[v1,02/23] benchtests: Add random benchmark in bench-strchr.c

Message ID 20220323215734.3927131-2-goldstein.w.n@gmail.com
State Accepted, archived
Headers
Series [v1,01/23] benchtests: Use json-lib in bench-strchr.c |

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent

Commit Message

Noah Goldstein March 23, 2022, 9:57 p.m. UTC
  Add benchmark that randomizes whether return should be NULL or pointer
to CHAR. The rationale is on many architectures there is a choice
between a predicate execution option (i.e cmovcc on x86) or a branch.

On x86 the results for cmovcc vs branch are something along the lines
of the following:

perc-zero, Br On Result, Time Br / Time cmov
     0.10,            1,              ,0.983
     0.10,            0,              ,1.246
     0.25,            1,              ,1.035
     0.25,            0,              ,1.49
     0.33,            1,              ,1.016
     0.33,            0,              ,1.579
     0.50,            1,              ,1.228
     0.50,            0,              ,1.739
     0.66,            1,              ,1.039
     0.66,            0,              ,1.764
     0.75,            1,              ,0.996
     0.75,            0,              ,1.642
     0.90,            1,              ,1.071
     0.90,            0,              ,1.409
     1.00,            1,              ,0.937
     1.00,            0,              ,0.999
---
 benchtests/bench-strchr.c | 143 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 143 insertions(+)
  

