P9 Float128 add/sub with round to odd, plus P8 equivalent.

Power9 provide round to odd versions for the QP arithmetic operations.
This patch provides P9/8 implementation for add/sub QP with round
to odd. Includes compile, unit, and performance tests.

	* src/pveclib/vec_f128_ppc.h [doxygen brief]:
	Added micro-benchmark data. Other clean up.
	[doxygen f128_softfloat_0_0]: Clarifications.
	[doxygen f128_softfloat_0_0_0]: Clarifications.
	[doxygen f128_softfloat_IRRN_0_0]: Clarifying Note.
	[doxygen f128_softfloat_IRRN_0_1]:
	Improve internal representation (IR) overview.
	[doxygen f128_softfloat_IRRN_0_2]:
	Expond on how the IR impacts rounding.
	[doxygen f128_softfloat_0_0_3_2]:
	Add overview of Add Quad-Precision with Round-to-Odd.
	[doxygen f128_softfloat_0_0_3_3]:
	Add overview of Subract Quad-Precision with Round-to-Odd.
	[__clang__]: Clang keeps changing, tried to fix it.
	(vec_negf128 [__FLOAT128__) && (__GNUC__ > 7]:
	Use C arithmetic syntax for negate.
	(vec_xsaddqpo, vec_xssubqpo): New inline operations.
	(vec_xsmulqpo): Update latency numbers.

	* src/testsuite/arith128_test_f128.c
	(db_vec_xsaddqpo, db_vec_xssubqpo): New debug implementations.
	(test_add_qpo, test_sub_qpo): New unit test functions.
	(test_sub_qpo): Add test_add_qpo and test_sub_qpo to test
	driver.

	* src/testsuite/pveclib_perf.c (test_time_f128):
	Add timed tests timed_gcc_addqpn_f128, timed_lib_addqpo_f128,
	timed_gcc_subqpn_f128, and timed_lib_subqpo_f128.

	* src/testsuite/vec_f128_dummy.c (force_eMin, force_eMin_V0,
	test_vec_xsaddqpo, test_vec_xssubqpo, test_genqpo_v0,
	test_vec_addqpo, test_vec_addqpo_V1, test_vec_addqpo_V0,
	test_negqp_nan_v0, test_vec_subqpo, test_vec_subqpo_V0):
	New compile tests. Some are used in unit and perf tests.
	(test_gcc_addqpn_f128, test_gcc_subqpn_f128):
	New combined compile and perf operation tests.
	(test_gcc_mulqpn_f128, test_gcc_mulqpo_f128): Fix
	combined compile and perf operation tests.

	* src/testsuite/vec_perf_f128.c (test_gcc_addqpn_f128,
	test_gcc_subqpn_f128, test_vec_addqpo, test_vec_subqpo):
	New externs.
 	(test_lib_addqpo_f128, test_lib_subqpo_f128):
 	New inner perf test.
 	(timed_lib_addqpo_f128, timed_lib_subqpo_f128,
 	timed_gcc_addqpn_f128, timed_gcc_subqpn_f128):
 	New timed perf tests.

 	* src/testsuite/vec_perf_f128.h (timed_gcc_addqpn_f128.
 	timed_lib_addqpo_f128, timed_gcc_subqpn_f128,
 	timed_lib_subqpo_f128): New externs.

 	*src/testsuite/vec_pwr10_dummy.c
 	[(__GNUC__ == 11) && (__GNUC_MINOR__ > 2)]:
 	Waiting for fixes in GCC 11.3?

 	* src/testsuite/vec_pwr9_dummy.c (test_negqp_PWR9):
 	New compile test.
 	Various -Wall clean-ups.

Signed-off-by: Steven Munroe <[email protected]>

src/testsuite/pveclib_perf.c

@@ -850,6 +850,62 @@ test_time_f128 (void)
   printf ("\n%s mulqpn_lib tb delta = %lu, sec = %10.6g\n", __FUNCTION__,
 	  t_delta, delta_sec);
 
+  printf ("\n%s addqpn_gcc start, ...\n", __FUNCTION__);
+  t_start = __builtin_ppc_get_timebase ();
+  for (i = 0; i < TIMING_ITERATIONS; i++)
+    {
+      rc += timed_gcc_addqpn_f128 ();
+    }
+  t_end = __builtin_ppc_get_timebase ();
+  t_delta = t_end - t_start;
+  delta_sec = TimeDeltaSec (t_delta);
+
+  printf ("\n%s addqpn_gcc end", __FUNCTION__);
+  printf ("\n%s addqpn_gcc tb delta = %lu, sec = %10.6g\n", __FUNCTION__,
+	  t_delta, delta_sec);
+
+  printf ("\n%s addqpo_lib start, ...\n", __FUNCTION__);
+  t_start = __builtin_ppc_get_timebase ();
+  for (i = 0; i < TIMING_ITERATIONS; i++)
+    {
+      rc += timed_lib_addqpo_f128 ();
+    }
+  t_end = __builtin_ppc_get_timebase ();
+  t_delta = t_end - t_start;
+  delta_sec = TimeDeltaSec (t_delta);
+
+  printf ("\n%s addqpo_lib end", __FUNCTION__);
+  printf ("\n%s addqpo_lib tb delta = %lu, sec = %10.6g\n", __FUNCTION__,
+	  t_delta, delta_sec);
+
+  printf ("\n%s subqpn_gcc start, ...\n", __FUNCTION__);
+  t_start = __builtin_ppc_get_timebase ();
+  for (i = 0; i < TIMING_ITERATIONS; i++)
+    {
+      rc += timed_gcc_subqpn_f128 ();
+    }
+  t_end = __builtin_ppc_get_timebase ();
+  t_delta = t_end - t_start;
+  delta_sec = TimeDeltaSec (t_delta);
+
+  printf ("\n%s subqpn_gcc end", __FUNCTION__);
+  printf ("\n%s subqpn_gcc tb delta = %lu, sec = %10.6g\n", __FUNCTION__,
+	  t_delta, delta_sec);
+
+  printf ("\n%s subqpo_lib start, ...\n", __FUNCTION__);
+  t_start = __builtin_ppc_get_timebase ();
+  for (i = 0; i < TIMING_ITERATIONS; i++)
+    {
+      rc += timed_lib_subqpo_f128 ();
+    }
+  t_end = __builtin_ppc_get_timebase ();
+  t_delta = t_end - t_start;
+  delta_sec = TimeDeltaSec (t_delta);
+
+  printf ("\n%s subqpo_lib end", __FUNCTION__);
+  printf ("\n%s subqpo_lib tb delta = %lu, sec = %10.6g\n", __FUNCTION__,
+	  t_delta, delta_sec);
+
   return (rc);
 }
 #endif