All examples, including more advanced ones, are shipped within cuBLASDx package.
This folder demonstrates cuBLASDx APIs usage.
- cuBLASDx/MathDx package
- See cuBLASDx requirements
- CMake 3.18 or newer
- Linux system with installed NVIDIA drivers
- NVIDIA GPU of Volta (SM70) or newer architecture
- You may specify
CUBLASDX_CUDA_ARCHITECTURESto limit CUDA architectures used for compilation (see CMake:CUDA_ARCHITECTURES) mathdx_ROOT- path to mathDx package (XX.Y - version of the package)
mkdir build && cd build
cmake -DCUBLASDX_CUDA_ARCHITECTURES=70-real -Dmathdx_ROOT=/opt/nvidia/mathdx/XX.Y ..
make
# Run
ctest
For the detailed descriptions of the examples please visit Examples section of the cuBLASDx documentation.
| Group | Subgroup | Example | Description |
|---|---|---|---|
| Introduction | GEMM | 01_introduction_example | Walks through descriptor creation and the shared-memory and register-result GEMM APIs. |
| Pipeline | 01_introduction_pipeline | Introduces the host pipeline object, device handle, tile pipeline, and epilogue flow. | |
| Simple GEMM | Basic | 02_simple_gemm_fp32 | Performs a checked fp32 GEMM and is the smallest general-purpose starting point. |
| Precision | 02_simple_gemm_mixed_precision | Uses different precisions and storage types for matrices A, B, and C. | |
| Precision | 02_simple_gemm_int8_int8_int32 | Performs integral GEMM using int8 inputs and int32 accumulation. | |
| Precision | 02_simple_gemm_fp8 | Performs fp8 GEMM on architectures that support the required fp8 operations. | |
| Complex | 03_simple_gemm_cfp16 | Performs complex half-precision GEMM. | |
| Complex | 03_simple_gemm_std_complex_fp32 | Uses cuda::std::complex as the matrix element type. | |
| Layout | 06_simple_gemm_leading_dimensions | Demonstrates non-default leading dimensions and padded matrix storage. | |
| Layout | 09_simple_gemm_custom_layout | Uses custom CuTe layouts for shared-memory matrices. | |
| Layout | 09_simple_gemm_aat | Computes C = A * A^T while reusing one shared-memory allocation for both views of A. | |
| Transform | 07_simple_gemm_transform | Applies element-wise load and store transform operators around GEMM. | |
| Transform | 18_gemm_conj_transpose | Applies tensor views such as conjugate transpose to GEMM inputs without materializing a copy. | |
| Register I/O | 08_simple_gemm_fp32_decoupled | Uses lower-precision input/output storage with higher-precision computation and register fragments. | |
| Runtime Compilation | NVRTC | 15_nvrtc_gemm | Compiles a GEMM kernel at runtime with NVRTC and passes cuBLASDx headers to device code. |
| NVRTC | 15_nvrtc_trsm | Compiles a block-level TRSM kernel at runtime with NVRTC. Requires cuBLASDx fatbin/LTO support. | |
| Performance | Block GEMM | 10_single_gemm_performance | Benchmarks a single block-level GEMM tile configuration. |
| Device GEMM | 11_device_gemm_performance | Builds a full-device GEMM from cuBLASDx tiles and compares with a reference path. | |
| Fusion | 14_fused_gemm_performance | Benchmarks two fused GEMMs against an unfused reference path. | |
| Advanced GEMM | Batching | 19_batched_gemm | Demonstrates rank-3 tensor batching for non-pipelined GEMM. Requires CUDA Toolkit 13.1 or newer. |
| Batching | 19_batched_gemm_pipeline | Demonstrates rank-3 tensor batching with the pipeline API. Requires CUDA Toolkit 13.1 or newer. | |
| Batching | 05_batched_gemm_fp64 | Shows manual batching inside one CUDA block with BlockDim. | |
| Block shape | 04_blockdim_gemm_fp16 | Shows how launch dimensions interact with BlockDim and participating threads. | |
| Accuracy | 12_gemm_device_partial_sums | Offloads partial accumulation to a higher-precision register array. | |
| Fusion | 14_gemm_fusion | Performs two dependent GEMMs in one CUDA kernel. | |
| cuFFTDx fusion | 13_gemm_fft | Fuses GEMM and FFT in one kernel. Requires cuFFTDx. | |
| cuFFTDx fusion | 13_gemm_fft_fp16 | Fuses half-precision complex GEMM and FFT. Requires cuFFTDx. | |
| cuFFTDx fusion | 13_gemm_fft_performance | Benchmarks GEMM and FFT fusion. Requires cuFFTDx. | |
| TRSM | Block | 17_trsm_block | Solves triangular systems cooperatively in a CUDA block using shared memory. |
| Thread | 17_trsm_thread | Solves many small triangular systems independently, one per CUDA thread. | |
| Tensor views | 18_trsm_conj_transpose | Uses conj_transpose_view for TRSM without copying or transposing matrix data. | |
| Emulation | Ozaki | 16_dgemm_emulation | Emulates double-precision GEMM using lower-precision GEMM operations. Requires CUDA Toolkit 13.1 or newer. |
| Pipeline | 20_dgemm_emulation_pipeline | Runs pipelined double-precision emulation controlled by required mantissa bits. Requires CUDA Toolkit 13.1 or newer. | |
| Pipeline | 21_sgemm_emulation_pipeline | Runs pipelined single-precision emulation using the same mantissa-bit mechanism. Requires CUDA Toolkit 13.1 or newer. | |
| Cluster | Persistent | 22_persistent_cluster_gemm | Runs a persistent GEMM kernel launched with thread block clusters (BlocksPerCluster) and rasterized tile scheduling. |