Path: blob/main/contrib/arm-optimized-routines/string/bench/memcpy.c
39500 views
/*1* memcpy benchmark.2*3* Copyright (c) 2020-2023, Arm Limited.4* SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception5*/67#define _GNU_SOURCE8#include <stdint.h>9#include <stdio.h>10#include <string.h>11#include <assert.h>12#include "stringlib.h"13#include "benchlib.h"1415#define ITERS 500016#define ITERS2 2000000017#define ITERS3 20000018#define NUM_TESTS 1638419#define MIN_SIZE 3276820#define MAX_SIZE (1024 * 1024)2122static uint8_t a[MAX_SIZE + 4096 + 64] __attribute__((__aligned__(4096)));23static uint8_t b[MAX_SIZE + 4096 + 64] __attribute__((__aligned__(4096)));2425#define DOTEST(STR,TESTFN) \26printf (STR); \27RUN (TESTFN, memcpy); \28RUNA64 (TESTFN, __memcpy_aarch64); \29RUNA64 (TESTFN, __memcpy_aarch64_simd); \30RUNSVE (TESTFN, __memcpy_aarch64_sve); \31RUNMOPS (TESTFN, __memcpy_aarch64_mops); \32RUNA32 (TESTFN, __memcpy_arm); \33printf ("\n");3435typedef struct { uint16_t size; uint16_t freq; } freq_data_t;36typedef struct { uint8_t align; uint16_t freq; } align_data_t;3738#define SIZE_NUM 6553639#define SIZE_MASK (SIZE_NUM-1)40static uint8_t size_arr[SIZE_NUM];4142/* Frequency data for memcpy of less than 4096 bytes based on SPEC2017. */43static freq_data_t size_freq[] =44{45{32,22320}, { 16,9554}, { 8,8915}, {152,5327}, { 4,2159}, {292,2035},46{ 12,1608}, { 24,1343}, {1152,895}, {144, 813}, {884, 733}, {284, 721},47{120, 661}, { 2, 649}, {882, 550}, { 5, 475}, { 7, 461}, {108, 460},48{ 10, 361}, { 9, 361}, { 6, 334}, { 3, 326}, {464, 308}, {2048,303},49{ 1, 298}, { 64, 250}, { 11, 197}, {296, 194}, { 68, 187}, { 15, 185},50{192, 184}, {1764,183}, { 13, 173}, {560, 126}, {160, 115}, {288, 96},51{104, 96}, {1144, 83}, { 18, 80}, { 23, 78}, { 40, 77}, { 19, 68},52{ 48, 63}, { 17, 57}, { 72, 54}, {1280, 51}, { 20, 49}, { 28, 47},53{ 22, 46}, {640, 45}, { 25, 41}, { 14, 40}, { 56, 37}, { 27, 35},54{ 35, 33}, {384, 33}, { 29, 32}, { 80, 30}, {4095, 22}, {232, 22},55{ 36, 19}, {184, 17}, { 21, 17}, {256, 16}, { 44, 15}, { 26, 15},56{ 31, 14}, { 88, 14}, {176, 13}, { 33, 12}, {1024, 12}, {208, 11},57{ 62, 11}, {128, 10}, {704, 10}, {324, 10}, { 96, 10}, { 60, 9},58{136, 9}, {124, 9}, { 34, 8}, { 30, 8}, {480, 8}, {1344, 8},59{273, 7}, {520, 7}, {112, 6}, { 52, 6}, {344, 6}, {336, 6},60{504, 5}, {168, 5}, {424, 5}, { 0, 4}, { 76, 3}, {200, 3},61{512, 3}, {312, 3}, {240, 3}, {960, 3}, {264, 2}, {672, 2},62{ 38, 2}, {328, 2}, { 84, 2}, { 39, 2}, {216, 2}, { 42, 2},63{ 37, 2}, {1608, 2}, { 70, 2}, { 46, 2}, {536, 2}, {280, 1},64{248, 1}, { 47, 1}, {1088, 1}, {1288, 1}, {224, 1}, { 41, 1},65{ 50, 1}, { 49, 1}, {808, 1}, {360, 1}, {440, 1}, { 43, 1},66{ 45, 1}, { 78, 1}, {968, 1}, {392, 1}, { 54, 1}, { 53, 1},67{ 59, 1}, {376, 1}, {664, 1}, { 58, 1}, {272, 1}, { 66, 1},68{2688, 1}, {472, 1}, {568, 1}, {720, 1}, { 51, 1}, { 63, 1},69{ 86, 1}, {496, 1}, {776, 1}, { 57, 1}, {680, 1}, {792, 1},70{122, 1}, {760, 1}, {824, 1}, {552, 1}, { 67, 1}, {456, 1},71{984, 1}, { 74, 1}, {408, 1}, { 75, 1}, { 92, 1}, {576, 1},72{116, 1}, { 65, 1}, {117, 1}, { 82, 1}, {352, 1}, { 55, 1},73{100, 1}, { 90, 1}, {696, 1}, {111, 1}, {880, 1}, { 79, 1},74{488, 1}, { 61, 1}, {114, 1}, { 94, 1}, {1032, 1}, { 98, 1},75{ 87, 1}, {584, 1}, { 85, 1}, {648, 1}, {0, 0}76};7778#define ALIGN_NUM 102479#define ALIGN_MASK (ALIGN_NUM-1)80static uint8_t src_align_arr[ALIGN_NUM];81static uint8_t dst_align_arr[ALIGN_NUM];8283/* Source alignment frequency for memcpy based on SPEC2017. */84static align_data_t src_align_freq[] =85{86{8, 300}, {16, 292}, {32, 168}, {64, 153}, {4, 79}, {2, 14}, {1, 18}, {0, 0}87};8889static align_data_t dst_align_freq[] =90{91{8, 265}, {16, 263}, {64, 209}, {32, 174}, {4, 90}, {2, 10}, {1, 13}, {0, 0}92};9394typedef struct95{96uint64_t src : 24;97uint64_t dst : 24;98uint64_t len : 16;99} copy_t;100101static copy_t test_arr[NUM_TESTS];102103typedef char *(*proto_t) (char *, const char *, size_t);104105static void106init_copy_distribution (void)107{108int i, j, freq, size, n;109110for (n = i = 0; (freq = size_freq[i].freq) != 0; i++)111for (j = 0, size = size_freq[i].size; j < freq; j++)112size_arr[n++] = size;113assert (n == SIZE_NUM);114115for (n = i = 0; (freq = src_align_freq[i].freq) != 0; i++)116for (j = 0, size = src_align_freq[i].align; j < freq; j++)117src_align_arr[n++] = size - 1;118assert (n == ALIGN_NUM);119120for (n = i = 0; (freq = dst_align_freq[i].freq) != 0; i++)121for (j = 0, size = dst_align_freq[i].align; j < freq; j++)122dst_align_arr[n++] = size - 1;123assert (n == ALIGN_NUM);124}125126static size_t127init_copies (size_t max_size)128{129size_t total = 0;130/* Create a random set of copies with the given size and alignment131distributions. */132for (int i = 0; i < NUM_TESTS; i++)133{134test_arr[i].dst = (rand32 (0) & (max_size - 1));135test_arr[i].dst &= ~dst_align_arr[rand32 (0) & ALIGN_MASK];136test_arr[i].src = (rand32 (0) & (max_size - 1));137test_arr[i].src &= ~src_align_arr[rand32 (0) & ALIGN_MASK];138test_arr[i].len = size_arr[rand32 (0) & SIZE_MASK];139total += test_arr[i].len;140}141142return total;143}144145static void inline __attribute ((always_inline))146memcpy_random (const char *name, void *(*fn)(void *, const void *, size_t))147{148printf ("%22s ", name);149uint64_t total = 0, tsum = 0;150for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2)151{152uint64_t copy_size = init_copies (size) * ITERS;153154for (int c = 0; c < NUM_TESTS; c++)155fn (b + test_arr[c].dst, a + test_arr[c].src, test_arr[c].len);156157uint64_t t = clock_get_ns ();158for (int i = 0; i < ITERS; i++)159for (int c = 0; c < NUM_TESTS; c++)160fn (b + test_arr[c].dst, a + test_arr[c].src, test_arr[c].len);161t = clock_get_ns () - t;162total += copy_size;163tsum += t;164printf ("%dK: %5.2f ", size / 1024, (double)copy_size / t);165}166printf( "avg %5.2f\n", (double)total / tsum);167}168169static void inline __attribute ((always_inline))170memcpy_medium_aligned (const char *name, void *(*fn)(void *, const void *, size_t))171{172printf ("%22s ", name);173174for (int size = 8; size <= 512; size *= 2)175{176uint64_t t = clock_get_ns ();177for (int i = 0; i < ITERS2; i++)178fn (b, a, size);179t = clock_get_ns () - t;180printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);181}182printf ("\n");183}184185static void inline __attribute ((always_inline))186memcpy_medium_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))187{188printf ("%22s ", name);189190for (int size = 8; size <= 512; size *= 2)191{192uint64_t t = clock_get_ns ();193for (int i = 0; i < ITERS2; i++)194fn (b + 3, a + 1, size);195t = clock_get_ns () - t;196printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);197}198printf ("\n");199}200201static void inline __attribute ((always_inline))202memcpy_large (const char *name, void *(*fn)(void *, const void *, size_t))203{204printf ("%22s ", name);205206for (int size = 1024; size <= 65536; size *= 2)207{208uint64_t t = clock_get_ns ();209for (int i = 0; i < ITERS3; i++)210fn (b, a, size);211t = clock_get_ns () - t;212printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);213}214printf ("\n");215}216217static void inline __attribute ((always_inline))218memmove_forward_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))219{220printf ("%22s ", name);221222for (int size = 1024; size <= 65536; size *= 2)223{224uint64_t t = clock_get_ns ();225for (int i = 0; i < ITERS3; i++)226fn (a, a + 256 + (i & 31), size);227t = clock_get_ns () - t;228printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);229}230231printf ("\n");232}233234static void inline __attribute ((always_inline))235memmove_backward_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))236{237printf ("%22s ", name);238239for (int size = 1024; size <= 65536; size *= 2)240{241uint64_t t = clock_get_ns ();242for (int i = 0; i < ITERS3; i++)243fn (a + 256 + (i & 31), a, size);244t = clock_get_ns () - t;245printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);246}247248printf ("\n");249}250251int main (void)252{253init_copy_distribution ();254255memset (a, 1, sizeof (a));256memset (b, 2, sizeof (b));257258DOTEST ("Random memcpy (bytes/ns):\n", memcpy_random);259DOTEST ("Medium memcpy aligned (bytes/ns):\n", memcpy_medium_aligned);260DOTEST ("Medium memcpy unaligned (bytes/ns):\n", memcpy_medium_unaligned);261DOTEST ("Large memcpy (bytes/ns):\n", memcpy_large);262DOTEST ("Forwards memmove unaligned (bytes/ns):\n", memmove_forward_unaligned);263DOTEST ("Backwards memmove unaligned (bytes/ns):\n", memmove_backward_unaligned);264265return 0;266}267268269