Revert "Delete GPU alpha clip masks"

This reverts commit a466228a616b1b02ede2d4389fefcfc839a54a25.

Reason for revert: Turned up a vulkan driver bug

Original change's description:
> Delete GPU alpha clip masks
> 
> The cost of switching render targets on each draw to make a custom
> clip is enormous. There are virtually no circumstances where this will
> outperform our cached, multi-threaded software mask generator. The
> tried-and-true approach to clipping on-GPU is with analytic FPs. And
> now that we support CCPR clip FPs, there ulitmately should be very few
> clip stacks that even require a mask as long as they don't use
> deprecated SkClipOps.
> 
> Bug: skia:
> Change-Id: I79c5558c93c1b99179f1e933d029f69b14ad1ce3
> Reviewed-on: https://skia-review.googlesource.com/116724
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Chris Dalton <csmartdalton@google.com>

TBR=robertphillips@google.com,brianosman@google.com,csmartdalton@google.com

# Not skipping CQ checks because original CL landed > 1 day ago.

Bug: skia:
Change-Id: Iba289e00ba2eca7084dc8517491cfb5f6ab6266f
Reviewed-on: https://skia-review.googlesource.com/117420
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>

5 files changed, 347 insertions, 16 deletions

diff --git a/gm/windowrectangles.cpp b/gm/windowrectangles.cpp
index 8b85554af8..931a154df4 100644
--- a/gm/windowrectangles.cpp
+++ b/gm/windowrectangles.cpp
@@ -107,8 +107,8 @@ DEF_GM( return new WindowRectanglesGM(); )
 constexpr static int kNumWindows = 8;
 
 /**
- * Visualizes the stencil mask for a clip with several window rectangles. The purpose of this test
- * is to verify that window rectangles are being used during clip mask generation, and to
+ * Visualizes the mask (alpha or stencil) for a clip with several window rectangles. The purpose of
+ * this test is to verify that window rectangles are being used during clip mask generation, and to
  * visualize where the window rectangles are placed.
  *
  * We use window rectangles when generating the clip mask because there is no need to invest time
@@ -123,6 +123,7 @@ private:
     constexpr static int kMaskCheckerSize = 5;
     SkString onShortName() final { return SkString("windowrectangles_mask"); }
     void onCoverClipStack(const SkClipStack&, SkCanvas*) final;
+    void visualizeAlphaMask(GrContext*, GrRenderTargetContext*, const GrReducedClip&, GrPaint&&);
     void visualizeStencilMask(GrContext*, GrRenderTargetContext*, const GrReducedClip&, GrPaint&&);
     void stencilCheckerboard(GrRenderTargetContext*, bool flip);
     void fail(SkCanvas*);
@@ -183,14 +184,48 @@ void WindowRectanglesMaskGM::onCoverClipStack(const SkClipStack& stack, SkCanvas
     const GrReducedClip reducedClip(stack, SkRect::Make(kCoverRect), rtc->caps()->shaderCaps(),
                                     kNumWindows);
 
