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 | #include "src/gpu/GrOpsTask.h" |
9 | |
10 | #include "include/private/GrRecordingContext.h" |
11 | #include "src/core/SkExchange.h" |
12 | #include "src/core/SkRectPriv.h" |
13 | #include "src/core/SkTraceEvent.h" |
14 | #include "src/gpu/GrAuditTrail.h" |
15 | #include "src/gpu/GrCaps.h" |
16 | #include "src/gpu/GrGpu.h" |
17 | #include "src/gpu/GrMemoryPool.h" |
18 | #include "src/gpu/GrOpFlushState.h" |
19 | #include "src/gpu/GrOpsRenderPass.h" |
20 | #include "src/gpu/GrRecordingContextPriv.h" |
21 | #include "src/gpu/GrRenderTarget.h" |
22 | #include "src/gpu/GrRenderTargetContext.h" |
23 | #include "src/gpu/GrRenderTargetPriv.h" |
24 | #include "src/gpu/GrResourceAllocator.h" |
25 | #include "src/gpu/GrStencilAttachment.h" |
26 | #include "src/gpu/GrTexturePriv.h" |
27 | #include "src/gpu/geometry/GrRect.h" |
28 | #include "src/gpu/ops/GrClearOp.h" |
29 | |
30 | //////////////////////////////////////////////////////////////////////////////// |
31 | |
32 | // Experimentally we have found that most combining occurs within the first 10 comparisons. |
33 | static const int kMaxOpMergeDistance = 10; |
34 | static const int kMaxOpChainDistance = 10; |
35 | |
36 | //////////////////////////////////////////////////////////////////////////////// |
37 | |
38 | using DstProxyView = GrXferProcessor::DstProxyView; |
39 | |
40 | //////////////////////////////////////////////////////////////////////////////// |
41 | |
42 | static inline bool can_reorder(const SkRect& a, const SkRect& b) { return !GrRectsOverlap(a, b); } |
43 | |
44 | //////////////////////////////////////////////////////////////////////////////// |
45 | |
46 | inline GrOpsTask::OpChain::List::List(std::unique_ptr<GrOp> op) |
47 | : fHead(std::move(op)), fTail(fHead.get()) { |
48 | this->validate(); |
49 | } |
50 | |
51 | inline GrOpsTask::OpChain::List::List(List&& that) { *this = std::move(that); } |
52 | |
53 | inline GrOpsTask::OpChain::List& GrOpsTask::OpChain::List::operator=(List&& that) { |
54 | fHead = std::move(that.fHead); |
55 | fTail = that.fTail; |
56 | that.fTail = nullptr; |
57 | this->validate(); |
58 | return *this; |
59 | } |
60 | |
61 | inline std::unique_ptr<GrOp> GrOpsTask::OpChain::List::popHead() { |
62 | SkASSERT(fHead); |
63 | auto temp = fHead->cutChain(); |
64 | std::swap(temp, fHead); |
65 | if (!fHead) { |
66 | SkASSERT(fTail == temp.get()); |
67 | fTail = nullptr; |
68 | } |
69 | return temp; |
70 | } |
71 | |
72 | inline std::unique_ptr<GrOp> GrOpsTask::OpChain::List::removeOp(GrOp* op) { |
73 | #ifdef SK_DEBUG |
74 | auto head = op; |
75 | while (head->prevInChain()) { head = head->prevInChain(); } |
76 | SkASSERT(head == fHead.get()); |
77 | #endif |
78 | auto prev = op->prevInChain(); |
79 | if (!prev) { |
80 | SkASSERT(op == fHead.get()); |
81 | return this->popHead(); |
82 | } |
83 | auto temp = prev->cutChain(); |
84 | if (auto next = temp->cutChain()) { |
85 | prev->chainConcat(std::move(next)); |
86 | } else { |
87 | SkASSERT(fTail == op); |
88 | fTail = prev; |
89 | } |
90 | this->validate(); |
91 | return temp; |
92 | } |
93 | |
94 | inline void GrOpsTask::OpChain::List::pushHead(std::unique_ptr<GrOp> op) { |
95 | SkASSERT(op); |
96 | SkASSERT(op->isChainHead()); |
97 | SkASSERT(op->isChainTail()); |
98 | if (fHead) { |
99 | op->chainConcat(std::move(fHead)); |
100 | fHead = std::move(op); |
101 | } else { |
102 | fHead = std::move(op); |
103 | fTail = fHead.get(); |
104 | } |
105 | } |
106 | |
107 | inline void GrOpsTask::OpChain::List::pushTail(std::unique_ptr<GrOp> op) { |
108 | SkASSERT(op->isChainTail()); |
109 | fTail->chainConcat(std::move(op)); |
110 | fTail = fTail->nextInChain(); |
111 | } |
112 | |
113 | inline void GrOpsTask::OpChain::List::validate() const { |
114 | #ifdef SK_DEBUG |
115 | if (fHead) { |
116 | SkASSERT(fTail); |
117 | fHead->validateChain(fTail); |
118 | } |
119 | #endif |
120 | } |
121 | |
122 | //////////////////////////////////////////////////////////////////////////////// |
123 | |
124 | GrOpsTask::OpChain::OpChain(std::unique_ptr<GrOp> op, |
125 | GrProcessorSet::Analysis processorAnalysis, |
126 | GrAppliedClip* appliedClip, const DstProxyView* dstProxyView) |
127 | : fList{std::move(op)} |
128 | , fProcessorAnalysis(processorAnalysis) |
129 | , fAppliedClip(appliedClip) { |
130 | if (fProcessorAnalysis.requiresDstTexture()) { |
131 | SkASSERT(dstProxyView && dstProxyView->proxy()); |
132 | fDstProxyView = *dstProxyView; |
133 | } |
134 | fBounds = fList.head()->bounds(); |
135 | } |
136 | |
