aboutsummaryrefslogtreecommitdiff
path: root/contrib/arm-optimized-routines/math/test/mathbench.c
diff options
context:
space:
mode:
Diffstat (limited to 'contrib/arm-optimized-routines/math/test/mathbench.c')
-rw-r--r--contrib/arm-optimized-routines/math/test/mathbench.c426
1 files changed, 117 insertions, 309 deletions
diff --git a/contrib/arm-optimized-routines/math/test/mathbench.c b/contrib/arm-optimized-routines/math/test/mathbench.c
index 0c17826e5296..653c58fbc484 100644
--- a/contrib/arm-optimized-routines/math/test/mathbench.c
+++ b/contrib/arm-optimized-routines/math/test/mathbench.c
@@ -1,10 +1,23 @@
/*
* Microbenchmark for math functions.
*
- * Copyright (c) 2018-2020, Arm Limited.
- * SPDX-License-Identifier: MIT
+ * Copyright (c) 2018-2024, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
*/
+#if WANT_SVE_TESTS
+# if __aarch64__ && __linux__
+# ifdef __clang__
+# pragma clang attribute push(__attribute__((target("sve"))), \
+ apply_to = any(function))
+# else
+# pragma GCC target("+sve")
+# endif
+# else
+# error "SVE not supported - please disable WANT_SVE_TESTS"
+# endif
+#endif
+
#undef _GNU_SOURCE
#define _GNU_SOURCE 1
#include <stdint.h>
@@ -15,11 +28,6 @@
#include <math.h>
#include "mathlib.h"
-#ifndef WANT_VMATH
-/* Enable the build of vector math code. */
-# define WANT_VMATH 1
-#endif
-
/* Number of measurements, best result is reported. */
#define MEASURE 60
/* Array size. */
@@ -34,50 +42,6 @@ static float Af[N];
static long measurecount = MEASURE;
static long itercount = ITER;
-#if __aarch64__ && WANT_VMATH
-typedef __f64x2_t v_double;
-
-#define v_double_len() 2
-
-static inline v_double
-v_double_load (const double *p)
-{
- return (v_double){p[0], p[1]};
-}
-
-static inline v_double
-v_double_dup (double x)
-{
- return (v_double){x, x};
-}
-
-typedef __f32x4_t v_float;
-
-#define v_float_len() 4
-
-static inline v_float
-v_float_load (const float *p)
-{
- return (v_float){p[0], p[1], p[2], p[3]};
-}
-
-static inline v_float
-v_float_dup (float x)
-{
- return (v_float){x, x, x, x};
-}
-#else
-/* dummy definitions to make things compile. */
-typedef double v_double;
-typedef float v_float;
-#define v_double_len(x) 1
-#define v_double_load(x) (x)[0]
-#define v_double_dup(x) (x)
-#define v_float_len(x) 1
-#define v_float_load(x) (x)[0]
-#define v_float_dup(x) (x)
-#endif
-
static double
dummy (double x)
{
@@ -89,128 +53,35 @@ dummyf (float x)
{
return x;
}
-
-#if WANT_VMATH
-#if __aarch64__
-static v_double
-__v_dummy (v_double x)
+#if __aarch64__ && __linux__
+__vpcs static float64x2_t
+__vn_dummy (float64x2_t x)
{
return x;
}
-static v_float
-__v_dummyf (v_float x)
+__vpcs static float32x4_t
+__vn_dummyf (float32x4_t x)
{
return x;
}
-
-#ifdef __vpcs
-__vpcs static v_double
-__vn_dummy (v_double x)
+#endif
+#if WANT_SVE_TESTS
+static svfloat64_t
+__sv_dummy (svfloat64_t x, svbool_t pg)
{
return x;
}
-__vpcs static v_float
-__vn_dummyf (v_float x)
+static svfloat32_t
