Note the full view's speed on real hardware

Co-Authored-By: Claude Opus 5.5 <[email protected]>

.gitea/scripts/release.py

@@ -0,0 +1,152 @@
+"""Publish the renderer's programs. Run by ci.yml.
+
+    release.py version         print version and channel
+    release.py publish <dir>   publish the programs in dir: a release for a vX.Y.Z tag, else the rolling nightly
+
+A vX.Y.Z tag is release X.Y.Z. Other builds are nightlies, "X.Y.Z-nightly.N+commit": N commits after the last
+release tag (0.0.0 before the first). The nightly pre-release (tag `nightly`) is replaced on every push to main.
+The same scheme as T5DE's own releases (t5de's .gitea/scripts/release.py).
+"""
+
+import hashlib
+import json
+import os
+import re
+import subprocess
+import sys
+import urllib.error
+import urllib.request
+import uuid
+from pathlib import Path
+
+ASSETS = ("t5deRenderEngine.dll", "t5de-renderer.exe")
+NIGHTLY_TAG = "nightly"
+USER_AGENT = "t5de-release (https://git.t5de.dev/t5de/t5de-renderer, 1)"
+RELEASE_TAG = re.compile(r"v(\d+)\.(\d+)\.(\d+)")
+
+
+def git(*args: str) -> str:
+    return subprocess.run(["git", *args], capture_output=True, text=True, check=True).stdout.strip()
+
+
+def version() -> dict[str, str]:
+    ref_name = os.environ.get("GITHUB_REF_NAME", "")
+    if os.environ.get("GITHUB_REF_TYPE") == "tag" and RELEASE_TAG.fullmatch(ref_name):
+        return {"version": ref_name[1:], "channel": "release", "tag": ref_name}
+    try:
+        last = git("describe", "--tags", "--abbrev=0", "--match", "v[0-9]*.[0-9]*.[0-9]*")
+        count = int(git("rev-list", "--count", f"{last}..HEAD"))
+        base = last[1:]
+    except subprocess.CalledProcessError:  # no release yet
+        count, base = int(git("rev-list", "--count", "HEAD")), "0.0.0"
+    commit = git("rev-parse", "--short=7", "HEAD")
+    return {"version": f"{base}-nightly.{count}+{commit}", "channel": "nightly", "tag": NIGHTLY_TAG}
+
+
+class Gitea:
+    def __init__(self):
+        self.api = f"{os.environ['GITHUB_SERVER_URL']}/api/v1/repos/{os.environ['GITHUB_REPOSITORY']}"
+        self.token = os.environ["GITEA_TOKEN"]
+
+    def request(self, method: str, path: str, body: dict | None = None, data: bytes | None = None,
+                content_type: str = "application/json"):
+        if body is not None:
+            data = json.dumps(body).encode()
+        req = urllib.request.Request(self.api + path, method=method, data=data,
+                                     # Cloudflare in front of the server refuses Python's default User-Agent.
+                                     headers={"Authorization": f"token {self.token}", "Content-Type": content_type,
+                                              "User-Agent": USER_AGENT})
+        try:
+            with urllib.request.urlopen(req, timeout=120) as response:
+                text = response.read()
+                return json.loads(text) if text else None
+        except urllib.error.HTTPError as e:
+            if e.code == 404 and method in ("GET", "DELETE"):
+                return None
+            raise SystemExit(f"{method} {path}: {e.code} {e.read().decode(errors='replace')}")
+
+    def upload(self, release_id: int, name: str, content: bytes):
+        boundary = uuid.uuid4().hex
+        data = (f"--{boundary}\r\nContent-Disposition: form-data; name=\"attachment\"; filename=\"{name}\"\r\n"
+                f"Content-Type: application/octet-stream\r\n\r\n").encode() + content + f"\r\n--{boundary}--\r\n".encode()
+        return self.request("POST", f"/releases/{release_id}/assets?name={name}", data=data,
+                            content_type=f"multipart/form-data; boundary={boundary}")
+
+
+def notes(info: dict[str, str], files: dict[str, bytes]) -> str:
+    nightly = ("A nightly build of the latest commit, for testing; it's replaced by the next one.\n\n"
+               if info["channel"] == "nightly" else "")
+    hashes = "\n".join(f"`{name}` sha256: `{hashlib.sha256(content).hexdigest()}`" for name, content in files.items())
+    return f"""{nightly}T5DE's renderer for IMVU Classic: copy both files into a T5DE install's folder, and choose OpenGL
+in the client's Settings > Advanced (with T5DE's renderer patch).
+
+Unofficial: not made by or affiliated with IMVU. Use at your own risk.
+
+Version {info['version']} · commit {os.environ.get('GITHUB_SHA', git('rev-parse', 'HEAD'))}
+{hashes}
+"""
+
+
+def publish(folder: str):
+    info = version()
+    files = {name: (Path(folder) / name).read_bytes() for name in ASSETS}
+    gitea = Gitea()
+    previous = None
+    if info["channel"] == "nightly":
+        old = gitea.request("GET", f"/releases/tags/{NIGHTLY_TAG}")
+        if old:
+            previous = old["target_commitish"]
+            gitea.request("DELETE", f"/releases/{old['id']}")
+        gitea.request("DELETE", f"/tags/{NIGHTLY_TAG}")
+        release = gitea.request("POST", "/releases", {
+            "tag_name": NIGHTLY_TAG, "target_commitish": os.environ["GITHUB_SHA"],
+            "name": f"Nightly {info['version']}", "body": notes(info, files), "prerelease": True})
+    else:
+        release = gitea.request("GET", f"/releases/tags/{info['tag']}")
+        fields = {"name": f"T5DE Renderer {info['version']}", "body": notes(info, files), "prerelease": False}
+        if release:  # a rerun: refresh it
+            release = gitea.request("PATCH", f"/releases/{release['id']}", fields)
+            for asset in release.get("assets") or []:
+                if asset["name"] in files:
+                    gitea.request("DELETE", f"/releases/{release['id']}/assets/{asset['id']}")
+        else:
+            release = gitea.request("POST", "/releases", {"tag_name": info["tag"], **fields})
+    for name, content in files.items():
+        gitea.upload(release["id"], name, content)
+    print(f"Published {release['html_url']}")
+    if info["channel"] == "nightly" and os.environ.get("DISCORD_NIGHTLY_WEBHOOK"):
+        announce(os.environ["DISCORD_NIGHTLY_WEBHOOK"], info, release, previous)
+
+
+def announce(webhook: str, info: dict[str, str], release: dict, previous: str | None):
+    """Post the release to a Discord channel. Failing to doesn't fail the release."""
+    commits = []
+    if previous:
+        try:
+            commits = git("log", "--format=%h %s", "--no-merges", f"{previous}..HEAD").splitlines()
+        except subprocess.CalledProcessError:  # the previous nightly's commit is gone (history rewritten)
+            pass
+    shown = commits[:10] + ([f"and {len(commits) - 10} more"] if len(commits) > 10 else [])
+    description = f"[Download]({release['html_url']})"
+    if shown:
+        description += "\n\n" + "\n".join(f"• {line}" for line in shown)
+    body = {"embeds": [{"title": f"Renderer Nightly {info['version']}", "url": release["html_url"],
+                        "description": description[:4000], "color": 0xFFD200}]}
+    req = urllib.request.Request(webhook, method="POST", data=json.dumps(body).encode(),
+                                 headers={"Content-Type": "application/json",
+                                          "User-Agent": USER_AGENT})
+    try:
+        urllib.request.urlopen(req, timeout=30).close()
+        print("Announced on Discord")
+    except (urllib.error.URLError, TimeoutError) as e:
+        print(f"warning: couldn't post to Discord: {e}")
+
+
+if __name__ == "__main__":
+    if sys.argv[1:] == ["version"]:
+        for key, value in version().items():
+            print(f"{key}={value}")
+    elif len(sys.argv) == 3 and sys.argv[1] == "publish":
+        publish(sys.argv[2])
+    else:
+        raise SystemExit(__doc__)

.gitea/workflows/ci.yml

@@ -0,0 +1,59 @@
+# Builds the adapter (32-bit) and the renderer process (64-bit), runs the tests, and publishes both programs: pushes to
+# main replace the nightly pre-release, vX.Y.Z tags make release X.Y.Z (.gitea/scripts/release.py).
+#
+# Runs on T5DE's Windows runner (label `windows`, jobs run directly on the machine: it needs Visual Studio's
+# compiler). The checkout lives in a fixed folder there ($T5DE_BUILD_ROOT, set in the runner's config), so Diligent
+# Engine is fetched and built once, not on every run. Never on pull requests: their code would run on that machine.
+name: CI
+
+on:
+  push:
+    branches: [main]
+    tags: ["v*"]
+  workflow_dispatch:
+
+jobs:
+  build:
+    runs-on: windows
+    defaults:
+      run:
+        # Steps run as script files, which Windows' default execution policy refuses.
+        shell: powershell -NoProfile -ExecutionPolicy Bypass -Command ". '{0}'"
+    steps:
+      - name: Check out
+        env:
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          function git { & git.exe -c "http.extraHeader=Authorization: token $env:TOKEN" @args; if ($LASTEXITCODE) { exit $LASTEXITCODE } }
+          $root = if ($env:T5DE_BUILD_ROOT) { $env:T5DE_BUILD_ROOT } else { $env:GITHUB_WORKSPACE }
+          $src = Join-Path $root ($env:GITHUB_REPOSITORY -replace '/', '-')
+          if (-not (Test-Path (Join-Path $src .git))) { git init -q $src }
+          Set-Location $src
+          git fetch -q --force --tags "$env:GITHUB_SERVER_URL/$env:GITHUB_REPOSITORY" "+${env:GITHUB_REF}:refs/remotes/ci/head"
+          git checkout -q --force --detach $env:GITHUB_SHA
+          git clean -q -ffdx -e /build/
+          "T5DE_SRC=$src" | Out-File -Append -Encoding utf8 $env:GITHUB_ENV
+
+      - name: Build and test the adapter
+        working-directory: ${{ env.T5DE_SRC }}
+        run: cmake --workflow --preset adapter
+
+      - name: Build the renderer process
+        working-directory: ${{ env.T5DE_SRC }}
+        run: cmake --workflow --preset renderer
+
+      - name: Collect the programs
+        working-directory: ${{ env.T5DE_SRC }}
+        run: |
+          cmake --install build/windows-x86 --config Release --prefix dist
+          if ($LASTEXITCODE) { exit $LASTEXITCODE }
+          cmake --install build/windows-x64 --config Release --prefix dist
+
+      - name: Publish the release
+        if: github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v')
+        working-directory: ${{ env.T5DE_SRC }}
+        env:
+          GITEA_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          # Optional: a Discord webhook URL that nightlies are announced on.
+          DISCORD_NIGHTLY_WEBHOOK: ${{ secrets.DISCORD_NIGHTLY_WEBHOOK }}
+        run: python .gitea/scripts/release.py publish dist

adapter/CMakeLists.txt

@@ -1,11 +1,14 @@
 # t5deRenderEngine.dll, the adapter: a renderer plugin for IMVU Classic (32-bit).
 
 add_library(t5de_render_engine SHARED
+    engine.cpp
     log.cpp
-    passthrough_engine.cpp
+    pixmap.cpp
     plugin.cpp
     plugin.def
-    renderer_process.cpp)
+    recorder.cpp
+    renderer_process.cpp
+    resources.cpp)
 set_target_properties(t5de_render_engine PROPERTIES OUTPUT_NAME t5deRenderEngine)
 target_link_libraries(t5de_render_engine PRIVATE t5de::imvu t5de::protocol t5de::options)
 

adapter/engine.cpp

@@ -0,0 +1,372 @@
+#include "engine.hpp"
+
+#include <windows.h>
+
+#include <bit>
+#include <cmath>
+#include <cstring>
+#include <vector>
+
+#include "log.hpp"
+#include "resources.hpp"
+
+namespace t5de {
+
+using protocol::Op;
+using protocol::State;
+
+namespace {
+
+protocol::SetTransform transform_command(const imvu::Transform& t)
+{
+    protocol::SetTransform command{};
+    std::memcpy(command.m, &t.linear[0][0], sizeof(float) * 9);
+    command.m[9] = t.translation.x;
+    command.m[10] = t.translation.y;
+    command.m[11] = t.translation.z;
+    return command;
+}
+
+} // namespace
+
+Engine::Engine(imvu::SharedPtr<imvu::PiRenderEngine> inner, std::shared_ptr<Recorder> recorder)
+    : inner_(std::move(inner)), recorder_(std::move(recorder))
+{
+}
+
+Engine::~Engine() { log::info("engine: destroyed after {} frames", frames_); }
+
+void Engine::setState(State state, std::uint32_t value)
+{
+    recorder_->add(Op::SetState, protocol::SetState{state, value});
+}
+
+void Engine::resizeFrameBuffer(imvu::Vec2i size) { inner_->resizeFrameBuffer(size); }
+
+void Engine::enable2dFrameBuffer() { inner_->enable2dFrameBuffer(); }
+
+void Engine::clear(std::uint32_t flags, imvu::Color color)
+{
+    if (innerDraws()) {
+        inner_->clear(flags, color);
+    }
+    recorder_->add(Op::Clear, protocol::Clear{flags, color});
+}
+
+void Engine::setProjection(imvu::Vec2i viewport, const imvu::Mat4f& projection)
+{
+    inner_->setProjection(viewport, projection);
+    if (projection.m[2][3] == 0.0f) {
+        ++stats_.flat;
+    }
+    protocol::SetProjection command{viewport.x, viewport.y, {}};
+    std::memcpy(command.m, &projection.m[0][0], sizeof(command.m));
+    recorder_->add(Op::SetProjection, command);
+}
+
+void Engine::setModelSpace(const imvu::Transform& model)
+{
+    inner_->setModelSpace(model);
+    recorder_->add(Op::SetModel, transform_command(model));
+}
+
+void Engine::setViewSpace(const imvu::Transform& view)
+{
+    inner_->setViewSpace(view);
+    recorder_->add(Op::SetView, transform_command(view));
+}
+
+void Engine::setTextures(imvu::PiRenderTexture* texture0, imvu::PiRenderTexture* texture1)
+{
+    inner_->setTextures(inner(texture0), inner(texture1));
+    for (imvu::PiRenderTexture* texture : {texture0, texture1}) {
+        if (texture != nullptr && id(texture) == 0 && foreignLogged_ < 20) {
+            ++foreignLogged_;
+            log::info("engine: frame {}: a texture the client renderer made ({})", frames_, static_cast<void*>(texture));
+        }
+    }
+    recorder_->add(Op::SetTextures, protocol::SetTextures{id(texture0), id(texture1)});
+}
+
+void Engine::setDepthTestingOn(bool on)
+{
+    inner_->setDepthTestingOn(on);
+    setState(State::DepthTest, on);
+}
+
+void Engine::setDepthWritingOn(bool on)
+{
+    inner_->setDepthWritingOn(on);
+    setState(State::DepthWrite, on);
+}
+
+void Engine::setBlendMode(imvu::RenderBlendMode mode)
+{
+    inner_->setBlendMode(mode);
+    setState(State::BlendMode, static_cast<std::uint32_t>(mode));
+}
+
+void Engine::setAlphaTestingOn(bool on)
+{
+    inner_->setAlphaTestingOn(on);
+    setState(State::AlphaTest, on);
+}
+
+void Engine::setAlphaTestValue(float value)
+{
+    inner_->setAlphaTestValue(value);
+    setState(State::AlphaReference, static_cast<std::uint32_t>(std::lround(value * 255.0f)));
+}
+
+void Engine::setBlendingOn(bool on)
+{
+    inner_->setBlendingOn(on);
+    setState(State::Blending, on);
+}
+
+void Engine::setBackFaceCullingOn(bool on)
+{
+    inner_->setBackFaceCullingOn(on);
+    setState(State::Culling, on);
+}
+
+void Engine::setNormalInvertingOn(bool on)
+{
+    inner_->setNormalInvertingOn(on);
+    setState(State::NormalInverting, on);
+}
+
+void Engine::setLightingOn(bool on)
+{
+    inner_->setLightingOn(on);
+    setState(State::Lighting, on);
+}
+
+void Engine::setLightProperties(std::uint32_t index, const imvu::LightProperties& light)
+{
+    inner_->setLightProperties(index, light);
+    protocol::SetLight command{index, light.enabled, light.diffuse,
+                               light.type == imvu::LightType::Directional,
+                               {light.vector.x, light.vector.y, light.vector.z}};
+    // IMVU's Direct3D renderer shines a directional light along -vector, along -x for a zero vector.
+    if (command.directional && light.vector.x == 0 && light.vector.y == 0 && light.vector.z == 0) {
+        command.vector[0] = 1;
+    }
+    recorder_->add(Op::SetLight, command);
+}
+
+void Engine::setFogOn(bool on)
+{
+    inner_->setFogOn(on);
+    setState(State::Fog, on);
+}
+
+void Engine::setFogAttributes(float start, float end, imvu::Color color)
+{
+    inner_->setFogAttributes(start, end, color);
+    recorder_->add(Op::SetFog, protocol::SetFog{start, end, color});
+}
+
+void Engine::setTextureTransformMatrixIdentity(std::uint32_t stage)
+{
+    inner_->setTextureTransformMatrixIdentity(stage);
+    recorder_->add(Op::SetTextureTransform, protocol::SetTextureTransform{stage, 1, {1, 0, 0, 1, 0, 0}});
+}
+
+void Engine::setTextureTransformMatrix(std::uint32_t stage, const imvu::TextureTransform& transform)
+{
+    inner_->setTextureTransformMatrix(stage, transform);
+    protocol::SetTextureTransform command{stage, 0, {}};
+    std::memcpy(command.m, transform.m, sizeof(command.m));
+    recorder_->add(Op::SetTextureTransform, command);
+}
+
+void Engine::setEnvironmentAmbient(imvu::Color color)
+{
+    inner_->setEnvironmentAmbient(color);
+    setState(State::Ambient, color);
+}
+
+imvu::PiRenderTexture* Engine::newRenderTexture(const imvu::Pixmap& pixmap, std::uint32_t flags)
+{
+    imvu::PiRenderTexture* texture = inner_->newRenderTexture(pixmap, flags);
+    if (texture == nullptr) {
+        return nullptr;
+    }
+    return new Texture(texture, recorder_, recorder_->createTexture(pixmap, flags));
+}
+
+imvu::PiInternalIndexBuffer* Engine::newIndexBuffer(std::uint32_t count, const std::uint16_t* indices)
+{
+    imvu::PiInternalIndexBuffer* buffer = inner_->newIndexBuffer(count, indices);
+    if (buffer == nullptr) {
+        return nullptr;
+    }
+    return new IndexBuffer(buffer, recorder_,
+                           recorder_->createBuffer(protocol::BufferKind::Index, indices, 2, 2, count));
+}
+
+imvu::PiInternalVertexBuffer* Engine::newStaticVertexBuffer(const void* data, std::uint32_t stride,
+                                                            std::uint32_t elementSize, std::uint32_t elementCount)
+{
+    imvu::PiInternalVertexBuffer* buffer = inner_->newStaticVertexBuffer(data, stride, elementSize, elementCount);
+    if (buffer == nullptr) {
+        return nullptr;
+    }
+    return new VertexBuffer(
+        buffer, recorder_,
+        recorder_->createBuffer(protocol::BufferKind::Vertex, data, stride, elementSize, elementCount));
+}
+
+imvu::PiInternalVertexBuffer* Engine::newDynamicVertexBuffer(std::uint32_t elementSize, std::uint32_t elementCount)
+{
+    imvu::PiInternalVertexBuffer* buffer = inner_->newDynamicVertexBuffer(elementSize, elementCount);
+    if (buffer == nullptr) {
+        return nullptr;
+    }
+    return new VertexBuffer(
+        buffer, recorder_,
+        recorder_->createBuffer(protocol::BufferKind::Vertex, nullptr, 0, elementSize, elementCount));
+}
+
+imvu::SharedPtr<imvu::PiPixmapRenderTarget> Engine::newPixmapRenderTarget(imvu::Vec2i size)
+{
+    log::info("engine: frame {}: a {}x{} pixmap render target", frames_, size.x, size.y);
+    return inner_->newPixmapRenderTarget(size);
+}
+
+imvu::SharedPtr<imvu::PiTextureRenderTarget> Engine::newTextureRenderTarget(imvu::Vec2i size)
+{
+    log::info("engine: frame {}: a {}x{} texture render target", frames_, size.x, size.y);
+    return inner_->newTextureRenderTarget(size);
+}
+
+void Engine::beginRendering(imvu::PiRenderTarget* target)
+{
+    inner_->beginRendering(target);
+    // Decided per frame: frames into render targets, and frames while the renderer process doesn't cover the view, are
+    // drawn by the client's renderer.
+    const bool takenOver = target == nullptr && recorder_->covering();
+    if (takenOver != takenOver_ && target == nullptr) {
+        log::info("engine: frame {}: {}", frames_, takenOver ? "drawn by the renderer process" : "drawn by the client's renderer");
+    }
+    takenOver_ = takenOver;
+    if (target != nullptr && foreignLogged_ < 20) {
+        ++foreignLogged_;
+        log::info("engine: frame {}: drawing into a render target ({})", frames_, static_cast<void*>(target));
+    }
+    recorder_->add(Op::BeginFrame, protocol::BeginFrame{target != nullptr});
+}
+
+void Engine::endRendering()
+{
+    inner_->endRendering();
+    LARGE_INTEGER now;
+    QueryPerformanceCounter(&now);
+    recorder_->add(Op::EndFrame, protocol::EndFrame{static_cast<std::uint64_t>(now.QuadPart)});
+    recorder_->flush(true);
+    ++frames_;
+    if (frames_ % 300 == 0 && statsLogged_ < 6) {
+        ++statsLogged_;
+        log::info("engine: last 300 frames: drawTriangles {} (and {} with buffers not ours), drawTrianglesFromMemory {} "
+                  "(and {} without indices), point sprites {}, lines {}, 2D projections {}",
+                  stats_.triangles, stats_.foreignTriangles, stats_.indexed, stats_.unindexed, stats_.sprites, stats_.lines,
+                  stats_.flat);
+        stats_ = {};
+    }
+}
+
+void Engine::drawTriangles(imvu::PiInternalIndexBuffer* indices, imvu::PiInternalVertexBuffer* positions,
+                           imvu::PiInternalVertexBuffer* normals, imvu::PiInternalVertexBuffer* texCoords,
+                           imvu::PiInternalVertexBuffer* colors)
+{
+    if (innerDraws()) {
+        inner_->drawTriangles(inner(indices), inner(positions), inner(normals), inner(texCoords), inner(colors));
+    }
+    ++(id(indices) != 0 && id(positions) != 0 ? stats_.triangles : stats_.foreignTriangles);
+    recorder_->add(Op::DrawTriangles,
+                   protocol::DrawTriangles{id(indices), id(positions), id(normals), id(texCoords), id(colors)});
+}
+
+void Engine::drawLine(imvu::Vec3f from, imvu::Vec3f to, imvu::Color color)
+{
+    if (innerDraws()) {
+        inner_->drawLine(from, to, color);
+    }
+    ++stats_.lines;
+}
+
+void Engine::drawPointSprites(const imvu::PointSpriteVector& sprites)
+{
+    if (innerDraws()) {
+        inner_->drawPointSprites(sprites);
+    }
+    ++stats_.sprites;
+    if (sprites.first == nullptr || sprites.last <= sprites.first) {
+        return;
+    }
+    const auto count = static_cast<std::uint32_t>(sprites.last - sprites.first);
+    if (std::byte* payload = recorder_->add(Op::DrawPointSprites, protocol::DrawPointSprites{count},
+                                            std::size_t{count} * sizeof(imvu::PointSprite))) {
+        std::memcpy(payload, sprites.first, std::size_t{count} * sizeof(imvu::PointSprite));
+    }
+}
+
+void Engine::drawTrianglesFromMemory(std::uint32_t vertexCount, std::uint32_t indexCount,
+                                     const std::uint16_t* indices,
+                                     const imvu::VertexAttributeArray<imvu::Vec3f>& positions,
+                                     const imvu::VertexAttributeArray<imvu::Vec3f>& normals,
+                                     const imvu::VertexAttributeArray<imvu::Vec2f>& texCoords,
+                                     const imvu::VertexAttributeArray<imvu::Color>& colors)
+{
+    if (innerDraws()) {
+        inner_->drawTrianglesFromMemory(vertexCount, indexCount, indices, positions, normals, texCoords, colors);
+    }
+    ++(indices != nullptr ? stats_.indexed : stats_.unindexed);
+    if (positions.data == nullptr || vertexCount == 0) {
+        return;
+    }
+    // Without indices: the vertices in order, as a triangle list.
+    std::vector<std::uint16_t> sequence;
+    if (indices == nullptr) {
+        if (vertexCount > 0xffff) {
+            return;
+        }
+        sequence.resize(vertexCount - vertexCount % 3);
+        for (std::size_t i = 0; i < sequence.size(); ++i) {
+            sequence[i] = static_cast<std::uint16_t>(i);
+        }
+        indices = sequence.data();
+        indexCount = static_cast<std::uint32_t>(sequence.size());
+    }
+    const bool hasNormals = normals.data != nullptr && normals.count >= vertexCount;
+    const bool hasTexCoords = texCoords.data != nullptr && texCoords.count >= vertexCount;
+    const bool hasColors = colors.data != nullptr && colors.count >= vertexCount;
+    const std::size_t indexBytes = protocol::padded(std::size_t{indexCount} * 2);
+    const std::size_t size = indexBytes + std::size_t{vertexCount} * (12 + (hasNormals ? 12 : 0) +
+                                                                      (hasTexCoords ? 8 : 0) + (hasColors ? 4 : 0));
+    std::byte* payload = recorder_->add(Op::DrawArrays,
+                                        protocol::DrawArrays{vertexCount, indexCount, hasNormals, hasTexCoords,
+                                                             hasColors},
+                                        size);
+    if (payload == nullptr) {
+        return;
+    }
+    std::memcpy(payload, indices, std::size_t{indexCount} * 2);
+    std::byte* next = payload + indexBytes;
+    pack(next, positions.data, positions.stride, 12, vertexCount);
+    next += std::size_t{vertexCount} * 12;
+    if (hasNormals) {
+        pack(next, normals.data, normals.stride, 12, vertexCount);
+        next += std::size_t{vertexCount} * 12;
+    }
+    if (hasTexCoords) {
+        pack(next, texCoords.data, texCoords.stride, 8, vertexCount);
+        next += std::size_t{vertexCount} * 8;
+    }
+    if (hasColors) {
+        pack(next, colors.data, colors.stride, 4, vertexCount);
+    }
+}
+
+} // namespace t5de

