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 <new> |
9 | |
10 | #include "include/core/SkPoint.h" |
11 | #include "include/core/SkPoint3.h" |
12 | #include "include/private/GrRecordingContext.h" |
13 | #include "include/private/SkFloatingPoint.h" |
14 | #include "include/private/SkTo.h" |
15 | #include "src/core/SkMathPriv.h" |
16 | #include "src/core/SkMatrixPriv.h" |
17 | #include "src/core/SkRectPriv.h" |
18 | #include "src/gpu/GrAppliedClip.h" |
19 | #include "src/gpu/GrCaps.h" |
20 | #include "src/gpu/GrDrawOpTest.h" |
21 | #include "src/gpu/GrGeometryProcessor.h" |
22 | #include "src/gpu/GrGpu.h" |
23 | #include "src/gpu/GrMemoryPool.h" |
24 | #include "src/gpu/GrOpFlushState.h" |
25 | #include "src/gpu/GrRecordingContextPriv.h" |
26 | #include "src/gpu/GrResourceProvider.h" |
27 | #include "src/gpu/GrResourceProviderPriv.h" |
28 | #include "src/gpu/GrShaderCaps.h" |
29 | #include "src/gpu/GrTexture.h" |
30 | #include "src/gpu/GrTexturePriv.h" |
31 | #include "src/gpu/GrTextureProxy.h" |
32 | #include "src/gpu/SkGr.h" |
33 | #include "src/gpu/effects/generated/GrClampFragmentProcessor.h" |
34 | #include "src/gpu/geometry/GrQuad.h" |
35 | #include "src/gpu/geometry/GrQuadBuffer.h" |
36 | #include "src/gpu/geometry/GrQuadUtils.h" |
37 | #include "src/gpu/glsl/GrGLSLVarying.h" |
38 | #include "src/gpu/ops/GrFillRectOp.h" |
39 | #include "src/gpu/ops/GrMeshDrawOp.h" |
40 | #include "src/gpu/ops/GrQuadPerEdgeAA.h" |
41 | #include "src/gpu/ops/GrSimpleMeshDrawOpHelper.h" |
42 | #include "src/gpu/ops/GrTextureOp.h" |
43 | |
44 | namespace { |
45 | |
46 | using Domain = GrQuadPerEdgeAA::Domain; |
47 | using VertexSpec = GrQuadPerEdgeAA::VertexSpec; |
48 | using ColorType = GrQuadPerEdgeAA::ColorType; |
49 | |
50 | // Extracts lengths of vertical and horizontal edges of axis-aligned quad. "width" is the edge |
51 | // between v0 and v2 (or v1 and v3), "height" is the edge between v0 and v1 (or v2 and v3). |
52 | static SkSize axis_aligned_quad_size(const GrQuad& quad) { |
53 | SkASSERT(quad.quadType() == GrQuad::Type::kAxisAligned); |
54 | // Simplification of regular edge length equation, since it's axis aligned and can avoid sqrt |
55 | float dw = sk_float_abs(quad.x(2) - quad.x(0)) + sk_float_abs(quad.y(2) - quad.y(0)); |
56 | float dh = sk_float_abs(quad.x(1) - quad.x(0)) + sk_float_abs(quad.y(1) - quad.y(0)); |
57 | return {dw, dh}; |
58 | } |
59 | |
60 | static bool filter_has_effect(const GrQuad& srcQuad, const GrQuad& dstQuad) { |
61 | // If not axis-aligned in src or dst, then always say it has an effect |
62 | if (srcQuad.quadType() != GrQuad::Type::kAxisAligned || |
63 | dstQuad.quadType() != GrQuad::Type::kAxisAligned) { |
64 | return true; |
65 | } |
66 | |
67 | SkRect srcRect; |
68 | SkRect dstRect; |
69 | if (srcQuad.asRect(&srcRect) && dstQuad.asRect(&dstRect)) { |
70 | // Disable filtering when there is no scaling (width and height are the same), and the |
71 | // top-left corners have the same fraction (so src and dst snap to the pixel grid |
72 | // identically). |
73 | SkASSERT(srcRect.isSorted()); |
74 | return srcRect.width() != dstRect.width() || srcRect.height() != dstRect.height() || |
75 | SkScalarFraction(srcRect.fLeft) != SkScalarFraction(dstRect.fLeft) || |
76 | SkScalarFraction(srcRect.fTop) != SkScalarFraction(dstRect.fTop); |
77 | } else { |
78 | // Although the quads are axis-aligned, the local coordinate system is transformed such |
79 | // that fractionally-aligned sample centers will not align with the device coordinate system |
80 | // So disable filtering when edges are the same length and both srcQuad and dstQuad |
81 | // 0th vertex is integer aligned. |
82 | if (SkScalarIsInt(srcQuad.x(0)) && SkScalarIsInt(srcQuad.y(0)) && |
83 | SkScalarIsInt(dstQuad.x(0)) && SkScalarIsInt(dstQuad.y(0))) { |
84 | // Extract edge lengths |
85 | SkSize srcSize = axis_aligned_quad_size(srcQuad); |
86 | SkSize dstSize = axis_aligned_quad_size(dstQuad); |
87 | return srcSize.fWidth != dstSize.fWidth || srcSize.fHeight != dstSize.fHeight; |
88 | } else { |
89 | return true; |
90 | } |
91 | } |
92 | } |
93 | |
94 | // Describes function for normalizing src coords: [x * iw, y * ih + yOffset] can represent |
95 | // regular and rectangular textures, w/ or w/o origin correction. |
96 | struct NormalizationParams { |
97 | float fIW; // 1 / width of texture, or 1.0 for texture rectangles |
98 | float fIH; // 1 / height of texture, or 1.0 for tex rects, X -1 if bottom-left origin |
99 | float fYOffset; // 0 for top-left origin, height of [normalized] tex if bottom-left |
100 | }; |
101 | static NormalizationParams proxy_normalization_params(const GrSurfaceProxy* proxy, |
102 | GrSurfaceOrigin origin) { |
103 | // Whether or not the proxy is instantiated, this is the size its texture will be, so we can |
104 | // normalize the src coordinates up front. |
105 | SkISize dimensions = proxy->backingStoreDimensions(); |
106 | float iw, ih, h; |
107 | if (proxy->backendFormat().textureType() == GrTextureType::kRectangle) { |
108 | iw = ih = 1.f; |
109 | h = dimensions.height(); |
110 | } else { |
111 | iw = 1.f / dimensions.width(); |
112 | ih = 1.f / dimensions.height(); |
113 | h = 1.f; |
114 | } |
115 | |
116 | if (origin == kBottomLeft_GrSurfaceOrigin) { |
117 | return {iw, -ih, h}; |
118 | } else { |
119 | return {iw, ih, 0.0f}; |
120 | } |
121 | } |
122 | |
123 | static SkRect inset_domain_for_bilerp(const NormalizationParams& params, const SkRect& domainRect) { |
124 | // Normalized pixel size is also equal to iw and ih, so the insets for bilerp are just |
125 | // in those units and can be applied safely after normalization. However, if the domain is |
126 | // smaller than a texel, it should clamp to the center of that axis. |
127 | float dw = domainRect.width() < params.fIW ? domainRect.width() : params.fIW; |
128 | float dh = domainRect.height() < params.fIH ? domainRect.height() : params.fIH; |
129 | return domainRect.makeInset(0.5f * dw, 0.5f * dh); |
130 | } |
131 | |
132 | // Normalize the domain. If 'domainRect' is null, it is assumed no domain constraint is desired, |
133 | // so a sufficiently large rect is returned even if the quad ends up batched with an op that uses |
134 | // domains overall. |
135 | static SkRect normalize_domain(GrSamplerState::Filter filter, |
136 | const NormalizationParams& params, |
137 | const SkRect* domainRect) { |
138 | static constexpr SkRect kLargeRect = {-100000, -100000, 1000000, 1000000}; |
139 | if (!domainRect) { |
140 | // Either the quad has no domain constraint and is batched with a domain constrained op |
141 | // (in which case we want a domain that doesn't restrict normalized tex coords), or the |
142 | // entire op doesn't use the domain, in which case the returned value is ignored. |
143 | return kLargeRect; |
144 | } |
145 | |
146 | auto ltrb = skvx::Vec<4, float>::Load(domainRect); |
147 | // Normalize and offset |
148 | ltrb = mad(ltrb, {params.fIW, params.fIH, params.fIW, params.fIH}, |
149 | {0.f, params.fYOffset, 0.f, params.fYOffset}); |
150 | if (params.fIH < 0.f) { |
151 | // Flip top and bottom to keep the rect sorted when loaded back to SkRect. |
152 | ltrb = skvx::shuffle<0, 3, 2, 1>(ltrb); |
153 | } |
154 | |
