renderer/src/rive_renderer.cpp

/*
 * Copyright 2022 Rive
 */

#include "rive/renderer/rive_renderer.hpp"

#include "gradient.hpp"
#include "rive_render_path.hpp"
#include "rive/math/math_types.hpp"
#include "rive/math/simd.hpp"
#include "rive/renderer/rive_render_image.hpp"
#include "rive/renderer/rive_render_paint.hpp"
#include "rive/profiler/profiler_macros.h"

namespace rive
{
static rcp<RiveRenderPath> invertClockwisePath(
    const RiveRenderPath* path,
    FillRule pathFillRule,
    const Mat2D& viewMatrix,
    IAABB bounds);

bool RiveRenderer::IsAABB(const RawPath& path, AABB* result)
{
  RIVE_PROF_SCOPE_L(3)
  // Any quadrilateral begins with a move plus 3 lines.
  constexpr static size_t kAABBVerbCount = 4;
  constexpr static PathVerb aabbVerbs[kAABBVerbCount] =
      {PathVerb::move, PathVerb::line, PathVerb::line, PathVerb::line};
  Span<const PathVerb> verbs = path.verbs();
  if (verbs.count() < kAABBVerbCount ||
      memcmp(verbs.data(), aabbVerbs, sizeof(aabbVerbs)) != 0)
  {
    return false;
  }

  // Only accept extra verbs and points if every point after the quadrilateral
  // is equal to p0.
  Span<const Vec2D> pts = path.points();
  for (size_t i = 4; i < pts.count(); ++i)
  {
    if (pts[i] != pts[0])
    {
      return false;
    }
  }

  // We have a quadrilateral! Now check if it is an axis-aligned rectangle.
  float4 corners = {pts[0].x, pts[0].y, pts[2].x, pts[2].y};
  float4 oppositeCorners = {pts[1].x, pts[1].y, pts[3].x, pts[3].y};
  if (simd::all(corners == oppositeCorners.zyxw) ||
      simd::all(corners == oppositeCorners.xwzy))
  {
    float4 r = simd::join(
        simd::min(corners.xy, corners.zw),
        simd::max(corners.xy, corners.zw));
    simd::store(result, r);
    return true;
  }
  return false;
}

RiveRenderer::ClipElement::ClipElement(
    const Mat2D& matrix_,
    const RiveRenderPath* path_,
    FillRule fillRule_,
    IAABB pixelBounds_,
    std::optional<StrokeParams> stroke_,
    float feather_,
    bool forceClosed_)
{
  reset(
      matrix_,
      path_,
      fillRule_,
      pixelBounds_,
      stroke_,
      feather_,
      forceClosed_);
}

RiveRenderer::ClipElement::~ClipElement() {}

void RiveRenderer::ClipElement::reset(
    const Mat2D& matrix_,
    const RiveRenderPath* path_,
    FillRule fillRule_,
    IAABB pixelBounds_,
    std::optional<StrokeParams> stroke_,
    float feather_,
    bool forceClosed_)
{
  matrix = matrix_;
  rawPathMutationID = path_->getRawPathMutationID();
  pathBounds = path_->getBounds();
  path = ref_rcp(path_);
  fillRule = fillRule_;
  clipID = 0; // This gets initialized lazily.
  pixelBounds = pixelBounds_;
  stroke = stroke_;
  feather = feather_;
  forceClosed = forceClosed_;
}

bool RiveRenderer::ClipElement::isEquivalent(
    const Mat2D& matrix_,
    const RiveRenderPath* path_,
    std::optional<StrokeParams> stroke_,
    float feather_,
    bool forceClosed_) const
{
  if (stroke.has_value())
  {
    if (!stroke_.has_value() || *stroke != *stroke_)
    {
      return false;
    }
  }
  else if (stroke_.has_value())
  {
    return false;
  }

  return matrix_ == matrix &&
         path_->getRawPathMutationID() == rawPathMutationID &&
         path_->getFillRule() == fillRule && feather == feather_ &&
         forceClosed == forceClosed_;
}

RiveRenderer::RiveRenderer(gpu::RenderContext* context) : m_context(context) {}

RiveRenderer::~RiveRenderer() {}

void RiveRenderer::save()
{
  // Copy the back of the stack before pushing, in case the vector grows and
  // invalidates the reference.
  RenderState copy = m_renderStateStack.back();
  m_renderStateStack.push_back(copy);
}

void RiveRenderer::restore()
{
  assert(m_renderStateStack.size() > 1);
  assert(
      m_renderStateStack.back().clipStackHeight >=
      m_renderStateStack[m_renderStateStack.size() - 2].clipStackHeight);
  m_renderStateStack.pop_back();
}

void RiveRenderer::transform(const Mat2D& matrix)
{
  m_renderStateStack.back().matrix = m_renderStateStack.back().matrix * matrix;
}

void RiveRenderer::modulateOpacity(float opacity)
{
  m_renderStateStack.back().modulatedOpacity = std::max(
      0.0f,
      m_renderStateStack.back().modulatedOpacity * opacity);
}