adapter/engine.hpp

@@ -1,21 +1,24 @@
 #pragma once
-// A renderer that forwards every call to another one (IMVU's Direct3D engine) and logs what a frame is made of: proof
+// The adapter's engine: what the client draws with. Everything is recorded for the renderer process, which draws the
-// that our declaration of the interface matches IMVU's, and a way to see how the client draws.
+// 3D view. The client's own Direct3D renderer gets everything too (state and resources: it still draws into render
+// targets and reads textures back, which the renderer process doesn't yet), but while the renderer process covers the
+// whole view, the client's renderer skips its frames' clears and draws. When the renderer process doesn't cover it
+// (it failed, or shows only part of the view, for comparing) the client's renderer draws everything, from the next
+// frame on.
 
-#include <array>
 #include <cstdint>
 #include <memory>
 
-#include "renderer_process.hpp"
+#include "recorder.hpp"
 #include "t5de/imvu/render_engine.hpp"
 
 namespace t5de {
 
-class PassthroughEngine final : public imvu::PiRenderEngine {
+class Engine final : public imvu::PiRenderEngine {
 public:
-    // renderer may be null.
+    // recorder may be inactive (no renderer process).
-    PassthroughEngine(imvu::SharedPtr<imvu::PiRenderEngine> inner, std::unique_ptr<RendererProcess> renderer);
+    Engine(imvu::SharedPtr<imvu::PiRenderEngine> inner, std::shared_ptr<Recorder> recorder);
-    ~PassthroughEngine() override;
+    ~Engine() override;
 
     void resizeFrameBuffer(imvu::Vec2i size) override;
     void enable2dFrameBuffer() override;
@@ -53,7 +56,7 @@ public:
     void drawTriangles(imvu::PiInternalIndexBuffer* indices, imvu::PiInternalVertexBuffer* positions,
                        imvu::PiInternalVertexBuffer* normals, imvu::PiInternalVertexBuffer* texCoords,
                        imvu::PiInternalVertexBuffer* colors) override;
-    void drawLine(imvu::Vec3f from, imvu::Vec3f to, std::uint32_t color) override;
+    void drawLine(imvu::Vec3f from, imvu::Vec3f to, imvu::Color color) override;
     void drawPointSprites(const imvu::PointSpriteVector& sprites) override;
     void drawTrianglesFromMemory(std::uint32_t vertexCount, std::uint32_t indexCount, const std::uint16_t* indices,
                                  const imvu::VertexAttributeArray<imvu::Vec3f>& positions,
@@ -61,16 +64,30 @@ public:
                                  const imvu::VertexAttributeArray<imvu::Vec2f>& texCoords,
                                  const imvu::VertexAttributeArray<imvu::Color>& colors) override;
 
-    // The interface's methods, by vtable slot (slot 0, the destructor, isn't counted).
-    static constexpr std::size_t slots = 35;
-
 private:
-    void count(std::size_t slot);
+    void setState(protocol::State state, std::uint32_t value);
+    // Whether the client's renderer should draw this frame's clears and draws.
+    [[nodiscard]] bool innerDraws() const { return !takenOver_; }
 
     imvu::SharedPtr<imvu::PiRenderEngine> inner_;
-    std::unique_ptr<RendererProcess> renderer_;
+    std::shared_ptr<Recorder> recorder_;
-    std::array<std::uint32_t, slots> calls_{};
     std::uint64_t frames_ = 0;
+    bool takenOver_ = false; // this frame is drawn by the renderer process only
+    int foreignLogged_ = 0; // diagnostics: textures the client renderer made, render targets
+
+    // Diagnostics: what the client draws, summed over 300 frames and logged (the first few times).
+    struct Stats {
+        std::uint32_t triangles = 0;        // drawTriangles with our buffers
+        std::uint32_t foreignTriangles = 0; // drawTriangles with buffers that aren't ours (not drawn)
+        std::uint32_t indexed = 0;          // drawTrianglesFromMemory with indices
+        std::uint32_t unindexed = 0;        // drawTrianglesFromMemory without
+        std::uint32_t sprites = 0;
+        std::uint32_t lines = 0;
+        std::uint32_t flat = 0; // setProjection with an orthographic projection (2D)
+    };
+    Stats stats_;
+    int statsLogged_ = 0;
+
 };
 
 } // namespace t5de

adapter/passthrough_engine.cpp

@@ -1,395 +0,0 @@
-#include "passthrough_engine.hpp"
-
-#include <windows.h>
-
-#include <atomic>
-#include <string>
-
-#include "log.hpp"
-
-namespace t5de {
-
-namespace {
-
-constexpr std::array<const char*, PassthroughEngine::slots> names = {
-    "~PiRenderEngine",
-    "resizeFrameBuffer",
-    "enable2dFrameBuffer",
-    "clear",
-    "setProjection",
-    "setModelSpace",
-    "setViewSpace",
-    "setTextures",
-    "setDepthTestingOn",
-    "setDepthWritingOn",
-    "setBlendMode",
-    "setAlphaTestingOn",
-    "setAlphaTestValue",
-    "setBlendingOn",
-    "setBackFaceCullingOn",
-    "setNormalInvertingOn",
-    "setLightingOn",
-    "setLightProperties",
-    "setFogOn",
-    "setFogAttributes",
-    "setTextureTransformMatrixIdentity",
-    "setTextureTransformMatrix",
-    "setEnvironmentAmbient",
-    "newRenderTexture",
-    "newIndexBuffer",
-    "newStaticVertexBuffer",
-    "newDynamicVertexBuffer",
-    "newPixmapRenderTarget",
-    "newTextureRenderTarget",
-    "beginRendering",
-    "endRendering",
-    "drawTriangles",
-    "drawLine",
-    "drawPointSprites",
-    "drawTrianglesFromMemory",
-};
-
-// Frames whose calls are logged: the first few, then one a minute or so at 30 FPS.
-bool logged(std::uint64_t frame) { return frame < 5 || frame % 1800 == 0; }
-
-// Whether to log a call's arguments: the first few calls of each method, in the whole process. Enough to check that
-// our layouts of IMVU's types read sensible values.
-bool sampled(std::size_t slot)
-{
-    static std::array<std::atomic<int>, PassthroughEngine::slots> calls{};
-    return calls[slot]++ < 3;
-}
-
-std::string text(const imvu::Vec3f& v) { return std::format("({:.3f}, {:.3f}, {:.3f})", v.x, v.y, v.z); }
-
-std::string text(const imvu::Transform& t)
-{
-    const auto& m = t.linear;
-    return std::format("[{:.3f} {:.3f} {:.3f}; {:.3f} {:.3f} {:.3f}; {:.3f} {:.3f} {:.3f}] + {}", m[0][0], m[0][1],
-                       m[0][2], m[1][0], m[1][1], m[1][2], m[2][0], m[2][1], m[2][2], text(t.translation));
-}
-
-// A pixmap's size, from pixmap.dll's exported Pixmap::width and Pixmap::height.
-imvu::Vec2i pixmap_size(const imvu::Pixmap& pixmap)
-{
-    using Getter = unsigned(__thiscall*)(const imvu::Pixmap*);
-    struct Getters {
-        Getter width;
-        Getter height;
-    };
-    static const Getters getters = [] {
-        HMODULE module = GetModuleHandleW(L"pixmap.dll");
-        if (module == nullptr) {
-            return Getters{};
-        }
-        return Getters{reinterpret_cast<Getter>(GetProcAddress(module, "?width@Pixmap@@QBEIXZ")),
-                       reinterpret_cast<Getter>(GetProcAddress(module, "?height@Pixmap@@QBEIXZ"))};
-    }();
-    if (getters.width == nullptr || getters.height == nullptr) {
-        return {-1, -1};
-    }
-    return {static_cast<int>(getters.width(&pixmap)), static_cast<int>(getters.height(&pixmap))};
-}
-
-} // namespace
-
-PassthroughEngine::PassthroughEngine(imvu::SharedPtr<imvu::PiRenderEngine> inner,
-                                     std::unique_ptr<RendererProcess> renderer)
-    : inner_(std::move(inner)), renderer_(std::move(renderer))
-{
-    log::info("passthrough: wrapping the engine at {}", static_cast<void*>(inner_.get()));
-}
-
-PassthroughEngine::~PassthroughEngine() { log::info("passthrough: destroyed after {} frames", frames_); }
-
-void PassthroughEngine::count(std::size_t slot) { ++calls_[slot]; }
-
-void PassthroughEngine::resizeFrameBuffer(imvu::Vec2i size)
-{
-    log::info("resizeFrameBuffer({}, {})", size.x, size.y);
-    count(1);
-    inner_->resizeFrameBuffer(size);
-}
-
-void PassthroughEngine::enable2dFrameBuffer()
-{
-    count(2);
-    inner_->enable2dFrameBuffer();
-}
-
-void PassthroughEngine::clear(std::uint32_t flags, imvu::Color color)
-{
-    if (sampled(3)) {
-        log::info("clear(flags {}, color {:08x})", flags, color);
-    }
-    count(3);
-    inner_->clear(flags, color);
-}
-
-void PassthroughEngine::setProjection(imvu::Vec2i viewport, const imvu::Mat4f& projection)
-{
-    if (sampled(4)) {
-        const auto& m = projection.m;
-        log::info("setProjection({}x{}, [{:.3f} {:.3f} {:.3f} {:.3f}; {:.3f} {:.3f} {:.3f} {:.3f}; "
-                  "{:.3f} {:.3f} {:.3f} {:.3f}; {:.3f} {:.3f} {:.3f} {:.3f}])",
-                  viewport.x, viewport.y, m[0][0], m[0][1], m[0][2], m[0][3], m[1][0], m[1][1], m[1][2],
-                  m[1][3], m[2][0], m[2][1], m[2][2], m[2][3], m[3][0], m[3][1], m[3][2], m[3][3]);
-    }
-    count(4);
-    inner_->setProjection(viewport, projection);
-}
-
-void PassthroughEngine::setModelSpace(const imvu::Transform& model)
-{
-    if (sampled(5)) {
-        log::info("setModelSpace({})", text(model));
-    }
-    count(5);
-    inner_->setModelSpace(model);
-}
-
-void PassthroughEngine::setViewSpace(const imvu::Transform& view)
-{
-    if (sampled(6)) {
-        log::info("setViewSpace({})", text(view));
-    }
-    count(6);
-    inner_->setViewSpace(view);
-}
-
-void PassthroughEngine::setTextures(imvu::PiRenderTexture* texture0, imvu::PiRenderTexture* texture1)
-{
-    count(7);
-    inner_->setTextures(texture0, texture1);
-}
-
-void PassthroughEngine::setDepthTestingOn(bool on)
-{
-    count(8);
-    inner_->setDepthTestingOn(on);
-}
-
-void PassthroughEngine::setDepthWritingOn(bool on)
-{
-    count(9);
-    inner_->setDepthWritingOn(on);
-}
-
-void PassthroughEngine::setBlendMode(imvu::RenderBlendMode mode)
-{
-    if (sampled(10)) {
-        log::info("setBlendMode({})", static_cast<int>(mode));
-    }
-    count(10);
-    inner_->setBlendMode(mode);
-}
-
-void PassthroughEngine::setAlphaTestingOn(bool on)
-{
-    count(11);
-    inner_->setAlphaTestingOn(on);
-}
-
-void PassthroughEngine::setAlphaTestValue(float value)
-{
-    if (sampled(12)) {
-        log::info("setAlphaTestValue({:.3f})", value);
-    }
-    count(12);
-    inner_->setAlphaTestValue(value);
-}
-
-void PassthroughEngine::setBlendingOn(bool on)
-{
-    count(13);
-    inner_->setBlendingOn(on);
-}
-
-void PassthroughEngine::setBackFaceCullingOn(bool on)
-{
-    count(14);
-    inner_->setBackFaceCullingOn(on);
-}
-
-void PassthroughEngine::setNormalInvertingOn(bool on)
-{
-    count(15);
-    inner_->setNormalInvertingOn(on);
-}
-
-void PassthroughEngine::setLightingOn(bool on)
-{
-    count(16);
-    inner_->setLightingOn(on);
-}
-
-void PassthroughEngine::setLightProperties(std::uint32_t index, const imvu::LightProperties& light)
-{
-    if (sampled(17)) {
-        log::info("setLightProperties({}, enabled {}, diffuse {:08x}, vector {}, type {})", index,
-                  light.enabled, light.diffuse, text(light.vector), static_cast<int>(light.type));
-    }
-    count(17);
-    inner_->setLightProperties(index, light);
-}
-
-void PassthroughEngine::setFogOn(bool on)
-{
-    count(18);
-    inner_->setFogOn(on);
-}
-
-void PassthroughEngine::setFogAttributes(float start, float end, imvu::Color color)
-{
-    if (sampled(19)) {
-        log::info("setFogAttributes(start {:.3f}, end {:.3f}, color {:08x})", start, end, color);
-    }
-    count(19);
-    inner_->setFogAttributes(start, end, color);
-}
-
-void PassthroughEngine::setTextureTransformMatrixIdentity(std::uint32_t stage)
-{
-    count(20);
-    inner_->setTextureTransformMatrixIdentity(stage);
-}
-
-void PassthroughEngine::setTextureTransformMatrix(std::uint32_t stage, const imvu::TextureTransform& transform)
-{
-    if (sampled(21)) {
-        const auto& m = transform.m;
-        log::info("setTextureTransformMatrix({}, [{:.3f} {:.3f} {:.3f} {:.3f} {:.3f} {:.3f}])", stage, m[0],
-                  m[1], m[2], m[3], m[4], m[5]);
-    }
-    count(21);
-    inner_->setTextureTransformMatrix(stage, transform);
-}
-
-void PassthroughEngine::setEnvironmentAmbient(imvu::Color color)
-{
-    if (sampled(22)) {
-        log::info("setEnvironmentAmbient({:08x})", color);
-    }
-    count(22);
-    inner_->setEnvironmentAmbient(color);
-}
-
-imvu::PiRenderTexture* PassthroughEngine::newRenderTexture(const imvu::Pixmap& pixmap, std::uint32_t flags)
-{
-    if (sampled(23)) {
-        const auto size = pixmap_size(pixmap);
-        log::info("newRenderTexture({}x{} pixmap, flags {:x})", size.x, size.y, flags);
-    }
-    count(23);
-    return inner_->newRenderTexture(pixmap, flags);
-}
-
-imvu::PiInternalIndexBuffer* PassthroughEngine::newIndexBuffer(std::uint32_t indexCount, const std::uint16_t* indices)
-{
-    if (sampled(24)) {
-        log::info("newIndexBuffer({} indices)", indexCount);
-    }
-    count(24);
-    return inner_->newIndexBuffer(indexCount, indices);
-}
-
-imvu::PiInternalVertexBuffer* PassthroughEngine::newStaticVertexBuffer(const void* data, std::uint32_t stride,
-                                                                       std::uint32_t elementSize, std::uint32_t elementCount)
-{
-    if (sampled(25)) {
-        log::info("newStaticVertexBuffer(stride {}, element size {}, {} elements)", stride, elementSize,
-                  elementCount);
-    }
-    count(25);
-    return inner_->newStaticVertexBuffer(data, stride, elementSize, elementCount);
-}
-
-imvu::PiInternalVertexBuffer* PassthroughEngine::newDynamicVertexBuffer(std::uint32_t elementSize, std::uint32_t elementCount)
-{
-    if (sampled(26)) {
-        log::info("newDynamicVertexBuffer(element size {}, {} elements)", elementSize, elementCount);
-    }
-    count(26);
-    return inner_->newDynamicVertexBuffer(elementSize, elementCount);
-}
-
-imvu::SharedPtr<imvu::PiPixmapRenderTarget> PassthroughEngine::newPixmapRenderTarget(imvu::Vec2i size)
-{
-    log::info("newPixmapRenderTarget({}, {})", size.x, size.y);
-    count(27);
-    return inner_->newPixmapRenderTarget(size);
-}
-
-imvu::SharedPtr<imvu::PiTextureRenderTarget> PassthroughEngine::newTextureRenderTarget(imvu::Vec2i size)
-{
-    log::info("newTextureRenderTarget({}, {})", size.x, size.y);
-    count(28);
-    return inner_->newTextureRenderTarget(size);
-}
-
-void PassthroughEngine::beginRendering(imvu::PiRenderTarget* target)
-{
-    count(29);
-    inner_->beginRendering(target);
-}
-
-void PassthroughEngine::endRendering()
-{
-    count(30);
-    inner_->endRendering();
-    if (renderer_) {
-        renderer_->frameDrawn();
-    }
-    if (logged(frames_)) {
-        std::string line = std::format("frame {}:", frames_);
-        for (std::size_t slot = 1; slot < slots; ++slot) {
-            if (calls_[slot] != 0) {
-                line += std::format(" {} x{},", names[slot], calls_[slot]);
-            }
-        }
-        line.pop_back();
-        log::write(line);
-    }
-    calls_ = {};
-    ++frames_;
-}
-
-void PassthroughEngine::drawTriangles(imvu::PiInternalIndexBuffer* indices, imvu::PiInternalVertexBuffer* positions,
-                                      imvu::PiInternalVertexBuffer* normals, imvu::PiInternalVertexBuffer* texCoords,
-                                      imvu::PiInternalVertexBuffer* colors)
-{
-    count(31);
-    inner_->drawTriangles(indices, positions, normals, texCoords, colors);
-}
-
-void PassthroughEngine::drawLine(imvu::Vec3f from, imvu::Vec3f to, std::uint32_t color)
-{
-    if (sampled(32)) {
-        log::info("drawLine({}, {}, {:08x})", text(from), text(to), color);
-    }
-    count(32);
-    inner_->drawLine(from, to, color);
-}
-
-void PassthroughEngine::drawPointSprites(const imvu::PointSpriteVector& sprites)
-{
-    count(33);
-    inner_->drawPointSprites(sprites);
-}
-
-void PassthroughEngine::drawTrianglesFromMemory(std::uint32_t vertexCount, std::uint32_t indexCount,
-                                                const std::uint16_t* indices,
-                                                const imvu::VertexAttributeArray<imvu::Vec3f>& positions,
-                                                const imvu::VertexAttributeArray<imvu::Vec3f>& normals,
-                                                const imvu::VertexAttributeArray<imvu::Vec2f>& texCoords,
-                                                const imvu::VertexAttributeArray<imvu::Color>& colors)
-{
-    if (sampled(34)) {
-        log::info("drawTrianglesFromMemory({} vertices, {} indices; positions {} stride {}, colors {} stride {})",
-                  vertexCount, indexCount, positions.count, positions.stride, colors.count, colors.stride);
-    }
-    count(34);
-    inner_->drawTrianglesFromMemory(vertexCount, indexCount, indices, positions, normals, texCoords, colors);
-}
-
-} // namespace t5de

