Reland "Reland "Add a tessellation mode that uses indirect draws""

This is a reland of 477512c0b3676347d163f3531d2e9df420adb063

Original change's description:
> Reland "Add a tessellation mode that uses indirect draws"
>
> This is a reland of 02d7cf79beec7f9fe0af4d0c4ad764264556727b
>
> Original change's description:
> > Add a tessellation mode that uses indirect draws
> >
> > This mode is oftentimes faster than tessellation, and other times it
> > serves as a polyfill when tessellation just isn't supported.
> >
> > Change-Id: I7b3d57fd0194c6869bfe28ee53ff0ff2e43df479
> > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/291036
> > Commit-Queue: Chris Dalton <csmartdalton@google.com>
> > Reviewed-by: Brian Osman <brianosman@google.com>
>
> TBR=brianosman@google.com
>
> Change-Id: Ia039d7897499a7dad55aff1072f9aa0d01001f9c
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/291693
> Reviewed-by: Chris Dalton <csmartdalton@google.com>
> Commit-Queue: Chris Dalton <csmartdalton@google.com>

TBR=brianosman@google.com

Change-Id: I55704b5adf5a7677a5382d07d39cadee9b252731
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/291738
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
diff --git a/bench/TessellatePathBench.cpp b/bench/TessellatePathBench.cpp
index 610eadc..12c3ce4 100644
--- a/bench/TessellatePathBench.cpp
+++ b/bench/TessellatePathBench.cpp
@@ -10,6 +10,8 @@
 #include "src/core/SkPathPriv.h"
 #include "src/gpu/GrContextPriv.h"
 #include "src/gpu/GrOpFlushState.h"
+#include "src/gpu/tessellate/GrMiddleOutPolygonTriangulator.h"
+#include "src/gpu/tessellate/GrResolveLevelCounter.h"
 #include "src/gpu/tessellate/GrTessellatePathOp.h"
 #include "src/gpu/tessellate/GrWangsFormula.h"
 #include "tools/ToolUtils.h"
@@ -43,6 +45,7 @@
                     "FATAL: wanted %zu bytes of static vertex data; only have %zu.\n",
                     vertexSize * vertexCount, SK_ARRAY_COUNT(fStaticVertexData)).c_str());
         }
+        *startVertex = 0;
         return fStaticVertexData;
     }
 
@@ -79,7 +82,7 @@
 #undef UNIMPL
 
 private:
-    SkPoint fStaticVertexData[(kNumCubicsInChalkboard + 2) * 5];
+    SkPoint fStaticVertexData[(kNumCubicsInChalkboard + 2) * 8];
     GrDrawIndexedIndirectCommand fStaticDrawIndexedIndirectData[32];
     SkSTArenaAlloc<1024 * 1024> fAllocator;
 };
@@ -95,10 +98,15 @@
     const char* onGetName() override { return fName.c_str(); }
     bool isSuitableFor(Backend backend) final { return backend == kNonRendering_Backend; }
 
-    class MiddleOutInnerTrianglesBench;
-    class OuterCubicsBench;
-    class CubicWedgesBench;
-    class WangsFormulaBench;
+    class prepareMiddleOutStencilGeometry;
+    class prepareMiddleOutStencilGeometry_indirect;
+    class prepareIndirectOuterCubics;
+    class prepareTessellatedOuterCubics;
+    class prepareTessellatedCubicWedges;
+    class wangs_formula_cubic_log2;
+    class wangs_formula_cubic_log2_scale;
+    class wangs_formula_cubic_log2_affine;
+    class middle_out_triangulation;
 
 private:
     void onDraw(int loops, SkCanvas*) final {
@@ -120,83 +128,92 @@
     SkString fName;
 };
 
-class GrTessellatePathOp::TestingOnly_Benchmark::MiddleOutInnerTrianglesBench
-        : public GrTessellatePathOp::TestingOnly_Benchmark {
-public:
-    MiddleOutInnerTrianglesBench()
-            : TestingOnly_Benchmark("prepareMiddleOutInnerTriangles",
-                                    ToolUtils::make_star(SkRect::MakeWH(100, 100),
-                                                         kNumCubicsInChalkboard),
-                                    SkMatrix::I()) {
-    }
-    void runBench(GrMeshDrawOp::Target* target, GrTessellatePathOp* op) override {
-        int numBeziers;
-        op->prepareMiddleOutInnerTriangles(target, &numBeziers);
-    }
-};
+#define DEF_TESS_BENCH(NAME, PATH, MATRIX, TARGET, OP) \
+    class GrTessellatePathOp::TestingOnly_Benchmark::NAME \
+            : public GrTessellatePathOp::TestingOnly_Benchmark { \
+    public: \
+        NAME() : TestingOnly_Benchmark(#NAME, (PATH), (MATRIX)) {} \
+        void runBench(GrMeshDrawOp::Target* target, GrTessellatePathOp* op) override; \
+    }; \
+    DEF_BENCH( return new GrTessellatePathOp::TestingOnly_Benchmark::NAME(); ); \
+    void GrTessellatePathOp::TestingOnly_Benchmark::NAME::runBench( \
+            GrMeshDrawOp::Target* TARGET, GrTessellatePathOp* op)
 
-DEF_BENCH( return new GrTessellatePathOp::TestingOnly_Benchmark::MiddleOutInnerTrianglesBench(); );
+DEF_TESS_BENCH(prepareMiddleOutStencilGeometry, make_cubic_path(), SkMatrix::I(), target, op) {
+    op->prepareMiddleOutTrianglesAndCubics(target);
+}
 
-class GrTessellatePathOp::TestingOnly_Benchmark::OuterCubicsBench
-        : public GrTessellatePathOp::TestingOnly_Benchmark {
-public:
-    OuterCubicsBench()
-            : TestingOnly_Benchmark("prepareOuterCubics", make_cubic_path(), SkMatrix::I()) {
-    }
-    void runBench(GrMeshDrawOp::Target* target, GrTessellatePathOp* op) override {
-        op->prepareOuterCubics(target, kNumCubicsInChalkboard,
-                               CubicDataAlignment::kVertexBoundary);
-    }
-};
+DEF_TESS_BENCH(prepareMiddleOutStencilGeometry_indirect, make_cubic_path(), SkMatrix::I(), target,
+               op) {
+    GrResolveLevelCounter resolveLevelCounter;
+    op->prepareMiddleOutTrianglesAndCubics(target, &resolveLevelCounter, true);
+}
 
-DEF_BENCH( return new GrTessellatePathOp::TestingOnly_Benchmark::OuterCubicsBench(); );
+DEF_TESS_BENCH(prepareIndirectOuterCubics, make_cubic_path(), SkMatrix::I(), target, op) {
+    GrResolveLevelCounter resolveLevelCounter;
+    resolveLevelCounter.reset(op->fPath, SkMatrix::I(), 4);
+    op->prepareIndirectOuterCubics(target, resolveLevelCounter);
+}
 
-class GrTessellatePathOp::TestingOnly_Benchmark::CubicWedgesBench
-        : public GrTessellatePathOp::TestingOnly_Benchmark {
-public:
-    CubicWedgesBench()
-            : TestingOnly_Benchmark("prepareCubicWedges", make_cubic_path(), SkMatrix::I()) {
-    }
-    void runBench(GrMeshDrawOp::Target* target, GrTessellatePathOp* op) override {
-        op->prepareCubicWedges(target);
-    }
-};
+DEF_TESS_BENCH(prepareTessellatedOuterCubics, make_cubic_path(), SkMatrix::I(), target, op) {
+    op->prepareTessellatedOuterCubics(target, kNumCubicsInChalkboard);
+}
 
-DEF_BENCH( return new GrTessellatePathOp::TestingOnly_Benchmark::CubicWedgesBench(););
+DEF_TESS_BENCH(prepareTessellatedCubicWedges, make_cubic_path(), SkMatrix::I(), target, op) {
+    op->prepareTessellatedCubicWedges(target);
+}
 
-class GrTessellatePathOp::TestingOnly_Benchmark::WangsFormulaBench
-        : public GrTessellatePathOp::TestingOnly_Benchmark {
-public:
-    WangsFormulaBench(const char* suffix, const SkMatrix& matrix)
-            : TestingOnly_Benchmark(SkStringPrintf("wangs_formula_cubic_log2%s", suffix).c_str(),
-                                    make_cubic_path(), SkMatrix::I())
-            , fMatrix(matrix) {
-    }
-    void runBench(GrMeshDrawOp::Target*, GrTessellatePathOp* op) override {
-        int sum = 0;
-        GrVectorXform xform(fMatrix);
-        for (auto [verb, pts, w] : SkPathPriv::Iterate(op->fPath)) {
-            if (verb == SkPathVerb::kCubic) {
-                sum += GrWangsFormula::cubic_log2(4, pts, xform);
-            }
-        }
-        // Don't let the compiler optimize away GrWangsFormula::cubic_log2.
-        if (sum <= 0) {
-            SK_ABORT("sum should be > 0.");
+static void benchmark_wangs_formula_cubic_log2(const SkMatrix& matrix, const SkPath& path) {
+    int sum = 0;
+    GrVectorXform xform(matrix);
+    for (auto [verb, pts, w] : SkPathPriv::Iterate(path)) {
+        if (verb == SkPathVerb::kCubic) {
+            sum += GrWangsFormula::cubic_log2(4, pts, xform);
         }
     }
-private:
-    SkMatrix fMatrix;
-};
+    // Don't let the compiler optimize away GrWangsFormula::cubic_log2.
+    if (sum <= 0) {
+        SK_ABORT("sum should be > 0.");
+    }
+}
 
