feat(render): Renderer with LOD selection, 3.3 and 4.3+ upload paths, headless smoke test

src/render/buffer.cpp

@@ -1,9 +1,13 @@
 #include <plot/render/buffer.hpp>
 
+#include <plot/gl/api.hpp>
+
+#include <cstring>
+
 namespace plot::render {
 namespace detail {
 
-bool InstanceStore::create(Status* st) noexcept {
+bool InstanceStore::create(UploadMode mode, Status* st) noexcept {
   // plot::gl::Buffer::create() already reports Unsupported while the entry
   // points are unresolved and GLFailure when the driver hands out no name, so
   // the wrapper only has to propagate its answer.
@@ -11,21 +15,103 @@ bool InstanceStore::create(Status* st) noexcept {
     return false;
   }
   capacity_ = 0;
+  slot_bytes_ = 0;
+  slot_ = kSlots - 1u;  // the first upload takes slot 0
+  // The mapped path additionally needs its entry points, which are resolved
+  // best-effort (see api.hpp): without them the store would never be able to
+  // write anything, so it stays on the baseline path instead.
+  mode_ = mode == UploadMode::PersistentMapped && gl::glBufferStorage != nullptr &&
+                  gl::glMapBufferRange != nullptr
+              ? UploadMode::PersistentMapped
+              : UploadMode::Orphan;
+  return true;
+}
+
+void InstanceStore::release_storage() noexcept {
+  if (mapped_ != nullptr) {
+    // A persistent mapping outlives the frame that created it, so it has to be
+    // dropped before the buffer it belongs to is deleted.
+    buffer_.unmap();
+    mapped_ = nullptr;
+  }
+  for (gl::Fence& fence : fences_) {
+    fence.reset();
+  }
+  buffer_.reset();
+  slot_bytes_ = 0;
+}
+
+bool InstanceStore::grow_persistent(usize bytes) noexcept {
+  // Deleting the old buffer is safe while the GPU may still be reading it: GL
+  // keeps the storage alive until the last command that references it retires.
+  release_storage();
+  Status st;
+  if (!buffer_.create(&st)) {
+    mode_ = UploadMode::Orphan;  // nothing left to draw with at all
+    capacity_ = 0;
+    return false;
+  }
+  const gl::GLbitfield flags =
+      gl::GL_MAP_WRITE_BIT | gl::GL_MAP_PERSISTENT_BIT | gl::GL_MAP_COHERENT_BIT;
+  buffer_.storage(kSlots * bytes, flags);
+  mapped_ = static_cast<u8*>(buffer_.map_range(0, kSlots * bytes, flags));
+  if (mapped_ == nullptr) {
+    // The driver refused the storage or the mapping. A frame that draws through
+    // the baseline path beats one that draws nothing, so re-create the empty
+    // buffer the orphan path needs and let the next upload use it.
+    release_storage();
+    mode_ = UploadMode::Orphan;
+    capacity_ = 0;
+    Status retry;
+    (void)buffer_.create(&retry);  // a failed retry leaves nothing to draw with
+    return false;
+  }
+  slot_bytes_ = bytes;
+  slot_ = kSlots - 1u;  // fresh storage: the first upload takes slot 0
+  ++epoch_;
   return true;
 }
 
 void InstanceStore::upload(const void* data, u32 count, usize instance_bytes) noexcept {
-  if (buffer_.id() == 0 || count == 0 || instance_bytes == 0) {
+  if (buffer_.id() == 0 || data == nullptr || count == 0 || instance_bytes == 0) {
     return;
   }
-  if (count > capacity_) {
-    // Only a workload larger than anything staged so far pays for an
-    // allocation; a frame that repeats the previous one writes into the store
-    // that is already there.
-    buffer_.upload(nullptr, static_cast<usize>(count) * instance_bytes, gl::GL_STREAM_DRAW);
-    capacity_ = count;
+  const usize bytes = usize(count) * instance_bytes;
+  if (mode_ == UploadMode::PersistentMapped && bytes > slot_bytes_ && !grow_persistent(bytes)) {
+    // The fallback already replaced the store with an orphan buffer.
   }
-  buffer_.upload_sub(data, static_cast<usize>(count) * instance_bytes, 0);
+
+  if (mode_ == UploadMode::Orphan) {
+    if (count > capacity_) {
+      capacity_ = count;  // grow only when the workload outgrows the store
+    }
+    buffer_.upload(nullptr, usize(capacity_) * instance_bytes, gl::GL_STREAM_DRAW);
+    buffer_.upload_sub(data, bytes, 0);
+    return;
+  }
+
+  // Advance the ring, then make sure the draw that last read this slot has
+  // retired. The non-blocking check is the common case; the bounded wait only
+  // costs anything when three frames' geometry is still in flight.
+  slot_ = (slot_ + 1u) % kSlots;
+  if (fences_[slot_]) {
+    if (!fences_[slot_].wait(0)) {
+      (void)fences_[slot_].wait(kFenceWaitNs);
+    }
+    fences_[slot_].reset();
+  }
+  // MAP_COHERENT storage needs no explicit flush, so this is the whole transfer.
+  std::memcpy(mapped_ + usize(slot_) * slot_bytes_, data, bytes);
+  ++epoch_;
+}
+
+void InstanceStore::mark_drawn() noexcept {
+  if (mode_ != UploadMode::PersistentMapped || mapped_ == nullptr) {
+    return;
+  }
+  Status st;
+  fences_[slot_].reset();
+  (void)fences_[slot_].create(&st);
 }
 