155 | SkRect out; |
156 | ltrb.store(&out); |
157 | return out; |
158 | } |
159 | |
160 | // Normalizes logical src coords and corrects for origin |
161 | static void normalize_src_quad(const NormalizationParams& params, |
162 | GrQuad* srcQuad) { |
163 | // The src quad should not have any perspective |
164 | SkASSERT(!srcQuad->hasPerspective()); |
165 | skvx::Vec<4, float> xs = srcQuad->x4f() * params.fIW; |
166 | skvx::Vec<4, float> ys = mad(srcQuad->y4f(), params.fIH, params.fYOffset); |
167 | xs.store(srcQuad->xs()); |
168 | ys.store(srcQuad->ys()); |
169 | } |
170 | |
171 | // Count the number of proxy runs in the entry set. This usually is already computed by |
172 | // SkGpuDevice, but when the BatchLengthLimiter chops the set up it must determine a new proxy count |
173 | // for each split. |
174 | static int proxy_run_count(const GrRenderTargetContext::TextureSetEntry set[], int count) { |
175 | int actualProxyRunCount = 0; |
176 | const GrSurfaceProxy* lastProxy = nullptr; |
177 | for (int i = 0; i < count; ++i) { |
178 | if (set[i].fProxyView.proxy() != lastProxy) { |
179 | actualProxyRunCount++; |
180 | lastProxy = set[i].fProxyView.proxy(); |
181 | } |
182 | } |
183 | return actualProxyRunCount; |
184 | } |
185 | |
186 | /** |
187 | * Op that implements GrTextureOp::Make. It draws textured quads. Each quad can modulate against a |
188 | * the texture by color. The blend with the destination is always src-over. The edges are non-AA. |
189 | */ |
190 | class TextureOp final : public GrMeshDrawOp { |
191 | public: |
192 | static std::unique_ptr<GrDrawOp> Make(GrRecordingContext* context, |
193 | GrSurfaceProxyView proxyView, |
194 | sk_sp<GrColorSpaceXform> textureXform, |
195 | GrSamplerState::Filter filter, |
196 | const SkPMColor4f& color, |
197 | GrTextureOp::Saturate saturate, |
198 | GrAAType aaType, |
199 | DrawQuad* quad, |
200 | const SkRect* domain) { |
201 | GrOpMemoryPool* pool = context->priv().opMemoryPool(); |
202 | return pool->allocate<TextureOp>(std::move(proxyView), std::move(textureXform), filter, |
203 | color, saturate, aaType, quad, domain); |
204 | } |
205 | |
206 | static std::unique_ptr<GrDrawOp> Make(GrRecordingContext* context, |
207 | GrRenderTargetContext::TextureSetEntry set[], |
208 | int cnt, |
209 | int proxyRunCnt, |
210 | GrSamplerState::Filter filter, |
211 | GrTextureOp::Saturate saturate, |
212 | GrAAType aaType, |
213 | SkCanvas::SrcRectConstraint constraint, |
214 | const SkMatrix& viewMatrix, |
215 | sk_sp<GrColorSpaceXform> textureColorSpaceXform) { |
216 | // Allocate size based on proxyRunCnt, since that determines number of ViewCountPairs. |
217 | SkASSERT(proxyRunCnt <= cnt); |
218 | |
219 | size_t size = sizeof(TextureOp) + sizeof(ViewCountPair) * (proxyRunCnt - 1); |
220 | GrOpMemoryPool* pool = context->priv().opMemoryPool(); |
221 | void* mem = pool->allocate(size); |
222 | return std::unique_ptr<GrDrawOp>( |
223 | new (mem) TextureOp(set, cnt, proxyRunCnt, filter, saturate, aaType, constraint, |
224 | viewMatrix, std::move(textureColorSpaceXform))); |
225 | } |
226 | |
227 | ~TextureOp() override { |
228 | for (unsigned p = 1; p < fMetadata.fProxyCount; ++p) { |
229 | fViewCountPairs[p].~ViewCountPair(); |
230 | } |
231 | } |
232 | |
233 | const char* name() const override { return "TextureOp" ; } |
234 | |
235 | void visitProxies(const VisitProxyFunc& func) const override { |
236 | bool mipped = (GrSamplerState::Filter::kMipMap == fMetadata.filter()); |
237 | for (unsigned p = 0; p < fMetadata.fProxyCount; ++p) { |
238 | func(fViewCountPairs[p].fProxy.get(), GrMipMapped(mipped)); |
239 | } |
240 | if (fDesc && fDesc->fProgramInfo) { |
241 | fDesc->fProgramInfo->visitFPProxies(func); |
242 | } |
243 | } |
244 | |
245 | #ifdef SK_DEBUG |
246 | SkString dumpInfo() const override { |
247 | SkString str; |
248 | str.appendf("# draws: %d\n" , fQuads.count()); |
249 | auto iter = fQuads.iterator(); |
250 | for (unsigned p = 0; p < fMetadata.fProxyCount; ++p) { |
251 | str.appendf("Proxy ID: %d, Filter: %d\n" , |
252 | fViewCountPairs[p].fProxy->uniqueID().asUInt(), |
253 | static_cast<int>(fMetadata.fFilter)); |
254 | int i = 0; |
255 | while(i < fViewCountPairs[p].fQuadCnt && iter.next()) { |
256 | const GrQuad* quad = iter.deviceQuad(); |
257 | GrQuad uv = iter.isLocalValid() ? *(iter.localQuad()) : GrQuad(); |
258 | const ColorDomainAndAA& info = iter.metadata(); |
259 | str.appendf( |
260 | "%d: Color: 0x%08x, Domain(%d): [L: %.2f, T: %.2f, R: %.2f, B: %.2f]\n" |
261 | " UVs [(%.2f, %.2f), (%.2f, %.2f), (%.2f, %.2f), (%.2f, %.2f)]\n" |
262 | " Quad [(%.2f, %.2f), (%.2f, %.2f), (%.2f, %.2f), (%.2f, %.2f)]\n" , |
263 | i, info.fColor.toBytes_RGBA(), fMetadata.fDomain, info.fDomainRect.fLeft, |
264 | info.fDomainRect.fTop, info.fDomainRect.fRight, info.fDomainRect.fBottom, |
265 | quad->point(0).fX, quad->point(0).fY, quad->point(1).fX, quad->point(1).fY, |
266 | quad->point(2).fX, quad->point(2).fY, quad->point(3).fX, quad->point(3).fY, |
267 | uv.point(0).fX, uv.point(0).fY, uv.point(1).fX, uv.point(1).fY, |
268 | uv.point(2).fX, uv.point(2).fY, uv.point(3).fX, uv.point(3).fY); |
269 | |
270 | i++; |
271 | } |
272 | } |
273 | str += INHERITED::dumpInfo(); |
274 | return str; |
275 | } |
276 | |
277 | static void ValidateResourceLimits() { |
278 | // The op implementation has an upper bound on the number of quads that it can represent. |
279 | // However, the resource manager imposes its own limit on the number of quads, which should |
280 | // always be lower than the numerical limit this op can hold. |
281 | using CountStorage = decltype(Metadata::fTotalQuadCount); |
282 | CountStorage maxQuadCount = std::numeric_limits<CountStorage>::max(); |
283 | // GrResourceProvider::Max...() is typed as int, so don't compare across signed/unsigned. |
284 | int resourceLimit = SkTo<int>(maxQuadCount); |
285 | SkASSERT(GrResourceProvider::MaxNumAAQuads() <= resourceLimit && |
286 | GrResourceProvider::MaxNumNonAAQuads() <= resourceLimit); |
287 | } |
288 | #endif |
289 | |
290 | GrProcessorSet::Analysis finalize( |
291 | const GrCaps& caps, const GrAppliedClip*, bool hasMixedSampledCoverage, |
292 | GrClampType clampType) override { |
293 | SkASSERT(fMetadata.colorType() == ColorType::kNone); |
294 | auto iter = fQuads.metadata(); |
295 | while(iter.next()) { |
296 | auto colorType = GrQuadPerEdgeAA::MinColorType(iter->fColor); |
297 | fMetadata.fColorType = std::max(fMetadata.fColorType, static_cast<uint16_t>(colorType)); |
298 | } |
299 | return GrProcessorSet::EmptySetAnalysis(); |
300 | } |
301 | |
302 | FixedFunctionFlags fixedFunctionFlags() const override { |
303 | return fMetadata.aaType() == GrAAType::kMSAA ? FixedFunctionFlags::kUsesHWAA |
304 | : FixedFunctionFlags::kNone; |
305 | } |
306 | |
307 | DEFINE_OP_CLASS_ID |
308 | |
309 | private: |
310 | friend class ::GrOpMemoryPool; |
311 | |
312 | struct ColorDomainAndAA { |
313 | ColorDomainAndAA(const SkPMColor4f& color, const SkRect& domainRect, GrQuadAAFlags aaFlags) |
314 | : fColor(color) |
315 | , fDomainRect(domainRect) |
316 | , fAAFlags(static_cast<uint16_t>(aaFlags)) { |
317 | SkASSERT(fAAFlags == static_cast<uint16_t>(aaFlags)); |
318 | } |
319 | |
320 | SkPMColor4f fColor; |
