mirror/libcpuid
1
0
Fork 0
mirror of https://github.com/anrieff/libcpuid synced 2024-11-10 22:59:13 +00:00
Code Issues Releases Wiki Activity

Fix cpu_clock_by_ic() for Skylake (it was 1.6 times too high).

Browse source 
The reason and fix is similar to what we did previously for Bulldozer.
This commit is contained in:
Veselin Georgiev 2016-05-17 22:29:32 +03:00
parent 7b9fe29cef
commit 14d6a9d875
1 changed files with 40 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

55
libcpuid/rdtsc.c
View file

@ -226,6 +226,45 @@ int cpu_clock_measure(int millis, int quad_check)
return (results[bi] + results[bj] + _zero) / 2; return (results[bi] + results[bj] + _zero) / 2;
} }
static void adjust_march_ic_multiplier(const struct cpu_id_t* id, int* numerator, int* denom)
{
/*
* for cpu_clock_by_ic: we need to know how many clocks does a typical ADDPS instruction
* take, when issued in rapid succesion without dependencies. The whole idea of
* cpu_clock_by_ic was that this is easy to determine, at least it was back in 2010. Now
* it's getting progressively more hairy, but here are the current measurements:
*
* 1. For CPUs with 64-bit SSE units, ADDPS issue rate is 0.5 IPC (one insn in 2 clocks)
* 2. For CPUs with 128-bit SSE units, issue rate is exactly 1.0 IPC
* 3. For Bulldozer and later, it is 1.4 IPC (we multiply by 5/7)
* 4. For Skylake and later, it is 1.6 IPC (we multiply by 5/8)
*/
//
if (id->sse_size < 128) {
debugf(1, "SSE execution path is 64-bit\n");
// on a CPU with half SSE unit length, SSE instructions execute at 0.5 IPC;
// the resulting value must be multiplied by 2:
*numerator = 2;
} else {
debugf(1, "SSE execution path is 128-bit\n");
}
//
// Bulldozer or later: assume 1.4 IPC
if (id->vendor == VENDOR_AMD && id->ext_family >= 21) {
debugf(1, "cpu_clock_by_ic: Bulldozer (or later) detected, dividing result by 1.4\n");
*numerator = 5;
*denom = 7; // multiply by 5/7, to divide by 1.4
}
//
// Skylake or later: assume 1.6 IPC
if (id->vendor == VENDOR_INTEL && id->ext_model >= 94) {
debugf(1, "cpu_clock_by_ic: Skylake (or later) detected, dividing result by 1.6\n");
*numerator = 5;
*denom = 8; // to divide by 1.6, multiply by 5/8
}
}
int cpu_clock_by_ic(int millis, int runs) int cpu_clock_by_ic(int millis, int runs)
{ {
int max_value = 0, cur_value, i, ri, cycles_inner, cycles_outer, c; int max_value = 0, cur_value, i, ri, cycles_inner, cycles_outer, c;
@ -237,21 +276,7 @@ int cpu_clock_by_ic(int millis, int runs)
// if there aren't SSE instructions - we can't run the test at all // if there aren't SSE instructions - we can't run the test at all
if (!id || !id->flags[CPU_FEATURE_SSE]) return -1; if (!id || !id->flags[CPU_FEATURE_SSE]) return -1;
// //
if (id->sse_size < 128) { adjust_march_ic_multiplier(id, &multiplier_numerator, &multiplier_denom);
debugf(1, "SSE execution path is 64-bit\n");
// on a CPU with half SSE unit length, SSE instructions execute at 0.5 IPC;
// the resulting value must be multiplied by 2:
multiplier_numerator = 2;
} else {
debugf(1, "SSE execution path is 128-bit\n");
}
//
// on a Bulldozer or later CPU, SSE instructions execute at 1.4 IPC, handle that as well:
if (id->vendor == VENDOR_AMD && id->ext_family >= 21) {
debugf(1, "cpu_clock_by_ic: Bulldozer (or later) detected, dividing result by 1.4\n");
multiplier_numerator = 5;
multiplier_denom = 7; // multiply by 5/7, to divide by 1.4
}
// //
tl = millis * 125; // (*1000 / 8) tl = millis * 125; // (*1000 / 8)
cycles_inner = 128; cycles_inner = 128;