 }  // namespace detail

src/render/primitives.cpp

@@ -27,6 +27,19 @@ void LineBatch::add_polyline(const Vec2* pts, u32 n) {
   }
 }
 
+void LineBatch::add_polyline_xy(const f32* xy, u32 n) {
+  if (xy == nullptr || n < 2u) {
+    return;
+  }
+  v_.reserve(v_.size() + 4u * (n - 1u));
+  for (u32 i = 1; i < n; ++i) {
+    v_.push_back(xy[2u * (i - 1u)]);
+    v_.push_back(xy[2u * (i - 1u) + 1u]);
+    v_.push_back(xy[2u * i]);
+    v_.push_back(xy[2u * i + 1u]);
+  }
+}
+
 // ----------------------------------------------------------------- TextBatch
 
 void TextBatch::begin() noexcept { v_.clear(); }

src/render/renderer.cpp

@@ -0,0 +1,619 @@
+#include <plot/render/renderer.hpp>
+
+#include <plot/gl/api.hpp>
+#include <plot/gl/caps.hpp>
+#include <plot/gl/loader.hpp>
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+namespace plot {
+namespace {
+
+/// `PLOT_FORCE_GL33_UPLOAD=1` forces the baseline transfer, like
+/// `PLOT_HEADLESS_FORCE_PBUFFER` forces the EGL fallback: presence is the switch,
+/// so both strategies can be exercised on one machine.
+bool force_gl33_upload() noexcept { return std::getenv("PLOT_FORCE_GL33_UPLOAD") != nullptr; }
+
+bool fail(Status* st, ErrorCode code, const char* message) noexcept {
+  if (st != nullptr) {
+    st->code = code;
+    st->message = message;
+  }
+  return false;
+}
+
+/// A data-space point in the coordinate the line program's affine part expects:
+/// the value itself on a linear axis, `log10` of it on a log axis -- exactly what
+/// `math::Transform::to_screen_*` applies before scaling. A non-positive value on
+/// a log axis takes the transform's own clamp, so it lands off-scale rather than
+/// as NaN.
+Vec2 pre_affine(f64 x, f64 y, bool log_x, bool log_y) noexcept {
+  return Vec2{static_cast<f32>(log_x ? Transform::log10_positive(x) : x),
+              static_cast<f32>(log_y ? Transform::log10_positive(y) : y)};
+}
+
+/// Sample index for a (possibly fractional, possibly non-finite) position in a
+/// series' index space, clamped to [0, n]: NaN and -inf to 0, +inf to n, so a
+/// range the axis cannot map still yields a valid window.
+u64 clamp_index(f64 v, u64 n) noexcept {
+  if (!(v > 0.0)) return 0;
+  if (v >= static_cast<f64>(n)) return n;
+  return static_cast<u64>(v);
+}
+
+/// One ASCII/UTF-8 line of the measurement readout. `format_si` leaves room for
+/// three significant digits plus a prefix and sign.
+constexpr usize kNumberChars = 24;
+constexpr usize kReadoutChars = 160;
+
+}  // namespace
+
+// --------------------------------------------------------------- AxisFont
+
+void Renderer::AxisFont::reset(const render::BitmapFont& font, f32 requested_px) noexcept {
+  font_ = &font;
+  const f32 native = static_cast<f32>(render::BitmapFont::kGlyphPx);
+  // Whole multiples of the native cell only: a 1-bit glyph drawn at 1.5x lands
+  // its strokes on uneven pixel counts, and the layout has to agree with what the
+  // pipeline draws, so the requested size picks the nearest multiple.
+  f32 steps = std::round(requested_px / native);
+  if (!(steps >= 1.0f)) steps = 1.0f;  // also catches NaN
+  cell_px_ = steps * native;
+  scale_ = cell_px_ / native;
+}
+
+f32 Renderer::AxisFont::measure(const char* utf8) const {
+  return font_ != nullptr ? font_->measure(utf8) * scale_ : 0.0f;
+}
+
+f32 Renderer::AxisFont::height() const {
+  return font_ != nullptr ? font_->height() * scale_ : 0.0f;
+}
+
+f32 Renderer::AxisFont::ascent() const {
+  return font_ != nullptr ? font_->ascent() * scale_ : 0.0f;
+}
+
+// ------------------------------------------------------------------ init
+
+bool Renderer::init(Status* st) noexcept {
+  if (!gl::available()) {
+    return fail(st, ErrorCode::Unsupported,
+                "plot::gl::load() has not resolved the GL entry points");
+  }
+
+  // The transfer strategy is decided once, from the capability probe, and never
+  // revisited. Both pointers are checked as well as the caps: they are resolved
+  // best-effort (see api.hpp), so a context can advertise the extension without
+  // the entry points being reachable.
+  const gl::Caps& caps = gl::caps();
+  persistent_ = caps.buffer_storage && caps.persistent_map && gl::glBufferStorage != nullptr &&
+                gl::glMapBufferRange != nullptr && !force_gl33_upload();
+  const render::UploadMode mode =
+      persistent_ ? render::UploadMode::PersistentMapped : render::UploadMode::Orphan;
+
+  font_backend_.reset(font_, style_.axis_font_px);
+