-DEF_BENCH(
-    return new GrTessellatePathOp::TestingOnly_Benchmark::WangsFormulaBench("", SkMatrix::I());
-);
-DEF_BENCH(
-    return new GrTessellatePathOp::TestingOnly_Benchmark::WangsFormulaBench(
-            "_scale", SkMatrix::Scale(1.1f, 0.9f));
-);
-DEF_BENCH(
-    return new GrTessellatePathOp::TestingOnly_Benchmark::WangsFormulaBench(
-            "_affine", SkMatrix::MakeAll(.9f,0.9f,0,  1.1f,1.1f,0, 0,0,1));
-);
+DEF_TESS_BENCH(wangs_formula_cubic_log2, make_cubic_path(), SkMatrix::I(), target, op) {
+    benchmark_wangs_formula_cubic_log2(op->fViewMatrix, op->fPath);
+}
+
+DEF_TESS_BENCH(wangs_formula_cubic_log2_scale, make_cubic_path(), SkMatrix::Scale(1.1f, 0.9f),
+               target, op) {
+    benchmark_wangs_formula_cubic_log2(op->fViewMatrix, op->fPath);
+}
+
+DEF_TESS_BENCH(wangs_formula_cubic_log2_affine, make_cubic_path(),
+               SkMatrix::MakeAll(.9f,0.9f,0,  1.1f,1.1f,0, 0,0,1), target, op) {
+    benchmark_wangs_formula_cubic_log2(op->fViewMatrix, op->fPath);
+}
+
+DEF_TESS_BENCH(middle_out_triangulation,
+               ToolUtils::make_star(SkRect::MakeWH(500, 500), kNumCubicsInChalkboard),
+               SkMatrix::I(), target, op) {
+    int baseVertex;
+    auto vertexData = static_cast<SkPoint*>(target->makeVertexSpace(
+            sizeof(SkPoint), kNumCubicsInChalkboard, nullptr, &baseVertex));
+    GrMiddleOutPolygonTriangulator middleOut(vertexData, 3, kNumCubicsInChalkboard + 2);
+    for (auto [verb, pts, w] : SkPathPriv::Iterate(op->fPath)) {
+        switch (verb) {
+            case SkPathVerb::kMove:
+                middleOut.closeAndMove(pts[0]);
+                break;
+            case SkPathVerb::kLine:
+                middleOut.pushVertex(pts[1]);
+                break;
+            case SkPathVerb::kClose:
+                middleOut.close();
+                break;
+            case SkPathVerb::kQuad:
+            case SkPathVerb::kConic:
+            case SkPathVerb::kCubic:
+                SkUNREACHABLE;
+        }
+        middleOut.closeAndMove(pts[0]);
+    }
+}
diff --git a/gn/gpu.gni b/gn/gpu.gni
index bc5bcf1..c13461a 100644
--- a/gn/gpu.gni
+++ b/gn/gpu.gni
@@ -444,6 +444,7 @@
   "$_src/gpu/tessellate/GrMiddleOutPolygonTriangulator.h",
   "$_src/gpu/tessellate/GrMidpointContourParser.h",
   "$_src/gpu/tessellate/GrPathShader.h",
+  "$_src/gpu/tessellate/GrResolveLevelCounter.h",
   "$_src/gpu/tessellate/GrStencilPathShader.cpp",
   "$_src/gpu/tessellate/GrStencilPathShader.h",
   "$_src/gpu/tessellate/GrTessellatePathOp.cpp",
diff --git a/src/gpu/GrPathRendererChain.cpp b/src/gpu/GrPathRendererChain.cpp
index ca2d7e4..cf7e271 100644
--- a/src/gpu/GrPathRendererChain.cpp
+++ b/src/gpu/GrPathRendererChain.cpp
@@ -32,7 +32,7 @@
         fChain.push_back(sk_make_sp<GrDashLinePathRenderer>());
     }
     if (options.fGpuPathRenderers & GpuPathRenderers::kTessellation) {
-        if (caps.shaderCaps()->tessellationSupport() && caps.drawInstancedSupport()) {
+        if (caps.drawInstancedSupport()) {
             auto tess = sk_make_sp<GrTessellationPathRenderer>(caps);
             context->priv().addOnFlushCallbackObject(tess.get());
             fChain.push_back(std::move(tess));
diff --git a/src/gpu/GrProcessor.h b/src/gpu/GrProcessor.h
index cb4711c..1a16595 100644
--- a/src/gpu/GrProcessor.h
+++ b/src/gpu/GrProcessor.h
@@ -162,9 +162,10 @@
         kTessellate_GrFillBoundingBoxShader_ClassID,
         kTessellate_GrFillCubicHullShader_ClassID,
         kTessellate_GrFillTriangleShader_ClassID,
-        kTessellate_GrStencilCubicShader_ClassID,
+        kTessellate_GrMiddleOutCubicShader_ClassID,
         kTessellate_GrStencilTriangleShader_ClassID,
-        kTessellate_GrStencilWedgeShader_ClassID,
+        kTessellate_GrTessellateCubicShader_ClassID,
+        kTessellate_GrTessellateWedgeShader_ClassID,
         kTestFP_ClassID,
         kTestRectOp_ClassID,
         kFlatNormalsFP_ClassID,
diff --git a/src/gpu/tessellate/GrResolveLevelCounter.h b/src/gpu/tessellate/GrResolveLevelCounter.h
new file mode 100644
index 0000000..c74f46c
--- /dev/null
+++ b/src/gpu/tessellate/GrResolveLevelCounter.h
@@ -0,0 +1,74 @@
+/*
+ * Copyright 2020 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef GrResolveLevelCounter_DEFINED
+#define GrResolveLevelCounter_DEFINED
+
+#include "src/core/SkPathPriv.h"
+#include "src/gpu/tessellate/GrStencilPathShader.h"
+#include "src/gpu/tessellate/GrWangsFormula.h"
+
+// This class helps bin cubics by log2 "resolveLevel" when we don't use hardware tessellation. It is
+// composed of simple counters that track how many cubics we intend to draw at each resolveLevel,
+// and how many resolveLevels there are that have at least one cubic.
+class GrResolveLevelCounter {
+public:
+    void reset() {
+        memset(fInstanceCounts, 0, sizeof(fInstanceCounts));
+        SkDEBUGCODE(fHasCalledReset = true;)
+    }
+
+    int reset(const SkPath& path, const SkMatrix& viewMatrix, float intolerance) {
+        this->reset();
+        GrVectorXform xform(viewMatrix);
+        for (auto [verb, pts, w] : SkPathPriv::Iterate(path)) {
+            switch (verb) {
+                case SkPathVerb::kQuad:
+                    // Quadratics get converted to cubics before rendering.
+                    this->countCubic(GrWangsFormula::quadratic_log2(intolerance, pts, xform));
+                    break;
+                case SkPathVerb::kCubic:
+                    this->countCubic(GrWangsFormula::cubic_log2(intolerance, pts, xform));
+                    break;
+                default:
+                    break;
+            }
+        }
+        return fTotalCubicInstanceCount;
+    }
+
+    void countCubic(int resolveLevel) {
+        SkASSERT(fHasCalledReset);
+        SkASSERT(resolveLevel >= 0);
+        if (resolveLevel == 0) {
+            // Cubics with 2^0=1 segments are empty (zero area). We ignore them completely.
+            return;
+        }
+        resolveLevel = std::min(resolveLevel, GrMiddleOutCubicShader::kMaxResolveLevel);
+        if (!fInstanceCounts[resolveLevel]++) {
+            ++fTotalCubicIndirectDrawCount;
+        }
+        ++fTotalCubicInstanceCount;
+    }
+
+    int operator[](int resolveLevel) const {
+        SkASSERT(fHasCalledReset);
+        SkASSERT(resolveLevel > 0);  // Empty cubics with 2^0=1 segments do not need to be drawn.
+        SkASSERT(resolveLevel <= GrMiddleOutCubicShader::kMaxResolveLevel);
+        return fInstanceCounts[resolveLevel];
+    }
+    int totalCubicInstanceCount() const { return fTotalCubicInstanceCount; }
+    int totalCubicIndirectDrawCount() const { return fTotalCubicIndirectDrawCount; }
+
+private:
+    SkDEBUGCODE(bool fHasCalledReset = false;)
+    int fInstanceCounts[GrMiddleOutCubicShader::kMaxResolveLevel + 1];
+    int fTotalCubicInstanceCount = 0;
+    int fTotalCubicIndirectDrawCount = 0;
+};
+
+#endif
diff --git a/src/gpu/tessellate/GrStencilPathShader.cpp b/src/gpu/tessellate/GrStencilPathShader.cpp
index 9b36c2a..fadc5b1 100644
--- a/src/gpu/tessellate/GrStencilPathShader.cpp
+++ b/src/gpu/tessellate/GrStencilPathShader.cpp
@@ -35,6 +35,7 @@
         })";
 
 class GrStencilPathShader::Impl : public GrGLSLGeometryProcessor {
+protected:
     void onEmitCode(EmitArgs& args, GrGPArgs* gpArgs) override {
         const auto& shader = args.fGP.cast<GrStencilPathShader>();
         args.fVaryingHandler->emitAttributes(shader);
@@ -45,7 +46,7 @@
             fViewMatrixUniform = args.fUniformHandler->addUniform(
                     nullptr, kVertex_GrShaderFlag, kFloat3x3_GrSLType, "view_matrix", &viewMatrix);
             args.fVertBuilder->codeAppendf(
-                    "float2 vertexpos = (%s * float3(point, 1)).xy;", viewMatrix);
+                    "float2 vertexpos = (%s * float3(inputPoint, 1)).xy;", viewMatrix);
             vertexPos.set(kFloat2_GrSLType, "vertexpos");
         }
 
