInstancedRendering() { SkASSERT(State::kRecordingShapes == fState); }
+
+    GrGpu* gpu() const { return fGpu; }
+

[bsalomon, 2016/04/25 13:22:38]

I think these could use a comment

[Chris Dalton, 2016/04/26 19:18:07]

Done.

+    enum Flags {
+        kStencilWrite_Flag       = (1 << 0),
+        kStencilBufferMSAA_Flag  = (1 << 1),
+        kColorWrite_Flag         = (1 << 2),
+        kColorBufferMSAA_Flag    = (1 << 3),
+        /**
+         * This should not be set if the fragment shader uses derivatives, automatic mipmap LOD, or
+         * other features that depend on neighboring pixels.
+         */
+        kUseDiscard_Flag         = (1 << 4)
+    };
+
+    /**
+     * These methods record a new instanced draw and return a batch that can render it. The client
+     * must call commitToGpu() before attempting to draw batches returned by this class. After

[bsalomon, 2016/04/25 13:22:38]

Maybe before attempting to flush batches returned by this class?

[Chris Dalton, 2016/04/26 19:18:07]

Done.

+     * commitToGpu(), it becomes invalid to record new draws until a subsequent call to restart().
+     */
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordRect(const SkRect&, const SkMatrix&, GrColor,

[bsalomon, 2016/04/25 13:22:38]

It seems to me like the "recording" aspect of these calls is relevant to the
implementation, but from the callers' perspectives these methods are factory
functions. Perhaps createFooBatch?

[Chris Dalton, 2016/04/25 17:01:19]

When naming these methods I did think this through, and decided I liked "record"
better than "create" :)

My rationale was that "record" signifies better that this class is internally
making a record for the desired shape, and the batch that gets returned is
literally just an index to that record. (Thus reinforcing the understanding that
this class is what actually does the rendering, not the batch, and the batch
depends on this class).

A traditional factory-style "create" name implies to me that the batch is more
autonomous, which is not the case here.

[Chris Dalton, 2016/04/26 19:18:07]

Updated the comment.

[bsalomon, 2016/04/27 14:10:32]

Ok, that makes sense to me.

+                                                  bool antialias, uint32_t flags, bool* useHWAA);
+
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordRect(const SkRect&, const SkMatrix&, GrColor,
+                                                  const SkRect& localRect, bool antialias,
+                                                  uint32_t flags, bool* useHWAA);
+
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordRect(const SkRect&, const SkMatrix&, GrColor,
+                                                  const SkMatrix& localMatrix, bool antialias,
+                                                  uint32_t flags, bool* useHWAA);
+
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordOval(const SkRect&, const SkMatrix&, GrColor,
+                                                  bool antialias, uint32_t flags, bool* useHWAA);
+
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordRRect(const SkRRect&, const SkMatrix&, GrColor,
+                                                   bool antialias, uint32_t flags, bool* useHWAA);
+
+    GrDrawBatch* SK_WARN_UNUSED_RESULT recordDRRect(const SkRRect& outer, const SkRRect& inner,
+                                                    const SkMatrix&, GrColor, bool antialias,
+                                                    uint32_t flags, bool* useHWAA);
+
+    /**
+     * Commits all recorded draws to GPU memory and allows the client to begin drawing the batches
+     * created by this class.
+     */
+    void commitToGpu(GrResourceProvider*);

[bsalomon, 2016/04/25 13:22:38]

Wonder if something like prepareToFlush or willFlush would more clearly tie this
into flushing the GrDTs

[Chris Dalton, 2016/04/26 19:18:08]

Done.

+
+    /**
+     * Called once the batches created previously by this class have all been released. Allows the
+     * client to begin recording draws again.
+     */
+    void restart();

[bsalomon, 2016/04/25 13:22:38]

Minor but I think we more commonly use reset()

[Chris Dalton, 2016/04/25 17:01:19]

I also thought this one through. I decided to go with "restart" specifically
because "reset" in Skia so commonly means "destroy all objects and reset
everything back to zero".

That's not the case here. This method simply releases a few scratch buffers that
were tied to the current flush and puts the class back into recording mode.

I'm not attached to the name "record" though, I'm just saying "reset" might
mislead somebody to want to call this method instead of clearGpuResources (name
in question), which is what you should actually call to destroy all gpu memory.

[Chris Dalton, 2016/04/26 19:18:07]

How's endFlush()?

[bsalomon, 2016/04/27 14:10:32]

sgtm

+
+    enum class ClearType {
+        kDestroy,
+        kAbandon
+    };
+
+    /**
+     * Clears all GPU resources, including those that are held long term. They will be lazily
+     * reinitialized if the class begins to be used again.
+     */
+    void clearGpuResources(ClearType);

[bsalomon, 2016/04/25 13:22:38]

freeGpuResources? again just trying to mentally map into when it is used by
naming it something like the higher level calls that use it.

[Chris Dalton, 2016/04/25 17:01:19]

clearGpuResources(ClearType::kDestroy) would be the equivalent of
freeGpuResources().

clearGpuResources(ClearType::kAbandon) would be the equivalent of
abandonGpuResources().

