Re: [Qemu-devel] [PATCH v1 01/14] tests: add fp-bench, a collection of simple floating-point microbenchmarks

[Alex Bennée]

Emilio G. Cota <address@hidden> writes:

> This will allow us to measure the performance impact of FP
> emulation optimizations.
>
> Signed-off-by: Emilio G. Cota <address@hidden>
> ---
>  tests/fp-bench.c       | 290 
> +++++++++++++++++++++++++++++++++++++++++++++++++
>  tests/.gitignore       |   1 +
>  tests/Makefile.include |   3 +-
>  3 files changed, 293 insertions(+), 1 deletion(-)
>  create mode 100644 tests/fp-bench.c
>
> diff --git a/tests/fp-bench.c b/tests/fp-bench.c
> new file mode 100644
> index 0000000..a782093
> --- /dev/null
> +++ b/tests/fp-bench.c
> @@ -0,0 +1,290 @@
> +/*
> + * fp-bench.c - A collection of simple floating point microbenchmarks.
> + *
> + * Copyright (C) 2018, Emilio G. Cota <address@hidden>
> + *
> + * License: GNU GPL, version 2 or later.
> + *   See the COPYING file in the top-level directory.
> + */
> +#include "qemu/osdep.h"
> +#include "qemu/atomic.h"
> +
> +#include <math.h>
> +
> +#include <sys/time.h>
> +#include <stdint.h>
> +#include <stdlib.h>
> +#include <unistd.h>
> +#include <stdio.h>
> +#include <time.h>
> +
> +/* amortize the computation of random inputs */
> +#define OPS_PER_ITER     (1000ULL)
> +
> +#define SEED_A 0xdeadfacedeadface
> +#define SEED_B 0xbadc0feebadc0fee
> +#define SEED_C 0xbeefdeadbeefdead
> +
> +enum op {
> +    OP_ADD,
> +    OP_SUB,
> +    OP_MUL,
> +    OP_DIV,
> +    OP_FMA,
> +    OP_SQRT,
> +};
> +
> +static const char * const op_names[] = {
> +    [OP_ADD] = "add",
> +    [OP_SUB] = "sub",
> +    [OP_MUL] = "mul",
> +    [OP_DIV] = "div",
> +    [OP_FMA] = "fma",
> +    [OP_SQRT] = "sqrt",
> +};
> +
> +static uint64_t n_ops = 10000000;
> +static enum op op;
> +static const char *precision = "float";
> +
> +static const char commands_string[] =
> +    " -n = number of floating point operations\n"
> +    " -o = floating point operation (add, sub, mul, div, fma, sqrt). 
> Default: add\n"
> +    " -p = precision (float|single, double). Default: float";
> +
> +static void usage_complete(int argc, char *argv[])
> +{
> +    fprintf(stderr, "Usage: %s [options]\n", argv[0]);
> +    fprintf(stderr, "options:\n%s\n", commands_string);
> +    exit(-1);
> +}
> +
> +static void set_op(const char *name)
> +{
> +    int i;
> +
> +    for (i = 0; i < ARRAY_SIZE(op_names); i++) {
> +        if (strcmp(name, op_names[i]) == 0) {
> +            op = i;
> +            return;
> +        }
> +    }
> +    fprintf(stderr, "Unsupported op '%s'\n", name);
> +    exit(EXIT_FAILURE);
> +}
> +
> +static inline int64_t get_clock_realtime(void)
> +{
> +    struct timeval tv;
> +
> +    gettimeofday(&tv, NULL);
> +    return tv.tv_sec * 1000000000LL + (tv.tv_usec * 1000);
> +}
> +
> +/*
> + * From: https://en.wikipedia.org/wiki/Xorshift
> + * This is faster than rand_r(), and gives us a wider range (RAND_MAX is only
> + * guaranteed to be >= INT_MAX).
> + */
> +static uint64_t xorshift64star(uint64_t x)
> +{
> +    x ^= x >> 12; /* a */
> +    x ^= x << 25; /* b */
> +    x ^= x >> 27; /* c */
> +    return x * UINT64_C(2685821657736338717);
> +}
> +
> +static inline bool u32_is_normal(uint32_t x)
> +{
> +    return ((x + 0x00800000) & 0x7fffffff) >= 0x01000000;
> +}
> +
> +static inline bool u64_is_normal(uint64_t x)
> +{
> +    return ((x + (1ULL << 52)) & -1ULL >> 1) >= 1ULL << 53;
> +}
> +
> +static inline float get_random_float(uint64_t *x)
> +{
> +    uint64_t r = *x;
> +    uint32_t r32;
> +
> +    do {
> +        r = xorshift64star(r);
> +    } while (!u32_is_normal(r));
> +    *x = r;
> +    r32 = r;
> +    return *(float *)&r32;
> +}
> +
> +static inline double get_random_double(uint64_t *x)
> +{
> +    uint64_t r = *x;
> +
> +    do {
> +        r = xorshift64star(r);
> +    } while (!u64_is_normal(r));
> +    *x = r;
> +    return *(double *)&r;
> +}
> +
> +/*
> + * Disable optimizations (e.g. "a OP b" outside of the inner loop) with
> + * volatile.
> + */
> +#define GEN_BENCH_1OPF(NAME, FUNC, PRECISION)                           \
> +    static void NAME(volatile PRECISION *res)                           \
> +    {                                                                   \
> +        uint64_t ra = SEED_A;                                           \
> +        uint64_t i, j;                                                  \
> +                                                                        \
> +        for (i = 0; i < n_ops; i += OPS_PER_ITER) {                     \
> +            volatile PRECISION a = glue(get_random_, PRECISION)(&ra);   \
> +                                                                        \
> +            for (j = 0; j < OPS_PER_ITER; j++) {                        \
> +                *res = FUNC(a);                                         \
> +            }                                                           \
> +        }                                                               \
> +    }
> +

