| 1 | /* |
| 2 | * Copyright 2019 Google Inc. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #ifndef GrOpsTask_DEFINED |
| 9 | #define GrOpsTask_DEFINED |
| 10 | |
| 11 | #include "include/core/SkMatrix.h" |
| 12 | #include "include/core/SkRefCnt.h" |
| 13 | #include "include/core/SkStrokeRec.h" |
| 14 | #include "include/core/SkTypes.h" |
| 15 | #include "include/private/GrRecordingContext.h" |
| 16 | #include "include/private/SkColorData.h" |
| 17 | #include "include/private/SkTArray.h" |
| 18 | #include "include/private/SkTDArray.h" |
| 19 | #include "src/core/SkArenaAlloc.h" |
| 20 | #include "src/core/SkClipStack.h" |
| 21 | #include "src/core/SkStringUtils.h" |
| 22 | #include "src/core/SkTLazy.h" |
| 23 | #include "src/gpu/GrAppliedClip.h" |
| 24 | #include "src/gpu/GrPathRendering.h" |
| 25 | #include "src/gpu/GrPrimitiveProcessor.h" |
| 26 | #include "src/gpu/GrRenderTask.h" |
| 27 | #include "src/gpu/ops/GrDrawOp.h" |
| 28 | #include "src/gpu/ops/GrOp.h" |
| 29 | |
| 30 | class GrAuditTrail; |
| 31 | class GrCaps; |
| 32 | class GrClearOp; |
| 33 | class GrGpuBuffer; |
| 34 | class GrRenderTargetProxy; |
| 35 | |
| 36 | class GrOpsTask : public GrRenderTask { |
| 37 | private: |
| 38 | using DstProxyView = GrXferProcessor::DstProxyView; |
| 39 | |
| 40 | public: |
| 41 | // The Arenas must outlive the GrOpsTask, either by preserving the context that owns |
| 42 | // the pool, or by moving the pool to the DDL that takes over the GrOpsTask. |
| 43 | GrOpsTask(GrRecordingContext::Arenas, GrSurfaceProxyView, GrAuditTrail*); |
| 44 | ~GrOpsTask() override; |
| 45 | |
| 46 | GrOpsTask* asOpsTask() override { return this; } |
| 47 | |
| 48 | bool isEmpty() const { return fOpChains.empty(); } |
| 49 | |
| 50 | /** |
| 51 | * Empties the draw buffer of any queued up draws. |
| 52 | */ |
| 53 | void endFlush() override; |
| 54 | |
| 55 | void onPrePrepare(GrRecordingContext*) override; |
| 56 | /** |
| 57 | * Together these two functions flush all queued up draws to GrCommandBuffer. The return value |
| 58 | * of executeOps() indicates whether any commands were actually issued to the GPU. |
| 59 | */ |
| 60 | void onPrepare(GrOpFlushState* flushState) override; |
| 61 | bool onExecute(GrOpFlushState* flushState) override; |
| 62 | |
| 63 | void addSampledTexture(GrSurfaceProxy* proxy) { |
| 64 | // This function takes a GrSurfaceProxy because all subsequent uses of the proxy do not |
| 65 | // require the specifics of GrTextureProxy, so this avoids a number of unnecessary virtual |
| 66 | // asTextureProxy() calls. However, sampling the proxy implicitly requires that the proxy |
| 67 | // be a texture. Eventually, when proxies are a unified type with flags, this can just |
| 68 | // assert that capability. |
| 69 | SkASSERT(proxy->asTextureProxy()); |
| 70 | fSampledProxies.push_back(proxy); |
| 71 | } |
| 72 | |
| 73 | void addOp(std::unique_ptr<GrOp> op, GrTextureResolveManager textureResolveManager, |
| 74 | const GrCaps& caps) { |
| 75 | auto addDependency = [ textureResolveManager, &caps, this ] ( |
| 76 | GrSurfaceProxy* p, GrMipMapped mipmapped) { |
| 77 | this->addDependency(p, mipmapped, textureResolveManager, caps); |
| 78 | }; |
| 79 | |
| 80 | op->visitProxies(addDependency); |
| 81 | |
| 82 | this->recordOp(std::move(op), GrProcessorSet::EmptySetAnalysis(), nullptr, nullptr, caps); |
| 83 | } |
| 84 | |
| 85 | void addWaitOp(std::unique_ptr<GrOp> op, GrTextureResolveManager textureResolveManager, |