137 | void GrOpsTask::OpChain::visitProxies(const GrOp::VisitProxyFunc& func) const { |
138 | if (fList.empty()) { |
139 | return; |
140 | } |
141 | for (const auto& op : GrOp::ChainRange<>(fList.head())) { |
142 | op.visitProxies(func); |
143 | } |
144 | if (fDstProxyView.proxy()) { |
145 | func(fDstProxyView.proxy(), GrMipMapped::kNo); |
146 | } |
147 | if (fAppliedClip) { |
148 | fAppliedClip->visitProxies(func); |
149 | } |
150 | } |
151 | |
152 | void GrOpsTask::OpChain::deleteOps(GrOpMemoryPool* pool) { |
153 | while (!fList.empty()) { |
154 | pool->release(fList.popHead()); |
155 | } |
156 | } |
157 | |
158 | // Concatenates two op chains and attempts to merge ops across the chains. Assumes that we know that |
159 | // the two chains are chainable. Returns the new chain. |
160 | GrOpsTask::OpChain::List GrOpsTask::OpChain::DoConcat( |
161 | List chainA, List chainB, const GrCaps& caps, GrRecordingContext::Arenas* arenas, |
162 | GrAuditTrail* auditTrail) { |
163 | // We process ops in chain b from head to tail. We attempt to merge with nodes in a, starting |
164 | // at chain a's tail and working toward the head. We produce one of the following outcomes: |
165 | // 1) b's head is merged into an op in a. |
166 | // 2) An op from chain a is merged into b's head. (In this case b's head gets processed again.) |
167 | // 3) b's head is popped from chain a and added at the tail of a. |
168 | // After result 3 we don't want to attempt to merge the next head of b with the new tail of a, |
169 | // as we assume merges were already attempted when chain b was created. So we keep track of the |
170 | // original tail of a and start our iteration of a there. We also track the bounds of the nodes |
171 | // appended to chain a that will be skipped for bounds testing. If the original tail of a is |
172 | // merged into an op in b (case 2) then we advance the "original tail" towards the head of a. |
173 | GrOp* origATail = chainA.tail(); |
174 | SkRect skipBounds = SkRectPriv::MakeLargestInverted(); |
175 | do { |
176 | int numMergeChecks = 0; |
177 | bool merged = false; |
178 | bool noSkip = (origATail == chainA.tail()); |
179 | SkASSERT(noSkip == (skipBounds == SkRectPriv::MakeLargestInverted())); |
180 | bool canBackwardMerge = noSkip || can_reorder(chainB.head()->bounds(), skipBounds); |
181 | SkRect forwardMergeBounds = skipBounds; |
182 | GrOp* a = origATail; |
183 | while (a) { |
184 | bool canForwardMerge = |
185 | (a == chainA.tail()) || can_reorder(a->bounds(), forwardMergeBounds); |
186 | if (canForwardMerge || canBackwardMerge) { |
187 | auto result = a->combineIfPossible(chainB.head(), arenas, caps); |
188 | SkASSERT(result != GrOp::CombineResult::kCannotCombine); |
189 | merged = (result == GrOp::CombineResult::kMerged); |
190 | GrOP_INFO("\t\t: (%s opID: %u) -> Combining with (%s, opID: %u)\n" , |
191 | chainB.head()->name(), chainB.head()->uniqueID(), a->name(), |
192 | a->uniqueID()); |
193 | } |
194 | if (merged) { |
195 | GR_AUDIT_TRAIL_OPS_RESULT_COMBINED(auditTrail, a, chainB.head()); |
196 | if (canBackwardMerge) { |
197 | arenas->opMemoryPool()->release(chainB.popHead()); |
198 | } else { |
199 | // We merged the contents of b's head into a. We will replace b's head with a in |
200 | // chain b. |
201 | SkASSERT(canForwardMerge); |
202 | if (a == origATail) { |
203 | origATail = a->prevInChain(); |
204 | } |
205 | std::unique_ptr<GrOp> detachedA = chainA.removeOp(a); |
206 | arenas->opMemoryPool()->release(chainB.popHead()); |
207 | chainB.pushHead(std::move(detachedA)); |
208 | if (chainA.empty()) { |
209 | // We merged all the nodes in chain a to chain b. |
210 | return chainB; |
211 | } |
212 | } |
213 | break; |
214 | } else { |
215 | if (++numMergeChecks == kMaxOpMergeDistance) { |
216 | break; |
217 | } |
218 | forwardMergeBounds.joinNonEmptyArg(a->bounds()); |
219 | canBackwardMerge = |
220 | canBackwardMerge && can_reorder(chainB.head()->bounds(), a->bounds()); |
221 | a = a->prevInChain(); |
222 | } |
223 | } |
224 | // If we weren't able to merge b's head then pop b's head from chain b and make it the new |
225 | // tail of a. |
226 | if (!merged) { |
227 | chainA.pushTail(chainB.popHead()); |
228 | skipBounds.joinNonEmptyArg(chainA.tail()->bounds()); |
229 | } |
230 | } while (!chainB.empty()); |
231 | return chainA; |
232 | } |
233 | |
234 | // Attempts to concatenate the given chain onto our own and merge ops across the chains. Returns |
235 | // whether the operation succeeded. On success, the provided list will be returned empty. |