ColorInt RiveRenderer::modulated(ColorInt color, float opacity) const
{
  return colorModulate(
      color,
      m_renderStateStack.back().modulatedColor,
      opacity);
}

void RiveRenderer::modulateColor(ColorInt color, bool replace)
{
  RenderState& state = m_renderStateStack.back();
  state.modulatedColor = replace ? color
                                 : colorModulate(state.modulatedColor, color);
}

void RiveRenderer::drawPath(RenderPath* renderPath, RenderPaint* renderPaint)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(path, RiveRenderPath*, renderPath);
  LITE_RTTI_CAST_OR_RETURN(paint, RiveRenderPaint*, renderPaint);

  if (path->getRawPath().empty())
  {
    return;
  }

  if (paint->getIsStroked() && m_context->frameDescriptor().strokesDisabled)
  {
    return;
  }
  if (!paint->getIsStroked() && m_context->frameDescriptor().fillsDisabled)
  {
    return;
  }
  if (paint->getIsStroked() &&
      // Use inverse logic to ensure we abort when stroke thickness is NaN.
      !(paint->getThickness() > 0))
  {
    return;
  }
  // Use inverse logic to ensure we abort when stroke thickness is NaN.
  if (!(paint->getFeather() >= 0))
  {
    return;
  }
  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  if (paint->getIsStroked() &&
      paint->getStrokePosition() != StrokePosition::center)
  {
    save();

    auto stroke = paint->getStrokeParams();
    stroke.thickness *= 2.0f;
    stroke.position = StrokePosition::center;

    if (paint->getFeather() == 0.0f)
    {
      // Without a feather, we can clip by the stroke, with double its
      // current thickness, then draw path into it.

      IAABB pixelBounds;
      clipPathImpl(
          path,
          stroke,
          paint->getFeather(),
          ForceClosed::yes,
          &pixelBounds);

      // Make a version of the paint that is the same as the current one,
      // except as a fill instead of stroke.
      auto fillPaint = paint->clone();
      fillPaint->style(RenderPaintStyle::fill);

      if (paint->getStrokePosition() == StrokePosition::inside)
      {
        drawPath(renderPath, fillPaint.get());
      }
      else
      {
        assert(paint->getStrokePosition() == StrokePosition::outside);

        if (m_context->frameInterlockMode() != gpu::InterlockMode::depthStencil)
        {
          // For non-depthStencil modes, there will be a gap between this outer
          // stroke and the same shape if filled normally, so expand this paint
          // by a full texel to cover the gap.
          fillPaint->additionalOutset(1.0f);
        }

        // For outside strokes we need to fill the *inverse* of the
        // path.
        auto fillPath = invertClockwisePath(
            path,
            path->getFillRule(),
            m_renderStateStack.back().matrix,
            pixelBounds);
        drawPath(fillPath.get(), fillPaint.get());
      }
    }
    else
    {
      // For a feathered stroke we need to invert this: clip by the path
      // (inverted for outer), then draw the (feathered) stroke

      // TODO: A possible optimization here when we have stencil is to
      // start with drawing the stroke, expanded by the feather amount (as
      // thickness) into stencil. For large paths this would eliminate a
      // lot of fill rate through the middle.
      if (paint->getStrokePosition() == StrokePosition::inside)
      {
        clipPathImpl(path);
      }
      else
      {
        auto strokeBounds = path->calculatePixelBounds(
            m_renderStateStack.back().matrix,
            stroke,
            paint->getFeather());
        auto clip = invertClockwisePath(
            path,
            path->getFillRule(),
            m_renderStateStack.back().matrix,
            strokeBounds);
        clipPathImpl(clip.get());
      }

      // Make a paint that has a centered stroke that's twice as thick,
      // and force it to draw the stroke as closed.
      auto thickenedCenteredStrokePaint = paint->clone();
      thickenedCenteredStrokePaint->stroke(stroke);
      thickenedCenteredStrokePaint->forceClosed(true);
      drawPath(path, thickenedCenteredStrokePaint.get());
    }
    restore();
    return;
  }

  Mat2D imageMatrix;
  Mat2D* imageMatrixPtr = nullptr;
  if (paint->getImageTexture() != nullptr)
  {
    if (!m_context->frameSupportsImagePaintForPaths())
    {
      // If we don't have image paint support we need to do this using
      // ImageRect (which can do all of the drawing of image +
      // color/gradient), clipped to the current path.
      save();

      AABB bounds;
      if (!IsAABB(path->getRawPath(), &bounds) || paint->getIsStroked() ||
          paint->getFeather() != 0.0f)
      {
        // If this is not an AABB directly we need to get the actual
        // bounds and then clip against the path.
        bounds = path->getBounds();

        std::optional<StrokeParams> stroke;
        if (paint->getIsStroked())
        {
          // We should only be here with centered strokes, since
          // inner/outer strokes are handled above (with an early
          // return)
          assert(paint->getStrokePosition() == StrokePosition::center);
          stroke = {
              .thickness = paint->getThickness(),
              .join = paint->getJoin(),
              .cap = paint->getCap(),
          };
        }

        float outset = RiveRenderPath::calculateBoundsOutset(
            stroke,
            paint->getFeather());
        bounds = bounds.outset(outset, outset);
        clipPathImpl(path, stroke, paint->getFeather());
      }

      const auto gradientMatrix = m_renderStateStack.back().matrix;

      // ImageRectDraw draws as a unit square with upper-left corner of 0,
      // 0 so we need to adjust our transform to set that up.
      Mat2D adjust = Mat2D::fromScaleAndTranslation(
          bounds.width(),
          bounds.height(),
          bounds.left(),
          bounds.top());
      transform(adjust);

      const auto& m = m_renderStateStack.back().matrix;

