| 1 | /* |
| 2 | * Copyright 2019 Google LLC. |
| 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/tessellate/GrPathTessellateOp.h" |
| 9 | |
| 10 | #include "src/gpu/GrEagerVertexAllocator.h" |
| 11 | #include "src/gpu/GrGpu.h" |
| 12 | #include "src/gpu/GrOpFlushState.h" |
| 13 | #include "src/gpu/GrTriangulator.h" |
| 14 | #include "src/gpu/tessellate/GrFillPathShader.h" |
| 15 | #include "src/gpu/tessellate/GrMiddleOutPolygonTriangulator.h" |
| 16 | #include "src/gpu/tessellate/GrMidpointContourParser.h" |
| 17 | #include "src/gpu/tessellate/GrResolveLevelCounter.h" |
| 18 | #include "src/gpu/tessellate/GrStencilPathShader.h" |
| 19 | #include "src/gpu/tessellate/GrTessellationPathRenderer.h" |
| 20 | |
| 21 | constexpr static float kLinearizationIntolerance = |
| 22 | GrTessellationPathRenderer::kLinearizationIntolerance; |
| 23 | |
| 24 | constexpr static int kMaxResolveLevel = GrTessellationPathRenderer::kMaxResolveLevel; |
| 25 | |
| 26 | using OpFlags = GrTessellationPathRenderer::OpFlags; |
| 27 | |
| 28 | GrPathTessellateOp::FixedFunctionFlags GrPathTessellateOp::fixedFunctionFlags() const { |
| 29 | auto flags = FixedFunctionFlags::kUsesStencil; |
| 30 | if (GrAAType::kNone != fAAType) { |
| 31 | flags |= FixedFunctionFlags::kUsesHWAA; |
| 32 | } |
| 33 | return flags; |
| 34 | } |
| 35 | |
| 36 | void GrPathTessellateOp::onPrePrepare(GrRecordingContext*, |
| 37 | const GrSurfaceProxyView* writeView, |
| 38 | GrAppliedClip*, |
| 39 | const GrXferProcessor::DstProxyView&) { |
| 40 | } |
| 41 | |
| 42 | void GrPathTessellateOp::onPrepare(GrOpFlushState* flushState) { |
| 43 | int numVerbs = fPath.countVerbs(); |
| 44 | if (numVerbs <= 0) { |
| 45 | return; |
| 46 | } |
| 47 | |
| 48 | // First check if the path is large and/or simple enough that we can actually triangulate the |
| 49 | // inner polygon(s) on the CPU. This is our fastest approach. It allows us to stencil only the |
| 50 | // curves, and then fill the internal polygons directly to the final render target, thus drawing |
| 51 | // the majority of pixels in a single render pass. |
| 52 | SkScalar scales[2]; |
| 53 | SkAssertResult(fViewMatrix.getMinMaxScales(scales)); // Will fail if perspective. |
| 54 | const SkRect& bounds = fPath.getBounds(); |
| 55 | float gpuFragmentWork = bounds.height() * scales[0] * bounds.width() * scales[1]; |
| 56 | float cpuTessellationWork = (float)numVerbs * SkNextLog2(numVerbs); // N log N. |
| 57 | if (cpuTessellationWork * 500 + (256 * 256) < gpuFragmentWork) { // Don't try below 256x256. |
| 58 | int numCountedCubics; |
| 59 | // This will fail if the inner triangles do not form a simple polygon (e.g., self |
| 60 | // intersection, double winding). |
| 61 | if (this->prepareNonOverlappingInnerTriangles(flushState, &numCountedCubics)) { |
| 62 | if (!numCountedCubics) { |
| 63 | return; |
| 64 | } |
| 65 | // Always use indirect draws for cubics instead of tessellation here. Our goal in this |
| 66 | // mode is to maximize GPU performance, and the middle-out topology used by our indirect |
| 67 | // draws is easier on the rasterizer than a tessellated fan. There also seems to be a |
| 68 | // small amount of fixed tessellation overhead that this avoids. |
| 69 | GrResolveLevelCounter resolveLevelCounter; |
| 70 | resolveLevelCounter.reset(fPath, fViewMatrix, kLinearizationIntolerance); |
| 71 | this->prepareIndirectOuterCubics(flushState, resolveLevelCounter); |
| 72 | return; |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | // When there are only a few verbs, it seems to always be fastest to make a single indirect draw |
| 77 | // that contains both the inner triangles and the outer cubics, instead of using hardware |
| 78 | // tessellation. Also take this path if tessellation is not supported. |
| 79 | bool drawTrianglesAsIndirectCubicDraw = (numVerbs < 50); |
| 80 | if (drawTrianglesAsIndirectCubicDraw || (fOpFlags & OpFlags::kDisableHWTessellation)) { |
| 81 | // Prepare outer cubics with indirect draws. |