236 | bool GrOpsTask::OpChain::tryConcat( |
237 | List* list, GrProcessorSet::Analysis processorAnalysis, const DstProxyView& dstProxyView, |
238 | const GrAppliedClip* appliedClip, const SkRect& bounds, const GrCaps& caps, |
239 | GrRecordingContext::Arenas* arenas, GrAuditTrail* auditTrail) { |
240 | SkASSERT(!fList.empty()); |
241 | SkASSERT(!list->empty()); |
242 | SkASSERT(fProcessorAnalysis.requiresDstTexture() == SkToBool(fDstProxyView.proxy())); |
243 | SkASSERT(processorAnalysis.requiresDstTexture() == SkToBool(dstProxyView.proxy())); |
244 | // All returns use explicit tuple constructor rather than {a, b} to work around old GCC bug. |
245 | if (fList.head()->classID() != list->head()->classID() || |
246 | SkToBool(fAppliedClip) != SkToBool(appliedClip) || |
247 | (fAppliedClip && *fAppliedClip != *appliedClip) || |
248 | (fProcessorAnalysis.requiresNonOverlappingDraws() != |
249 | processorAnalysis.requiresNonOverlappingDraws()) || |
250 | (fProcessorAnalysis.requiresNonOverlappingDraws() && |
251 | // Non-overlaping draws are only required when Ganesh will either insert a barrier, |
252 | // or read back a new dst texture between draws. In either case, we can neither |
253 | // chain nor combine overlapping Ops. |
254 | GrRectsTouchOrOverlap(fBounds, bounds)) || |
255 | (fProcessorAnalysis.requiresDstTexture() != processorAnalysis.requiresDstTexture()) || |
256 | (fProcessorAnalysis.requiresDstTexture() && fDstProxyView != dstProxyView)) { |
257 | return false; |
258 | } |
259 | |
260 | SkDEBUGCODE(bool first = true;) |
261 | do { |
262 | switch (fList.tail()->combineIfPossible(list->head(), arenas, caps)) { |
263 | case GrOp::CombineResult::kCannotCombine: |
264 | // If an op supports chaining then it is required that chaining is transitive and |
265 | // that if any two ops in two different chains can merge then the two chains |
266 | // may also be chained together. Thus, we should only hit this on the first |
267 | // iteration. |
268 | SkASSERT(first); |
269 | return false; |
270 | case GrOp::CombineResult::kMayChain: |
271 | fList = DoConcat(std::move(fList), skstd::exchange(*list, List()), caps, arenas, |
272 | auditTrail); |
273 | // The above exchange cleared out 'list'. The list needs to be empty now for the |
274 | // loop to terminate. |
275 | SkASSERT(list->empty()); |
276 | break; |
277 | case GrOp::CombineResult::kMerged: { |
278 | GrOP_INFO("\t\t: (%s opID: %u) -> Combining with (%s, opID: %u)\n" , |
279 | list->tail()->name(), list->tail()->uniqueID(), list->head()->name(), |
280 | list->head()->uniqueID()); |
281 | GR_AUDIT_TRAIL_OPS_RESULT_COMBINED(auditTrail, fList.tail(), list->head()); |
282 | arenas->opMemoryPool()->release(list->popHead()); |
283 | break; |
284 | } |
285 | } |
286 | SkDEBUGCODE(first = false); |
287 | } while (!list->empty()); |
288 | |
289 | // The new ops were successfully merged and/or chained onto our own. |
290 | fBounds.joinPossiblyEmptyRect(bounds); |
291 | return true; |
292 | } |
293 | |
294 | bool GrOpsTask::OpChain::prependChain(OpChain* that, const GrCaps& caps, |
295 | GrRecordingContext::Arenas* arenas, |
296 | GrAuditTrail* auditTrail) { |
297 | if (!that->tryConcat(&fList, fProcessorAnalysis, fDstProxyView, fAppliedClip, fBounds, caps, |
298 | arenas, auditTrail)) { |
299 | this->validate(); |
300 | // append failed |
301 | return false; |
302 | } |
303 | |
304 | // 'that' owns the combined chain. Move it into 'this'. |
305 | SkASSERT(fList.empty()); |
306 | fList = std::move(that->fList); |
307 | fBounds = that->fBounds; |
308 | |
309 | that->fDstProxyView.setProxyView({}); |
310 | if (that->fAppliedClip) { |
311 | for (int i = 0; i < that->fAppliedClip->numClipCoverageFragmentProcessors(); ++i) { |
312 | that->fAppliedClip->detachClipCoverageFragmentProcessor(i); |
313 | } |
314 | } |
315 | this->validate(); |
316 | return true; |
317 | } |
318 | |
319 | std::unique_ptr<GrOp> GrOpsTask::OpChain::appendOp( |
320 | std::unique_ptr<GrOp> op, GrProcessorSet::Analysis processorAnalysis, |
321 | const DstProxyView* dstProxyView, const GrAppliedClip* appliedClip, const GrCaps& caps, |
322 | GrRecordingContext::Arenas* arenas, GrAuditTrail* auditTrail) { |
323 | const GrXferProcessor::DstProxyView noDstProxyView; |
324 | if (!dstProxyView) { |
325 | dstProxyView = &noDstProxyView; |
326 | } |
327 | SkASSERT(op->isChainHead() && op->isChainTail()); |
328 | SkRect opBounds = op->bounds(); |
329 | List chain(std::move(op)); |
330 | if (!this->tryConcat( |
331 | &chain, processorAnalysis, *dstProxyView, appliedClip, opBounds, caps, |
