Add hip support to mem_ref

[SimeonEhrig]

Similar to the define ENABLE_CUDA, there is a define ENABLE_HIP if hip is enabled.

[SimeonEhrig]

ping @afif-ishamsyah

[SimeonEhrig]

I profile the main.py width only the function example_context_copy (example_manual_copy was commented).

I recognized, that are 7 hipMemcpy was applied, but only 5 was expected (2 for the function open_sync and one for the function compute). Can you please find out, how is causing the 2 extra mem copies. In the end of the comment, you find a manual to trace the application with rocprof.

This causes also the idea to add new context manager functions. Can you please rename sync_open to sync_open_rw, create the functions sync_open_r (reads only data) and sync_open_w (writes only data) and used it in the main.py to reduce memory operations.

Using rocprof

1. comment the function example_manual_copy
2. build your application
3. go to the build folder and create the trace with rocprof --hip-trace -o binding_trace.csv python src/main.py

check how many time hipMemcpy was executed

• cat binding_trace.hip_stats.csv shows how many times hipMemcpy was executed.

check when hipMemcpy was executed

• copy the file binding_trace.json to your computer
• open the chrome profiler in Chrome/Chromium (mayby new Egde) via the URL: chrome://tracing/
• open the binding_trace.json

[afif-ishamsyah]

2 extra hipMemcpy comes from algo.initialize_array function. It create 2 hip_mem object, each containing an array of zeros. 1 hip_mem will be used as input object, the other one is for output object. Creating a hip_mem object automatically run a hipMemcpy. It is on algo.hpp, line 166.