Matrix multiplication for tensors

Matrix multiplication for tensors

Usage

tensor_matmul(x, y, precision = "standard")

Arguments

x, y

Two-dimensional cudatensor objects or numeric matrices.

precision

Compute policy: "standard" preserves the existing dtype and backend arithmetic. "tf32" explicitly permits reduced-precision TensorFloat-32 multiplication with float32 output on the native CUDA backend. It requires two native CUDA float32 tensors and an NVIDIA GPU with compute capability 8.0 or newer.

Value

A cudatensor.

Details

Row names come from x and column names come from y. When both operands name the contracted dimension, those names must be identical.

Standard native float32 multiplication uses cuBLAS SGEMM; float64 uses DGEMM. Float32 inputs are never automatically promoted to float64 for accuracy. TF32 permits cuBLAS to use Tensor Cores with reduced input precision; the selected kernel and any speedup depend on the workload. An unsupported TF32 request raises an error, including when NVIDIA_TF32_OVERRIDE=0. The policy applies only to this call, and cuda_provenance() records matmul_standard or matmul_tf32_permitted as its selection reason. %*% uses the standard policy.

Rounding and cancellation can cause float32 dot products to differ from a double-precision reference, especially near zero. TF32 allows larger differences. Use float64 tensors when the analysis requires greater accuracy; assess precision against the needs of the complete workflow.

Examples

x <- cuda_tensor(matrix(1:6, 2, 3), device = "cpu")
y <- cuda_tensor(matrix(1:6, 3, 2), device = "cpu")
tensor_matmul(x, y)
#> <cudatensor[2x2] device=cpu backend=base dtype=float64>
#>      [,1] [,2]
#> [1,]   22   49
#> [2,]   28   64