#include "rive/offscreen_raster.hpp"

#ifdef RIVE_CANVAS

#include "rive/renderer.hpp"
#include "rive/renderer/cmd/deferred_canvas_host.hpp"
#include "rive/renderer/render_canvas.hpp"

#include <algorithm>
#include <atomic>
#include <cassert>
#include <cmath>

namespace rive
{
namespace offscreen
{

namespace
{
// NaN passes through (std::max/min return their first argument when the
// comparison fails), which the callers' finiteness checks then reject.
float clampResolution(float resolution)
{
  constexpr float kMinRes = 0.01f;
  constexpr float kMaxRes = 8.0f;
  return std::min<float>(std::max<float>(resolution, kMinRes), kMaxRes);
}
} // namespace

bool planRasterScale(Renderer* renderer, float resolution, RasterPlan* out)
{
  // How many device pixels one local unit covers right now: viewport zoom
  // times window density times whatever scale the mount transform adds. This
  // is the number naive sizing ignores -- rasterizing in local units means
  // that on a 2x display at 100% zoom a resolution-1 raster is already a
  // half-resolution image being magnified, which is what softens text.
  // Reading it here (rather than at replay) is why DeferredRenderer shadows
  // its CTM: the raster size has to be chosen while recording.
  out->haveCtm = renderer->currentTransform(&out->ctm);
  // Read before rasterizing, for the same reason as the CTM: if the host
  // hands back a fresh renderer to composite through, that renderer starts
  // at opacity 1 and the enclosing modulateOpacity() scope has to be carried
  // over by hand.
  out->haveOpacity = renderer->currentModulatedOpacity(&out->modulatedOpacity);

  float deviceScale = 1.0f;
  if (out->haveCtm)
  {
    const float s = out->ctm.findMaxScale();
    if (std::isfinite(s) && s > 0.0f)
    {
      // Quantize to sixteenths so scrubbing zoom does not re-raster and
      // re-allocate on every frame. Rounding up keeps the raster at least
      // as fine as the screen, and the scales a user actually rests at
      // (1, 1.5, 2, 3, 4) are already multiples of 1/16, so the cases
      // that matter stay pixel exact rather than merely close.
      // The slack absorbs float noise at a boundary (a rotated 1.5 reads
      // as 1.50000012), which would otherwise flip buckets per frame.
      constexpr float kScaleQuantum = 16.0f;
      constexpr float kSlack = 1.0f / 1024.0f;
      deviceScale = std::ceil(s * kScaleQuantum - kSlack) / kScaleQuantum;
    }
  }

  const float res = clampResolution(resolution);

  // Texels per local unit. resolution 1 means one texel per screen pixel, so
  // the composite is a 1:1 blit; 2 is a genuine 2x supersample of what the
  // screen shows. A renderer that cannot report a transform leaves
  // deviceScale at 1, which is the old local-unit meaning.
  const float rasterScale = res * deviceScale;
  if (!(rasterScale > 0.0f) || !std::isfinite(rasterScale))
  {
    return false;
  }
  out->rasterScale = rasterScale;
  return true;
}

uint32_t bucketedRasterDim(uint32_t n)
{
  // A fixed 64 rather than a ladder that widens with n. The obvious
  // refinement -- a quantum of n/16, so the slack stays a fixed *fraction* --
  // degenerates at kMaxDim = 2048: the next step up, 128, would only be
  // reached at n >= 2048, which is the cap itself. So 64 is the whole ladder
  // until kMaxDim grows.
  constexpr uint32_t kQuantum = 64;
  if (n >= kMaxDim)
  {
    return kMaxDim;
  }
  const uint32_t bucketed = ((n + kQuantum - 1) / kQuantum) * kQuantum;
  return std::min(bucketed, kMaxDim);
}

RasterResize decideRasterResize(
    uint32_t needW,
    uint32_t needH,
    uint32_t haveW,
    uint32_t haveH,
    uint8_t* streak)
{
  const uint32_t wantW = bucketedRasterDim(needW);
  const uint32_t wantH = bucketedRasterDim(needH);

