| 1 | /* |
| 2 | * Copyright 2017 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/ccpr/GrCoverageCountingPathRenderer.h" |
| 9 | |
| 10 | #include "include/pathops/SkPathOps.h" |
| 11 | #include "src/gpu/GrCaps.h" |
| 12 | #include "src/gpu/GrClip.h" |
| 13 | #include "src/gpu/GrProxyProvider.h" |
| 14 | #include "src/gpu/ccpr/GrCCClipProcessor.h" |
| 15 | #include "src/gpu/ccpr/GrCCDrawPathsOp.h" |
| 16 | #include "src/gpu/ccpr/GrCCPathCache.h" |
| 17 | |
| 18 | using PathInstance = GrCCPathProcessor::Instance; |
| 19 | |
| 20 | bool GrCoverageCountingPathRenderer::IsSupported(const GrCaps& caps, CoverageType* coverageType) { |
| 21 | const GrShaderCaps& shaderCaps = *caps.shaderCaps(); |
| 22 | GrBackendFormat defaultA8Format = caps.getDefaultBackendFormat(GrColorType::kAlpha_8, |
| 23 | GrRenderable::kYes); |
| 24 | if (caps.driverBlacklistCCPR() || !shaderCaps.integerSupport() || |
| 25 | !caps.drawInstancedSupport() || !shaderCaps.floatIs32Bits() || |
| 26 | !defaultA8Format.isValid() || // This checks both texturable and renderable |
| 27 | !caps.halfFloatVertexAttributeSupport()) { |
| 28 | return false; |
| 29 | } |
| 30 | |
| 31 | GrBackendFormat defaultAHalfFormat = caps.getDefaultBackendFormat(GrColorType::kAlpha_F16, |
| 32 | GrRenderable::kYes); |
| 33 | if (caps.allowCoverageCounting() && |
| 34 | defaultAHalfFormat.isValid()) { // This checks both texturable and renderable |
| 35 | if (coverageType) { |
| 36 | *coverageType = CoverageType::kFP16_CoverageCount; |
| 37 | } |
| 38 | return true; |
| 39 | } |
| 40 | |
| 41 | if (!caps.driverBlacklistMSAACCPR() && |
| 42 | caps.internalMultisampleCount(defaultA8Format) > 1 && |
| 43 | caps.sampleLocationsSupport() && |
| 44 | shaderCaps.sampleMaskSupport()) { |
| 45 | if (coverageType) { |
| 46 | *coverageType = CoverageType::kA8_Multisample; |
| 47 | } |
| 48 | return true; |
| 49 | } |
| 50 | |
| 51 | return false; |
| 52 | } |
| 53 | |
| 54 | sk_sp<GrCoverageCountingPathRenderer> GrCoverageCountingPathRenderer::CreateIfSupported( |
| 55 | const GrCaps& caps, AllowCaching allowCaching, uint32_t contextUniqueID) { |
| 56 | CoverageType coverageType; |
| 57 | if (IsSupported(caps, &coverageType)) { |
| 58 | return sk_sp<GrCoverageCountingPathRenderer>(new GrCoverageCountingPathRenderer( |
| 59 | coverageType, allowCaching, contextUniqueID)); |
| 60 | } |
| 61 | return nullptr; |
| 62 | } |
| 63 | |
| 64 | GrCoverageCountingPathRenderer::GrCoverageCountingPathRenderer( |
| 65 | CoverageType coverageType, AllowCaching allowCaching, uint32_t contextUniqueID) |
| 66 | : fCoverageType(coverageType) { |
| 67 | if (AllowCaching::kYes == allowCaching) { |
| 68 | fPathCache = std::make_unique<GrCCPathCache>(contextUniqueID); |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | GrCCPerOpsTaskPaths* GrCoverageCountingPathRenderer::lookupPendingPaths(uint32_t opsTaskID) { |
| 73 | auto it = fPendingPaths.find(opsTaskID); |
| 74 | if (fPendingPaths.end() == it) { |
| 75 | sk_sp<GrCCPerOpsTaskPaths> paths = sk_make_sp<GrCCPerOpsTaskPaths>(); |
| 76 | it = fPendingPaths.insert(std::make_pair(opsTaskID, std::move(paths))).first; |
| 77 | } |
| 78 | return it->second.get(); |
| 79 | } |
| 80 | |
| 81 | GrPathRenderer::CanDrawPath GrCoverageCountingPathRenderer::onCanDrawPath( |