332 | arenas, auditTrail)) { |
333 | // append failed, give the op back to the caller. |
334 | this->validate(); |
335 | return chain.popHead(); |
336 | } |
337 | |
338 | SkASSERT(chain.empty()); |
339 | this->validate(); |
340 | return nullptr; |
341 | } |
342 | |
343 | inline void GrOpsTask::OpChain::validate() const { |
344 | #ifdef SK_DEBUG |
345 | fList.validate(); |
346 | for (const auto& op : GrOp::ChainRange<>(fList.head())) { |
347 | // Not using SkRect::contains because we allow empty rects. |
348 | SkASSERT(fBounds.fLeft <= op.bounds().fLeft && fBounds.fTop <= op.bounds().fTop && |
349 | fBounds.fRight >= op.bounds().fRight && fBounds.fBottom >= op.bounds().fBottom); |
350 | } |
351 | #endif |
352 | } |
353 | |
354 | //////////////////////////////////////////////////////////////////////////////// |
355 | |
356 | GrOpsTask::GrOpsTask(GrRecordingContext::Arenas arenas, |
357 | GrSurfaceProxyView view, |
358 | GrAuditTrail* auditTrail) |
359 | : GrRenderTask(std::move(view)) |
360 | , fArenas(arenas) |
361 | , fAuditTrail(auditTrail) |
362 | , fLastClipStackGenID(SK_InvalidUniqueID) |
363 | SkDEBUGCODE(, fNumClips(0)) { |
364 | fTargetView.proxy()->setLastRenderTask(this); |
365 | } |
366 | |
367 | void GrOpsTask::deleteOps() { |
368 | for (auto& chain : fOpChains) { |
369 | chain.deleteOps(fArenas.opMemoryPool()); |
370 | } |
371 | fOpChains.reset(); |
372 | } |
373 | |
374 | GrOpsTask::~GrOpsTask() { |
375 | this->deleteOps(); |
376 | } |
377 | |
378 | //////////////////////////////////////////////////////////////////////////////// |
379 | |
380 | void GrOpsTask::endFlush() { |
381 | fLastClipStackGenID = SK_InvalidUniqueID; |
382 | this->deleteOps(); |
383 | fClipAllocator.reset(); |
384 | |
385 | GrSurfaceProxy* proxy = fTargetView.proxy(); |
386 | if (proxy && this == proxy->getLastRenderTask()) { |
387 | proxy->setLastRenderTask(nullptr); |
388 | } |
389 | |
390 | fTargetView.reset(); |
391 | fDeferredProxies.reset(); |
392 | fSampledProxies.reset(); |
393 | fAuditTrail = nullptr; |
394 | } |
395 | |
396 | void GrOpsTask::onPrePrepare(GrRecordingContext* context) { |
397 | SkASSERT(this->isClosed()); |
398 | #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK |
399 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
400 | #endif |
401 | // TODO: remove the check for discard here once reduced op splitting is turned on. Currently we |
402 | // can end up with GrOpsTasks that only have a discard load op and no ops. For vulkan validation |
403 | // we need to keep that discard and not drop it. Once we have reduce op list splitting enabled |
404 | // we shouldn't end up with GrOpsTasks with only discard. |
405 | if (this->isNoOp() || (fClippedContentBounds.isEmpty() && fColorLoadOp != GrLoadOp::kDiscard)) { |
406 | return; |
407 | } |
408 | |
409 | for (const auto& chain : fOpChains) { |
410 | if (chain.shouldExecute()) { |
411 | chain.head()->prePrepare(context, |
412 | &fTargetView, |
413 | chain.appliedClip(), |
414 | chain.dstProxyView()); |
415 | } |
416 | } |
417 | } |
418 | |
419 | void GrOpsTask::onPrepare(GrOpFlushState* flushState) { |
420 | SkASSERT(fTargetView.proxy()->peekRenderTarget()); |
421 | SkASSERT(this->isClosed()); |
422 | #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK |
423 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
424 | #endif |
425 | // TODO: remove the check for discard here once reduced op splitting is turned on. Currently we |
426 | // can end up with GrOpsTasks that only have a discard load op and no ops. For vulkan validation |
427 | // we need to keep that discard and not drop it. Once we have reduce op list splitting enabled |
428 | // we shouldn't end up with GrOpsTasks with only discard. |
429 | if (this->isNoOp() || (fClippedContentBounds.isEmpty() && fColorLoadOp != GrLoadOp::kDiscard)) { |
430 | return; |
431 | } |
432 | |
433 | flushState->setSampledProxyArray(&fSampledProxies); |
434 | // Loop over the ops that haven't yet been prepared. |
435 | for (const auto& chain : fOpChains) { |
436 | if (chain.shouldExecute()) { |
437 | #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK |
438 | TRACE_EVENT0("skia.gpu" , chain.head()->name()); |
439 | #endif |
440 | GrOpFlushState::OpArgs opArgs(chain.head(), |
441 | &fTargetView, |
442 | chain.appliedClip(), |
443 | chain.dstProxyView()); |
444 | |
445 | flushState->setOpArgs(&opArgs); |
446 | |
447 | // Temporary debugging helper: for debugging prePrepare w/o going through DDLs |
448 | // Delete once most of the GrOps have an onPrePrepare. |