@@ -75,8 +76,8 @@
     return new Impl;
 }
 
-SkString GrStencilCubicShader::getTessControlShaderGLSL(const char* versionAndExtensionDecls,
-                                                        const GrShaderCaps&) const {
+SkString GrTessellateCubicShader::getTessControlShaderGLSL(const char* versionAndExtensionDecls,
+                                                           const GrShaderCaps&) const {
     SkString code(versionAndExtensionDecls);
     code.append(kWangsFormulaCubicFn);
     code.append(R"(
@@ -115,8 +116,8 @@
     return code;
 }
 
-SkString GrStencilCubicShader::getTessEvaluationShaderGLSL(const char* versionAndExtensionDecls,
-                                                           const GrShaderCaps&) const {
+SkString GrTessellateCubicShader::getTessEvaluationShaderGLSL(const char* versionAndExtensionDecls,
+                                                              const GrShaderCaps&) const {
     SkString code(versionAndExtensionDecls);
     code.append(kEvalCubicFn);
     code.append(R"(
@@ -147,8 +148,8 @@
     return code;
 }
 
-SkString GrStencilWedgeShader::getTessControlShaderGLSL(const char* versionAndExtensionDecls,
-                                                        const GrShaderCaps&) const {
+SkString GrTessellateWedgeShader::getTessControlShaderGLSL(const char* versionAndExtensionDecls,
+                                                           const GrShaderCaps&) const {
     SkString code(versionAndExtensionDecls);
     code.append(kWangsFormulaCubicFn);
     code.append(R"(
@@ -183,8 +184,8 @@
     return code;
 }
 
-SkString GrStencilWedgeShader::getTessEvaluationShaderGLSL(const char* versionAndExtensionDecls,
-                                                           const GrShaderCaps&) const {
+SkString GrTessellateWedgeShader::getTessEvaluationShaderGLSL(const char* versionAndExtensionDecls,
+                                                              const GrShaderCaps&) const {
     SkString code(versionAndExtensionDecls);
     code.append(kEvalCubicFn);
     code.append(R"(
@@ -218,3 +219,109 @@
 
     return code;
 }
+
+GR_DECLARE_STATIC_UNIQUE_KEY(gMiddleOutIndexBufferKey);
+
+sk_sp<const GrGpuBuffer> GrMiddleOutCubicShader::FindOrMakeMiddleOutIndexBuffer(
+        GrResourceProvider* resourceProvider) {
+    GR_DEFINE_STATIC_UNIQUE_KEY(gMiddleOutIndexBufferKey);
+    if (auto buffer = resourceProvider->findByUniqueKey<GrGpuBuffer>(gMiddleOutIndexBufferKey)) {
+        return std::move(buffer);
+    }
+
+    // One explicit triangle at index 0, and one middle-out cubic with kMaxResolveLevel line
+    // segments beginning at index 3.
+    constexpr static int indexCount = 3 + NumVerticesAtResolveLevel(kMaxResolveLevel);
+    auto buffer = resourceProvider->createBuffer(
+            indexCount * sizeof(uint16_t), GrGpuBufferType::kIndex, kStatic_GrAccessPattern);
+    if (!buffer) {
+        return nullptr;
+    }
+
+    // We shouldn't bin and/or cache static buffers.
+    SkASSERT(buffer->size() == indexCount * sizeof(uint16_t));
+    SkASSERT(!buffer->resourcePriv().getScratchKey().isValid());
+    auto indexData = static_cast<uint16_t*>(buffer->map());
+    SkAutoTMalloc<uint16_t> stagingBuffer;
+    if (!indexData) {
+        SkASSERT(!buffer->isMapped());
+        indexData = stagingBuffer.reset(indexCount);
+    }
+
+    // Indices 0,1,2 contain special values that emit points P0, P1, and P2 respectively. (When the
+    // vertex shader is fed an index value larger than (1 << kMaxResolveLevel), it emits
+    // P[index % 4].)
+    int i = 0;
+    indexData[i++] = (1 << kMaxResolveLevel) + 4;  // % 4 == 0
+    indexData[i++] = (1 << kMaxResolveLevel) + 5;  // % 4 == 1
+    indexData[i++] = (1 << kMaxResolveLevel) + 6;  // % 4 == 2
+
+    // Starting at index 3, we triangulate a cubic with 2^kMaxResolveLevel line segments. Each
+    // index value corresponds to parametric value T=(index / 2^kMaxResolveLevel). Since the
+    // triangles are arranged in "middle-out" order, we will be able to conveniently control the
+    // resolveLevel by changing only the indexCount.
+    for (uint16_t advance = 1 << (kMaxResolveLevel - 1); advance; advance >>= 1) {
+        uint16_t T = 0;
+        do {
+            indexData[i++] = T;
+            indexData[i++] = (T += advance);
+            indexData[i++] = (T += advance);
+        } while (T != (1 << kMaxResolveLevel));
+    }
+    SkASSERT(i == indexCount);
+
+    if (buffer->isMapped()) {
+        buffer->unmap();
+    } else {
+        buffer->updateData(stagingBuffer, indexCount * sizeof(uint16_t));
+    }
+    buffer->resourcePriv().setUniqueKey(gMiddleOutIndexBufferKey);
+    return std::move(buffer);
+}
+
+class GrMiddleOutCubicShader::Impl : public GrStencilPathShader::Impl {
+    void onEmitCode(EmitArgs& args, GrGPArgs* gpArgs) override {
+        const auto& shader = args.fGP.cast<GrMiddleOutCubicShader>();
+        args.fVaryingHandler->emitAttributes(shader);
+        args.fVertBuilder->defineConstant("kMaxResolveLevel", kMaxResolveLevel);
+        args.fVertBuilder->codeAppend(R"(
+                float4x2 P = float4x2(inputPoints_0_1, inputPoints_2_3);
+                float2 point;
+                if (sk_VertexID > (1 << kMaxResolveLevel)) {
+                    // This is a special index value that wants us to emit a specific point.
+                    point = P[sk_VertexID & 3];
+                } else {)");
+        // Evaluate the cubic at T=(sk_VertexID / 2^kMaxResolveLevel).
+        if (args.fShaderCaps->fpManipulationSupport()) {
+            args.fVertBuilder->codeAppend(R"(
+                    float T = ldexp(sk_VertexID, -kMaxResolveLevel);)");
+        } else {
+            args.fVertBuilder->codeAppend(R"(
+                    float T = sk_VertexID / float(1 << kMaxResolveLevel);)");
+        }
+        args.fVertBuilder->codeAppend(R"(
+                    float2 ab = mix(P[0], P[1], T);
+                    float2 bc = mix(P[1], P[2], T);
+                    float2 cd = mix(P[2], P[3], T);
+                    float2 abc = mix(ab, bc, T);
+                    float2 bcd = mix(bc, cd, T);
+                    point = mix(abc, bcd, T);
+                })");
+
+        GrShaderVar vertexPos("point", kFloat2_GrSLType);
+        if (!shader.viewMatrix().isIdentity()) {
+            const char* viewMatrix;
+            fViewMatrixUniform = args.fUniformHandler->addUniform(
+                    nullptr, kVertex_GrShaderFlag, kFloat3x3_GrSLType, "view_matrix", &viewMatrix);
+            args.fVertBuilder->codeAppendf(R"(
+                    float2 transformedPoint = (%s * float3(point, 1)).xy;)", viewMatrix);
+            vertexPos.set(kFloat2_GrSLType, "transformedPoint");
+        }
+        gpArgs->fPositionVar = vertexPos;
+        // No fragment shader.
+    }
+};
+
+GrGLSLPrimitiveProcessor* GrMiddleOutCubicShader::createGLSLInstance(const GrShaderCaps&) const {
+    return new Impl;
+}
diff --git a/src/gpu/tessellate/GrStencilPathShader.h b/src/gpu/tessellate/GrStencilPathShader.h
index 0aa0b59..c331726 100644
--- a/src/gpu/tessellate/GrStencilPathShader.h
+++ b/src/gpu/tessellate/GrStencilPathShader.h
@@ -17,16 +17,15 @@
     GrStencilPathShader(ClassID classID, const SkMatrix& viewMatrix, GrPrimitiveType primitiveType,
                         int tessellationPatchVertexCount = 0)
             : GrPathShader(classID, viewMatrix, primitiveType, tessellationPatchVertexCount) {
-        constexpr static Attribute kPointAttrib = {
-                "point", kFloat2_GrVertexAttribType, kFloat2_GrSLType};
-        this->setVertexAttributes(&kPointAttrib, 1);
     }
 