| 82 | const CanDrawPathArgs& args) const { |
| 83 | const GrShape& shape = *args.fShape; |
| 84 | // We use "kCoverage", or analytic AA, no mater what the coverage type of our atlas: Even if the |
| 85 | // atlas is multisampled, that resolves into analytic coverage before we draw the path to the |
| 86 | // main canvas. |
| 87 | if (GrAAType::kCoverage != args.fAAType || shape.style().hasPathEffect() || |
| 88 | args.fViewMatrix->hasPerspective() || shape.inverseFilled()) { |
| 89 | return CanDrawPath::kNo; |
| 90 | } |
| 91 | |
| 92 | SkPath path; |
| 93 | shape.asPath(&path); |
| 94 | |
| 95 | const SkStrokeRec& stroke = shape.style().strokeRec(); |
| 96 | switch (stroke.getStyle()) { |
| 97 | case SkStrokeRec::kFill_Style: { |
| 98 | SkRect devBounds; |
| 99 | args.fViewMatrix->mapRect(&devBounds, path.getBounds()); |
| 100 | |
| 101 | SkIRect clippedIBounds; |
| 102 | devBounds.roundOut(&clippedIBounds); |
| 103 | if (!clippedIBounds.intersect(*args.fClipConservativeBounds)) { |
| 104 | // The path is completely clipped away. Our code will eventually notice this before |
| 105 | // doing any real work. |
| 106 | return CanDrawPath::kYes; |
| 107 | } |
| 108 | |
| 109 | int64_t numPixels = sk_64_mul(clippedIBounds.height(), clippedIBounds.width()); |
| 110 | if (path.countVerbs() > 1000 && path.countPoints() > numPixels) { |
| 111 | // This is a complicated path that has more vertices than pixels! Let's let the SW |
| 112 | // renderer have this one: It will probably be faster and a bitmap will require less |
| 113 | // total memory on the GPU than CCPR instance buffers would for the raw path data. |
| 114 | return CanDrawPath::kNo; |
| 115 | } |
| 116 | |
| 117 | if (numPixels > 256 * 256) { |
| 118 | // Large paths can blow up the atlas fast. And they are not ideal for a two-pass |
| 119 | // rendering algorithm. Give the simpler direct renderers a chance before we commit |
| 120 | // to drawing it. |
| 121 | return CanDrawPath::kAsBackup; |
| 122 | } |
| 123 | |
| 124 | if (args.fShape->hasUnstyledKey() && path.countVerbs() > 50) { |
| 125 | // Complex paths do better cached in an SDF, if the renderer will accept them. |
| 126 | return CanDrawPath::kAsBackup; |
| 127 | } |
| 128 | |
| 129 | return CanDrawPath::kYes; |
| 130 | } |
| 131 | |
| 132 | case SkStrokeRec::kStroke_Style: |
| 133 | if (!args.fViewMatrix->isSimilarity()) { |
| 134 | // The stroker currently only supports rigid-body transfoms for the stroke lines |
| 135 | // themselves. This limitation doesn't affect hairlines since their stroke lines are |
| 136 | // defined relative to device space. |
| 137 | return CanDrawPath::kNo; |
| 138 | } |
| 139 | // fallthru |
| 140 | case SkStrokeRec::kHairline_Style: { |
| 141 | if (CoverageType::kFP16_CoverageCount != fCoverageType) { |
| 142 | // Stroking is not yet supported in MSAA atlas mode. |
| 143 | return CanDrawPath::kNo; |
| 144 | } |
| 145 | float inflationRadius; |
| 146 | GetStrokeDevWidth(*args.fViewMatrix, stroke, &inflationRadius); |
| 147 | if (!(inflationRadius <= kMaxBoundsInflationFromStroke)) { |
| 148 | // Let extremely wide strokes be converted to fill paths and drawn by the CCPR |
| 149 | // filler instead. (Cast the logic negatively in order to also catch r=NaN.) |
| 150 | return CanDrawPath::kNo; |
| 151 | } |