321 | // If the op doesn't use domains, this is ignored. If the op uses domains and the specific |
322 | // entry does not, this rect will equal kLargeRect, so it automatically has no effect. |
323 | SkRect fDomainRect; |
324 | unsigned fAAFlags : 4; |
325 | |
326 | GrQuadAAFlags aaFlags() const { return static_cast<GrQuadAAFlags>(fAAFlags); } |
327 | }; |
328 | |
329 | struct ViewCountPair { |
330 | // Normally this would be a GrSurfaceProxyView, but GrTextureOp applies the GrOrigin right |
331 | // away so it doesn't need to be stored, and all ViewCountPairs in an op have the same |
332 | // swizzle so that is stored in the op metadata. |
333 | sk_sp<GrSurfaceProxy> fProxy; |
334 | int fQuadCnt; |
335 | }; |
336 | |
337 | // TextureOp and ViewCountPair are 8 byte aligned. This is packed into 8 bytes to minimally |
338 | // increase the size of the op; increasing the op size can have a surprising impact on |
339 | // performance (since texture ops are one of the most commonly used in an app). |
340 | struct Metadata { |
341 | // AAType must be filled after initialization; ColorType is determined in finalize() |
342 | Metadata(const GrSwizzle& swizzle, GrSamplerState::Filter filter, |
343 | GrQuadPerEdgeAA::Domain domain, GrTextureOp::Saturate saturate) |
344 | : fSwizzle(swizzle) |
345 | , fProxyCount(1) |
346 | , fTotalQuadCount(1) |
347 | , fFilter(static_cast<uint16_t>(filter)) |
348 | , fAAType(static_cast<uint16_t>(GrAAType::kNone)) |
349 | , fColorType(static_cast<uint16_t>(ColorType::kNone)) |
350 | , fDomain(static_cast<uint16_t>(domain)) |
351 | , fSaturate(static_cast<uint16_t>(saturate)) {} |
352 | |
353 | GrSwizzle fSwizzle; // sizeof(GrSwizzle) == uint16_t |
354 | uint16_t fProxyCount; |
355 | // This will be >= fProxyCount, since a proxy may be drawn multiple times |
356 | uint16_t fTotalQuadCount; |
357 | |
358 | // These must be based on uint16_t to help MSVC's pack bitfields optimally |
359 | uint16_t fFilter : 2; // GrSamplerState::Filter |
360 | uint16_t fAAType : 2; // GrAAType |
361 | uint16_t fColorType : 2; // GrQuadPerEdgeAA::ColorType |
362 | uint16_t fDomain : 1; // bool |
363 | uint16_t fSaturate : 1; // bool |
364 | uint16_t fUnused : 8; // # of bits left before Metadata exceeds 8 bytes |
365 | |
366 | GrSamplerState::Filter filter() const { |
367 | return static_cast<GrSamplerState::Filter>(fFilter); |
368 | } |
369 | GrAAType aaType() const { return static_cast<GrAAType>(fAAType); } |
370 | ColorType colorType() const { return static_cast<ColorType>(fColorType); } |
371 | Domain domain() const { return static_cast<Domain>(fDomain); } |
372 | GrTextureOp::Saturate saturate() const { |
373 | return static_cast<GrTextureOp::Saturate>(fSaturate); |
374 | } |
375 | |
376 | static_assert(GrSamplerState::kFilterCount <= 4); |
377 | static_assert(kGrAATypeCount <= 4); |
378 | static_assert(GrQuadPerEdgeAA::kColorTypeCount <= 4); |
379 | }; |
380 | static_assert(sizeof(Metadata) == 8); |
381 | |
382 | // This descriptor is used to store the draw info we decide on during on(Pre)PrepareDraws. We |
383 | // store the data in a separate struct in order to minimize the size of the TextureOp. |
384 | // Historically, increasing the TextureOp's size has caused surprising perf regressions, but we |
385 | // may want to re-evaluate whether this is still necessary. |
386 | // |
387 | // In the onPrePrepareDraws case it is allocated in the creation-time opData arena, and |
388 | // allocatePrePreparedVertices is also called. |
389 | // |
390 | // In the onPrepareDraws case this descriptor is allocated in the flush-time arena (i.e., as |
391 | // part of the flushState). |
392 | struct Desc { |
393 | VertexSpec fVertexSpec; |
394 | int fNumProxies = 0; |
395 | int fNumTotalQuads = 0; |
396 | |
397 | // This member variable is only used by 'onPrePrepareDraws'. |
398 | char* fPrePreparedVertices = nullptr; |
399 | |
400 | GrProgramInfo* fProgramInfo = nullptr; |
401 | |
402 | sk_sp<const GrBuffer> fIndexBuffer; |
403 | sk_sp<const GrBuffer> fVertexBuffer; |
404 | int fBaseVertex; |
405 | |
406 | // How big should 'fVertices' be to hold all the vertex data? |
407 | size_t totalSizeInBytes() const { |
408 | return this->totalNumVertices() * fVertexSpec.vertexSize(); |
409 | } |
410 | |
411 | int totalNumVertices() const { |
412 | return fNumTotalQuads * fVertexSpec.verticesPerQuad(); |
413 | } |
414 | |
415 | void allocatePrePreparedVertices(SkArenaAlloc* arena) { |
416 | fPrePreparedVertices = arena->makeArrayDefault<char>(this->totalSizeInBytes()); |
417 | } |
418 | |
419 | }; |
420 | |
421 | // If domainRect is not null it will be used to apply a strict src rect-style constraint. |
422 | TextureOp(GrSurfaceProxyView proxyView, |
423 | sk_sp<GrColorSpaceXform> textureColorSpaceXform, |
424 | GrSamplerState::Filter filter, |
425 | const SkPMColor4f& color, |
426 | GrTextureOp::Saturate saturate, |
427 | GrAAType aaType, |
428 | DrawQuad* quad, |
429 | const SkRect* domainRect) |
430 | : INHERITED(ClassID()) |
431 | , fQuads(1, true /* includes locals */) |
432 | , fTextureColorSpaceXform(std::move(textureColorSpaceXform)) |
433 | , fDesc(nullptr) |
434 | , fMetadata(proxyView.swizzle(), filter, Domain(!!domainRect), saturate) { |
435 | |
436 | // Clean up disparities between the overall aa type and edge configuration and apply |
437 | // optimizations based on the rect and matrix when appropriate |
438 | GrQuadUtils::ResolveAAType(aaType, quad->fEdgeFlags, quad->fDevice, |
439 | &aaType, &quad->fEdgeFlags); |
440 | fMetadata.fAAType = static_cast<uint16_t>(aaType); |
441 | |
442 | // We expect our caller to have already caught this optimization. |
443 | SkASSERT(!domainRect || |
444 | !domainRect->contains(proxyView.proxy()->backingStoreBoundsRect())); |
445 | |
446 | // We may have had a strict constraint with nearest filter solely due to possible AA bloat. |
447 | // If we don't have (or determined we don't need) coverage AA then we can skip using a |
448 | // domain. |
449 | if (domainRect && filter == GrSamplerState::Filter::kNearest && |
450 | aaType != GrAAType::kCoverage) { |
451 | domainRect = nullptr; |
452 | fMetadata.fDomain = static_cast<uint16_t>(Domain::kNo); |
453 | } |
454 | |
455 | // Normalize src coordinates and the domain (if set) |
456 | NormalizationParams params = proxy_normalization_params(proxyView.proxy(), |
457 | proxyView.origin()); |
458 | normalize_src_quad(params, &quad->fLocal); |
459 | SkRect domain = normalize_domain(filter, params, domainRect); |
460 | |
461 | // Set bounds before clipping so we don't have to worry about unioning the bounds of |
462 | // the two potential quads (GrQuad::bounds() is perspective-safe). |
463 | this->setBounds(quad->fDevice.bounds(), HasAABloat(aaType == GrAAType::kCoverage), |
464 | IsHairline::kNo); |
465 | |
466 | int quadCount = this->appendQuad(quad, color, domain); |
467 | fViewCountPairs[0] = {proxyView.detachProxy(), quadCount}; |
468 | } |
469 | |
470 | TextureOp(GrRenderTargetContext::TextureSetEntry set[], |
471 | int cnt, |
472 | int proxyRunCnt, |
473 | GrSamplerState::Filter filter, |
474 | GrTextureOp::Saturate saturate, |
475 | GrAAType aaType, |
476 | SkCanvas::SrcRectConstraint constraint, |
477 | const SkMatrix& viewMatrix, |