| 82 | GrResolveLevelCounter resolveLevelCounter; |
| 83 | this->prepareMiddleOutTrianglesAndCubics(flushState, &resolveLevelCounter, |
| 84 | drawTrianglesAsIndirectCubicDraw); |
| 85 | return; |
| 86 | } |
| 87 | |
| 88 | // The caller should have sent Flags::kDisableHWTessellation if it was not supported. |
| 89 | SkASSERT(flushState->caps().shaderCaps()->tessellationSupport()); |
| 90 | |
| 91 | // Next see if we can split up the inner triangles and outer cubics into two draw calls. This |
| 92 | // allows for a more efficient inner triangle topology that can reduce the rasterizer load by a |
| 93 | // large margin on complex paths, but also causes greater CPU overhead due to the extra shader |
| 94 | // switches and draw calls. |
| 95 | // NOTE: Raster-edge work is 1-dimensional, so we sum height and width instead of multiplying. |
| 96 | float rasterEdgeWork = (bounds.height() + bounds.width()) * scales[1] * fPath.countVerbs(); |
| 97 | if (rasterEdgeWork > 300 * 300) { |
| 98 | this->prepareMiddleOutTrianglesAndCubics(flushState); |
| 99 | return; |
| 100 | } |
| 101 | |
| 102 | // Fastest CPU approach: emit one cubic wedge per verb, fanning out from the center. |
| 103 | this->prepareTessellatedCubicWedges(flushState); |
| 104 | } |
| 105 | |
| 106 | bool GrPathTessellateOp::prepareNonOverlappingInnerTriangles(GrMeshDrawOp::Target* target, |
| 107 | int* numCountedCurves) { |
| 108 | SkASSERT(!fTriangleBuffer); |
| 109 | SkASSERT(!fDoStencilTriangleBuffer); |
| 110 | SkASSERT(!fDoFillTriangleBuffer); |
| 111 | |
| 112 | using GrTriangulator::Mode; |
| 113 | |
| 114 | GrEagerDynamicVertexAllocator vertexAlloc(target, &fTriangleBuffer, &fBaseTriangleVertex); |
| 115 | fTriangleVertexCount = GrTriangulator::PathToTriangles(fPath, 0, SkRect::MakeEmpty(), |
| 116 | &vertexAlloc, Mode::kSimpleInnerPolygons, |
| 117 | numCountedCurves); |
| 118 | if (fTriangleVertexCount == 0) { |
| 119 | // Mode::kSimpleInnerPolygons causes PathToTriangles to fail if the inner polygon(s) are not |
| 120 | // simple. |
| 121 | return false; |
| 122 | } |
| 123 | if (((OpFlags::kStencilOnly | OpFlags::kWireframe) & fOpFlags) || |
| 124 | GrAAType::kCoverage == fAAType || |
| 125 | (target->appliedClip() && target->appliedClip()->hasStencilClip())) { |
| 126 | // If we have certain flags, mixed samples, or a stencil clip then we unfortunately |
| 127 | // can't fill the inner polygon directly. Indicate that these triangles need to be |
| 128 | // stencilled. |
| 129 | fDoStencilTriangleBuffer = true; |
| 130 | } |
| 131 | if (!(OpFlags::kStencilOnly & fOpFlags)) { |
| 132 | fDoFillTriangleBuffer = true; |
| 133 | } |
| 134 | return true; |
| 135 | } |
| 136 | |
| 137 | void GrPathTessellateOp::prepareMiddleOutTrianglesAndCubics( |
| 138 | GrMeshDrawOp::Target* target, GrResolveLevelCounter* resolveLevelCounter, |
| 139 | bool drawTrianglesAsIndirectCubicDraw) { |
| 140 | SkASSERT(!fTriangleBuffer); |
| 141 | SkASSERT(!fDoStencilTriangleBuffer); |
| 142 | SkASSERT(!fDoFillTriangleBuffer); |
| 143 | SkASSERT(!fCubicBuffer); |
| 144 | SkASSERT(!fStencilCubicsShader); |
| 145 | SkASSERT(!fIndirectDrawBuffer); |
| 146 | |
| 147 | // No initial moveTo, plus an implicit close at the end; n-2 triangles fill an n-gon. |
| 148 | int maxInnerTriangles = fPath.countVerbs() - 1; |
| 149 | int maxCubics = fPath.countVerbs(); |
| 150 | |
| 151 | SkPoint* vertexData; |
| 152 | int vertexAdvancePerTriangle; |
| 153 | if (drawTrianglesAsIndirectCubicDraw) { |
| 154 | // Allocate the triangles as 4-point instances at the beginning of the cubic buffer. |
| 155 | SkASSERT(resolveLevelCounter); |
| 156 | vertexAdvancePerTriangle = 4; |
| 157 | int baseTriangleInstance; |
| 158 | vertexData = static_cast<SkPoint*>(target->makeVertexSpace( |
| 159 | sizeof(SkPoint) * 4, maxInnerTriangles + maxCubics, &fCubicBuffer, |
| 160 | &baseTriangleInstance)); |
| 161 | fBaseCubicVertex = baseTriangleInstance * 4; |
| 162 | } else { |