-     GrPaint paint;
+    GrPaint paint;
     if (GrFSAAType::kNone == rtc->fsaaType()) {
-        // Use different colors so we don't confuse masks that don't have AA with ones that should.
         paint.setColor4f(GrColor4f(0, 0.25f, 1, 1));
+        this->visualizeAlphaMask(ctx, rtc, reducedClip, std::move(paint));
     } else {
         paint.setColor4f(GrColor4f(1, 0.25f, 0.25f, 1));
+        this->visualizeStencilMask(ctx, rtc, reducedClip, std::move(paint));
     }
-    this->visualizeStencilMask(ctx, rtc, reducedClip, std::move(paint));
+}
+
+void WindowRectanglesMaskGM::visualizeAlphaMask(GrContext* ctx, GrRenderTargetContext* rtc,
+                                                const GrReducedClip& reducedClip, GrPaint&& paint) {
+    const int padRight = (kDeviceRect.right() - kCoverRect.right()) / 2;
+    const int padBottom = (kDeviceRect.bottom() - kCoverRect.bottom()) / 2;
+    sk_sp<GrRenderTargetContext> maskRTC(
+        ctx->contextPriv().makeDeferredRenderTargetContextWithFallback(
+                                                         SkBackingFit::kExact,
+                                                         kCoverRect.width() + padRight,
+                                                         kCoverRect.height() + padBottom,
+                                                         kAlpha_8_GrPixelConfig, nullptr));
+    if (!maskRTC) {
+        return;
+    }
+
+    // Draw a checker pattern into the alpha mask so we can visualize the regions left untouched by
+    // the clip mask generation.
+    this->stencilCheckerboard(maskRTC.get(), true);
+    maskRTC->clear(nullptr, GrColorPackA4(0xff), GrRenderTargetContext::CanClearFullscreen::kYes);
+    maskRTC->priv().drawAndStencilRect(make_stencil_only_clip(), &GrUserStencilSettings::kUnused,
+                                       SkRegion::kDifference_Op, false, GrAA::kNo, SkMatrix::I(),
+                                       SkRect::MakeIWH(maskRTC->width(), maskRTC->height()));
+    reducedClip.drawAlphaClipMask(maskRTC.get());
+
+    int x = kCoverRect.x() - kDeviceRect.x(),
+        y = kCoverRect.y() - kDeviceRect.y();
+
+    // Now visualize the alpha mask by drawing a rect over the area where it is defined. The regions
+    // inside window rectangles or outside the scissor should still have the initial checkerboard
+    // intact. (This verifies we didn't spend any time modifying those pixels in the mask.)
+    AlphaOnlyClip clip(maskRTC->asTextureProxyRef(), x, y);
+    rtc->drawRect(clip, std::move(paint), GrAA::kYes, SkMatrix::I(),
+                  SkRect::Make(SkIRect::MakeXYWH(x, y, maskRTC->width(), maskRTC->height())));
 }
 
 void WindowRectanglesMaskGM::visualizeStencilMask(GrContext* ctx, GrRenderTargetContext* rtc,
diff --git a/src/gpu/GrClipStackClip.cpp b/src/gpu/GrClipStackClip.cpp
index ed4a098c8d..fe8d10095c 100644
--- a/src/gpu/GrClipStackClip.cpp
+++ b/src/gpu/GrClipStackClip.cpp
@@ -73,6 +73,111 @@ void GrClipStackClip::getConservativeBounds(int width, int height, SkIRect* devR
     devBounds.roundOut(devResult);
 }
 