I guess I got "clear" from the Blink-style naming conventions on smart pointers.
In this context though "clear" is a loaded name. How about:

resetGpuResources(ResetType)

[Chris Dalton, 2016/04/26 19:18:07]

Done.

+
+protected:
+    class Batch : public GrDrawBatch {
+    public:
+        virtual ~Batch() { fInUse = false; } // fInUse will continue to be accessed.
+
+        const char* name() const override { return "Instanced Batch"; }
+
+        void computePipelineOptimizations(GrInitInvariantOutput* color,
+                                          GrInitInvariantOutput* coverage,
+                                          GrBatchToXPOverrides*) const override;
+
+    protected:
+        Batch(uint32_t classID, GrInstancedRendering* ir, AntialiasMode aa, int instanceIdx)
+            : INHERITED(classID),
+              fInstancedRendering(ir),
+              fAntialiasMode(aa),
+              fFirstInstanceIdx(instanceIdx),
+              fInUse(true) {
+#ifdef SK_DEBUG
+            fIsCombined = false;
+#endif
+        }
+
+        void initBatchTracker(const GrXPOverridesForBatch&) override;
+
+        void onPrepare(GrBatchFlushState*) override {}
+        void onDraw(GrBatchFlushState*) override;
+        void onDelete() const override;
+
+        GrInstancedRendering* const   fInstancedRendering;
+        const AntialiasMode           fAntialiasMode;
+        const int                     fFirstInstanceIdx;
+        BatchTracker                  fTracker;
+        bool                          fInUse;
+#ifdef SK_DEBUG
+        bool                          fIsCombined;
+#endif
+
+        typedef GrDrawBatch INHERITED;
+
+        friend class GrInstancedRendering;
+    };
+
+    class BatchAllocator : public GrAllocator {
+    public:
+        BatchAllocator(size_t sizeofBatchClass)
+            : INHERITED(sizeofBatchClass, kBatchesPerBlock, nullptr) {
+            fFirstBlock = sk_malloc_throw(kBatchesPerBlock * sizeofBatchClass);
+            this->setInitialBlock(fFirstBlock);
+        }
+
+        ~BatchAllocator() {
+            sk_free(fFirstBlock);
+        }
+
+    private:
+        enum { kBatchesPerBlock = 128 };
+
+        void* fFirstBlock;
+
+        typedef GrAllocator INHERITED;
+    };
+
+    GrInstancedRendering(GrGpu* gpu, uint32_t supportedAAModes, size_t sizeofBatchClass);
+
+    const GrBuffer* vertexBuffer() const { SkASSERT(fVertexBuffer); return fVertexBuffer; }
+    const GrBuffer* indexBuffer() const { SkASSERT(fIndexBuffer); return fIndexBuffer; }
+    const GrBuffer* instanceBuffer() const { SkASSERT(fInstanceBuffer); return fInstanceBuffer; }
+    const BatchAllocator* batchAllocator() const { return &fBatchAllocator; }
+
+    virtual void onCommitToGpu(GrResourceProvider*) = 0;
+    virtual void onDraw(const GrPipeline&, const GrInstanceProcessor&, const Batch*) = 0;
+    virtual void onRestart() = 0;
+    virtual void onClearGpuResources(ClearType) = 0;
+
+#ifdef SK_DEBUG
+    int fInUseBatchCount;
+#endif
+
+private:
+    enum class State {
+        kRecordingShapes,
+        kDrawingBatches
+    };
+
+    Batch* SK_WARN_UNUSED_RESULT recordShape(ShapeType, const SkRect& bounds,
+                                             const SkMatrix& viewMatrix, GrColor,
+                                             const SkRect& localRect, bool antialias,
+                                             uint32_t flags, bool* requireHWAA);
+
+    bool selectAntialiasMode(const SkMatrix& viewMatrix, bool antialias, uint32_t flags,
+                             AntialiasMode*, bool* requireHWAA);
+
+    void appendRRectParams(const SkRRect&, BatchTracker*);
+    void appendParamsTexel(const SkScalar* vals, int count);
+    void appendParamsTexel(SkScalar x, SkScalar y, SkScalar z, SkScalar w);
+    void appendParamsTexel(SkScalar x, SkScalar y, SkScalar z);
+
+    virtual Batch* constructBatch(void* storage, AntialiasMode, int instanceIdx) = 0;
+
+    const SkAutoTUnref<GrGpu>            fGpu;
+    const uint32_t                       fSupportedAAModes;
+    State                                fState;
+    SkSTArray<1024, Instance, true>      fInstances;
+    SkSTArray<1024, ParamsTexel, true>   fParams;
+    BatchAllocator                       fBatchAllocator;
+    SkAutoTUnref<const GrBuffer>         fVertexBuffer;
+    SkAutoTUnref<const GrBuffer>         fIndexBuffer;
+    SkAutoTUnref<const GrBuffer>         fInstanceBuffer;
+    SkAutoTUnref<GrBuffer>               fParamsBuffer;
+};
+
+#endif