+__sv_dummyf (svfloat32_t x, svbool_t pg)
{
return x;
}
-__vpcs static v_float
-xy__vn_powf (v_float x)
-{
- return __vn_powf (x, x);
-}
-
-__vpcs static v_float
-xy_Z_powf (v_float x)
-{
- return _ZGVnN4vv_powf (x, x);
-}
-
-__vpcs static v_double
-xy__vn_pow (v_double x)
-{
- return __vn_pow (x, x);
-}
-
-__vpcs static v_double
-xy_Z_pow (v_double x)
-{
- return _ZGVnN2vv_pow (x, x);
-}
#endif
-static v_float
-xy__v_powf (v_float x)
-{
- return __v_powf (x, x);
-}
-
-static v_double
-xy__v_pow (v_double x)
-{
- return __v_pow (x, x);
-}
-#endif
-
-static float
-xy__s_powf (float x)
-{
- return __s_powf (x, x);
-}
-
-static double
-xy__s_pow (double x)
-{
- return __s_pow (x, x);
-}
-#endif
-
-static double
-xypow (double x)
-{
- return pow (x, x);
-}
-
-static float
-xypowf (float x)
-{
- return powf (x, x);
-}
-
-static double
-xpow (double x)
-{
- return pow (x, 23.4);
-}
-
-static float
-xpowf (float x)
-{
- return powf (x, 23.4f);
-}
-
-static double
-ypow (double x)
-{
- return pow (2.34, x);
-}
-
-static float
-ypowf (float x)
-{
- return powf (2.34f, x);
-}
-
-static float
-sincosf_wrap (float x)
-{
- float s, c;
- sincosf (x, &s, &c);
- return s + c;
-}
+#include "test/mathbench_wrappers.h"
static const struct fun
{
@@ -223,127 +94,42 @@ static const struct fun
{
double (*d) (double);
float (*f) (float);
- v_double (*vd) (v_double);
- v_float (*vf) (v_float);
-#ifdef __vpcs
- __vpcs v_double (*vnd) (v_double);
- __vpcs v_float (*vnf) (v_float);
+#if __aarch64__ && __linux__
+ __vpcs float64x2_t (*vnd) (float64x2_t);
+ __vpcs float32x4_t (*vnf) (float32x4_t);
+#endif
+#if WANT_SVE_TESTS
+ svfloat64_t (*svd) (svfloat64_t, svbool_t);
+ svfloat32_t (*svf) (svfloat32_t, svbool_t);
#endif
} fun;
} funtab[] = {
+// clang-format off
#define D(func, lo, hi) {#func, 'd', 0, lo, hi, {.d = func}},
#define F(func, lo, hi) {#func, 'f', 0, lo, hi, {.f = func}},
-#define VD(func, lo, hi) {#func, 'd', 'v', lo, hi, {.vd = func}},
-#define VF(func, lo, hi) {#func, 'f', 'v', lo, hi, {.vf = func}},
#define VND(func, lo, hi) {#func, 'd', 'n', lo, hi, {.vnd = func}},
#define VNF(func, lo, hi) {#func, 'f', 'n', lo, hi, {.vnf = func}},
+#define SVD(func, lo, hi) {#func, 'd', 's', lo, hi, {.svd = func}},
+#define SVF(func, lo, hi) {#func, 'f', 's', lo, hi, {.svf = func}},
D (dummy, 1.0, 2.0)
-D (exp, -9.9, 9.9)
-D (exp, 0.5, 1.0)
-D (exp2, -9.9, 9.9)
-D (log, 0.01, 11.1)
-D (log, 0.999, 1.001)
-D (log2, 0.01, 11.1)
-D (log2, 0.999, 1.001)
-{"pow", 'd', 0, 0.01, 11.1, {.d = xypow}},
-D (xpow, 0.01, 11.1)
-D (ypow, -9.9, 9.9)
-D (erf, -6.0, 6.0)
-
F (dummyf, 1.0, 2.0)
-F (expf, -9.9, 9.9)
-F (exp2f, -9.9, 9.9)
-F (logf, 0.01, 11.1)
-F (log2f, 0.01, 11.1)
-{"powf", 'f', 0, 0.01, 11.1, {.f = xypowf}},
-F (xpowf, 0.01, 11.1)
-F (ypowf, -9.9, 9.9)
-{"sincosf", 'f', 0, 0.1, 0.7, {.f = sincosf_wrap}},
-{"sincosf", 'f', 0, 0.8, 3.1, {.f = sincosf_wrap}},
-{"sincosf", 'f', 0, -3.1, 3.1, {.f = sincosf_wrap}},