Comments

H.J. Lu March 24, 2022, 6:44 p.m. UTC | #1
On Wed, Mar 23, 2022 at 2:58 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> Add benchmark that randomizes whether return should be NULL or pointer
> to CHAR. The rationale is on many architectures there is a choice
> between a predicate execution option (i.e cmovcc on x86) or a branch.
>
> On x86 the results for cmovcc vs branch are something along the lines
> of the following:
>
> perc-zero, Br On Result, Time Br / Time cmov
>      0.10,            1,              ,0.983
>      0.10,            0,              ,1.246
>      0.25,            1,              ,1.035
>      0.25,            0,              ,1.49
>      0.33,            1,              ,1.016
>      0.33,            0,              ,1.579
>      0.50,            1,              ,1.228
>      0.50,            0,              ,1.739
>      0.66,            1,              ,1.039
>      0.66,            0,              ,1.764
>      0.75,            1,              ,0.996
>      0.75,            0,              ,1.642
>      0.90,            1,              ,1.071
>      0.90,            0,              ,1.409
>      1.00,            1,              ,0.937
>      1.00,            0,              ,0.999
> ---
>  benchtests/bench-strchr.c | 143 ++++++++++++++++++++++++++++++++++++++
>  1 file changed, 143 insertions(+)
>
> diff --git a/benchtests/bench-strchr.c b/benchtests/bench-strchr.c
> index 203900d4ad..54640bde7e 100644
> --- a/benchtests/bench-strchr.c
> +++ b/benchtests/bench-strchr.c
> @@ -53,6 +53,11 @@
>  # define SMALL_CHAR 851
>  #endif /* WIDE */
>
> +#ifdef USE_FOR_STRCHRNUL
> +# define DO_RAND_TEST(...)
> +#else
> +# define DO_RAND_TEST(...) do_rand_test(__VA_ARGS__)
> +#endif
>  #ifdef USE_FOR_STRCHRNUL
>  # define NULLRET(endptr) endptr
>  #else
> @@ -74,6 +79,133 @@ simple_STRCHR (const CHAR *s, int c)
>  IMPL (simple_STRCHR, 0)
>  IMPL (STRCHR, 1)
>
> +#ifndef USE_FOR_STRCHRNUL
> +/* Random benchmarks for strchr (if return is CHAR or NULL).  The
> +   rational for the benchmark is returning null/char can be done with
> +   predicate execution (i.e cmovcc on x86) or a branch. */
> +
> +
> +/* Large enough that full history can't be stored in BHT. */
> +#define NUM_SEARCH_CHARS 2048
> +
> +/* Expectation is usecases of strchr check the return. Otherwise
> +   strchrnul would almost always be better. Since there is another
> +   branch coming we want to test the case where a potential branch in
> +   strchr can be used to skip a later mispredict because of the
> +   relationship between the two branches. */
> +static void __attribute__ ((noinline, noclone))
> +do_one_rand_plus_branch_test (json_ctx_t *json_ctx, impl_t *impl,
> +                              const CHAR *s, const CHAR *c)
> +{
> +  size_t i, iters = INNER_LOOP_ITERS_LARGE;
> +  int must_execute = 0;
> +  timing_t start, stop, cur;
> +  TIMING_NOW (start);
> +  for (i = 0; i < iters; ++i)
> +    {
> +      if (CALL (impl, s, c[i % NUM_SEARCH_CHARS]))
> +        {
> +          /* We just need something that will force compiler to emit
> +             a branch instead of conditional execution. */
> +          ++must_execute;
> +          asm volatile("" : : :);
> +        }
> +    }
> +  TIMING_NOW (stop);
> +
> +  TIMING_DIFF (cur, start, stop);
> +
> +  json_element_double (json_ctx, (double)cur / (double)iters);
> +}
> +
> +static void __attribute__ ((noinline, noclone))
> +do_one_rand_test (json_ctx_t *json_ctx, impl_t *impl, const CHAR *s,
> +                  const CHAR *c)
> +{
> +  size_t i, iters = INNER_LOOP_ITERS_LARGE;