+  if (!atlas_.create_r8(render::BitmapFont::kAtlasWidth, render::BitmapFont::kAtlasHeight, st)) {
+    return false;
+  }
+  atlas_.upload_r8(render::BitmapFont::kAtlasWidth, render::BitmapFont::kAtlasHeight,
+                   render::BitmapFont::atlas_image());
+  // Texel-for-texel: the text shader samples the nearest texel centre of the
+  // glyph cell, and linear filtering would smear one cell into its neighbour.
+  atlas_.set_filter(gl::GL_NEAREST, gl::GL_NEAREST);
+
+  if (!line_program_.create(st) || !text_program_.create(st)) {
+    return false;
+  }
+  if (!line_vao_.create(st, mode) || !text_vao_.create(st, mode)) {
+    return false;
+  }
+  ready_ = true;
+  return true;
+}
+
+// ------------------------------------------------------------ frame cycle
+
+void Renderer::begin_frame(u32 fb_w, u32 fb_h, Rect viewport_px) noexcept {
+  fb_w_ = fb_w;
+  fb_h_ = fb_h;
+  viewport_ = viewport_px;
+  lines_.clear();
+  glyphs_.clear();
+  stats_ = FrameStats{};
+  gl_error_ = false;
+
+  // Rect is top-left origin, GL is bottom-left: flip once, here, and never again.
+  gl_viewport_ = Rect{viewport_px.x,
+                      static_cast<f32>(fb_h) - (viewport_px.y + viewport_px.h),
+                      viewport_px.w, viewport_px.h};
+  if (!ready_) {
+    return;
+  }
+
+  const gl::GLint vx = static_cast<gl::GLint>(gl_viewport_.x);
+  const gl::GLint vy = static_cast<gl::GLint>(gl_viewport_.y);
+  const gl::GLsizei vw = static_cast<gl::GLsizei>(gl_viewport_.w);
+  const gl::GLsizei vh = static_cast<gl::GLsizei>(gl_viewport_.h);
+
+  // Clear only the plot's region: a host that embeds the plot in a surface it has
+  // already painted keeps everything outside the rect. The scissor box is in
+  // framebuffer coordinates and is not affected by glViewport, which is why the
+  // clear happens before the viewport is set.
+  gl::glEnable(gl::GL_SCISSOR_TEST);
+  gl::glScissor(vx, vy, vw, vh);
+  const Color& background = style_.background;
+  gl::glClearColor(background.r, background.g, background.b, background.a);
+  gl::glClear(gl::GL_COLOR_BUFFER_BIT);
+  gl::glDisable(gl::GL_SCISSOR_TEST);
+  gl::glViewport(vx, vy, vw, vh);
+
+  // The pipelines write straight alpha and rely on blending being on; setting it
+  // per frame costs nothing and makes the renderer independent of what the host
+  // left behind.
+  gl::glEnable(gl::GL_BLEND);
+  gl::glBlendFunc(gl::GL_SRC_ALPHA, gl::GL_ONE_MINUS_SRC_ALPHA);
+}
+
+void Renderer::draw(Plot& plot) {
+  stats_ = FrameStats{};
+  if (!ready_) {
+    return;
+  }
+  // The plot reserves its label strips from TextBackend metrics, so it must
+  // measure with the size the text pipeline draws with, and re-read the style
+  // each frame so a changed axis_font_px takes effect on the next one.
+  font_backend_.reset(font_, style_.axis_font_px);
+  if (plot.text_backend() == nullptr) {
+    plot.set_text_backend(&font_backend_);
+  }
+  // One style, not two: the plot reserves its strips and picks its tick density
+  // from the values the renderer draws with, so the label pad is never one number
+  // for the layout and another for the text placed inside it.
+  plot.set_style(style_);
+  (void)plot.layout();  // autoscale, ticks and grid for the final axes size
+
+  // Draw order: chrome, then data, then the annotations over it, and the text
+  // last (it is texture-bound, so it is one draw call after every line batch).
+  draw_grid(plot);
+  draw_axis_lines(plot);
+  draw_tick_labels(plot);
+  draw_titles(plot);
+  draw_traces(plot);
+  draw_cursors(plot);
+  draw_measures(plot);
+  draw_annotations(plot);
+  flush_text();
+}
+
+void Renderer::end_frame() noexcept {
+  if (!ready_) {
+    return;
+  }
+  // Leave the host's context as close to neutral as a frame can: nothing bound to
+  // a pipeline stage, and off every state bit the frame turned on. What remains
+  // is the viewport (the caller's to set for its next pass), the pixel-pack
+  // alignment a PPM dump may have set, and the vertex buffers' own contents.
+  gl::glBindVertexArray(0);
+  gl::glUseProgram(0);
+  gl::glBindBuffer(gl::GL_ARRAY_BUFFER, 0);
+  gl::glActiveTexture(gl::GL_TEXTURE0);
+  gl::glBindTexture(gl::GL_TEXTURE_2D, 0);
+  gl::glDisable(gl::GL_BLEND);
+  // Drain, do not peek: the first error hides every later one, and a frame that
+  // produced several must be reported as having produced one.
+  while (gl::glGetError() != gl::GL_NO_ERROR) {
+    gl_error_ = true;
+  }
+}
+
+// ------------------------------------------------------------ submission
+
