CUBLAS GEMM Example

A CUBLAS Matrix Multiply (GEMM) example.

Requirements:

- Nvidia GPU supporting CUDA
- CUDA v11.0 or greater
- CUBLAS v11.0 (should come with CUDA)
- openblas (max-perf CPU test)

Install and Compile:

a) Clone Repo:
    git clone https://github.com/temporal-hpc/cublas-gemm

b) Compile:
    cd cublas-gemm
    make

c) Data Types:
    You can specify the data type (half, float) for each matrix
    Example:
    make ATYPE=half BTYPE=half CTYPE=hal

Run:

a) Run:
    run as ./prog dev nt n comptype mode
    dev:      Device ID
    nt:       Number of CPU threads (accelerates data init and CPU mode)
    n:        Matrix size of n x n
    comptype: GPU CUBLAS mode
    mode:     CPU=0,  GPU=1

b) CUBLAS Compute Types:
        0  = CUBLAS_COMPUTE_16F
        1  = CUBLAS_COMPUTE_16F_PEDANTIC
        2  = CUBLAS_COMPUTE_32F
        3  = CUBLAS_COMPUTE_32F_PEDANTIC
        4  = CUBLAS_COMPUTE_32F_FAST_16F
        5  = CUBLAS_COMPUTE_32F_FAST_16BF
        6  = CUBLAS_COMPUTE_32F_FAST_TF32
        7  = CUBLAS_COMPUTE_64F
        8  = CUBLAS_COMPUTE_64F_PEDANTIC
        9  = CUBLAS_COMPUTE_32I
        10 = CUBLAS_COMPUTE_32I_PEDANTIC

Example executions:

a) [GPU CUBLAS] Default CUBLAS math (FP32 CUDA cores)
    make ATYPE=float BTYPE=float CTYPE=float
    ./prog 0 4 $((2**13)) 2 1

b) [GPU CUBLAS] Tensor Cores with mixed precision
    make ATYPE=half BTYPE=half CTYPE=float
    ./prog 0 4 $((2**13)) 4 1

c) [GPU CUBLAS] Tensor Cores with FP16
    make ATYPE=half BTYPE=half CTYPE=half
    ./prog 0 4 $((2**13)) 0 1

d) [CPU CBLAS] FP32 Using 8 CPU threads 
    make
    ./prog 0 8 $((2**13)) 0 0

e) [CPU CBLAS] FP64 Using 8 CPU threads 
    make CPUFP64=CPUFP64
    ./prog 0 8 $((2**13)) 0 0