adapter/pixmap.cpp

@@ -0,0 +1,45 @@
+#include "pixmap.hpp"
+
+#include <windows.h>
+
+namespace t5de {
+
+namespace {
+
+using Getter = unsigned(__thiscall*)(const imvu::Pixmap*);
+using PixelsGetter = std::uint32_t*(__thiscall*)(const imvu::Pixmap*);
+
+struct Methods {
+    Getter width = nullptr;
+    Getter height = nullptr;
+    Getter rowBytes = nullptr;
+    PixelsGetter pixels = nullptr;
+};
+
+const Methods& methods()
+{
+    static const Methods methods = [] {
+        Methods found;
+        if (HMODULE module = GetModuleHandleW(L"pixmap.dll")) {
+            found.width = reinterpret_cast<Getter>(GetProcAddress(module, "?width@Pixmap@@QBEIXZ"));
+            found.height = reinterpret_cast<Getter>(GetProcAddress(module, "?height@Pixmap@@QBEIXZ"));
+            found.rowBytes = reinterpret_cast<Getter>(GetProcAddress(module, "?rowBytes@Pixmap@@QBEIXZ"));
+            found.pixels = reinterpret_cast<PixelsGetter>(GetProcAddress(module, "?pixels@Pixmap@@QBEPAIXZ"));
+        }
+        return found;
+    }();
+    return methods;
+}
+
+} // namespace
+
+std::optional<PixmapPixels> pixmap_pixels(const imvu::Pixmap& pixmap)
+{
+    const Methods& m = methods();
+    if (m.width == nullptr || m.height == nullptr || m.rowBytes == nullptr || m.pixels == nullptr) {
+        return std::nullopt;
+    }
+    return PixmapPixels{m.width(&pixmap), m.height(&pixmap), m.rowBytes(&pixmap), m.pixels(&pixmap)};
+}
+
+} // namespace t5de

adapter/pixmap.hpp

@@ -0,0 +1,21 @@
+#pragma once
+// Reading IMVU's images (Pixmap) through pixmap.dll's exported methods, rather than depending on their layout.
+
+#include <cstdint>
+#include <optional>
+
+#include "t5de/imvu/types.hpp"
+
+namespace t5de {
+
+struct PixmapPixels {
+    std::uint32_t width;
+    std::uint32_t height;
+    std::uint32_t rowBytes;
+    const std::uint32_t* pixels; // 0xAARRGGBB, rows rowBytes apart
+};
+
+// Null if pixmap.dll's methods can't be found.
+std::optional<PixmapPixels> pixmap_pixels(const imvu::Pixmap& pixmap);
+
+} // namespace t5de

adapter/plugin.cpp

@@ -1,12 +1,15 @@
 // The renderer plugin's entry points, which SceneWindow.dll's createDLLRenderer looks up by name (exported
-// undecorated by plugin.def). For now the renderer is the pass-through: IMVU's Direct3D engine, wrapped.
+// undecorated by plugin.def). The engine is the adapter's (engine.hpp): IMVU's Direct3D engine draws, and the
+// renderer process draws over part of it.
 
 #include <windows.h>
 
 #include <exception>
+#include <string_view>
 
 #include "log.hpp"
-#include "passthrough_engine.hpp"
+#include "engine.hpp"
+#include "recorder.hpp"
 #include "renderer_process.hpp"
 #include "t5de/imvu/render_engine.hpp"
 
@@ -30,6 +33,19 @@ imvu::SharedPtr<imvu::PiRenderEngine> create_imvu_engine(imvu::SharedPtr<imvu::I
     return create(std::move(window), imvu::RendererType::Direct3D);
 }
 
+// What the renderer process draws: T5DE_RENDERER_VIEW=off, split (the right half), full (the default) or compare.
+// (The renderer process can also be switched while it runs: renderer/main.cpp.)
+protocol::View view()
+{
+    wchar_t value[16] = {};
+    GetEnvironmentVariableW(L"T5DE_RENDERER_VIEW", value, 16);
+    const std::wstring_view setting = value;
+    return setting == L"off"       ? protocol::View::Off
+           : setting == L"split"   ? protocol::View::Split
+           : setting == L"compare" ? protocol::View::Compare
+                                   : protocol::View::Full;
+}
+
 } // namespace
 
 extern "C" int __cdecl isRenderEngineSupported() { return 1; }
@@ -45,9 +61,10 @@ extern "C" imvu::SharedPtr<imvu::PiRenderEngine> __cdecl createRenderEngine(imvu
         if (!inner) {
             return {};
         }
-        auto renderer = hwnd != nullptr ? RendererProcess::start(hwnd) : nullptr;
+        const protocol::View shown = view();
-        return imvu::SharedPtr<imvu::PiRenderEngine>(
+        auto renderer = hwnd != nullptr && shown != protocol::View::Off ? RendererProcess::start(hwnd, shown) : nullptr;
-            std::make_unique<PassthroughEngine>(std::move(inner), std::move(renderer)));
+        return imvu::SharedPtr<imvu::PiRenderEngine>(std::make_unique<Engine>(
+            std::move(inner), std::make_shared<Recorder>(std::move(renderer))));
     } catch (const std::exception& e) {
         // Exceptions mustn't cross into IMVU's code (a different C++ runtime).
         log::info("createRenderEngine failed: {}", e.what());

adapter/recorder.cpp

@@ -0,0 +1,111 @@
+#include "recorder.hpp"
+
+#include <cstring>
+
+#include "log.hpp"
+#include "pixmap.hpp"
+
+namespace t5de {
+
+namespace {
+
+// Between frames, recorded commands (mostly textures and buffers being loaded) are sent once they reach this size.
+constexpr std::size_t flushSize = 4 * 1024 * 1024;
+
+} // namespace
+
+void pack(std::byte* destination, const void* source, std::uint32_t stride, std::uint32_t elementSize,
+          std::uint32_t count)
+{
+    const auto* from = static_cast<const std::byte*>(source);
+    if (stride == 0 || stride == elementSize) {
+        std::memcpy(destination, from, std::size_t{elementSize} * count);
+        return;
+    }
+    for (std::uint32_t i = 0; i < count; ++i) {
+        std::memcpy(destination + std::size_t{i} * elementSize, from + std::size_t{i} * stride, elementSize);
+    }
+}
+
+Recorder::Recorder(std::unique_ptr<RendererProcess> process) : process_(std::move(process)) {}
+
+void Recorder::flush(bool endsFrame)
+{
+    if (!active()) {
+        writer_.clear();
+        return;
+    }
+    process_->send(writer_.bytes(), endsFrame);
+    writer_.clear();
+}
+
+void Recorder::flushIfLarge()
+{
+    if (writer_.size() >= flushSize) {
+        flush(false);
+    }
+}
+
+void Recorder::addPixels(protocol::Op op, std::uint32_t id, const imvu::Pixmap& pixmap, std::uint32_t flags)
+{
+    const auto pixels = pixmap_pixels(pixmap);
+    if (!pixels || pixels->pixels == nullptr) {
+        return;
+    }
+    const std::size_t rowSize = std::size_t{pixels->width} * 4;
+    std::byte* payload = nullptr;
+    if (op == protocol::Op::CreateTexture) {
+        payload = add(op, protocol::CreateTexture{id, pixels->width, pixels->height, flags}, rowSize * pixels->height);
+    } else {
+        payload = add(op, protocol::UpdateTexture{id, pixels->width, pixels->height}, rowSize * pixels->height);
+    }
+    if (payload == nullptr) {
+        return;
+    }
+    const auto* rows = reinterpret_cast<const std::byte*>(pixels->pixels);
+    for (std::uint32_t y = 0; y < pixels->height; ++y) {
+        std::memcpy(payload + y * rowSize, rows + std::size_t{y} * pixels->rowBytes, rowSize);
+    }
+    flushIfLarge();
+}
+
+std::uint32_t Recorder::createTexture(const imvu::Pixmap& pixmap, std::uint32_t flags)
+{
+    const std::uint32_t id = nextId_++;
+    addPixels(protocol::Op::CreateTexture, id, pixmap, flags);
+    return id;
+}
+
+void Recorder::updateTexture(std::uint32_t id, const imvu::Pixmap& pixmap)
+{
+    addPixels(protocol::Op::UpdateTexture, id, pixmap, 0);
+}
+
+void Recorder::destroyTexture(std::uint32_t id) { add(protocol::Op::DestroyTexture, protocol::DestroyTexture{id}); }
+
+std::uint32_t Recorder::createBuffer(protocol::BufferKind kind, const void* data, std::uint32_t stride,
+                                     std::uint32_t elementSize, std::uint32_t count)
+{
+    const std::uint32_t id = nextId_++;
+    const std::size_t size = data != nullptr ? std::size_t{elementSize} * count : 0;
+    if (std::byte* payload = add(protocol::Op::CreateBuffer, protocol::CreateBuffer{id, kind, elementSize, count}, size);
+        payload != nullptr && size != 0) {
+        pack(payload, data, stride, elementSize, count);
+    }
+    flushIfLarge();
+    return id;
+}
+
+void Recorder::updateBuffer(std::uint32_t id, const void* data, std::uint32_t stride, std::uint32_t elementSize,
+                            std::uint32_t count)
+{
+    const std::size_t size = std::size_t{elementSize} * count;
+    if (std::byte* payload = add(protocol::Op::UpdateBuffer, protocol::UpdateBuffer{id, elementSize, count}, size);
+        payload != nullptr && data != nullptr) {
+        pack(payload, data, stride, elementSize, count);
+    }
+}
+
+void Recorder::destroyBuffer(std::uint32_t id) { add(protocol::Op::DestroyBuffer, protocol::DestroyBuffer{id}); }
+
+} // namespace t5de

adapter/recorder.hpp

@@ -0,0 +1,62 @@
+#pragma once
+// Records what the client draws as commands (include/t5de/protocol/commands.hpp) and sends them to the renderer
+// process. Shared by an engine and the resources it creates (which can outlive it); once the renderer process is
+// gone, it records nothing.
+
+#include <cstdint>
+#include <memory>
+
+#include "renderer_process.hpp"
+#include "t5de/imvu/types.hpp"
+#include "t5de/protocol/commands.hpp"
+
+namespace t5de {
+
+class Recorder {
+public:
+    explicit Recorder(std::unique_ptr<RendererProcess> process);
+
+    [[nodiscard]] bool active() const { return process_ && process_->alive(); }
+
+    // Whether the renderer process draws the whole 3D view, so the client's renderer needn't.
+    [[nodiscard]] bool covering() const { return process_ && process_->covering(); }
+
+    // Records a command (if active); returns its payload to fill, or null.
+    template <class Body>
+    std::byte* add(protocol::Op op, const Body& body, std::size_t payloadSize = 0)
+    {
+        if (!active()) {
+            return nullptr;
+        }
+        std::byte* payload = writer_.add(op, body, payloadSize);
+        return payload;
+    }
+
+    // Sends what's recorded: at the end of each frame, and between frames when a lot is waiting (resources).
+    void flush(bool endsFrame);
+    void flushIfLarge();
+
+    std::uint32_t createTexture(const imvu::Pixmap& pixmap, std::uint32_t flags);
+    void updateTexture(std::uint32_t id, const imvu::Pixmap& pixmap);
+    void destroyTexture(std::uint32_t id);
+
+    // Copies count elements of elementSize bytes, stride bytes apart (0: packed), from data (null: none yet).
+    std::uint32_t createBuffer(protocol::BufferKind kind, const void* data, std::uint32_t stride,
+                               std::uint32_t elementSize, std::uint32_t count);
+    void updateBuffer(std::uint32_t id, const void* data, std::uint32_t stride, std::uint32_t elementSize,
+                      std::uint32_t count);
+    void destroyBuffer(std::uint32_t id);
+
+private:
+    void addPixels(protocol::Op op, std::uint32_t id, const imvu::Pixmap& pixmap, std::uint32_t flags);
+
+    std::unique_ptr<RendererProcess> process_;
+    protocol::Writer writer_;
+    std::uint32_t nextId_ = 1;
+};
+
+// Copies count elements of elementSize bytes, stride bytes apart (0: packed), into destination, packed.
+void pack(std::byte* destination, const void* source, std::uint32_t stride, std::uint32_t elementSize,
+          std::uint32_t count);
+
+} // namespace t5de

adapter/renderer_process.cpp

@@ -1,6 +1,8 @@
 #include "renderer_process.hpp"
 
+#include <algorithm>
 #include <atomic>
+#include <cstring>
 #include <filesystem>
 #include <format>
 #include <string>
@@ -11,6 +13,9 @@ namespace t5de {
 
 namespace {
 
+// How long to wait for the renderer process to take a chunk before giving up on it.
+constexpr DWORD consumeTimeout = 5000;
+
 // The folder this DLL is in: the client's program folder.
 std::filesystem::path module_folder()
 {
@@ -22,36 +27,73 @@ std::filesystem::path module_folder()
     return std::filesystem::path(path).parent_path();
 }
 
+std::string narrow(const std::wstring& text) { return std::filesystem::path(text).string(); }
+
+// The renderer process runs from a copy in %TEMP%\t5de-renderer, one per version of the program (by size and
+// modification time), so the program in the client's folder can be replaced (updated) while the client runs: the
+// next engine runs the new one. Copies no longer running are removed. Falls back to the program itself.
+std::filesystem::path running_copy(const std::filesystem::path& exe)
+{
+    std::error_code error;
+    const auto size = std::filesystem::file_size(exe, error);
+    const auto modified = std::filesystem::last_write_time(exe, error);
+    if (error) {
+        return exe;
+    }
+    const auto folder = std::filesystem::temp_directory_path(error) / L"t5de-renderer";
+    std::filesystem::create_directories(folder, error);
+    const auto copy = folder / std::format(L"t5de-renderer-{}-{}.exe", size, modified.time_since_epoch().count());
+    if (!std::filesystem::exists(copy, error)) {
+        for (const auto& old : std::filesystem::directory_iterator(folder, error)) {
+            std::filesystem::remove(old.path(), error); // fails, harmlessly, for copies still running
+        }
+        if (!CopyFileW(exe.c_str(), copy.c_str(), FALSE)) {
+            return exe;
+        }
+    }
+    return copy;
+}
+
 } // namespace
 