-private:
-    void getGLSLProcessorKey(const GrShaderCaps&, GrProcessorKeyBuilder* b) const final {
+protected:
+    constexpr static Attribute kSinglePointAttrib{"inputPoint", kFloat2_GrVertexAttribType,
+                                                  kFloat2_GrSLType};
+    void getGLSLProcessorKey(const GrShaderCaps&, GrProcessorKeyBuilder* b) const override {
         b->add32(this->viewMatrix().isIdentity());
     }
-    GrGLSLPrimitiveProcessor* createGLSLInstance(const GrShaderCaps&) const final;
+    GrGLSLPrimitiveProcessor* createGLSLInstance(const GrShaderCaps&) const override;
 
     class Impl;
 };
@@ -35,18 +34,21 @@
 class GrStencilTriangleShader : public GrStencilPathShader {
 public:
     GrStencilTriangleShader(const SkMatrix& viewMatrix) : GrStencilPathShader(
-            kTessellate_GrStencilTriangleShader_ClassID, viewMatrix, GrPrimitiveType::kTriangles) {}
+            kTessellate_GrStencilTriangleShader_ClassID, viewMatrix, GrPrimitiveType::kTriangles) {
+        this->setVertexAttributes(&kSinglePointAttrib, 1);
+    }
     const char* name() const override { return "tessellate_GrStencilTriangleShader"; }
 };
 
-// Uses GPU tessellation shaders to linearize, triangulate, and render standalone cubics. Here, a
-// "cubic" is a standalone closed contour consisting of a single cubic bezier.
+// Uses GPU tessellation shaders to linearize, triangulate, and render standalone closed cubics.
 // TODO: Eventually we want to use rational cubic wedges in order to support perspective and conics.
-class GrStencilCubicShader : public GrStencilPathShader {
+class GrTessellateCubicShader : public GrStencilPathShader {
 public:
-    GrStencilCubicShader(const SkMatrix& viewMatrix) : GrStencilPathShader(
-            kTessellate_GrStencilCubicShader_ClassID, viewMatrix, GrPrimitiveType::kPatches, 4) {}
-    const char* name() const override { return "tessellate_GrStencilCubicShader"; }
+    GrTessellateCubicShader(const SkMatrix& viewMatrix) : GrStencilPathShader(
+            kTessellate_GrTessellateCubicShader_ClassID, viewMatrix, GrPrimitiveType::kPatches, 4) {
+        this->setVertexAttributes(&kSinglePointAttrib, 1);
+    }
+    const char* name() const override { return "tessellate_GrTessellateCubicShader"; }
 
 private:
     SkString getTessControlShaderGLSL(const char* versionAndExtensionDecls,
@@ -59,11 +61,13 @@
 // wedge is a 5-point patch consisting of 4 cubic control points, plus an anchor point fanning from
 // the center of the curve's resident contour.
 // TODO: Eventually we want to use rational cubic wedges in order to support perspective and conics.
-class GrStencilWedgeShader : public GrStencilPathShader {
+class GrTessellateWedgeShader : public GrStencilPathShader {
 public:
-    GrStencilWedgeShader(const SkMatrix& viewMatrix) : GrStencilPathShader(
-            kTessellate_GrStencilWedgeShader_ClassID, viewMatrix, GrPrimitiveType::kPatches, 5) {}
-    const char* name() const override { return "tessellate_GrStencilWedgeShader"; }
+    GrTessellateWedgeShader(const SkMatrix& viewMatrix) : GrStencilPathShader(
+            kTessellate_GrTessellateWedgeShader_ClassID, viewMatrix, GrPrimitiveType::kPatches, 5) {
+        this->setVertexAttributes(&kSinglePointAttrib, 1);
+    }
+    const char* name() const override { return "tessellate_GrTessellateWedgeShader"; }
 
 private:
     SkString getTessControlShaderGLSL(const char* versionAndExtensionDecls,
@@ -72,4 +76,69 @@
                                          const GrShaderCaps&) const override;
 };
 
+// Uses indirect (instanced) draws to triangulate standalone closed cubics with a "middle-out"
+// topology. The caller must compute each cubic's resolveLevel on the CPU (i.e., the log2 number of
+// line segments it will be divided into; see GrWangsFormula::cubic_log2/quadratic_log2), and then
+// sort the instance buffer by resolveLevel for efficient batching of indirect draws.
+class GrMiddleOutCubicShader : public GrStencilPathShader {
+public:
+    // Each resolveLevel linearizes the curve into 2^resolveLevel line segments. The finest
+    // supported resolveLevel is therefore 2^12=4096 line segments.
+    constexpr static int kMaxResolveLevel = 12;
+
+    // How many vertices do we need to draw in order to triangulate a cubic with 2^resolveLevel
+    // line segments?
+    constexpr static int NumVerticesAtResolveLevel(int resolveLevel) {
+        // resolveLevel=0 -> 0 line segments -> 0 triangles -> 0 vertices
+        // resolveLevel=1 -> 2 line segments -> 1 triangle -> 3 vertices
+        // resolveLevel=2 -> 4 line segments -> 3 triangles -> 9 vertices
+        // resolveLevel=3 -> 8 line segments -> 7 triangles -> 21 vertices
+        // ...
+        return ((1 << resolveLevel) - 1) * 3;
+    }
+
+    // Configures an indirect draw to render cubic instances with 2^resolveLevel evenly-spaced (in
+    // the parametric sense) line segments.
+    static GrDrawIndexedIndirectCommand MakeDrawCubicsIndirectCmd(int resolveLevel,
+                                                                  uint32_t instanceCount,
+                                                                  uint32_t baseInstance) {
+        SkASSERT(resolveLevel > 0 && resolveLevel <= kMaxResolveLevel);
+        // Starting at baseIndex=3, the index buffer triangulates a cubic with 2^kMaxResolveLevel
+        // line segments. Each index value corresponds to a parametric T value on the curve. Since
+        // the triangles are arranged in "middle-out" order, we can conveniently control the
+        // resolveLevel by changing only the indexCount.
+        uint32_t indexCount = NumVerticesAtResolveLevel(resolveLevel);
+        return {indexCount, instanceCount, 3, 0, baseInstance};
+    }
+
+    // For performance reasons we can often express triangles as an indirect cubic draw and sneak
+    // them in alongside the other indirect draws. This method configures an indirect draw to emit
+    // the triangle [P0, P1, P2] from a 4-point instance.
+    static GrDrawIndexedIndirectCommand MakeDrawTrianglesIndirectCmd(uint32_t instanceCount,
+                                                                     uint32_t baseInstance) {
+        // Indices 0,1,2 have special index values that emit points P0, P1, and P2 respectively.
+        return {3, instanceCount, 0, 0, baseInstance};
+    }
+
+    // Returns the index buffer that should be bound when drawing with this shader.
+    // (Our vertex shader uses raw index values directly, so there is no vertex buffer.)
+    static sk_sp<const GrGpuBuffer> FindOrMakeMiddleOutIndexBuffer(GrResourceProvider*);
+
+    GrMiddleOutCubicShader(const SkMatrix& viewMatrix)
+            : GrStencilPathShader(kTessellate_GrMiddleOutCubicShader_ClassID, viewMatrix,
+                                  GrPrimitiveType::kTriangles) {
+        constexpr static Attribute kInputPtsAttribs[] = {
+                {"inputPoints_0_1", kFloat4_GrVertexAttribType, kFloat4_GrSLType},
+                {"inputPoints_2_3", kFloat4_GrVertexAttribType, kFloat4_GrSLType}};
+        this->setInstanceAttributes(kInputPtsAttribs, 2);
+    }
+
+    const char* name() const override { return "tessellate_GrMiddleOutCubicShader"; }
+
+private:
+    GrGLSLPrimitiveProcessor* createGLSLInstance(const GrShaderCaps&) const override;
+
+    class Impl;
+};
+
 #endif
diff --git a/src/gpu/tessellate/GrTessellatePathOp.cpp b/src/gpu/tessellate/GrTessellatePathOp.cpp
index cf79daa..cb1bf37 100644
--- a/src/gpu/tessellate/GrTessellatePathOp.cpp
+++ b/src/gpu/tessellate/GrTessellatePathOp.cpp
@@ -14,8 +14,12 @@
 #include "src/gpu/tessellate/GrFillPathShader.h"
 #include "src/gpu/tessellate/GrMiddleOutPolygonTriangulator.h"
 #include "src/gpu/tessellate/GrMidpointContourParser.h"
+#include "src/gpu/tessellate/GrResolveLevelCounter.h"
 #include "src/gpu/tessellate/GrStencilPathShader.h"
 
+constexpr static int kMaxResolveLevel = GrMiddleOutCubicShader::kMaxResolveLevel;
+constexpr static float kTessellationIntolerance = 4;  // 1/4 of a pixel.
+
 GrTessellatePathOp::FixedFunctionFlags GrTessellatePathOp::fixedFunctionFlags() const {
     auto flags = FixedFunctionFlags::kUsesStencil;
     if (GrAAType::kNone != fAAType) {
@@ -30,50 +34,69 @@
                                       const GrXferProcessor::DstProxyView&) {
 }
 