+////////////////////////////////////////////////////////////////////////////////
+// set up the draw state to enable the aa clipping mask.
+static std::unique_ptr<GrFragmentProcessor> create_fp_for_mask(sk_sp<GrTextureProxy> mask,
+                                                               const SkIRect& devBound) {
+    SkIRect domainTexels = SkIRect::MakeWH(devBound.width(), devBound.height());
+    return GrDeviceSpaceTextureDecalFragmentProcessor::Make(std::move(mask), domainTexels,
+                                                            {devBound.fLeft, devBound.fTop});
+}
+
+// Does the path in 'element' require SW rendering? If so, return true (and,
+// optionally, set 'prOut' to NULL. If not, return false (and, optionally, set
+// 'prOut' to the non-SW path renderer that will do the job).
+bool GrClipStackClip::PathNeedsSWRenderer(GrContext* context,
+                                          const SkIRect& scissorRect,
+                                          bool hasUserStencilSettings,
+                                          const GrRenderTargetContext* renderTargetContext,
+                                          const SkMatrix& viewMatrix,
+                                          const Element* element,
+                                          GrPathRenderer** prOut,
+                                          bool needsStencil) {
+    if (Element::DeviceSpaceType::kRect == element->getDeviceSpaceType()) {
+        // rects can always be drawn directly w/o using the software path
+        // TODO: skip rrects once we're drawing them directly.
+        if (prOut) {
+            *prOut = nullptr;
+        }
+        return false;
+    } else {
+        // We shouldn't get here with an empty clip element.
+        SkASSERT(Element::DeviceSpaceType::kEmpty != element->getDeviceSpaceType());
+
+        // the gpu alpha mask will draw the inverse paths as non-inverse to a temp buffer
+        SkPath path;
+        element->asDeviceSpacePath(&path);
+        if (path.isInverseFillType()) {
+            path.toggleInverseFillType();
+        }
+
+        GrPathRendererChain::DrawType type =
+                needsStencil ? GrPathRendererChain::DrawType::kStencilAndColor
+                             : GrPathRendererChain::DrawType::kColor;
+
+        GrShape shape(path, GrStyle::SimpleFill());
+        GrPathRenderer::CanDrawPathArgs canDrawArgs;
+        canDrawArgs.fCaps = context->caps();
+        canDrawArgs.fClipConservativeBounds = &scissorRect;
+        canDrawArgs.fViewMatrix = &viewMatrix;
+        canDrawArgs.fShape = &shape;
+        canDrawArgs.fAAType = GrChooseAAType(GrAA(element->isAA()),
+                                             renderTargetContext->fsaaType(),
+                                             GrAllowMixedSamples::kYes,
+                                             *context->caps());
+        canDrawArgs.fHasUserStencilSettings = hasUserStencilSettings;
+
+        // the 'false' parameter disallows use of the SW path renderer
+        GrPathRenderer* pr =
+            context->contextPriv().drawingManager()->getPathRenderer(canDrawArgs, false, type);
+        if (prOut) {
+            *prOut = pr;
+        }
+        return SkToBool(!pr);
+    }
+}
+
+/*
+ * This method traverses the clip stack to see if the GrSoftwarePathRenderer
+ * will be used on any element. If so, it returns true to indicate that the
+ * entire clip should be rendered in SW and then uploaded en masse to the gpu.
+ */
+bool GrClipStackClip::UseSWOnlyPath(GrContext* context,
+                                    bool hasUserStencilSettings,
+                                    const GrRenderTargetContext* renderTargetContext,
+                                    const GrReducedClip& reducedClip) {
+    // TODO: generalize this function so that when
+    // a clip gets complex enough it can just be done in SW regardless
+    // of whether it would invoke the GrSoftwarePathRenderer.
+
+    // If we're avoiding stencils, always use SW:
+    if (context->caps()->avoidStencilBuffers())
+        return true;
+
+    // Set the matrix so that rendered clip elements are transformed to mask space from clip
+    // space.
+    SkMatrix translate;
+    translate.setTranslate(SkIntToScalar(-reducedClip.left()), SkIntToScalar(-reducedClip.top()));
+
+    for (ElementList::Iter iter(reducedClip.maskElements()); iter.get(); iter.next()) {
+        const Element* element = iter.get();
+
+        SkClipOp op = element->getOp();
+        bool invert = element->isInverseFilled();
+        bool needsStencil = invert ||
+                            kIntersect_SkClipOp == op || kReverseDifference_SkClipOp == op;
+
+        if (PathNeedsSWRenderer(context, reducedClip.scissor(), hasUserStencilSettings,
+                                renderTargetContext, translate, element, nullptr, needsStencil)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// sort out what kind of clip mask needs to be created: alpha, stencil,
+// scissor, or entirely software
 bool GrClipStackClip::apply(GrContext* context, GrRenderTargetContext* renderTargetContext,
                             bool useHWAA, bool hasUserStencilSettings, GrAppliedClip* out,
                             SkRect* bounds) const {
@@ -148,18 +253,24 @@ bool GrClipStackClip::applyClipMask(GrContext* context, GrRenderTargetContext* r
     // If the stencil buffer is multisampled we can use it to do everything.
     if ((GrFSAAType::kNone == renderTargetContext->fsaaType() && reducedClip.maskRequiresAA()) ||
         context->caps()->avoidStencilBuffers()) {
-        if (auto mask = this->createSoftwareClipMask(context, reducedClip, renderTargetContext)) {
-            // The mask should fill the clip's scissor.
-            SkIRect domainTexels = SkIRect::MakeWH(reducedClip.width(), reducedClip.height());
-            SkIPoint maskOffset = SkIPoint::Make(reducedClip.left(), reducedClip.top());
-            out->addCoverageFP(GrDeviceSpaceTextureDecalFragmentProcessor::Make(std::move(mask),
-                                                                                domainTexels,
-                                                                                maskOffset));
+        sk_sp<GrTextureProxy> result;
+        if (UseSWOnlyPath(context, hasUserStencilSettings, renderTargetContext, reducedClip)) {
+            // The clip geometry is complex enough that it will be more efficient to create it
+            // entirely in software
+            result = this->createSoftwareClipMask(context, reducedClip, renderTargetContext);
+        } else {
+            result = this->createAlphaClipMask(context, reducedClip);
+        }
+
+        if (result) {
+            // The mask's top left coord should be pinned to the rounded-out top left corner of
+            // the clip's device space bounds.
+            out->addCoverageFP(create_fp_for_mask(std::move(result), reducedClip.scissor()));
             return true;
         }
 