| 163 | // Allocate the triangles as normal 3-point instances in the triangle buffer. |
| 164 | vertexAdvancePerTriangle = 3; |
| 165 | vertexData = static_cast<SkPoint*>(target->makeVertexSpace( |
| 166 | sizeof(SkPoint), maxInnerTriangles * 3, &fTriangleBuffer, &fBaseTriangleVertex)); |
| 167 | } |
| 168 | if (!vertexData) { |
| 169 | return; |
| 170 | } |
| 171 | |
| 172 | GrVectorXform xform(fViewMatrix); |
| 173 | GrMiddleOutPolygonTriangulator middleOut(vertexData, vertexAdvancePerTriangle, |
| 174 | fPath.countVerbs()); |
| 175 | if (resolveLevelCounter) { |
| 176 | resolveLevelCounter->reset(); |
| 177 | } |
| 178 | int numCountedCurves = 0; |
| 179 | for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) { |
| 180 | switch (verb) { |
| 181 | case SkPathVerb::kMove: |
| 182 | middleOut.closeAndMove(pts[0]); |
| 183 | break; |
| 184 | case SkPathVerb::kLine: |
| 185 | middleOut.pushVertex(pts[1]); |
| 186 | break; |
| 187 | case SkPathVerb::kQuad: |
| 188 | middleOut.pushVertex(pts[2]); |
| 189 | if (resolveLevelCounter) { |
| 190 | // Quadratics get converted to cubics before rendering. |
| 191 | resolveLevelCounter->countCubic(GrWangsFormula::quadratic_log2( |
| 192 | kLinearizationIntolerance, pts, xform)); |
| 193 | break; |
| 194 | } |
| 195 | ++numCountedCurves; |
| 196 | break; |
| 197 | case SkPathVerb::kCubic: |
| 198 | middleOut.pushVertex(pts[3]); |
| 199 | if (resolveLevelCounter) { |
| 200 | resolveLevelCounter->countCubic(GrWangsFormula::cubic_log2( |
| 201 | kLinearizationIntolerance, pts, xform)); |
| 202 | break; |
| 203 | } |
| 204 | ++numCountedCurves; |
| 205 | break; |
| 206 | case SkPathVerb::kClose: |
| 207 | middleOut.close(); |
| 208 | break; |
| 209 | case SkPathVerb::kConic: |
| 210 | SkUNREACHABLE; |
| 211 | } |
| 212 | } |
| 213 | int triangleCount = middleOut.close(); |
| 214 | SkASSERT(triangleCount <= maxInnerTriangles); |
| 215 | |
| 216 | if (drawTrianglesAsIndirectCubicDraw) { |
| 217 | SkASSERT(resolveLevelCounter); |
| 218 | int totalInstanceCount = triangleCount + resolveLevelCounter->totalCubicInstanceCount(); |
| 219 | SkASSERT(vertexAdvancePerTriangle == 4); |
| 220 | target->putBackVertices(maxInnerTriangles + maxCubics - totalInstanceCount, |
| 221 | sizeof(SkPoint) * 4); |
| 222 | if (totalInstanceCount) { |
| 223 | this->prepareIndirectOuterCubicsAndTriangles(target, *resolveLevelCounter, vertexData, |
| 224 | triangleCount); |
| 225 | } |
| 226 | } else { |
| 227 | SkASSERT(vertexAdvancePerTriangle == 3); |
| 228 | target->putBackVertices(maxInnerTriangles - triangleCount, sizeof(SkPoint) * 3); |
| 229 | fTriangleVertexCount = triangleCount * 3; |
| 230 | if (fTriangleVertexCount) { |
| 231 | fDoStencilTriangleBuffer = true; |
| 232 | } |
| 233 | if (resolveLevelCounter) { |
| 234 | this->prepareIndirectOuterCubics(target, *resolveLevelCounter); |
| 235 | } else { |
| 236 | this->prepareTessellatedOuterCubics(target, numCountedCurves); |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | static SkPoint lerp(const SkPoint& a, const SkPoint& b, float T) { |
| 242 | SkASSERT(1 != T); // The below does not guarantee lerp(a, b, 1) === b. |
| 243 | return (b - a) * T + a; |
| 244 | } |
| 245 | |
| 246 | static void line2cubic(const SkPoint& p0, const SkPoint& p1, SkPoint* out) { |
| 247 | out[0] = p0; |
| 248 | out[1] = lerp(p0, p1, 1/3.f); |
| 249 | out[2] = lerp(p0, p1, 2/3.f); |
| 250 | out[3] = p1; |
| 251 | } |
| 252 | |
| 253 | static void quad2cubic(const SkPoint pts[], SkPoint* out) { |
| 254 | out[0] = pts[0]; |
| 255 | out[1] = lerp(pts[0], pts[1], 2/3.f); |
| 256 | out[2] = lerp(pts[1], pts[2], 1/3.f); |
| 257 | out[3] = pts[2]; |
| 258 | } |
| 259 | |
| 260 | void GrPathTessellateOp::prepareIndirectOuterCubics( |
| 261 | GrMeshDrawOp::Target* target, const GrResolveLevelCounter& resolveLevelCounter) { |
| 262 | SkASSERT(resolveLevelCounter.totalCubicInstanceCount() >= 0); |
| 263 | if (resolveLevelCounter.totalCubicInstanceCount() == 0) { |
| 264 | return; |