-std::unique_ptr<RendererProcess> RendererProcess::start(HWND window)
+std::unique_ptr<RendererProcess> RendererProcess::start(HWND window, protocol::View view)
 {
     static std::atomic<int> engines = 0;
     const std::wstring name = std::format(L"Local\\t5de-renderer-{}-{}", GetCurrentProcessId(), engines++);
+    constexpr DWORD size = sizeof(protocol::Channel) + protocol::channelCapacity;
 
     std::unique_ptr<RendererProcess> renderer(new RendererProcess());
-    renderer->mapping_ = CreateFileMappingW(INVALID_HANDLE_VALUE, nullptr, PAGE_READWRITE, 0,
+    renderer->mapping_ = CreateFileMappingW(INVALID_HANDLE_VALUE, nullptr, PAGE_READWRITE, 0, size, name.c_str());
-                                            sizeof(protocol::Channel), name.c_str());
     if (renderer->mapping_ == nullptr) {
         log::info("renderer process: can't create the channel: error {}", GetLastError());
         return nullptr;
     }
-    renderer->channel_ = static_cast<protocol::Channel*>(
+    renderer->channel_ =
-        MapViewOfFile(renderer->mapping_, FILE_MAP_ALL_ACCESS, 0, 0, sizeof(protocol::Channel)));
+        static_cast<protocol::Channel*>(MapViewOfFile(renderer->mapping_, FILE_MAP_ALL_ACCESS, 0, 0, size));
-    renderer->frameEvent_ = CreateEventW(nullptr, FALSE, FALSE, (name + protocol::frameEventSuffix).c_str());
+    renderer->ready_ = CreateEventW(nullptr, FALSE, FALSE, (name + protocol::readyEventSuffix).c_str());
-    if (renderer->channel_ == nullptr || renderer->frameEvent_ == nullptr) {
+    renderer->consumed_ = CreateEventW(nullptr, FALSE, TRUE, (name + protocol::consumedEventSuffix).c_str());
+    if (renderer->channel_ == nullptr || renderer->ready_ == nullptr || renderer->consumed_ == nullptr) {
         log::info("renderer process: can't set up the channel: error {}", GetLastError());
         return nullptr;
     }
+    renderer->data_ = reinterpret_cast<std::byte*>(renderer->channel_ + 1);
     *renderer->channel_ = {.version = protocol::version,
                            .closing = 0,
                            .window = reinterpret_cast<std::uintptr_t>(window),
-                           .frames = 0};
+                           .frames = 0,
+                           .view = view,
+                           .capacity = protocol::channelCapacity,
+                           .chunkSize = 0,
+                           .chunkFlags = 0,
+                           .covering = 0,
+                           .reserved = 0};
 
-    // The renderer's window goes on top of the client's 3D view: keep the client's drawing out of it.
+    // The renderer's window goes on top of the client's 3D view, as a child window. The client's window mustn't clip
-    SetWindowLongW(window, GWL_STYLE, GetWindowLongW(window, GWL_STYLE) | WS_CLIPCHILDREN);
+    // its children (WS_CLIPCHILDREN, which the client creates it with): the client draws a frame when its window gets
-
+    // WM_PAINT, which a window entirely covered by clipped children never gets (it then drew about once a second). The
-    const auto exe = module_folder() / L"t5de-renderer.exe";
+    // renderer's swap chain is a flip model one, which the compositor shows above the client's drawing anyway.
+    SetWindowLongW(window, GWL_STYLE, GetWindowLongW(window, GWL_STYLE) & ~WS_CLIPCHILDREN);
+    const auto exe = running_copy(module_folder() / L"t5de-renderer.exe");
     std::wstring command = std::format(L"\"{}\" --channel {} --parent {}", exe.wstring(), name, GetCurrentProcessId());
     STARTUPINFOW startup{.cb = sizeof(STARTUPINFOW)};
     PROCESS_INFORMATION process{};
@@ -62,8 +104,7 @@ std::unique_ptr<RendererProcess> RendererProcess::start(HWND window)
     }
     CloseHandle(process.hThread);
     renderer->process_ = process.hProcess;
-    log::info("renderer process: started {} (pid {}) on channel {}", exe.string(), process.dwProcessId,
+    log::info("renderer process: started {} (pid {}) on channel {}", exe.string(), process.dwProcessId, narrow(name));
-              std::filesystem::path(name).string());
     return renderer;
 }
 
@@ -71,22 +112,42 @@ RendererProcess::~RendererProcess()
 {
     if (channel_ != nullptr) {
         channel_->closing = 1;
-        if (frameEvent_ != nullptr) {
+        if (ready_ != nullptr) {
-            SetEvent(frameEvent_);
+            SetEvent(ready_);
         }
         UnmapViewOfFile(channel_);
     }
-    for (HANDLE handle : {frameEvent_, mapping_, process_}) {
+    for (HANDLE handle : {ready_, consumed_, mapping_, process_}) {
         if (handle != nullptr) {
             CloseHandle(handle);
         }
     }
 }
 
-void RendererProcess::frameDrawn()
+bool RendererProcess::send(std::span<const std::byte> commands, bool endsFrame)
 {
-    ++channel_->frames;
+    if (!alive_) {
-    SetEvent(frameEvent_);
+        return false;
+    }
+    do {
+        const HANDLE waits[] = {consumed_, process_};
+        const DWORD result = WaitForMultipleObjects(2, waits, FALSE, consumeTimeout);
+        if (result != WAIT_OBJECT_0) {
+            log::info("renderer process: {}; giving up on it", result == WAIT_OBJECT_0 + 1 ? "exited" : "not responding");
+            alive_ = false;
+            return false;
+        }
+        const std::size_t size = std::min<std::size_t>(commands.size(), channel_->capacity);
+        std::memcpy(data_, commands.data(), size);
+        commands = commands.subspan(size);
+        channel_->chunkSize = static_cast<std::uint32_t>(size);
+        channel_->chunkFlags = endsFrame && commands.empty() ? protocol::chunkEndsFrame : 0;
+        if (endsFrame && commands.empty()) {
+            ++channel_->frames;
+        }
+        SetEvent(ready_);
+    } while (!commands.empty());
+    return true;
 }
 
 } // namespace t5de

adapter/renderer_process.hpp

@@ -1,11 +1,12 @@
 #pragma once
-// The renderer process (t5de-renderer.exe, 64-bit) an engine draws through: started with the engine, told about
+// The renderer process (t5de-renderer.exe, 64-bit) an engine draws through: started with the engine, sent the
-// each frame through a shared memory channel (include/t5de/protocol/channel.hpp), and told to exit with it.
+// command stream through a shared memory channel (include/t5de/protocol/channel.hpp), and told to exit with it.
 
 #include <windows.h>
 
-#include <cstdint>
+#include <cstddef>
 #include <memory>
+#include <span>
 
 #include "t5de/protocol/channel.hpp"
 
@@ -13,23 +14,32 @@ namespace t5de {
 
 class RendererProcess {
 public:
-    // Starts t5de-renderer.exe from this DLL's folder, drawing into window; null if it can't.
+    // Starts t5de-renderer.exe from this DLL's folder, drawing into part of window; null if it can't.
-    static std::unique_ptr<RendererProcess> start(HWND window);
+    static std::unique_ptr<RendererProcess> start(HWND window, protocol::View view);
 
     RendererProcess(const RendererProcess&) = delete;
     RendererProcess& operator=(const RendererProcess&) = delete;
     ~RendererProcess();
 
-    // The client drew a frame.
+    // Sends commands, in as many chunks as they take; endsFrame marks the end of a frame. False if the renderer
-    void frameDrawn();
+    // process is gone or stopped taking them (it's then given up on).
+    bool send(std::span<const std::byte> commands, bool endsFrame);
+
+    [[nodiscard]] bool alive() const { return alive_; }
+
+    // Whether the renderer process covers the whole 3D view (protocol::Channel::covering).
+    [[nodiscard]] bool covering() const { return alive_ && channel_->covering != 0; }
 
 private:
     RendererProcess() = default;
 
     HANDLE mapping_ = nullptr;
-    HANDLE frameEvent_ = nullptr;
+    HANDLE ready_ = nullptr;
+    HANDLE consumed_ = nullptr;
     HANDLE process_ = nullptr;
     protocol::Channel* channel_ = nullptr;
+    std::byte* data_ = nullptr;
+    bool alive_ = true;
 };
 
 } // namespace t5de

adapter/resources.cpp

@@ -0,0 +1,108 @@
+#include "resources.hpp"
+
+#include <mutex>
+#include <unordered_set>
+
+namespace t5de {
+
+namespace {
+
+// Our live resources, by address.
+struct Registry {
+    std::mutex mutex;
+    std::unordered_set<const void*> objects;
+};
+
+Registry& registry()
+{
+    static Registry instance;
+    return instance;
+}
+
+void add(const void* object)
+{
+    const std::scoped_lock lock(registry().mutex);
+    registry().objects.insert(object);
+}
+
+void remove(const void* object)
+{
+    const std::scoped_lock lock(registry().mutex);
+    registry().objects.erase(object);
+}
+
+} // namespace
+
+bool ours(const void* object)
+{
+    if (object == nullptr) {
+        return false;
+    }
+    const std::scoped_lock lock(registry().mutex);
+    return registry().objects.contains(object);
+}
+
+Texture::Texture(imvu::PiRenderTexture* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id)
+    : inner_(inner), recorder_(std::move(recorder)), id_(id)
+{
+    add(static_cast<imvu::PiRenderTexture*>(this));
+}
+
+Texture::~Texture()
+{
+    remove(static_cast<imvu::PiRenderTexture*>(this));
+    delete inner_; // through its virtual destructor: IMVU's code frees it
+    recorder_->destroyTexture(id_);
+}
+
+imvu::Vec2i Texture::size() const { return inner_->size(); }
+
+void Texture::updateRegions(const imvu::Pixmap& pixmap, const imvu::CompoundRegion& regions)
+{
+    inner_->updateRegions(pixmap, regions);
+    recorder_->updateTexture(id_, pixmap); // all of it: the regions' layout isn't known
+}
+
+void Texture::update(const imvu::Pixmap& pixmap)
+{
+    inner_->update(pixmap);
+    recorder_->updateTexture(id_, pixmap);
+}
+
+imvu::Pixmap* Texture::read(imvu::Pixmap* result) const { return inner_->read(result); }
+
+IndexBuffer::IndexBuffer(imvu::PiInternalIndexBuffer* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id)
+    : inner_(inner), recorder_(std::move(recorder)), id_(id)
+{
+    add(static_cast<imvu::PiInternalIndexBuffer*>(this));
+}
+
+IndexBuffer::~IndexBuffer()
+{
+    remove(static_cast<imvu::PiInternalIndexBuffer*>(this));
+    delete inner_;
+    recorder_->destroyBuffer(id_);
+}
+
+VertexBuffer::VertexBuffer(imvu::PiInternalVertexBuffer* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id)
+    : inner_(inner), recorder_(std::move(recorder)), id_(id)
+{
+    add(static_cast<imvu::PiInternalVertexBuffer*>(this));
+}
+
+VertexBuffer::~VertexBuffer()
+{
+    remove(static_cast<imvu::PiInternalVertexBuffer*>(this));
+    delete inner_;
+    recorder_->destroyBuffer(id_);
+}
+
+void VertexBuffer::update(const void* data, std::uint32_t stride, std::uint32_t elementSize, std::uint32_t elementCount)
+{
+    inner_->update(data, stride, elementSize, elementCount);
+    recorder_->updateBuffer(id_, data, stride, elementSize, elementCount);
+}
+
+void VertexBuffer::unknown2() { inner_->unknown2(); }
+
+} // namespace t5de

adapter/resources.hpp

@@ -0,0 +1,95 @@
+#pragma once
+// The textures and buffers the adapter's engine hands to the client: each wraps the client renderer's own (which the
+// client renderer still draws with) and mirrors it, by id, to the renderer process. The client deletes them through
+// their virtual destructors.
+
+#include <cstdint>
+#include <memory>
+
+#include "recorder.hpp"
+#include "t5de/imvu/resources.hpp"
+
+namespace t5de {
+
+class Texture final : public imvu::PiRenderTexture {
+public:
+    Texture(imvu::PiRenderTexture* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id);
+    ~Texture() override;
+
+    imvu::Vec2i size() const override;
+    void updateRegions(const imvu::Pixmap& pixmap, const imvu::CompoundRegion& regions) override;
+    void update(const imvu::Pixmap& pixmap) override;
+    imvu::Pixmap* read(imvu::Pixmap* result) const override;
+
+    [[nodiscard]] imvu::PiRenderTexture* inner() const { return inner_; }
+    [[nodiscard]] std::uint32_t id() const { return id_; }
+
+private:
+    imvu::PiRenderTexture* inner_;
+    std::shared_ptr<Recorder> recorder_;
+    std::uint32_t id_;
+};
+
+class IndexBuffer final : public imvu::PiInternalIndexBuffer {
+public:
+    IndexBuffer(imvu::PiInternalIndexBuffer* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id);
+    ~IndexBuffer() override;
+
+    [[nodiscard]] imvu::PiInternalIndexBuffer* inner() const { return inner_; }
+    [[nodiscard]] std::uint32_t id() const { return id_; }
+
+private:
+    imvu::PiInternalIndexBuffer* inner_;
+    std::shared_ptr<Recorder> recorder_;
+    std::uint32_t id_;
+};
+
+class VertexBuffer final : public imvu::PiInternalVertexBuffer {
+public:
+    VertexBuffer(imvu::PiInternalVertexBuffer* inner, std::shared_ptr<Recorder> recorder, std::uint32_t id);
+    ~VertexBuffer() override;
+
+    void update(const void* data, std::uint32_t stride, std::uint32_t elementSize,
+                std::uint32_t elementCount) override;
+    void unknown2() override;
+
+    [[nodiscard]] imvu::PiInternalVertexBuffer* inner() const { return inner_; }
+    [[nodiscard]] std::uint32_t id() const { return id_; }
+
+private:
+    imvu::PiInternalVertexBuffer* inner_;
+    std::shared_ptr<Recorder> recorder_;
+    std::uint32_t id_;
+};
+
+// Whether an object is one of ours (live Texture, IndexBuffer or VertexBuffer): the client can also hand the engine
+// objects the client renderer made itself (render targets' textures), which must reach it unchanged.
+bool ours(const void* object);
+
+// The client renderer's object, and our id, for what the client passes in: ours unwrapped, others unchanged (id 0).
+inline imvu::PiRenderTexture* inner(imvu::PiRenderTexture* texture)
+{
+    return ours(texture) ? static_cast<Texture*>(texture)->inner() : texture;
+}
+inline std::uint32_t id(imvu::PiRenderTexture* texture)
+{
+    return ours(texture) ? static_cast<Texture*>(texture)->id() : 0;
+}
+inline imvu::PiInternalIndexBuffer* inner(imvu::PiInternalIndexBuffer* buffer)
+{
+    return ours(buffer) ? static_cast<IndexBuffer*>(buffer)->inner() : buffer;
+}
+inline std::uint32_t id(imvu::PiInternalIndexBuffer* buffer)
+{
+    return ours(buffer) ? static_cast<IndexBuffer*>(buffer)->id() : 0;
+}
+inline imvu::PiInternalVertexBuffer* inner(imvu::PiInternalVertexBuffer* buffer)
+{
+    return ours(buffer) ? static_cast<VertexBuffer*>(buffer)->inner() : buffer;
+}
+inline std::uint32_t id(imvu::PiInternalVertexBuffer* buffer)
+{
+    return ours(buffer) ? static_cast<VertexBuffer*>(buffer)->id() : 0;
+}
+
+} // namespace t5de

docs/architecture.md

@@ -29,21 +29,49 @@ creates (each 3D view, and again on room changes), the adapter:
 
 The renderer process draws into a child window it creates inside the client's window: window handles are valid in
 every process, and Windows allows child windows across processes (Chrome's GPU process works this way). The child
-is disabled so mouse input goes to the client's window under it, and the adapter gives the client's window
+is disabled so mouse input goes to the client's window under it. Its swap chain is a flip model one, which the
-`WS_CLIPCHILDREN` so the client's own drawing stays out of it. The process exits when the engine is destroyed or the
+compositor shows above the client's own drawing. The client's window must not clip its children (`WS_CLIPCHILDREN`,
-client exits.
+which the client creates it with, so the adapter clears it):
+the client draws a frame when its window gets `WM_PAINT` (`SceneWindow.dll` invalidates it, and draws on the paint),
+and a window entirely covered by clipped children never gets one, so the client fell to about one frame a second.
+The process exits when the engine is destroyed or the client exits.
 
 ## Status
 
-A spike: the adapter still has IMVU's Direct3D 9 engine draw the scene (the pass-through), and the renderer process
+The renderer process draws the whole 3D view in place of the client's renderer (the full view, the default). The
-only draws a box in the 3D view's corner, pulsing with the client's frames. Checked in the client (Direct3D 11 in
+adapter still forwards everything to the client's Direct3D renderer (state and resources: it draws render targets and
-software, on a machine without a graphics driver): the box draws over the scene from the 64-bit process, keeps in
+reads textures back, which the renderer process doesn't yet), but while the renderer process covers the whole view
-step with the client's frames, clicks on it reach the scene, and each renderer process exits with its engine (room
+(`Channel::covering`, set after each frame it presents) the client's renderer skips its clears and draws, decided per
-changes) and with the client. The client keeps two engines alive in a room, so two processes (about 15 MB each);
+frame; it draws again as soon as the renderer process stops covering the view (the view changes, or the process
-one process serving all of a client's engines would be leaner.
+fails). Checked in rooms with screenshots of both renderers side by side (the compare view): they match, particles
+and the 2D overlay (name tags) included.
 
-Next: the protocol carrying each frame's calls and the resources (buffers, textures), and the renderer process
+For comparing while the client runs (development):
-drawing the scene itself, with IMVU's Direct3D engine left as the fallback.
+
+- Views: `T5DE_RENDERER_VIEW` (off, split, full (the default) or compare) when the client starts, or
+  `%TEMP%\t5de-renderer-view.txt` holding one of those at any time. Compare draws the left half's scene over the
+  right half, so both renderers show the same thing side by side.
+- `%TEMP%\t5de-renderer-shaders\fixed_function.hlsl`, if present, replaces the built-in shaders and is reloaded
+  when it changes.
+- The renderer process runs from a copy (`%TEMP%\t5de-renderer`), so `t5de-renderer.exe` can be replaced while the
+  client runs; engines started afterwards run the new one (the room's view keeps its engine until the client
+  restarts).
+- The adapter's log (`%TEMP%\t5de-renderer.log`) sums up what the client draws every 300 frames (the first few
+  times), and logs textures and render targets the client renderer made itself.
+
+- The command stream (`include/t5de/protocol/commands.hpp`) carries state, transforms, lights, fog, textures,
+  buffers (copied when created or updated) and draws, in chunks through the channel; the adapter waits for each to
+  be taken, and gives up on a renderer process that stops taking them (the client's renderer carries on).
+- The renderer process reproduces IMVU's Direct3D 9 setup (`renderer/shaders/fixed_function.hlsl`), particles
+  included (point sprites, drawn as instanced squares), with a pipeline per combination of states.
+- Speed: on a PC with a graphics card, drawing the whole view (3440x1059, a room of some 220 draws) takes 1.3 to
+  3.8 ms a frame, and the time from the client finishing a frame to ours being presented is the same (frames don't
+  queue, none skipped). Earlier, in the split view, the halves couldn't be told apart, even while the camera moved. On the development VM (software rendering, over Remote Desktop) drawing takes 8 to 60 ms, and
+  the right half visibly trails, mostly from compositing another process's window over Remote Desktop. The
+  renderer process draws only the newest frame when it falls behind, and keeps at most one frame queued.
+
+Next: render targets (offscreen frames are skipped, and drawn by the client's renderer), and one renderer process for
+all of a client's views.
 