| 152 | SkASSERT(!SkScalarIsNaN(inflationRadius)); |
| 153 | if (SkPathPriv::ConicWeightCnt(path)) { |
| 154 | // The stroker does not support conics yet. |
| 155 | return CanDrawPath::kNo; |
| 156 | } |
| 157 | return CanDrawPath::kYes; |
| 158 | } |
| 159 | |
| 160 | case SkStrokeRec::kStrokeAndFill_Style: |
| 161 | return CanDrawPath::kNo; |
| 162 | } |
| 163 | |
| 164 | SK_ABORT("Invalid stroke style." ); |
| 165 | } |
| 166 | |
| 167 | bool GrCoverageCountingPathRenderer::onDrawPath(const DrawPathArgs& args) { |
| 168 | SkASSERT(!fFlushing); |
| 169 | |
| 170 | SkIRect clipIBounds; |
| 171 | GrRenderTargetContext* rtc = args.fRenderTargetContext; |
| 172 | args.fClip->getConservativeBounds(rtc->width(), rtc->height(), &clipIBounds, nullptr); |
| 173 | |
| 174 | auto op = GrCCDrawPathsOp::Make(args.fContext, clipIBounds, *args.fViewMatrix, *args.fShape, |
| 175 | std::move(args.fPaint)); |
| 176 | this->recordOp(std::move(op), args); |
| 177 | return true; |
| 178 | } |
| 179 | |
| 180 | void GrCoverageCountingPathRenderer::recordOp(std::unique_ptr<GrCCDrawPathsOp> op, |
| 181 | const DrawPathArgs& args) { |
| 182 | if (op) { |
| 183 | auto addToOwningPerOpsTaskPaths = [this](GrOp* op, uint32_t opsTaskID) { |
| 184 | op->cast<GrCCDrawPathsOp>()->addToOwningPerOpsTaskPaths( |
| 185 | sk_ref_sp(this->lookupPendingPaths(opsTaskID))); |
| 186 | }; |
| 187 | args.fRenderTargetContext->addDrawOp(*args.fClip, std::move(op), |
| 188 | addToOwningPerOpsTaskPaths); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | std::unique_ptr<GrFragmentProcessor> GrCoverageCountingPathRenderer::makeClipProcessor( |
| 193 | uint32_t opsTaskID, const SkPath& deviceSpacePath, const SkIRect& accessRect, |
| 194 | const GrCaps& caps) { |
| 195 | SkASSERT(!fFlushing); |
| 196 | |
| 197 | uint32_t key = deviceSpacePath.getGenerationID(); |
| 198 | if (CoverageType::kA8_Multisample == fCoverageType) { |
| 199 | // We only need to consider fill rule in MSAA mode. In coverage count mode Even/Odd and |
| 200 | // Nonzero both reference the same coverage count mask. |
| 201 | key = (key << 1) | (uint32_t)GrFillRuleForSkPath(deviceSpacePath); |
| 202 | } |
| 203 | GrCCClipPath& clipPath = |
| 204 | this->lookupPendingPaths(opsTaskID)->fClipPaths[key]; |
| 205 | if (!clipPath.isInitialized()) { |
| 206 | // This ClipPath was just created during lookup. Initialize it. |
| 207 | const SkRect& pathDevBounds = deviceSpacePath.getBounds(); |
| 208 | if (std::max(pathDevBounds.height(), pathDevBounds.width()) > kPathCropThreshold) { |
| 209 | // The path is too large. Crop it or analytic AA can run out of fp32 precision. |
| 210 | SkPath croppedPath; |
| 211 | int maxRTSize = caps.maxRenderTargetSize(); |
| 212 | CropPath(deviceSpacePath, SkIRect::MakeWH(maxRTSize, maxRTSize), &croppedPath); |
| 213 | clipPath.init(croppedPath, accessRect, fCoverageType, caps); |
| 214 | } else { |
| 215 | clipPath.init(deviceSpacePath, accessRect, fCoverageType, caps); |
| 216 | } |
| 217 | } else { |
| 218 | clipPath.addAccess(accessRect); |
| 219 | } |
| 220 | |
| 221 | auto isCoverageCount = GrCCClipProcessor::IsCoverageCount( |
| 222 | CoverageType::kFP16_CoverageCount == fCoverageType); |
| 223 | auto mustCheckBounds = GrCCClipProcessor::MustCheckBounds( |
| 224 | !clipPath.pathDevIBounds().contains(accessRect)); |
