microbench/profile: rename and few other scripts

Author: paugier

.gitignore

@@ -94,5 +94,7 @@ docs/_build/
 microbench/pixi.lock
 microbench/.venv_*
 microbench/tmp_*
-microbench/profile_alloc/log
-microbench/profile_alloc/*.prof*
+microbench/profile/log*
+microbench/profile/*.prof*
+microbench/profile/*.log
+microbench/profile/callgrind.out.*

microbench/pixi.toml

@@ -12,3 +12,4 @@ gcc = ">=15.1.0,<15.2"
 valgrind = ">=3.25.0,<4"
 libffi = ">=3.4.6,<4"
 libxcrypt = ">=4.4.36,<5"
+qcachegrind = ">=0.7.4,<0.8"

microbench/profile/Makefile

@@ -0,0 +1,3 @@
+
+clean:
+	rm -f *.prof* logfile* callgrind.out.* tmp*

microbench/profile/README.md

@@ -0,0 +1,40 @@
+# Profile
+
+## Disable the JIT
+
+```sh
+python --jit off profile_alloc.py
+```
+
+## With vmprof
+
+```sh
+pip install vmprof
+sudo apt install libunwind-dev
+python -m vmprof -o tmp_output.log profile_alloc.py
+vmprofshow tmp_output.log tree
+```
+
+## with vmprof-firefox-converter
+
+```sh
+pip install vmprof-firefox-converter
+vmprofconvert -run profile_alloc.py cpy
+# or
+vmprofconvert -run profile_alloc.py
+```
+
+## With `PYPYLOG`
+
+```sh
+PYPYLOG=jit-log-opt,jit-backend:logfile python profile_alloc_and_die.py
+```
+
+## With Valgrind / Callgrind
+
+```sh
+valgrind --tool=callgrind --dump-instr=yes --simulate-cache=yes --collect-jumps=yes python profile_alloc_and_die.py --short
+```
+
+The Valgrind output (a file starting with "callgrind.out.") can for example be
+opened with `qcachegrind`.

microbench/profile/profile_alloc.py

@@ -0,0 +1,19 @@
+import gc
+
+from time import perf_counter as time
+
+from util import cls, N
+
+
+def main():
+    objs = [None] * N
+    for i in range(N):
+        objs[i] = cls()
+    return objs
+
+
+gc.collect()
+
+t_start = time()
+main()
+print(f"time per allocation: {(time() - t_start)/N:.1e} s")

microbench/profile/profile_alloc_and_die.py

@@ -0,0 +1,18 @@
+import gc
+
+from time import perf_counter as time
+
+from util import cls, N
+
+
+def main():
+    for _ in range(N):
+        obj = cls()
+        obj.onearg(None)
+
+
+gc.collect()
+
+t_start = time()
+main()
+print(f"time per allocation: {(time() - t_start)/N:.1e} s")

microbench/profile/profile_allocate_and_collect.py

@@ -0,0 +1,21 @@
+import gc
+
+from time import perf_counter as time
+
+from util import cls, N
+
+
+def main():
+    objs = [None] * N
+    for i in range(N):
+        objs[i] = cls()
+    return objs
+
+
+gc.collect()
+
+t_start = time()
+_objs = main()
+del _objs
+gc.collect()
+print(f"time per object: {(time() - t_start)/N:.1e} s")

microbench/profile/profile_allocate_tuple.py

@@ -0,0 +1,16 @@
+import gc
+
+from time import perf_counter as time
+
+from util import simple, N
+
+
+def main():
+    for _ in range(N):
+        simple.allocate_tuple()
+
+gc.collect()
+
+t_start = time()
+main()
+print(f"time per allocation: {(time() - t_start)/N:.1e} s")

microbench/profile/profile_method_lookup.py

@@ -0,0 +1,21 @@
+import gc
+
+from time import perf_counter as time
+
+from util import cls, N
+
+N *= 40
+
+def main():
+    obj = cls()
+    for _ in range(N):
+        # note: here we are NOT calling it, we want to measure just
+        # the lookup
+        obj.noargs  # pylint: disable=W0104
+
+
+gc.collect()
+
+t_start = time()
+main()
+print(f"time per lookup: {(time() - t_start)/N:.1e} s")

microbench/profile_alloc/profile_alloc.py renamed to microbench/profile/util.py

@@ -1,9 +1,5 @@
-
-import gc
 import sys
 
-from time import perf_counter as time
-
 if "purepy" in sys.argv:
     import purepy_simple as simple
 elif "cpy" in sys.argv:
@@ -17,18 +13,12 @@
 print(simple)
 
 cls = simple.HTFoo
+
 N = 10000000
 if sys.implementation.name == "cpython":
     N *= 10
+elif sys.implementation.name == "pypy" and "cpy" in sys.argv:
+    N *= 4
 
-def main():
-    objs = [None] * N
-    for i in range(N):
-        objs[i] = cls()
-    return objs
-
-gc.collect()
-
-t_start = time()
-main()
-print(f"time per allocation: {(time() - t_start)/N:.1e} s")
+if "--short" in sys.argv:
+    N //= 100