 ## Building
 

docs/interface.md

@@ -71,7 +71,19 @@ virtual methods, IMVU's code only uses an engine through helpers that call its v
   vertical field of view, Direct3D's 0 to 1 depth) and an orthographic one for 2D, orthonormal view rotations,
   opaque clear colors, ambient light with zero alpha, linear fog (1 to 50), scrolling texture transforms. Furniture
   vertices are interleaved: positions and normals are 12-byte elements 32 bytes apart, so strides matter.
-- Still unknown: point sprites (`drawPointSprites`, particles), render targets, `CompoundRegion` (texture update
+- Projections give depth from -1 to 1, as OpenGL's do: IMVU's Direct3D renderer multiplies them by a matrix
+  mapping it to 0 to 1 (z' = z / 2 + w / 2) and shifting them half a pixel (-1 / width, +1 / height, for Direct3D
+  9's pixel centers). Without the depth mapping, everything in the nearer half of the depth range is clipped: the
+  2D overlay (name tags, the frame rate, the room's title) is drawn at depth -0.98.
+- The 2D overlay is ordinary `drawTriangles` with an orthographic projection (a few per frame), textured with text
+  IMVU draws with GDI into bitmaps.
+- Particles (`drawPointSprites`): a Visual C++ 2010 `std::vector` of 16-byte points (position, color), drawn as
+  Direct3D 9 point sprites: squares viewport height / distance pixels across (IMVU sets the size, 1, and the scale
+  factors 0, 0, 1), textured over the whole square. Direct3D 9's documented 64-pixel default limit doesn't apply in
+  practice: clamped to it they look much smaller and fainter than the client's.
+- Fog: linear table fog, start 1 and end 50 in rooms. Fogging by view depth matches the client; by depth buffer
+  value (as Direct3D 9's documentation suggests for projections like IMVU's) doesn't.
+- Still unknown: render targets, `CompoundRegion` (texture update
   regions), `newRenderTexture`'s flags, and what `beginRendering` does with a null target. MSVC orders overloaded
   virtual methods unlike their declaration, so our declarations never overload virtual methods.
 - Memory crosses the boundary only through virtual methods (destructors, control blocks), so each side can use its

include/t5de/imvu/types.hpp

@@ -90,8 +90,21 @@ static_assert(sizeof(VertexAttributeArray<Vec3f>) == 12);
 // (width, height, pitch, rowBytes, pixels, ...).
 class Pixmap;
 
-// Unknown so far: used only by drawPointSprites (particles), as a Visual C++ 2010 std::vector<PointSprite>.
+// A particle (drawPointSprites): a point drawn as a camera-facing square, textured, about viewport height / distance
-struct PointSpriteVector;
+// pixels across (1 to 64), as the Direct3D engine's point sprites are.
+struct PointSprite {
+    Vec3f position;
+    Color color;
+};
+
+static_assert(sizeof(PointSprite) == 16);
+
+// std::vector<PointSprite> as Visual C++ 2010 lays it out (release builds): its elements are [first, last).
+struct PointSpriteVector {
+    const PointSprite* first;
+    const PointSprite* last;
+    const PointSprite* end; // of the allocation
+};
 
 // Region lists IMVU passes to texture updates; unknown so far.
 class CompoundRegion;

include/t5de/protocol/channel.hpp

@@ -7,19 +7,42 @@
 
 namespace t5de::protocol {
 
-inline constexpr std::uint32_t version = 1;
+inline constexpr std::uint32_t version = 3;
 
-// A shared memory section the adapter creates for each engine. Its name is passed to the renderer process as
+// Which part of the client's 3D view the renderer process draws. Full is the default; the others are for comparing
-// `--channel <name>`; the event "<name>-frame" (auto-reset) is set after each frame the client draws.
+// with the client's renderer, which then draws the whole view under the renderer process's window.
+enum class View : std::uint32_t {
+    Off,     // none
+    Split,   // the right half, to compare with the client's renderer on the left
+    Full,
+    Compare, // over the right half, the left half's scene: both renderers side by side, showing the same
+};
+
+// A shared memory section the adapter creates for each engine, named by the renderer process's `--channel <name>`
+// argument: this header, then `capacity` bytes of data. The adapter sends the command stream (commands.hpp) through
+// it in chunks: it waits for the event "<name>-consumed", writes a chunk and sets "<name>-ready"; the renderer
+// process copies the chunk out and sets "<name>-consumed" (both auto-reset; consumed starts set).
 struct Channel {
     std::uint32_t version;
     std::uint32_t closing; // nonzero: the engine is gone, the renderer process should exit
     std::uint64_t window;  // IMVU's 3D window (an HWND; window handles are valid in every process)
     std::uint64_t frames;  // frames the client has drawn
+    View view;
+    std::uint32_t capacity;
+    std::uint32_t chunkSize;
+    std::uint32_t chunkFlags; // chunkEndsFrame
+    // Set by the renderer process while its window covers the whole 3D view and it's presenting frames: the client's
+    // renderer then needn't draw (the adapter skips its draws; it draws again as soon as this is cleared).
+    std::uint32_t covering;
+    std::uint32_t reserved;
 };
 
-static_assert(sizeof(Channel) == 24, "same layout in 32-bit and 64-bit builds");
+static_assert(sizeof(Channel) == 48, "same layout in 32-bit and 64-bit builds");
 
-inline constexpr wchar_t frameEventSuffix[] = L"-frame";
+inline constexpr std::uint32_t chunkEndsFrame = 1;
+inline constexpr std::uint32_t channelCapacity = 16 * 1024 * 1024;
+
+inline constexpr wchar_t readyEventSuffix[] = L"-ready";
+inline constexpr wchar_t consumedEventSuffix[] = L"-consumed";
 
 } // namespace t5de::protocol

include/t5de/protocol/commands.hpp

@@ -0,0 +1,266 @@
+#pragma once
+// The command stream: what the adapter records from IMVU's renderer interface, and the renderer process carries out.
+// Each command is a Header, its fixed-size body, then for some a variable-size payload (pixels, vertex data),
+// padded to 4 bytes. Built for 32-bit and 64-bit: fixed-size fields only.
+
+#include <cstddef>
+#include <cstdint>
+#include <cstring>
+#include <optional>
+#include <span>
+#include <vector>
+
+namespace t5de::protocol {
+
+enum class Op : std::uint16_t {
+    BeginFrame = 1,
+    EndFrame,
+    Clear,
+    SetProjection,
+    SetModel,
+    SetView,
+    SetTextures,
+    SetState,
+    SetLight,
+    SetFog,
+    SetTextureTransform,
+    CreateTexture,
+    UpdateTexture,
+    DestroyTexture,
+    CreateBuffer,
+    UpdateBuffer,
+    DestroyBuffer,
+    DrawTriangles,
+    DrawArrays,
+    DrawPointSprites,
+};
+
+struct Header {
+    Op op;
+    std::uint16_t reserved;
+    std::uint32_t size; // of the body and payload, padded to 4 bytes
+};
+
+// Colors are 0xAARRGGBB. Resource ids are the adapter's, per engine; 0 means none.
+
+struct BeginFrame {
+    std::uint32_t offscreen; // drawn into a render target, not the window: nothing to show
+};
+
+struct EndFrame {
+    std::uint64_t sentAt; // the adapter's QueryPerformanceCounter (the same clock in every process)
+};
+
+struct Clear {
+    std::uint32_t flags; // 1: color, 2: depth (to 1)
+    std::uint32_t color;
+};
+
+// The viewport's size, and a Direct3D projection matrix (row-major, row vectors).
+struct SetProjection {
+    std::int32_t width;
+    std::int32_t height;
+    float m[16];
+};
+
+// An affine transform (SetModel, SetView): a 3x3 matrix, row-major, then the translation.
+struct SetTransform {
+    float m[12];
+};
+
+struct SetTextures {
+    std::uint32_t texture0;
+    std::uint32_t texture1;
+};
+
+enum class State : std::uint32_t {
+    DepthTest,
+    DepthWrite,
+    Blending,
+    BlendMode,      // 1: alpha, 2: additive, 3: premultiplied, 4: additive alpha
+    AlphaTest,
+    AlphaReference, // 0-255: pixels pass if their alpha is greater
+    Culling,
+    NormalInverting, // flips which winding faces front
+    Lighting,
+    Fog,
+    Ambient,         // a color
+};
+
+struct SetState {
+    State state;
+    std::uint32_t value;
+};
+
+struct SetLight {
+    std::uint32_t index;
+    std::uint32_t enabled;
+    std::uint32_t color;
+    std::uint32_t directional; // 0: a point light at vector; 1: light coming from direction vector
+    float vector[3];
+};
+
+struct SetFog {
+    float start;
+    float end;
+    std::uint32_t color;
+};
+
+// Texture coordinates for a stage: u' = u * m[0] + v * m[1] + m[4], v' = u * m[2] + v * m[3] + m[5].
+struct SetTextureTransform {
+    std::uint32_t stage;
+    std::uint32_t identity;
+    float m[6];
+};
+
+// Payload: width * height pixels (0xAARRGGBB), rows packed.
+struct CreateTexture {
+    std::uint32_t id;
+    std::uint32_t width;
+    std::uint32_t height;
+    std::uint32_t flags; // 2: wrap u, 4: wrap v (else clamp)
+};
+
+struct UpdateTexture {
+    std::uint32_t id;
+    std::uint32_t width;
+    std::uint32_t height;
+};
+
+struct DestroyTexture {
+    std::uint32_t id;
+};
+
+enum class BufferKind : std::uint32_t {
+    Index,  // 16-bit indices
+    Vertex, // positions and normals (3 floats), texture coordinates (2 floats) or colors (0xAARRGGBB)
+};
+
+// Payload: count elements of elementSize bytes, packed; empty for a buffer filled later (UpdateBuffer).
+struct CreateBuffer {
+    std::uint32_t id;
+    BufferKind kind;
+    std::uint32_t elementSize;
+    std::uint32_t count;
+};
+
+struct UpdateBuffer {
+    std::uint32_t id;
+    std::uint32_t elementSize;
+    std::uint32_t count;
+};
+
+struct DestroyBuffer {
+    std::uint32_t id;
+};
+
+// Triangles from buffers: indices and positions are required, the others optional.
+struct DrawTriangles {
+    std::uint32_t indices;
+    std::uint32_t positions;
+    std::uint32_t normals;
+    std::uint32_t texCoords;
+    std::uint32_t colors;
+};
+
+// Triangles from data in the payload, packed: indexCount 16-bit indices (padded to 4 bytes), then vertexCount
+// positions, and normals, texture coordinates and colors if present.
+struct DrawArrays {
+    std::uint32_t vertexCount;
+    std::uint32_t indexCount;
+    std::uint32_t hasNormals;
+    std::uint32_t hasTexCoords;
+    std::uint32_t hasColors;
+};
+
+// Particles: payload count sprites of 16 bytes, a position (3 floats) and a color. Each is a camera-facing square,
+// viewport height / distance pixels across (1 to 64), textured with the first texture over the whole square.
+struct DrawPointSprites {
+    std::uint32_t count;
+};
+
+static_assert(sizeof(Header) == 8 && sizeof(SetProjection) == 72 && sizeof(SetLight) == 28
+              && sizeof(DrawArrays) == 20);
+
+inline constexpr std::size_t padded(std::size_t size) { return (size + 3) & ~std::size_t{3}; }
+
+// Appends commands to a byte buffer.
+class Writer {
+public:
+    // Appends a command and returns its payload, payloadSize bytes for the caller to fill.
+    template <class Body>
+    std::byte* add(Op op, const Body& body, std::size_t payloadSize = 0)
+    {
+        static_assert(sizeof(Body) % 4 == 0);
+        const std::size_t size = sizeof(Body) + padded(payloadSize);
+        const std::size_t start = bytes_.size();
+        bytes_.resize(start + sizeof(Header) + size);
+        const Header header{op, 0, static_cast<std::uint32_t>(size)};
+        std::memcpy(bytes_.data() + start, &header, sizeof(Header));
+        std::memcpy(bytes_.data() + start + sizeof(Header), &body, sizeof(Body));
+        return bytes_.data() + start + sizeof(Header) + sizeof(Body);
+    }
+
+    [[nodiscard]] std::span<const std::byte> bytes() const { return bytes_; }
+    [[nodiscard]] std::size_t size() const { return bytes_.size(); }
+    void clear() { bytes_.clear(); }
+
+private:
+    std::vector<std::byte> bytes_;
+};
+
+// One command, as read back.
+class Command {
+public:
+    Command(Op op, std::span<const std::byte> data) : op_(op), data_(data) {}
+
+    [[nodiscard]] Op op() const { return op_; }
+
+    // The body, if the command is long enough to hold one.
+    template <class Body>
+    [[nodiscard]] std::optional<Body> body() const
+    {
+        if (data_.size() < sizeof(Body)) {
+            return std::nullopt;
+        }
+        Body body;
+        std::memcpy(&body, data_.data(), sizeof(Body));
+        return body;
+    }
+
+    template <class Body>
+    [[nodiscard]] std::span<const std::byte> payload() const
+    {
+        return data_.size() < sizeof(Body) ? std::span<const std::byte>{} : data_.subspan(sizeof(Body));
+    }
+
+private:
+    Op op_;
+    std::span<const std::byte> data_;
+};
+
+// Reads commands back from a byte buffer; stops at the end, or at a truncated command.
+class Reader {
+public:
+    explicit Reader(std::span<const std::byte> bytes) : bytes_(bytes) {}
+
+    std::optional<Command> next()
+    {
+        if (bytes_.size() < sizeof(Header)) {
+            return std::nullopt;
+        }
+        Header header;
+        std::memcpy(&header, bytes_.data(), sizeof(Header));
+        if (bytes_.size() - sizeof(Header) < header.size) {
+            return std::nullopt;
+        }
+        Command command(header.op, bytes_.subspan(sizeof(Header), header.size));
+        bytes_ = bytes_.subspan(sizeof(Header) + header.size);
+        return command;
+    }
+
+private:
+    std::span<const std::byte> bytes_;
+};
+
+} // namespace t5de::protocol

renderer/CMakeLists.txt

@@ -20,8 +20,21 @@ FetchContent_Declare(DiligentCore
     SYSTEM)
 FetchContent_MakeAvailable(DiligentCore)
 
-add_executable(t5de_renderer WIN32 main.cpp)
+# Shader sources, embedded in the program (Diligent compiles them at run time): each shaders/<name>.hlsl becomes
+# <name>_hlsl.hpp, defining <camelCaseName>Hlsl.
+function(t5de_embed_shader file name)
+    set(T5DE_SHADER_FILE ${file})
+    set(T5DE_SHADER_NAME ${name})
+    file(READ ${CMAKE_CURRENT_SOURCE_DIR}/shaders/${file} T5DE_SHADER_SOURCE)
+    get_filename_component(stem ${file} NAME_WE)
+    configure_file(shaders/embed.hpp.in ${CMAKE_CURRENT_BINARY_DIR}/generated/${stem}_hlsl.hpp @ONLY)
+    set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS shaders/${file})
+endfunction()
+t5de_embed_shader(fixed_function.hlsl fixedFunctionHlsl)
+
+add_executable(t5de_renderer WIN32 log.cpp main.cpp scene.cpp)
 set_target_properties(t5de_renderer PROPERTIES OUTPUT_NAME t5de-renderer)
+target_include_directories(t5de_renderer PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/generated)
 target_link_libraries(t5de_renderer PRIVATE t5de::protocol t5de::options Diligent-GraphicsEngineD3D11-static Diligent-Common)
 
 install(TARGETS t5de_renderer RUNTIME DESTINATION .)

renderer/log.cpp

@@ -0,0 +1,30 @@
+#include "log.hpp"
+
+#include <windows.h>
+
+#include <filesystem>
+#include <string>
+
+namespace t5de::log {
+
+// Each line is one appending write, so lines from several processes don't mix.
+void write(std::string_view line)
+{
+    static const HANDLE file = [] {
+        wchar_t temp[MAX_PATH + 1];
+        const DWORD length = GetTempPathW(MAX_PATH + 1, temp);
+        if (length == 0 || length > MAX_PATH) {
+            return INVALID_HANDLE_VALUE;
+        }
+        return CreateFileW((std::filesystem::path(temp) / L"t5de-renderer-process.log").c_str(), FILE_APPEND_DATA,
+                           FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, nullptr, OPEN_ALWAYS,
+                           FILE_ATTRIBUTE_NORMAL, nullptr);
+    }();
+    if (file != INVALID_HANDLE_VALUE) {
+        const std::string text = std::format("{}: {}\n", GetCurrentProcessId(), line);
+        DWORD written = 0;
+        WriteFile(file, text.data(), static_cast<DWORD>(text.size()), &written, nullptr);
+    }
+}
+
+} // namespace t5de::log

renderer/log.hpp

@@ -0,0 +1,18 @@
+#pragma once
+// The renderer processes' log: %TEMP%\t5de-renderer-process.log, shared by all of them (lines start with the pid).
+
+#include <format>
+#include <string_view>
+#include <utility>
+
+namespace t5de::log {
+
+void write(std::string_view line);
+
+template <class... Args>
+void info(std::format_string<Args...> format, Args&&... args)
+{
+    write(std::format(format, std::forward<Args>(args)...));
+}
+
+} // namespace t5de::log

renderer/main.cpp

@@ -1,31 +1,45 @@
 // t5de-renderer.exe: the renderer process. Started by the adapter (t5deRenderEngine.dll) for each of the client's
-// 3D views; draws into a child window it places in the client's window, frame by frame as the client draws.
+// 3D views: receives the client's drawing as commands through a shared memory channel, and draws it with Diligent
-//
+// into a child window it places in the client's window (over the right half, for now, to compare with the client's
-// For now a spike proving the architecture: a small box in the 3D view's corner, its color changing with each frame
+// own renderer on the left).
-// the client draws, drawn with Diligent (Direct3D 11) from this 64-bit process.
 
