我正在尝试找到一个 PMC（性能监控计数器），它将显示prefetcht0指令命中 L1 dcache（或未命中）的次数。
icelake 客户端：Intel(R) Core(TM) i7-1065G7 CPU @ 1.30GHz
我正在尝试制作这种细粒度的即（注意应包括lfence周围prefetcht0）
xorl %ecx, %ecx
rdpmc
movl %eax, %edi
prefetcht0 (%rsi)
rdpmc
testl %eax, %edi
// jump depending on if it was a miss or not
目标是检查预取是否命中 L1。如果没有执行一些准备好的代码，否则继续。
仅基于可用的内容，它似乎必须是一个错过事件。
我尝试了libpfm4和 intel 手册中的一些事件，但没有成功：
L1-DCACHE-LOAD-MISSES, emask=0x00, umask=0x10000
L1D.REPLACEMENT, emask=0x51, umask=0x1
L2_RQSTS.SWPF_HIT, emask=0x24, umask=0xc8
L2_RQSTS.SWPF_MISS, emask=0x24, umask=0x28
LOAD_HIT_PREFETCH.SWPF, emask=0x01, umask=0x4c (this very misleadingly is non-sw prefetch hits)
L1D.REPLACEMENT和L1-DCACHE-LOAD-MISSES一种作品，如果我延迟它会起作用，rdpmc但如果它们一个接一个，它似乎充其量是不可靠的。其他的都是完整的半身像。
问题：
- 这些是否可以用于检测预取是否命中 L1 dcache？（即我的测试很糟糕）
- 如果不。什么事件可用于检测预取是否命中 L1 dcache？
编辑：MEM_LOAD_RETIRED.L1_HIT似乎不适用于软件预取。
这是我用来做测试的代码：
#include <asm/unistd.h>
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <linux/perf_event.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#define HIT 0
#define MISS 1
#define TODO MISS
#define PAGE_SIZE 4096
// to force hit make TSIZE low
#define TSIZE 10000
#define err_assert(cond) \
if (__builtin_expect(!(cond), 0)) { \
fprintf(stderr, "%d:%d: %s\n", __LINE__, errno, strerror(errno)); \
exit(-1); \
}
uint64_t
get_addr() {
uint8_t * addr =
(uint8_t *)mmap(NULL, TSIZE * PAGE_SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
err_assert(addr != NULL);
for (uint32_t i = 0; i < TSIZE; ++i) {
addr[i * PAGE_SIZE + (PAGE_SIZE - 1)] = 0;
#if TODO == HIT
addr[i * PAGE_SIZE] = 0;
#endif
}
return uint64_t(addr);
}
int
perf_event_open(struct perf_event_attr * hw_event,
pid_t pid,
int cpu,
int group_fd,
unsigned long flags) {
int ret;
ret = syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
return ret;
}
void
init_perf_event_struct(struct perf_event_attr * pe,
const uint32_t type,
const uint64_t ev_config,
int lead) {
__builtin_memset(pe, 0, sizeof(struct perf_event_attr));
pe->type = type;
pe->size = sizeof(struct perf_event_attr);
pe->config = ev_config;
pe->disabled = !!lead;
pe->exclude_kernel = 1;
pe->exclude_hv = 1;
}
/* Fixed Counters */
static constexpr uint32_t core_instruction_ev = 0x003c;
static constexpr uint32_t core_instruction_idx = (1 << 30) + 0;
static constexpr uint32_t core_cycles_ev = 0x00c0;
static constexpr uint32_t core_cycles_idx = (1 << 30) + 1;
static constexpr uint32_t ref_cycles_ev = 0x0300;
static constexpr uint32_t ref_cycles_idx = (1 << 30) + 2;
/* programmable counters */
static constexpr uint32_t mem_load_retired_l1_hit = 0x01d1;
static constexpr uint32_t mem_load_retired_l1_miss = 0x08d1;
int
init_perf_tracking() {
struct perf_event_attr pe;
init_perf_event_struct(&pe, PERF_TYPE_RAW, core_instruction_ev, 1);
int leadfd = perf_event_open(&pe, 0, -1, -1, 0);