  RasterResize out;
  if (haveW == 0 || haveH == 0)
  {
    *streak = 0;
    return {true, wantW, wantH};
  }

  if (needW > haveW || needH > haveH)
  {
    // Grow immediately: a raster smaller than its content crops, and a crop
    // is visible where an oversized texture is only wasteful. Only the axis
    // that is actually short grows, so growing one axis never sneaks an
    // un-hysteresised shrink onto the other.
    *streak = 0;
    out.reallocate = true;
    out.widthPx = needW > haveW ? wantW : haveW;
    out.heightPx = needH > haveH ? wantH : haveH;
    return out;
  }

  if (wantW < haveW || wantH < haveH)
  {
    // Wants to shrink. Wait: content animating through a size would
    // otherwise re-allocate every frame, and re-allocating is far more
    // expensive than carrying a texture that is one bucket too big.
    // Increment before comparing, so the kShrinkDraws'th consecutive
    // shrink request is the one that acts rather than the one after it.
    ++(*streak);
    if (*streak < kShrinkDraws)
    {
      out.widthPx = haveW;
      out.heightPx = haveH;
      return out;
    }
    *streak = 0;
    return {true, wantW, wantH};
  }

  *streak = 0;
  out.widthPx = haveW;
  out.heightPx = haveH;
  return out;
}

bool fitRasterToBox(
    const AABB& box,
    uint32_t guardTexels,
    RasterFit fit,
    RasterPlan* io)
{
  assert(fit == RasterFit::stableGrid || guardTexels == 0);

  const float w = box.width();
  const float h = box.height();
  if (!(w > 0.0f) || !(h > 0.0f))
  {
    return false;
  }

  // Cap by shrinking the scale, not by clamping one axis: clamping a single
  // dimension would make the two axes disagree about the mapping, and the
  // composite inverts one scale for both. Uniform coarsening at least keeps
  // the image undistorted.
  //
  // stableGrid holds back room for the guard band and for the texel either
  // end of the origin snap can add, so the cap still holds after they are
  // applied rather than being clamped away afterwards.
  const uint32_t reserve = fit == RasterFit::stableGrid ? 2u * guardTexels + 2u
                                                        : 0u;
  const float usableDim = static_cast<float>(
      kMaxDim > reserve ? kMaxDim - reserve : 1u);
  const float maxScale = std::min<float>(usableDim / w, usableDim / h);

  const float rasterScale = std::min<float>(io->rasterScale, maxScale);
  if (!(rasterScale > 0.0f) || !std::isfinite(rasterScale))
  {
    return false;
  }
  io->rasterScale = rasterScale;

  if (fit == RasterFit::exact)
  {
    io->box = box;
    io->widthPx = static_cast<uint32_t>(std::min<float>(
        std::max<float>(std::ceil(w * rasterScale), 1.0f),
        static_cast<float>(kMaxDim)));
    io->heightPx = static_cast<uint32_t>(std::min<float>(
        std::max<float>(std::ceil(h * rasterScale), 1.0f),
        static_cast<float>(kMaxDim)));
    return true;
  }

  // Snap the origin down to the raster's own texel grid, anchored at local 0
  // -- NOT to the tight box's own edge. If the origin tracked the content,
  // the content would land at the same sub-texel phase in every raster, and
  // beginComposite's pixel snap would then quantize the composite to whole
  // device pixels: sub-pixel motion would judder. A grid fixed in local space
  // lets the phase vary continuously, exactly as a fixed box does.
  const float guard = static_cast<float>(guardTexels);
  const float leftTx = std::floor(box.left() * rasterScale) - guard;
  const float topTx = std::floor(box.top() * rasterScale) - guard;
  const float rightTx = std::ceil(box.right() * rasterScale) + guard;
  const float bottomTx = std::ceil(box.bottom() * rasterScale) + guard;
  if (!std::isfinite(leftTx) || !std::isfinite(topTx) ||
      !std::isfinite(rightTx) || !std::isfinite(bottomTx))
  {
    return false;
  }

  const uint32_t widthPx = bucketedRasterDim(
      static_cast<uint32_t>(std::min<float>(
          std::max<float>(rightTx - leftTx, 1.0f),
          static_cast<float>(kMaxDim))));
  const uint32_t heightPx = bucketedRasterDim(
      static_cast<uint32_t>(std::min<float>(
          std::max<float>(bottomTx - topTx, 1.0f),
          static_cast<float>(kMaxDim))));