      // The image matrix needs to map from the desired image space to the
      // "box space" (where the upper-left and lower-right coordinates are
      // (0, 0) and (1, 1) respectively). This is effectively the inverse
      // of undoing the adjust matrix then applying the paint's
      // imageMatrix, which becomes the following:
      const auto imageMatrix = paint->getImageTransform().invertOrIdentity() *
                               adjust;

      ColorInt paintColor = paint->getColor();
      if (paint->getGradient() != nullptr)
      {
        // Paints with gradients have no color data so use solid white.
        paintColor = 0xFFFFFFFF;
      }

      clipAndPushDraw(gpu::DrawUniquePtr(m_context->make<gpu::ImageRectDraw>(
          m_context,
          m.mapBoundingBox(AABB{0, 0, 1, 1}).roundOut(),
          m,
          paint->getBlendMode(),
          paint->getAdditiveness(),
          ref_rcp(paint->getImageTexture()),
          ref_rcp(paint->getGradient()),
          paint->getImageSampler(),
          modulated(paintColor, m_renderStateStack.back().modulatedOpacity),
          imageMatrix,
          gradientMatrix,
          paint->getInverseGradientTransform())));

      restore();
      return;
    }

    imageMatrix = m_renderStateStack.back().matrix * paint->getImageTransform();
    imageMatrixPtr = &imageMatrix;
  }

  if (paint->getFeather() != 0 && !paint->getIsStroked())
  {
    if (path->getFillRule() != FillRule::clockwise &&
        !m_context->frameDescriptor().clockwiseFillOverride)
    {
      // Don't draw feathered fills that aren't clockwise.
      return;
    }
    float matrixMaxScale = m_renderStateStack.back().matrix.findMaxScale();
    if (paint->getFeather() * matrixMaxScale > 1)
    {
      clipAndPushDraw(gpu::PathDraw::Make(
          m_context,
          m_renderStateStack.back().matrix,
          imageMatrixPtr,
          path->makeSoftenedCopyForFeathering(
              paint->getFeather(),
              matrixMaxScale),
          path->getFillRule(),
          paint,
          m_renderStateStack.back().modulatedOpacity,
          m_renderStateStack.back().modulatedColor));
      return;
    }
  }

  clipAndPushDraw(gpu::PathDraw::Make(
      m_context,
      m_renderStateStack.back().matrix,
      imageMatrixPtr,
      ref_rcp(path),
      path->getFillRule(),
      paint,
      m_renderStateStack.back().modulatedOpacity,
      m_renderStateStack.back().modulatedColor));
}

void RiveRenderer::clipPath(RenderPath* renderPath)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(path, RiveRenderPath*, renderPath);

  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  if (path->getRawPath().empty())
  {
    m_renderStateStack.back().overallClipPixelBounds = {};
    return;
  }

  // First try to handle axis-aligned rectangles using the "ENABLE_CLIP_RECT"
  // shader feature. Multiple axis-aligned rectangles can be intersected into
  // a single rectangle if their matrices are compatible.
  AABB clipRectCandidate;
  if (m_context->frameSupportsClipRects() &&
      IsAABB(path->getRawPath(), &clipRectCandidate))
  {
    clipRectImpl(clipRectCandidate, path);
  }
  else
  {
    clipPathImpl(path);
  }
}

void RiveRenderer::clipStroke(
    RenderPath* renderPath,
    const StrokeParams& params)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(path, RiveRenderPath*, renderPath);

  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  if (path->getRawPath().empty())
  {
    m_renderStateStack.back().overallClipPixelBounds = {};
    return;
  }

  // For centered strokes we can just clip against the path.
  clipPathImpl(path, params);
}

// Finds a new rect, if such a rect exists, such that:
//
//     currentMatrix * rect == newMatrix * newRect
//
// Returns true if *rect was replaced with newRect.
static bool transform_rect_to_new_space(
    AABB* rect,
    const Mat2D& currentMatrix,
    const Mat2D& newMatrix)
{
  if (currentMatrix == newMatrix)
  {
    return true;
  }
  Mat2D currentToNew;
  if (!newMatrix.invert(&currentToNew))
  {
    return false;
  }
  currentToNew = currentToNew * currentMatrix;
  float maxSkew = fmaxf(fabsf(currentToNew.xy()), fabsf(currentToNew.yx()));
  float maxScale = fmaxf(fabsf(currentToNew.xx()), fabsf(currentToNew.yy()));
  if (maxSkew > math::EPSILON && maxScale > math::EPSILON)
  {
    // Transforming this rect to the new view matrix would turn it into
    // something that isn't a rect.
    return false;
  }
  Vec2D pts[2] = {{rect->left(), rect->top()}, {rect->right(), rect->bottom()}};
  currentToNew.mapPoints(pts, pts, 2);
  float4 p = simd::load4f(pts);
  float2 topLeft = simd::min(p.xy, p.zw);
  float2 botRight = simd::max(p.xy, p.zw);
  *rect = {topLeft.x, topLeft.y, botRight.x, botRight.y};
  return true;
}

void RiveRenderer::clipRectImpl(AABB rect, const RiveRenderPath* originalPath)
{
  RIVE_PROF_SCOPE_L(3)

  auto& renderState = m_renderStateStack.back();
  bool hasClipRect = renderState.clipRectInverseMatrix != nullptr;
  if (rect.isEmptyOrNaN())
  {
    renderState.overallClipPixelBounds = {};
    return;
  }

