• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 OpsTask_DEFINED
9 #define OpsTask_DEFINED
10 
11 #include "include/core/SkMatrix.h"
12 #include "include/core/SkRefCnt.h"
13 #include "include/core/SkSpan.h"
14 #include "include/core/SkStrokeRec.h"
15 #include "include/core/SkTypes.h"
16 #include "include/gpu/GrRecordingContext.h"
17 #include "include/private/base/SkTArray.h"
18 #include "include/private/base/SkTDArray.h"
19 #include "include/private/base/SkTypeTraits.h"
20 #include "src/base/SkArenaAlloc.h"
21 #include "src/base/SkTLazy.h"
22 #include "src/core/SkClipStack.h"
23 #include "src/core/SkStringUtils.h"
24 #include "src/gpu/ganesh/GrAppliedClip.h"
25 #include "src/gpu/ganesh/GrDstProxyView.h"
26 #include "src/gpu/ganesh/GrGeometryProcessor.h"
27 #include "src/gpu/ganesh/GrProcessorSet.h"
28 #include "src/gpu/ganesh/GrRenderTask.h"
29 #include "src/gpu/ganesh/ops/GrOp.h"
30 
31 class GrAuditTrail;
32 class GrCaps;
33 class GrClearOp;
34 class GrGpuBuffer;
35 class GrRenderTargetProxy;
36 class OpsTaskTestingAccess;
37 
38 namespace skgpu::v1 {
39 
40 class SurfaceDrawContext;
41 
42 class OpsTask : public GrRenderTask {
43 public:
44     // Manage the arenas life time by maintaining are reference to it.
45     OpsTask(GrDrawingManager*, GrSurfaceProxyView, GrAuditTrail*, sk_sp<GrArenas>);
46     ~OpsTask() override;
47 
asOpsTask()48     OpsTask* asOpsTask() override { return this; }
49 
isEmpty()50     bool isEmpty() const { return fOpChains.empty(); }
usesMSAASurface()51     bool usesMSAASurface() const { return fUsesMSAASurface; }
renderPassXferBarriers()52     GrXferBarrierFlags renderPassXferBarriers() const { return fRenderPassXferBarriers; }
53 
54     /**
55      * Empties the draw buffer of any queued up draws.
56      */
57     void endFlush(GrDrawingManager*) override;
58 
59     void onPrePrepare(GrRecordingContext*) override;
60     /**
61      * Together these two functions flush all queued up draws to GrCommandBuffer. The return value
62      * of onExecute() indicates whether any commands were actually issued to the GPU.
63      */
64     void onPrepare(GrOpFlushState* flushState) override;
65     bool onExecute(GrOpFlushState* flushState) override;
66 
addSampledTexture(GrSurfaceProxy * proxy)67     void addSampledTexture(GrSurfaceProxy* proxy) {
68         // This function takes a GrSurfaceProxy because all subsequent uses of the proxy do not
69         // require the specifics of GrTextureProxy, so this avoids a number of unnecessary virtual
70         // asTextureProxy() calls. However, sampling the proxy implicitly requires that the proxy
71         // be a texture. Eventually, when proxies are a unified type with flags, this can just
72         // assert that capability.
73         SkASSERT(proxy->asTextureProxy());
74         fSampledProxies.push_back(proxy);
75     }
76 
77     void addOp(GrDrawingManager*, GrOp::Owner, GrTextureResolveManager, const GrCaps&);
78 
79     void addDrawOp(GrDrawingManager*, GrOp::Owner, bool usesMSAA, const GrProcessorSet::Analysis&,
80                    GrAppliedClip&&, const GrDstProxyView&, GrTextureResolveManager, const GrCaps&);
81 
82     void discard();
83 
84     enum class CanDiscardPreviousOps : bool {
85         kYes = true,
86         kNo = false
87     };
88 
89     // Perform book-keeping for a fullscreen clear, regardless of how the clear is implemented later
90     // (i.e. setColorLoadOp(), adding a ClearOp, or adding a FillRectOp that covers the device).
91     // Returns true if the clear can be converted into a load op (barring device caps).
92     bool resetForFullscreenClear(CanDiscardPreviousOps);
93 
94     // Must only be called if native color buffer clearing is enabled.
95     void setColorLoadOp(GrLoadOp op, std::array<float, 4> color = {0, 0, 0, 0});
96 
97     // Returns whether the given opsTask can be appended at the end of this one.
98     bool canMerge(const OpsTask*) const;
99 
100     // Merge as many opsTasks as possible from the head of 'tasks'. They should all be
101     // renderPass compatible. Return the number of tasks merged into 'this'.