449 | // chain.head()->prePrepare(flushState->gpu()->getContext(), &fTargetView, |
450 | // chain.appliedClip()); |
451 | |
452 | // GrOp::prePrepare may or may not have been called at this point |
453 | chain.head()->prepare(flushState); |
454 | flushState->setOpArgs(nullptr); |
455 | } |
456 | } |
457 | flushState->setSampledProxyArray(nullptr); |
458 | } |
459 | |
460 | static GrOpsRenderPass* create_render_pass( |
461 | GrGpu* gpu, GrRenderTarget* rt, GrSurfaceOrigin origin, const SkIRect& bounds, |
462 | GrLoadOp colorLoadOp, const SkPMColor4f& loadClearColor, GrLoadOp stencilLoadOp, |
463 | GrStoreOp stencilStoreOp, const SkTArray<GrSurfaceProxy*, true>& sampledProxies) { |
464 | const GrOpsRenderPass::LoadAndStoreInfo kColorLoadStoreInfo { |
465 | colorLoadOp, |
466 | GrStoreOp::kStore, |
467 | loadClearColor |
468 | }; |
469 | |
470 | // TODO: |
471 | // We would like to (at this level) only ever clear & discard. We would need |
472 | // to stop splitting up higher level OpsTasks for copyOps to achieve that. |
473 | // Note: we would still need SB loads and stores but they would happen at a |
474 | // lower level (inside the VK command buffer). |
475 | const GrOpsRenderPass::StencilLoadAndStoreInfo stencilLoadAndStoreInfo { |
476 | stencilLoadOp, |
477 | stencilStoreOp, |
478 | }; |
479 | |
480 | return gpu->getOpsRenderPass(rt, origin, bounds, kColorLoadStoreInfo, stencilLoadAndStoreInfo, |
481 | sampledProxies); |
482 | } |
483 | |
484 | // TODO: this is where GrOp::renderTarget is used (which is fine since it |
485 | // is at flush time). However, we need to store the RenderTargetProxy in the |
486 | // Ops and instantiate them here. |
487 | bool GrOpsTask::onExecute(GrOpFlushState* flushState) { |
488 | // TODO: remove the check for discard here once reduced op splitting is turned on. Currently we |
489 | // can end up with GrOpsTasks that only have a discard load op and no ops. For vulkan validation |
490 | // we need to keep that discard and not drop it. Once we have reduce op list splitting enabled |
491 | // we shouldn't end up with GrOpsTasks with only discard. |
492 | if (this->isNoOp() || (fClippedContentBounds.isEmpty() && fColorLoadOp != GrLoadOp::kDiscard)) { |
493 | return false; |
494 | } |
495 | |
496 | SkASSERT(fTargetView.proxy()); |
497 | GrRenderTargetProxy* proxy = fTargetView.proxy()->asRenderTargetProxy(); |
498 | SkASSERT(proxy); |
499 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
500 | |
501 | // Make sure load ops are not kClear if the GPU needs to use draws for clears |
502 | SkASSERT(fColorLoadOp != GrLoadOp::kClear || |
503 | !flushState->gpu()->caps()->performColorClearsAsDraws()); |
504 | |
505 | const GrCaps& caps = *flushState->gpu()->caps(); |
506 | GrRenderTarget* renderTarget = proxy->peekRenderTarget(); |
507 | SkASSERT(renderTarget); |
508 | |
509 | GrStencilAttachment* stencil = nullptr; |
510 | if (int numStencilSamples = proxy->numStencilSamples()) { |
511 | if (!flushState->resourceProvider()->attachStencilAttachment( |
512 | renderTarget, numStencilSamples)) { |
513 | SkDebugf("WARNING: failed to attach a stencil buffer. Rendering will be skipped.\n" ); |
514 | return false; |
515 | } |
516 | stencil = renderTarget->renderTargetPriv().getStencilAttachment(); |
517 | } |
518 | |
519 | SkASSERT(!stencil || stencil->numSamples() == proxy->numStencilSamples()); |
520 | |
521 | GrLoadOp stencilLoadOp; |
522 | switch (fInitialStencilContent) { |
523 | case StencilContent::kDontCare: |
524 | stencilLoadOp = GrLoadOp::kDiscard; |
525 | break; |
526 | case StencilContent::kUserBitsCleared: |
527 | SkASSERT(!caps.performStencilClearsAsDraws()); |
528 | SkASSERT(stencil); |
529 | if (caps.discardStencilValuesAfterRenderPass()) { |
530 | // Always clear the stencil if it is being discarded after render passes. This is |
531 | // also an optimization because we are on a tiler and it avoids loading the values |
532 | // from memory. |
533 | stencilLoadOp = GrLoadOp::kClear; |
534 | break; |
535 | } |
536 | if (!stencil->hasPerformedInitialClear()) { |
537 | stencilLoadOp = GrLoadOp::kClear; |
538 | stencil->markHasPerformedInitialClear(); |
539 | break; |
540 | } |
541 | // renderTargetContexts are required to leave the user stencil bits in a cleared state |
542 | // once finished, meaning the stencil values will always remain cleared after the |
543 | // initial clear. Just fall through to reloading the existing (cleared) stencil values |
544 | // from memory. |
545 | case StencilContent::kPreserved: |
546 | SkASSERT(stencil); |
547 | stencilLoadOp = GrLoadOp::kLoad; |