Have you had a chance to look at if this will vectorise? I have a
similar benchmark which I compile with multiple options to test normal,
NEON/AdvSIMD and SVE enabled loops.

> +GEN_BENCH_1OPF(bench_float_sqrt, sqrtf, float)
> +GEN_BENCH_1OPF(bench_double_sqrt, sqrt, double)
> +#undef GEN_BENCH_1OPF
> +
> +#define GEN_BENCH_2OP(NAME, OP, PRECISION)                              \
> +    static void NAME(volatile PRECISION *res)                           \
> +    {                                                                   \
> +        uint64_t ra = SEED_A;                                           \
> +        uint64_t rb = SEED_B;                                           \
> +        uint64_t i, j;                                                  \
> +                                                                        \
> +        for (i = 0; i < n_ops; i += OPS_PER_ITER) {                     \
> +            volatile PRECISION a = glue(get_random_, PRECISION)(&ra);   \
> +            volatile PRECISION b = glue(get_random_, PRECISION)(&rb);   \
> +                                                                        \
> +            for (j = 0; j < OPS_PER_ITER; j++) {                        \
> +                *res = a OP b;                                          \
> +            }                                                           \
> +        }                                                               \
> +    }
> +
> +GEN_BENCH_2OP(bench_float_add, +, float)
> +GEN_BENCH_2OP(bench_float_sub, -, float)
> +GEN_BENCH_2OP(bench_float_mul, *, float)
> +GEN_BENCH_2OP(bench_float_div, /, float)
> +
> +GEN_BENCH_2OP(bench_double_add, +, double)
> +GEN_BENCH_2OP(bench_double_sub, -, double)
> +GEN_BENCH_2OP(bench_double_mul, *, double)
> +GEN_BENCH_2OP(bench_double_div, /, double)
> +
> +#define GEN_BENCH_3OPF(NAME, FUNC, PRECISION)                           \
> +    static void NAME(volatile PRECISION *res)                           \
> +    {                                                                   \
> +        uint64_t ra = SEED_A;                                           \
> +        uint64_t rb = SEED_B;                                           \
> +        uint64_t rc = SEED_C;                                           \
> +        uint64_t i, j;                                                  \
> +                                                                        \
> +        for (i = 0; i < n_ops; i += OPS_PER_ITER) {                     \
> +            volatile PRECISION a = glue(get_random_, PRECISION)(&ra);   \
> +            volatile PRECISION b = glue(get_random_, PRECISION)(&rb);   \
> +            volatile PRECISION c = glue(get_random_, PRECISION)(&rc);   \
> +                                                                        \
> +            for (j = 0; j < OPS_PER_ITER; j++) {                        \
> +                *res = FUNC(a, b, c);                                   \
> +            }                                                           \
> +        }                                                               \
> +    }
> +
> +GEN_BENCH_3OPF(bench_float_fma, fmaf, float)
> +GEN_BENCH_3OPF(bench_double_fma, fma, double)
> +#undef GEN_BENCH_3OPF
> +
> +static void parse_args(int argc, char *argv[])
> +{
> +    int c;
> +
> +    for (;;) {
> +        c = getopt(argc, argv, "n:ho:p:");
> +        if (c < 0) {
> +            break;
> +        }
> +        switch (c) {
> +        case 'h':
> +            usage_complete(argc, argv);
> +            exit(0);
> +        case 'n':
> +            n_ops = atoll(optarg);
> +            if (n_ops < OPS_PER_ITER) {
> +                n_ops = OPS_PER_ITER;
> +            }
> +            n_ops -= n_ops % OPS_PER_ITER;
> +            break;
> +        case 'o':
> +            set_op(optarg);
> +            break;
> +        case 'p':
> +            precision = optarg;
> +            if (strcmp(precision, "float") &&
> +                strcmp(precision, "single") &&
> +                strcmp(precision, "double")) {
> +                fprintf(stderr, "Unsupported precision '%s'\n", precision);
> +                exit(EXIT_FAILURE);

Supporting half-precision if the compiler does would also be useful here.