+void Renderer::upload_vertices(std::span<const f32> interleaved) noexcept {
+  // Four floats per instance is the line pipeline's layout (see LineInstance).
+  // Both transfer strategies live behind the vertex array, so this is the only
+  // place the line pipeline's staged bytes reach the GPU and no caller of it can
+  // tell which path ran.
+  line_vao_.upload(interleaved.data(), static_cast<u32>(interleaved.size() / 4u));
+}
+
+void Renderer::upload_text_vertices(std::span<const f32> interleaved) noexcept {
+  text_vao_.upload(interleaved.data(), static_cast<u32>(interleaved.size() / 4u));
+}
+
+void Renderer::flush_lines(const Mapping& map, Color color, f32 width) noexcept {
+  const u32 segments = lines_.segment_count();
+  if (segments != 0 && ready_) {
+    upload_vertices(lines_.data());
+    line_program_.bind();
+    line_program_.set_scale(map.scale);
+    line_program_.set_offset(map.offset);
+    line_program_.set_viewport({viewport_.w, viewport_.h});
+    line_program_.set_width(width);
+    line_program_.set_color(color);
+    line_vao_.bind();
+    // One triangle strip per instance, four corners from gl_VertexID.
+    gl::glDrawArraysInstanced(gl::GL_TRIANGLE_STRIP, 0, 4,
+                              static_cast<gl::GLsizei>(line_vao_.instance_count()));
+    // The draw is in the command stream now, so this is the point at which the
+    // mapped path can let a fence follow it.
+    line_vao_.mark_drawn();
+    stats_.staged_segments += segments;
+    ++stats_.draw_calls;
+  }
+  lines_.clear();  // keeps the capacity, so the next group re-uses the store
+}
+
+void Renderer::flush_text() noexcept {
+  const u32 glyphs = glyphs_.glyph_count();
+  if (glyphs != 0 && ready_) {
+    upload_text_vertices(glyphs_.data());
+    text_program_.bind();
+    text_program_.set_viewport({viewport_.w, viewport_.h});
+    const f32 cell = font_backend_.cell_px();
+    text_program_.set_px_range({cell, cell});
+    text_program_.set_texture_size(
+        {static_cast<f32>(render::BitmapFont::kAtlasWidth),
+         static_cast<f32>(render::BitmapFont::kAtlasHeight)});
+    text_program_.set_color(style_.axis_text);
+    atlas_.bind();
+    text_vao_.bind();
+    gl::glDrawArraysInstanced(gl::GL_TRIANGLE_STRIP, 0, 4,
+                              static_cast<gl::GLsizei>(text_vao_.instance_count()));
+    text_vao_.mark_drawn();
+    ++stats_.draw_calls;
+  }
+  glyphs_.clear();
+}
+
+// ----------------------------------------------------------- coordinates
+
+Renderer::Mapping Renderer::mapping(const Plot& plot, u32 axis_x, u32 axis_y) const noexcept {
+  const Transform tx = plot.x_transform(axis_x);
+  const Transform ty = plot.y_transform(axis_y);
+  // The plot's transforms already map data to framebuffer pixels; the shaders
+  // emit NDC relative to the GL viewport, so the viewport origin comes off here
+  // -- once per mapping, never per primitive.
+  return Mapping{Vec2{static_cast<f32>(tx.sx), static_cast<f32>(ty.sy)},
+                 Vec2{static_cast<f32>(tx.tx) - viewport_.x,
+                      static_cast<f32>(ty.ty) - viewport_.y}};
+}
+
+Renderer::Mapping Renderer::pixel_mapping() const noexcept {
+  // Geometry already in framebuffer pixels needs no scaling, and the line
+  // program's offset uniform does the viewport shift. (The text pipeline has no
+  // offset uniform, so stage_text() shifts its positions itself.)
+  return Mapping{Vec2{1.0f, 1.0f}, Vec2{-viewport_.x, -viewport_.y}};
+}
+
+Vec2 Renderer::to_viewport(Vec2 px) const noexcept {
+  return Vec2{px.x - viewport_.x, px.y - viewport_.y};
+}
+
+void Renderer::stage_segment(f64 x0, f64 y0, f64 x1, f64 y1, bool log_x, bool log_y) {
+  lines_.add_segment(pre_affine(x0, y0, log_x, log_y), pre_affine(x1, y1, log_x, log_y));
+}
+
+void Renderer::stage_xy(const f32* xy, u32 n, bool log_x, bool log_y) {
+  if (!log_x && !log_y) {
+    // The common case: gathered points are already interleaved `x,y`, which is
+    // the batch's own payload, so this is a straight copy.
+    lines_.add_polyline_xy(xy, n);
+    return;
+  }
+  for (u32 i = 0; i + 1u < n; ++i) {
+    stage_segment(xy[2u * i], xy[2u * i + 1u], xy[2u * i + 2u], xy[2u * i + 3u], log_x, log_y);
+  }
+}
+
+Vec2 Renderer::stage_text(Vec2 pos_px, const char* utf8) {
+  // Text is staged in viewport pixels: the text program has no offset uniform.
+  return glyphs_.add_text(to_viewport(pos_px), utf8, font_, font_backend_.cell_px());
+}
+
+// -------------------------------------------------------------------- grid
+