478 | sk_sp<GrColorSpaceXform> textureColorSpaceXform) |
479 | : INHERITED(ClassID()) |
480 | , fQuads(cnt, true /* includes locals */) |
481 | , fTextureColorSpaceXform(std::move(textureColorSpaceXform)) |
482 | , fDesc(nullptr) |
483 | , fMetadata(set[0].fProxyView.swizzle(), GrSamplerState::Filter::kNearest, |
484 | Domain::kNo, saturate) { |
485 | // Update counts to reflect the batch op |
486 | fMetadata.fProxyCount = SkToUInt(proxyRunCnt); |
487 | fMetadata.fTotalQuadCount = SkToUInt(cnt); |
488 | |
489 | SkRect bounds = SkRectPriv::MakeLargestInverted(); |
490 | |
491 | GrAAType netAAType = GrAAType::kNone; // aa type maximally compatible with all dst rects |
492 | Domain netDomain = Domain::kNo; |
493 | GrSamplerState::Filter netFilter = GrSamplerState::Filter::kNearest; |
494 | |
495 | const GrSurfaceProxy* curProxy = nullptr; |
496 | |
497 | // 'q' is the index in 'set' and fQuadBuffer; 'p' is the index in fViewCountPairs and only |
498 | // increases when set[q]'s proxy changes. |
499 | int p = 0; |
500 | for (int q = 0; q < cnt; ++q) { |
501 | if (q == 0) { |
502 | // We do not placement new the first ViewCountPair since that one is allocated and |
503 | // initialized as part of the GrTextureOp creation. |
504 | fViewCountPairs[0].fProxy = set[0].fProxyView.detachProxy(); |
505 | fViewCountPairs[0].fQuadCnt = 0; |
506 | curProxy = fViewCountPairs[0].fProxy.get(); |
507 | } else if (set[q].fProxyView.proxy() != curProxy) { |
508 | // We must placement new the ViewCountPairs here so that the sk_sps in the |
509 | // GrSurfaceProxyView get initialized properly. |
510 | new(&fViewCountPairs[++p])ViewCountPair({set[q].fProxyView.detachProxy(), 0}); |
511 | |
512 | curProxy = fViewCountPairs[p].fProxy.get(); |
513 | SkASSERT(GrTextureProxy::ProxiesAreCompatibleAsDynamicState( |
514 | curProxy, fViewCountPairs[0].fProxy.get())); |
515 | SkASSERT(fMetadata.fSwizzle == set[q].fProxyView.swizzle()); |
516 | } // else another quad referencing the same proxy |
517 | |
518 | SkMatrix ctm = viewMatrix; |
519 | if (set[q].fPreViewMatrix) { |
520 | ctm.preConcat(*set[q].fPreViewMatrix); |
521 | } |
522 | |
523 | // Use dstRect/srcRect unless dstClip is provided, in which case derive new source |
524 | // coordinates by mapping dstClipQuad by the dstRect to srcRect transform. |
525 | DrawQuad quad; |
526 | if (set[q].fDstClipQuad) { |
527 | quad.fDevice = GrQuad::MakeFromSkQuad(set[q].fDstClipQuad, ctm); |
528 | |
529 | SkPoint srcPts[4]; |
530 | GrMapRectPoints(set[q].fDstRect, set[q].fSrcRect, set[q].fDstClipQuad, srcPts, 4); |
531 | quad.fLocal = GrQuad::MakeFromSkQuad(srcPts, SkMatrix::I()); |
532 | } else { |
533 | quad.fDevice = GrQuad::MakeFromRect(set[q].fDstRect, ctm); |
534 | quad.fLocal = GrQuad(set[q].fSrcRect); |
535 | } |
536 | |
537 | if (netFilter != filter && filter_has_effect(quad.fLocal, quad.fDevice)) { |
538 | // The only way netFilter != filter is if bilerp is requested and we haven't yet |
539 | // found a quad that requires bilerp (so net is still nearest). |
540 | SkASSERT(netFilter == GrSamplerState::Filter::kNearest && |
541 | filter == GrSamplerState::Filter::kBilerp); |
542 | netFilter = GrSamplerState::Filter::kBilerp; |
543 | } |
544 | |
545 | // Normalize the src quads and apply origin |
546 | NormalizationParams proxyParams = proxy_normalization_params( |
547 | curProxy, set[q].fProxyView.origin()); |
548 | normalize_src_quad(proxyParams, &quad.fLocal); |
549 | |
550 | // Update overall bounds of the op as the union of all quads |
551 | bounds.joinPossiblyEmptyRect(quad.fDevice.bounds()); |
552 | |
553 | // Determine the AA type for the quad, then merge with net AA type |
554 | GrAAType aaForQuad; |
555 | GrQuadUtils::ResolveAAType(aaType, set[q].fAAFlags, quad.fDevice, |
556 | &aaForQuad, &quad.fEdgeFlags); |
557 | // Resolve sets aaForQuad to aaType or None, there is never a change between aa methods |
558 | SkASSERT(aaForQuad == GrAAType::kNone || aaForQuad == aaType); |
559 | if (netAAType == GrAAType::kNone && aaForQuad != GrAAType::kNone) { |
560 | netAAType = aaType; |
561 | } |
562 | |
563 | // Calculate metadata for the entry |
564 | const SkRect* domainForQuad = nullptr; |
565 | if (constraint == SkCanvas::kStrict_SrcRectConstraint) { |
566 | // Check (briefly) if the strict constraint is needed for this set entry |
567 | if (!set[q].fSrcRect.contains(curProxy->backingStoreBoundsRect()) && |
568 | (filter == GrSamplerState::Filter::kBilerp || |
569 | aaForQuad == GrAAType::kCoverage)) { |
570 | // Can't rely on hardware clamping and the draw will access outer texels |
571 | // for AA and/or bilerp. Unlike filter quality, this op still has per-quad |
572 | // control over AA so that can check aaForQuad, not netAAType. |
573 | netDomain = Domain::kYes; |
574 | domainForQuad = &set[q].fSrcRect; |
575 | } |
576 | } |
577 | // This domain may represent a no-op, otherwise it will have the origin and dimensions |
578 | // of the texture applied to it. Insetting for bilinear filtering is deferred until |
579 | // on[Pre]Prepare so that the overall filter can be lazily determined. |
580 | SkRect domain = normalize_domain(filter, proxyParams, domainForQuad); |
581 | |
582 | // Always append a quad (or 2 if perspective clipped), it just may refer back to a prior |
583 | // ViewCountPair (this frequently happens when Chrome draws 9-patches). |
584 | float alpha = SkTPin(set[q].fAlpha, 0.f, 1.f); |
585 | fViewCountPairs[p].fQuadCnt += this->appendQuad( |
586 | &quad, {alpha, alpha, alpha, alpha}, domain); |
587 | } |
588 | // The # of proxy switches should match what was provided (+1 because we incremented p |
589 | // when a new proxy was encountered). |
590 | SkASSERT((p + 1) == fMetadata.fProxyCount); |
591 | SkASSERT(fQuads.count() == fMetadata.fTotalQuadCount); |
592 | |
593 | fMetadata.fAAType = static_cast<uint16_t>(netAAType); |
594 | fMetadata.fFilter = static_cast<uint16_t>(netFilter); |
595 | fMetadata.fDomain = static_cast<uint16_t>(netDomain); |
596 | |
597 | this->setBounds(bounds, HasAABloat(netAAType == GrAAType::kCoverage), IsHairline::kNo); |
598 | } |
599 | |
600 | int appendQuad(DrawQuad* quad, const SkPMColor4f& color, const SkRect& domain) { |
601 | DrawQuad ; |
602 | // Only clip when there's anti-aliasing. When non-aa, the GPU clips just fine and there's |
603 | // no inset/outset math that requires w > 0. |
604 | int quadCount = quad->fEdgeFlags != GrQuadAAFlags::kNone ? |
605 | GrQuadUtils::ClipToW0(quad, &extra) : 1; |
606 | if (quadCount == 0) { |
607 | // We can't discard the op at this point, but disable AA flags so it won't go through |
608 | // inset/outset processing |
609 | quad->fEdgeFlags = GrQuadAAFlags::kNone; |
610 | quadCount = 1; |
611 | } |
612 | fQuads.append(quad->fDevice, {color, domain, quad->fEdgeFlags}, &quad->fLocal); |
613 | if (quadCount > 1) { |
614 | fQuads.append(extra.fDevice, {color, domain, extra.fEdgeFlags}, &extra.fLocal); |
615 | fMetadata.fTotalQuadCount++; |
616 | } |
617 | return quadCount; |
618 | } |
619 | |
620 | GrProgramInfo* programInfo() override { |
621 | // Although this Op implements its own onPrePrepareDraws it calls GrMeshDrawOps' version so |
622 | // this entry point will be called. |