-{"sincosf", 'f', 0, 3.3, 33.3, {.f = sincosf_wrap}},
-{"sincosf", 'f', 0, 100, 1000, {.f = sincosf_wrap}},
-{"sincosf", 'f', 0, 1e6, 1e32, {.f = sincosf_wrap}},
-F (sinf, 0.1, 0.7)
-F (sinf, 0.8, 3.1)
-F (sinf, -3.1, 3.1)
-F (sinf, 3.3, 33.3)
-F (sinf, 100, 1000)
-F (sinf, 1e6, 1e32)
-F (cosf, 0.1, 0.7)
-F (cosf, 0.8, 3.1)
-F (cosf, -3.1, 3.1)
-F (cosf, 3.3, 33.3)
-F (cosf, 100, 1000)
-F (cosf, 1e6, 1e32)
-F (erff, -4.0, 4.0)
-#if WANT_VMATH
-D (__s_sin, -3.1, 3.1)
-D (__s_cos, -3.1, 3.1)
-D (__s_exp, -9.9, 9.9)
-D (__s_log, 0.01, 11.1)
-{"__s_pow", 'd', 0, 0.01, 11.1, {.d = xy__s_pow}},
-F (__s_expf, -9.9, 9.9)
-F (__s_expf_1u, -9.9, 9.9)
-F (__s_exp2f, -9.9, 9.9)
-F (__s_exp2f_1u, -9.9, 9.9)
-F (__s_logf, 0.01, 11.1)
-{"__s_powf", 'f', 0, 0.01, 11.1, {.f = xy__s_powf}},
-F (__s_sinf, -3.1, 3.1)
-F (__s_cosf, -3.1, 3.1)
-#if __aarch64__
-VD (__v_dummy, 1.0, 2.0)
-VD (__v_sin, -3.1, 3.1)
-VD (__v_cos, -3.1, 3.1)
-VD (__v_exp, -9.9, 9.9)
-VD (__v_log, 0.01, 11.1)
-{"__v_pow", 'd', 'v', 0.01, 11.1, {.vd = xy__v_pow}},
-VF (__v_dummyf, 1.0, 2.0)
-VF (__v_expf, -9.9, 9.9)
-VF (__v_expf_1u, -9.9, 9.9)
-VF (__v_exp2f, -9.9, 9.9)
-VF (__v_exp2f_1u, -9.9, 9.9)
-VF (__v_logf, 0.01, 11.1)
-{"__v_powf", 'f', 'v', 0.01, 11.1, {.vf = xy__v_powf}},
-VF (__v_sinf, -3.1, 3.1)
-VF (__v_cosf, -3.1, 3.1)
-#ifdef __vpcs
+#if __aarch64__ && __linux__
VND (__vn_dummy, 1.0, 2.0)
-VND (__vn_exp, -9.9, 9.9)
-VND (_ZGVnN2v_exp, -9.9, 9.9)
-VND (__vn_log, 0.01, 11.1)
-VND (_ZGVnN2v_log, 0.01, 11.1)
-{"__vn_pow", 'd', 'n', 0.01, 11.1, {.vnd = xy__vn_pow}},
-{"_ZGVnN2vv_pow", 'd', 'n', 0.01, 11.1, {.vnd = xy_Z_pow}},
-VND (__vn_sin, -3.1, 3.1)
-VND (_ZGVnN2v_sin, -3.1, 3.1)
-VND (__vn_cos, -3.1, 3.1)
-VND (_ZGVnN2v_cos, -3.1, 3.1)
VNF (__vn_dummyf, 1.0, 2.0)
-VNF (__vn_expf, -9.9, 9.9)
-VNF (_ZGVnN4v_expf, -9.9, 9.9)
-VNF (__vn_expf_1u, -9.9, 9.9)
-VNF (__vn_exp2f, -9.9, 9.9)
-VNF (_ZGVnN4v_exp2f, -9.9, 9.9)
-VNF (__vn_exp2f_1u, -9.9, 9.9)
-VNF (__vn_logf, 0.01, 11.1)
-VNF (_ZGVnN4v_logf, 0.01, 11.1)
-{"__vn_powf", 'f', 'n', 0.01, 11.1, {.vnf = xy__vn_powf}},
-{"_ZGVnN4vv_powf", 'f', 'n', 0.01, 11.1, {.vnf = xy_Z_powf}},
-VNF (__vn_sinf, -3.1, 3.1)
-VNF (_ZGVnN4v_sinf, -3.1, 3.1)
-VNF (__vn_cosf, -3.1, 3.1)
-VNF (_ZGVnN4v_cosf, -3.1, 3.1)
-#endif
#endif
+#if WANT_SVE_TESTS
+SVD (__sv_dummy, 1.0, 2.0)
+SVF (__sv_dummyf, 1.0, 2.0)
#endif
+#include "test/mathbench_funcs.h"
{0},
#undef F
#undef D
-#undef VF
-#undef VD
#undef VNF
#undef VND
+#undef SVF
+#undef SVD
+ // clang-format on
};
static void
@@ -442,69 +228,77 @@ runf_latency (float f (float))
prev = f (Af[i] + prev * z);
}
+#if __aarch64__ && __linux__
static void
-run_v_thruput (v_double f (v_double))
+run_vn_thruput (__vpcs float64x2_t f (float64x2_t))
{
- for (int i = 0; i < N; i += v_double_len ())
- f (v_double_load (A+i));
+ for (int i = 0; i < N; i += 2)
+ f (vld1q_f64 (A + i));
}
static void
-runf_v_thruput (v_float f (v_float))
+runf_vn_thruput (__vpcs float32x4_t f (float32x4_t))