 #include <windows.h>
 
+#include <d3d11.h>
+#include <dxgi.h>
 #include <shellapi.h>
 
-#include <cmath>
 #include <cstdint>
+#include <cstring>
+#include <filesystem>
+#include <format>
 #include <string>
 #include <string_view>
+#include <vector>
 
 #include "EngineFactoryD3D11.h"
 #include "RefCntAutoPtr.hpp"
+#include "RenderDeviceD3D11.h"
+#include "log.hpp"
+#include "scene.hpp"
 #include "t5de/protocol/channel.hpp"
 
 namespace {
 
 using namespace Diligent;
+using t5de::protocol::Channel;
+using t5de::protocol::View;
 
-// The box: where in the client's 3D view, and how big.
+using t5de::log::info;
-constexpr int boxX = 16;
+
-constexpr int boxY = 16;
+// Diligent's messages (shader compilation errors, invalid states, ...), into the log.
-constexpr int boxWidth = 240;
+void DILIGENT_CALL_TYPE diligent_message(DEBUG_MESSAGE_SEVERITY severity, const Char* message, const Char* function,
-constexpr int boxHeight = 80;
+                                         const Char*, int)
+{
+    static constexpr const char* severities[] = {"info", "warning", "error", "fatal error"};
+    info("Diligent {}: {}{}{}", severities[severity < 4 ? severity : 3], function != nullptr ? function : "",
+        function != nullptr ? ": " : "", message != nullptr ? message : "");
+}
 
 struct Arguments {
     std::wstring channel;
@@ -67,7 +81,93 @@ HWND create_child_window(HWND parent)
                             .lpszClassName = L"T5DERenderer"};
     RegisterClassExW(&windowClass);
     return CreateWindowExW(0, windowClass.lpszClassName, L"", WS_CHILD | WS_VISIBLE | WS_DISABLED | WS_CLIPSIBLINGS,
-                           boxX, boxY, boxWidth, boxHeight, parent, nullptr, windowClass.hInstance, nullptr);
+                           0, 0, 1, 1, parent, nullptr, windowClass.hInstance, nullptr);
+}
+
+// The view, as the adapter set it, or as %TEMP%\t5de-renderer-view.txt says while the client runs (off, split, full
+// or compare): for comparing the renderers without restarting the client. Read every second or so.
+View current_view(View adapter)
+{
+    static View view = adapter;
+    static ULONGLONG checked = 0;
+    if (GetTickCount64() - checked < 1000) {
+        return view;
+    }
+    checked = GetTickCount64();
+    view = adapter;
+    wchar_t temp[MAX_PATH + 1];
+    const DWORD length = GetTempPathW(MAX_PATH + 1, temp);
+    if (length == 0 || length > MAX_PATH) {
+        return view;
+    }
+    const HANDLE file = CreateFileW((std::filesystem::path(temp) / L"t5de-renderer-view.txt").c_str(), GENERIC_READ,
+                                    FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, nullptr, OPEN_EXISTING,
+                                    FILE_ATTRIBUTE_NORMAL, nullptr);
+    if (file == INVALID_HANDLE_VALUE) {
+        return view;
+    }
+    char text[16] = {};
+    DWORD read = 0;
+    ReadFile(file, text, sizeof(text) - 1, &read, nullptr);
+    CloseHandle(file);
+    const std::string_view setting(text, std::strcspn(text, "\r\n "));
+    view = setting == "off"       ? View::Off
+           : setting == "split"   ? View::Split
+           : setting == "full"    ? View::Full
+           : setting == "compare" ? View::Compare
+                                  : adapter;
+    return view;
+}
+
+// Shader changes to try while the client runs: %TEMP%\t5de-renderer-shaders\fixed_function.hlsl, if there is
+// one, replaces the built-in shaders, and is read again when it changes (checked every second or so).
+void reload_shaders(t5de::Scene& scene)
+{
+    static ULONGLONG checked = 0;
+    static std::filesystem::file_time_type loaded{};
+    if (GetTickCount64() - checked < 1000) {
+        return;
+    }
+    checked = GetTickCount64();
+    std::error_code error;
+    const auto path = std::filesystem::temp_directory_path(error) / L"t5de-renderer-shaders" / L"fixed_function.hlsl";
+    const auto modified = std::filesystem::last_write_time(path, error);
+    if (error || modified == loaded) {
+        return;
+    }
+    const HANDLE file = CreateFileW(path.c_str(), GENERIC_READ, FILE_SHARE_READ | FILE_SHARE_WRITE, nullptr,
+                                    OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, nullptr);
+    if (file == INVALID_HANDLE_VALUE) {
+        return;
+    }
+    std::string source(static_cast<std::size_t>(std::filesystem::file_size(path, error)), ' ');
+    DWORD read = 0;
+    ReadFile(file, source.data(), static_cast<DWORD>(source.size()), &read, nullptr);
+    CloseHandle(file);
+    source.resize(read);
+    loaded = modified;
+    scene.setShaderSource(std::move(source));
+    info("shaders reloaded from {}", path.string());
+}
+
+// Where the child window goes in the client's 3D view, and where in the 3D view its scene starts (offset).
+struct Placement {
+    RECT window{};
+    int offset = 0;
+};
+
+Placement placement(HWND parent, View view)
+{
+    Placement place;
+    GetClientRect(parent, &place.window);
+    if (view == View::Split || view == View::Compare) {
+        place.window.left = place.window.right / 2;
+    }
+    place.offset = view == View::Compare ? 0 : place.window.left;
+    if (view == View::Off) {
+        place.window = {};
+    }
+    return place;
 }
 
 } // namespace
@@ -80,37 +180,74 @@ int WINAPI wWinMain(HINSTANCE, HINSTANCE, PWSTR, int)
     }
     HANDLE mapping = OpenFileMappingW(FILE_MAP_ALL_ACCESS, FALSE, arguments.channel.c_str());
     auto* channel = mapping == nullptr ? nullptr
-                                       : static_cast<t5de::protocol::Channel*>(MapViewOfFile(
+                                       : static_cast<Channel*>(MapViewOfFile(mapping, FILE_MAP_ALL_ACCESS, 0, 0, 0));
-                                             mapping, FILE_MAP_ALL_ACCESS, 0, 0, sizeof(t5de::protocol::Channel)));
+    HANDLE ready = OpenEventW(SYNCHRONIZE, FALSE, (arguments.channel + t5de::protocol::readyEventSuffix).c_str());
-    HANDLE frameEvent = OpenEventW(SYNCHRONIZE, FALSE, (arguments.channel + t5de::protocol::frameEventSuffix).c_str());
+    HANDLE consumed =
+        OpenEventW(EVENT_MODIFY_STATE, FALSE, (arguments.channel + t5de::protocol::consumedEventSuffix).c_str());
     HANDLE parent = OpenProcess(SYNCHRONIZE, FALSE, arguments.parent);
-    if (channel == nullptr || frameEvent == nullptr || parent == nullptr || channel->version != t5de::protocol::version) {
+    if (channel == nullptr || ready == nullptr || consumed == nullptr || parent == nullptr
+        || channel->version != t5de::protocol::version) {
         return 3;
     }
+    const auto* data = reinterpret_cast<const std::byte*>(channel + 1);
 
-    HWND window = create_child_window(reinterpret_cast<HWND>(static_cast<std::uintptr_t>(channel->window)));
+    const HWND clientWindow = reinterpret_cast<HWND>(static_cast<std::uintptr_t>(channel->window));
+    const HWND window = create_child_window(clientWindow);
     if (window == nullptr) {
         return 4;
     }
 
+    info("started for window {:#x}, view {}", channel->window, static_cast<int>(channel->view));
     IEngineFactoryD3D11* factory = GetEngineFactoryD3D11();
+    factory->SetMessageCallback(diligent_message);
     RefCntAutoPtr<IRenderDevice> device;
     RefCntAutoPtr<IDeviceContext> context;
     EngineD3D11CreateInfo engineInfo;
     factory->CreateDeviceAndContextsD3D11(engineInfo, &device, &context);
     if (!device) {
+        info("can't create a Direct3D 11 device");
         return 5;
     }
+    // Show each frame as soon as it's drawn: Direct3D 11 otherwise lets 3 frames queue, which the view would trail by.
+    if (RefCntAutoPtr<IRenderDeviceD3D11> d3d11{device, IID_RenderDeviceD3D11}) {
+        IDXGIDevice1* dxgi = nullptr;
+        if (SUCCEEDED(d3d11->GetD3D11Device()->QueryInterface(__uuidof(IDXGIDevice1), reinterpret_cast<void**>(&dxgi)))) {
+            dxgi->SetMaximumFrameLatency(1);
+            dxgi->Release();
+        }
+    }
     RefCntAutoPtr<ISwapChain> swapChain;
     SwapChainDesc swapChainDesc;
+    swapChainDesc.ColorBufferFormat = TEX_FORMAT_RGBA8_UNORM; // no sRGB conversion, like the client's renderer
+    swapChainDesc.DepthBufferFormat = TEX_FORMAT_D32_FLOAT;
     factory->CreateSwapChainD3D11(device, context, swapChainDesc, FullScreenModeDesc{}, Win32NativeWindow{window},
                                   &swapChain);
     if (!swapChain) {
+        info("can't create a swap chain");
         return 6;
     }
+    t5de::Scene scene(device, context, swapChain);
 
+    // Takes the chunk waiting in the channel (after "ready"), and lets the adapter send the next: appends it to
+    // frame, and returns whether it ends a frame.
+    std::vector<std::byte> frame;
+    const auto take_chunk = [&] {
+        const std::uint32_t size = channel->chunkSize <= channel->capacity ? channel->chunkSize : 0;
+        const bool endsFrame = (channel->chunkFlags & t5de::protocol::chunkEndsFrame) != 0;
+        frame.insert(frame.end(), data, data + size);
+        SetEvent(consumed);
+        return endsFrame;
+    };
+
+    LARGE_INTEGER frequency;
+    QueryPerformanceFrequency(&frequency);
+    double drawTime = 0;
+    double latency = 0;
+    int drawn = 0;
+    int skipped = 0;
+    RECT placed{};
     for (;;) {
-        const HANDLE handles[] = {frameEvent, parent};
+        const HANDLE handles[] = {ready, parent};
         const DWORD result = MsgWaitForMultipleObjects(2, handles, FALSE, 1000, QS_ALLINPUT);
         MSG message;
         while (PeekMessageW(&message, nullptr, 0, 0, PM_REMOVE)) {
@@ -119,19 +256,70 @@ int WINAPI wWinMain(HINSTANCE, HINSTANCE, PWSTR, int)
         if (result == WAIT_OBJECT_0 + 1 || channel->closing != 0) {
             break; // the client exited, or the engine is gone
         }
-        if (result != WAIT_OBJECT_0) {
+        if (result != WAIT_OBJECT_0 || !take_chunk()) {
             continue;
         }
 
-        // T5DE gold, pulsing with the client's frames.
+        // If the client has sent more frames meanwhile (drawing fell behind), catch up: carry out the older ones'
-        const float pulse = 0.5f + 0.5f * std::sin(static_cast<float>(channel->frames) * 0.1f);
+        // changes without drawing them, and draw only the newest. (An incomplete frame stays for the next round.)
-        const float color[] = {1.0f, 0.82f * pulse + 0.1f, 0.0f, 1.0f};
+        std::size_t complete = frame.size();
-        ITextureView* backBuffer = swapChain->GetCurrentBackBufferRTV();
+        while (WaitForSingleObject(ready, 0) == WAIT_OBJECT_0) {
-        context->SetRenderTargets(1, &backBuffer, nullptr, RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+            bool ended = take_chunk();
-        context->ClearRenderTarget(backBuffer, color, RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+            while (!ended && WaitForSingleObject(ready, 1000) == WAIT_OBJECT_0) {
-        swapChain->Present(0);
+                ended = take_chunk();
+            }
+            if (!ended) {
+                break;
+            }
+            scene.execute(std::span(frame).first(complete), 0, false);
+            frame.erase(frame.begin(), frame.begin() + static_cast<std::ptrdiff_t>(complete));
+            complete = frame.size();
+            ++skipped;
+        }
+
+        // Follow the client's 3D view's size.
+        const auto [place, offset] = placement(clientWindow, current_view(channel->view));
+        const int width = place.right - place.left;
+        const int height = place.bottom - place.top;
+        if (std::memcmp(&place, &placed, sizeof(RECT)) != 0) {
+            placed = place;
+            if (width > 0 && height > 0) {
+                SetWindowPos(window, nullptr, place.left, place.top, width, height,
+                             SWP_NOZORDER | SWP_NOACTIVATE | SWP_SHOWWINDOW);
+                swapChain->Resize(static_cast<Uint32>(width), static_cast<Uint32>(height));
+                info("drawing {}x{} at {} in the client's 3D view, its scene from {}", width, height, place.left, offset);
+            } else {
+                ShowWindow(window, SW_HIDE);
+                channel->covering = 0;
+                info("hidden");
+            }
+        }
+        reload_shaders(scene);
+        LARGE_INTEGER start, end;
+        QueryPerformanceCounter(&start);
+        const bool covers = current_view(channel->view) == View::Full && width > 0 && height > 0;
+        if (scene.execute(std::span(frame).first(complete), offset, width > 0 && height > 0)) {
+            swapChain->Present(0);
+            channel->covering = covers ? 1 : 0; // the client's renderer can stop drawing (or must start again)
+            QueryPerformanceCounter(&end);
+            drawTime += static_cast<double>(end.QuadPart - start.QuadPart) / static_cast<double>(frequency.QuadPart);
+            latency += static_cast<double>(end.QuadPart - static_cast<LONGLONG>(scene.lastSentAt())) /
+                       static_cast<double>(frequency.QuadPart);
+            if (++drawn == 300) {
+                info("last 300 frames: {:.1f} ms each to draw and present, {:.1f} ms from the client's end of frame to "
+                     "presented, {} skipped to catch up",
+                     drawTime * 1000 / drawn, latency * 1000 / drawn, skipped);
+                drawTime = 0;
+                latency = 0;
+                drawn = 0;
+                skipped = 0;
+            }
+        }
+        frame.erase(frame.begin(), frame.begin() + static_cast<std::ptrdiff_t>(complete));
     }
 
+    channel->covering = 0;
+    info("exiting");
     DestroyWindow(window);
     return 0;
 }

