TheRock/rocm-libraries/shared/rocroller/scripts/README.md

Scripts

trace_memory.py

This script is intended to be used from within rocGDB. e.g.:

rocgdb -ex "source trace_memory.py" -ex "gpu_memory_trace -h /work/tensile/Tensile/working/0_Build/client/tensile_client

Using this Script:

1. The first step is to get the executable and inputs you're going to execute. For Tensile this will look something like:

   ./working/0_Build/client/tensile_client --config-file ./working/1_BenchmarkProblems/Cijk_Ailk_Bjlk_HHS_BH_00/00_Final/source/ClientParameters.ini
   

   The kernel only needs to be launched once, so consider modifying the num-* parameters in the ClientParameters.ini file. You can move the ClientParameters.ini to a more convenient place if you'd like.

2. The second step is to get the label for the kernel your working on, this is the first label in the kernel. For Tensile kernels this will look something like:

   Cijk_Ailk_Bjlk_HHS_BH_MT128x256x16_MI32x32x8x1_SE_K1
   

3. Next is to get the addresses matrices you are interested in tracing. To do this you can use the get_to_kernel option to this script to stop once the kernel is reached to enable inspection.

   Setting the environment variable TENSILE_DB=0x40 will output the Tensile kernel arguments, including the addresses of all of the input matrices. For the A matrix it will have something like:

   [40..47] a: 00 00 a0 b0 f7 7f 00 00 (0x7ff7b0a00000)
   

4. Now we you are ready to collect the trace:

   rocgdb -ex "source /data/scripts/trace_memory.py" -ex "gpu_memory_trace full_memory_trace --kernel_label Cijk_Ailk_Bjlk_HHS_BH_MT128x256x16_MI32x32x8x1_SE_K1 --instruction=load --workgroup=38,1,0 --base_address=0x7ff7b0a00000 --csv_file=/data/build/memory.csv --run_command=\"--config-file ClientParameters.ini\"" /work/tensile/Tensile/working/0_Build/client/tensile_client
   

5. Finally, you can visualize the trace using the trace_to_png.py script:

   ./trace_to_png.py -c -m 7680 -n 8448 -s 2 -w 8 -t wf memory.csv