-void GrTessellatePathOp::onPrepare(GrOpFlushState* state) {
-    // First check if the path is large and/or simple enough that we can actually triangulate the
-    // inner polygon(s) on the CPU. This is our fastest approach. It allows us to stencil only the
-    // curves, and then fill the internal polygons directly to the final render target, thus filling
-    // in the majority of pixels in a single render pass.
-    SkScalar scales[2];
-    SkAssertResult(fViewMatrix.getMinMaxScales(scales));  // Will fail if perspective.
-    const SkRect& bounds = fPath.getBounds();
+void GrTessellatePathOp::onPrepare(GrOpFlushState* flushState) {
     int numVerbs = fPath.countVerbs();
     if (numVerbs <= 0) {
         return;
     }
+
+    // First check if the path is large and/or simple enough that we can actually triangulate the
+    // inner polygon(s) on the CPU. This is our fastest approach. It allows us to stencil only the
+    // curves, and then fill the internal polygons directly to the final render target, thus drawing
+    // the majority of pixels in a single render pass.
+    SkScalar scales[2];
+    SkAssertResult(fViewMatrix.getMinMaxScales(scales));  // Will fail if perspective.
+    const SkRect& bounds = fPath.getBounds();
     float gpuFragmentWork = bounds.height() * scales[0] * bounds.width() * scales[1];
     float cpuTessellationWork = (float)numVerbs * SkNextLog2(numVerbs);  // N log N.
     if (cpuTessellationWork * 500 + (256 * 256) < gpuFragmentWork) {  // Don't try below 256x256.
-        int numCountedCurves;
+        int numCountedCubics;
         // This will fail if the inner triangles do not form a simple polygon (e.g., self
         // intersection, double winding).
-        if (this->prepareNonOverlappingInnerTriangles(state, &numCountedCurves)) {
-            // Prepare cubics on an instance boundary so we can use the buffer to fill local convex
-            // hulls as well.
-            this->prepareOuterCubics(state, numCountedCurves,
-                                     CubicDataAlignment::kInstanceBoundary);
+        if (this->prepareNonOverlappingInnerTriangles(flushState, &numCountedCubics)) {
+            if (!numCountedCubics) {
+                return;
+            }
+            // Always use indirect draws for cubics instead of tessellation here. Our goal in this
+            // mode is to maximize GPU performance, and the middle-out topology used by our indirect
+            // draws is easier on the rasterizer than a tessellated fan. There also seems to be a
+            // small amount of fixed tessellation overhead that this avoids.
+            //
+            // NOTE: This will count fewer cubics than above if it discards any for resolveLevel=0.
+            GrResolveLevelCounter resolveLevelCounter;
+            numCountedCubics = resolveLevelCounter.reset(fPath, fViewMatrix,
+                                                         kTessellationIntolerance);
+            this->prepareIndirectOuterCubics(flushState, resolveLevelCounter);
             return;
         }
     }
 
-    // Next see if we can split up inner polygon triangles and curves, and triangulate the inner
-    // polygon(s) more efficiently. This causes greater CPU overhead due to the extra shaders and
-    // draw calls, but the better triangulation can reduce the rasterizer load by a great deal on
-    // complex paths.
+    // When there are only a few verbs, it seems to always be fastest to make a single indirect draw
+    // that contains both the inner triangles and the outer cubics, instead of using hardware
+    // tessellation. Also take this path if tessellation is not supported.
+    bool drawTrianglesAsIndirectCubicDraw = (numVerbs < 50);
+    if (drawTrianglesAsIndirectCubicDraw ||
+        !flushState->caps().shaderCaps()->tessellationSupport()) {
+        // Prepare outer cubics with indirect draws.
+        GrResolveLevelCounter resolveLevelCounter;
+        this->prepareMiddleOutTrianglesAndCubics(flushState, &resolveLevelCounter,
+                                                 drawTrianglesAsIndirectCubicDraw);
+        return;
+    }
+
+    // Next see if we can split up the inner triangles and outer cubics into two draw calls. This
+    // allows for a more efficient inner triangle topology that can reduce the rasterizer load by a
+    // large margin on complex paths, but also causes greater CPU overhead due to the extra shader
+    // switches and draw calls.
     // NOTE: Raster-edge work is 1-dimensional, so we sum height and width instead of multiplying.
     float rasterEdgeWork = (bounds.height() + bounds.width()) * scales[1] * fPath.countVerbs();
-    if (rasterEdgeWork > 1000 * 1000) {
-        int numCountedCurves;
-        this->prepareMiddleOutInnerTriangles(state, &numCountedCurves);
-        // We will fill the path with a bounding box instead local cubic convex hulls, so there is
-        // no need to prepare the cubics on an instance boundary.
-        this->prepareOuterCubics(state, numCountedCurves, CubicDataAlignment::kVertexBoundary);
+    if (rasterEdgeWork > 300 * 300) {
+        this->prepareMiddleOutTrianglesAndCubics(flushState);
         return;
     }
 
     // Fastest CPU approach: emit one cubic wedge per verb, fanning out from the center.
-    this->prepareCubicWedges(state);
+    this->prepareTessellatedCubicWedges(flushState);
 }
 
 bool GrTessellatePathOp::prepareNonOverlappingInnerTriangles(GrMeshDrawOp::Target* target,
@@ -106,40 +129,74 @@
     return true;
 }
 
-void GrTessellatePathOp::prepareMiddleOutInnerTriangles(GrMeshDrawOp::Target* target,
-                                                        int* numCountedCurves) {
+void GrTessellatePathOp::prepareMiddleOutTrianglesAndCubics(
+        GrMeshDrawOp::Target* target, GrResolveLevelCounter* resolveLevelCounter,
+        bool drawTrianglesAsIndirectCubicDraw) {
     SkASSERT(!fTriangleBuffer);
     SkASSERT(!fDoStencilTriangleBuffer);
     SkASSERT(!fDoFillTriangleBuffer);
+    SkASSERT(!fCubicBuffer);
+    SkASSERT(!fStencilCubicsShader);
+    SkASSERT(!fIndirectDrawBuffer);
 
     // No initial moveTo, plus an implicit close at the end; n-2 triangles fill an n-gon.
-    // Each triangle has 3 vertices.
-    int maxVertices = (fPath.countVerbs() - 1) * 3;
+    int maxInnerTriangles = fPath.countVerbs() - 1;
+    int maxCubics = fPath.countVerbs();
 
-    GrEagerDynamicVertexAllocator vertexAlloc(target, &fTriangleBuffer, &fBaseTriangleVertex);
-    auto* vertexData = vertexAlloc.lock<SkPoint>(maxVertices);
+    SkPoint* vertexData;
+    int vertexAdvancePerTriangle;
+    if (drawTrianglesAsIndirectCubicDraw) {
+        // Allocate the triangles as 4-point instances at the beginning of the cubic buffer.
+        SkASSERT(resolveLevelCounter);
+        vertexAdvancePerTriangle = 4;
+        int baseTriangleInstance;
+        vertexData = static_cast<SkPoint*>(target->makeVertexSpace(
+                sizeof(SkPoint) * 4, maxInnerTriangles + maxCubics, &fCubicBuffer,
+                &baseTriangleInstance));
+        fBaseCubicVertex = baseTriangleInstance * 4;
+    } else {
+        // Allocate the triangles as normal 3-point instances in the triangle buffer.
+        vertexAdvancePerTriangle = 3;
+        vertexData = static_cast<SkPoint*>(target->makeVertexSpace(
+                sizeof(SkPoint), maxInnerTriangles * 3, &fTriangleBuffer, &fBaseTriangleVertex));
+    }
     if (!vertexData) {
         return;
     }
 
-    constexpr static int kNumVerticesPerTriangle = 3;
-    GrMiddleOutPolygonTriangulator middleOut(vertexData, kNumVerticesPerTriangle, maxVertices);
-    int localCurveCount = 0;
+    GrVectorXform xform(fViewMatrix);
+    GrMiddleOutPolygonTriangulator middleOut(vertexData, vertexAdvancePerTriangle,
+                                             fPath.countVerbs());
+    if (resolveLevelCounter) {
+        resolveLevelCounter->reset();
+    }
+    int numCountedCurves = 0;
     for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) {
         switch (verb) {
             case SkPathVerb::kMove:
-                middleOut.closeAndMove(*pts++);
+                middleOut.closeAndMove(pts[0]);
                 break;
             case SkPathVerb::kLine:
                 middleOut.pushVertex(pts[1]);
                 break;
             case SkPathVerb::kQuad:
                 middleOut.pushVertex(pts[2]);
-                ++localCurveCount;
+                if (resolveLevelCounter) {
+                    // Quadratics get converted to cubics before rendering.
+                    resolveLevelCounter->countCubic(GrWangsFormula::quadratic_log2(
+                            kTessellationIntolerance, pts, xform));
+                    break;
+                }
+                ++numCountedCurves;
                 break;
             case SkPathVerb::kCubic:
                 middleOut.pushVertex(pts[3]);
-                ++localCurveCount;
+                if (resolveLevelCounter) {
+                    resolveLevelCounter->countCubic(GrWangsFormula::cubic_log2(
+                            kTessellationIntolerance, pts, xform));
+                    break;
+                }
+                ++numCountedCurves;
                 break;
             case SkPathVerb::kClose:
                 middleOut.close();
@@ -148,13 +205,31 @@
                 SkUNREACHABLE;
         }
     }
-    fTriangleVertexCount = middleOut.close() * kNumVerticesPerTriangle;
-    *numCountedCurves = localCurveCount;
+    int triangleCount = middleOut.close();
+    SkASSERT(triangleCount <= maxInnerTriangles);
 