-        // If software clip mask creation fails, fall through to the stencil code paths, unless
-        // stencils are disallowed.
+        // If alpha or software clip mask creation fails, fall through to the stencil code paths,
+        // unless stencils are disallowed.
         if (context->caps()->avoidStencilBuffers()) {
             SkDebugf("WARNING: Clip mask requires stencil, but stencil unavailable. "
                      "Clip will be ignored.\n");
@@ -213,6 +324,45 @@ static void add_invalidate_on_pop_message(const SkClipStack& stack, uint32_t cli
     SkDEBUGFAIL("Gen ID was not found in stack.");
 }
 
+sk_sp<GrTextureProxy> GrClipStackClip::createAlphaClipMask(GrContext* context,
+                                                           const GrReducedClip& reducedClip) const {
+    GrProxyProvider* proxyProvider = context->contextPriv().proxyProvider();
+    GrUniqueKey key;
+    create_clip_mask_key(reducedClip.maskGenID(), reducedClip.scissor(),
+                         reducedClip.numAnalyticFPs(), &key);
+
+    sk_sp<GrTextureProxy> proxy(proxyProvider->findOrCreateProxyByUniqueKey(
+                                                                key, kBottomLeft_GrSurfaceOrigin));
+    if (proxy) {
+        return proxy;
+    }
+
+    sk_sp<GrRenderTargetContext> rtc(
+        context->contextPriv().makeDeferredRenderTargetContextWithFallback(SkBackingFit::kApprox,
+                                                                           reducedClip.width(),
+                                                                           reducedClip.height(),
+                                                                           kAlpha_8_GrPixelConfig,
+                                                                           nullptr));
+    if (!rtc) {
+        return nullptr;
+    }
+
+    if (!reducedClip.drawAlphaClipMask(rtc.get())) {
+        return nullptr;
+    }
+
+    sk_sp<GrTextureProxy> result(rtc->asTextureProxyRef());
+    if (!result) {
+        return nullptr;
+    }
+
+    SkASSERT(result->origin() == kBottomLeft_GrSurfaceOrigin);
+    proxyProvider->assignUniqueKeyToProxy(key, result.get());
+    add_invalidate_on_pop_message(*fStack, reducedClip.maskGenID(), key);
+
+    return result;
+}
+
 namespace {
 
 /**
diff --git a/src/gpu/GrClipStackClip.h b/src/gpu/GrClipStackClip.h
index 05d9f18c88..aeb983468c 100644
--- a/src/gpu/GrClipStackClip.h
+++ b/src/gpu/GrClipStackClip.h
@@ -37,14 +37,31 @@ public:
     static const char kMaskTestTag[];
 
 private:
+    static bool PathNeedsSWRenderer(GrContext* context,
+                                    const SkIRect& scissorRect,
+                                    bool hasUserStencilSettings,
+                                    const GrRenderTargetContext*,
+                                    const SkMatrix& viewMatrix,
+                                    const SkClipStack::Element* element,
+                                    GrPathRenderer** prOut,
+                                    bool needsStencil);
+
     bool applyClipMask(GrContext*, GrRenderTargetContext*, const GrReducedClip&,
                        bool hasUserStencilSettings, GrAppliedClip*) const;
 