102     int mergeFrom(SkSpan<const sk_sp<GrRenderTask>> tasks);
103 
104 #ifdef SK_DEBUG
numClips()105     int numClips() const override { return fNumClips; }
106     void visitProxies_debugOnly(const GrVisitProxyFunc&) const override;
107 #endif
108 
109 #if GR_TEST_UTILS
110     void dump(const SkString& label,
111               SkString indent,
112               bool printDependencies,
113               bool close) const override;
name()114     const char* name() const final { return "Ops"; }
numOpChains()115     int numOpChains() const { return fOpChains.size(); }
getChain(int index)116     const GrOp* getChain(int index) const { return fOpChains[index].head(); }
117 #endif
118 
119 protected:
120     enum class StencilContent {
121         kDontCare,
122         kUserBitsCleared,  // User bits: cleared
123                            // Clip bit: don't care (Ganesh always pre-clears the clip bit.)
124         kPreserved
125     };
126 
127     // Lets the caller specify what the content of the stencil buffer should be at the beginning
128     // of the render pass.
129     //
130     // When requesting kClear: Tilers will load the stencil buffer with a "clear" op; non-tilers
131     // will clear the stencil on first load, and then preserve it on subsequent loads. (Preserving
132     // works because SurfaceDrawContexts are required to leave the user bits in a cleared state
133     // once finished.)
134     //
135     // NOTE: initialContent must not be kClear if caps.performStencilClearsAsDraws() is true.
setInitialStencilContent(StencilContent initialContent)136     void setInitialStencilContent(StencilContent initialContent) {
137         fInitialStencilContent = initialContent;
138     }
139 
140     void recordOp(GrOp::Owner, bool usesMSAA, GrProcessorSet::Analysis, GrAppliedClip*,
141                   const GrDstProxyView*, const GrCaps&);
142 
143     ExpectedOutcome onMakeClosed(GrRecordingContext*, SkIRect* targetUpdateBounds) override;
144 
145 private:
isColorNoOp()146     bool isColorNoOp() const {
147         // TODO: GrLoadOp::kDiscard (i.e., storing a discard) should also be grounds for skipping
148         // execution. We currently don't because of Vulkan. See http://skbug.com/9373.
149         return fOpChains.empty() && GrLoadOp::kLoad == fColorLoadOp;
150     }
151 
152     void deleteOps();
153 
154     // If a surfaceDrawContext splits its opsTask, it uses this method to guarantee stencil values
155     // get preserved across its split tasks.
setMustPreserveStencil()156     void setMustPreserveStencil() { fMustPreserveStencil = true; }
157 
158     // Prevents this opsTask from merging backward. This is used by DMSAA when a non-multisampled
159     // opsTask cannot be promoted to MSAA, or when we split a multisampled opsTask in order to
160     // resolve its texture.
setCannotMergeBackward()161     void setCannotMergeBackward() { fCannotMergeBackward = true; }
162 
163     class OpChain {
164     public:
165         OpChain(GrOp::Owner, GrProcessorSet::Analysis, GrAppliedClip*, const GrDstProxyView*);
~OpChain()166         ~OpChain() {
167             // The ops are stored in a GrMemoryPool and must be explicitly deleted via the pool.
168             SkASSERT(fList.empty());
169         }
170 
171         OpChain(const OpChain&) = delete;
172         OpChain& operator=(const OpChain&) = delete;
173         OpChain(OpChain&&) = default;
174         OpChain& operator=(OpChain&&) = default;
175 
176         void visitProxies(const GrVisitProxyFunc&) const;
177 
head()178         GrOp* head() const { return fList.head(); }
179 
appliedClip()180         GrAppliedClip* appliedClip() const { return fAppliedClip; }
dstProxyView()181         const GrDstProxyView& dstProxyView() const { return fDstProxyView; }
bounds()182         const SkRect& bounds() const { return fBounds; }
183 
184         // Deletes all the ops in the chain.
185         void deleteOps();
186 
187         // Attempts to move the ops from the passed chain to this chain at the head. Also attempts
188         // to merge ops between the chains. Upon success the passed chain is empty.
189         // Fails when the chains aren't of the same op type, have different clips or dst proxies.
190         bool prependChain(OpChain*, const GrCaps&, SkArenaAlloc* opsTaskArena, GrAuditTrail*);
191 
192         // Attempts to add 'op' to this chain either by merging or adding to the tail. Returns
193         // 'op' to the caller upon failure, otherwise null. Fails when the op and chain aren't of
194         // the same op type, have different clips or dst proxies.