-    vertexAlloc.unlock(fTriangleVertexCount);
-
-    if (fTriangleVertexCount) {
-        fDoStencilTriangleBuffer = true;
+    if (drawTrianglesAsIndirectCubicDraw) {
+        SkASSERT(resolveLevelCounter);
+        int totalInstanceCount = triangleCount + resolveLevelCounter->totalCubicInstanceCount();
+        SkASSERT(vertexAdvancePerTriangle == 4);
+        target->putBackVertices(maxInnerTriangles + maxCubics - totalInstanceCount,
+                                sizeof(SkPoint) * 4);
+        if (totalInstanceCount) {
+            this->prepareIndirectOuterCubicsAndTriangles(target, *resolveLevelCounter, vertexData,
+                                                         triangleCount);
+        }
+    } else {
+        SkASSERT(vertexAdvancePerTriangle == 3);
+        target->putBackVertices(maxInnerTriangles - triangleCount, sizeof(SkPoint) * 3);
+        fTriangleVertexCount = triangleCount * 3;
+        if (fTriangleVertexCount) {
+            fDoStencilTriangleBuffer = true;
+        }
+        if (resolveLevelCounter) {
+            this->prepareIndirectOuterCubics(target, *resolveLevelCounter);
+        } else {
+            this->prepareTessellatedOuterCubics(target, numCountedCurves);
+        }
     }
 }
 
@@ -177,8 +252,133 @@
     out[3] = pts[2];
 }
 
-void GrTessellatePathOp::prepareOuterCubics(GrMeshDrawOp::Target* target, int numCountedCurves,
-                                            CubicDataAlignment alignment) {
+void GrTessellatePathOp::prepareIndirectOuterCubics(
+        GrMeshDrawOp::Target* target, const GrResolveLevelCounter& resolveLevelCounter) {
+    SkASSERT(resolveLevelCounter.totalCubicInstanceCount() >= 0);
+    if (resolveLevelCounter.totalCubicInstanceCount() == 0) {
+        return;
+    }
+    // Allocate a buffer to store the cubic data.
+    SkPoint* cubicData;
+    int baseInstance;
+    cubicData = static_cast<SkPoint*>(target->makeVertexSpace(
+            sizeof(SkPoint) * 4, resolveLevelCounter.totalCubicInstanceCount(), &fCubicBuffer,
+            &baseInstance));
+    if (!cubicData) {
+        return;
+    }
+    fBaseCubicVertex = baseInstance * 4;
+    this->prepareIndirectOuterCubicsAndTriangles(target, resolveLevelCounter, cubicData,
+                                                 /*numTrianglesAtBeginningOfData=*/0);
+}
+
+void GrTessellatePathOp::prepareIndirectOuterCubicsAndTriangles(
+        GrMeshDrawOp::Target* target, const GrResolveLevelCounter& resolveLevelCounter,
+        SkPoint* cubicData, int numTrianglesAtBeginningOfData) {
+    SkASSERT(numTrianglesAtBeginningOfData + resolveLevelCounter.totalCubicInstanceCount() > 0);
+    SkASSERT(!fStencilCubicsShader);
+    SkASSERT(cubicData);
+
+    // Here we treat fCubicBuffer as an instance buffer. It should have been prepared with the base
+    // vertex on an instance boundary in order to accommodate this.
+    SkASSERT(fBaseCubicVertex % 4 == 0);
+    int baseInstance = fBaseCubicVertex >> 2;
+
+    // Start preparing the indirect draw buffer.
+    fIndirectDrawCount = resolveLevelCounter.totalCubicIndirectDrawCount();
+    if (numTrianglesAtBeginningOfData) {
+        ++fIndirectDrawCount;  // Add an indirect draw for the triangles at the beginning.
+    }
+
+    // Allocate space for the GrDrawIndexedIndirectCommand structs.
+    GrDrawIndexedIndirectCommand* indirectData = target->makeDrawIndexedIndirectSpace(
+            fIndirectDrawCount, &fIndirectDrawBuffer, &fIndirectDrawOffset);
+    if (!indirectData) {
+        SkASSERT(!fIndirectDrawBuffer);
+        return;
+    }
+
+    // Fill out the GrDrawIndexedIndirectCommand structs and determine the starting instance data
+    // location at each resolve level.
+    SkPoint* instanceLocations[kMaxResolveLevel + 1];
+    int indirectIdx = 0;
+    int runningInstanceCount = 0;
+    if (numTrianglesAtBeginningOfData) {
+        // The caller has already packed "triangleInstanceCount" triangles into 4-point instances
+        // at the beginning of the instance buffer. Add a special-case indirect draw here that will
+        // emit the triangles [P0, P1, P2] from these 4-point instances.
+        indirectData[0] = GrMiddleOutCubicShader::MakeDrawTrianglesIndirectCmd(
+                numTrianglesAtBeginningOfData, baseInstance);
+        indirectIdx = 1;
+        runningInstanceCount = numTrianglesAtBeginningOfData;
+    }
+    for (int resolveLevel = 1; resolveLevel <= kMaxResolveLevel; ++resolveLevel) {
+        instanceLocations[resolveLevel] = cubicData + runningInstanceCount * 4;
+        if (int instanceCountAtCurrLevel = resolveLevelCounter[resolveLevel]) {
+            indirectData[indirectIdx++] = GrMiddleOutCubicShader::MakeDrawCubicsIndirectCmd(
+                    resolveLevel, instanceCountAtCurrLevel, baseInstance + runningInstanceCount);
+            runningInstanceCount += instanceCountAtCurrLevel;
+        }
+    }
+
+#ifdef SK_DEBUG
+    SkASSERT(indirectIdx == fIndirectDrawCount);
+    SkASSERT(runningInstanceCount == numTrianglesAtBeginningOfData +
+                                     resolveLevelCounter.totalCubicInstanceCount());
+    SkASSERT(fIndirectDrawCount > 0);
+
+    SkPoint* endLocations[kMaxResolveLevel + 1];
+    memcpy(endLocations, instanceLocations + 1, kMaxResolveLevel * sizeof(SkPoint*));
+    int totalInstanceCount = numTrianglesAtBeginningOfData +
+                             resolveLevelCounter.totalCubicInstanceCount();
+    endLocations[kMaxResolveLevel] = cubicData + totalInstanceCount * 4;
+#endif
+
+    fCubicVertexCount = numTrianglesAtBeginningOfData * 4;
+
+    if (resolveLevelCounter.totalCubicInstanceCount()) {
+        GrVectorXform xform(fViewMatrix);
+        for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) {
+            int level;
+            switch (verb) {
+                default:
+                    continue;
+                case SkPathVerb::kQuad:
+                    level = GrWangsFormula::quadratic_log2(kTessellationIntolerance, pts, xform);
+                    if (level == 0) {
+                        continue;
+                    }
+                    level = std::min(level, kMaxResolveLevel);
+                    quad2cubic(pts, instanceLocations[level]);
+                    break;
+                case SkPathVerb::kCubic:
+                    level = GrWangsFormula::cubic_log2(kTessellationIntolerance, pts, xform);
+                    if (level == 0) {
+                        continue;
+                    }
+                    level = std::min(level, kMaxResolveLevel);
+                    memcpy(instanceLocations[level], pts, sizeof(SkPoint) * 4);
+                    break;
+            }
+            instanceLocations[level] += 4;
+            fCubicVertexCount += 4;
+        }
+    }
+
+#ifdef SK_DEBUG
+    for (int i = 1; i <= kMaxResolveLevel; ++i) {
+        SkASSERT(instanceLocations[i] == endLocations[i]);
+    }
+    SkASSERT(fCubicVertexCount == (numTrianglesAtBeginningOfData +
+                                   resolveLevelCounter.totalCubicInstanceCount()) * 4);
+#endif
+
+    fStencilCubicsShader = target->allocator()->make<GrMiddleOutCubicShader>(fViewMatrix);
+}
+
+void GrTessellatePathOp::prepareTessellatedOuterCubics(GrMeshDrawOp::Target* target,
+                                                       int numCountedCurves) {
+    SkASSERT(numCountedCurves >= 0);
     SkASSERT(!fCubicBuffer);
     SkASSERT(!fStencilCubicsShader);
 