623 | return (fDesc) ? fDesc->fProgramInfo : nullptr; |
624 | } |
625 | |
626 | void onCreateProgramInfo(const GrCaps* caps, |
627 | SkArenaAlloc* arena, |
628 | const GrSurfaceProxyView* writeView, |
629 | GrAppliedClip&& appliedClip, |
630 | const GrXferProcessor::DstProxyView& dstProxyView) override { |
631 | SkASSERT(fDesc); |
632 | |
633 | GrGeometryProcessor* gp; |
634 | |
635 | { |
636 | const GrBackendFormat& backendFormat = |
637 | fViewCountPairs[0].fProxy->backendFormat(); |
638 | |
639 | GrSamplerState samplerState = GrSamplerState(GrSamplerState::WrapMode::kClamp, |
640 | fMetadata.filter()); |
641 | |
642 | gp = GrQuadPerEdgeAA::MakeTexturedProcessor( |
643 | arena, fDesc->fVertexSpec, *caps->shaderCaps(), backendFormat, samplerState, |
644 | fMetadata.fSwizzle, std::move(fTextureColorSpaceXform), fMetadata.saturate()); |
645 | |
646 | SkASSERT(fDesc->fVertexSpec.vertexSize() == gp->vertexStride()); |
647 | } |
648 | |
649 | auto pipelineFlags = (GrAAType::kMSAA == fMetadata.aaType()) ? |
650 | GrPipeline::InputFlags::kHWAntialias : GrPipeline::InputFlags::kNone; |
651 | |
652 | fDesc->fProgramInfo = GrSimpleMeshDrawOpHelper::CreateProgramInfo( |
653 | caps, arena, writeView, std::move(appliedClip), dstProxyView, gp, |
654 | GrProcessorSet::MakeEmptySet(), fDesc->fVertexSpec.primitiveType(), |
655 | pipelineFlags); |
656 | } |
657 | |
658 | void onPrePrepareDraws(GrRecordingContext* context, |
659 | const GrSurfaceProxyView* writeView, |
660 | GrAppliedClip* clip, |
661 | const GrXferProcessor::DstProxyView& dstProxyView) override { |
662 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
663 | |
664 | SkDEBUGCODE(this->validate();) |
665 | SkASSERT(!fDesc); |
666 | |
667 | SkArenaAlloc* arena = context->priv().recordTimeAllocator(); |
668 | |
669 | fDesc = arena->make<Desc>(); |
670 | this->characterize(fDesc); |
671 | fDesc->allocatePrePreparedVertices(arena); |
672 | FillInVertices(*context->priv().caps(), this, fDesc, fDesc->fPrePreparedVertices); |
673 | |
674 | // This will call onCreateProgramInfo and register the created program with the DDL. |
675 | this->INHERITED::onPrePrepareDraws(context, writeView, clip, dstProxyView); |
676 | } |
677 | |
678 | static void FillInVertices(const GrCaps& caps, TextureOp* texOp, Desc* desc, char* vertexData) { |
679 | SkASSERT(vertexData); |
680 | |
681 | int totQuadsSeen = 0; |
682 | SkDEBUGCODE(int totVerticesSeen = 0;) |
683 | SkDEBUGCODE(const size_t vertexSize = desc->fVertexSpec.vertexSize()); |
684 | |
685 | GrQuadPerEdgeAA::Tessellator tessellator(desc->fVertexSpec, vertexData); |
686 | for (const auto& op : ChainRange<TextureOp>(texOp)) { |
687 | auto iter = op.fQuads.iterator(); |
688 | for (unsigned p = 0; p < op.fMetadata.fProxyCount; ++p) { |
689 | const int quadCnt = op.fViewCountPairs[p].fQuadCnt; |
690 | SkDEBUGCODE(int meshVertexCnt = quadCnt * desc->fVertexSpec.verticesPerQuad()); |
691 | |
692 | // Can just use top-left for origin here since we only need the dimensions to |
693 | // determine the texel size for insetting. |
694 | NormalizationParams params = proxy_normalization_params( |
695 | op.fViewCountPairs[p].fProxy.get(), kTopLeft_GrSurfaceOrigin); |
696 | |
697 | bool inset = texOp->fMetadata.filter() != GrSamplerState::Filter::kNearest; |
698 | |
699 | for (int i = 0; i < quadCnt && iter.next(); ++i) { |
700 | SkASSERT(iter.isLocalValid()); |
701 | const ColorDomainAndAA& info = iter.metadata(); |
702 | |
703 | tessellator.append(iter.deviceQuad(), iter.localQuad(), info.fColor, |
704 | inset ? inset_domain_for_bilerp(params, info.fDomainRect) |
705 | : info.fDomainRect, |
706 | info.aaFlags()); |
707 | } |
708 | |
709 | SkASSERT((totVerticesSeen + meshVertexCnt) * vertexSize |
710 | == (size_t)(tessellator.vertices() - vertexData)); |
711 | |
712 | totQuadsSeen += quadCnt; |
713 | SkDEBUGCODE(totVerticesSeen += meshVertexCnt); |
714 | SkASSERT(totQuadsSeen * desc->fVertexSpec.verticesPerQuad() == totVerticesSeen); |
715 | } |
716 | |
717 | // If quad counts per proxy were calculated correctly, the entire iterator |
718 | // should have been consumed. |
719 | SkASSERT(!iter.next()); |
720 | } |
721 | |
722 | SkASSERT(desc->totalSizeInBytes() == (size_t)(tessellator.vertices() - vertexData)); |
723 | SkASSERT(totQuadsSeen == desc->fNumTotalQuads); |
724 | SkASSERT(totVerticesSeen == desc->totalNumVertices()); |
725 | } |
726 | |
727 | #ifdef SK_DEBUG |
728 | void validate() const override { |
729 | // NOTE: Since this is debug-only code, we use the virtual asTextureProxy() |
730 | auto textureType = fViewCountPairs[0].fProxy->asTextureProxy()->textureType(); |
731 | GrAAType aaType = fMetadata.aaType(); |
732 | |
733 | int quadCount = 0; |
734 | for (const auto& op : ChainRange<TextureOp>(this)) { |
735 | SkASSERT(op.fMetadata.fSwizzle == fMetadata.fSwizzle); |
736 | |
737 | for (unsigned p = 0; p < op.fMetadata.fProxyCount; ++p) { |
738 | auto* proxy = op.fViewCountPairs[p].fProxy->asTextureProxy(); |
739 | quadCount += op.fViewCountPairs[p].fQuadCnt; |
740 | SkASSERT(proxy); |
741 | SkASSERT(proxy->textureType() == textureType); |
742 | } |
743 | |
744 | // Each individual op must be a single aaType. kCoverage and kNone ops can chain |
745 | // together but kMSAA ones do not. |
746 | if (aaType == GrAAType::kCoverage || aaType == GrAAType::kNone) { |
747 | SkASSERT(op.fMetadata.aaType() == GrAAType::kCoverage || |
748 | op.fMetadata.aaType() == GrAAType::kNone); |
749 | } else { |
750 | SkASSERT(aaType == GrAAType::kMSAA && op.fMetadata.aaType() == GrAAType::kMSAA); |
751 | } |
752 | } |
753 | |
754 | SkASSERT(quadCount == this->numChainedQuads()); |
755 | } |
756 | #endif |
757 | |
758 | #if GR_TEST_UTILS |
759 | int numQuads() const final { return this->totNumQuads(); } |
760 | #endif |
761 | |
762 | void characterize(Desc* desc) const { |
763 | GrQuad::Type quadType = GrQuad::Type::kAxisAligned; |
764 | ColorType colorType = ColorType::kNone; |
765 | GrQuad::Type srcQuadType = GrQuad::Type::kAxisAligned; |
766 | Domain domain = Domain::kNo; |
767 | GrAAType overallAAType = fMetadata.aaType(); |
768 | |
769 | desc->fNumProxies = 0; |
770 | desc->fNumTotalQuads = 0; |
771 | int maxQuadsPerMesh = 0; |
772 | |
773 | for (const auto& op : ChainRange<TextureOp>(this)) { |
774 | if (op.fQuads.deviceQuadType() > quadType) { |
775 | quadType = op.fQuads.deviceQuadType(); |
776 | } |
777 | if (op.fQuads.localQuadType() > srcQuadType) { |
778 | srcQuadType = op.fQuads.localQuadType(); |
779 | } |
780 | if (op.fMetadata.domain() == Domain::kYes) { |
781 | domain = Domain::kYes; |
782 | } |
783 | colorType = std::max(colorType, op.fMetadata.colorType()); |
784 | desc->fNumProxies += op.fMetadata.fProxyCount; |
785 | |
786 | for (unsigned p = 0; p < op.fMetadata.fProxyCount; ++p) { |
787 | maxQuadsPerMesh = std::max(op.fViewCountPairs[p].fQuadCnt, maxQuadsPerMesh); |
788 | } |
789 | desc->fNumTotalQuads += op.totNumQuads(); |
790 | |
791 | if (op.fMetadata.aaType() == GrAAType::kCoverage) { |
792 | overallAAType = GrAAType::kCoverage; |
793 | } |
794 | } |
795 | |
796 | SkASSERT(desc->fNumTotalQuads == this->numChainedQuads()); |
797 | |
798 | SkASSERT(!CombinedQuadCountWillOverflow(overallAAType, false, desc->fNumTotalQuads)); |