> +  timing_t start, stop, cur;
> +  TIMING_NOW (start);
> +  for (i = 0; i < iters; ++i)
> +    {
> +      CALL (impl, s, c[i % NUM_SEARCH_CHARS]);
> +    }
> +  TIMING_NOW (stop);
> +
> +  TIMING_DIFF (cur, start, stop);
> +
> +  json_element_double (json_ctx, (double)cur / (double)iters);
> +}
> +
> +static void
> +do_rand_test (json_ctx_t *json_ctx, size_t align, size_t pos, size_t len,
> +              float perc_zero)
> +{
> +  size_t i;
> +  int perc_zero_int;
> +  CHAR *buf = (CHAR *)buf1;
> +  CHAR *c = (CHAR *)buf2;
> +  align &= 127;
> +  if ((align + len) * sizeof (CHAR) >= page_size)
> +    return;
> +
> +  /* Test is only interesting if we can hit both cases. */
> +  if (pos >= len)
> +    return;
> +
> +  /* Segfault if we run the test. */
> +  if (NUM_SEARCH_CHARS * sizeof (CHAR) > page_size)
> +    return;
> +
> +  for (i = 0; i < len; ++i)
> +    {
> +      buf[align + i] = 2;
> +    }
> +  buf[align + len] = 0;
> +  buf[align + pos] = 1;
> +
> +  perc_zero_int = perc_zero * RAND_MAX;
> +  for (i = 0; i < NUM_SEARCH_CHARS; ++i)
> +    {
> +      if (rand () > perc_zero_int)
> +        c[i] = 0;
> +      else
> +        c[i] = 1;
> +    }
> +  {
> +    json_element_object_begin (json_ctx);
> +    json_attr_uint (json_ctx, "rand", 1);
> +    json_attr_uint (json_ctx, "branch", 1);
> +    json_attr_double (json_ctx, "perc-zero", perc_zero);
> +    json_attr_uint (json_ctx, "length", len);
> +    json_attr_uint (json_ctx, "pos", pos);
> +    json_attr_uint (json_ctx, "alignment", align);
> +    json_array_begin (json_ctx, "timings");
> +
> +    FOR_EACH_IMPL (impl, 0)
> +      do_one_rand_plus_branch_test (json_ctx, impl, buf + align, c);
> +
> +    json_array_end (json_ctx);
> +    json_element_object_end (json_ctx);
> +  }
> +  {
> +    json_element_object_begin (json_ctx);
> +    json_attr_uint (json_ctx, "rand", 1);
> +    json_attr_uint (json_ctx, "branch", 0);
> +    json_attr_double (json_ctx, "perc-zero", perc_zero);
> +    json_attr_uint (json_ctx, "length", len);
> +    json_attr_uint (json_ctx, "pos", pos);
> +    json_attr_uint (json_ctx, "alignment", align);
> +    json_array_begin (json_ctx, "timings");
> +
> +    FOR_EACH_IMPL (impl, 0)
> +      do_one_rand_test (json_ctx, impl, buf + align, c);
> +
> +    json_array_end (json_ctx);
> +    json_element_object_end (json_ctx);
> +  }
> +}
> +#endif
> +
>  static void
>  do_one_test (json_ctx_t *json_ctx, impl_t *impl, const CHAR *s, int c,
>               const CHAR *exp_res)
> @@ -136,6 +268,7 @@ do_test (json_ctx_t *json_ctx, size_t align, size_t pos, size_t len,
>      result = NULLRET (buf + align + len);
>
>    json_element_object_begin (json_ctx);
> +  json_attr_uint (json_ctx, "rand", 0);
>    json_attr_uint (json_ctx, "length", len);
>    json_attr_uint (json_ctx, "pos", pos);
>    json_attr_uint (json_ctx, "seek_char", seek_char);
> @@ -234,6 +367,16 @@ test_main (void)
>        do_test (&json_ctx, 0, i, i + 1, 0, BIG_CHAR);
>      }
>
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.0);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.1);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.25);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.33);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.5);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.66);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.75);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.9);
> +  DO_RAND_TEST(&json_ctx, 0, 15, 16, 1.0);
> +
>    json_array_end (&json_ctx);
>    json_attr_object_end (&json_ctx);
>    json_attr_object_end (&json_ctx);
> --
> 2.25.1
>