@@ -186,41 +386,34 @@
         return;
     }
 
-    bool instanceAligned = (alignment == CubicDataAlignment::kInstanceBoundary);
-    int instanceOrVertexStride = (instanceAligned) ? sizeof(SkPoint) * 4 : sizeof(SkPoint);
-    int instanceOrVertexCount = (instanceAligned) ? numCountedCurves : numCountedCurves * 4;
-    int baseInstanceOrVertex;
-
     auto* vertexData = static_cast<SkPoint*>(target->makeVertexSpace(
-            instanceOrVertexStride, instanceOrVertexCount, &fCubicBuffer, &baseInstanceOrVertex));
+            sizeof(SkPoint), numCountedCurves * 4, &fCubicBuffer, &fBaseCubicVertex));
     if (!vertexData) {
         return;
     }
-    fBaseCubicVertex = (instanceAligned) ? baseInstanceOrVertex * 4 : baseInstanceOrVertex;
     fCubicVertexCount = 0;
 
     for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) {
         switch (verb) {
+            default:
+                continue;
             case SkPathVerb::kQuad:
                 SkASSERT(fCubicVertexCount < numCountedCurves * 4);
                 quad2cubic(pts, vertexData + fCubicVertexCount);
-                fCubicVertexCount += 4;
                 break;
             case SkPathVerb::kCubic:
                 SkASSERT(fCubicVertexCount < numCountedCurves * 4);
                 memcpy(vertexData + fCubicVertexCount, pts, sizeof(SkPoint) * 4);
-                fCubicVertexCount += 4;
-                break;
-            default:
                 break;
         }
+        fCubicVertexCount += 4;
     }
     SkASSERT(fCubicVertexCount == numCountedCurves * 4);
 
-    fStencilCubicsShader = target->allocator()->make<GrStencilCubicShader>(fViewMatrix);
+    fStencilCubicsShader = target->allocator()->make<GrTessellateCubicShader>(fViewMatrix);
 }
 
-void GrTessellatePathOp::prepareCubicWedges(GrMeshDrawOp::Target* target) {
+void GrTessellatePathOp::prepareTessellatedCubicWedges(GrMeshDrawOp::Target* target) {
     SkASSERT(!fCubicBuffer);
     SkASSERT(!fStencilCubicsShader);
 
@@ -275,18 +468,18 @@
     vertexAlloc.unlock(fCubicVertexCount);
 
     if (fCubicVertexCount) {
-        fStencilCubicsShader = target->allocator()->make<GrStencilWedgeShader>(fViewMatrix);
+        fStencilCubicsShader = target->allocator()->make<GrTessellateWedgeShader>(fViewMatrix);
     }
 }
 
-void GrTessellatePathOp::onExecute(GrOpFlushState* state, const SkRect& chainBounds) {
-    this->drawStencilPass(state);
+void GrTessellatePathOp::onExecute(GrOpFlushState* flushState, const SkRect& chainBounds) {
+    this->drawStencilPass(flushState);
     if (!(Flags::kStencilOnly & fFlags)) {
-        this->drawCoverPass(state);
+        this->drawCoverPass(flushState);
     }
 }
 
-void GrTessellatePathOp::drawStencilPass(GrOpFlushState* state) {
+void GrTessellatePathOp::drawStencilPass(GrOpFlushState* flushState) {
     // Increments clockwise triangles and decrements counterclockwise. Used for "winding" fill.
     constexpr static GrUserStencilSettings kIncrDecrStencil(
         GrUserStencilSettings::StaticInitSeparate<
@@ -311,41 +504,49 @@
     if (GrAAType::kNone != fAAType) {
         initArgs.fInputFlags |= GrPipeline::InputFlags::kHWAntialias;
     }
-    if (state->caps().wireframeSupport() && (Flags::kWireframe & fFlags)) {
+    if (flushState->caps().wireframeSupport() && (Flags::kWireframe & fFlags)) {
         initArgs.fInputFlags |= GrPipeline::InputFlags::kWireframe;
     }
     SkASSERT(SkPathFillType::kWinding == fPath.getFillType() ||
              SkPathFillType::kEvenOdd == fPath.getFillType());
     initArgs.fUserStencil = (SkPathFillType::kWinding == fPath.getFillType()) ?
             &kIncrDecrStencil : &kInvertStencil;
-    initArgs.fCaps = &state->caps();
+    initArgs.fCaps = &flushState->caps();
     GrPipeline pipeline(initArgs, GrDisableColorXPFactory::MakeXferProcessor(),
-                        state->appliedHardClip());
+                        flushState->appliedHardClip());
 
     if (fDoStencilTriangleBuffer) {
         SkASSERT(fTriangleBuffer);
         GrStencilTriangleShader stencilTriangleShader(fViewMatrix);
-        GrPathShader::ProgramInfo programInfo(state->writeView(), &pipeline,
+        GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline,
                                               &stencilTriangleShader);
-        state->bindPipelineAndScissorClip(programInfo, this->bounds());
-        state->bindBuffers(nullptr, nullptr, fTriangleBuffer.get());
-        state->draw(fTriangleVertexCount, fBaseTriangleVertex);
+        flushState->bindPipelineAndScissorClip(programInfo, this->bounds());
+        flushState->bindBuffers(nullptr, nullptr, fTriangleBuffer.get());
+        flushState->draw(fTriangleVertexCount, fBaseTriangleVertex);
     }
 
     if (fStencilCubicsShader) {
-        GrPathShader::ProgramInfo programInfo(state->writeView(), &pipeline, fStencilCubicsShader);
-        state->bindPipelineAndScissorClip(programInfo, this->bounds());
-        state->bindBuffers(nullptr, nullptr, fCubicBuffer.get());
-        state->draw(fCubicVertexCount, fBaseCubicVertex);
-    }
-
-    // http://skbug.com/9739
-    if (state->caps().requiresManualFBBarrierAfterTessellatedStencilDraw()) {
-        state->gpu()->insertManualFramebufferBarrier();
+        SkASSERT(fCubicBuffer);
+        GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline,
+                                              fStencilCubicsShader);
+        flushState->bindPipelineAndScissorClip(programInfo, this->bounds());
+        if (fIndirectDrawBuffer) {
+            auto indexBuffer = GrMiddleOutCubicShader::FindOrMakeMiddleOutIndexBuffer(
+                    flushState->resourceProvider());
+            flushState->bindBuffers(indexBuffer.get(), fCubicBuffer.get(), nullptr);
+            flushState->drawIndexedIndirect(fIndirectDrawBuffer.get(), fIndirectDrawOffset,
+                                            fIndirectDrawCount);
+        } else {
+            flushState->bindBuffers(nullptr, nullptr, fCubicBuffer.get());
+            flushState->draw(fCubicVertexCount, fBaseCubicVertex);
+            if (flushState->caps().requiresManualFBBarrierAfterTessellatedStencilDraw()) {
+                flushState->gpu()->insertManualFramebufferBarrier();  // http://skbug.com/9739
+            }
+        }
     }
 }
 
-void GrTessellatePathOp::drawCoverPass(GrOpFlushState* state) {
+void GrTessellatePathOp::drawCoverPass(GrOpFlushState* flushState) {
     // Allows non-zero stencil values to pass and write a color, and resets the stencil value back
     // to zero; discards immediately on stencil values of zero.
     // NOTE: It's ok to not check the clip here because the previous stencil pass only wrote to
@@ -362,7 +563,7 @@
     GrPipeline::InitArgs initArgs;
     if (GrAAType::kNone != fAAType) {
         initArgs.fInputFlags |= GrPipeline::InputFlags::kHWAntialias;
-        if (1 == state->proxy()->numSamples()) {
+        if (1 == flushState->proxy()->numSamples()) {
             SkASSERT(GrAAType::kCoverage == fAAType);
             // We are mixed sampled. Use conservative raster to make the sample coverage mask 100%
             // at every fragment. This way we will still get a double hit on shared edges, but
@@ -371,10 +572,10 @@
             initArgs.fInputFlags |= GrPipeline::InputFlags::kConservativeRaster;
         }
     }
-    initArgs.fCaps = &state->caps();
-    initArgs.fDstProxyView = state->drawOpArgs().dstProxyView();
-    initArgs.fWriteSwizzle = state->drawOpArgs().writeSwizzle();
-    GrPipeline pipeline(initArgs, std::move(fProcessors), state->detachAppliedClip());
+    initArgs.fCaps = &flushState->caps();
+    initArgs.fDstProxyView = flushState->drawOpArgs().dstProxyView();
+    initArgs.fWriteSwizzle = flushState->drawOpArgs().writeSwizzle();
+    GrPipeline pipeline(initArgs, std::move(fProcessors), flushState->detachAppliedClip());
 
     if (fDoFillTriangleBuffer) {
         SkASSERT(fTriangleBuffer);
@@ -421,29 +622,32 @@
         }
 
         GrFillTriangleShader fillTriangleShader(fViewMatrix, fColor);
-        GrPathShader::ProgramInfo programInfo(state->writeView(), &pipeline, &fillTriangleShader);
-        state->bindPipelineAndScissorClip(programInfo, this->bounds());
-        state->bindTextures(fillTriangleShader, nullptr, pipeline);
-        state->bindBuffers(nullptr, nullptr, fTriangleBuffer.get());
-        state->draw(fTriangleVertexCount, fBaseTriangleVertex);
+        GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline,
+                                              &fillTriangleShader);
+        flushState->bindPipelineAndScissorClip(programInfo, this->bounds());
+        flushState->bindTextures(fillTriangleShader, nullptr, pipeline);
+        flushState->bindBuffers(nullptr, nullptr, fTriangleBuffer.get());
+        flushState->draw(fTriangleVertexCount, fBaseTriangleVertex);
 
         if (fStencilCubicsShader) {
+            SkASSERT(fCubicBuffer);
+
             // At this point, every pixel is filled in except the ones touched by curves. Issue a
             // final cover pass over the curves by drawing their convex hulls. This will fill in any
             // remaining samples and reset the stencil buffer.
             pipeline.setUserStencil(&kTestAndResetStencil);
             GrFillCubicHullShader fillCubicHullShader(fViewMatrix, fColor);
-            GrPathShader::ProgramInfo programInfo(state->writeView(), &pipeline,
+            GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline,
                                                   &fillCubicHullShader);
-            state->bindPipelineAndScissorClip(programInfo, this->bounds());
-            state->bindTextures(fillCubicHullShader, nullptr, pipeline);
+            flushState->bindPipelineAndScissorClip(programInfo, this->bounds());
+            flushState->bindTextures(fillCubicHullShader, nullptr, pipeline);
 