renderer/scene.cpp

@@ -0,0 +1,687 @@
+#include "scene.hpp"
+
+#include <cstring>
+#include <string>
+
+#include "fixed_function_hlsl.hpp"
+
+namespace t5de {
+
+using namespace Diligent;
+using protocol::Op;
+
+namespace {
+
+void set_identity(float (&m)[16])
+{
+    std::memset(m, 0, sizeof(m));
+    m[0] = m[5] = m[10] = m[15] = 1;
+}
+
+// An affine transform (3x3 row-major, then translation) as a 4x4 matrix for row vectors.
+void set_transform(float (&m)[16], const float (&t)[12])
+{
+    for (int row = 0; row < 3; ++row) {
+        m[row * 4 + 0] = t[row * 3 + 0];
+        m[row * 4 + 1] = t[row * 3 + 1];
+        m[row * 4 + 2] = t[row * 3 + 2];
+        m[row * 4 + 3] = 0;
+    }
+    m[12] = t[9];
+    m[13] = t[10];
+    m[14] = t[11];
+    m[15] = 1;
+}
+
+// u' = u * m[0] + v * m[1] + m[4], v' = u * m[2] + v * m[3] + m[5], as two rows dotted with (u, v, 1).
+void set_texture_transform(float (&rows)[8], const float (&m)[6])
+{
+    const float values[8] = {m[0], m[1], m[4], 0, m[2], m[3], m[5], 0};
+    std::memcpy(rows, values, sizeof(rows));
+}
+
+void set_color(float (&rgba)[4], std::uint32_t color)
+{
+    rgba[0] = static_cast<float>((color >> 16) & 0xff) / 255.0f;
+    rgba[1] = static_cast<float>((color >> 8) & 0xff) / 255.0f;
+    rgba[2] = static_cast<float>(color & 0xff) / 255.0f;
+    rgba[3] = static_cast<float>(color >> 24) / 255.0f;
+}
+
+constexpr std::uint32_t hasNormals = 1;
+constexpr std::uint32_t hasTexCoords = 2;
+constexpr std::uint32_t hasColors = 4;
+
+} // namespace
+
+Scene::Scene(IRenderDevice* device, IDeviceContext* context, ISwapChain* swapChain)
+    : device_(device), context_(context), swapChain_(swapChain), shaderSource_(fixedFunctionHlsl)
+{
+    BufferDesc constants;
+    constants.Name = "Constants";
+    constants.Size = sizeof(Constants);
+    constants.Usage = USAGE_DYNAMIC;
+    constants.BindFlags = BIND_UNIFORM_BUFFER;
+    constants.CPUAccessFlags = CPU_ACCESS_WRITE;
+    device_->CreateBuffer(constants, nullptr, &constantBuffer_);
+
+    // IMVU's textures: bilinear, no mipmaps, wrapped or clamped per texture.
+    for (int wrap = 0; wrap < 4; ++wrap) {
+        SamplerDesc sampler;
+        sampler.MinFilter = FILTER_TYPE_LINEAR;
+        sampler.MagFilter = FILTER_TYPE_LINEAR;
+        sampler.MipFilter = FILTER_TYPE_POINT;
+        sampler.AddressU = (wrap & 1) != 0 ? TEXTURE_ADDRESS_WRAP : TEXTURE_ADDRESS_CLAMP;
+        sampler.AddressV = (wrap & 2) != 0 ? TEXTURE_ADDRESS_WRAP : TEXTURE_ADDRESS_CLAMP;
+        device_->CreateSampler(sampler, &samplers_[wrap]);
+    }
+
+    set_identity(constants_.world);
+    set_identity(constants_.view);
+    set_identity(constants_.projection);
+    const float identity[6] = {1, 0, 0, 1, 0, 0};
+    set_texture_transform(constants_.texTransform0, identity);
+    set_texture_transform(constants_.texTransform1, identity);
+    constants_.flags[2] = 127; // the client's alpha reference
+}
+
+void Scene::setShaderSource(std::string source)
+{
+    shaderSource_ = std::move(source);
+    shaders_.clear();
+    pipelines_.clear();
+}
+
+bool Scene::execute(std::span<const std::byte> commands, int offsetX, bool draw)
+{
+    offsetX_ = offsetX;
+    draw_ = draw;
+    bool drew = false;
+    protocol::Reader reader(commands);
+    while (const auto command = reader.next()) {
+        switch (command->op()) {
+        case Op::BeginFrame:
+            if (const auto begin = command->body<protocol::BeginFrame>()) {
+                onscreen_ = begin->offscreen == 0 && draw_;
+                if (onscreen_) {
+                    ITextureView* target = swapChain_->GetCurrentBackBufferRTV();
+                    context_->SetRenderTargets(1, &target, swapChain_->GetDepthBufferDSV(),
+                                               RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+                }
+            }
+            break;
+        case Op::EndFrame:
+            if (const auto end = command->body<protocol::EndFrame>()) {
+                lastSentAt_ = end->sentAt;
+            }
+            drew = drew || onscreen_;
+            onscreen_ = false;
+            break;
+        case Op::Clear:
+            if (const auto clear = command->body<protocol::Clear>(); clear && onscreen_) {
+                if ((clear->flags & 1) != 0) {
+                    float color[4];
+                    set_color(color, clear->color);
+                    context_->ClearRenderTarget(swapChain_->GetCurrentBackBufferRTV(), color,
+                                                RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+                }
+                if ((clear->flags & 2) != 0) {
+                    context_->ClearDepthStencil(swapChain_->GetDepthBufferDSV(), CLEAR_DEPTH_FLAG, 1.0f, 0,
+                                                RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+                }
+            }
+            break;
+        case Op::SetProjection:
+            if (const auto projection = command->body<protocol::SetProjection>()) {
+                setProjection(*projection);
+            }
+            break;
+        case Op::SetModel:
+            if (const auto model = command->body<protocol::SetTransform>()) {
+                set_transform(constants_.world, model->m);
+            }
+            break;
+        case Op::SetView:
+            if (const auto view = command->body<protocol::SetTransform>()) {
+                set_transform(constants_.view, view->m);
+            }
+            break;
+        case Op::SetTextures:
+            if (const auto textures = command->body<protocol::SetTextures>()) {
+                state_.texture0 = textures->texture0;
+                state_.texture1 = textures->texture1;
+            }
+            break;
+        case Op::SetState:
+            if (const auto state = command->body<protocol::SetState>()) {
+                setState(*state);
+            }
+            break;
+        case Op::SetLight:
+            if (const auto light = command->body<protocol::SetLight>(); light && light->index < 8) {
+                set_color(constants_.lightColor[light->index], light->color);
+                constants_.lightColor[light->index][3] = light->enabled != 0 ? 1.0f : 0.0f;
+                std::memcpy(constants_.lightVector[light->index], light->vector, sizeof(light->vector));
+                constants_.lightVector[light->index][3] = light->directional != 0 ? 1.0f : 0.0f;
+            }
+            break;
+        case Op::SetFog:
+            if (const auto fog = command->body<protocol::SetFog>()) {
+                constants_.fog[0] = fog->start;
+                constants_.fog[1] = fog->end;
+                set_color(constants_.fogColor, fog->color);
+            }
+            break;
+        case Op::SetTextureTransform:
+            if (const auto transform = command->body<protocol::SetTextureTransform>(); transform && transform->stage < 2) {
+                set_texture_transform(transform->stage == 0 ? constants_.texTransform0 : constants_.texTransform1,
+                                      transform->m);
+            }
+            break;
+        case Op::CreateTexture:
+            createTexture(*command);
+            break;
+        case Op::UpdateTexture:
+            updateTexture(*command);
+            break;
+        case Op::DestroyTexture:
+            if (const auto destroy = command->body<protocol::DestroyTexture>()) {
+                textures_.erase(destroy->id);
+            }
+            break;
+        case Op::CreateBuffer:
+            createBuffer(*command);
+            break;
+        case Op::UpdateBuffer:
+            updateBuffer(*command);
+            break;
+        case Op::DestroyBuffer:
+            if (const auto destroy = command->body<protocol::DestroyBuffer>()) {
+                buffers_.erase(destroy->id);
+            }
+            break;
+        case Op::DrawTriangles:
+            if (const auto triangles = command->body<protocol::DrawTriangles>(); triangles && onscreen_) {
+                drawTriangles(*triangles);
+            }
+            break;
+        case Op::DrawArrays:
+            if (onscreen_) {
+                drawArrays(*command);
+            }
+            break;
+        case Op::DrawPointSprites:
+            if (onscreen_) {
+                drawPointSprites(*command);
+            }
+            break;
+        }
+    }
+    return drew;
+}
+
+void Scene::setProjection(const protocol::SetProjection& projection)
+{
+    std::memcpy(constants_.projection, projection.m, sizeof(constants_.projection));
+    // The client's projections give depth from -1 to 1, as OpenGL does; IMVU's Direct3D renderer maps it to 0 to 1
+    // (z' = z / 2 + w / 2) before setting them, and so do we. (It also shifts them half a pixel, which Direct3D 9
+    // needs and Direct3D 11 doesn't.)
+    for (int row = 0; row < 4; ++row) {
+        float* m = constants_.projection + row * 4;
+        m[2] = 0.5f * m[2] + 0.5f * m[3];
+    }
+    constants_.viewport[0] = static_cast<float>(projection.width);
+    constants_.viewport[1] = static_cast<float>(projection.height);
+    if (!onscreen_) {
+        return;
+    }
+    // The client's projection covers its whole 3D view; the window shows part of it, from offsetX.
+    const SwapChainDesc& window = swapChain_->GetDesc();
+    Viewport viewport;
+    viewport.TopLeftX = static_cast<float>(-offsetX_);
+    viewport.TopLeftY = 0;
+    viewport.Width = static_cast<float>(projection.width);
+    viewport.Height = static_cast<float>(projection.height);
+    viewport.MinDepth = 0;
+    viewport.MaxDepth = 1;
+    context_->SetViewports(1, &viewport, window.Width, window.Height);
+}
+
+void Scene::setState(const protocol::SetState& state)
+{
+    const bool on = state.value != 0;
+    switch (state.state) {
+    case protocol::State::DepthTest:
+        state_.depthTest = on;
+        break;
+    case protocol::State::DepthWrite:
+        state_.depthWrite = on;
+        break;
+    case protocol::State::Blending:
+        state_.blending = on;
+        break;
+    case protocol::State::BlendMode:
+        if (state.value >= 1 && state.value <= 4) {
+            state_.blendMode = state.value;
+        }
+        break;
+    case protocol::State::AlphaTest:
+        constants_.flags[1] = on ? 1.0f : 0.0f;
+        break;
+    case protocol::State::AlphaReference:
+        constants_.flags[2] = static_cast<float>(state.value);
+        break;
+    case protocol::State::Culling:
+        state_.culling = on;
+        break;
+    case protocol::State::NormalInverting:
+        state_.normalInverting = on;
+        break;
+    case protocol::State::Lighting:
+        constants_.flags[0] = on ? 1.0f : 0.0f;
+        break;
+    case protocol::State::Fog:
+        constants_.fog[2] = on ? 1.0f : 0.0f;
+        break;
+    case protocol::State::Ambient:
+        set_color(constants_.ambient, state.value);
+        break;
+    }
+}
+
+void Scene::createTexture(const protocol::Command& command)
+{
+    const auto create = command.body<protocol::CreateTexture>();
+    const auto pixels = command.payload<protocol::CreateTexture>();
+    if (!create || create->width == 0 || create->height == 0
+        || pixels.size() < std::size_t{create->width} * create->height * 4) {
+        return;
+    }
+    TextureDesc desc;
+    desc.Name = "Client texture";
+    desc.Type = RESOURCE_DIM_TEX_2D;
+    desc.Width = create->width;
+    desc.Height = create->height;
+    desc.MipLevels = 1;
+    desc.Format = TEX_FORMAT_BGRA8_UNORM; // 0xAARRGGBB in memory
+    desc.BindFlags = BIND_SHADER_RESOURCE;
+    desc.Usage = USAGE_DEFAULT;
+    TextureSubResData level;
+    level.pData = pixels.data();
+    level.Stride = std::uint64_t{create->width} * 4;
+    TextureData data;
+    data.pSubResources = &level;
+    data.NumSubresources = 1;
+    Texture texture{.width = create->width, .height = create->height, .flags = create->flags};
+    device_->CreateTexture(desc, &data, &texture.texture);
+    if (!texture.texture) {
+        return;
+    }
+    texture.texture->GetDefaultView(TEXTURE_VIEW_SHADER_RESOURCE)
+        ->SetSampler(samplers_[((create->flags & 2) != 0 ? 1 : 0) | ((create->flags & 4) != 0 ? 2 : 0)]);
+    textures_[create->id] = std::move(texture);
+}
+
+void Scene::updateTexture(const protocol::Command& command)
+{
+    const auto update = command.body<protocol::UpdateTexture>();
+    const auto pixels = command.payload<protocol::UpdateTexture>();
+    if (!update || pixels.size() < std::size_t{update->width} * update->height * 4) {
+        return;
+    }
+    const auto found = textures_.find(update->id);
+    if (found == textures_.end()) {
+        return;
+    }
+    Texture& texture = found->second;
+    if (texture.width != update->width || texture.height != update->height) {
+        // A different size: make it again, keeping its sampling.
+        protocol::Writer writer;
+        std::memcpy(writer.add(Op::CreateTexture,
+                               protocol::CreateTexture{update->id, update->width, update->height, texture.flags},
+                               pixels.size()),
+                    pixels.data(), pixels.size());
+        protocol::Reader reader(writer.bytes());
+        createTexture(*reader.next());
+        return;
+    }
+    Box box;
+    box.MaxX = update->width;
+    box.MaxY = update->height;
+    TextureSubResData level;
+    level.pData = pixels.data();
+    level.Stride = std::uint64_t{update->width} * 4;
+    context_->UpdateTexture(texture.texture, 0, 0, box, level, RESOURCE_STATE_TRANSITION_MODE_TRANSITION,
+                            RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+}
+
+void Scene::createBuffer(const protocol::Command& command)
+{
+    const auto create = command.body<protocol::CreateBuffer>();
+    if (!create || create->count == 0 || create->elementSize == 0) {
+        return;
+    }
+    const std::size_t size = std::size_t{create->elementSize} * create->count;
+    const auto contents = command.payload<protocol::CreateBuffer>();
+    const bool filled = contents.size() >= size;
+    BufferDesc desc;
+    desc.Name = create->kind == protocol::BufferKind::Index ? "Client indices" : "Client vertices";
+    desc.Size = size;
+    desc.BindFlags = create->kind == protocol::BufferKind::Index ? BIND_INDEX_BUFFER : BIND_VERTEX_BUFFER;
+    desc.Usage = filled ? USAGE_IMMUTABLE : USAGE_DEFAULT; // dynamic buffers are filled by updates
+    BufferData data;
+    data.pData = contents.data();
+    data.DataSize = size;
+    Buffer buffer{.kind = create->kind, .elementSize = create->elementSize, .count = create->count};
+    device_->CreateBuffer(desc, filled ? &data : nullptr, &buffer.buffer);
+    if (buffer.buffer) {
+        buffers_[create->id] = std::move(buffer);
+    }
+}
+
+void Scene::updateBuffer(const protocol::Command& command)
+{
+    const auto update = command.body<protocol::UpdateBuffer>();
+    const auto contents = command.payload<protocol::UpdateBuffer>();
+    if (!update) {
+        return;
+    }
+    const auto found = buffers_.find(update->id);
+    const std::size_t size = std::size_t{update->elementSize} * update->count;
+    if (found == buffers_.end() || contents.size() < size || found->second.elementSize != update->elementSize
+        || found->second.count != update->count || found->second.buffer->GetDesc().Usage != USAGE_DEFAULT) {
+        return;
+    }
+    context_->UpdateBuffer(found->second.buffer, 0, size, contents.data(), RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+}
+
+ITextureView* Scene::textureView(std::uint32_t id)
+{
+    const auto found = textures_.find(id);
+    return found == textures_.end() ? nullptr : found->second.texture->GetDefaultView(TEXTURE_VIEW_SHADER_RESOURCE);
+}
+
+void Scene::drawTriangles(const protocol::DrawTriangles& draw)
+{
+    const auto buffer = [this](std::uint32_t id, protocol::BufferKind kind, std::uint32_t elementSize) -> Buffer* {
+        const auto found = buffers_.find(id);
+        if (id == 0 || found == buffers_.end() || found->second.kind != kind || found->second.elementSize != elementSize) {
+            return nullptr;
+        }
+        return &found->second;
+    };
+    Buffer* indices = buffer(draw.indices, protocol::BufferKind::Index, 2);
+    Buffer* positions = buffer(draw.positions, protocol::BufferKind::Vertex, 12);
+    if (indices == nullptr || positions == nullptr) {
+        return;
+    }
+    // Optional streams count only if they cover every vertex.
+    const auto optional = [&](std::uint32_t id, std::uint32_t elementSize) -> IBuffer* {
+        Buffer* found = buffer(id, protocol::BufferKind::Vertex, elementSize);
+        return found != nullptr && found->count >= positions->count ? found->buffer.RawPtr() : nullptr;
+    };
+    this->draw(indices->buffer, indices->count, positions->buffer, optional(draw.normals, 12),
+               optional(draw.texCoords, 8), optional(draw.colors, 4));
+}
+
+void Scene::drawArrays(const protocol::Command& command)
+{
+    const auto arrays = command.body<protocol::DrawArrays>();
+    const auto payload = command.payload<protocol::DrawArrays>();
+    if (!arrays || arrays->vertexCount == 0 || arrays->indexCount == 0) {
+        return;
+    }
+    const std::size_t indexBytes = protocol::padded(std::size_t{arrays->indexCount} * 2);
+    const std::size_t vertices = arrays->vertexCount;
+    if (payload.size() < indexBytes + vertices * (12 + (arrays->hasNormals ? 12 : 0) + (arrays->hasTexCoords ? 8 : 0) +
+                                                  (arrays->hasColors ? 4 : 0))) {
+        return;
+    }
+    const auto make = [this](const std::byte* data, std::size_t size, BIND_FLAGS bind) {
+        BufferDesc desc;
+        desc.Name = "Client arrays";
+        desc.Size = size;
+        desc.BindFlags = bind;
+        desc.Usage = USAGE_IMMUTABLE;
+        BufferData contents;
+        contents.pData = data;
+        contents.DataSize = size;
+        RefCntAutoPtr<IBuffer> buffer;
+        device_->CreateBuffer(desc, &contents, &buffer);
+        return buffer;
+    };
+    const std::byte* next = payload.data();
+    const auto indices = make(next, std::size_t{arrays->indexCount} * 2, BIND_INDEX_BUFFER);
+    next += indexBytes;
+    const auto positions = make(next, vertices * 12, BIND_VERTEX_BUFFER);
+    next += vertices * 12;
+    RefCntAutoPtr<IBuffer> normals, texCoords, colors;
+    if (arrays->hasNormals) {
+        normals = make(next, vertices * 12, BIND_VERTEX_BUFFER);
+        next += vertices * 12;
+    }
+    if (arrays->hasTexCoords) {
+        texCoords = make(next, vertices * 8, BIND_VERTEX_BUFFER);
+        next += vertices * 8;
+    }
+    if (arrays->hasColors) {
+        colors = make(next, vertices * 4, BIND_VERTEX_BUFFER);
+    }
+    draw(indices, arrays->indexCount, positions, normals, texCoords, colors);
+}
+
+void Scene::draw(IBuffer* indices, std::uint32_t indexCount, IBuffer* positions, IBuffer* normals, IBuffer* texCoords,
+                 IBuffer* colors)
+{
+    if (indices == nullptr || positions == nullptr || indexCount == 0) {
+        return;
+    }
+    const std::uint32_t streams =
+        (normals != nullptr ? hasNormals : 0) | (texCoords != nullptr ? hasTexCoords : 0) | (colors != nullptr ? hasColors : 0);
+    if (prepare(streams, false) == nullptr) {
+        return;
+    }
+
+    IBuffer* vertexBuffers[4];
+    std::uint32_t count = 0;
+    for (IBuffer* stream : {positions, normals, texCoords, colors}) {
+        if (stream != nullptr) {
+            vertexBuffers[count++] = stream;
+        }
+    }
+    context_->SetVertexBuffers(0, count, vertexBuffers, nullptr, RESOURCE_STATE_TRANSITION_MODE_TRANSITION,
+                               SET_VERTEX_BUFFERS_FLAG_RESET);
+    context_->SetIndexBuffer(indices, 0, RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+    DrawIndexedAttribs attributes;
+    attributes.NumIndices = indexCount;
+    attributes.IndexType = VT_UINT16;
+    context_->DrawIndexed(attributes);
+}
+
+Scene::Pipeline* Scene::prepare(std::uint32_t streams, bool sprites)
+{
+    // Like the client: no textures, the first, or both (the second only with a first).
+    ITextureView* texture0 = textureView(state_.texture0);
+    ITextureView* texture1 = texture0 != nullptr ? textureView(state_.texture1) : nullptr;
+    const std::uint32_t textures = texture0 == nullptr ? 0 : texture1 == nullptr ? 1 : 2;
+    Pipeline* selected = pipeline(streams, textures, sprites);
+    if (selected == nullptr) {
+        return nullptr;
+    }
+
+    void* mapped = nullptr;
+    context_->MapBuffer(constantBuffer_, MAP_WRITE, MAP_FLAG_DISCARD, mapped);
+    if (mapped == nullptr) {
+        return nullptr;
+    }
+    std::memcpy(mapped, &constants_, sizeof(constants_));
+    context_->UnmapBuffer(constantBuffer_, MAP_WRITE);
+
+    if (selected->texture0 != nullptr) {
+        selected->texture0->Set(texture0);
+    }
+    if (selected->texture1 != nullptr) {
+        selected->texture1->Set(texture1);
+    }
+    context_->SetPipelineState(selected->state);
+    context_->CommitShaderResources(selected->bindings, RESOURCE_STATE_TRANSITION_MODE_TRANSITION);
+    return selected;
+}
+
+void Scene::drawPointSprites(const protocol::Command& command)
+{
+    const auto sprites = command.body<protocol::DrawPointSprites>();
+    const auto payload = command.payload<protocol::DrawPointSprites>();
+    if (!sprites || sprites->count == 0 || payload.size() < std::size_t{sprites->count} * 16) {
+        return;
+    }
+    BufferDesc desc;
+    desc.Name = "Client particles";
+    desc.Size = std::size_t{sprites->count} * 16;
+    desc.BindFlags = BIND_VERTEX_BUFFER;
+    desc.Usage = USAGE_IMMUTABLE;
+    BufferData contents;
+    contents.pData = payload.data();
+    contents.DataSize = desc.Size;
+    RefCntAutoPtr<IBuffer> buffer;
+    device_->CreateBuffer(desc, &contents, &buffer);
+    if (!buffer || prepare(0, true) == nullptr) {
+        return;
+    }
+    IBuffer* buffers[] = {buffer};
+    context_->SetVertexBuffers(0, 1, buffers, nullptr, RESOURCE_STATE_TRANSITION_MODE_TRANSITION,
+                               SET_VERTEX_BUFFERS_FLAG_RESET);
+    DrawAttribs attributes;
+    attributes.NumVertices = 4;
+    attributes.NumInstances = sprites->count;
+    context_->Draw(attributes);
+}
+
+const Scene::Shaders& Scene::shaders(std::uint32_t streams, std::uint32_t textures, bool sprites)
+{
+    const std::uint32_t key = streams | (textures << 3) | (sprites ? 1u << 5 : 0);
+    if (const auto found = shaders_.find(key); found != shaders_.end()) {
+        return found->second;
+    }
+    const std::string source = std::string("#define HAS_NORMALS ") + ((streams & hasNormals) != 0 ? "1" : "0") +
+                               "\n#define HAS_TEX_COORDS " + ((streams & hasTexCoords) != 0 ? "1" : "0") +
+                               "\n#define HAS_COLORS " + ((streams & hasColors) != 0 ? "1" : "0") +
+                               "\n#define TEXTURES " + std::to_string(textures) + "\n" + shaderSource_;
+    Shaders created;
+    ShaderCreateInfo info;
+    info.SourceLanguage = SHADER_SOURCE_LANGUAGE_HLSL;
+    info.Desc.UseCombinedTextureSamplers = true; // g_TextureN with g_TextureN_sampler, set from each texture's view
+    info.Source = source.c_str();
+    info.Desc.ShaderType = SHADER_TYPE_VERTEX;
+    info.Desc.Name = sprites ? "Fixed function particles (vertex)" : "Fixed function (vertex)";
+    info.EntryPoint = sprites ? "vs_sprite" : "vs_main";
+    device_->CreateShader(info, &created.vertex);
+    info.Desc.ShaderType = SHADER_TYPE_PIXEL;
+    info.Desc.Name = "Fixed function (pixel)";
+    info.EntryPoint = "ps_main";
+    device_->CreateShader(info, &created.pixel);
+    return shaders_[key] = std::move(created);
+}
+
+Scene::Pipeline* Scene::pipeline(std::uint32_t streams, std::uint32_t textures, bool sprites)
+{
+    const std::uint32_t blend = state_.blending ? state_.blendMode : 0;
+    const bool culling = state_.culling && !sprites; // point sprites aren't culled
+    const std::uint32_t key = streams | (textures << 3) | (blend << 5) | (state_.depthTest ? 1u << 8 : 0) |
+                              (state_.depthWrite ? 1u << 9 : 0) | (culling ? 1u << 10 : 0) |
+                              (state_.normalInverting ? 1u << 11 : 0) | (sprites ? 1u << 12 : 0);
+    if (const auto found = pipelines_.find(key); found != pipelines_.end()) {
+        return found->second.state ? &found->second : nullptr;
+    }
+    Pipeline& entry = pipelines_[key];
+    const Shaders& compiled = shaders(streams, textures, sprites);
+    if (!compiled.vertex || !compiled.pixel) {
+        return nullptr;
+    }
+
+    GraphicsPipelineStateCreateInfo info;
+    info.PSODesc.Name = "Fixed function";
+    info.PSODesc.PipelineType = PIPELINE_TYPE_GRAPHICS;
+    const SwapChainDesc& window = swapChain_->GetDesc();
+    GraphicsPipelineDesc& graphics = info.GraphicsPipeline;
+    graphics.NumRenderTargets = 1;
+    graphics.RTVFormats[0] = window.ColorBufferFormat;
+    graphics.DSVFormat = window.DepthBufferFormat;
+    graphics.PrimitiveTopology = sprites ? PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP : PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
+    // The client culls clockwise triangles, or counter-clockwise ones when normals are inverted (mirrored objects).
+    graphics.RasterizerDesc.CullMode = culling ? CULL_MODE_BACK : CULL_MODE_NONE;
+    graphics.RasterizerDesc.FrontCounterClockwise = !state_.normalInverting;
+    graphics.DepthStencilDesc.DepthEnable = state_.depthTest;
+    graphics.DepthStencilDesc.DepthWriteEnable = state_.depthWrite;
+    graphics.DepthStencilDesc.DepthFunc = COMPARISON_FUNC_LESS;
+    RenderTargetBlendDesc& blending = graphics.BlendDesc.RenderTargets[0];
+    blending.BlendEnable = blend != 0;
+    if (blend != 0) {
+        const BLEND_FACTOR source = blend == 1 || blend == 4 ? BLEND_FACTOR_SRC_ALPHA : BLEND_FACTOR_ONE;
+        const BLEND_FACTOR destination = blend == 2 || blend == 4 ? BLEND_FACTOR_ONE : BLEND_FACTOR_INV_SRC_ALPHA;
+        blending.SrcBlend = blending.SrcBlendAlpha = source;
+        blending.DestBlend = blending.DestBlendAlpha = destination;
+        blending.BlendOp = blending.BlendOpAlpha = BLEND_OPERATION_ADD;
+    }
+
+    LayoutElement elements[4];
+    std::uint32_t count = 0;
+    if (sprites) {
+        // One buffer of sprites, each a position and a color, read once per square (instance).
+        elements[0] = LayoutElement{0, 0, 3, VT_FLOAT32, False};
+        elements[1] = LayoutElement{3, 0, 4, VT_UINT8, True};
+        elements[1].RelativeOffset = 12;
+        for (LayoutElement& element : std::span(elements).first(2)) {
+            element.Stride = 16;
+            element.Frequency = INPUT_ELEMENT_FREQUENCY_PER_INSTANCE;
+            element.InstanceDataStepRate = 1;
+        }
+        count = 2;
+    } else {
+        elements[count] = LayoutElement{0, count, 3, VT_FLOAT32, False};
+        ++count;
+    }
+    if (!sprites && (streams & hasNormals) != 0) {
+        elements[count] = LayoutElement{1, count, 3, VT_FLOAT32, False};
+        ++count;
+    }
+    if (!sprites && (streams & hasTexCoords) != 0) {
+        elements[count] = LayoutElement{2, count, 2, VT_FLOAT32, False};
+        ++count;
+    }
+    if (!sprites && (streams & hasColors) != 0) {
+        elements[count] = LayoutElement{3, count, 4, VT_UINT8, True};
+        ++count;
+    }
+    graphics.InputLayout.LayoutElements = elements;
+    graphics.InputLayout.NumElements = count;
+
+    info.pVS = compiled.vertex;
+    info.pPS = compiled.pixel;
+    const ShaderResourceVariableDesc variables[] = {
+        {SHADER_TYPE_PIXEL, "g_Texture0", SHADER_RESOURCE_VARIABLE_TYPE_DYNAMIC},
+        {SHADER_TYPE_PIXEL, "g_Texture1", SHADER_RESOURCE_VARIABLE_TYPE_DYNAMIC},
+    };
+    info.PSODesc.ResourceLayout.DefaultVariableType = SHADER_RESOURCE_VARIABLE_TYPE_STATIC;
+    info.PSODesc.ResourceLayout.Variables = variables;
+    info.PSODesc.ResourceLayout.NumVariables = textures;
+    device_->CreateGraphicsPipelineState(info, &entry.state);
+    if (!entry.state) {
+        return nullptr;
+    }
+    for (const SHADER_TYPE stage : {SHADER_TYPE_VERTEX, SHADER_TYPE_PIXEL}) {
+        if (IShaderResourceVariable* constants = entry.state->GetStaticVariableByName(stage, "Constants")) {
+            constants->Set(constantBuffer_);
+        }
+    }
+    entry.state->CreateShaderResourceBinding(&entry.bindings, true);
+    if (textures >= 1) {
+        entry.texture0 = entry.bindings->GetVariableByName(SHADER_TYPE_PIXEL, "g_Texture0");
+    }
+    if (textures >= 2) {
+        entry.texture1 = entry.bindings->GetVariableByName(SHADER_TYPE_PIXEL, "g_Texture1");
+    }
+    return &entry;
+}
+
+} // namespace t5de

