Skip to main content

Command Line Arguments

Most of the benchmarks in the samples directory have similar command line arguments. Overall, we follow the Go's command line argument convention, where each argument are lead by a single dash. Values of the argument can be either separated by a space or a equal sign. Users can use -h to get help information. The arguments are mainly divided into the following categories:

  • Simulation arguments:

    • -timing: Enable detailed timing simulation.
    • -parallel: Enable parallel simulation (not recommended).
    • -magic-memory-copy: Do not simulate the memory copy process. All the memory copy are completed without taking virtual time.
    • -debug-isa: Dump a file that records the execution of the GPU instructions and the corresponding hardware state after the execution of each instruction.
    • -trace-vis: Collect very detailed traces of the GPU execution. The visualization traces can be examined with Daisen.
    • -verify: Run the execution again in the simulator to verify if the execution is correct. Most benchmarks, but not all, support this argument. This benchmark cannot be used together with -magic-memory-copy.

  • Hardware-related arguments:

    • -gpus: Specify the GPUs to be used by the benchmark. Note that not all the benchmark support real multi-GPU execution. For example, if 4 GPUs wants to be used, the argument should be -gpus=1,2,3,4. Note that the GPU ID starts from 1, as device ID 0 is reserved for the host CPU.
    • -unified-gpus: Specify the GPUs to be used by the benchmark. This is different from the -gpus argument. With unified GPUs, the driver will create a virtual GPU interface that represent all the physical GPUs specified in the list. The workload will feel like it is running on a single GPU. The driver will split the kernel into sub-kernels and execute them on the physical GPUs. All benchmarks support this argument.

  • Metrics-related arguments:

    • -report-all: Report all the performance metrics. By default, only the total execution time and each GPU's kernel execution time are reported. The -report-all argument enables hardware performance metrics (e.g, cache hit rate).
    • -report-busy-time: Report the time that each SIMD unit is being used. This can be used to calculate GPU ALU utilization.
    • -report-cache-hit-rate: Report the number of transactions that hit and miss in each cache. Note that we separate read and write operations. Also, we include a special class named as MSHR hit, which generally should be considered as a miss, but the performance impact is not as high as regular misses.
    • -report-cache-latency: Report the average access latency at each cache unit.
    • -report-cpi-stack: Report the CPI stack recorded from each compute unit.
    • -report-dram-transaction-count: Report the number of transactions that hit the memory controllers.
    • -report-inst-count: Report the number of instructions executed by each compute unit.
    • -report-rdma-transaction-count: Report the number of transactions that hit the RDMA controllers (inter-GPU cache-line level memory access).
    • -report-tlb-hit-rate: Report the number of transactions that hit and miss in each TLB. Similar to cache hit rate, we also report MSHR hit as a special class.

  • Benchmark-specific arguments: These arguments are specific to each benchmark. For example, many benchmark has the -length argument to specify the length of the input data.