# 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
#   Set parallel execution to false
#   Set matrix_size to 8640 (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: gpustress-9000-sgemm-false
  device: all
  module: gst
  parallel: false
  count: 1
  duration: 10000
  copy_matrix: false
  target_stress: 9000
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: sgemm
  lda: 8640
  ldb: 8640
  ldc: 8640


- name: gpustress-8000-sgemm-true
  device: all
  module: gst
  parallel: true
  count: 1
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 8000
  tolerance: 0.07
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: sgemm
  lda: 8640
  ldb: 8640
  ldc: 8640

- name: gpustress-8000-hgemm-false
  device: all
  module: gst
  parallel: false
  count: 1
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 16000
  tolerance: 0.1
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: hgemm
  lda: 8640
  ldb: 8640
  ldc: 8640

- name: gpustress-8000-hgemm-true
  device: all
  module: gst
  parallel: true
  count: 2
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 16000
  tolerance: 0.07
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: hgemm
  lda: 8640
  ldb: 8640
  ldc: 8640

- name: gpustress-8000-dgemm-false
  device: all
  module: gst
  parallel: false
  count: 2
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 4000
  tolerance: 0.07
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: dgemm
  lda: 8640
  ldb: 8640
  ldc: 8640

- name: gpustress-8000-dgemm-true
  device: all
  module: gst
  parallel: true
  count: 2
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 4000
  tolerance: 0.07
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: dgemm
  lda: 8640
  ldb: 8640
  ldc: 8640

- name: gpustress-8000-device-false
  device: all
  module: gst
  parallel: false
  count: 2
  wait: 100
  duration: 18000
  ramp_interval: 7000
  log_interval: 1000
  max_violations: 1
  copy_matrix: true
  target_stress: 8000
  tolerance: 0.07
  matrix_size_a: 8640
  matrix_size_b: 8640
  matrix_size_c: 8640
  ops_type: sgemm
  lda: 8640
  ldb: 8640
  ldc: 8640