{
- for (int i = 0; i < N; i += v_float_len ())
- f (v_float_load (Af+i));
+ for (int i = 0; i < N; i += 4)
+ f (vld1q_f32 (Af + i));
}
static void
-run_v_latency (v_double f (v_double))
+run_vn_latency (__vpcs float64x2_t f (float64x2_t))
{
- v_double z = v_double_dup (zero);
- v_double prev = z;
- for (int i = 0; i < N; i += v_double_len ())
- prev = f (v_double_load (A+i) + prev * z);
+ volatile uint64x2_t vsel = (uint64x2_t) { 0, 0 };
+ uint64x2_t sel = vsel;
+ float64x2_t prev = vdupq_n_f64 (0);
+ for (int i = 0; i < N; i += 2)
+ prev = f (vbslq_f64 (sel, prev, vld1q_f64 (A + i)));
}
static void
-runf_v_latency (v_float f (v_float))
+runf_vn_latency (__vpcs float32x4_t f (float32x4_t))
{
- v_float z = v_float_dup (zero);
- v_float prev = z;
- for (int i = 0; i < N; i += v_float_len ())
- prev = f (v_float_load (Af+i) + prev * z);
+ volatile uint32x4_t vsel = (uint32x4_t) { 0, 0, 0, 0 };
+ uint32x4_t sel = vsel;
+ float32x4_t prev = vdupq_n_f32 (0);
+ for (int i = 0; i < N; i += 4)
+ prev = f (vbslq_f32 (sel, prev, vld1q_f32 (Af + i)));
}
+#endif
-#ifdef __vpcs
+#if WANT_SVE_TESTS
static void
-run_vn_thruput (__vpcs v_double f (v_double))
+run_sv_thruput (svfloat64_t f (svfloat64_t, svbool_t))
{
- for (int i = 0; i < N; i += v_double_len ())
- f (v_double_load (A+i));
+ for (int i = 0; i < N; i += svcntd ())
+ f (svld1_f64 (svptrue_b64 (), A + i), svptrue_b64 ());
}
static void
-runf_vn_thruput (__vpcs v_float f (v_float))
+runf_sv_thruput (svfloat32_t f (svfloat32_t, svbool_t))
{
- for (int i = 0; i < N; i += v_float_len ())
- f (v_float_load (Af+i));
+ for (int i = 0; i < N; i += svcntw ())
+ f (svld1_f32 (svptrue_b32 (), Af + i), svptrue_b32 ());
}
static void
-run_vn_latency (__vpcs v_double f (v_double))
+run_sv_latency (svfloat64_t f (svfloat64_t, svbool_t))
{
- v_double z = v_double_dup (zero);
- v_double prev = z;
- for (int i = 0; i < N; i += v_double_len ())
- prev = f (v_double_load (A+i) + prev * z);
+ volatile svbool_t vsel = svptrue_b64 ();
+ svbool_t sel = vsel;
+ svfloat64_t prev = svdup_f64 (0);
+ for (int i = 0; i < N; i += svcntd ())
+ prev = f (svsel_f64 (sel, svld1_f64 (svptrue_b64 (), A + i), prev),
+ svptrue_b64 ());
}
static void
-runf_vn_latency (__vpcs v_float f (v_float))
+runf_sv_latency (svfloat32_t f (svfloat32_t, svbool_t))
{
- v_float z = v_float_dup (zero);
- v_float prev = z;
- for (int i = 0; i < N; i += v_float_len ())
- prev = f (v_float_load (Af+i) + prev * z);
+ volatile svbool_t vsel = svptrue_b32 ();
+ svbool_t sel = vsel;
+ svfloat32_t prev = svdup_f32 (0);
+ for (int i = 0; i < N; i += svcntw ())
+ prev = f (svsel_f32 (sel, svld1_f32 (svptrue_b32 (), Af + i), prev),
+ svptrue_b32 ());
}
#endif
@@ -512,7 +306,11 @@ static uint64_t
tic (void)
{
struct timespec ts;
+#if defined(_MSC_VER)
+ if (!timespec_get (&ts, TIME_UTC))
+#else
if (clock_gettime (CLOCK_REALTIME, &ts))
+#endif
abort ();
return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
