Instrumenting POCL memory allocation

instructions.md

Tracing POCL Memory Allocation

To use this, build pocl with the patch below and then run your OpenCL program with

POCL_DEBUG=memory

Since you'll be using a patched pocl, maybe you will need ocl-icd to see that. To get that to happen, you can use this shell script. Save as, maybe, with-my-pocl:

#! /bin/bash
OCL_ICD_VENDORS=$HOME/shared/pack/pocl-master-install LD_LIBRARY_PATH=$HOME/shared/pack/pocl-master-install/lib:$LD_LIBRARY_PATH exec "$@"

Then run:

POCL_DEBUG=memory with-my-pocl ./my-program 2>out

The resulting file out can then be read by the parsing script.

parse_pocl_log.py

import re
import datetime

POCL_LEADER_RE = re.compile(r"^\[([-0-9: ]+)(\.[0-9]+)\] POCL:(.*)$")


def main():
    with open("out", "r") as inf:
        lines = inf.readlines()

    pocl_logs = []
    i = 0
    while i < len(lines):
        ln = lines[i]
        if m := POCL_LEADER_RE.match(ln):
            i += 1
            dt = (
                    datetime.datetime.strptime(m.group(1), "%Y-%m-%d %H:%M:%S").timestamp()
                    + float(m.group(2)))

            level, what, msg = lines[i].strip().split("|")
            pocl_logs.append((dt, m.group(3).strip(), level.strip(" *"), what.strip(), msg.strip()))

        i += 1
        if i % 100_000 in [0, 1]:
            print(i)

    mem_timestamps = []
    allocated_sizes = []

    allocated = 0

    alloc_timestamps = []
    addresses = []

    for tstamp, where, level, what, msg in pocl_logs:
        if msg.startswith("mem_alloc"):
            addr, size = msg.split(", ")
            size = int(size.strip().removeprefix("size").strip())
            addr = int(addr.strip().removeprefix("mem_alloc").strip(), 16)
            allocated += size

            mem_timestamps.append(tstamp)
            allocated_sizes.append(allocated)

            alloc_timestamps.append(tstamp)
            addresses.append(addr)

        elif msg.startswith("mem_free"):
            mem_timestamps.append(tstamp)
            _, size = msg.split(", ")
            size = int(size.strip().removeprefix("size").strip())
            allocated -= size
            allocated_sizes.append(allocated)

        else:
            pass

    import matplotlib.pyplot as plt
    if 1:
        plt.plot(mem_timestamps, allocated_sizes)
    else:
        plt.plot(alloc_timestamps, addresses, "x")
    plt.show()

if __name__ == "__main__":
    main()

pocl-instrumentation.diff

diff --git a/lib/CL/devices/common.c b/lib/CL/devices/common.c
index a62eca66..bb7ca332 100644
--- a/lib/CL/devices/common.c
+++ b/lib/CL/devices/common.c
@@ -1280,6 +1280,8 @@ pocl_aligned_malloc_global_mem(cl_device_id device, size_t align, size_t size)
   if (!retval)
     goto ERROR;
 
+  POCL_MSG_PRINT_MEMORY (" mem_alloc %p, size %d\n", retval, size);
+
   mem->currently_allocated += size;
   if (mem->max_ever_allocated < mem->currently_allocated)
     mem->max_ever_allocated = mem->currently_allocated;
@@ -1301,6 +1303,8 @@ pocl_free_global_mem(cl_device_id device, void* ptr, size_t size)
   mem->currently_allocated -= size;
   POCL_UNLOCK (mem->pocl_lock);
 
+  POCL_MSG_PRINT_MEMORY (" mem_free %p, size %d\n", ptr, size);
+
   POCL_MEM_FREE(ptr);
 }