195         GrOp::Owner appendOp(GrOp::Owner op, GrProcessorSet::Analysis, const GrDstProxyView*,
196                              const GrAppliedClip*, const GrCaps&, SkArenaAlloc* opsTaskArena,
197                              GrAuditTrail*);
198 
shouldExecute()199         bool shouldExecute() const {
200             return SkToBool(this->head());
201         }
202 
203         using sk_is_trivially_relocatable = std::true_type;
204 
205     private:
206         class List {
207         public:
208             List() = default;
209             List(GrOp::Owner);
210             List(List&&);
211             List& operator=(List&& that);
212 
empty()213             bool empty() const { return !SkToBool(fHead); }
head()214             GrOp* head() const { return fHead.get(); }
tail()215             GrOp* tail() const { return fTail; }
216 
217             GrOp::Owner popHead();
218             GrOp::Owner removeOp(GrOp* op);
219             void pushHead(GrOp::Owner op);
220             void pushTail(GrOp::Owner);
221 
222             void validate() const;
223 
224             using sk_is_trivially_relocatable = std::true_type;
225 
226         private:
227             GrOp::Owner fHead{nullptr};
228             GrOp* fTail{nullptr};
229 
230             static_assert(::sk_is_trivially_relocatable<decltype(fHead)>::value);
231             static_assert(::sk_is_trivially_relocatable<decltype(fTail)>::value);
232         };
233 
234         void validate() const;
235 
236         bool tryConcat(List*, GrProcessorSet::Analysis, const GrDstProxyView&, const GrAppliedClip*,
237                        const SkRect& bounds, const GrCaps&, SkArenaAlloc* opsTaskArena,
238                        GrAuditTrail*);
239         static List DoConcat(List, List, const GrCaps&, SkArenaAlloc* opsTaskArena, GrAuditTrail*);
240 
241         List fList;
242         GrProcessorSet::Analysis fProcessorAnalysis;
243         GrDstProxyView fDstProxyView;
244         GrAppliedClip* fAppliedClip;
245         SkRect fBounds;
246 
247         static_assert(::sk_is_trivially_relocatable<decltype(fProcessorAnalysis)>::value);
248         static_assert(::sk_is_trivially_relocatable<decltype(fDstProxyView)>::value);
249         static_assert(::sk_is_trivially_relocatable<decltype(fAppliedClip)>::value);
250         static_assert(::sk_is_trivially_relocatable<decltype(fBounds)>::value);
251     };
252 
253     void onMakeSkippable() override;
254 
255     bool onIsUsed(GrSurfaceProxy*) const override;
256 
257     void gatherProxyIntervals(GrResourceAllocator*) const override;
258 
259     void forwardCombine(const GrCaps&);
260 
261     // Remove all ops, proxies, etc. Used in the merging algorithm when tasks can be skipped.
262     void reset();
263 
264     friend class ::OpsTaskTestingAccess;
265 
266     // The SDC and OpsTask have to work together to handle buffer clears. In most cases, buffer
267     // clearing can be done natively, in which case the op list's load ops are sufficient. In other
268     // cases, draw ops must be used, which makes the SDC the best place for those decisions. This,
269     // however, requires that the SDC be able to coordinate with the op list to achieve similar ends
270     friend class skgpu::v1::SurfaceDrawContext;
271 
272     GrAuditTrail* fAuditTrail;
273 
274     bool fUsesMSAASurface;
275     skgpu::Swizzle fTargetSwizzle;
276     GrSurfaceOrigin fTargetOrigin;
277 
278     GrLoadOp fColorLoadOp = GrLoadOp::kLoad;
279     std::array<float, 4> fLoadClearColor = {0, 0, 0, 0};
280     StencilContent fInitialStencilContent = StencilContent::kDontCare;
281     bool fMustPreserveStencil = false;
282     bool fCannotMergeBackward = false;
283 
284     uint32_t fLastClipStackGenID = SK_InvalidUniqueID;
285     SkIRect fLastDevClipBounds;
286     int fLastClipNumAnalyticElements;
287 
288     GrXferBarrierFlags fRenderPassXferBarriers = GrXferBarrierFlags::kNone;
289 
290     // For ops/opsTask we have mean: 5 stdDev: 28
291     SkSTArray<25, OpChain> fOpChains;
292 
293     sk_sp<GrArenas> fArenas;
294     SkDEBUGCODE(int fNumClips;)
295 
296     // TODO: We could look into this being a set if we find we're adding a lot of duplicates that is
297     // causing slow downs.
298     SkTArray<GrSurfaceProxy*, true> fSampledProxies;
299 
300     SkRect fTotalBounds = SkRect::MakeEmpty();
301     SkIRect fClippedContentBounds = SkIRect::MakeEmpty();
302 };
303 
304 } // namespace skgpu::v1
305 
306 #endif // OpsTask_DEFINED
307