+void Renderer::draw_grid(const Plot& plot) {
+  const Grid& grid = plot.grid();
+  const Mapping map = mapping(plot, 0, 0);
+  // The grid is built from the primary axis pair, so it is mapped with their
+  // transforms even when a trace uses another pair.
+  const bool log_x = plot.x_axis(0).log;
+  const bool log_y = plot.y_axis(0).log;
+
+  // Two colour groups, one pass over the segments each: `LineBatch::clear()`
+  // keeps its capacity, so the second group re-uses the first one's storage.
+  for (u32 pass = 0; pass < 2u; ++pass) {
+    const bool majors = pass == 0u;
+    for (const GridLine& line : grid.lines()) {
+      if (line.major != majors) continue;  // grid.build() already dropped non-finite ones
+      stage_segment(line.x0, line.y0, line.x1, line.y1, log_x, log_y);
+    }
+    flush_lines(map, majors ? style_.grid_major : style_.grid_minor, 1.0f);
+  }
+}
+
+// ------------------------------------------------------------- axis chrome
+
+void Renderer::draw_axis_lines(const Plot& plot) {
+  const Range y_range = plot.y_range(0);
+  const Range x_range = plot.x_range(0);
+  const bool log_x0 = plot.x_axis(0).log;
+  const bool log_y0 = plot.y_axis(0).log;
+
+  // An x axis draws along its edge of the axes rect, spanning its own range; the
+  // perpendicular extent is the primary y axis' range, which is exactly what
+  // maps to the rect's top and bottom.
+  for (u32 i = 0; i < plot.x_axis_count(); ++i) {
+    const Axis& axis = plot.x_axis(i);
+    const f64 y = axis.place == AxisPlace::Top ? y_range.max : y_range.min;
+    stage_segment(axis.range.min, y, axis.range.max, y, axis.log, log_y0);
+    flush_lines(mapping(plot, i, 0), style_.axis_line, style_.line_width_px);
+  }
+  for (u32 i = 0; i < plot.y_axis_count(); ++i) {
+    const Axis& axis = plot.y_axis(i);
+    const f64 x = axis.place == AxisPlace::Right ? x_range.max : x_range.min;
+    stage_segment(x, axis.range.min, x, axis.range.max, log_x0, axis.log);
+    flush_lines(mapping(plot, 0, i), style_.axis_line, style_.line_width_px);
+  }
+}
+
+void Renderer::draw_tick_labels(const Plot& plot) {
+  const PlotLayout& layout = plot.layout();
+  const f32 line_h = font_backend_.height();
+
+  // X labels: centred under their tick, on the strip's single text line.
+  for (u32 i = 0; i < plot.x_axis_count(); ++i) {
+    const Transform transform = plot.x_transform(i);
+    for (const Tick& tick : plot.x_axis(i).ticks) {
+      if (!tick.major || tick.label.empty()) continue;
+      const f32 width = font_backend_.measure(tick.label.c_str());
+      stage_text({static_cast<f32>(transform.to_screen_x(tick.value)) - 0.5f * width,
+                  layout.x_label_area.y + 0.5f * (layout.x_label_area.h - line_h)},
+                 tick.label.c_str());
+    }
+  }
+
+  // Y labels: right-aligned in their strip, centred on the tick and pulled back
+  // inside the strip at the two extremes (a label centred on the top tick would
+  // otherwise hang above the plot area).
+  const f32 top_limit =
+      std::max(layout.y_label_area.y, layout.y_label_area.bottom() - line_h);
+  for (u32 i = 0; i < plot.y_axis_count(); ++i) {
+    const Transform transform = plot.y_transform(i);
+    for (const Tick& tick : plot.y_axis(i).ticks) {
+      if (!tick.major || tick.label.empty()) continue;
+      const f32 width = font_backend_.measure(tick.label.c_str());
+      const f32 centre = static_cast<f32>(transform.to_screen_y(tick.value));
+      const f32 top = std::clamp(centre - 0.5f * line_h, layout.y_label_area.y, top_limit);
+      stage_text({layout.y_label_area.right() - style_.grid_label_pad - width, top},
+                 tick.label.c_str());
+    }
+  }
+}
+
+void Renderer::draw_titles(const Plot& plot) {
+  // The layout reserves one text line per strip for the tick labels, so an axis
+  // title has no room outside the axes rect. It hugs the inside of its axis line
+  // instead, at the far end of the axis, where it needs no extra layout space and
+  // cannot collide with the labels of the axis it names.
+  const Rect& axes = plot.layout().axes;
+  const f32 line_h = font_backend_.height();
+
+  for (u32 i = 0; i < plot.x_axis_count(); ++i) {
+    const Axis& axis = plot.x_axis(i);
+    if (axis.label.empty()) continue;
+    const f32 width = font_backend_.measure(axis.label.c_str());
+    const f32 top = axis.place == AxisPlace::Top ? axes.y + 2.0f : axes.bottom() - line_h - 2.0f;
+    stage_text({axes.right() - width - 2.0f, top}, axis.label.c_str());
+  }