-    // Creates an alpha mask of the remaining reduced clip elements that could not be handled
-    // analytically on the GPU. The mask fills the reduced clip's scissor rect.
+    // Creates an alpha mask of the clip. The mask is a rasterization of elements through the
+    // rect specified by clipSpaceIBounds.
+    sk_sp<GrTextureProxy> createAlphaClipMask(GrContext*, const GrReducedClip&) const;
+
+    // Similar to createAlphaClipMask but it rasterizes in SW and uploads to the result texture.
     sk_sp<GrTextureProxy> createSoftwareClipMask(GrContext*, const GrReducedClip&,
                                                  GrRenderTargetContext*) const;
 
+    static bool UseSWOnlyPath(GrContext*,
+                              bool hasUserStencilSettings,
+                              const GrRenderTargetContext*,
+                              const GrReducedClip&);
+
     const SkClipStack*  fStack;
 };
 
diff --git a/src/gpu/GrReducedClip.cpp b/src/gpu/GrReducedClip.cpp
index 2f60dad28f..9d27f7696e 100644
--- a/src/gpu/GrReducedClip.cpp
+++ b/src/gpu/GrReducedClip.cpp
@@ -673,6 +673,134 @@ void GrReducedClip::makeEmpty() {
 }
 
 ////////////////////////////////////////////////////////////////////////////////
+// Create a 8-bit clip mask in alpha
+
+static bool stencil_element(GrRenderTargetContext* rtc,
+                            const GrFixedClip& clip,
+                            const GrUserStencilSettings* ss,
+                            const SkMatrix& viewMatrix,
+                            const SkClipStack::Element* element) {
+    GrAA aa = GrAA(element->isAA());
+    switch (element->getDeviceSpaceType()) {
+        case SkClipStack::Element::DeviceSpaceType::kEmpty:
+            SkDEBUGFAIL("Should never get here with an empty element.");
+            break;
+        case SkClipStack::Element::DeviceSpaceType::kRect:
+            return rtc->priv().drawAndStencilRect(clip, ss, (SkRegion::Op)element->getOp(),
+                                                  element->isInverseFilled(), aa, viewMatrix,
+                                                  element->getDeviceSpaceRect());
+            break;
+        default: {
+            SkPath path;
+            element->asDeviceSpacePath(&path);
+            if (path.isInverseFillType()) {
+                path.toggleInverseFillType();
+            }
+
+            return rtc->priv().drawAndStencilPath(clip, ss, (SkRegion::Op)element->getOp(),
+                                                  element->isInverseFilled(), aa, viewMatrix, path);
+            break;
+        }
+    }
+
+    return false;
+}
+
+static void draw_element(GrRenderTargetContext* rtc,
+                         const GrClip& clip,  // TODO: can this just always be WideOpen?
+                         GrPaint&& paint,
+                         GrAA aa,
+                         const SkMatrix& viewMatrix,
+                         const SkClipStack::Element* element) {
+    // TODO: Draw rrects directly here.
+    switch (element->getDeviceSpaceType()) {
+        case SkClipStack::Element::DeviceSpaceType::kEmpty:
+            SkDEBUGFAIL("Should never get here with an empty element.");
+            break;
+        case SkClipStack::Element::DeviceSpaceType::kRect:
+            rtc->drawRect(clip, std::move(paint), aa, viewMatrix, element->getDeviceSpaceRect());
+            break;
+        default: {
+            SkPath path;
+            element->asDeviceSpacePath(&path);
+            if (path.isInverseFillType()) {
+                path.toggleInverseFillType();
+            }
+
+            rtc->drawPath(clip, std::move(paint), aa, viewMatrix, path, GrStyle::SimpleFill());
+            break;
+        }
+    }
+}
+
+bool GrReducedClip::drawAlphaClipMask(GrRenderTargetContext* rtc) const {