  // If there already is a clipRect, we can only accept another one by
  // intersecting it with the existing one. This means the new rect must be
  // axis-aligned with the existing clipRect.
  if (hasClipRect && !transform_rect_to_new_space(
                         &rect,
                         renderState.matrix,
                         renderState.clipRectMatrix))
  {
    // 'rect' is not axis-aligned with the existing clipRect. Fall back to
    // clipPath.
    clipPathImpl(originalPath);
    return;
  }

  if (!hasClipRect)
  {
    // There wasn't an existing clipRect. This is the one!
    renderState.clipRect = rect;
    renderState.clipRectMatrix = renderState.matrix;
  }
  else
  {
    // Both rects are in the same space now. Intersect the two
    // geometrically.
    float4 a = simd::load4f(&renderState.clipRect);
    float4 b = simd::load4f(&rect);
    float4 intersection = simd::join(
        simd::max(a.xy, b.xy),
        simd::min(a.zw, b.zw));
    simd::store(&renderState.clipRect, intersection);
  }

  // Grab the pixel bounds of the new combined (intersected) clip rect
  renderState.clipRectPixelBounds = renderState.clipRectMatrix
                                        .mapBoundingBox(renderState.clipRect)
                                        .roundOut();

  renderState.overallClipPixelBounds = renderState.overallClipPixelBounds
                                           .intersect(
                                               renderState.clipRectPixelBounds);

  renderState.clipRectInverseMatrix = m_context
                                          ->make<gpu::ClipRectInverseMatrix>(
                                              renderState.clipRectMatrix,
                                              renderState.clipRect);
}

void RiveRenderer::clipPathImpl(
    const RiveRenderPath* path,
    std::optional<StrokeParams> stroke,
    float feather,
    ForceClosed forceClosed,
    IAABB* boundsOut)
{
  RIVE_PROF_SCOPE_L(3)
  auto& renderState = m_renderStateStack.back();
  if (path->getBounds().isEmptyOrNaN())
  {
    if (boundsOut != nullptr)
    {
      *boundsOut = IAABB{};
    }

    renderState.overallClipPixelBounds = {};
    return;
  }

  if (stroke.has_value() && stroke->position != StrokePosition::center)
  {
    // To handle an inner or outer stroke it's actually two clips: we need
    // to clip by the stroke then *additionally* the path by itself (which
    // needs to be inverted for outside strokes)
    auto clipStroke = *stroke;
    clipStroke.thickness *= 2.0f;
    clipStroke.position = StrokePosition::center;
    IAABB outerBounds;
    clipPathImpl(path, clipStroke, 0.0f, ForceClosed::yes, &outerBounds);

    if (boundsOut != nullptr)
    {
      *boundsOut = outerBounds;
    }

    if (stroke->position == StrokePosition::inside)
    {
      clipPathImpl(path);
    }
    else
    {
      // TODO: There's an optimization that could be done here in some
      // render modes where instead of inverting the path we could instead
      // render them "subtractively" (in depthStencil this would be
      // rendering with a "clear the upper bit" flag)
      assert(stroke->position == StrokePosition::outside);
      auto inverted = invertClockwisePath(
          path,
          path->getFillRule(),
          m_renderStateStack.back().matrix,
          outerBounds);
      clipPath(inverted.get());
    }

    return;
  }

  // Only write a new clip element if this path isn't already on the stack
  // from before. e.g.:
  //
  //     clipPath(samePath);
  //     restore();
  //     save();
  //     clipPath(samePath); // <-- reuse the ClipElement (and clipID!)
  //     already in m_clipStack.
  //
  const size_t clipStackHeight = renderState.clipStackHeight;
  assert(m_clipStack.size() >= clipStackHeight);
  if (m_clipStack.size() == clipStackHeight ||
      !m_clipStack[clipStackHeight].isEquivalent(
          renderState.matrix,
          path,
          stroke,
          feather,
          bool(forceClosed)))
  {
    // Calculate the pixel bounds for this clip path before we push it into
    // the stack to ensure that we even need to do so
    const auto pixelBounds = path->calculatePixelBounds(
        renderState.matrix,
        stroke,
        feather);
    renderState.overallClipPixelBounds = renderState.overallClipPixelBounds
                                             .intersect(pixelBounds);

    if (boundsOut != nullptr)
    {
      *boundsOut = pixelBounds;
    }
    if (renderState.overallClipPixelBounds.empty())
    {
      // Nothing can draw under this, so no need to add to the stack.
      return;
    }

    m_clipStack.resize(clipStackHeight);
    m_clipStack.emplace_back(
        renderState.matrix,
        path,
        path->getFillRule(),
        pixelBounds,
        stroke,
        feather,
        bool(forceClosed));
  }
  else
  {
    if (boundsOut != nullptr)
    {
      *boundsOut = m_clipStack[clipStackHeight].pixelBounds;
    }