799 | |
800 | auto indexBufferOption = GrQuadPerEdgeAA::CalcIndexBufferOption(overallAAType, |
801 | maxQuadsPerMesh); |
802 | |
803 | desc->fVertexSpec = VertexSpec(quadType, colorType, srcQuadType, /* hasLocal */ true, |
804 | domain, overallAAType, /* alpha as coverage */ true, |
805 | indexBufferOption); |
806 | |
807 | SkASSERT(desc->fNumTotalQuads <= GrQuadPerEdgeAA::QuadLimit(indexBufferOption)); |
808 | } |
809 | |
810 | int totNumQuads() const { |
811 | #ifdef SK_DEBUG |
812 | int tmp = 0; |
813 | for (unsigned p = 0; p < fMetadata.fProxyCount; ++p) { |
814 | tmp += fViewCountPairs[p].fQuadCnt; |
815 | } |
816 | SkASSERT(tmp == fMetadata.fTotalQuadCount); |
817 | #endif |
818 | |
819 | return fMetadata.fTotalQuadCount; |
820 | } |
821 | |
822 | int numChainedQuads() const { |
823 | int numChainedQuads = this->totNumQuads(); |
824 | |
825 | for (const GrOp* tmp = this->prevInChain(); tmp; tmp = tmp->prevInChain()) { |
826 | numChainedQuads += ((const TextureOp*)tmp)->totNumQuads(); |
827 | } |
828 | |
829 | for (const GrOp* tmp = this->nextInChain(); tmp; tmp = tmp->nextInChain()) { |
830 | numChainedQuads += ((const TextureOp*)tmp)->totNumQuads(); |
831 | } |
832 | |
833 | return numChainedQuads; |
834 | } |
835 | |
836 | // onPrePrepareDraws may or may not have been called at this point |
837 | void onPrepareDraws(Target* target) override { |
838 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
839 | |
840 | SkDEBUGCODE(this->validate();) |
841 | |
842 | SkASSERT(!fDesc || fDesc->fPrePreparedVertices); |
843 | |
844 | if (!fDesc) { |
845 | SkArenaAlloc* arena = target->allocator(); |
846 | fDesc = arena->make<Desc>(); |
847 | this->characterize(fDesc); |
848 | SkASSERT(!fDesc->fPrePreparedVertices); |
849 | } |
850 | |
851 | size_t vertexSize = fDesc->fVertexSpec.vertexSize(); |
852 | |
853 | void* vdata = target->makeVertexSpace(vertexSize, fDesc->totalNumVertices(), |
854 | &fDesc->fVertexBuffer, &fDesc->fBaseVertex); |
855 | if (!vdata) { |
856 | SkDebugf("Could not allocate vertices\n" ); |
857 | return; |
858 | } |
859 | |
860 | if (fDesc->fVertexSpec.needsIndexBuffer()) { |
861 | fDesc->fIndexBuffer = GrQuadPerEdgeAA::GetIndexBuffer( |
862 | target, fDesc->fVertexSpec.indexBufferOption()); |
863 | if (!fDesc->fIndexBuffer) { |
864 | SkDebugf("Could not allocate indices\n" ); |
865 | return; |
866 | } |
867 | } |
868 | |
869 | if (fDesc->fPrePreparedVertices) { |
870 | memcpy(vdata, fDesc->fPrePreparedVertices, fDesc->totalSizeInBytes()); |
871 | } else { |
872 | FillInVertices(target->caps(), this, fDesc, (char*) vdata); |
873 | } |
874 | } |
875 | |
876 | void onExecute(GrOpFlushState* flushState, const SkRect& chainBounds) override { |
877 | if (!fDesc->fVertexBuffer) { |
878 | return; |
879 | } |
880 | |
881 | if (fDesc->fVertexSpec.needsIndexBuffer() && !fDesc->fIndexBuffer) { |
882 | return; |
883 | } |
884 | |
885 | if (!fDesc->fProgramInfo) { |
886 | this->createProgramInfo(flushState); |
887 | SkASSERT(fDesc->fProgramInfo); |
888 | } |
889 | |
890 | flushState->bindPipelineAndScissorClip(*fDesc->fProgramInfo, chainBounds); |
891 | flushState->bindBuffers(fDesc->fIndexBuffer.get(), nullptr, fDesc->fVertexBuffer.get()); |
892 | |
893 | int totQuadsSeen = 0; |
894 | SkDEBUGCODE(int numDraws = 0;) |
895 | for (const auto& op : ChainRange<TextureOp>(this)) { |
896 | for (unsigned p = 0; p < op.fMetadata.fProxyCount; ++p) { |
897 | const int quadCnt = op.fViewCountPairs[p].fQuadCnt; |
898 | SkASSERT(numDraws < fDesc->fNumProxies); |
899 | flushState->bindTextures(fDesc->fProgramInfo->primProc(), |
900 | *op.fViewCountPairs[p].fProxy, |
901 | fDesc->fProgramInfo->pipeline()); |
902 | GrQuadPerEdgeAA::IssueDraw(flushState->caps(), flushState->opsRenderPass(), |
903 | fDesc->fVertexSpec, totQuadsSeen, quadCnt, |
904 | fDesc->totalNumVertices(), fDesc->fBaseVertex); |
905 | totQuadsSeen += quadCnt; |
906 | SkDEBUGCODE(++numDraws;) |
907 | } |
908 | } |
909 | |
910 | SkASSERT(totQuadsSeen == fDesc->fNumTotalQuads); |
911 | SkASSERT(numDraws == fDesc->fNumProxies); |
912 | } |
913 | |
914 | CombineResult onCombineIfPossible(GrOp* t, GrRecordingContext::Arenas*, |
915 | const GrCaps& caps) override { |
916 | TRACE_EVENT0("skia.gpu" , TRACE_FUNC); |
917 | const auto* that = t->cast<TextureOp>(); |
918 | |
919 | if (fDesc || that->fDesc) { |
920 | // This should never happen (since only DDL recorded ops should be prePrepared) |
921 | // but, in any case, we should never combine ops that that been prePrepared |
922 | return CombineResult::kCannotCombine; |
923 | } |
924 | |
925 | if (fMetadata.domain() != that->fMetadata.domain()) { |
926 | // It is technically possible to combine operations across domain modes, but performance |
927 | // testing suggests it's better to make more draw calls where some take advantage of |
928 | // the more optimal shader path without coordinate clamping. |
929 | return CombineResult::kCannotCombine; |
930 | } |
931 | if (!GrColorSpaceXform::Equals(fTextureColorSpaceXform.get(), |
932 | that->fTextureColorSpaceXform.get())) { |
933 | return CombineResult::kCannotCombine; |
934 | } |
935 | |
936 | bool upgradeToCoverageAAOnMerge = false; |
937 | if (fMetadata.aaType() != that->fMetadata.aaType()) { |
938 | if (!CanUpgradeAAOnMerge(fMetadata.aaType(), that->fMetadata.aaType())) { |
939 | return CombineResult::kCannotCombine; |
940 | } |
941 | upgradeToCoverageAAOnMerge = true; |
942 | } |
943 | |
944 | if (CombinedQuadCountWillOverflow(fMetadata.aaType(), upgradeToCoverageAAOnMerge, |
945 | this->numChainedQuads() + that->numChainedQuads())) { |
946 | return CombineResult::kCannotCombine; |
947 | } |
948 | |
949 | if (fMetadata.saturate() != that->fMetadata.saturate()) { |
950 | return CombineResult::kCannotCombine; |
951 | } |
952 | if (fMetadata.filter() != that->fMetadata.filter()) { |
953 | return CombineResult::kCannotCombine; |
954 | } |
955 | if (fMetadata.fSwizzle != that->fMetadata.fSwizzle) { |
956 | return CombineResult::kCannotCombine; |
957 | } |
958 | const auto* thisProxy = fViewCountPairs[0].fProxy.get(); |
959 | const auto* thatProxy = that->fViewCountPairs[0].fProxy.get(); |
960 | if (fMetadata.fProxyCount > 1 || that->fMetadata.fProxyCount > 1 || |
961 | thisProxy != thatProxy) { |
962 | // We can't merge across different proxies. Check if 'this' can be chained with 'that'. |
963 | if (GrTextureProxy::ProxiesAreCompatibleAsDynamicState(thisProxy, thatProxy) && |
964 | caps.dynamicStateArrayGeometryProcessorTextureSupport()) { |
965 | return CombineResult::kMayChain; |
966 | } |
967 | return CombineResult::kCannotCombine; |
968 | } |
969 | |
970 | fMetadata.fDomain |= that->fMetadata.fDomain; |
971 | fMetadata.fColorType = std::max(fMetadata.fColorType, that->fMetadata.fColorType); |
972 | if (upgradeToCoverageAAOnMerge) { |
973 | fMetadata.fAAType = static_cast<uint16_t>(GrAAType::kCoverage); |
974 | } |
975 | |
976 | // Concatenate quad lists together |
977 | fQuads.concat(that->fQuads); |
978 | fViewCountPairs[0].fQuadCnt += that->fQuads.count(); |
979 | fMetadata.fTotalQuadCount += that->fQuads.count(); |
980 | |
981 | return CombineResult::kMerged; |
982 | } |
983 | |
984 | GrQuadBuffer<ColorDomainAndAA> fQuads; |