| 86 | const GrCaps& caps) { |
| 87 | fHasWaitOp = true; |
| 88 | this->addOp(std::move(op), textureResolveManager, caps); |
| 89 | } |
| 90 | |
| 91 | void addDrawOp(std::unique_ptr<GrDrawOp> op, const GrProcessorSet::Analysis& processorAnalysis, |
| 92 | GrAppliedClip&& clip, const DstProxyView& dstProxyView, |
| 93 | GrTextureResolveManager textureResolveManager, const GrCaps& caps) { |
| 94 | auto addDependency = [ textureResolveManager, &caps, this ] ( |
| 95 | GrSurfaceProxy* p, GrMipMapped mipmapped) { |
| 96 | this->addSampledTexture(p); |
| 97 | this->addDependency(p, mipmapped, textureResolveManager, caps); |
| 98 | }; |
| 99 | |
| 100 | op->visitProxies(addDependency); |
| 101 | clip.visitProxies(addDependency); |
| 102 | if (dstProxyView.proxy()) { |
| 103 | this->addSampledTexture(dstProxyView.proxy()); |
| 104 | addDependency(dstProxyView.proxy(), GrMipMapped::kNo); |
| 105 | } |
| 106 | |
| 107 | this->recordOp(std::move(op), processorAnalysis, clip.doesClip() ? &clip : nullptr, |
| 108 | &dstProxyView, caps); |
| 109 | } |
| 110 | |
| 111 | void discard(); |
| 112 | |
| 113 | SkDEBUGCODE(void dump(bool printDependencies) const override;) |
| 114 | SkDEBUGCODE(int numClips() const override { return fNumClips; }) |
| 115 | SkDEBUGCODE(void visitProxies_debugOnly(const GrOp::VisitProxyFunc&) const override;) |
| 116 | |
| 117 | #if GR_TEST_UTILS |
| 118 | int numOpChains() const { return fOpChains.count(); } |
| 119 | const GrOp* getChain(int index) const { return fOpChains[index].head(); } |
| 120 | #endif |
| 121 | |
| 122 | private: |
| 123 | bool isNoOp() const { |
| 124 | // TODO: GrLoadOp::kDiscard (i.e., storing a discard) should also be grounds for skipping |
| 125 | // execution. We currently don't because of Vulkan. See http://skbug.com/9373. |
| 126 | // |
| 127 | // TODO: We should also consider stencil load/store here. We get away with it for now |
| 128 | // because we never discard stencil buffers. |
| 129 | return fOpChains.empty() && GrLoadOp::kLoad == fColorLoadOp; |
| 130 | } |
| 131 | |
| 132 | void deleteOps(); |
| 133 | |
| 134 | enum class StencilContent { |
| 135 | kDontCare, |
| 136 | kUserBitsCleared, // User bits: cleared |
| 137 | // Clip bit: don't care (Ganesh always pre-clears the clip bit.) |
| 138 | kPreserved |
| 139 | }; |
| 140 | |
| 141 | // Lets the caller specify what the content of the stencil buffer should be at the beginning |
| 142 | // of the render pass. |
| 143 | // |
| 144 | // When requesting kClear: Tilers will load the stencil buffer with a "clear" op; non-tilers |
| 145 | // will clear the stencil on first load, and then preserve it on subsequent loads. (Preserving |
| 146 | // works because renderTargetContexts are required to leave the user bits in a cleared state |
| 147 | // once finished.) |
| 148 | // |
| 149 | // NOTE: initialContent must not be kClear if caps.performStencilClearsAsDraws() is true. |
| 150 | void setInitialStencilContent(StencilContent initialContent) { |
| 151 | fInitialStencilContent = initialContent; |
| 152 | } |
| 153 | |
| 154 | // If a renderTargetContext splits its opsTask, it uses this method to guarantee stencil values |
| 155 | // get preserved across its split tasks. |
| 156 | void setMustPreserveStencil() { fMustPreserveStencil = true; } |
| 157 | |
| 158 | // Must only be called if native color buffer clearing is enabled. |
| 159 | void setColorLoadOp(GrLoadOp op, const SkPMColor4f& color); |
