FSP.DMRCrack v0.3.3

Full Changelog: v0.3.2...v0.3.3

FSP.DMRCrack v0.3.2

Full Changelog: v0.3.1...v0.3.2

FSP.DMRCrack v0.3.1

Full Changelog: v0.3.0...v0.3.1

Full Changelog: v0.3.0...v0.3.1

FSP.DMRCrack v0.3.0

Fixed

• CUDA atomic race on best_key / best_score (#1) — the two-step
  atomicCAS(score) + atomicExch(key) update was replaced with a single atomicMax
  on a packed unsigned long long that encodes both the score (top 24 sortable bits)
  and the 40-bit key. Eliminates the race window where a polling read could observe a
  new score paired with a stale key, which caused the blank best-candidate display
  reported in #4.
• CUDA error auto-fallback to CPU — if the GPU returns an error mid-run, the engine
  now stops the CUDA path, restarts from the same offset on CPU threads, and shows a
  banner in the status strip. Replaces the v0.2.1 behaviour of aborting with
  "NVIDIA CUDA Error o no hay GPUs compatibles."
• GPU/CPU keyspace split inverted — the engine was giving 20 % of the keyspace to
  the GPU and 80 % to the CPU; now it's 80 / 20 as documented, configurable via the
  new GPU % spin control.
• Demodulation silent failure — run_process_stderr_redirect only redirected stderr;
  stdout was set to GetStdHandle(STD_OUTPUT_HANDLE), which returns NULL in a GUI
  process. Cygwin apps (dsd-fme) received a NULL stdout handle and could exit 0
  without producing DSP output. Both streams now go to the log file.
• DSP output search too narrow — added two more candidate paths under the
  dsd-fme.exe directory so captures from non-default installs are found.
• Test/production parity — test_strict_score.c used 31.0f * k as the absolute
  pruning floor while the GPU kernel uses 33.0f * k; the test could pass keys the
  kernel would have rejected. Aligned to 33.0f.
• #undef BCNT_INC typo — BCNT_ADD was leaking out of the strict kernel
  translation unit because the #undef referenced a macro that never existed.
• Documentation drift — CONTRIBUTING.md referenced non-existent files
  (build_test_bin.bat, test_bin_score.exe) and pointed contributors at the wrong
  scoring file. README claimed a CPU fallback that, in practice, takes days; both
  corrected.

Added

• Operator Dashboard GUI — full visual redesign:
  • Header bar with GPU status dot (running / paused / stopped) and the GPU's
    sm_XY · N SMs summary.
  • Metric tiles with hierarchy: THROUGHPUT dominant (Consolas 22 pt, 58 % wide),
    BEST CANDIDATE beside it, PROGRESS as a full-width row with a large bar
    and one-line metadata.
  • Side-by-side line charts (keys/s, best score) with area fill, grid lines, and a
    Z = 400 threshold marker on the score chart.
  • Status strip anchored at the bottom of the window (Win32 convention).
• KPA silence-frame detection — DMR voice frames whose first 24 bits are
  consistent silence are flagged at load time (SILENCE= tag in .bin). The CUDA
  kernel uses these as a pre-filter to reject wrong keys earlier in the search.
  A badge in the GUI shows how many silence frames were detected.
• Configurable GPU/CPU split — spin control (50–95 %, step 5) lets the user
  adjust how the keyspace is divided between the GPU and CPU workers at scan start.
• ILP-2 kernel enabled on sm_75+ — the dual-key kernel that was previously
  restricted to sm_89+ now runs on Turing and Ampere too, improving throughput on
  GTX 16xx / RTX 20xx and RTX 30xx cards.
• Multi-arch build — single dmrcrack.exe ships with native SASS for sm_75 /
  sm_86 / sm_89 plus a compute_75 PTX fallback for RTX 50xx and future GPUs
  (JIT-compiled by the driver on first run).
• .bin validation on load — payload set is inspected on file open: KMI9
  coverage stats (how many lines carry MI metadata), KID histogram, low-payload
  warning when count < 64, alignment warning if the first frame isn't at
  burst_pos == 0.
• Full EN/ES i18n with runtime toggle — every user-facing string routed through
  the Lang struct, including engine startup errors emitted by bruteforce.cu.
  The language toggle button in the header switches g_lang_ptr and persists the
  choice to dmrcrack.ini. The help dialog is also fully localized.
• Help dialog — quick-start guide accessible from the Help button: capture
  with SDR# → demodulate → brute-force, with NFM / de-emphasis caveats.
• Copy-key button — copies the current best-candidate key to the clipboard.
• Key-found notification — taskbar flash + system beep when the search completes
  with a candidate above the Z threshold.
• Resizable window with per-monitor DPI awareness — minimum 980 × 720; all
  controls and graphs reposition via WM_SIZE.
• WinSparkle auto-updater — replaced the custom WinHTTP + GitHub-API updater
  (no signature verification) with WinSparkle 0.9.2: built-in UI, EdDSA (Ed25519)
  signature verification, Sparkle appcast support. Each release now publishes an
  appcast.xml alongside the installer.
• gpu_keys_tested counter — engine snapshot now exposes GPU and CPU
  contributions separately, used by the THROUGHPUT tile to display
  "GPU 1.6 G/s · CPU 240 M/s".
• ILP test bench — build_bench.bat builds bin/test_bench.exe, a standalone
  micro-benchmark for the strict vs ILP-2 kernels.