             // Here we treat fCubicBuffer as an instance buffer. It should have been prepared with
             // the base vertex on an instance boundary in order to accommodate this.
             SkASSERT((fCubicVertexCount % 4) == 0);
             SkASSERT((fBaseCubicVertex % 4) == 0);
-            state->bindBuffers(nullptr, fCubicBuffer.get(), nullptr);
-            state->drawInstanced(fCubicVertexCount >> 2, fBaseCubicVertex >> 2, 4, 0);
+            flushState->bindBuffers(nullptr, fCubicBuffer.get(), nullptr);
+            flushState->drawInstanced(fCubicVertexCount >> 2, fBaseCubicVertex >> 2, 4, 0);
         }
         return;
     }
@@ -451,9 +655,10 @@
     // There are no triangles to fill. Just draw a bounding box.
     pipeline.setUserStencil(&kTestAndResetStencil);
     GrFillBoundingBoxShader fillBoundingBoxShader(fViewMatrix, fColor, fPath.getBounds());
-    GrPathShader::ProgramInfo programInfo(state->writeView(), &pipeline, &fillBoundingBoxShader);
-    state->bindPipelineAndScissorClip(programInfo, this->bounds());
-    state->bindTextures(fillBoundingBoxShader, nullptr, pipeline);
-    state->bindBuffers(nullptr, nullptr, nullptr);
-    state->draw(4, 0);
+    GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline,
+                                          &fillBoundingBoxShader);
+    flushState->bindPipelineAndScissorClip(programInfo, this->bounds());
+    flushState->bindTextures(fillBoundingBoxShader, nullptr, pipeline);
+    flushState->bindBuffers(nullptr, nullptr, nullptr);
+    flushState->draw(4, 0);
 }
diff --git a/src/gpu/tessellate/GrTessellatePathOp.h b/src/gpu/tessellate/GrTessellatePathOp.h
index 8ee6d21..89c9f7f 100644
--- a/src/gpu/tessellate/GrTessellatePathOp.h
+++ b/src/gpu/tessellate/GrTessellatePathOp.h
@@ -12,10 +12,11 @@
 
 class GrAppliedHardClip;
 class GrStencilPathShader;
+class GrResolveLevelCounter;
 
-// Renders paths using a hybrid Red Book "stencil, then cover" method. Curves get linearized by
-// GPU tessellation shaders. This Op doesn't apply analytic AA, so it requires a render target that
-// supports either MSAA or mixed samples if AA is desired.
+// Renders paths using a hybrid "Red Book" (stencil, then cover) method. Curves get linearized by
+// either GPU tessellation shaders or indirect draws. This Op doesn't apply analytic AA, so it
+// requires a render target that supports either MSAA or mixed samples if AA is desired.
 class GrTessellatePathOp : public GrDrawOp {
 public:
     enum class Flags {
@@ -66,25 +67,41 @@
     // and this is not an option as it would introduce T-junctions with the outer cubics.
     bool prepareNonOverlappingInnerTriangles(GrMeshDrawOp::Target*, int* numCountedCurves);
 
-    // Produces a "Red Book" style triangulation of the SkPath's inner polygon(s). The inner
-    // polygons connect the endpoints of each verb. (i.e., they are the path that would result from
-    // collapsing all curves to single lines.) Stencilled together with the outer cubics, these
-    // define the complete path.
+    // Produces a "Red Book" style triangulation of the SkPath's inner polygon(s) using a
+    // "middle-out" topology (See GrMiddleOutPolygonTriangulator), and then prepares outer cubics in
+    // the cubic buffer. The inner triangles and outer cubics stencilled together define the
+    // complete path.
     //
-    // This method emits the inner triangles with a "middle-out" topology. Middle-out can reduce
-    // the load on the rasterizer by a great deal as compared to a linear triangle strip or fan.
-    // See GrMiddleOutPolygonTriangulator.
-    void prepareMiddleOutInnerTriangles(GrMeshDrawOp::Target*, int* numCountedCurves);
+    // If a resolveLevel counter is provided, this method resets it and uses it to count and
+    // prepares the outer cubics as indirect draws. Otherwise they are prepared as hardware
+    // tessellation patches.
+    //
+    // If drawTrianglesAsIndirectCubicDraw is true, then the resolveLevel counter must be non-null,
+    // and we express the inner triangles as an indirect cubic draw and sneak them in alongside the
+    // other cubic draws.
+    void prepareMiddleOutTrianglesAndCubics(GrMeshDrawOp::Target*, GrResolveLevelCounter* = nullptr,
+                                            bool drawTrianglesAsIndirectCubicDraw = false);
 
-    enum class CubicDataAlignment : bool {
-        kVertexBoundary,
-        kInstanceBoundary
-    };
+    // Prepares a list of indirect draw commands and instance data for the path's "outer cubics",
+    // converting any quadratics to cubics. An outer cubic is an independent, 4-point closed contour
+    // consisting of a single cubic curve. Stencilled together with the inner triangles, these
+    // define the complete path.
+    void prepareIndirectOuterCubics(GrMeshDrawOp::Target*, const GrResolveLevelCounter&);
 
-    // Writes an array of "outer" cubics from each bezier in the SkPath, converting any quadratics
-    // to cubics. An outer cubic is an independent, 4-point closed contour consisting of a single
-    // cubic curve. Stencilled together with the inner triangles, these define the complete path.
-    void prepareOuterCubics(GrMeshDrawOp::Target*, int numCountedCurves, CubicDataAlignment);
+    // For performance reasons we can often express triangles as an indirect cubic draw and sneak
+    // them in alongside the other indirect draws. This prepareIndirectOuterCubics variant allows
+    // the caller to provide a mapped cubic buffer with triangles already written into 4-point
+    // instances at the beginning. If numTrianglesAtBeginningOfData is nonzero, we add an extra
+    // indirect draw that renders these triangles.
+    void prepareIndirectOuterCubicsAndTriangles(GrMeshDrawOp::Target*, const GrResolveLevelCounter&,
+                                                SkPoint* cubicData,
+                                                int numTrianglesAtBeginningOfData);
+
+    // Writes an array of "outer cubic" tessellation patches from each bezier in the SkPath,
+    // converting any quadratics to cubics. An outer cubic is an independent, 4-point closed contour
+    // consisting of a single cubic curve. Stencilled together with the inner triangles, these
+    // define the complete path.
+    void prepareTessellatedOuterCubics(GrMeshDrawOp::Target*, int numCountedCurves);
 
     // Writes an array of cubic "wedges" from the SkPath, converting any lines or quadratics to
     // cubics. A wedge is an independent, 5-point closed contour consisting of 4 cubic control
@@ -92,7 +109,7 @@
     // stencilled, these wedges alone define the complete path.
     //
     // TODO: Eventually we want to use rational cubic wedges in order to support conics.
-    void prepareCubicWedges(GrMeshDrawOp::Target*);
+    void prepareTessellatedCubicWedges(GrMeshDrawOp::Target*);
 
     void onExecute(GrOpFlushState*, const SkRect& chainBounds) override;
     void drawStencilPass(GrOpFlushState*);
@@ -137,6 +154,13 @@
     int fCubicVertexCount;
     GrStencilPathShader* fStencilCubicsShader = nullptr;
 
+    // If fIndirectDrawBuffer is non-null, then we issue an indexed-indirect draw instead of using
+    // hardware tessellation. This is oftentimes faster than tessellation, and other times it serves
+    // as a polyfill when tessellation just isn't supported.
+    sk_sp<const GrBuffer> fIndirectDrawBuffer;
+    size_t fIndirectDrawOffset;
+    int fIndirectDrawCount;
+
     friend class GrOpMemoryPool;  // For ctor.
 
 public:
diff --git a/src/gpu/tessellate/GrTessellationPathRenderer.cpp b/src/gpu/tessellate/GrTessellationPathRenderer.cpp
index 85fa43e..ab21f28 100644
--- a/src/gpu/tessellate/GrTessellationPathRenderer.cpp
+++ b/src/gpu/tessellate/GrTessellationPathRenderer.cpp
@@ -37,8 +37,6 @@
 
 GrPathRenderer::CanDrawPath GrTessellationPathRenderer::onCanDrawPath(
         const CanDrawPathArgs& args) const {
-    // This class should not have been added to the chain without tessellation support.
-    SkASSERT(args.fCaps->shaderCaps()->tessellationSupport());
     if (!args.fShape->style().isSimpleFill() || args.fShape->inverseFilled() ||
         args.fViewMatrix->hasPerspective()) {
         return CanDrawPath::kNo;