| 160 | // Sets the clear color to transparent black |
| 161 | void setColorLoadOp(GrLoadOp op) { |
| 162 | static const SkPMColor4f kDefaultClearColor = {0.f, 0.f, 0.f, 0.f}; |
| 163 | this->setColorLoadOp(op, kDefaultClearColor); |
| 164 | } |
| 165 | |
| 166 | enum class CanDiscardPreviousOps : bool { |
| 167 | kYes = true, |
| 168 | kNo = false |
| 169 | }; |
| 170 | |
| 171 | // Perform book-keeping for a fullscreen clear, regardless of how the clear is implemented later |
| 172 | // (i.e. setColorLoadOp(), adding a ClearOp, or adding a GrFillRectOp that covers the device). |
| 173 | // Returns true if the clear can be converted into a load op (barring device caps). |
| 174 | bool resetForFullscreenClear(CanDiscardPreviousOps); |
| 175 | |
| 176 | class OpChain { |
| 177 | public: |
| 178 | OpChain(const OpChain&) = delete; |
| 179 | OpChain& operator=(const OpChain&) = delete; |
| 180 | OpChain(std::unique_ptr<GrOp>, GrProcessorSet::Analysis, GrAppliedClip*, |
| 181 | const DstProxyView*); |
| 182 | |
| 183 | ~OpChain() { |
| 184 | // The ops are stored in a GrMemoryPool and must be explicitly deleted via the pool. |
| 185 | SkASSERT(fList.empty()); |
| 186 | } |
| 187 | |
| 188 | void visitProxies(const GrOp::VisitProxyFunc&) const; |
| 189 | |
| 190 | GrOp* head() const { return fList.head(); } |
| 191 | |
| 192 | GrAppliedClip* appliedClip() const { return fAppliedClip; } |
| 193 | const DstProxyView& dstProxyView() const { return fDstProxyView; } |
| 194 | const SkRect& bounds() const { return fBounds; } |
| 195 | |
| 196 | // Deletes all the ops in the chain via the pool. |
| 197 | void deleteOps(GrOpMemoryPool* pool); |
| 198 | |
| 199 | // Attempts to move the ops from the passed chain to this chain at the head. Also attempts |
| 200 | // to merge ops between the chains. Upon success the passed chain is empty. |
| 201 | // Fails when the chains aren't of the same op type, have different clips or dst proxies. |
| 202 | bool prependChain(OpChain*, const GrCaps&, GrRecordingContext::Arenas*, GrAuditTrail*); |
| 203 | |
| 204 | // Attempts to add 'op' to this chain either by merging or adding to the tail. Returns |
| 205 | // 'op' to the caller upon failure, otherwise null. Fails when the op and chain aren't of |
| 206 | // the same op type, have different clips or dst proxies. |
| 207 | std::unique_ptr<GrOp> appendOp(std::unique_ptr<GrOp> op, GrProcessorSet::Analysis, |
| 208 | const DstProxyView*, const GrAppliedClip*, const GrCaps&, |
| 209 | GrRecordingContext::Arenas*, GrAuditTrail*); |
| 210 | |
| 211 | void setSkipExecuteFlag() { fSkipExecute = true; } |
| 212 | bool shouldExecute() const { |
| 213 | return SkToBool(this->head()) && !fSkipExecute; |
| 214 | } |
| 215 | |
| 216 | private: |
| 217 | class List { |
| 218 | public: |
| 219 | List() = default; |
| 220 | List(std::unique_ptr<GrOp>); |
| 221 | List(List&&); |
| 222 | List& operator=(List&& that); |
| 223 | |
| 224 | bool empty() const { return !SkToBool(fHead); } |
| 225 | GrOp* head() const { return fHead.get(); } |
| 226 | GrOp* tail() const { return fTail; } |
| 227 | |
| 228 | std::unique_ptr<GrOp> popHead(); |
| 229 | std::unique_ptr<GrOp> removeOp(GrOp* op); |
| 230 | void pushHead(std::unique_ptr<GrOp> op); |
| 231 | void pushTail(std::unique_ptr<GrOp>); |
| 232 | |
| 233 | void validate() const; |
| 234 | |
| 235 | private: |
| 236 | std::unique_ptr<GrOp> fHead; |
| 237 | GrOp* fTail = nullptr; |
| 238 | }; |
| 239 | |