| 265 | } |
| 266 | // Allocate a buffer to store the cubic data. |
| 267 | SkPoint* cubicData; |
| 268 | int baseInstance; |
| 269 | cubicData = static_cast<SkPoint*>(target->makeVertexSpace( |
| 270 | sizeof(SkPoint) * 4, resolveLevelCounter.totalCubicInstanceCount(), &fCubicBuffer, |
| 271 | &baseInstance)); |
| 272 | if (!cubicData) { |
| 273 | return; |
| 274 | } |
| 275 | fBaseCubicVertex = baseInstance * 4; |
| 276 | this->prepareIndirectOuterCubicsAndTriangles(target, resolveLevelCounter, cubicData, |
| 277 | /*numTrianglesAtBeginningOfData=*/0); |
| 278 | } |
| 279 | |
| 280 | void GrPathTessellateOp::prepareIndirectOuterCubicsAndTriangles( |
| 281 | GrMeshDrawOp::Target* target, const GrResolveLevelCounter& resolveLevelCounter, |
| 282 | SkPoint* cubicData, int numTrianglesAtBeginningOfData) { |
| 283 | SkASSERT(target->caps().drawInstancedSupport()); |
| 284 | SkASSERT(numTrianglesAtBeginningOfData + resolveLevelCounter.totalCubicInstanceCount() > 0); |
| 285 | SkASSERT(!fStencilCubicsShader); |
| 286 | SkASSERT(cubicData); |
| 287 | |
| 288 | fIndirectIndexBuffer = GrMiddleOutCubicShader::FindOrMakeMiddleOutIndexBuffer( |
| 289 | target->resourceProvider()); |
| 290 | if (!fIndirectIndexBuffer) { |
| 291 | return; |
| 292 | } |
| 293 | |
| 294 | // Here we treat fCubicBuffer as an instance buffer. It should have been prepared with the base |
| 295 | // vertex on an instance boundary in order to accommodate this. |
| 296 | SkASSERT(fBaseCubicVertex % 4 == 0); |
| 297 | int baseInstance = fBaseCubicVertex >> 2; |
| 298 | |
| 299 | // Start preparing the indirect draw buffer. |
| 300 | fIndirectDrawCount = resolveLevelCounter.totalCubicIndirectDrawCount(); |
| 301 | if (numTrianglesAtBeginningOfData) { |
| 302 | ++fIndirectDrawCount; // Add an indirect draw for the triangles at the beginning. |
| 303 | } |
| 304 | |
| 305 | // Allocate space for the GrDrawIndexedIndirectCommand structs. |
| 306 | GrDrawIndexedIndirectCommand* indirectData = target->makeDrawIndexedIndirectSpace( |
| 307 | fIndirectDrawCount, &fIndirectDrawBuffer, &fIndirectDrawOffset); |
| 308 | if (!indirectData) { |
| 309 | SkASSERT(!fIndirectDrawBuffer); |
| 310 | return; |
| 311 | } |
| 312 | |
| 313 | // Fill out the GrDrawIndexedIndirectCommand structs and determine the starting instance data |
| 314 | // location at each resolve level. |
| 315 | SkPoint* instanceLocations[kMaxResolveLevel + 1]; |
| 316 | int indirectIdx = 0; |
| 317 | int runningInstanceCount = 0; |
| 318 | if (numTrianglesAtBeginningOfData) { |
| 319 | // The caller has already packed "triangleInstanceCount" triangles into 4-point instances |
| 320 | // at the beginning of the instance buffer. Add a special-case indirect draw here that will |
| 321 | // emit the triangles [P0, P1, P2] from these 4-point instances. |
| 322 | indirectData[0] = GrMiddleOutCubicShader::MakeDrawTrianglesIndirectCmd( |
| 323 | numTrianglesAtBeginningOfData, baseInstance); |
| 324 | indirectIdx = 1; |
| 325 | runningInstanceCount = numTrianglesAtBeginningOfData; |
| 326 | } |
| 327 | for (int resolveLevel = 1; resolveLevel <= kMaxResolveLevel; ++resolveLevel) { |
| 328 | int instanceCountAtCurrLevel = resolveLevelCounter[resolveLevel]; |
| 329 | if (!instanceCountAtCurrLevel) { |
| 330 | SkDEBUGCODE(instanceLocations[resolveLevel] = nullptr;) |
| 331 | continue; |
| 332 | } |
| 333 | instanceLocations[resolveLevel] = cubicData + runningInstanceCount * 4; |
| 334 | indirectData[indirectIdx++] = GrMiddleOutCubicShader::MakeDrawCubicsIndirectCmd( |
| 335 | resolveLevel, instanceCountAtCurrLevel, baseInstance + runningInstanceCount); |
| 336 | runningInstanceCount += instanceCountAtCurrLevel; |
| 337 | } |
| 338 | |
| 339 | #ifdef SK_DEBUG |
| 340 | SkASSERT(indirectIdx == fIndirectDrawCount); |
| 341 | SkASSERT(runningInstanceCount == numTrianglesAtBeginningOfData + |
| 342 | resolveLevelCounter.totalCubicInstanceCount()); |
| 343 | SkASSERT(fIndirectDrawCount > 0); |
| 344 | |
