parallel-dc

parallel-dc is a Sega Dreamcast, NAOMI, NAOMI 2 and Atomiswave emulator core for libretro frontends such as RetroArch.

It started life as libretro's own fork of Flycast (itself a fork of reicast) and has since been developed independently. It is actively maintained and is no longer deprecated: much of the emulator has been audited against upstream Flycast and other references, the useful upstream fixes were brought over after being verified, and a great deal has been added or rewritten that upstream does not have.

The same goals apply throughout the code base:

• Accuracy - emulate what the hardware does, not what one game needs.
• Performance - every change has to be at least as fast as what it replaces, and is measured. No allocation, shader compilation or blocking wait in the middle of a game.
• No locks - threads talk through lock-free queues and atomics (libretro-common's retro_atomic), never mutexes or sleeps.
• Portability - plain C/C89 where possible, built for every libretro platform, with libretro-common used for threads, file I/O and archives.

What it runs

• Dreamcast discs: CHD, GDI, CUE/BIN, CDI, and ELF programs, plus M3U playlists for multi-disc games.
• NAOMI, NAOMI 2 and Atomiswave games, as MAME romsets or flat .bin/.dat images. A romset is recognised by its contents, not its file name, so non-MAME sets work too, and a game can be started out of a merged set (a core option picks which one).
• Any of the above inside zip, 7z or rar archives (RAR 2.9 to 5). Large compressed disc tracks are read as they are needed instead of being unpacked into memory first.
• No BIOS required. The Dreamcast and NAOMI BIOS routines are reimplemented (HLE), including the NAOMI test menu. With Use Real BIOS (If Available) on (the default), a real BIOS in the system directory is used when present.

What is different from upstream Flycast

Video

• A renderer for every frontend driver: OpenGL / OpenGL ES, Vulkan, Direct3D 9, Direct3D 11 and Direct3D 12. The core uses whichever the frontend prefers.
• Per-pixel translucency sorting (order-independent transparency) on OpenGL 4.3+, Vulkan, Direct3D 11 and Direct3D 12, chosen with a core option rather than a build flag. Frames with more translucent pixels than the buffers hold are drawn in bands instead of losing pixels.
• Compute renderer (Vulkan, experimental, Rendering Method option): draws the screen tile by tile with compute shaders the way the PowerVR2 does, with exact per-pixel sorting. The rasterizer remains the default.
• TA processing on the GPU (TA Processing option, Vulkan and OpenGL 4.3+): the display lists the game sends are turned into vertices on the GPU, byte-identical to the CPU path.
• SSAA (2x/4x/8x, Vulkan), sized to the video memory the driver reports.
• Texture decoding on the GPU, with textures written straight into an upload ring (no staging copy per texture). Shaders and pipelines are built at start-up and cached on disk, so there are no mid-game compile hitches.
• Render-to-texture kept on the GPU and copied back to emulated video memory only when the game actually reads it; the per-game RTT setting is gone.
• PowerVR features are always on, as on hardware: fog, modifier volumes (shadows, including in the per-triangle renderers) and mipmapping. Force Texture LOD0 replaces the old mipmapping switch.
• Fog, shadows, depth, palettes, texture formats and clipping audited and fixed against the hardware; host-side precision loss in shaders and colour conversion removed. glm replaced by a small C89 matrix header.
• Vulkan memory handled by a lean C89 block heap instead of the Vulkan Memory Allocator, sized from the driver's real memory budget.
• Threaded rendering reworked: one emulated frame per retro_run, no skipped frames, output identical to non-threaded rendering.

CPU and timing

• SH4: instruction fixes from an audit against upstream and the hardware (DIV1, FTRC, FIPR/FTRV and more); every instruction now has a recompiled form; faster x86-64 and ARM64 recompilers (direct block links, loads/stores through the host memory mapping, fewer instructions per access); and an optional cycle-accurate timing model (SH4 Timing: Legacy / Accurate; NAOMI 2 always uses Accurate).
• MMU (Windows CE games): supported by the 32-bit ARM recompiler as well, with translation tables on every host, correct TLB invalidation, write protection and self-modifying-code handling.
• ARM7 recompiler for x86-64 (Sound CPU Recompiler option), with an interpreter fallback.
• Scanline and frame timing fixed so frames come out at the right rate without lost lines.

NAOMI 2

• The ELAN geometry processor is emulated in C89, with SIMD (SSE2/NEON) vertex lighting and its memory mapped for the recompilers.
• 45 NAOMI 2 games in the game table, with input maps; romsets and cartridge sets start without a BIOS.

Sound

• AICA audited: two-pole low-pass filter, 20-bit interpolation, faster channel loop, save states that keep pending notes and CD audio.
• The DSP is always emulated. Each video frame carries exactly the number of samples the hardware produces for it, not fixed-size chunks.

Frontend and I/O

• Lock-free threading throughout: emulation thread, CHD read-ahead, modem / broadband adapter / NAOMI network, save writer, archive decoding. No C++11 threads, <mutex> or <chrono>.
• File I/O goes through the frontend's VFS and libretro-common, with sectors read in place rather than copied and archives decoded in the background.
• Obsolete core options removed (Framerate, Fog Effects, Volume Modifier, Enable DSP, Enable RTT Buffer, the OSD/FPS options and others): the hardware behaviour they switched off is now always emulated.

Building

make -j$(nproc)

Use platform=<name> for other targets (for example platform=win, platform=libnx, platform=emscripten; see the Makefile). The core is built as flycast_libretro, and its options keep the reicast_ prefix so existing configurations still apply.

Tests

tools/ holds the regression tests used during development: a headless runner and test disc (tools/threads), and checks for the SH4 and ARM7 recompilers, AICA/DSP, the NAOMI 2 geometry chip, the tile accelerator on CPU and GPU, the texture decoders on every API, a reference software renderer used as an oracle (tools/rend/refsw), the Vulkan heap, and the CHD and archive readers.

License

GPL version 2; see LICENSE.

All code contributed to this fork is not bound by the Individual Contributor License Agreement of the reicast repository (https://github.com/reicast/reicast-emulator) and shall not be considered a contribution to that repository.