Docs  /  Visual Studio build configurations

Visual Studio build configurations

Project: _AugmentedIntelligence.vcxproj · toolset v145 · platform x64 (primary).

Configurations

Configuration GPU/CUDA TensorRT TF32 SIMD Llama build dir Preprocessor highlights
Debug | x64 Toolkit present no no* default default Dev
Release | x64 CUDA linked no no* default llama_build USE_CUDA, AI_MATHNN_LEARN, AI_REALITY_CONSTRUCT
Release-GPU-AVX512 | x64 CUDA no yes AVX-512 llama_build USE_CUDA, AI_USE_TF32, AI_MATHNN_LARGE
Release-GPU-AVX2 | x64 CUDA no yes AVX2 llama_build_cuda_avx2 USE_CUDA, AI_USE_TF32, AI_MATHNN_LARGE
Release-GPU-AVX512-Tensor | x64 CUDA yes yes AVX-512 llama_build above + AI_USE_TENSORRT, AI_TENSOR_ACCEL, AI_USE_TF32
Release-GPU-AVX2-Tensor | x64 CUDA yes yes AVX2 llama_build_cuda_avx2 above + AI_USE_TENSORRT, AI_TENSOR_ACCEL, AI_USE_TF32
Release-CPU-AVX2 | x64 No CUDA libs no no AVX2 llama_build_cpu_avx2 AI_CPU_ONLY
Release-CPU-NoAVX | x64 No GPU no no NoExtensions llama_build_cpu_noavx AI_CPU_ONLY

* TF32 can be force-enabled on any config with msbuild /p:EnableTf32=true.

TF32 (optional, configuration-managed)

TensorFloat-32 cuBLAS path is compile-optional via preprocessor AI_USE_TF32.

Layer Behavior
VS Configuration Manager GPU AVX* and *-Tensor configs set property EnableTf32=true by default
Preprocessor EnableTf32=true → injects AI_USE_TF32 into ClCompile
Runtime Still defaults to FP32; enable with cutlass set precision tf32
CPU / plain Release No TF32 unless you pass /p:EnableTf32=true
Disable on a GPU config msbuild /p:EnableTf32=false or unset in a custom PropertyGroup 
# Ship GPU build with TF32 available at runtime
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX2 -Platform x64

# Tensor stack (TRT + TF32 compile flag)
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX2-Tensor -Platform x64

# Deterministic FP32-only GPU (no AI_USE_TF32)
msbuild _AugmentedIntelligence.vcxproj /p:Configuration=Release-GPU-AVX2 /p:EnableTf32=false

# Opt-in TF32 on plain Release
msbuild _AugmentedIntelligence.vcxproj /p:Configuration=Release /p:EnableTf32=true

In-app:

cutlass status          # shows tf32_compiled= / tf32_available=
cutlass set precision tf32   # no-op → fp32 if not compiled
cutlass check 64        # TF32 probe only if AI_USE_TF32

See TF32MatrixMultiply.md · Cutlass.md.

Tensor configs (AVX2 Tensor / AVX512 Tensor)

These ship NVIDIA TensorRT inference (TensorRT.cpp) on top of the matching GPU+SIMD baseline, and also enable TF32 compile support.

  1. Install TensorRT 8.x+ for your CUDA version.
  2. Either: - Copy/extract under third_party/tensorrt/ with include/ and lib/, or - Build with /p:TensorRtDir=C:\Path\To\TensorRT
  3. Select configuration in VS or:
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX2-Tensor -Platform x64
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX512-Tensor -Platform x64

Linked: nvinfer.lib, nvonnxparser.lib, nvinfer_plugin.lib.
Post-build copies nvinfer*.dll when present under $(TensorRtDir)\lib or \bin.

Without TensorRT installed, non-Tensor GPU configs still build; Tensor configs will fail at link until TensorRtDir is valid.

Compile-type notes

  • UseCuda — GPU configs true; CPU false.
  • USE_CUDA — RealityConstruct hybrid, MathNNDevice GPU, RemoteCuda, cuBLAS FP32 GEMM.
  • AI_USE_TF32 — optional TF32 Tensor Core math in CutlassAccel (see above).
  • AI_USE_TENSORRT / AI_TENSOR_ACCEL — TensorRT OD path enabled at compile time.
  • AI_CPU_ONLY — no NVIDIA driver required.
  • AI_MATHNN_LEARN, AI_REALITY_CONSTRUCT, AI_MATHNN_LARGE — feature markers.

Build scripts

.\scripts\build_v145.ps1
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX2 -Platform x64
.\scripts\build_v145.ps1 -Configuration Release-GPU-AVX2-Tensor -Platform x64
.\scripts\build_v145.ps1 -Configuration Release-CPU-AVX2 -Platform x64
Machine Configuration TF32
NVIDIA Ampere+ + modern CPU + TensorRT Release-GPU-AVX2-Tensor or AVX512-Tensor compiled; opt-in runtime
NVIDIA + modern CPU (no TRT) Release-GPU-AVX2 or Release-GPU-AVX512 compiled; opt-in runtime
Deterministic GPU numerics GPU config + /p:EnableTf32=false off
No GPU laptop Release-CPU-AVX2 off
Day-to-day dev Release x64 off by default
Generated from the project markdown docs on 2026-07-24. This is a static, self-contained site.