On Thu, 26 May 2022 15:20:29 +0200
Mattias Rönnblom <hof...@lysator.liu.se> wrote:
> On 2022-05-25 22:31, Stephen Hemminger wrote:
> > The PIE code and other applications can benefit from having a
> > fast way to get a random floating point value. This new function
> > is equivalent to drand() in the standard library.
> >
> > Signed-off-by: Stephen Hemminger <step...@networkplumber.org>
> > ---
> > app/test/test_rand_perf.c | 7 +++++
> > doc/guides/rel_notes/release_22_07.rst | 5 ++++
> > lib/eal/common/rte_random.c | 41 ++++++++++++++++++++++++++
> > lib/eal/include/rte_random.h | 18 +++++++++++
> > lib/eal/meson.build | 3 ++
> > lib/eal/version.map | 1 +
> > 6 files changed, 75 insertions(+)
> >
> > diff --git a/app/test/test_rand_perf.c b/app/test/test_rand_perf.c
> > index fe797ebfa1ca..26fb1d9a586e 100644
> > --- a/app/test/test_rand_perf.c
> > +++ b/app/test/test_rand_perf.c
> > @@ -20,6 +20,7 @@ static volatile uint64_t vsum;
> >
> > enum rand_type {
> > rand_type_64,
> > + rand_type_float,
> > rand_type_bounded_best_case,
> > rand_type_bounded_worst_case
> > };
> > @@ -30,6 +31,8 @@ rand_type_desc(enum rand_type rand_type)
> > switch (rand_type) {
> > case rand_type_64:
> > return "Full 64-bit [rte_rand()]";
> > + case rand_type_float:
> > + return "Floating point [rte_drand()]";
> > case rand_type_bounded_best_case:
> > return "Bounded average best-case [rte_rand_max()]";
> > case rand_type_bounded_worst_case:
> > @@ -55,6 +58,9 @@ test_rand_perf_type(enum rand_type rand_type)
> > case rand_type_64:
> > sum += rte_rand();
> > break;
> > + case rand_type_float:
> > + sum += 1000. * rte_drand();
>
> Including this floating point multiplication will lead to an
> overestimation of rte_drand() latency.
>
> You could refactor this function to be a macro, and pass the return type
> to as a parameter to this macro. I did just that, and on both an AMD
> 5900X and a Cortex-A72 it didn't add more than ~5%, so I don't think
> it's necessary.
>
> > + break;
> > case rand_type_bounded_best_case:
> > sum += rte_rand_max(BEST_CASE_BOUND);
> > break;
> > @@ -83,6 +89,7 @@ test_rand_perf(void)
> > printf("Pseudo-random number generation latencies:\n");
> >
> > test_rand_perf_type(rand_type_64);
> > + test_rand_perf_type(rand_type_float);
> > test_rand_perf_type(rand_type_bounded_best_case);
> > test_rand_perf_type(rand_type_bounded_worst_case);
> >
> > diff --git a/doc/guides/rel_notes/release_22_07.rst
> > b/doc/guides/rel_notes/release_22_07.rst
> > index e49cacecefd4..b131ea577226 100644
> > --- a/doc/guides/rel_notes/release_22_07.rst
> > +++ b/doc/guides/rel_notes/release_22_07.rst
> > @@ -104,6 +104,11 @@ New Features
> > * ``RTE_EVENT_QUEUE_ATTR_WEIGHT``
> > * ``RTE_EVENT_QUEUE_ATTR_AFFINITY``
> >
> > +* ** Added function get random floating point number.**
> > +
> > + Added the function ``rte_drand()`` to provide a pseudo-random
> > + floating point number.
> > +
> >
> > Removed Items
> > -------------
> > diff --git a/lib/eal/common/rte_random.c b/lib/eal/common/rte_random.c
> > index 4535cc980cec..3dc3484ee655 100644
> > --- a/lib/eal/common/rte_random.c
> > +++ b/lib/eal/common/rte_random.c
> > @@ -6,6 +6,9 @@
> > #include <x86intrin.h>
> > #endif
> > #include <unistd.h>
> > +#ifdef RTE_LIBEAL_USE_IEEE754
> > +#include <ieee754.h>
> > +#endif
> >
> > #include <rte_branch_prediction.h>
> > #include <rte_cycles.h>
> > @@ -173,6 +176,44 @@ rte_rand_max(uint64_t upper_bound)
> > return res;
> > }
> >
> > +double
> > +rte_drand(void)
> > +{
> > + struct rte_rand_state *state = __rte_rand_get_state();
> > + uint64_t rand64 = __rte_rand_lfsr258(state);
> > +#ifdef RTE_LIBEAL_USE_IEEE754
> > + union ieee754_double u = {
> > + .ieee = {
> > + .negative = 0,
> > + .exponent = IEEE754_DOUBLE_BIAS,
> > + },
> > + };
> > +
> > + /* Take 64 bit random value and put it into the mantissa
> > + * This uses direct access to IEEE format to avoid doing
> > + * any direct floating point math here.
> > + */
> > + u.ieee.mantissa0 = rand64 >> 32;
> > + u.ieee.mantissa1 = rand64;
> > +
> > + return u.d - 1.0;
> > +#else
> > + /* Slower method requiring floating point divide
> > + *
>
> Do you know how much slower? I ran rand_perf_test on two of my systems.
>
> AMD 5900X Pi4 (ARM Cortex-A72)
> IEEE754 version 12 1.19
> Non-IEEE754 version 11 1.16
> Naive version* 24 1.16
>
> * (double)rte_rand() / (double)UINT64_MAX
>
> Numbers are TSC cycles/op.
On AMD Ryzen 7 both versions take 9 cycles/op with the rand_perf_autotest
So it is a toss up.
The 754 version is:
ubfx r1, r1, #0, #20
orr r3, r1, #1069547520 << mantissa0
mov r2, r0
orr r3, r3, #3145728
vmov.f64 d0, #1.0e+0
vmov d16, r2, r3
vsub.f64 d0, d16, d0 << return u.d - 1.0
Note: the compiler is doing smart optimization on the divide version.
It knows that since denominator is fixed value it can use multiply.
vmov d16, r0, r1
vmul.f64 d0, d16, d0