985 | sk_sp<GrColorSpaceXform> fTextureColorSpaceXform; |
986 | // Most state of TextureOp is packed into these two field to minimize the op's size. |
987 | // Historically, increasing the size of TextureOp has caused surprising perf regressions, so |
988 | // consider/measure changes with care. |
989 | Desc* fDesc; |
990 | Metadata fMetadata; |
991 | |
992 | // This field must go last. When allocating this op, we will allocate extra space to hold |
993 | // additional ViewCountPairs immediately after the op's allocation so we can treat this |
994 | // as an fProxyCnt-length array. |
995 | ViewCountPair fViewCountPairs[1]; |
996 | |
997 | typedef GrMeshDrawOp INHERITED; |
998 | }; |
999 | |
1000 | } // anonymous namespace |
1001 | |
1002 | #if GR_TEST_UTILS |
1003 | uint32_t GrTextureOp::ClassID() { |
1004 | return TextureOp::ClassID(); |
1005 | } |
1006 | #endif |
1007 | |
1008 | std::unique_ptr<GrDrawOp> GrTextureOp::Make(GrRecordingContext* context, |
1009 | GrSurfaceProxyView proxyView, |
1010 | SkAlphaType alphaType, |
1011 | sk_sp<GrColorSpaceXform> textureXform, |
1012 | GrSamplerState::Filter filter, |
1013 | const SkPMColor4f& color, |
1014 | Saturate saturate, |
1015 | SkBlendMode blendMode, |
1016 | GrAAType aaType, |
1017 | DrawQuad* quad, |
1018 | const SkRect* domain) { |
1019 | // Apply optimizations that are valid whether or not using GrTextureOp or GrFillRectOp |
1020 | if (domain && domain->contains(proxyView.proxy()->backingStoreBoundsRect())) { |
1021 | // No need for a shader-based domain if hardware clamping achieves the same effect |
1022 | domain = nullptr; |
1023 | } |
1024 | |
1025 | if (filter != GrSamplerState::Filter::kNearest && |
1026 | !filter_has_effect(quad->fLocal, quad->fDevice)) { |
1027 | filter = GrSamplerState::Filter::kNearest; |
1028 | } |
1029 | |
1030 | if (blendMode == SkBlendMode::kSrcOver) { |
1031 | return TextureOp::Make(context, std::move(proxyView), std::move(textureXform), filter, |
1032 | color, saturate, aaType, std::move(quad), domain); |
1033 | } else { |
1034 | // Emulate complex blending using GrFillRectOp |
1035 | GrPaint paint; |
1036 | paint.setColor4f(color); |
1037 | paint.setXPFactory(SkBlendMode_AsXPFactory(blendMode)); |
1038 | |
1039 | std::unique_ptr<GrFragmentProcessor> fp; |
1040 | if (domain) { |
1041 | const auto& caps = *context->priv().caps(); |
1042 | SkRect localRect; |
1043 | if (quad->fLocal.asRect(&localRect)) { |
1044 | fp = GrTextureEffect::MakeSubset(std::move(proxyView), alphaType, SkMatrix::I(), filter, |
1045 | *domain, localRect, caps); |
1046 | } else { |
1047 | fp = GrTextureEffect::MakeSubset(std::move(proxyView), alphaType, SkMatrix::I(), filter, |
1048 | *domain, caps); |
1049 | } |
1050 | } else { |
1051 | fp = GrTextureEffect::Make(std::move(proxyView), alphaType, SkMatrix::I(), filter); |
1052 | } |
1053 | fp = GrColorSpaceXformEffect::Make(std::move(fp), std::move(textureXform)); |
1054 | paint.addColorFragmentProcessor(std::move(fp)); |
1055 | if (saturate == GrTextureOp::Saturate::kYes) { |
1056 | paint.addColorFragmentProcessor(GrClampFragmentProcessor::Make(false)); |
1057 | } |
1058 | |
1059 | return GrFillRectOp::Make(context, std::move(paint), aaType, quad); |
1060 | } |
1061 | } |
1062 | |
1063 | // A helper class that assists in breaking up bulk API quad draws into manageable chunks. |
1064 | class GrTextureOp::BatchSizeLimiter { |
1065 | public: |
1066 | BatchSizeLimiter(GrRenderTargetContext* rtc, |
1067 | const GrClip& clip, |
1068 | GrRecordingContext* context, |
1069 | int numEntries, |
1070 | GrSamplerState::Filter filter, |
1071 | GrTextureOp::Saturate saturate, |
1072 | SkCanvas::SrcRectConstraint constraint, |
1073 | const SkMatrix& viewMatrix, |
1074 | sk_sp<GrColorSpaceXform> textureColorSpaceXform) |
1075 | : fRTC(rtc) |
1076 | , fClip(clip) |
1077 | , fContext(context) |
1078 | , fFilter(filter) |
1079 | , fSaturate(saturate) |
1080 | , fConstraint(constraint) |
1081 | , fViewMatrix(viewMatrix) |
1082 | , fTextureColorSpaceXform(textureColorSpaceXform) |
1083 | , fNumLeft(numEntries) { |
1084 | } |
1085 | |
1086 | void createOp(GrRenderTargetContext::TextureSetEntry set[], |
1087 | int clumpSize, |
1088 | GrAAType aaType) { |
1089 | int clumpProxyCount = proxy_run_count(&set[fNumClumped], clumpSize); |
1090 | std::unique_ptr<GrDrawOp> op = TextureOp::Make(fContext, &set[fNumClumped], clumpSize, |
1091 | clumpProxyCount, fFilter, fSaturate, aaType, |
1092 | fConstraint, fViewMatrix, |
1093 | fTextureColorSpaceXform); |
1094 | fRTC->addDrawOp(fClip, std::move(op)); |
1095 | |
1096 | fNumLeft -= clumpSize; |
1097 | fNumClumped += clumpSize; |
1098 | } |
1099 | |
1100 | int numLeft() const { return fNumLeft; } |
1101 | int baseIndex() const { return fNumClumped; } |
1102 | |
1103 | private: |
1104 | GrRenderTargetContext* fRTC; |
1105 | const GrClip& fClip; |
1106 | GrRecordingContext* fContext; |
1107 | GrSamplerState::Filter fFilter; |
1108 | GrTextureOp::Saturate fSaturate; |
1109 | SkCanvas::SrcRectConstraint fConstraint; |
1110 | const SkMatrix& fViewMatrix; |
1111 | sk_sp<GrColorSpaceXform> fTextureColorSpaceXform; |
1112 | |
1113 | int fNumLeft; |
1114 | int fNumClumped = 0; // also the offset for the start of the next clump |
1115 | }; |
1116 | |
1117 | // Greedily clump quad draws together until the index buffer limit is exceeded. |
1118 | void GrTextureOp::AddTextureSetOps(GrRenderTargetContext* rtc, |
1119 | const GrClip& clip, |
1120 | GrRecordingContext* context, |
1121 | GrRenderTargetContext::TextureSetEntry set[], |
1122 | int cnt, |
1123 | int proxyRunCnt, |
1124 | GrSamplerState::Filter filter, |
1125 | Saturate saturate, |
1126 | SkBlendMode blendMode, |
1127 | GrAAType aaType, |
1128 | SkCanvas::SrcRectConstraint constraint, |
1129 | const SkMatrix& viewMatrix, |
1130 | sk_sp<GrColorSpaceXform> textureColorSpaceXform) { |
1131 | // Ensure that the index buffer limits are lower than the proxy and quad count limits of |
1132 | // the op's metadata so we don't need to worry about overflow. |
1133 | SkDEBUGCODE(TextureOp::ValidateResourceLimits();) |
1134 | SkASSERT(proxy_run_count(set, cnt) == proxyRunCnt); |
1135 | |
1136 | // First check if we can support batches as a single op |
1137 | if (blendMode != SkBlendMode::kSrcOver || |
1138 | !context->priv().caps()->dynamicStateArrayGeometryProcessorTextureSupport()) { |
1139 | // Append each entry as its own op; these may still be GrTextureOps if the blend mode is |
1140 | // src-over but the backend doesn't support dynamic state changes. Otherwise Make() |
1141 | // automatically creates the appropriate GrFillRectOp to emulate GrTextureOp. |
1142 | SkMatrix ctm; |
1143 | for (int i = 0; i < cnt; ++i) { |
1144 | float alpha = set[i].fAlpha; |
1145 | ctm = viewMatrix; |
1146 | if (set[i].fPreViewMatrix) { |
1147 | ctm.preConcat(*set[i].fPreViewMatrix); |
1148 | } |
1149 | |
1150 | DrawQuad quad; |
1151 | quad.fEdgeFlags = set[i].fAAFlags; |
1152 | if (set[i].fDstClipQuad) { |
1153 | quad.fDevice = GrQuad::MakeFromSkQuad(set[i].fDstClipQuad, ctm); |
1154 | |
1155 | SkPoint srcPts[4]; |
1156 | GrMapRectPoints(set[i].fDstRect, set[i].fSrcRect, set[i].fDstClipQuad, srcPts, 4); |
1157 | quad.fLocal = GrQuad::MakeFromSkQuad(srcPts, SkMatrix::I()); |