548 | break; |
549 | } |
550 | |
551 | // NOTE: If fMustPreserveStencil is set, then we are executing a renderTargetContext that split |
552 | // its opsTask. |
553 | // |
554 | // FIXME: We don't currently flag render passes that don't use stencil at all. In that case |
555 | // their store op might be "discard", and we currently make the assumption that a discard will |
556 | // not invalidate what's already in main memory. This is probably ok for now, but certainly |
557 | // something we want to address soon. |
558 | GrStoreOp stencilStoreOp = (caps.discardStencilValuesAfterRenderPass() && !fMustPreserveStencil) |
559 | ? GrStoreOp::kDiscard |
560 | : GrStoreOp::kStore; |
561 | |
562 | GrOpsRenderPass* renderPass = create_render_pass( |
563 | flushState->gpu(), proxy->peekRenderTarget(), fTargetView.origin(), |
564 | fClippedContentBounds, fColorLoadOp, fLoadClearColor, stencilLoadOp, stencilStoreOp, |
565 | fSampledProxies); |
566 | if (!renderPass) { |
567 | return false; |
568 | } |
569 | flushState->setOpsRenderPass(renderPass); |
570 | renderPass->begin(); |
571 | |
572 | // Draw all the generated geometry. |
573 | for (const auto& chain : fOpChains) { |
574 | if (!chain.shouldExecute()) { |
575 | continue; |
576 | } |
577 | #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK |
578 | TRACE_EVENT0("skia.gpu" , chain.head()->name()); |
579 | #endif |
580 | |
581 | GrOpFlushState::OpArgs opArgs(chain.head(), |
582 | &fTargetView, |
583 | chain.appliedClip(), |
584 | chain.dstProxyView()); |
585 | |
586 | flushState->setOpArgs(&opArgs); |
587 | chain.head()->execute(flushState, chain.bounds()); |
588 | flushState->setOpArgs(nullptr); |
589 | } |
590 | |
591 | renderPass->end(); |
592 | flushState->gpu()->submit(renderPass); |
593 | flushState->setOpsRenderPass(nullptr); |
594 | |
595 | return true; |
596 | } |
597 | |
598 | void GrOpsTask::setColorLoadOp(GrLoadOp op, const SkPMColor4f& color) { |
599 | fColorLoadOp = op; |
600 | fLoadClearColor = color; |
601 | if (GrLoadOp::kClear == fColorLoadOp) { |
602 | GrSurfaceProxy* proxy = fTargetView.proxy(); |
603 | SkASSERT(proxy); |
604 | fTotalBounds = proxy->getBoundsRect(); |
605 | } |
606 | } |
607 | |
608 | bool GrOpsTask::resetForFullscreenClear(CanDiscardPreviousOps canDiscardPreviousOps) { |
609 | // If we previously recorded a wait op, we cannot delete the wait op. Until we track the wait |
610 | // ops separately from normal ops, we have to avoid clearing out any ops in this case as well. |
611 | if (fHasWaitOp) { |
612 | canDiscardPreviousOps = CanDiscardPreviousOps::kNo; |
613 | } |
614 | |
615 | if (CanDiscardPreviousOps::kYes == canDiscardPreviousOps || this->isEmpty()) { |
616 | this->deleteOps(); |
617 | fDeferredProxies.reset(); |
618 | fSampledProxies.reset(); |
619 | |
620 | // If the opsTask is using a render target which wraps a vulkan command buffer, we can't do |
621 | // a clear load since we cannot change the render pass that we are using. Thus we fall back |
622 | // to making a clear op in this case. |
623 | return !fTargetView.asRenderTargetProxy()->wrapsVkSecondaryCB(); |
624 | } |
625 | |
626 | // Could not empty the task, so an op must be added to handle the clear |
627 | return false; |
628 | } |
629 | |
630 | void GrOpsTask::discard() { |
631 | // Discard calls to in-progress opsTasks are ignored. Calls at the start update the |
632 | // opsTasks' color & stencil load ops. |
633 | if (this->isEmpty()) { |
634 | fColorLoadOp = GrLoadOp::kDiscard; |
635 | fInitialStencilContent = StencilContent::kDontCare; |
636 | fTotalBounds.setEmpty(); |
637 | } |
638 | } |
639 | |
640 | //////////////////////////////////////////////////////////////////////////////// |
641 | |
642 | #ifdef SK_DEBUG |
643 | void GrOpsTask::dump(bool printDependencies) const { |
644 | GrRenderTask::dump(printDependencies); |
645 | |
646 | SkDebugf("fColorLoadOp: " ); |
647 | switch (fColorLoadOp) { |
648 | case GrLoadOp::kLoad: |
649 | SkDebugf("kLoad\n" ); |
650 | break; |
651 | case GrLoadOp::kClear: |
652 | SkDebugf("kClear (0x%x)\n" , fLoadClearColor.toBytes_RGBA()); |
653 | break; |
654 | case GrLoadOp::kDiscard: |
655 | SkDebugf("kDiscard\n" ); |
656 | break; |
657 | } |
658 | |
659 | SkDebugf("fInitialStencilContent: " ); |
660 | switch (fInitialStencilContent) { |
661 | case StencilContent::kDontCare: |
662 | SkDebugf("kDontCare\n" ); |
663 | break; |
664 | case StencilContent::kUserBitsCleared: |
665 | SkDebugf("kUserBitsCleared\n" ); |
666 | break; |
667 | case StencilContent::kPreserved: |