| 225 | return std::make_unique<GrCCClipProcessor>(caps, &clipPath, isCoverageCount, mustCheckBounds); |
| 226 | } |
| 227 | |
| 228 | void GrCoverageCountingPathRenderer::preFlush( |
| 229 | GrOnFlushResourceProvider* onFlushRP, const uint32_t* opsTaskIDs, int numOpsTaskIDs) { |
| 230 | using DoCopiesToA8Coverage = GrCCDrawPathsOp::DoCopiesToA8Coverage; |
| 231 | SkASSERT(!fFlushing); |
| 232 | SkASSERT(fFlushingPaths.empty()); |
| 233 | SkDEBUGCODE(fFlushing = true); |
| 234 | |
| 235 | if (fPathCache) { |
| 236 | fPathCache->doPreFlushProcessing(); |
| 237 | } |
| 238 | |
| 239 | if (fPendingPaths.empty()) { |
| 240 | return; // Nothing to draw. |
| 241 | } |
| 242 | |
| 243 | GrCCPerFlushResourceSpecs specs; |
| 244 | int maxPreferredRTSize = onFlushRP->caps()->maxPreferredRenderTargetSize(); |
| 245 | specs.fCopyAtlasSpecs.fMaxPreferredTextureSize = std::min(2048, maxPreferredRTSize); |
| 246 | SkASSERT(0 == specs.fCopyAtlasSpecs.fMinTextureSize); |
| 247 | specs.fRenderedAtlasSpecs.fMaxPreferredTextureSize = maxPreferredRTSize; |
| 248 | specs.fRenderedAtlasSpecs.fMinTextureSize = std::min(512, maxPreferredRTSize); |
| 249 | |
| 250 | // Move the per-opsTask paths that are about to be flushed from fPendingPaths to fFlushingPaths, |
| 251 | // and count them up so we can preallocate buffers. |
| 252 | fFlushingPaths.reserve(numOpsTaskIDs); |
| 253 | for (int i = 0; i < numOpsTaskIDs; ++i) { |
| 254 | auto iter = fPendingPaths.find(opsTaskIDs[i]); |
| 255 | if (fPendingPaths.end() == iter) { |
| 256 | continue; // No paths on this opsTask. |
| 257 | } |
| 258 | |
| 259 | fFlushingPaths.push_back(std::move(iter->second)); |
| 260 | fPendingPaths.erase(iter); |
| 261 | |
| 262 | for (GrCCDrawPathsOp* op : fFlushingPaths.back()->fDrawOps) { |
| 263 | op->accountForOwnPaths(fPathCache.get(), onFlushRP, &specs); |
| 264 | } |
| 265 | for (const auto& clipsIter : fFlushingPaths.back()->fClipPaths) { |
| 266 | clipsIter.second.accountForOwnPath(&specs); |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | if (specs.isEmpty()) { |
| 271 | return; // Nothing to draw. |
| 272 | } |
| 273 | |
| 274 | // Determine if there are enough reusable paths from last flush for it to be worth our time to |
| 275 | // copy them to cached atlas(es). |
| 276 | int numCopies = specs.fNumCopiedPaths[GrCCPerFlushResourceSpecs::kFillIdx] + |
| 277 | specs.fNumCopiedPaths[GrCCPerFlushResourceSpecs::kStrokeIdx]; |
| 278 | auto doCopies = DoCopiesToA8Coverage(numCopies > 100 || |
| 279 | specs.fCopyAtlasSpecs.fApproxNumPixels > 256 * 256); |
| 280 | if (numCopies && DoCopiesToA8Coverage::kNo == doCopies) { |
| 281 | specs.cancelCopies(); |
| 282 | } |
| 283 | |
| 284 | auto resources = sk_make_sp<GrCCPerFlushResources>(onFlushRP, fCoverageType, specs); |
| 285 | if (!resources->isMapped()) { |
| 286 | return; // Some allocation failed. |
| 287 | } |
| 288 | |
| 289 | // Layout the atlas(es) and parse paths. |
| 290 | for (const auto& flushingPaths : fFlushingPaths) { |
| 291 | for (GrCCDrawPathsOp* op : flushingPaths->fDrawOps) { |
| 292 | op->setupResources(fPathCache.get(), onFlushRP, resources.get(), doCopies); |
| 293 | } |
| 294 | for (auto& clipsIter : flushingPaths->fClipPaths) { |