| 240 | void validate() const; |
| 241 | |
| 242 | bool tryConcat(List*, GrProcessorSet::Analysis, const DstProxyView&, const GrAppliedClip*, |
| 243 | const SkRect& bounds, const GrCaps&, GrRecordingContext::Arenas*, |
| 244 | GrAuditTrail*); |
| 245 | static List DoConcat(List, List, const GrCaps&, GrRecordingContext::Arenas*, GrAuditTrail*); |
| 246 | |
| 247 | List fList; |
| 248 | GrProcessorSet::Analysis fProcessorAnalysis; |
| 249 | DstProxyView fDstProxyView; |
| 250 | GrAppliedClip* fAppliedClip; |
| 251 | SkRect fBounds; |
| 252 | |
| 253 | // We set this flag to true if any of the ops' proxies fail to instantiate so that we know |
| 254 | // not to try and draw the op. |
| 255 | bool fSkipExecute = false; |
| 256 | }; |
| 257 | |
| 258 | |
| 259 | bool onIsUsed(GrSurfaceProxy*) const override; |
| 260 | |
| 261 | void handleInternalAllocationFailure() override; |
| 262 | |
| 263 | void gatherProxyIntervals(GrResourceAllocator*) const override; |
| 264 | |
| 265 | void recordOp(std::unique_ptr<GrOp>, GrProcessorSet::Analysis, GrAppliedClip*, |
| 266 | const DstProxyView*, const GrCaps& caps); |
| 267 | |
| 268 | void forwardCombine(const GrCaps&); |
| 269 | |
| 270 | ExpectedOutcome onMakeClosed(const GrCaps& caps, SkIRect* targetUpdateBounds) override; |
| 271 | |
| 272 | friend class GrRenderTargetContextPriv; // for stencil clip state. TODO: this is invasive |
| 273 | |
| 274 | // The RTC and OpsTask have to work together to handle buffer clears. In most cases, buffer |
| 275 | // clearing can be done natively, in which case the op list's load ops are sufficient. In other |
| 276 | // cases, draw ops must be used, which makes the RTC the best place for those decisions. This, |
| 277 | // however, requires that the RTC be able to coordinate with the op list to achieve similar ends |
| 278 | friend class GrRenderTargetContext; |
| 279 | |
| 280 | // This is a backpointer to the Arenas that holds the memory for this GrOpsTask's ops. In the |
| 281 | // DDL case, the Arenas must have been detached from the original recording context and moved |
| 282 | // into the owning DDL. |
| 283 | GrRecordingContext::Arenas fArenas; |
| 284 | GrAuditTrail* fAuditTrail; |
| 285 | |
| 286 | GrLoadOp fColorLoadOp = GrLoadOp::kLoad; |
| 287 | SkPMColor4f fLoadClearColor = SK_PMColor4fTRANSPARENT; |
| 288 | StencilContent fInitialStencilContent = StencilContent::kDontCare; |
| 289 | bool fMustPreserveStencil = false; |
| 290 | |
| 291 | uint32_t fLastClipStackGenID; |
| 292 | SkIRect fLastDevClipBounds; |
| 293 | int fLastClipNumAnalyticFPs; |
| 294 | |
| 295 | // We must track if we have a wait op so that we don't delete the op when we have a full clear. |
| 296 | bool fHasWaitOp = false;; |
| 297 | |
| 298 | // For ops/opsTask we have mean: 5 stdDev: 28 |
| 299 | SkSTArray<25, OpChain, true> fOpChains; |
| 300 | |
| 301 | // MDB TODO: 4096 for the first allocation of the clip space will be huge overkill. |
| 302 | // Gather statistics to determine the correct size. |
| 303 | SkArenaAlloc fClipAllocator{4096}; |
| 304 | SkDEBUGCODE(int fNumClips;) |
| 305 | |
| 306 | // TODO: We could look into this being a set if we find we're adding a lot of duplicates that is |
| 307 | // causing slow downs. |
| 308 | SkTArray<GrSurfaceProxy*, true> fSampledProxies; |
| 309 | |
| 310 | SkRect fTotalBounds = SkRect::MakeEmpty(); |
| 311 | SkIRect fClippedContentBounds = SkIRect::MakeEmpty(); |
| 312 | }; |
| 313 | |
| 314 | #endif |
| 315 | |