| 345 | SkPoint* endLocations[kMaxResolveLevel + 1]; |
| 346 | int lastResolveLevel = 0; |
| 347 | for (int resolveLevel = 1; resolveLevel <= kMaxResolveLevel; ++resolveLevel) { |
| 348 | if (!instanceLocations[resolveLevel]) { |
| 349 | endLocations[resolveLevel] = nullptr; |
| 350 | continue; |
| 351 | } |
| 352 | endLocations[lastResolveLevel] = instanceLocations[resolveLevel]; |
| 353 | lastResolveLevel = resolveLevel; |
| 354 | } |
| 355 | int totalInstanceCount = numTrianglesAtBeginningOfData + |
| 356 | resolveLevelCounter.totalCubicInstanceCount(); |
| 357 | endLocations[lastResolveLevel] = cubicData + totalInstanceCount * 4; |
| 358 | #endif |
| 359 | |
| 360 | fCubicVertexCount = numTrianglesAtBeginningOfData * 4; |
| 361 | |
| 362 | if (resolveLevelCounter.totalCubicInstanceCount()) { |
| 363 | GrVectorXform xform(fViewMatrix); |
| 364 | for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) { |
| 365 | int level; |
| 366 | switch (verb) { |
| 367 | default: |
| 368 | continue; |
| 369 | case SkPathVerb::kQuad: |
| 370 | level = GrWangsFormula::quadratic_log2(kLinearizationIntolerance, pts, xform); |
| 371 | if (level == 0) { |
| 372 | continue; |
| 373 | } |
| 374 | level = std::min(level, kMaxResolveLevel); |
| 375 | quad2cubic(pts, instanceLocations[level]); |
| 376 | break; |
| 377 | case SkPathVerb::kCubic: |
| 378 | level = GrWangsFormula::cubic_log2(kLinearizationIntolerance, pts, xform); |
| 379 | if (level == 0) { |
| 380 | continue; |
| 381 | } |
| 382 | level = std::min(level, kMaxResolveLevel); |
| 383 | memcpy(instanceLocations[level], pts, sizeof(SkPoint) * 4); |
| 384 | break; |
| 385 | } |
| 386 | instanceLocations[level] += 4; |
| 387 | fCubicVertexCount += 4; |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | #ifdef SK_DEBUG |
| 392 | for (int i = 1; i <= kMaxResolveLevel; ++i) { |
| 393 | SkASSERT(instanceLocations[i] == endLocations[i]); |
| 394 | } |
| 395 | SkASSERT(fCubicVertexCount == (numTrianglesAtBeginningOfData + |
| 396 | resolveLevelCounter.totalCubicInstanceCount()) * 4); |
| 397 | #endif |
| 398 | |
| 399 | fStencilCubicsShader = target->allocator()->make<GrMiddleOutCubicShader>(fViewMatrix); |
| 400 | } |
| 401 | |
| 402 | void GrPathTessellateOp::prepareTessellatedOuterCubics(GrMeshDrawOp::Target* target, |
| 403 | int numCountedCurves) { |
| 404 | SkASSERT(target->caps().shaderCaps()->tessellationSupport()); |
| 405 | SkASSERT(numCountedCurves >= 0); |
| 406 | SkASSERT(!fCubicBuffer); |
| 407 | SkASSERT(!fStencilCubicsShader); |
| 408 | |
| 409 | if (numCountedCurves == 0) { |
| 410 | return; |
| 411 | } |
| 412 | |
| 413 | auto* vertexData = static_cast<SkPoint*>(target->makeVertexSpace( |
| 414 | sizeof(SkPoint), numCountedCurves * 4, &fCubicBuffer, &fBaseCubicVertex)); |
| 415 | if (!vertexData) { |
| 416 | return; |
| 417 | } |
| 418 | fCubicVertexCount = 0; |
| 419 | |
| 420 | for (auto [verb, pts, w] : SkPathPriv::Iterate(fPath)) { |
| 421 | switch (verb) { |
| 422 | default: |
| 423 | continue; |
| 424 | case SkPathVerb::kQuad: |
| 425 | SkASSERT(fCubicVertexCount < numCountedCurves * 4); |
| 426 | quad2cubic(pts, vertexData + fCubicVertexCount); |
| 427 | break; |
| 428 | case SkPathVerb::kCubic: |
| 429 | SkASSERT(fCubicVertexCount < numCountedCurves * 4); |
| 430 | memcpy(vertexData + fCubicVertexCount, pts, sizeof(SkPoint) * 4); |
| 431 | break; |
| 432 | } |
| 433 | fCubicVertexCount += 4; |
| 434 | } |
| 435 | SkASSERT(fCubicVertexCount == numCountedCurves * 4); |
| 436 | |
| 437 | fStencilCubicsShader = target->allocator()->make<GrCubicTessellateShader>(fViewMatrix); |
| 438 | } |
| 439 | |
| 440 | void GrPathTessellateOp::prepareTessellatedCubicWedges(GrMeshDrawOp::Target* target) { |
| 441 | SkASSERT(target->caps().shaderCaps()->tessellationSupport()); |
| 442 | SkASSERT(!fCubicBuffer); |
| 443 | SkASSERT(!fStencilCubicsShader); |
| 444 | |
| 445 | // No initial moveTo, one wedge per verb, plus an implicit close at the end. |
| 446 | // Each wedge has 5 vertices. |