Changed

• Capture module simplified — WAV input + Demodulate + Export kept;
  RTL-SDR mode, slot selector (Both / 1 / 2), inverted-polarity checkbox and the
  WAV/RTL mode toggle removed. DSD-FME is invoked with -V 3 (both slots) by
  default. Users who need slot-specific or polarity-inverted demodulation can run
  dsd-fme manually and load the resulting .bin.
• All English source code — Spanish comments and identifiers translated to
  English across bruteforce.cu, bruteforce.c, bruteforce.h,
  test_score_windows.c and the DSP-converter helper script.
• build.bat rewrite — auto-detects Visual Studio via vswhere, sets
  INCLUDE / PATH for the CUDA toolkit, compiles with -O3 /arch:AVX2.

Internal

• 13 new GitHub issues filed for community contributors (good first issue,
  help wanted): magic-number cleanup, __restrict__ annotations, CMake build,
  POSIX HAL, headless CLI, KPA pre-filter optimisation, legacy-kernel test
  coverage, and more.

⚠️ The releases v0.1.2, v0.2.0 and v0.2.1 were tagged without updating this
changelog. Their content is rolled forward into the 0.3.0 entry above; the
individual tags remain on GitHub for reference.

FSP.DMRCrack v0.2.1

Full Changelog: v0.2.0...v0.2.1

FSP.DMRCrack v0.2.0

Full Changelog: v0.1.2...v0.2.0

Appcast

WinSparkle auto-update feed — do not delete

v0.1.2

What's new

• Inline error diagnostics: demodulation errors now show a MessageBox with the dsd-fme log output, so users can see exactly what happened without opening log files manually
• Improved dsd-fme output search: searches both DSP/ subdirectory and flat locations for DSP output
• Better error messages: all error states include actionable context

Download

Download FSP.DMRCrack-0.1.2-Setup.exe below and run the installer.

FSP.DMRCrack v0.1.0

FSP.DMRCrack v0.1.0

GPU-accelerated ARC4 key recovery for DMR Enhanced Privacy voice communications.

Download

FSP.DMRCrack-0.1.0-Setup.exe — Windows installer (includes DSD-FME and all required Cygwin runtime DLLs)

Requirements

• Windows 10/11 x64
• NVIDIA GPU (GTX 16xx / RTX 20xx or newer, sm_75+)
• No additional installation required — everything is bundled

What's included

• dmrcrack.exe — main application (Win32 GUI, CUDA brute-force engine)
• tools/dsd-fme.exe — DMR demodulator backend
• tools/cyg*.dll — Cygwin runtime libraries required by dsd-fme
• tools/dsdfme_dsp_to_bin.py — DSP output converter
• tools/verify_decrypt.py — decryption validation tool