    // We are going to reuse the element that is already in the clip stack,
    // but need to re-update the overall clip pixel bounds.
    renderState
        .overallClipPixelBounds = renderState.overallClipPixelBounds.intersect(
        m_clipStack[clipStackHeight].pixelBounds);

    if (renderState.overallClipPixelBounds.empty())
    {
      // Nothing can draw under this, so no need to increment the stack
      // height.
      return;
    }
  }

  renderState.clipStackHeight = clipStackHeight + 1;
}

void RiveRenderer::drawImage(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    BlendMode blendMode,
    float opacity)
{
  drawImage(renderImage, imageSampler, blendMode, opacity, 0);
}

void RiveRenderer::drawImage(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    BlendMode blendMode,
    float opacity,
    float additiveness)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(image, const RiveRenderImage*, renderImage);

  // ImageRectDraw below takes the blend mode directly rather than through a
  // RiveRenderPaint, so fold additive here too.
  blendMode = foldAdditiveToSrcOver(blendMode);

  rcp<gpu::Texture> imageTexture = image->refTexture();
  if (imageTexture == nullptr)
  {
    // imageTexture may be null if the backend uses a custom factory and/or
    // updates out-of-band assets asynchronously. If there's no texture yet,
    // just don't draw anything.
    return;
  }

  // Apply modulated opacity (clamp to prevent negative values)
  float finalOpacity = std::max(
      0.0f,
      opacity * m_renderStateStack.back().modulatedOpacity);

  // Scale the view matrix so we can draw this image as the rect [0, 0, 1, 1].
  save();
  scale(image->width(), image->height());

  if (!m_context->frameSupportsImagePaintForPaths())
  {
    // Fall back on ImageRectDraw if the current frame doesn't support
    // drawing paths with image paints.
    if (!m_renderStateStack.back().overallClipPixelBounds.empty())
    {
      const Mat2D& m = m_renderStateStack.back().matrix;
      clipAndPushDraw(gpu::DrawUniquePtr(m_context->make<gpu::ImageRectDraw>(
          m_context,
          m.mapBoundingBox(AABB{0, 0, 1, 1}).roundOut(),
          m,
          blendMode,
          additiveness,
          std::move(imageTexture),
          nullptr, // gradient
          imageSampler,
          modulated(0xFFFFFFFF, finalOpacity),
          Mat2D{},    // imageMatrix
          Mat2D{},    // gradientMatrix
          Mat2D{}))); // inverseGradientTransform
    }
  }
  else
  {
    // Implement drawImage() as drawPath() with a rectangular path and an
    // image paint.
    if (m_unitRectPath == nullptr)
    {
      m_unitRectPath = make_rcp<RiveRenderPath>();
      m_unitRectPath->line({1, 0});
      m_unitRectPath->line({1, 1});
      m_unitRectPath->line({0, 1});
    }

    RiveRenderPaint paint;
    paint.image(std::move(imageTexture), finalOpacity);
    paint.blendMode(blendMode);
    paint.additiveness(additiveness);
    paint.imageSampler(imageSampler);
    drawPath(m_unitRectPath.get(), &paint);
  }

  restore();
}

void RiveRenderer::applyLayerMask(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    LayerMaskMode mode)
{
  RIVE_PROF_SCOPE_L(2)
  if (!m_context->frameSupportsLayerMask())
  {
    // Leaves the layer unmasked. The alternative on a mode whose blend step
    // has not been taught the op is worse than unmasked: find_paint_color
    // would hand back the coverage texel and src-over would paint it over
    // the content.
    return;
  }
  LITE_RTTI_CAST_OR_RETURN(image, const RiveRenderImage*, renderImage);
  rcp<gpu::Texture> maskTexture = image->refTexture();
  if (maskTexture == nullptr)
  {
    return;
  }
  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  // Same shape as the image-paint branch of drawImage: a unit rect scaled to
  // the mask's pixel extent. The caller has already put the renderer in the
  // canvas's own space, so this covers exactly the target.
  if (m_unitRectPath == nullptr)
  {
    m_unitRectPath = make_rcp<RiveRenderPath>();
    m_unitRectPath->line({1, 0});
    m_unitRectPath->line({1, 1});
    m_unitRectPath->line({0, 1});
  }

  save();
  scale(
      static_cast<float>(image->width()),
      static_cast<float>(image->height()));
  RiveRenderPaint paint;
  paint.layerMask(std::move(maskTexture), mode);
  paint.imageSampler(imageSampler);
  drawPath(m_unitRectPath.get(), &paint);
  restore();
}

void RiveRenderer::drawImageMesh(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    rcp<RenderBuffer> vertices_f32,
    rcp<RenderBuffer> uvCoords_f32,
    rcp<RenderBuffer> indices_u16,
    uint32_t vertexCount,
    uint32_t indexCount,
    BlendMode blendMode,
    float opacity)
{
  drawImageMesh(
      renderImage,
      imageSampler,
      std::move(vertices_f32),
      std::move(uvCoords_f32),
      std::move(indices_u16),
      vertexCount,
      indexCount,
      blendMode,
      opacity,
      0);
}

void RiveRenderer::drawImageMesh(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    rcp<RenderBuffer> vertices_f32,
    rcp<RenderBuffer> uvCoords_f32,
    rcp<RenderBuffer> indices_u16,
    uint32_t vertexCount,
    uint32_t indexCount,
    BlendMode blendMode,
    float opacity,
    float additiveness)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(image, const RiveRenderImage*, renderImage);