668 | SkDebugf("kPreserved\n" ); |
669 | break; |
670 | } |
671 | |
672 | SkDebugf("ops (%d):\n" , fOpChains.count()); |
673 | for (int i = 0; i < fOpChains.count(); ++i) { |
674 | SkDebugf("*******************************\n" ); |
675 | if (!fOpChains[i].head()) { |
676 | SkDebugf("%d: <combined forward or failed instantiation>\n" , i); |
677 | } else { |
678 | SkDebugf("%d: %s\n" , i, fOpChains[i].head()->name()); |
679 | SkRect bounds = fOpChains[i].bounds(); |
680 | SkDebugf("ClippedBounds: [L: %.2f, T: %.2f, R: %.2f, B: %.2f]\n" , bounds.fLeft, |
681 | bounds.fTop, bounds.fRight, bounds.fBottom); |
682 | for (const auto& op : GrOp::ChainRange<>(fOpChains[i].head())) { |
683 | SkString info = SkTabString(op.dumpInfo(), 1); |
684 | SkDebugf("%s\n" , info.c_str()); |
685 | bounds = op.bounds(); |
686 | SkDebugf("\tClippedBounds: [L: %.2f, T: %.2f, R: %.2f, B: %.2f]\n" , bounds.fLeft, |
687 | bounds.fTop, bounds.fRight, bounds.fBottom); |
688 | } |
689 | } |
690 | } |
691 | } |
692 | |
693 | void GrOpsTask::visitProxies_debugOnly(const GrOp::VisitProxyFunc& func) const { |
694 | auto textureFunc = [ func ] (GrSurfaceProxy* tex, GrMipMapped mipmapped) { |
695 | func(tex, mipmapped); |
696 | }; |
697 | |
698 | for (const OpChain& chain : fOpChains) { |
699 | chain.visitProxies(textureFunc); |
700 | } |
701 | } |
702 | |
703 | #endif |
704 | |
705 | //////////////////////////////////////////////////////////////////////////////// |
706 | |
707 | bool GrOpsTask::onIsUsed(GrSurfaceProxy* proxyToCheck) const { |
708 | bool used = false; |
709 | |
710 | auto visit = [ proxyToCheck, &used ] (GrSurfaceProxy* p, GrMipMapped) { |
711 | if (p == proxyToCheck) { |
712 | used = true; |
713 | } |
714 | }; |
715 | for (const OpChain& recordedOp : fOpChains) { |
716 | recordedOp.visitProxies(visit); |
717 | } |
718 | |
719 | return used; |
720 | } |
721 | |
722 | void GrOpsTask::handleInternalAllocationFailure() { |
723 | bool hasUninstantiatedProxy = false; |
724 | auto checkInstantiation = [&hasUninstantiatedProxy](GrSurfaceProxy* p, GrMipMapped) { |
725 | if (!p->isInstantiated()) { |
726 | hasUninstantiatedProxy = true; |
727 | } |
728 | }; |
729 | for (OpChain& recordedOp : fOpChains) { |
730 | hasUninstantiatedProxy = false; |
731 | recordedOp.visitProxies(checkInstantiation); |
732 | if (hasUninstantiatedProxy) { |
733 | recordedOp.setSkipExecuteFlag(); |
734 | } |
735 | } |
736 | } |
737 | |
738 | void GrOpsTask::gatherProxyIntervals(GrResourceAllocator* alloc) const { |
739 | for (int i = 0; i < fDeferredProxies.count(); ++i) { |
740 | SkASSERT(!fDeferredProxies[i]->isInstantiated()); |
741 | // We give all the deferred proxies a write usage at the very start of flushing. This |
742 | // locks them out of being reused for the entire flush until they are read - and then |
743 | // they can be recycled. This is a bit unfortunate because a flush can proceed in waves |
744 | // with sub-flushes. The deferred proxies only need to be pinned from the start of |
745 | // the sub-flush in which they appear. |
746 | alloc->addInterval(fDeferredProxies[i], 0, 0, GrResourceAllocator::ActualUse::kNo); |
747 | } |
748 | |
749 | GrSurfaceProxy* targetProxy = fTargetView.proxy(); |
750 | |
751 | // Add the interval for all the writes to this GrOpsTasks's target |
752 | if (fOpChains.count()) { |
753 | unsigned int cur = alloc->curOp(); |
754 | |
755 | alloc->addInterval(targetProxy, cur, cur + fOpChains.count() - 1, |
756 | GrResourceAllocator::ActualUse::kYes); |
757 | } else { |
758 | // This can happen if there is a loadOp (e.g., a clear) but no other draws. In this case we |
759 | // still need to add an interval for the destination so we create a fake op# for |
760 | // the missing clear op. |
761 | alloc->addInterval(targetProxy, alloc->curOp(), alloc->curOp(), |
762 | GrResourceAllocator::ActualUse::kYes); |
763 | alloc->incOps(); |
764 | } |
765 | |
766 | auto gather = [ alloc SkDEBUGCODE(, this) ] (GrSurfaceProxy* p, GrMipMapped) { |
767 | alloc->addInterval(p, alloc->curOp(), alloc->curOp(), GrResourceAllocator::ActualUse::kYes |
768 | SkDEBUGCODE(, fTargetView.proxy() == p)); |
769 | }; |
770 | for (const OpChain& recordedOp : fOpChains) { |
771 | recordedOp.visitProxies(gather); |
772 | |
773 | // Even though the op may have been (re)moved we still need to increment the op count to |
774 | // keep all the math consistent. |
775 | alloc->incOps(); |
776 | } |
777 | } |
778 | |
779 | void GrOpsTask::recordOp( |
780 | std::unique_ptr<GrOp> op, GrProcessorSet::Analysis processorAnalysis, GrAppliedClip* clip, |
