LAME 4.0 cVBRb GPU (CUDA) for Windows x64
================================================

Package name:

  lame-4.0-cVBRb-20260806-CUDA-VS-x64.zip

Release type:

  Specialist NVIDIA CUDA build

About this build:

  This is a 64-bit Windows command-line build of the LAME 4.0 cVBRb branch
  with NVIDIA CUDA acceleration.

  CUDA accelerates selected psychoacoustic and bitrate-decision work. LAME
  continues to perform the final MP3 encoding and bitstream generation.

  Most users should use a standard CPU build. This specialist release is
  intended for systems with a supported NVIDIA CUDA GPU and compatible driver.

Contents:

  lame.exe
      LAME 4.0 cVBRb CUDA command-line encoder and MP3 decoder.

  cufft64_11.dll
      NVIDIA cuFFT runtime required for CUDA encoding.

  libmpg123-0.dll
      mpg123 runtime used by --decode mode.

  README.txt
      Package usage and compatibility information.

  BUILD-INFO.txt
      Toolchain, architecture, dependency, and verification details.

  SOURCE.txt
      Upstream source, fork, licence, and source-offer information.

System requirements:

  Encoding:
    - 64-bit Windows
    - Supported NVIDIA CUDA GPU
    - Compatible NVIDIA driver
    - cufft64_11.dll beside lame.exe

  Decoding:
    - 64-bit Windows
    - libmpg123-0.dll beside lame.exe
    - No NVIDIA GPU is required

Compiled CUDA targets:

  sm_61   Pascal
  sm_75   Turing
  sm_80   Ampere
  sm_86   Ampere
  sm_89   Ada
  sm_120  Blackwell

  The package also contains compute_120 PTX for forward JIT compatibility.

  The build should not be described as supporting every NVIDIA GPU generation.
  GPUs with compute capabilities not listed above have not been included as
  native targets in this package.

Basic encoding usage:

  lame.exe input.wav output.mp3

cVBRb example:

  lame.exe -V0 -b 192 --vbr-min-strict --vbr-max=256 ^
    --bitrate-boost=2 input.wav output.mp3

CUDA authority example:

  lame.exe -V0 -b 192 --vbr-min-strict --vbr-max=256 ^
    --bitrate-boost=2 --cuda-mask=granule-cuda-authority ^
    input.wav output.mp3

Exact Audio Copy:

  The encoder accepts common EAC/LAME switches including:

    -v
    -h
    -p

  Long options that take values accept both forms:

    --bitrate-boost 2
    --bitrate-boost=2

  Select this lame.exe as EAC's external encoder only on a system with a
  supported NVIDIA GPU. Use a standard CPU build for broader compatibility.

Decoding:

  lame.exe --decode input.mp3 output.wav

  Decode mode uses libmpg123, handles Xing/LAME gapless information, writes
  signed 16-bit PCM WAV output, and bypasses CUDA initialization.

User-facing output:

  The encoder provides conventional LAME-style progress, elapsed/estimated
  time, end-to-end play/REAL speed, bitrate distribution, LR/MS statistics,
  and block-type statistics.

Diagnostics:

  Add the following switch to display CUDA engineering counters and internal
  validation information:

    --cuda-diagnostics

Runtime files:

  Keep these files together:

    lame.exe
    cufft64_11.dll
    libmpg123-0.dll

  The full NVIDIA CUDA Toolkit is not required on the target computer. A
  compatible NVIDIA driver is still required for CUDA encoding.

  The CUDA runtime is linked statically.

This package does NOT include:

  - libmp3lame.dll
  - lame_enc.dll
  - libiconv-2.dll
  - development headers
  - import libraries
  - a CPU-only fallback encoder executable

Important limitations:

  - CUDA encoding requires a supported NVIDIA GPU.
  - The active CUDA authority path requires mapped pinned host-memory support.
  - Performance varies by GPU, driver, CPU, PCIe configuration, and Windows
    scheduling behaviour.
  - Native targets are limited to the compute capabilities listed above.
  - Decode mode is CPU-based and does not require CUDA.

About LAME and this fork:

  LAME is an MP3 encoder. The official LAME project distributes source code.

  This specialist binary package is based on LAME 4.0 and includes cVBRb and
  NVIDIA CUDA work maintained by Cole Williams.

  Upstream:
    https://lame.sourceforge.io/

  Project and package information:
    https://www.lamemp3.co.uk/

Licence and source notes:

  LAME is licensed under the GNU LGPL.

  NVIDIA cuFFT and mpg123 remain subject to their own licence and
  redistribution terms.

  See SOURCE.txt for source-code and redistribution information.