  // ImageMeshDraw below takes the blend mode directly rather than through a
  // RiveRenderPaint, so fold additive here too.
  blendMode = foldAdditiveToSrcOver(blendMode);

  rcp<gpu::Texture> imageTexture = image->refTexture();
  if (imageTexture == nullptr)
  {
    // imageTexture may be null if the backend uses a custom factory and/or
    // updates out-of-band assets asynchronously. If there's no texture yet,
    // just don't draw anything.
    return;
  }

  assert(vertices_f32);
  assert(uvCoords_f32);
  assert(indices_u16);

  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  // Apply modulated opacity (clamp to prevent negative values)
  float finalOpacity = std::max(
      0.0f,
      opacity * m_renderStateStack.back().modulatedOpacity);

  clipAndPushDraw(gpu::DrawUniquePtr(m_context->make<gpu::ImageMeshDraw>(
      gpu::Draw::FULLSCREEN_PIXEL_BOUNDS,
      m_renderStateStack.back().matrix,
      blendMode,
      additiveness,
      std::move(imageTexture),
      imageSampler,
      std::move(vertices_f32),
      std::move(uvCoords_f32),
      std::move(indices_u16),
      indexCount,
      modulated(0xFFFFFFFF, finalOpacity),
      Vec2D{0.0f, 0.0f},
      Vec2D{1.0f, 1.0f})));
}

void RiveRenderer::drawImageMeshInstanced(
    const RenderImage* renderImage,
    ImageSampler imageSampler,
    rcp<RenderBuffer> vertices_f32,
    rcp<RenderBuffer> uvCoords_f32,
    rcp<RenderBuffer> indices_u16,
    uint32_t vertexCount,
    uint32_t indexCount,
    rcp<ImageMeshInstances> instances)
{
  RIVE_PROF_SCOPE_L(2)
  LITE_RTTI_CAST_OR_RETURN(image, const RiveRenderImage*, renderImage);

  rcp<gpu::Texture> imageTexture = image->refTexture();
  if (imageTexture == nullptr)
  {
    // imageTexture may be null if the backend uses a custom factory and/or
    // updates out-of-band assets asynchronously. If there's no texture yet,
    // just don't draw anything.
    return;
  }

  assert(vertices_f32);
  assert(uvCoords_f32);
  assert(indices_u16);

  if (instances == nullptr || instances->count() == 0)
  {
    return;
  }

  if (m_renderStateStack.back().overallClipPixelBounds.empty())
  {
    return;
  }

  if (m_context->frameInterlockMode() == gpu::InterlockMode::atomics)
  {
    // If we're in atomic mode, meshes can't self-overlap, so we can't
    // safely issue instanced calls without barriers since the individual
    // instances may overlap. As a workaround, issue individual draw calls
    // for each instance instead. This is slow, and I'm only doing it this
    // way because atomic mode is soon going away.
    for (const auto& instanceData : instances->instanceData())
    {
      float finalOpacity = std::max(
          0.0f,
          instanceData.opacity * m_renderStateStack.back().modulatedOpacity);

      clipAndPushDraw(gpu::DrawUniquePtr(m_context->make<gpu::ImageMeshDraw>(
          gpu::Draw::FULLSCREEN_PIXEL_BOUNDS,
          m_renderStateStack.back().matrix * instanceData.transform,
          BlendMode::srcOver,
          instanceData.additiveness,
          imageTexture,
          imageSampler,
          vertices_f32,
          uvCoords_f32,
          indices_u16,
          indexCount,
          modulated(0xFFFFFFFF, finalOpacity),
          instanceData.uvTranslate,
          instanceData.uvScale)));
    }
  }
  else
  {
    // Unlike in drawImageMesh, we wait to resolve the modulated
    // color/opacity because each instance has its own opacity, so we avoid
    // precision loss if we do the float -> u8 conversion after it's been
    // multiplied in.
    clipAndPushDraw(
        gpu::DrawUniquePtr(m_context->make<gpu::ImageMeshInstancedDraw>(
            gpu::Draw::FULLSCREEN_PIXEL_BOUNDS,
            m_renderStateStack.back().matrix,
            std::move(imageTexture),
            imageSampler,
            std::move(vertices_f32),
            std::move(uvCoords_f32),
            std::move(indices_u16),
            indexCount,
            std::move(instances),
            m_renderStateStack.back().modulatedColor,
            m_renderStateStack.back().modulatedOpacity)));
  }
}

void RiveRenderer::clipAndPushDraw(gpu::DrawUniquePtr draw)
{
  RIVE_PROF_SCOPE_L(3)
  assert(!m_renderStateStack.back().overallClipPixelBounds.empty());
  if (draw.get() == nullptr)
  {
    return;
  }
  if (m_context->isOutsideCurrentFrame(draw->pixelBounds()))
  {
    return;
  }

  // Make two attempts to issue the draw: once on the context as-is and once
  // with a clean flush.
  for (int i = 0; i < 2; ++i)
  {
    // Always make sure we begin this loop with the internal draw batch
    // empty, and clear it when we're done.
    struct AutoResetInternalDrawBatch
    {
    public:
      AutoResetInternalDrawBatch(RiveRenderer* renderer) : m_renderer(renderer)
      {
        assert(m_renderer->m_internalDrawBatch.empty());
      }
      ~AutoResetInternalDrawBatch() { m_renderer->m_internalDrawBatch.clear(); }

    private:
      RiveRenderer* m_renderer;
    };