+    // The texture may be larger than necessary, this rect represents the part of the texture
+    // we populate with a rasterization of the clip.
+    GrFixedClip clip(SkIRect::MakeWH(fScissor.width(), fScissor.height()));
+
+    if (!fWindowRects.empty()) {
+        clip.setWindowRectangles(fWindowRects.makeOffset(-fScissor.left(), -fScissor.top()),
+                                 GrWindowRectsState::Mode::kExclusive);
+    }
+
+    // The scratch texture that we are drawing into can be substantially larger than the mask. Only
+    // clear the part that we care about.
+    GrColor initialCoverage = InitialState::kAllIn == this->initialState() ? -1 : 0;
+    rtc->priv().clear(clip, initialCoverage, GrRenderTargetContext::CanClearFullscreen::kYes);
+
+    // Set the matrix so that rendered clip elements are transformed to mask space from clip space.
+    SkMatrix translate;
+    translate.setTranslate(SkIntToScalar(-fScissor.left()), SkIntToScalar(-fScissor.top()));
+
+    // walk through each clip element and perform its set op
+    for (ElementList::Iter iter(fMaskElements); iter.get(); iter.next()) {
+        const Element* element = iter.get();
+        SkRegion::Op op = (SkRegion::Op)element->getOp();
+        GrAA aa = GrAA(element->isAA());
+        bool invert = element->isInverseFilled();
+        if (invert || SkRegion::kIntersect_Op == op || SkRegion::kReverseDifference_Op == op) {
+            // draw directly into the result with the stencil set to make the pixels affected
+            // by the clip shape be non-zero.
+            static constexpr GrUserStencilSettings kStencilInElement(
+                 GrUserStencilSettings::StaticInit<
+                     0xffff,
+                     GrUserStencilTest::kAlways,
+                     0xffff,
+                     GrUserStencilOp::kReplace,
+                     GrUserStencilOp::kReplace,
+                     0xffff>()
+            );
+            if (!stencil_element(rtc, clip, &kStencilInElement, translate, element)) {
+                return false;
+            }
+
+            // Draw to the exterior pixels (those with a zero stencil value).
+            static constexpr GrUserStencilSettings kDrawOutsideElement(
+                 GrUserStencilSettings::StaticInit<
+                     0x0000,
+                     GrUserStencilTest::kEqual,
+                     0xffff,
+                     GrUserStencilOp::kZero,
+                     GrUserStencilOp::kZero,
+                     0xffff>()
+            );
+            if (!rtc->priv().drawAndStencilRect(clip, &kDrawOutsideElement, op, !invert, GrAA::kNo,
+                                                translate, SkRect::Make(fScissor))) {
+                return false;
+            }
+        } else {
+            // all the remaining ops can just be directly draw into the accumulation buffer
+            GrPaint paint;
+            paint.setCoverageSetOpXPFactory(op, false);
+
+            draw_element(rtc, clip, std::move(paint), aa, translate, element);
+        }
+    }
+
+    return true;
+}
+
+////////////////////////////////////////////////////////////////////////////////
 // Create a 1-bit clip mask in the stencil buffer.
 
 bool GrReducedClip::drawStencilClipMask(GrContext* context,
diff --git a/src/gpu/GrReducedClip.h b/src/gpu/GrReducedClip.h
index 5447652137..c3838d8802 100644
--- a/src/gpu/GrReducedClip.h
+++ b/src/gpu/GrReducedClip.h
@@ -83,6 +83,7 @@ public:
      */
     bool maskRequiresAA() const { SkASSERT(!fMaskElements.isEmpty()); return fMaskRequiresAA; }
 
+    bool drawAlphaClipMask(GrRenderTargetContext*) const;
     bool drawStencilClipMask(GrContext*, GrRenderTargetContext*) const;
 
     int numAnalyticFPs() const { return fAnalyticFPs.count() + fCCPRClipPaths.count(); }