+  for (u32 i = 0; i < plot.y_axis_count(); ++i) {
+    const Axis& axis = plot.y_axis(i);
+    if (axis.label.empty()) continue;
+    const f32 width = font_backend_.measure(axis.label.c_str());
+    const f32 left = axis.place == AxisPlace::Right ? axes.right() - width - 2.0f : axes.x + 2.0f;
+    stage_text({left, axes.y + 2.0f}, axis.label.c_str());
+  }
+}
+
+// ------------------------------------------------------------------- data
+
+void Renderer::draw_traces(const Plot& plot) {
+  const PlotLayout& layout = plot.layout();
+  if (!(layout.axes.w > 0.0f)) {
+    return;  // a collapsed plot area has no pixel to select a level for
+  }
+
+  for (u32 i = 0; i < plot.trace_count(); ++i) {
+    const Trace& trace = plot.trace(i);
+    if (!trace.visible()) continue;  // null series, or one with no sample yet
+    const data::Series& series = *trace.series;
+    const f64 dx = series.dx();
+    const u64 n = series.size();
+    if (!(dx > 0.0) || n == 0) continue;
+
+    // Map the visible x range onto sample indices. One sample of slack on each
+    // side covers a partially visible end bucket; the range is clamped to the
+    // series, so a zoomed-out view never asks for more than exists.
+    const Range x_range = plot.x_range(trace.axis_x);
+    f64 lo = (x_range.min - series.x0()) / dx;
+    f64 hi = (x_range.max - series.x0()) / dx;
+    if (hi < lo) std::swap(lo, hi);
+    const u64 first = clamp_index(std::floor(lo) - 1.0, n);
+    const u64 last = clamp_index(std::ceil(hi) + 1.0, n);
+    if (last <= first) continue;
+
+    // This is the whole performance contract: the window, never the series.
+    const u64 visible = last - first;
+    const f64 samples_per_pixel = static_cast<f64>(visible) / static_cast<f64>(layout.axes.w);
+    const u32 level = samples_per_pixel >= static_cast<f64>(style_.dense_threshold_spp)
+                          ? series.pyramid().level_for(n, samples_per_pixel)
+                          : 0u;
+
+    gather_.clear();
+    series.gather(level, first, last, gather_);
+    if (gather_.size() < 4u) continue;  // fewer than two points is no segment
+
+    stats_.visible_samples += visible;
+    if (level == 0u) {
+      ++stats_.raw_traces;
+    } else {
+      ++stats_.lod_traces;
+    }
+
+    stage_xy(gather_.data(), static_cast<u32>(gather_.size() / 2u),
+             plot.x_axis(trace.axis_x).log, plot.y_axis(trace.axis_y).log);
+    flush_lines(mapping(plot, trace.axis_x, trace.axis_y), trace.color, trace.width);
+  }
+}
+
+void Renderer::draw_cursors(const Plot& plot) {
+  const PlotLayout& layout = plot.layout();
+  const Range y_range = plot.y_range(0);
+  const bool log_y = plot.y_axis(0).log;
+
+  for (u32 i = 0; i < plot.cursor_count(); ++i) {
+    const Cursor& cursor = plot.cursor(i);
+    if (!cursor.visible) continue;
+    const Transform transform = plot.x_transform(cursor.axis_x);
+    const f32 px = static_cast<f32>(transform.to_screen_x(cursor.x));
+    // A cursor scrolled out of the visible x range is off the plot: drawing it
+    // would put a marker over the label strip.
+    if (px < layout.axes.x || px >= layout.axes.right()) continue;
+
+    stage_segment(cursor.x, y_range.min, cursor.x, y_range.max, plot.x_axis(cursor.axis_x).log,
+                  log_y);
+    flush_lines(mapping(plot, cursor.axis_x, 0), cursor.color, style_.line_width_px);
+    if (!cursor.label.empty()) {
+      stage_text({px + 3.0f, layout.axes.y + 2.0f}, cursor.label.c_str());
+    }
+  }
+}
+
+void Renderer::draw_measures(const Plot& plot) {
+  const PlotLayout& layout = plot.layout();
+  const Rect& axes = layout.axes;
+  const Range y_range = plot.y_range(0);
+  const bool log_y = plot.y_axis(0).log;
+  const f32 line_h = font_backend_.height();
+
+  for (u32 i = 0; i < plot.measure_count(); ++i) {
+    const MeasureSpan span = plot.measure_span(i);
+    const Measurement& measure = plot.measurement(i);
+    if (!span.valid || !measure.valid) continue;
+
+    const Cursor& a = plot.cursor(span.cursor_a);
+    const Cursor& b = plot.cursor(span.cursor_b);
+    const Transform transform = plot.x_transform(a.axis_x);
+    f64 left_px = transform.to_screen_x(a.x);
+    f64 right_px = transform.to_screen_x(b.x);
+    if (right_px < left_px) std::swap(left_px, right_px);
+    const f32 x0 = std::max(static_cast<f32>(left_px), axes.x);
+    const f32 x1 = std::min(static_cast<f32>(right_px), axes.right());
+
+    if (x1 > x0) {
+      // The shaded band: one vertical stroke per pixel column of the interval.