781 | const DstProxyView* dstProxyView, const GrCaps& caps) { |
782 | SkDEBUGCODE(op->validate();) |
783 | SkASSERT(processorAnalysis.requiresDstTexture() == (dstProxyView && dstProxyView->proxy())); |
784 | GrSurfaceProxy* proxy = fTargetView.proxy(); |
785 | SkASSERT(proxy); |
786 | |
787 | // A closed GrOpsTask should never receive new/more ops |
788 | SkASSERT(!this->isClosed()); |
789 | if (!op->bounds().isFinite()) { |
790 | fArenas.opMemoryPool()->release(std::move(op)); |
791 | return; |
792 | } |
793 | |
794 | // Account for this op's bounds before we attempt to combine. |
795 | // NOTE: The caller should have already called "op->setClippedBounds()" by now, if applicable. |
796 | fTotalBounds.join(op->bounds()); |
797 | |
798 | // Check if there is an op we can combine with by linearly searching back until we either |
799 | // 1) check every op |
800 | // 2) intersect with something |
801 | // 3) find a 'blocker' |
802 | GR_AUDIT_TRAIL_ADD_OP(fAuditTrail, op.get(), proxy->uniqueID()); |
803 | GrOP_INFO("opsTask: %d Recording (%s, opID: %u)\n" |
804 | "\tBounds [L: %.2f, T: %.2f R: %.2f B: %.2f]\n" , |
805 | this->uniqueID(), |
806 | op->name(), |
807 | op->uniqueID(), |
808 | op->bounds().fLeft, op->bounds().fTop, |
809 | op->bounds().fRight, op->bounds().fBottom); |
810 | GrOP_INFO(SkTabString(op->dumpInfo(), 1).c_str()); |
811 | GrOP_INFO("\tOutcome:\n" ); |
812 | int maxCandidates = std::min(kMaxOpChainDistance, fOpChains.count()); |
813 | if (maxCandidates) { |
814 | int i = 0; |
815 | while (true) { |
816 | OpChain& candidate = fOpChains.fromBack(i); |
817 | op = candidate.appendOp(std::move(op), processorAnalysis, dstProxyView, clip, caps, |
818 | &fArenas, fAuditTrail); |
819 | if (!op) { |
820 | return; |
821 | } |
822 | // Stop going backwards if we would cause a painter's order violation. |
823 | if (!can_reorder(candidate.bounds(), op->bounds())) { |
824 | GrOP_INFO("\t\tBackward: Intersects with chain (%s, head opID: %u)\n" , |
825 | candidate.head()->name(), candidate.head()->uniqueID()); |
826 | break; |
827 | } |
828 | if (++i == maxCandidates) { |
829 | GrOP_INFO("\t\tBackward: Reached max lookback or beginning of op array %d\n" , i); |
830 | break; |
831 | } |
832 | } |
833 | } else { |
834 | GrOP_INFO("\t\tBackward: FirstOp\n" ); |
835 | } |
836 | if (clip) { |
837 | clip = fClipAllocator.make<GrAppliedClip>(std::move(*clip)); |
838 | SkDEBUGCODE(fNumClips++;) |
839 | } |
840 | fOpChains.emplace_back(std::move(op), processorAnalysis, clip, dstProxyView); |
841 | } |
842 | |
843 | void GrOpsTask::forwardCombine(const GrCaps& caps) { |
844 | SkASSERT(!this->isClosed()); |
845 | GrOP_INFO("opsTask: %d ForwardCombine %d ops:\n" , this->uniqueID(), fOpChains.count()); |
846 | |
847 | for (int i = 0; i < fOpChains.count() - 1; ++i) { |
848 | OpChain& chain = fOpChains[i]; |
849 | int maxCandidateIdx = std::min(i + kMaxOpChainDistance, fOpChains.count() - 1); |
850 | int j = i + 1; |
851 | while (true) { |
852 | OpChain& candidate = fOpChains[j]; |
853 | if (candidate.prependChain(&chain, caps, &fArenas, fAuditTrail)) { |
854 | break; |
855 | } |
856 | // Stop traversing if we would cause a painter's order violation. |
857 | if (!can_reorder(chain.bounds(), candidate.bounds())) { |
858 | GrOP_INFO( |
859 | "\t\t%d: chain (%s head opID: %u) -> " |
860 | "Intersects with chain (%s, head opID: %u)\n" , |
861 | i, chain.head()->name(), chain.head()->uniqueID(), candidate.head()->name(), |
862 | candidate.head()->uniqueID()); |
863 | break; |
864 | } |
865 | if (++j > maxCandidateIdx) { |
866 | GrOP_INFO("\t\t%d: chain (%s opID: %u) -> Reached max lookahead or end of array\n" , |
867 | i, chain.head()->name(), chain.head()->uniqueID()); |
868 | break; |
869 | } |
870 | } |
871 | } |
872 | } |
873 | |
874 | GrRenderTask::ExpectedOutcome GrOpsTask::onMakeClosed( |
875 | const GrCaps& caps, SkIRect* targetUpdateBounds) { |
876 | this->forwardCombine(caps); |
877 | if (!this->isNoOp()) { |
878 | GrSurfaceProxy* proxy = fTargetView.proxy(); |
879 | SkRect clippedContentBounds = proxy->getBoundsRect(); |
880 | // TODO: If we can fix up GLPrograms test to always intersect the fTargetView proxy bounds |
881 | // then we can simply assert here that the bounds intersect. |
882 | if (clippedContentBounds.intersect(fTotalBounds)) { |
883 | clippedContentBounds.roundOut(&fClippedContentBounds); |
884 | *targetUpdateBounds = fClippedContentBounds; |
885 | return ExpectedOutcome::kTargetDirty; |
886 | } |
887 | } |
888 | return ExpectedOutcome::kTargetUnchanged; |
889 | } |
890 | |