    AutoResetInternalDrawBatch aridb(this);

    auto applyClipResult = applyClip(draw.get());
    if (applyClipResult == ApplyClipResult::failure)
    {
      // There wasn't room in the GPU buffers for this path draw. Flush
      // and try again.
      m_context->logicalFlush();
      continue;
    }
    else if (applyClipResult == ApplyClipResult::fullyClipped)
    {
      return;
    }

    m_internalDrawBatch.push_back(std::move(draw));
    if (!m_context->pushDraws(
            m_internalDrawBatch.data(),
            m_internalDrawBatch.size()))
    {
      // There wasn't room in the GPU buffers for this path draw. Flush
      // and try again.
      m_context->logicalFlush();
      // Reclaim "draw" because we will use it again on the next
      // iteration.
      draw = std::move(m_internalDrawBatch.back());
      assert(draw != nullptr);
      m_internalDrawBatch.pop_back();
      continue;
    }

    // Success!
    return;
  }

  // We failed to process the draw. Release its refs.
  fprintf(
      stderr,
      "RiveRenderer::clipAndPushDraw failed. The draw and/or clip stack "
      "are too complex.\n");
}

// Used by clipping in clockwiseAtomic mode, as well as outer strokes.
//
// Returns the inverse of a path, meaning, regions that were filled in the old
// path are now empty, and empty regions in the old path are now filled.
//
// NOTE: A true inverse path would expand infinitely in all directions, but this
// function limits it by the provided "bounds".
//
// NOTE: The returned path is always clockwise, even if the given path was not.
// If the given path is not clockwise, we attempt to convert it to clockwise
// based on its dominant winding direction. This may or may not be accurate.
static rcp<RiveRenderPath> invertClockwisePath(
    const RiveRenderPath* path,
    FillRule pathFillRule,
    const Mat2D& viewMatrix,
    IAABB bounds)
{
  auto inversePath = make_rcp<RiveRenderPath>();
  inversePath->fillRule(FillRule::clockwise);
  Mat2D viewInverseMatrix;
  if (viewMatrix.invert(&viewInverseMatrix))
  {
    // Add the pre-viewMatrix "bounds" rect to the new path.
    std::array<Vec2D, 4> boundsVertices = {
        Vec2D(bounds.left, bounds.top),
        Vec2D(bounds.right, bounds.top),
        Vec2D(bounds.right, bounds.bottom),
        Vec2D(bounds.left, bounds.bottom),
    };
    viewInverseMatrix.mapPoints(
        boundsVertices.data(),
        boundsVertices.data(),
        4);
    inversePath->move(boundsVertices[0]);
    if (const float viewMatrixDeterminant = viewMatrix[0] * viewMatrix[3] -
                                            viewMatrix[2] * viewMatrix[1];
        viewMatrixDeterminant >= 0)
    {
      inversePath->line(boundsVertices[1]);
      inversePath->line(boundsVertices[2]);
      inversePath->line(boundsVertices[3]);
    }
    else
    {
      inversePath->line(boundsVertices[3]);
      inversePath->line(boundsVertices[2]);
      inversePath->line(boundsVertices[1]);
    }
    // Subtract the given path out of the bounds rect.
    if (pathFillRule == FillRule::clockwise || path->getCoarseArea() >= 0)
    {
      inversePath->addRenderPathBackwards(path, Mat2D());
    }
    else
    {
      inversePath->addRenderPath(path, Mat2D());
    }
  }
  return inversePath;
}

RiveRenderer::ApplyClipResult RiveRenderer::applyClip(gpu::Draw* draw)
{
  RIVE_PROF_SCOPE_L(3)
  auto& renderState = m_renderStateStack.back();

  draw->setClipRect(
      renderState.clipRectInverseMatrix,
      renderState.overallClipPixelBounds);

  if (draw->clippedPixelBounds().empty() ||
      m_context->isOutsideCurrentFrame(draw->clippedPixelBounds()))
  {
    // Either this draw is outside of the current frame (in which case it's
    // not visible) or it was completely clipped by the current overall clip
    // bounds.
    return ApplyClipResult::fullyClipped;
  }

  const size_t clipStackHeight = renderState.clipStackHeight;
  if (clipStackHeight == 0)
  {
    assert(draw->clipID() == 0);
    return ApplyClipResult::success;
  }

  // Find which clip element in the stack (if any) is currently rendered to
  // the clip buffer.
  size_t clipIdxCurrentlyInClipBuffer = -1; // i.e., "none".
  if (m_context->getClipContentID() != 0)
  {
    for (size_t i = clipStackHeight - 1; i != -1; --i)
    {
      if (m_clipStack[i].clipID == m_context->getClipContentID())
      {
        clipIdxCurrentlyInClipBuffer = i;
        break;
      }
    }
  }

  // Draw the necessary updates to the clip buffer (i.e., draw every clip
  // element after clipIdxCurrentlyInClipBuffer).
  uint32_t parentClipID = clipIdxCurrentlyInClipBuffer == -1
                              ? 0 // The next clip to be drawn is not nested.
                              : m_clipStack[clipIdxCurrentlyInClipBuffer]
                                    .clipID;
  if (m_context->frameInterlockMode() == gpu::InterlockMode::clockwiseAtomic ||
      m_context->frameInterlockMode() == gpu::InterlockMode::depthStencil)
  {
    if (parentClipID == 0 && m_context->getClipContentID() != 0)
    {
      // Time for a new stencil clip! Erase the clip currently in the
      // stencil buffer before we draw the new one.
      auto stencilClipClear = gpu::DrawUniquePtr(
          m_context->make<gpu::ClipReset>(
              m_context,
              m_context->getClipContentID(),
              gpu::DrawContents::none,
              gpu::ClipReset::ResetAction::clearPreviousClip));
      if (!m_context->isOutsideCurrentFrame(stencilClipClear->pixelBounds()))
      {
        m_internalDrawBatch.push_back(std::move(stencilClipClear));
      }
    }
  }

  for (size_t i = clipIdxCurrentlyInClipBuffer + 1; i < clipStackHeight; ++i)
  {
    ClipElement& clip = m_clipStack[i];
    assert(clip.pathBounds == clip.path->getBounds());

