niansa-misc/glibc
1
0
Fork 0
mirror of git://sourceware.org/git/glibc.git synced 2025-03-06 20:58:33 +01:00
Code Issues Projects Releases Packages Wiki Activity

aarch64: Fix libmvec benchmarks

Browse source 
These were broken by the new atan2 functions, as they were only
set up for univariate functions. Arity is now detected from the
input file - this revealed a mistake that the double-precision
inputs were being used for both single- and double-precision
routines, which is now remedied.
This commit is contained in:
Joe Ramsay 2023-11-21 14:39:39 +00:00 • committed by Szabolcs Nagy
parent 5d7f1bce7d
commit bd70d3bacf
2 changed files with 78 additions and 46 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

62
sysdeps/aarch64/fpu/scripts/bench_libmvec_advsimd.py
View file

@ -22,40 +22,49 @@ TEMPLATE = """
#include <math.h> #include <math.h>
#include <arm_neon.h> #include <arm_neon.h>
#define STRIDE {stride} #define STRIDE {rowlen}
#define CALL_BENCH_FUNC(v, i) (__extension__ ({{ \\ #define CALL_BENCH_FUNC_1(v, i) (__extension__ ({{ \\
{rtype} mx0 = {fname}(vld1q_f{prec_short} (variants[v].in[i].arg0)); \\ {rtype} mx0 = {fname}(vld1q_f{prec_short} (&variants[v].in->arg0[i * STRIDE])); \\
mx0; }})) mx0; }}))
struct args #define CALL_BENCH_FUNC_2(v, i) (__extension__ ({{ \\
{rtype} mx0 = {fname}(vld1q_f{prec_short} (&variants[v].in->arg0[i * STRIDE]), \\
vld1q_f{prec_short} (&variants[v].in->arg1[i * STRIDE])); \\
mx0; }}))
struct args_1
{{ {{
{stype} arg0[STRIDE]; {stype} arg0[{nelems}];
double timing; }};
struct args_2
{{
{stype} arg0[{nelems}];
{stype} arg1[{nelems}];
}}; }};
struct _variants struct _variants
{{ {{
const char *name; const char *name;
int count; const struct args_{arity} *in;
const struct args *in;
}}; }};
static const struct args in0[{rowcount}] = {{ static const struct args_{arity} in0 = {{
{in_data} {in_data}
}}; }};
static const struct _variants variants[1] = {{ static const struct _variants variants[1] = {{
{{"", {rowcount}, in0}}, {{"", &in0}},
}}; }};
#define NUM_VARIANTS 1 #define NUM_VARIANTS 1
#define NUM_SAMPLES(i) (variants[i].count) #define NUM_SAMPLES(i) ({nelems} / STRIDE)
#define VARIANT(i) (variants[i].name) #define VARIANT(i) (variants[i].name)
static {rtype} volatile ret; static {rtype} volatile ret;
#define BENCH_FUNC(i, j) ({{ ret = CALL_BENCH_FUNC(i, j); }}) #define BENCH_FUNC(i, j) ({{ ret = CALL_BENCH_FUNC_{arity}(i, j); }})
#define FUNCNAME "{fname}" #define FUNCNAME "{fname}"
#include <bench-libmvec-skeleton.c> #include <bench-libmvec-skeleton.c>
""" """
@ -63,27 +72,34 @@ static {rtype} volatile ret;
def main(name): def main(name):
_, prec, _, func = name.split("-") _, prec, _, func = name.split("-")
scalar_to_advsimd_type = {"double": "float64x2_t", "float": "float32x4_t"} scalar_to_advsimd_type = {"double": "float64x2_t", "float": "float32x4_t"}
rowlen = {"double": 2, "float": 4}[prec]
stride = {"double": 2, "float": 4}[prec]
rtype = scalar_to_advsimd_type[prec] rtype = scalar_to_advsimd_type[prec]
atype = scalar_to_advsimd_type[prec] atype = scalar_to_advsimd_type[prec]
fname = f"_ZGVnN{stride}v_{func}{'f' if prec == 'float' else ''}"
prec_short = {"double": 64, "float": 32}[prec] prec_short = {"double": 64, "float": 32}[prec]
input_filename = {"double": f"{func}-inputs", "float": f"{func}f-inputs"}[prec]
with open(f"../benchtests/libmvec/{func}-inputs") as f: with open(f"../benchtests/libmvec/{input_filename}") as f:
in_vals = [l.strip() for l in f.readlines() if l and not l.startswith("#")] input_file = f.readlines()
in_vals = [in_vals[i:i+stride] for i in range(0, len(in_vals), stride)] in_vals = (l.strip() for l in input_file if l and not l.startswith("#"))
rowcount= len(in_vals) # Split in case of multivariate signature
in_data = ",\n".join("{{" + ", ".join(row) + "}, 0}" for row in in_vals) in_vals = (l.split(", ") for l in in_vals)
# Transpose
in_vals = list(zip(*in_vals))
in_data = ",\n".join("{" + (", ".join(val for val in col) + "}")
for col in in_vals)
print(TEMPLATE.format(stride=stride, arity = [l for l in input_file if l.startswith("## args: ")][0].count(prec)
fname = f"_ZGVnN{rowlen}{'v' * arity}_{func}{'f' if prec == 'float' else ''}"
print(TEMPLATE.format(rowlen=rowlen,
rtype=rtype, rtype=rtype,
atype=atype, atype=atype,
fname=fname, fname=fname,
prec_short=prec_short, prec_short=prec_short,
in_data=in_data, in_data=in_data,
rowcount=rowcount, stype=prec,
stype=prec)) arity=arity,
nelems=len(in_vals[0])))
if __name__ == "__main__": if __name__ == "__main__":