  io->widthPx = widthPx;
  io->heightPx = heightPx;
  // The box spans exactly the chosen pixel size, so the composite's
  // 1 / rasterScale inverse maps the raster onto it with nothing left over.
  io->box = AABB(
      leftTx / rasterScale,
      topTx / rasterScale,
      (leftTx + static_cast<float>(widthPx)) / rasterScale,
      (topTx + static_cast<float>(heightPx)) / rasterScale);
  return true;
}

void holdRasterAllocation(uint32_t widthPx, uint32_t heightPx, RasterPlan* io)
{
  if (widthPx <= io->widthPx && heightPx <= io->heightPx)
  {
    return;
  }
  const uint32_t w = std::max(widthPx, io->widthPx);
  const uint32_t h = std::max(heightPx, io->heightPx);
  // Right and bottom only. io->box.left/top is what beginComposite's pixel
  // snap reads; moving it would shift the composite on screen.
  io->box = AABB(
      io->box.left(),
      io->box.top(),
      io->box.left() + static_cast<float>(w) / io->rasterScale,
      io->box.top() + static_cast<float>(h) / io->rasterScale);
  io->widthPx = w;
  io->heightPx = h;
}

bool planRaster(
    Renderer* renderer,
    const AABB& box,
    float resolution,
    RasterPlan* out)
{
  RasterPlan plan;
  if (!planRasterScale(renderer, resolution, &plan))
  {
    return false;
  }
  if (!fitRasterToBox(box, 0, RasterFit::exact, &plan))
  {
    return false;
  }
  *out = plan;
  return true;
}

namespace
{
std::atomic<float> s_displayScale{1.0f};
} // namespace

void setDisplayScale(float scale)
{
  if (!std::isfinite(scale) || scale <= 0.0f)
  {
    return;
  }
  s_displayScale.store(
      std::min(std::max(scale, 1.0f), 4.0f),
      std::memory_order_relaxed);
}

float displayScale() { return s_displayScale.load(std::memory_order_relaxed); }

bool planFixedRasterSize(const AABB& box, float resolution, RasterPlan* out)
{
  RasterPlan plan;
  const float res = clampResolution(resolution) * displayScale();
  if (!(res > 0.0f) || !std::isfinite(res))
  {
    return false;
  }
  plan.rasterScale = res;
  if (!fitRasterToBox(box, 0, RasterFit::exact, &plan))
  {
    return false;
  }
  *out = plan;
  return true;
}

bool planFixedRaster(
    Renderer* renderer,
    const AABB& box,
    float resolution,
    RasterPlan* out)
{
  RasterPlan plan;
  if (!planFixedRasterSize(box, resolution, &plan))
  {
    return false;
  }
  plan.haveCtm = renderer->currentTransform(&plan.ctm);
  plan.haveOpacity = renderer->currentModulatedOpacity(&plan.modulatedOpacity);
  *out = plan;
  return true;
}

CanvasContentScope::CanvasContentScope(
    cmd::DeferredCanvasHost* host,
    gpu::RenderCanvas* canvas,
    const RasterPlan& plan,
    uint32_t clearColor)
    : m_host(host)
    , m_canvas(canvas)
    , m_renderer(nullptr)
{
  if (host == nullptr || canvas == nullptr)
  {
    return;
  }
  // beginCanvasContent hands back a recording renderer whose frame targets
  // the canvas texture. Draw content into it at exactly rasterScale texels
  // per local unit -- not widthPx/w, which would bake ceil()'s rounding into
  // the mapping. Curve flattening and feathering read this transform, so this
  // is also what decides how finely the content is tessellated.
  m_renderer = host->beginCanvasContent(canvas, clearColor);
  if (m_renderer == nullptr)
  {
    return;
  }
  m_renderer->save();
  m_renderer->transform(Mat2D::fromScale(plan.rasterScale, plan.rasterScale));
  m_renderer->translate(-plan.box.left(), -plan.box.top());
}

CanvasContentScope::~CanvasContentScope()
{
  if (m_renderer == nullptr)
  {
    // Nothing was drawn, so never end a bracket that never began.
    return;
  }
  m_renderer->restore();
  m_host->endCanvasContent(m_canvas);
}