| 295 | clipsIter.second.renderPathInAtlas(resources.get(), onFlushRP); |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | if (fPathCache) { |
| 300 | // Purge invalidated textures from previous atlases *before* calling finalize(). That way, |
| 301 | // the underlying textures objects can be freed up and reused for the next atlases. |
| 302 | fPathCache->purgeInvalidatedAtlasTextures(onFlushRP); |
| 303 | } |
| 304 | |
| 305 | // Allocate resources and then render the atlas(es). |
| 306 | if (!resources->finalize(onFlushRP)) { |
| 307 | return; |
| 308 | } |
| 309 | |
| 310 | // Commit flushing paths to the resources once they are successfully completed. |
| 311 | for (auto& flushingPaths : fFlushingPaths) { |
| 312 | SkASSERT(!flushingPaths->fFlushResources); |
| 313 | flushingPaths->fFlushResources = resources; |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | void GrCoverageCountingPathRenderer::postFlush(GrDeferredUploadToken, const uint32_t* opsTaskIDs, |
| 318 | int numOpsTaskIDs) { |
| 319 | SkASSERT(fFlushing); |
| 320 | |
| 321 | if (!fFlushingPaths.empty()) { |
| 322 | // In DDL mode these aren't guaranteed to be deleted so we must clear out the perFlush |
| 323 | // resources manually. |
| 324 | for (auto& flushingPaths : fFlushingPaths) { |
| 325 | flushingPaths->fFlushResources = nullptr; |
| 326 | } |
| 327 | |
| 328 | // We wait to erase these until after flush, once Ops and FPs are done accessing their data. |
| 329 | fFlushingPaths.reset(); |
| 330 | } |
| 331 | |
| 332 | SkDEBUGCODE(fFlushing = false); |
| 333 | } |
| 334 | |
| 335 | void GrCoverageCountingPathRenderer::purgeCacheEntriesOlderThan( |
| 336 | GrProxyProvider* proxyProvider, const GrStdSteadyClock::time_point& purgeTime) { |
| 337 | if (fPathCache) { |
| 338 | fPathCache->purgeEntriesOlderThan(proxyProvider, purgeTime); |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | void GrCoverageCountingPathRenderer::CropPath(const SkPath& path, const SkIRect& cropbox, |
| 343 | SkPath* out) { |
| 344 | SkPath cropboxPath; |
| 345 | cropboxPath.addRect(SkRect::Make(cropbox)); |
| 346 | if (!Op(cropboxPath, path, kIntersect_SkPathOp, out)) { |
| 347 | // This can fail if the PathOps encounter NaN or infinities. |
| 348 | out->reset(); |
| 349 | } |
| 350 | out->setIsVolatile(true); |
| 351 | } |
| 352 | |
| 353 | float GrCoverageCountingPathRenderer::GetStrokeDevWidth(const SkMatrix& m, |
| 354 | const SkStrokeRec& stroke, |
| 355 | float* inflationRadius) { |
| 356 | float strokeDevWidth; |
| 357 | if (stroke.isHairlineStyle()) { |
| 358 | strokeDevWidth = 1; |
| 359 | } else { |
| 360 | SkASSERT(SkStrokeRec::kStroke_Style == stroke.getStyle()); |
| 361 | SkASSERT(m.isSimilarity()); // Otherwise matrixScaleFactor = m.getMaxScale(). |
| 362 | float matrixScaleFactor = SkVector::Length(m.getScaleX(), m.getSkewY()); |
| 363 | strokeDevWidth = stroke.getWidth() * matrixScaleFactor; |
| 364 | } |
| 365 | if (inflationRadius) { |
| 366 | // Inflate for a minimum stroke width of 1. In some cases when the stroke is less than 1px |
| 367 | // wide, we may inflate it to 1px and instead reduce the opacity. |
| 368 | *inflationRadius = SkStrokeRec::GetInflationRadius( |
| 369 | stroke.getJoin(), stroke.getMiter(), stroke.getCap(), std::max(strokeDevWidth, 1.f)); |
| 370 | } |
| 371 | return strokeDevWidth; |
| 372 | } |
| 373 | |