62
sysdeps/aarch64/fpu/scripts/bench_libmvec_sve.py
View file

@ -22,46 +22,55 @@ TEMPLATE = """
#include <math.h> #include <math.h>
#include <arm_sve.h> #include <arm_sve.h>
#define MAX_STRIDE {max_stride}
#define STRIDE {stride} #define STRIDE {stride}
#define PTRUE svptrue_b{prec_short} #define PTRUE svptrue_b{prec_short}
#define SV_LOAD svld1_f{prec_short} #define SV_LOAD svld1_f{prec_short}
#define SV_STORE svst1_f{prec_short} #define SV_STORE svst1_f{prec_short}
#define REQUIRE_SVE #define REQUIRE_SVE
#define CALL_BENCH_FUNC(v, i) (__extension__ ({{ \\ #define CALL_BENCH_FUNC_1(v, i) (__extension__ ({{ \\
{rtype} mx0 = {fname}(SV_LOAD (PTRUE(), variants[v].in[i].arg0), PTRUE()); \\ {rtype} mx0 = {fname}(SV_LOAD (PTRUE(), &variants[v].in->arg0[i * STRIDE]), PTRUE()); \\
mx0; }})) mx0; }}))
struct args #define CALL_BENCH_FUNC_2(v, i) (__extension__ ({{ \\
{rtype} mx0 = {fname}(SV_LOAD (PTRUE(), &variants[v].in->arg0[i * STRIDE]), \\
SV_LOAD (PTRUE(), &variants[v].in->arg1[i * STRIDE]), \\
PTRUE()); \\
mx0; }}))
struct args_1
{{ {{
{stype} arg0[MAX_STRIDE]; {stype} arg0[{nelems}];
double timing; }};
struct args_2
{{
{stype} arg0[{nelems}];
{stype} arg1[{nelems}];
}}; }};
struct _variants struct _variants
{{ {{
const char *name; const char *name;
int count; const struct args_{arity} *in;
const struct args *in;
}}; }};
static const struct args in0[{rowcount}] = {{ static const struct args_{arity} in0 = {{
{in_data} {in_data}
}}; }};
static const struct _variants variants[1] = {{ static const struct _variants variants[1] = {{
{{"", {rowcount}, in0}}, {{"", &in0}},
}}; }};
#define NUM_VARIANTS 1 #define NUM_VARIANTS 1
#define NUM_SAMPLES(i) (variants[i].count) #define NUM_SAMPLES(i) ({nelems} / STRIDE)
#define VARIANT(i) (variants[i].name) #define VARIANT(i) (variants[i].name)
// Cannot pass volatile pointer to svst1. This still does not appear to get optimised out. // Cannot pass volatile pointer to svst1. This still does not appear to get optimised out.
static {stype} /*volatile*/ ret[MAX_STRIDE]; static {stype} /*volatile*/ ret[{rowlen}];
#define BENCH_FUNC(i, j) ({{ SV_STORE(PTRUE(), ret, CALL_BENCH_FUNC(i, j)); }}) #define BENCH_FUNC(i, j) ({{ SV_STORE(PTRUE(), ret, CALL_BENCH_FUNC_{arity}(i, j)); }})
#define FUNCNAME "{fname}" #define FUNCNAME "{fname}"
#include <bench-libmvec-skeleton.c> #include <bench-libmvec-skeleton.c>
""" """
@ -69,23 +78,29 @@ static {stype} /*volatile*/ ret[MAX_STRIDE];
def main(name): def main(name):
