This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH, AArch64] Adjust tuning parameters for Falkor


Disabling software prefetching and switching the autoprefetcher to weak improves
CPU2017 rate and speed benchmarks for both int and fp sets on Falkor.

SPECrate 2017 fp is up 0.38%
SPECspeed 2017 fp is up 0.54%
SPECrate 2017 int is up 3.02%
SPECspeed 2017 int is up 3.16%

There are only a couple individual regressions. The biggest one being about 4%
in parest.

For SPEC2006, we've noticed the following:

SPECint is up 0.91%
SPECfp is stable

In the case of SPEC2006 we noticed both a big regression in mcf (about 20%)
and a big improvement for hmmer (about 40%).

Since the overall result is positive, we would like to make these new tuning
settings the default for Falkor.

We may revisit the software prefetcher setting in the future, in case we
can adjust it enough so it provides us a good balance between improvements and
regressions (mcf). But for now it is best if it stays off.

I understand the freeze is happening soon, so it would be great to have this
in before then.

OK?

Thanks,
Luis

2017-11-14  Luis Machado  <luis.machado@linaro.org>

	* config/aarch64/aarch64.c (qdf24xx_prefetch_tune): Remove.
	(qdf24xx_tunings): Replace qdf24xx_prefetch_tune with
	generic_prefetch_tune and tune_params::AUTOPREFETCHER_STRONG with
	tune_params::AUTOPREFETCHER_WEAK.
---
 gcc/ChangeLog                |  7 +++++++
 gcc/config/aarch64/aarch64.c | 13 ++-----------
 2 files changed, 9 insertions(+), 11 deletions(-)

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index b80a421..4dbfda0 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2017-11-14  Luis Machado  <luis.machado@linaro.org>
+
+	* config/aarch64/aarch64.c (qdf24xx_prefetch_tune): Remove.
+	(qdf24xx_tunings): Replace qdf24xx_prefetch_tune with
+	generic_prefetch_tune and tune_params::AUTOPREFETCHER_STRONG with
+	tune_params::AUTOPREFETCHER_WEAK.
+
 2017-11-14  Carl Love  <cel@us.ibm.com>
 
 	* config/rs6000/rs6000.c (swap_endian_selector_for_mode): Remove
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 0c67e2b..171a230 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -502,15 +502,6 @@ static const cpu_prefetch_tune exynosm1_prefetch_tune =
   -1			/* default_opt_level  */
 };
 
-static const cpu_prefetch_tune qdf24xx_prefetch_tune =
-{
-  4,			/* num_slots  */
-  32,			/* l1_cache_size  */
-  64,			/* l1_cache_line_size  */
-  1024,			/* l2_cache_size  */
-  3			/* default_opt_level  */
-};
-
 static const cpu_prefetch_tune thunderxt88_prefetch_tune =
 {
   8,			/* num_slots  */
@@ -817,9 +808,9 @@ static const struct tune_params qdf24xx_tunings =
   2,	/* min_div_recip_mul_sf.  */
   2,	/* min_div_recip_mul_df.  */
   0,	/* max_case_values.  */
-  tune_params::AUTOPREFETCHER_STRONG,	/* autoprefetcher_model.  */
+  tune_params::AUTOPREFETCHER_WEAK,	/* autoprefetcher_model.  */
   (AARCH64_EXTRA_TUNE_NONE),		/* tune_flags.  */
-  &qdf24xx_prefetch_tune
+  &generic_prefetch_tune
 };
 
 /* Tuning structure for the Qualcomm Saphira core.  Default to falkor values
-- 
2.7.4


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]