Skip to content

README.md

Latest commit

 

History

History
70 lines (53 loc) · 2.93 KB

README.md

File metadata and controls

70 lines (53 loc) · 2.93 KB

cuBLAS Examples

19 examples covering BLAS Level 1, 2, and 3 operations — from scalar vector ops to batched mixed-precision matrix multiply. Enable with -Dcublas=true.

Build & Run

zig build run-cublas-<name> -Dcublas=true

# Examples:
zig build run-cublas-gemm -Dcublas=true
zig build run-cublas-axpy -Dcublas=true

Level 1 — Vector–Vector Operations

Example File BLAS Op Description
axpy axpy.zig SAXPY y = α·x + y
dot dot.zig SDOT Dot product: Σ xᵢyᵢ
nrm2_asum nrm2_asum.zig SNRM2, SASUM L2 norm and L1 norm
scal scal.zig SSCAL Vector scaling: x = α·x
amax_amin amax_amin.zig ISAMAX, ISAMIN Index of max/min absolute value
swap_copy swap_copy.zig SSWAP, SCOPY Vector swap and copy
rot rot.zig SROT, SROTG Givens plane rotation
cosine_similarity cosine_similarity.zig SDOT, SNRM2 Cosine similarity via L1 ops

Level 2 — Matrix–Vector Operations

Example File BLAS Op Description
gemv gemv.zig SGEMV y = α·A·x + β·y (matrix-vector)
symv_syr symv_syr.zig SSYMV, SSYR Symmetric matrix-vector ops
trmv_trsv trmv_trsv.zig STRMV, STRSV Triangular multiply and solve

Level 3 — Matrix–Matrix Operations

Example File BLAS Op Description
gemm gemm.zig SGEMM C = α·A·B + β·C (single precision)
gemm_batched gemm_batched.zig SGEMMStridedBatched Strided batched GEMM
gemm_ex gemm_ex.zig cublasGemmEx Mixed-precision GEMM (FP16 in, FP32 out)
symm symm.zig SSYMM Symmetric matrix multiply: C = α·A·B + β·C
trsm trsm.zig STRSM Triangular solve: op(A)·X = α·B
syrk syrk.zig SSYRK Symmetric rank-k update: C = α·A·Aᵀ + β·C
geam geam.zig SGEAM Matrix add / scale / transpose
dgmm dgmm.zig SDGMM Diagonal matrix multiply

Row-major Note

cuBLAS uses column-major storage. For row-major buffers (Zig/C default), swap A↔B and M↔N in sgemm:

// ✅ Correct for row-major C = A × B (M×N = M×K × K×N):
try blas.sgemm(.no_transpose, .no_transpose,
    n, m, k,          // ← swap n and m
    alpha,
    b, n,             // ← B first, leading dim n
    a, k,             // ← A second, leading dim k
    beta, c, n);

See docs/cublas/README.md for the full explanation.

→ Full API reference: docs/cublas/README.md