| 447 | int maxVertices = (fPath.countVerbs() + 1) * 5; |
| 448 | |
| 449 | GrEagerDynamicVertexAllocator vertexAlloc(target, &fCubicBuffer, &fBaseCubicVertex); |
| 450 | auto* vertexData = vertexAlloc.lock<SkPoint>(maxVertices); |
| 451 | if (!vertexData) { |
| 452 | return; |
| 453 | } |
| 454 | fCubicVertexCount = 0; |
| 455 | |
| 456 | GrMidpointContourParser parser(fPath); |
| 457 | while (parser.parseNextContour()) { |
| 458 | SkPoint midpoint = parser.currentMidpoint(); |
| 459 | SkPoint startPoint = {0, 0}; |
| 460 | SkPoint lastPoint = startPoint; |
| 461 | for (auto [verb, pts, w] : parser.currentContour()) { |
| 462 | switch (verb) { |
| 463 | case SkPathVerb::kMove: |
| 464 | startPoint = lastPoint = pts[0]; |
| 465 | continue; |
| 466 | case SkPathVerb::kClose: |
| 467 | continue; // Ignore. We can assume an implicit close at the end. |
| 468 | case SkPathVerb::kLine: |
| 469 | line2cubic(pts[0], pts[1], vertexData + fCubicVertexCount); |
| 470 | lastPoint = pts[1]; |
| 471 | break; |
| 472 | case SkPathVerb::kQuad: |
| 473 | quad2cubic(pts, vertexData + fCubicVertexCount); |
| 474 | lastPoint = pts[2]; |
| 475 | break; |
| 476 | case SkPathVerb::kCubic: |
| 477 | memcpy(vertexData + fCubicVertexCount, pts, sizeof(SkPoint) * 4); |
| 478 | lastPoint = pts[3]; |
| 479 | break; |
| 480 | case SkPathVerb::kConic: |
| 481 | SkUNREACHABLE; |
| 482 | } |
| 483 | vertexData[fCubicVertexCount + 4] = midpoint; |
| 484 | fCubicVertexCount += 5; |
| 485 | } |
| 486 | if (lastPoint != startPoint) { |
| 487 | line2cubic(lastPoint, startPoint, vertexData + fCubicVertexCount); |
| 488 | vertexData[fCubicVertexCount + 4] = midpoint; |
| 489 | fCubicVertexCount += 5; |
| 490 | } |
| 491 | } |
| 492 | |
| 493 | vertexAlloc.unlock(fCubicVertexCount); |
| 494 | |
| 495 | if (fCubicVertexCount) { |
| 496 | fStencilCubicsShader = target->allocator()->make<GrWedgeTessellateShader>(fViewMatrix); |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | void GrPathTessellateOp::onExecute(GrOpFlushState* flushState, const SkRect& chainBounds) { |
| 501 | this->drawStencilPass(flushState); |
| 502 | if (!(OpFlags::kStencilOnly & fOpFlags)) { |
| 503 | this->drawCoverPass(flushState); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | void GrPathTessellateOp::drawStencilPass(GrOpFlushState* flushState) { |
| 508 | // Increments clockwise triangles and decrements counterclockwise. Used for "winding" fill. |
| 509 | constexpr static GrUserStencilSettings kIncrDecrStencil( |
| 510 | GrUserStencilSettings::StaticInitSeparate< |
| 511 | 0x0000, 0x0000, |
| 512 | GrUserStencilTest::kAlwaysIfInClip, GrUserStencilTest::kAlwaysIfInClip, |
| 513 | 0xffff, 0xffff, |
| 514 | GrUserStencilOp::kIncWrap, GrUserStencilOp::kDecWrap, |
| 515 | GrUserStencilOp::kKeep, GrUserStencilOp::kKeep, |
| 516 | 0xffff, 0xffff>()); |
| 517 | |
| 518 | // Inverts the bottom stencil bit. Used for "even/odd" fill. |
| 519 | constexpr static GrUserStencilSettings kInvertStencil( |
| 520 | GrUserStencilSettings::StaticInit< |
| 521 | 0x0000, |
| 522 | GrUserStencilTest::kAlwaysIfInClip, |
| 523 | 0xffff, |
| 524 | GrUserStencilOp::kInvert, |
| 525 | GrUserStencilOp::kKeep, |
| 526 | 0x0001>()); |
| 527 | |
| 528 | GrPipeline::InitArgs initArgs; |
| 529 | if (GrAAType::kNone != fAAType) { |
| 530 | initArgs.fInputFlags |= GrPipeline::InputFlags::kHWAntialias; |
| 531 | } |
| 532 | if (flushState->caps().wireframeSupport() && (OpFlags::kWireframe & fOpFlags)) { |
| 533 | initArgs.fInputFlags |= GrPipeline::InputFlags::kWireframe; |
| 534 | } |
| 535 | SkASSERT(SkPathFillType::kWinding == fPath.getFillType() || |
| 536 | SkPathFillType::kEvenOdd == fPath.getFillType()); |
| 537 | initArgs.fUserStencil = (SkPathFillType::kWinding == fPath.getFillType()) ? |
| 538 | &kIncrDecrStencil : &kInvertStencil; |
| 539 | initArgs.fCaps = &flushState->caps(); |
| 540 | GrPipeline pipeline(initArgs, GrDisableColorXPFactory::MakeXferProcessor(), |
| 541 | flushState->appliedHardClip()); |