    IAABB clipDrawBounds;
    RiveRenderPaint clipUpdatePaint;
    clipUpdatePaint.clipUpdate(
        /*clip THIS clipDraw against:*/ parentClipID);
    clipUpdatePaint.feather(clip.feather);
    if (clip.stroke.has_value())
    {
      clipUpdatePaint.stroke(*clip.stroke);
    }

    clipUpdatePaint.forceClosed(clip.forceClosed);

    rcp clipPath = clip.path;
    FillRule clipFillRule = clip.fillRule;
    std::optional pixelBounds = clip.pixelBounds;

    if (m_context->frameInterlockMode() ==
            gpu::InterlockMode::clockwiseAtomic &&
        parentClipID != 0)
    {
      // clockwiseAtomic implements nested clips by erasing the inverse
      // of the inner path from the outer clip.
      clipPath = invertClockwisePath(
          clipPath.get(),
          clipFillRule,
          clip.matrix,
          m_context->getClipContentBounds(parentClipID));
      clipFillRule = FillRule::clockwise;

      // Clear this because the inverted path has different pixel bounds
      // (as it now contains the clip content bounds as a box around the
      // path)
      pixelBounds = std::nullopt;
    }

    gpu::DrawUniquePtr clipDraw = gpu::PathDraw::Make(
        m_context,
        clip.matrix,
        nullptr, // imageMatrix, unneeded for clips
        std::move(clipPath),
        clipFillRule,
        &clipUpdatePaint,
        1.0f, // no modulation for clips
        0xFFFFFFFF,
        pixelBounds);

    // We have already validated that the clip path is within the screen
    // bounds, so null means the clip path is empty and clips out everything.
    if (clipDraw == nullptr)
    {
      return ApplyClipResult::fullyClipped;
    }

    clipDrawBounds = clipDraw->pixelBounds();

    // Generate a new clipID every time we (re-)render an element to the
    // clip buffer. (Each embodiment of the element needs its own
    // separate readBounds.)
    {
      // if we have a parent, use its current adjusted write bounds as its
      // outer bounds, otherwise limit it to the screen area.
      // TODO: This should take into account any clip rect that might be
      // applied.
      const auto outerBounds = (parentClipID != 0)
                                   ? m_context->getTightenedClipBounds(
                                         parentClipID)
                                   : AABBu16::MakeWH(
                                         m_context->frameDescriptor()
                                             .renderTargetWidth,
                                         m_context->frameDescriptor()
                                             .renderTargetHeight);

      // Trim the draw bounds to the outer bounds as the initial minimal
      // clip bounds (we shouldn't need to write to or read from anywhere
      // that is outside of the screen or a parent clip's box, if one
      // exists).
      const auto tightenedBounds = outerBounds.intersect(clipDrawBounds);

      // If there is a parent clip, the next element up the clip stack
      // should have its ID.
      assert(
          parentClipID == 0 ||
          (i != 0 && m_clipStack[i - 1].clipID == parentClipID));

      clip.clipID = m_context->generateClipID(
          clipDrawBounds,
          parentClipID,
          tightenedBounds);
    }
    assert(clip.clipID != m_context->getClipContentID());
    if (clip.clipID == 0)
    {
      return ApplyClipResult::failure; // The context is out of
                                       // clipIDs. We will flush and
                                       // try again.
    }
    clipDraw->setClipID(clip.clipID);

    gpu::DrawContents clipDrawContents = clipDraw->combinedDrawContents();
    if (!m_context->isOutsideCurrentFrame(clipDrawBounds))
    {
      m_internalDrawBatch.push_back(std::move(clipDraw));
    }

    if (parentClipID != 0)
    {
      if (m_context->frameInterlockMode() == gpu::InterlockMode::depthStencil)
      {
        // When drawing nested stencil clips, we need to intersect them,
        // which involves erasing the region of the current clip in the
        // stencil buffer that is outside the the one we just drew.
        auto stencilClipIntersect = gpu::DrawUniquePtr(
            m_context->make<gpu::ClipReset>(
                m_context,
                parentClipID,
                clipDrawContents,
                gpu::ClipReset::ResetAction::intersectPreviousClip));
        if (!m_context->isOutsideCurrentFrame(
                stencilClipIntersect->pixelBounds()))
        {
          m_internalDrawBatch.push_back(std::move(stencilClipIntersect));
        }
      }
    }

    parentClipID = clip.clipID; // Nest the next clip (if any) inside the
                                // one we just rendered.
  }

  assert(parentClipID == m_clipStack[clipStackHeight - 1].clipID);
  draw->setClipID(parentClipID);
  m_context->setClipContentID(parentClipID);

  return ApplyClipResult::success;
}
} // namespace rive