CompositePlacement beginComposite(
    Renderer* renderer,
    cmd::DeferredCanvasHost* host,
    const RasterPlan& plan,
    float compositedRasterScale,
    const AABB& compositedBox)
{
  // Some hosts cannot composite through the renderer that was drawing when
  // the offscreen frame interrupted it, and hand back a clean one instead.
  // It shares no state, so the current transform and the modulated opacity
  // both have to be re-applied by hand -- and it inherits no clip. Known
  // limitation: on that path (WebGL today) an ancestor's clip does not
  // constrain the composite, so cached content under a clipping shape can
  // paint outside it where the vector draw would have been cropped.
  Renderer* composite = renderer;
  float compositeOpacity = 1.0f;
  if (host != nullptr)
  {
    if (Renderer* fresh = host->compositeRenderer())
    {
      if (plan.haveCtm && plan.haveOpacity)
      {
        composite = fresh;
        compositeOpacity = plan.modulatedOpacity;
      }
    }
  }
  composite->save();
  if (composite != renderer)
  {
    composite->transform(plan.ctm);
  }
  // Pixel snap. Even at a perfectly matched scale, an origin that lands on a
  // half pixel makes every bilinear tap a blend of two texels -- a box blur
  // over the whole image, and the reason cache-as-bitmap has historically
  // needed a pixelSnapping switch. Only meaningful when the transform is axis
  // aligned; under rotation or skew there is no pixel grid to snap to.
  if (plan.haveCtm && plan.ctm.xy() == 0.0f && plan.ctm.yx() == 0.0f &&
      plan.ctm.xx() != 0.0f && plan.ctm.yy() != 0.0f)
  {
    const Vec2D deviceOrigin = plan.ctm *
                               Vec2D(compositedBox.left(), compositedBox.top());
    const Vec2D delta = {
        std::round(deviceOrigin.x) - deviceOrigin.x,
        std::round(deviceOrigin.y) - deviceOrigin.y};
    // The nudge is in device space but transform() concatenates in local
    // space, so push it back through the (diagonal) linear part.
    composite->translate(delta.x / plan.ctm.xx(), delta.y / plan.ctm.yy());
  }
  composite->translate(compositedBox.left(), compositedBox.top());
  // The exact inverse of the scale the content was rasterized at. Deriving it
  // from w / widthPx instead would fold ceil()'s rounding into the mapping,
  // leaving a fractional-width box permanently resampled at ~0.995x -- a blur
  // on every pixel for nothing. The cost is that the quad reaches up to one
  // texel past the box on the right and bottom, which is the sliver that
  // rounding allocated anyway.
  const float invScale = 1.0f / compositedRasterScale;
  composite->transform(Mat2D::fromScale(invScale, invScale));

  return {composite, compositeOpacity};
}

CompositePlacement beginComposite(
    Renderer* renderer,
    cmd::DeferredCanvasHost* host,
    const RasterPlan& plan,
    float compositedRasterScale)
{
  return beginComposite(renderer, host, plan, compositedRasterScale, plan.box);
}

void endComposite(const CompositePlacement& placement)
{
  placement.renderer->restore();
}

CompositePlacement beginPlacement(
    Renderer* renderer,
    cmd::DeferredCanvasHost* host,
    const RasterPlan& plan)
{
  // Same renderer choice as beginComposite.
  Renderer* target = renderer;
  float opacity = 1.0f;
  if (host != nullptr)
  {
    if (Renderer* fresh = host->compositeRenderer())
    {
      if (plan.haveCtm && plan.haveOpacity)
      {
        target = fresh;
        opacity = plan.modulatedOpacity;
      }
    }
  }
  target->save();
  if (target != renderer)
  {
    target->transform(plan.ctm);
  }
  return {target, opacity};
}

CachedRaster::CachedRaster() = default;
CachedRaster::~CachedRaster() = default;

void CachedRaster::release()
{
  canvas.reset();
  widthPx = 0;
  heightPx = 0;
  contentRevision = 0;
  dirty = true;
}

} // namespace offscreen
} // namespace rive

#endif // RIVE_CANVAS