| 542 | |
| 543 | if (fDoStencilTriangleBuffer) { |
| 544 | SkASSERT(fTriangleBuffer); |
| 545 | GrStencilTriangleShader stencilTriangleShader(fViewMatrix); |
| 546 | GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline, |
| 547 | &stencilTriangleShader); |
| 548 | flushState->bindPipelineAndScissorClip(programInfo, this->bounds()); |
| 549 | flushState->bindBuffers(nullptr, nullptr, fTriangleBuffer); |
| 550 | flushState->draw(fTriangleVertexCount, fBaseTriangleVertex); |
| 551 | } |
| 552 | |
| 553 | if (fStencilCubicsShader) { |
| 554 | SkASSERT(fCubicBuffer); |
| 555 | GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline, |
| 556 | fStencilCubicsShader); |
| 557 | flushState->bindPipelineAndScissorClip(programInfo, this->bounds()); |
| 558 | if (fIndirectDrawBuffer) { |
| 559 | SkASSERT(fIndirectIndexBuffer); |
| 560 | flushState->bindBuffers(fIndirectIndexBuffer, fCubicBuffer, nullptr); |
| 561 | flushState->drawIndexedIndirect(fIndirectDrawBuffer.get(), fIndirectDrawOffset, |
| 562 | fIndirectDrawCount); |
| 563 | } else { |
| 564 | flushState->bindBuffers(nullptr, nullptr, fCubicBuffer); |
| 565 | flushState->draw(fCubicVertexCount, fBaseCubicVertex); |
| 566 | if (flushState->caps().requiresManualFBBarrierAfterTessellatedStencilDraw()) { |
| 567 | flushState->gpu()->insertManualFramebufferBarrier(); // http://skbug.com/9739 |
| 568 | } |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | void GrPathTessellateOp::drawCoverPass(GrOpFlushState* flushState) { |
| 574 | // Allows non-zero stencil values to pass and write a color, and resets the stencil value back |
| 575 | // to zero; discards immediately on stencil values of zero. |
| 576 | // NOTE: It's ok to not check the clip here because the previous stencil pass only wrote to |
| 577 | // samples already inside the clip. |
| 578 | constexpr static GrUserStencilSettings kTestAndResetStencil( |
| 579 | GrUserStencilSettings::StaticInit< |
| 580 | 0x0000, |
| 581 | GrUserStencilTest::kNotEqual, |
| 582 | 0xffff, |
| 583 | GrUserStencilOp::kZero, |
| 584 | GrUserStencilOp::kKeep, |
| 585 | 0xffff>()); |
| 586 | |
| 587 | GrPipeline::InitArgs initArgs; |
| 588 | if (GrAAType::kNone != fAAType) { |
| 589 | if (flushState->proxy()->numSamples() == 1) { |
| 590 | // We are mixed sampled. We need to either enable conservative raster (preferred) or |
| 591 | // disable MSAA in order to avoid double blend artifacts. (Even if we disable MSAA for |
| 592 | // the cover geometry, the stencil test is still multisampled and will still produce |
| 593 | // smooth results.) |
| 594 | SkASSERT(GrAAType::kCoverage == fAAType); |
| 595 | if (flushState->caps().conservativeRasterSupport()) { |
| 596 | initArgs.fInputFlags |= GrPipeline::InputFlags::kHWAntialias; |
| 597 | initArgs.fInputFlags |= GrPipeline::InputFlags::kConservativeRaster; |
| 598 | } |
| 599 | } else { |
| 600 | // We are standard MSAA. Leave MSAA enabled for the cover geometry. |
| 601 | initArgs.fInputFlags |= GrPipeline::InputFlags::kHWAntialias; |
| 602 | } |
| 603 | } |
| 604 | initArgs.fCaps = &flushState->caps(); |
| 605 | initArgs.fDstProxyView = flushState->drawOpArgs().dstProxyView(); |
| 606 | initArgs.fWriteSwizzle = flushState->drawOpArgs().writeSwizzle(); |
| 607 | GrPipeline pipeline(initArgs, std::move(fProcessors), flushState->detachAppliedClip()); |
| 608 | |
| 609 | if (fDoFillTriangleBuffer) { |
| 610 | SkASSERT(fTriangleBuffer); |
| 611 | |
| 612 | // These are a twist on the standard red book stencil settings that allow us to fill the |
| 613 | // inner polygon directly to the final render target. At this point, the curves are already |
| 614 | // stencilled in. So if the stencil value is zero, then it means the path at our sample is |
| 615 | // not affected by any curves and we fill the path in directly. If the stencil value is |
| 616 | // nonzero, then we don't fill and instead continue the standard red book stencil process. |
| 617 | // |