1158 | } else { |
1159 | quad.fDevice = GrQuad::MakeFromRect(set[i].fDstRect, ctm); |
1160 | quad.fLocal = GrQuad(set[i].fSrcRect); |
1161 | } |
1162 | |
1163 | const SkRect* domain = constraint == SkCanvas::kStrict_SrcRectConstraint |
1164 | ? &set[i].fSrcRect : nullptr; |
1165 | |
1166 | auto op = Make(context, set[i].fProxyView, set[i].fSrcAlphaType, textureColorSpaceXform, |
1167 | filter, {alpha, alpha, alpha, alpha}, saturate, blendMode, aaType, |
1168 | &quad, domain); |
1169 | rtc->addDrawOp(clip, std::move(op)); |
1170 | } |
1171 | return; |
1172 | } |
1173 | |
1174 | // Second check if we can always just make a single op and avoid the extra iteration |
1175 | // needed to clump things together. |
1176 | if (cnt <= std::min(GrResourceProvider::MaxNumNonAAQuads(), |
1177 | GrResourceProvider::MaxNumAAQuads())) { |
1178 | auto op = TextureOp::Make(context, set, cnt, proxyRunCnt, filter, saturate, aaType, |
1179 | constraint, viewMatrix, std::move(textureColorSpaceXform)); |
1180 | rtc->addDrawOp(clip, std::move(op)); |
1181 | return; |
1182 | } |
1183 | |
1184 | BatchSizeLimiter state(rtc, clip, context, cnt, filter, saturate, constraint, viewMatrix, |
1185 | std::move(textureColorSpaceXform)); |
1186 | |
1187 | // kNone and kMSAA never get altered |
1188 | if (aaType == GrAAType::kNone || aaType == GrAAType::kMSAA) { |
1189 | // Clump these into series of MaxNumNonAAQuads-sized GrTextureOps |
1190 | while (state.numLeft() > 0) { |
1191 | int clumpSize = std::min(state.numLeft(), GrResourceProvider::MaxNumNonAAQuads()); |
1192 | |
1193 | state.createOp(set, clumpSize, aaType); |
1194 | } |
1195 | } else { |
1196 | // kCoverage can be downgraded to kNone. Note that the following is conservative. kCoverage |
1197 | // can also get downgraded to kNone if all the quads are on integer coordinates and |
1198 | // axis-aligned. |
1199 | SkASSERT(aaType == GrAAType::kCoverage); |
1200 | |
1201 | while (state.numLeft() > 0) { |
1202 | GrAAType runningAA = GrAAType::kNone; |
1203 | bool clumped = false; |
1204 | |
1205 | for (int i = 0; i < state.numLeft(); ++i) { |
1206 | int absIndex = state.baseIndex() + i; |
1207 | |
1208 | if (set[absIndex].fAAFlags != GrQuadAAFlags::kNone) { |
1209 | |
1210 | if (i >= GrResourceProvider::MaxNumAAQuads()) { |
1211 | // Here we either need to boost the AA type to kCoverage, but doing so with |
1212 | // all the accumulated quads would overflow, or we have a set of AA quads |
1213 | // that has just gotten too large. In either case, calve off the existing |
1214 | // quads as their own TextureOp. |
1215 | state.createOp( |
1216 | set, |
1217 | runningAA == GrAAType::kNone ? i : GrResourceProvider::MaxNumAAQuads(), |
1218 | runningAA); // maybe downgrading AA here |
1219 | clumped = true; |
1220 | break; |
1221 | } |
1222 | |
1223 | runningAA = GrAAType::kCoverage; |
1224 | } else if (runningAA == GrAAType::kNone) { |
1225 | |
1226 | if (i >= GrResourceProvider::MaxNumNonAAQuads()) { |
1227 | // Here we've found a consistent batch of non-AA quads that has gotten too |
1228 | // large. Calve it off as its own GrTextureOp. |
1229 | state.createOp(set, GrResourceProvider::MaxNumNonAAQuads(), |
1230 | GrAAType::kNone); // definitely downgrading AA here |
1231 | clumped = true; |
1232 | break; |
1233 | } |
1234 | } |
1235 | } |
1236 | |
1237 | if (!clumped) { |
1238 | // We ran through the above loop w/o hitting a limit. Spit out this last clump of |
1239 | // quads and call it a day. |
1240 | state.createOp(set, state.numLeft(), runningAA); // maybe downgrading AA here |
1241 | } |
1242 | } |
1243 | } |
1244 | } |
1245 | |
1246 | #if GR_TEST_UTILS |
1247 | #include "include/private/GrRecordingContext.h" |
1248 | #include "src/gpu/GrProxyProvider.h" |
1249 | #include "src/gpu/GrRecordingContextPriv.h" |
1250 | |
1251 | GR_DRAW_OP_TEST_DEFINE(TextureOp) { |
1252 | SkISize dims; |
1253 | dims.fHeight = random->nextULessThan(90) + 10; |
1254 | dims.fWidth = random->nextULessThan(90) + 10; |
1255 | auto origin = random->nextBool() ? kTopLeft_GrSurfaceOrigin : kBottomLeft_GrSurfaceOrigin; |
1256 | GrMipMapped mipMapped = random->nextBool() ? GrMipMapped::kYes : GrMipMapped::kNo; |
1257 | SkBackingFit fit = SkBackingFit::kExact; |
1258 | if (mipMapped == GrMipMapped::kNo) { |
1259 | fit = random->nextBool() ? SkBackingFit::kApprox : SkBackingFit::kExact; |
1260 | } |
1261 | const GrBackendFormat format = |
1262 | context->priv().caps()->getDefaultBackendFormat(GrColorType::kRGBA_8888, |
1263 | GrRenderable::kNo); |
1264 | GrProxyProvider* proxyProvider = context->priv().proxyProvider(); |
1265 | sk_sp<GrTextureProxy> proxy = proxyProvider->createProxy( |
1266 | format, dims, GrRenderable::kNo, 1, mipMapped, fit, SkBudgeted::kNo, GrProtected::kNo, |
1267 | GrInternalSurfaceFlags::kNone); |
1268 | |
1269 | SkRect rect = GrTest::TestRect(random); |
1270 | SkRect srcRect; |
1271 | srcRect.fLeft = random->nextRangeScalar(0.f, proxy->width() / 2.f); |
1272 | srcRect.fRight = random->nextRangeScalar(0.f, proxy->width()) + proxy->width() / 2.f; |
1273 | srcRect.fTop = random->nextRangeScalar(0.f, proxy->height() / 2.f); |
1274 | srcRect.fBottom = random->nextRangeScalar(0.f, proxy->height()) + proxy->height() / 2.f; |
1275 | SkMatrix viewMatrix = GrTest::TestMatrixPreservesRightAngles(random); |
1276 | SkPMColor4f color = SkPMColor4f::FromBytes_RGBA(SkColorToPremulGrColor(random->nextU())); |
1277 | GrSamplerState::Filter filter = (GrSamplerState::Filter)random->nextULessThan( |
1278 | static_cast<uint32_t>(GrSamplerState::Filter::kMipMap) + 1); |
1279 | while (mipMapped == GrMipMapped::kNo && filter == GrSamplerState::Filter::kMipMap) { |
1280 | filter = (GrSamplerState::Filter)random->nextULessThan( |
1281 | static_cast<uint32_t>(GrSamplerState::Filter::kMipMap) + 1); |
1282 | } |
1283 | auto texXform = GrTest::TestColorXform(random); |
1284 | GrAAType aaType = GrAAType::kNone; |
1285 | if (random->nextBool()) { |
1286 | aaType = (numSamples > 1) ? GrAAType::kMSAA : GrAAType::kCoverage; |
1287 | } |
1288 | GrQuadAAFlags aaFlags = GrQuadAAFlags::kNone; |
1289 | aaFlags |= random->nextBool() ? GrQuadAAFlags::kLeft : GrQuadAAFlags::kNone; |
1290 | aaFlags |= random->nextBool() ? GrQuadAAFlags::kTop : GrQuadAAFlags::kNone; |
1291 | aaFlags |= random->nextBool() ? GrQuadAAFlags::kRight : GrQuadAAFlags::kNone; |
1292 | aaFlags |= random->nextBool() ? GrQuadAAFlags::kBottom : GrQuadAAFlags::kNone; |
1293 | bool useDomain = random->nextBool(); |
1294 | auto saturate = random->nextBool() ? GrTextureOp::Saturate::kYes : GrTextureOp::Saturate::kNo; |
1295 | GrSurfaceProxyView proxyView( |
1296 | std::move(proxy), origin, |
1297 | context->priv().caps()->getReadSwizzle(format, GrColorType::kRGBA_8888)); |
1298 | auto alphaType = static_cast<SkAlphaType>( |
1299 | random->nextRangeU(kUnknown_SkAlphaType + 1, kLastEnum_SkAlphaType)); |
1300 | |
1301 | DrawQuad quad = {GrQuad::MakeFromRect(rect, viewMatrix), GrQuad(srcRect), aaFlags}; |
1302 | return GrTextureOp::Make(context, std::move(proxyView), alphaType, std::move(texXform), filter, |
1303 | color, saturate, SkBlendMode::kSrcOver, aaType, |
1304 | &quad, useDomain ? &srcRect : nullptr); |
1305 | } |
1306 | |
1307 | #endif |
1308 | |