# GST test # # Preconditions: # Set device to all. If you need to run the rvs only on a subset of GPUs, please run rvs with -g # option, collect the GPUs IDs (e.g.: GPU[ 5 - 50599] -> 50599 is the GPU ID) and then specify # all the GPUs IDs separated by white space (e.g.: device: 50599 3245) # Set parallel execution to false # Set matrix_size to 5760 (for Vega 10 cards). For Vega 20, the recommended matrix_size is 8640 # Set run count to 2 (each test will run twice) # Set copy_matrix to false (the matrices will be copied to GPUs only once) # # Run test with: # cd bin # sudo ./rvs -c conf/gst_1.conf -d 3 # # Expected result: # The test on each GPU passes (TRUE) if the GPU achieves 5000 gflops # in maximum 7 seconds and then the GPU sustains the gflops # for the rest of the test duration (total duration is 18 seconds). # A single Gflops violation (with a 7% tolerance) is allowed. # FALSE otherwise actions: - name: action_1 device: all #Runs on all devices module: gst parallel: true #True/False, Runs on all 8 GPUs count: 1 #Number of times you want to repeat the test duration: 7000 #Duration of the test, but hit_calls takes over duration ( down below) hot_calls: 20000 #Hot calls inner iteration, it blasts the GPU, 5000 will run for >=30mins approximately copy_matrix: false #True , generates the matrix everytime and does a BLAS operation, false, does it only at the start target_stress: 2000 #End of the test it checks wether it reached target stress or not matrix_size_a: 4096 #Size of matrix a matrix_size_b: 4096 #Size of matrix b matrix_size_c: 4096 #Size of matrix c ops_type: sgemm #Double precission GEMM transa: 0 #Matrix Transpose A, can be 0 or 1 transb: 0 #Matrix Transpose B, can be 0 or 1 alpha: 1 #Scalar value 1 beta: 1 #Scalar value 1 lda: 4096 #Leading offset A ldb: 4096 #Leading offset B ldc: 4096 #Leading offset C