> +            }
> +            break;
> +        }
> +    }
> +}
> +
> +#define CALL_BENCH(OP, PRECISION, RESP)                 \
> +    do {                                                \
> +        switch (OP) {                                   \
> +        case OP_ADD:                                    \
> +            glue(glue(bench_, PRECISION), _add)(RESP);  \
> +            break;                                      \
> +        case OP_SUB:                                    \
> +            glue(glue(bench_, PRECISION), _sub)(RESP);  \
> +            break;                                      \
> +        case OP_MUL:                                    \
> +            glue(glue(bench_, PRECISION), _mul)(RESP);  \
> +            break;                                      \
> +        case OP_DIV:                                    \
> +            glue(glue(bench_, PRECISION), _div)(RESP);  \
> +            break;                                      \
> +        case OP_FMA:                                    \
> +            glue(glue(bench_, PRECISION), _fma)(RESP);  \
> +            break;                                      \
> +        case OP_SQRT:                                   \
> +            glue(glue(bench_, PRECISION), _sqrt)(RESP); \
> +            break;                                      \
> +        default:                                        \
> +            g_assert_not_reached();                     \
> +        }                                               \
> +    } while (0)
> +
> +int main(int argc, char *argv[])
> +{
> +    int64_t t0, t1;
> +    double resd;
> +
> +    parse_args(argc, argv);
> +    if (!strcmp(precision, "float") || !strcmp(precision, "single")) {
> +        float res;
> +        t0 = get_clock_realtime();
> +        CALL_BENCH(op, float, &res);
> +        t1 = get_clock_realtime();
> +        resd = res;
> +    } else if (!strcmp(precision, "double")) {
> +        t0 = get_clock_realtime();
> +        CALL_BENCH(op, double, &resd);
> +        t1 = get_clock_realtime();
> +    } else {
> +        g_assert_not_reached();
> +    }
> +    printf("%.2f MFlops\n", (double)n_ops / (t1 - t0) * 1e3);
> +    if (resd) {
> +        return 0;
> +    }
> +    return 0;
> +}
> diff --git a/tests/.gitignore b/tests/.gitignore
> index 18e58b2..df69175 100644
> --- a/tests/.gitignore
> +++ b/tests/.gitignore
> @@ -12,6 +12,7 @@ check-qobject
>  check-qstring
>  check-qom-interface
>  check-qom-proplist
> +fp-bench
>  qht-bench
>  rcutorture
>  test-aio
> diff --git a/tests/Makefile.include b/tests/Makefile.include
> index ef9b88c..f6121ee 100644
> --- a/tests/Makefile.include
> +++ b/tests/Makefile.include
> @@ -587,7 +587,7 @@ test-obj-y = tests/check-qnum.o tests/check-qstring.o 
> tests/check-qdict.o \
>       tests/rcutorture.o tests/test-rcu-list.o \
>       tests/test-qdist.o tests/test-shift128.o \
>       tests/test-qht.o tests/qht-bench.o tests/test-qht-par.o \
> -     tests/atomic_add-bench.o
> +     tests/atomic_add-bench.o tests/fp-bench.o

Not sure why but "make check" didn't build this. I had to explicitly
"make tests/fp-bench". I guess along with atomic_add_bench though these
are explicitly guest facing tests so maybe we should move them once
tests/tcg is working again. I'll have another run at that this week.

>
>  $(test-obj-y): QEMU_INCLUDES += -Itests
>  QEMU_CFLAGS += -I$(SRC_PATH)/tests
> @@ -639,6 +639,7 @@ tests/test-qht-par$(EXESUF): tests/test-qht-par.o 
> tests/qht-bench$(EXESUF) $(tes
>  tests/qht-bench$(EXESUF): tests/qht-bench.o $(test-util-obj-y)
>  tests/test-bufferiszero$(EXESUF): tests/test-bufferiszero.o 
> $(test-util-obj-y)
>  tests/atomic_add-bench$(EXESUF): tests/atomic_add-bench.o $(test-util-obj-y)
> +tests/fp-bench$(EXESUF): tests/fp-bench.o $(test-util-obj-y)
>
>  tests/test-qdev-global-props$(EXESUF): tests/test-qdev-global-props.o \
>       hw/core/qdev.o hw/core/qdev-properties.o hw/core/hotplug.o\

Anyway for this version:

Reviewed-by: Alex Bennée <address@hidden>

--
Alex Bennée