LGTM.

Reviewed-by: H.J. Lu <hjl.tools@gmail.com>

Thanks.
  

Patch

diff --git a/benchtests/bench-strchr.c b/benchtests/bench-strchr.c
index 203900d4ad..54640bde7e 100644
--- a/benchtests/bench-strchr.c
+++ b/benchtests/bench-strchr.c
@@ -53,6 +53,11 @@ 
 # define SMALL_CHAR 851
 #endif /* WIDE */
 
+#ifdef USE_FOR_STRCHRNUL
+# define DO_RAND_TEST(...)
+#else
+# define DO_RAND_TEST(...) do_rand_test(__VA_ARGS__)
+#endif
 #ifdef USE_FOR_STRCHRNUL
 # define NULLRET(endptr) endptr
 #else
@@ -74,6 +79,133 @@  simple_STRCHR (const CHAR *s, int c)
 IMPL (simple_STRCHR, 0)
 IMPL (STRCHR, 1)
 
+#ifndef USE_FOR_STRCHRNUL
+/* Random benchmarks for strchr (if return is CHAR or NULL).  The
+   rational for the benchmark is returning null/char can be done with
+   predicate execution (i.e cmovcc on x86) or a branch. */
+
+
+/* Large enough that full history can't be stored in BHT. */
+#define NUM_SEARCH_CHARS 2048
+
+/* Expectation is usecases of strchr check the return. Otherwise
+   strchrnul would almost always be better. Since there is another
+   branch coming we want to test the case where a potential branch in
+   strchr can be used to skip a later mispredict because of the
+   relationship between the two branches. */
+static void __attribute__ ((noinline, noclone))
+do_one_rand_plus_branch_test (json_ctx_t *json_ctx, impl_t *impl,
+                              const CHAR *s, const CHAR *c)
+{
+  size_t i, iters = INNER_LOOP_ITERS_LARGE;
+  int must_execute = 0;
+  timing_t start, stop, cur;
+  TIMING_NOW (start);
+  for (i = 0; i < iters; ++i)
+    {
+      if (CALL (impl, s, c[i % NUM_SEARCH_CHARS]))
+        {
+          /* We just need something that will force compiler to emit
+             a branch instead of conditional execution. */
+          ++must_execute;
+          asm volatile("" : : :);
+        }
+    }
+  TIMING_NOW (stop);
+
+  TIMING_DIFF (cur, start, stop);
+
+  json_element_double (json_ctx, (double)cur / (double)iters);
+}
+
+static void __attribute__ ((noinline, noclone))
+do_one_rand_test (json_ctx_t *json_ctx, impl_t *impl, const CHAR *s,
+                  const CHAR *c)
+{
+  size_t i, iters = INNER_LOOP_ITERS_LARGE;
+  timing_t start, stop, cur;
+  TIMING_NOW (start);
+  for (i = 0; i < iters; ++i)
+    {
+      CALL (impl, s, c[i % NUM_SEARCH_CHARS]);
+    }
+  TIMING_NOW (stop);
+
+  TIMING_DIFF (cur, start, stop);
+
+  json_element_double (json_ctx, (double)cur / (double)iters);
+}
+
+static void
+do_rand_test (json_ctx_t *json_ctx, size_t align, size_t pos, size_t len,
+              float perc_zero)
+{
+  size_t i;
+  int perc_zero_int;
+  CHAR *buf = (CHAR *)buf1;
+  CHAR *c = (CHAR *)buf2;
+  align &= 127;
+  if ((align + len) * sizeof (CHAR) >= page_size)
+    return;
+
+  /* Test is only interesting if we can hit both cases. */
+  if (pos >= len)
+    return;
+
+  /* Segfault if we run the test. */
+  if (NUM_SEARCH_CHARS * sizeof (CHAR) > page_size)
+    return;
+
+  for (i = 0; i < len; ++i)
+    {
+      buf[align + i] = 2;
+    }
+  buf[align + len] = 0;
+  buf[align + pos] = 1;
+
+  perc_zero_int = perc_zero * RAND_MAX;
+  for (i = 0; i < NUM_SEARCH_CHARS; ++i)
+    {
+      if (rand () > perc_zero_int)
+        c[i] = 0;
+      else
+        c[i] = 1;
+    }
+  {
+    json_element_object_begin (json_ctx);
+    json_attr_uint (json_ctx, "rand", 1);
+    json_attr_uint (json_ctx, "branch", 1);
+    json_attr_double (json_ctx, "perc-zero", perc_zero);
+    json_attr_uint (json_ctx, "length", len);
+    json_attr_uint (json_ctx, "pos", pos);
+    json_attr_uint (json_ctx, "alignment", align);
+    json_array_begin (json_ctx, "timings");
+
+    FOR_EACH_IMPL (impl, 0)
+      do_one_rand_plus_branch_test (json_ctx, impl, buf + align, c);
+
+    json_array_end (json_ctx);
+    json_element_object_end (json_ctx);
+  }
+  {
+    json_element_object_begin (json_ctx);
+    json_attr_uint (json_ctx, "rand", 1);
+    json_attr_uint (json_ctx, "branch", 0);
+    json_attr_double (json_ctx, "perc-zero", perc_zero);
+    json_attr_uint (json_ctx, "length", len);
+    json_attr_uint (json_ctx, "pos", pos);
+    json_attr_uint (json_ctx, "alignment", align);
+    json_array_begin (json_ctx, "timings");
+
+    FOR_EACH_IMPL (impl, 0)
+      do_one_rand_test (json_ctx, impl, buf + align, c);
+
+    json_array_end (json_ctx);
+    json_element_object_end (json_ctx);
+  }
+}
+#endif
+
 static void
 do_one_test (json_ctx_t *json_ctx, impl_t *impl, const CHAR *s, int c,
              const CHAR *exp_res)
@@ -136,6 +268,7 @@  do_test (json_ctx_t *json_ctx, size_t align, size_t pos, size_t len,
     result = NULLRET (buf + align + len);
 
   json_element_object_begin (json_ctx);
+  json_attr_uint (json_ctx, "rand", 0);
   json_attr_uint (json_ctx, "length", len);
   json_attr_uint (json_ctx, "pos", pos);
   json_attr_uint (json_ctx, "seek_char", seek_char);
@@ -234,6 +367,16 @@  test_main (void)
       do_test (&json_ctx, 0, i, i + 1, 0, BIG_CHAR);
     }
 
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.0);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.1);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.25);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.33);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.5);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.66);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.75);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 0.9);
+  DO_RAND_TEST(&json_ctx, 0, 15, 16, 1.0);
+
   json_array_end (&json_ctx);
   json_attr_object_end (&json_ctx);
   json_attr_object_end (&json_ctx);