+      // The line pipeline has no filled primitive, and the band is bounded by the
+      // plot's width, so a per-column fill stays O(pixels) and needs no second
+      // pipeline. Strokes are staged in data space like everything else, so the
+      // band follows a pan or a zoom through the uniform alone.
+      for (f32 px = x0; px < x1; px += 1.0f) {
+        const f64 x = transform.to_data_x(static_cast<f64>(px) + 0.5);
+        stage_segment(x, y_range.min, x, y_range.max, plot.x_axis(a.axis_x).log, log_y);
+      }
+      flush_lines(mapping(plot, a.axis_x, 0), style_.measure_fill, 1.0f);
+    }
+
+    char dx[kNumberChars];
+    char dy[kNumberChars];
+    char rms[kNumberChars];
+    char mean[kNumberChars];
+    format_si(measure.dx, dx, kNumberChars);
+    format_si(measure.dy, dy, kNumberChars);
+    format_si(measure.rms, rms, kNumberChars);
+    format_si(measure.mean, mean, kNumberChars);
+    char readout[kReadoutChars];
+    std::snprintf(readout, kReadoutChars, "dx %s  dy %s  rms %s  mean %s", dx, dy, rms, mean);
+
+    const f32 width = font_backend_.measure(readout);
+    const f32 limit = std::max(axes.x + 2.0f, axes.right() - width - 2.0f);
+    const f32 left = std::clamp(0.5f * (x0 + x1) - 0.5f * width, axes.x + 2.0f, limit);
+    // The second line of the plot area, not the first: the first belongs to the
+    // y axis title (and to cursor labels), and a wide readout clamped to the left
+    // edge would overprint them.
+    stage_text({left, axes.y + line_h + 2.0f}, readout);
+  }
+}
+
+void Renderer::draw_annotations(const Plot& plot) {
+  const f32 line_h = font_backend_.height();
+  for (u32 i = 0; i < plot.annotation_count(); ++i) {
+    const Annotation& annotation = plot.annotation(i);
+    if (!annotation.visible) continue;
+    const f32 px = static_cast<f32>(plot.x_transform(annotation.axis_x).to_screen_x(annotation.x));
+    const f32 py = static_cast<f32>(plot.y_transform(annotation.axis_y).to_screen_y(annotation.y));
+
+    // A small cross pins the point; the caption sits to its right. Both are in
+    // framebuffer pixels, hence the identity mapping for the marker.
+    lines_.add_segment({px - 3.0f, py}, {px + 3.0f, py});
+    lines_.add_segment({px, py - 3.0f}, {px, py + 3.0f});
+    flush_lines(pixel_mapping(), annotation.color, 1.0f);
+    if (!annotation.text.empty()) {
+      stage_text({px + 5.0f, py - 0.5f * line_h}, annotation.text.c_str());
+    }
+  }
+}
+
+// -------------------------------------------------------------------- dump
+
+bool Renderer::dump_frame_ppm(const char* path, Status* st) const noexcept {
+  if (path == nullptr || path[0] == '\0') {
+    return fail(st, ErrorCode::InvalidArgument, "PPM path is empty");
+  }
+  if (!ready_) {
+    return fail(st, ErrorCode::Unsupported, "renderer was never initialised");
+  }
+  const u32 width = static_cast<u32>(gl_viewport_.w);
+  const u32 height = static_cast<u32>(gl_viewport_.h);
+  if (width == 0 || height == 0) {
+    return fail(st, ErrorCode::InvalidArgument, "no frame has been drawn yet");
+  }
+
+  // RGBA, not RGB: 4-byte pixels have no row padding whatever the pack alignment,
+  // so a host's GL_PACK_ALIGNMENT cannot make the read overrun the buffer.
+  readback_.assign(usize(width) * height * 4u, 0u);
+  gl::glPixelStorei(gl::GL_PACK_ALIGNMENT, 4);
+  gl::glReadPixels(static_cast<gl::GLint>(gl_viewport_.x), static_cast<gl::GLint>(gl_viewport_.y),
+                   static_cast<gl::GLsizei>(width), static_cast<gl::GLsizei>(height), gl::GL_RGBA,
+                   gl::GL_UNSIGNED_BYTE, readback_.data());
+
+  std::FILE* file = std::fopen(path, "wb");
+  if (file == nullptr) {
+    return fail(st, ErrorCode::InvalidArgument, "cannot open the PPM path for writing");
+  }
+  std::fprintf(file, "P6\n%u %u\n255\n", width, height);
+  std::vector<u8> row(usize(width) * 3u);
+  bool ok = true;
+  for (u32 y = 0; y < height && ok; ++y) {
+    // glReadPixels hands back rows bottom-up; a PPM's rows run top-down.
+    const u8* src = readback_.data() + usize(height - 1u - y) * usize(width) * 4u;
+    for (u32 x = 0; x < width; ++x) {
+      row[usize(x) * 3u] = src[usize(x) * 4u];
+      row[usize(x) * 3u + 1u] = src[usize(x) * 4u + 1u];
+      row[usize(x) * 3u + 2u] = src[usize(x) * 4u + 2u];
+    }
+    ok = std::fwrite(row.data(), 1, row.size(), file) == row.size();
+  }
+  if (std::fclose(file) != 0) ok = false;
+  if (!ok) {
+    return fail(st, ErrorCode::InvalidArgument, "writing the PPM failed");
+  }
+  return true;
+}
+
+}  // namespace plot