| 618 | // NOTE: These settings are currently incompatible with a stencil clip. |
| 619 | constexpr static GrUserStencilSettings kFillOrIncrDecrStencil( |
| 620 | GrUserStencilSettings::StaticInitSeparate< |
| 621 | 0x0000, 0x0000, |
| 622 | GrUserStencilTest::kEqual, GrUserStencilTest::kEqual, |
| 623 | 0xffff, 0xffff, |
| 624 | GrUserStencilOp::kKeep, GrUserStencilOp::kKeep, |
| 625 | GrUserStencilOp::kIncWrap, GrUserStencilOp::kDecWrap, |
| 626 | 0xffff, 0xffff>()); |
| 627 | |
| 628 | constexpr static GrUserStencilSettings kFillOrInvertStencil( |
| 629 | GrUserStencilSettings::StaticInit< |
| 630 | 0x0000, |
| 631 | GrUserStencilTest::kEqual, |
| 632 | 0xffff, |
| 633 | GrUserStencilOp::kKeep, |
| 634 | GrUserStencilOp::kZero, |
| 635 | 0xffff>()); |
| 636 | |
| 637 | if (fDoStencilTriangleBuffer) { |
| 638 | // The path was already stencilled. Here we just need to do a cover pass. |
| 639 | pipeline.setUserStencil(&kTestAndResetStencil); |
| 640 | } else if (!fStencilCubicsShader) { |
| 641 | // There are no stencilled curves. We can ignore stencil and fill the path directly. |
| 642 | pipeline.setUserStencil(&GrUserStencilSettings::kUnused); |
| 643 | } else if (SkPathFillType::kWinding == fPath.getFillType()) { |
| 644 | // Fill in the path pixels not touched by curves, incr/decr stencil otherwise. |
| 645 | SkASSERT(!pipeline.hasStencilClip()); |
| 646 | pipeline.setUserStencil(&kFillOrIncrDecrStencil); |
| 647 | } else { |
| 648 | // Fill in the path pixels not touched by curves, invert stencil otherwise. |
| 649 | SkASSERT(!pipeline.hasStencilClip()); |
| 650 | pipeline.setUserStencil(&kFillOrInvertStencil); |
| 651 | } |
| 652 | |
| 653 | GrFillTriangleShader fillTriangleShader(fViewMatrix, fColor); |
| 654 | GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline, |
| 655 | &fillTriangleShader); |
| 656 | flushState->bindPipelineAndScissorClip(programInfo, this->bounds()); |
| 657 | flushState->bindTextures(fillTriangleShader, nullptr, pipeline); |
| 658 | flushState->bindBuffers(nullptr, nullptr, fTriangleBuffer); |
| 659 | flushState->draw(fTriangleVertexCount, fBaseTriangleVertex); |
| 660 | |
| 661 | if (fStencilCubicsShader) { |
| 662 | SkASSERT(fCubicBuffer); |
| 663 | |
| 664 | // At this point, every pixel is filled in except the ones touched by curves. Issue a |
| 665 | // final cover pass over the curves by drawing their convex hulls. This will fill in any |
| 666 | // remaining samples and reset the stencil buffer. |
| 667 | pipeline.setUserStencil(&kTestAndResetStencil); |
| 668 | GrFillCubicHullShader fillCubicHullShader(fViewMatrix, fColor); |
| 669 | GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline, |
| 670 | &fillCubicHullShader); |
| 671 | flushState->bindPipelineAndScissorClip(programInfo, this->bounds()); |
| 672 | flushState->bindTextures(fillCubicHullShader, nullptr, pipeline); |
| 673 | |
| 674 | // Here we treat fCubicBuffer as an instance buffer. It should have been prepared with |
| 675 | // the base vertex on an instance boundary in order to accommodate this. |
| 676 | SkASSERT((fCubicVertexCount % 4) == 0); |
| 677 | SkASSERT((fBaseCubicVertex % 4) == 0); |
| 678 | flushState->bindBuffers(nullptr, fCubicBuffer, nullptr); |
| 679 | flushState->drawInstanced(fCubicVertexCount >> 2, fBaseCubicVertex >> 2, 4, 0); |
| 680 | } |
| 681 | return; |
| 682 | } |
| 683 | |
| 684 | // There are no triangles to fill. Just draw a bounding box. |
| 685 | pipeline.setUserStencil(&kTestAndResetStencil); |
| 686 | GrFillBoundingBoxShader fillBoundingBoxShader(fViewMatrix, fColor, fPath.getBounds()); |
| 687 | GrPathShader::ProgramInfo programInfo(flushState->writeView(), &pipeline, |
| 688 | &fillBoundingBoxShader); |
| 689 | flushState->bindPipelineAndScissorClip(programInfo, this->bounds()); |
| 690 | flushState->bindTextures(fillBoundingBoxShader, nullptr, pipeline); |
| 691 | flushState->bindBuffers(nullptr, nullptr, nullptr); |
| 692 | flushState->draw(4, 0); |
| 693 | } |
| 694 | |