renderer/scene.hpp

@@ -0,0 +1,128 @@
+#pragma once
+// Carries out the command stream (include/t5de/protocol/commands.hpp) with Diligent, reproducing IMVU's Direct3D 9
+// look (shaders/fixed_function.hlsl): the client's textures and buffers by id, its state as it sets it, a pipeline
+// for each combination of states it draws with.
+
+#include <cstddef>
+#include <cstdint>
+#include <span>
+#include <string>
+#include <unordered_map>
+
+#include "EngineFactoryD3D11.h"
+#include "RefCntAutoPtr.hpp"
+#include "t5de/protocol/commands.hpp"
+
+namespace t5de {
+
+class Scene {
+public:
+    Scene(Diligent::IRenderDevice* device, Diligent::IDeviceContext* context, Diligent::ISwapChain* swapChain);
+
+    // Carries out one frame's commands. offsetX is where the window starts within the client's 3D view; without draw,
+    // only the frame's changes are made (textures, buffers, state). True if the frame was drawn into the window (to
+    // present).
+    bool execute(std::span<const std::byte> commands, int offsetX, bool draw);
+
+    // Replaces the shaders' source (the built-in one by default): for trying changes while the client runs.
+    void setShaderSource(std::string source);
+
+    // When the adapter sent the last frame carried out (QueryPerformanceCounter).
+    [[nodiscard]] std::uint64_t lastSentAt() const { return lastSentAt_; }
+
+private:
+    struct Texture {
+        Diligent::RefCntAutoPtr<Diligent::ITexture> texture;
+        std::uint32_t width = 0;
+        std::uint32_t height = 0;
+        std::uint32_t flags = 0;
+    };
+
+    struct Buffer {
+        Diligent::RefCntAutoPtr<Diligent::IBuffer> buffer;
+        protocol::BufferKind kind = protocol::BufferKind::Vertex;
+        std::uint32_t elementSize = 0;
+        std::uint32_t count = 0;
+    };
+
+    struct Pipeline {
+        Diligent::RefCntAutoPtr<Diligent::IPipelineState> state;
+        Diligent::RefCntAutoPtr<Diligent::IShaderResourceBinding> bindings;
+        Diligent::IShaderResourceVariable* texture0 = nullptr;
+        Diligent::IShaderResourceVariable* texture1 = nullptr;
+    };
+
+    struct Shaders {
+        Diligent::RefCntAutoPtr<Diligent::IShader> vertex;
+        Diligent::RefCntAutoPtr<Diligent::IShader> pixel;
+    };
+
+    // Mirrors the shader's constant buffer.
+    struct Constants {
+        float world[16];
+        float view[16];
+        float projection[16];
+        float texTransform0[8];
+        float texTransform1[8];
+        float ambient[4];
+        float lightColor[8][4];
+        float lightVector[8][4];
+        float fogColor[4];
+        float fog[4];
+        float flags[4];
+        float viewport[4];
+    };
+
+    // What a draw needs from the state, as set by the client.
+    struct State {
+        bool depthTest = true;
+        bool depthWrite = true;
+        bool blending = false;
+        std::uint32_t blendMode = 1;
+        bool culling = true;
+        bool normalInverting = false;
+        std::uint32_t texture0 = 0;
+        std::uint32_t texture1 = 0;
+    };
+
+    void createTexture(const protocol::Command& command);
+    void updateTexture(const protocol::Command& command);
+    void createBuffer(const protocol::Command& command);
+    void updateBuffer(const protocol::Command& command);
+    void setProjection(const protocol::SetProjection& projection);
+    void setState(const protocol::SetState& state);
+    void drawTriangles(const protocol::DrawTriangles& draw);
+    void drawArrays(const protocol::Command& command);
+    void drawPointSprites(const protocol::Command& command);
+
+    // Draws indexCount indices from index, with vertex streams (positions required, the others may be null).
+    void draw(Diligent::IBuffer* indices, std::uint32_t indexCount, Diligent::IBuffer* positions,
+              Diligent::IBuffer* normals, Diligent::IBuffer* texCoords, Diligent::IBuffer* colors);
+    // The pipeline for the current state, with these vertex streams and textures, for triangles or particles.
+    Pipeline* pipeline(std::uint32_t streams, std::uint32_t textures, bool sprites);
+    const Shaders& shaders(std::uint32_t streams, std::uint32_t textures, bool sprites);
+    // Picks the pipeline for the current state and sets it up for a draw: constants, textures. Null if it can't.
+    Pipeline* prepare(std::uint32_t streams, bool sprites);
+    Diligent::ITextureView* textureView(std::uint32_t id);
+
+    Diligent::RefCntAutoPtr<Diligent::IRenderDevice> device_;
+    Diligent::RefCntAutoPtr<Diligent::IDeviceContext> context_;
+    Diligent::RefCntAutoPtr<Diligent::ISwapChain> swapChain_;
+    Diligent::RefCntAutoPtr<Diligent::IBuffer> constantBuffer_;
+    Diligent::RefCntAutoPtr<Diligent::ISampler> samplers_[4]; // by wrap u (1) and wrap v (2)
+
+    std::unordered_map<std::uint32_t, Texture> textures_;
+    std::unordered_map<std::uint32_t, Buffer> buffers_;
+    std::unordered_map<std::uint32_t, Shaders> shaders_;
+    std::unordered_map<std::uint32_t, Pipeline> pipelines_;
+    std::string shaderSource_;
+
+    Constants constants_{};
+    State state_;
+    bool onscreen_ = false; // in a frame drawn into the window
+    bool draw_ = true;
+    std::uint64_t lastSentAt_ = 0;
+    int offsetX_ = 0;
+};
+
+} // namespace t5de

renderer/shaders/embed.hpp.in

@@ -0,0 +1,4 @@
+#pragma once
+// Generated by CMake from renderer/shaders/@T5DE_SHADER_FILE@: the shader source, compiled at run time by Diligent.
+
+inline constexpr const char* @T5DE_SHADER_NAME@ = R"hlsl(@T5DE_SHADER_SOURCE@)hlsl";

renderer/shaders/fixed_function.hlsl

@@ -0,0 +1,146 @@
+// IMVU's Direct3D 9 fixed-function look, as its renderer sets it up (docs/interface.md):
+// - per-vertex (Gouraud) lighting, with the vertex color as diffuse and ambient material (white without one): the
+//   environment's ambient light, plus each light's color times max(0, N.L); no specular or emissive light;
+// - textures: none (the lit color), one (times the lit color), or two (the second blended over the first by its
+//   alpha, then times the lit color), both on the vertex's texture coordinates with their own transforms;
+// - an alpha test passing pixels whose alpha is greater than the reference, then linear fog by view depth (checked
+//   against the client: fogging by depth buffer value instead looks nothing like it).
+//
+// Particles (vs_sprite) are Direct3D 9 point sprites there: camera-facing squares, viewport height / distance pixels
+// across (1 to 64), textured over the whole square; here each is drawn as an instance of a 4-vertex strip.
+//
+// Compiled for each combination of the vertex streams present (HAS_NORMALS, HAS_TEX_COORDS, HAS_COLORS) and the
+// number of textures (TEXTURES).
+
+cbuffer Constants
+{
+    row_major float4x4 g_World;
+    row_major float4x4 g_View;
+    row_major float4x4 g_Projection;
+    float4 g_TexTransform0[2]; // u' = dot((u, v, 1), [0].xyz), v' = dot((u, v, 1), [1].xyz)
+    float4 g_TexTransform1[2];
+    float4 g_Ambient;          // rgb
+    float4 g_LightColor[8];    // rgb, a: enabled
+    float4 g_LightVector[8];   // xyz: position, or direction toward the light; w: 1 if directional
+    float4 g_FogColor;         // rgb
+    float4 g_Fog;              // x: start, y: end, z: on
+    float4 g_Flags;            // x: lighting on, y: alpha test on, z: alpha reference (0-255)
+    float4 g_Viewport;         // xy: size in pixels
+};
+
+struct VSInput
+{
+    float3 Position : ATTRIB0;
+#if HAS_NORMALS
+    float3 Normal : ATTRIB1;
+#endif
+#if HAS_TEX_COORDS
+    float2 TexCoord : ATTRIB2;
+#endif
+#if HAS_COLORS
+    float4 Color : ATTRIB3; // bytes B, G, R, A
+#endif
+};
+
+struct PSInput
+{
+    float4 Position : SV_POSITION;
+    float4 Color : COLOR0;
+    float2 TexCoord0 : TEXCOORD0;
+    float2 TexCoord1 : TEXCOORD1;
+    float Depth : TEXCOORD2;
+};
+
+void vs_main(in VSInput input, out PSInput output)
+{
+    float4 world = mul(float4(input.Position, 1.0), g_World);
+    float4 view = mul(world, g_View);
+    output.Position = mul(view, g_Projection);
+    output.Depth = abs(view.z);
+
+    float4 material = float4(1.0, 1.0, 1.0, 1.0);
+#if HAS_COLORS
+    material = input.Color.zyxw;
+#endif
+    float4 color = material;
+    if (g_Flags.x > 0.5)
+    {
+        float3 normal = float3(0.0, 0.0, 0.0);
+#if HAS_NORMALS
+        normal = normalize(mul(input.Normal, (float3x3)g_World));
+#endif
+        float3 lit = g_Ambient.rgb * material.rgb;
+        for (int i = 0; i < 8; ++i)
+        {
+            if (g_LightColor[i].a > 0.5)
+            {
+                float3 toLight = g_LightVector[i].w > 0.5 ? g_LightVector[i].xyz : g_LightVector[i].xyz - world.xyz;
+                lit += g_LightColor[i].rgb * material.rgb * max(0.0, dot(normal, normalize(toLight)));
+            }
+        }
+        color = float4(saturate(lit), material.a);
+    }
+    output.Color = color;
+
+    float3 uv = float3(0.0, 0.0, 1.0);
+#if HAS_TEX_COORDS
+    uv.xy = input.TexCoord;
+#endif
+    output.TexCoord0 = float2(dot(uv, g_TexTransform0[0].xyz), dot(uv, g_TexTransform0[1].xyz));
+    output.TexCoord1 = float2(dot(uv, g_TexTransform1[0].xyz), dot(uv, g_TexTransform1[1].xyz));
+}
+
+struct SpriteInput
+{
+    float3 Position : ATTRIB0;
+    float4 Color : ATTRIB3; // bytes B, G, R, A
+    uint Corner : SV_VertexID;
+};
+
+void vs_sprite(in SpriteInput input, out PSInput output)
+{
+    float4 view = mul(mul(float4(input.Position, 1.0), g_World), g_View);
+    float4 clip = mul(view, g_Projection);
+    float size = clamp(g_Viewport.y / max(length(view.xyz), 1e-4), 1.0, 2048.0);
+    float2 corner = float2((input.Corner & 1) != 0 ? 1.0 : -1.0, (input.Corner & 2) != 0 ? -1.0 : 1.0);
+    clip.xy += corner * size / g_Viewport.xy * clip.w; // half the size each way, in clip space
+    output.Position = clip;
+    output.Depth = abs(view.z);
+
+    // No normals: lit by the ambient light alone.
+    float4 material = input.Color.zyxw;
+    output.Color = g_Flags.x > 0.5 ? float4(saturate(g_Ambient.rgb * material.rgb), material.a) : material;
+    output.TexCoord0 = float2(corner.x * 0.5 + 0.5, 0.5 - corner.y * 0.5);
+    output.TexCoord1 = output.TexCoord0;
+}
+
+#if TEXTURES >= 1
+Texture2D g_Texture0;
+SamplerState g_Texture0_sampler;
+#endif
+#if TEXTURES >= 2
+Texture2D g_Texture1;
+SamplerState g_Texture1_sampler;
+#endif
+
+float4 ps_main(in PSInput input) : SV_Target
+{
+    float4 color = input.Color;
+#if TEXTURES == 1
+    color *= g_Texture0.Sample(g_Texture0_sampler, input.TexCoord0);
+#elif TEXTURES == 2
+    float4 first = g_Texture0.Sample(g_Texture0_sampler, input.TexCoord0);
+    float4 second = g_Texture1.Sample(g_Texture1_sampler, input.TexCoord1);
+    color *= lerp(first, second, second.a);
+#endif
+    if (g_Flags.y > 0.5 && !(color.a * 255.0 > g_Flags.z))
+    {
+        discard;
+    }
+    if (g_Fog.z > 0.5)
+    {
+        float visible = saturate((g_Fog.y - input.Depth) / max(g_Fog.y - g_Fog.x, 1e-6));
+        color.rgb = lerp(g_FogColor.rgb, color.rgb, visible);
+    }
+    return color;
+}

tests/CMakeLists.txt

@@ -1,3 +1,7 @@
 add_executable(shared_ptr_test shared_ptr_test.cpp)
 target_link_libraries(shared_ptr_test PRIVATE t5de::imvu t5de::options)
 add_test(NAME shared_ptr COMMAND shared_ptr_test)
+
+add_executable(protocol_test protocol_test.cpp)
+target_link_libraries(protocol_test PRIVATE t5de::protocol t5de::options)
+add_test(NAME protocol COMMAND protocol_test)

tests/protocol_test.cpp

@@ -0,0 +1,61 @@
+// The command stream: what the adapter writes, the renderer process reads back the same.
+
+#include <cstdio>
+#include <cstring>
+
+#include "t5de/protocol/commands.hpp"
+
+using namespace t5de::protocol;
+
+namespace {
+
+int failures = 0;
+
+void check(bool ok, const char* what)
+{
+    if (!ok) {
+        std::printf("FAIL: %s\n", what);
+        ++failures;
+    }
+}
+
+} // namespace
+
+int main()
+{
+    Writer writer;
+    writer.add(Op::Clear, Clear{3, 0xff131313});
+    const unsigned char pixels[] = {1, 2, 3, 4, 5, 6, 7, 8};
+    std::memcpy(writer.add(Op::CreateTexture, CreateTexture{7, 2, 1, 6}, sizeof(pixels)), pixels, sizeof(pixels));
+    std::byte* odd = writer.add(Op::UpdateBuffer, UpdateBuffer{9, 1, 3}, 3); // padded to 4
+    std::memset(odd, 0xab, 3);
+    writer.add(Op::EndFrame, EndFrame{});
+    check(writer.size() % 4 == 0, "commands stay 4-byte aligned");
+
+    Reader reader(writer.bytes());
+    auto clear = reader.next();
+    check(clear && clear->op() == Op::Clear && clear->body<Clear>()->color == 0xff131313, "clear");
+    auto texture = reader.next();
+    check(texture && texture->op() == Op::CreateTexture && texture->body<CreateTexture>()->id == 7, "texture");
+    check(texture && texture->payload<CreateTexture>().size() == sizeof(pixels)
+              && std::memcmp(texture->payload<CreateTexture>().data(), pixels, sizeof(pixels)) == 0,
+          "texture pixels");
+    auto update = reader.next();
+    check(update && update->payload<UpdateBuffer>().size() == 4, "payload padded");
+    auto end = reader.next();
+    check(end && end->op() == Op::EndFrame, "end of frame");
+    check(!reader.next(), "nothing after");
+
+    // A truncated stream (a chunk cut short) yields only whole commands.
+    Reader truncated(writer.bytes().first(writer.size() - 2));
+    int count = 0;
+    while (truncated.next()) {
+        ++count;
+    }
+    check(count == 3, "truncated: whole commands only");
+
+    if (failures == 0) {
+        std::printf("ok\n");
+    }
+    return failures == 0 ? 0 : 1;
+}