src/render/vertex_array.cpp

@@ -14,11 +14,11 @@ constexpr gl::GLuint kPerInstance = 1;
 
 }  // namespace
 
-bool LineVertexArray::create(Status* st) noexcept {
+bool LineVertexArray::create(Status* st, UploadMode mode) noexcept {
   if (!vao_.create(st)) {
     return false;
   }
-  if (!buffer_.create(st)) {
+  if (!buffer_.create(st, mode)) {
     vao_.reset();  // do not leave a vertex array behind on the failure path
     return false;
   }
@@ -32,6 +32,7 @@ bool LineVertexArray::create(Status* st) noexcept {
   gl::glVertexAttribDivisor(0u, kPerInstance);
   gl::glVertexAttribDivisor(1u, kPerInstance);
   gl::glBindVertexArray(0);  // leave the binding as we found it
+  epoch_ = buffer_.epoch();
   return true;
 }
 
@@ -43,14 +44,26 @@ void LineVertexArray::upload(const f32* interleaved, u32 segments) noexcept {
     return;
   }
   buffer_.upload(interleaved, segments);
+  // Only the mapped transfer moves the instances between slots, and the
+  // attribute offsets have to follow them. The generation changes for a new slot
+  // or a re-created buffer, never otherwise, so the baseline path re-points at
+  // most once.
+  if (buffer_.epoch() != epoch_) {
+    const usize base = buffer_.slot_offset();
+    vao_.bind();
+    buffer_.bind();
+    vao_.attrib(0u, 2, sizeof(LineInstance), base + offsetof(LineInstance, x0));
+    vao_.attrib(1u, 2, sizeof(LineInstance), base + offsetof(LineInstance, x1));
+    epoch_ = buffer_.epoch();
+  }
   instances_ = segments;
 }
 
-bool TextVertexArray::create(Status* st) noexcept {
+bool TextVertexArray::create(Status* st, UploadMode mode) noexcept {
   if (!vao_.create(st)) {
     return false;
   }
-  if (!buffer_.create(st)) {
+  if (!buffer_.create(st, mode)) {
     vao_.reset();
     return false;
   }
@@ -61,6 +74,7 @@ bool TextVertexArray::create(Status* st) noexcept {
   gl::glVertexAttribDivisor(0u, kPerInstance);
   gl::glVertexAttribDivisor(1u, kPerInstance);
   gl::glBindVertexArray(0);
+  epoch_ = buffer_.epoch();
   return true;
 }
 
@@ -70,6 +84,14 @@ void TextVertexArray::upload(const f32* interleaved, u32 glyphs) noexcept {
     return;
   }
   buffer_.upload(interleaved, glyphs);
+  if (buffer_.epoch() != epoch_) {
+    const usize base = buffer_.slot_offset();
+    vao_.bind();
+    buffer_.bind();
+    vao_.attrib(0u, 2, sizeof(TextInstance), base + offsetof(TextInstance, x));
+    vao_.attrib(1u, 2, sizeof(TextInstance), base + offsetof(TextInstance, u));
+    epoch_ = buffer_.epoch();
+  }
   instances_ = glyphs;
 }