}
@@ -539,10 +337,12 @@ bench1 (const struct fun *f, int type, double lo, double hi)
const char *s = type == 't' ? "rthruput" : "latency";
int vlen = 1;
- if (f->vec && f->prec == 'd')
- vlen = v_double_len();
- else if (f->vec && f->prec == 'f')
- vlen = v_float_len();
+ if (f->vec == 'n')
+ vlen = f->prec == 'd' ? 2 : 4;
+#if WANT_SVE_TESTS
+ else if (f->vec == 's')
+ vlen = f->prec == 'd' ? svcntd () : svcntw ();
+#endif
if (f->prec == 'd' && type == 't' && f->vec == 0)
TIMEIT (run_thruput, f->fun.d);
@@ -552,15 +352,7 @@ bench1 (const struct fun *f, int type, double lo, double hi)
TIMEIT (runf_thruput, f->fun.f);
else if (f->prec == 'f' && type == 'l' && f->vec == 0)
TIMEIT (runf_latency, f->fun.f);
- else if (f->prec == 'd' && type == 't' && f->vec == 'v')
- TIMEIT (run_v_thruput, f->fun.vd);
- else if (f->prec == 'd' && type == 'l' && f->vec == 'v')
- TIMEIT (run_v_latency, f->fun.vd);
- else if (f->prec == 'f' && type == 't' && f->vec == 'v')
- TIMEIT (runf_v_thruput, f->fun.vf);
- else if (f->prec == 'f' && type == 'l' && f->vec == 'v')
- TIMEIT (runf_v_latency, f->fun.vf);
-#ifdef __vpcs
+#if __aarch64__ && __linux__
else if (f->prec == 'd' && type == 't' && f->vec == 'n')
TIMEIT (run_vn_thruput, f->fun.vnd);
else if (f->prec == 'd' && type == 'l' && f->vec == 'n')
@@ -570,20 +362,32 @@ bench1 (const struct fun *f, int type, double lo, double hi)
else if (f->prec == 'f' && type == 'l' && f->vec == 'n')
TIMEIT (runf_vn_latency, f->fun.vnf);
#endif
+#if WANT_SVE_TESTS
+ else if (f->prec == 'd' && type == 't' && f->vec == 's')
+ TIMEIT (run_sv_thruput, f->fun.svd);
+ else if (f->prec == 'd' && type == 'l' && f->vec == 's')
+ TIMEIT (run_sv_latency, f->fun.svd);
+ else if (f->prec == 'f' && type == 't' && f->vec == 's')
+ TIMEIT (runf_sv_thruput, f->fun.svf);
+ else if (f->prec == 'f' && type == 'l' && f->vec == 's')
+ TIMEIT (runf_sv_latency, f->fun.svf);
+#endif
if (type == 't')
{
ns100 = (100 * dt + itercount * N / 2) / (itercount * N);
- printf ("%9s %8s: %4u.%02u ns/elem %10llu ns in [%g %g]\n", f->name, s,
+ printf ("%9s %8s: %4u.%02u ns/elem %10llu ns in [%g %g] vlen %d\n",
+ f->name, s,
(unsigned) (ns100 / 100), (unsigned) (ns100 % 100),
- (unsigned long long) dt, lo, hi);
+ (unsigned long long) dt, lo, hi, vlen);
}
else if (type == 'l')
{
ns100 = (100 * dt + itercount * N / vlen / 2) / (itercount * N / vlen);
- printf ("%9s %8s: %4u.%02u ns/call %10llu ns in [%g %g]\n", f->name, s,
+ printf ("%9s %8s: %4u.%02u ns/call %10llu ns in [%g %g] vlen %d\n",
+ f->name, s,
(unsigned) (ns100 / 100), (unsigned) (ns100 % 100),
- (unsigned long long) dt, lo, hi);
+ (unsigned long long) dt, lo, hi, vlen);
}
fflush (stdout);
}
@@ -771,3 +575,7 @@ main (int argc, char *argv[])
}
return 0;
}
+
+#if __aarch64__ && __linux__ && WANT_SVE_TESTS && defined(__clang__)
+# pragma clang attribute pop
+#endif
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-14 03:55:11 +0000