_, prec, _, func = name.split("-") _, prec, _, func = name.split("-")
scalar_to_sve_type = {"double": "svfloat64_t", "float": "svfloat32_t"} scalar_to_sve_type = {"double": "svfloat64_t", "float": "svfloat32_t"}
stride = {"double": "svcntd()", "float": "svcntw()"}[prec] stride = {"double": "svcntd()", "float": "svcntw()"}[prec]
rtype = scalar_to_sve_type[prec] rtype = scalar_to_sve_type[prec]
atype = scalar_to_sve_type[prec] atype = scalar_to_sve_type[prec]
fname = f"_ZGVsMxv_{func}{'f' if prec == 'float' else ''}"
prec_short = {"double": 64, "float": 32}[prec] prec_short = {"double": 64, "float": 32}[prec]
# Max SVE vector length is 2048 bits. To ensure benchmarks are # Max SVE vector length is 2048 bits. To ensure benchmarks are
# vector-length-agnostic, but still use as wide vectors as # vector-length-agnostic, but still use as wide vectors as
# possible on any given target, divide input data into 2048-bit # possible on any given target, divide input data into 2048-bit
# rows, then load/store as many elements as the target will allow. # rows, then load/store as many elements as the target will allow.
max_stride = 2048 // prec_short rowlen = {"double": 32, "float": 64}[prec]
input_filename = {"double": f"{func}-inputs", "float": f"{func}f-inputs"}[prec]
with open(f"../benchtests/libmvec/{func}-inputs") as f: with open(f"../benchtests/libmvec/{input_filename}") as f:
in_vals = [l.strip() for l in f.readlines() if l and not l.startswith("#")] input_file = f.readlines()
in_vals = [in_vals[i:i+max_stride] for i in range(0, len(in_vals), max_stride)] in_vals = (l.strip() for l in input_file if l and not l.startswith("#"))
rowcount= len(in_vals) # Split in case of multivariate signature
in_data = ",\n".join("{{" + ", ".join(row) + "}, 0}" for row in in_vals) in_vals = (l.split(", ") for l in in_vals)
# Transpose
in_vals = list(zip(*in_vals))
in_data = ",\n".join("{" + (", ".join(val for val in col) + "}")
for col in in_vals)
arity = [l for l in input_file if l.startswith("## args: ")][0].count(prec)
fname = f"_ZGVsMx{'v' * arity}_{func}{'f' if prec == 'float' else ''}"
print(TEMPLATE.format(stride=stride, print(TEMPLATE.format(stride=stride,
rtype=rtype, rtype=rtype,
@ -93,9 +108,10 @@ def main(name):
fname=fname, fname=fname,
prec_short=prec_short, prec_short=prec_short,
in_data=in_data, in_data=in_data,
rowcount=rowcount,
stype=prec, stype=prec,
max_stride=max_stride)) rowlen=rowlen,
arity=arity,
nelems=len(in_vals[0])))
if __name__ == "__main__": if __name__ == "__main__":