err_assert(leadfd >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, core_cycles_ev, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, ref_cycles_ev, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, mem_load_retired_l1_hit, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
return leadfd;
}
void
start_perf_tracking(int leadfd) {
ioctl(leadfd, PERF_EVENT_IOC_RESET, 0);
ioctl(leadfd, PERF_EVENT_IOC_ENABLE, 0);
}
#define _V_TO_STR(X) #X
#define V_TO_STR(X) _V_TO_STR(X)
//#define DO_PREFETCH
#ifdef DO_PREFETCH
#define DO_MEMORY_OP(addr) "prefetcht0 (%[" V_TO_STR(addr) "])\n\t"
#else
#define DO_MEMORY_OP(addr) "movl (%[" V_TO_STR(addr) "]), %%eax\n\t"
#endif
int
main() {
int fd = init_perf_tracking();
start_perf_tracking(fd);
uint64_t addr = get_addr();
uint32_t prefetch_miss, cycles_to_detect;
asm volatile(
"lfence\n\t"
"movl %[core_cycles_idx], %%ecx\n\t"
"rdpmc\n\t"
"movl %%eax, %[cycles_to_detect]\n\t"
"xorl %%ecx, %%ecx\n\t"
"rdpmc\n\t"
"movl %%eax, %[prefetch_miss]\n\t"
"lfence\n\t"
DO_MEMORY_OP(prefetch_addr)
"lfence\n\t"
"xorl %%ecx, %%ecx\n\t"
"rdpmc\n\t"
"subl %[prefetch_miss], %%eax\n\t"
"movl %%eax, %[prefetch_miss]\n\t"
"movl %[core_cycles_idx], %%ecx\n\t"
"rdpmc\n\t"
"subl %[cycles_to_detect], %%eax\n\t"
"movl %%eax, %[cycles_to_detect]\n\t"
"lfence\n\t"
: [ prefetch_miss ] "=&r"(prefetch_miss),
[ cycles_to_detect ] "=&r"(cycles_to_detect)
: [ prefetch_addr ] "r"(addr), [ core_cycles_idx ] "i"(core_cycles_idx)
: "eax", "edx", "ecx");
fprintf(stderr, "Hit : %d\n", prefetch_miss);
fprintf(stderr, "Cycles : %d\n", cycles_to_detect);
}
如果我将DO_PREFETCH结果定义为MEM_LOAD_RETIRED.L1_HIT始终为 1（似乎总是受到打击）。如果我注释掉DO_PREFETCH结果与我所期望的一致（当地址显然不在缓存中时报告未命中，当它显然是报告命中时）。
与DO_PREFETCH：
g++ -DDO_PREFETCH -O3 -march=native -mtune=native prefetch_hits.cc -o prefetch_hits
$> ./prefetch_hits
Hit : 1
Cycles : 554
并且没有DO_PREFETCH
g++ -DDO_PREFETCH -O3 -march=native -mtune=native prefetch_hits.cc -o prefetch_hits
$> ./prefetch_hits
Hit : 0
Cycles : 888
有了L2_RQSTS.SWPF_HIT并且L2_RQSTS.SWPF_MISS能够让它工作。非常感谢哈迪·布雷斯。值得注意的是，L1D_PEND_MISS.PENDING不起作用的原因可能与 Icelake 有关。Hadi Brais 报告说，它可以用于预测 Haswell 上的 L1D 缓存未命中。
为了试图确定为什么不工作L1_PEND_MISS.PENDING，MEM_LOAD_RETIRED.L1_HIT发布了我用来测试它们的确切代码：
#include <asm/unistd.h>
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <linux/perf_event.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#define HIT 0
#define MISS 1
#define TODO MISS
#define PAGE_SIZE 4096
#define TSIZE 1000
#define err_assert(cond) \
if (__builtin_expect(!(cond), 0)) { \
fprintf(stderr, "%d:%d: %s\n", __LINE__, errno, strerror(errno)); \
exit(-1); \
}
uint64_t
get_addr() {
uint8_t * addr =
(uint8_t *)mmap(NULL, TSIZE * PAGE_SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
err_assert(addr != NULL);
__builtin_memset(addr, -1, TSIZE * PAGE_SIZE);
return uint64_t(addr);
}
int
perf_event_open(struct perf_event_attr * hw_event,
pid_t pid,
int cpu,
int group_fd,
unsigned long flags) {
int ret;
ret = syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
return ret;
}
void
init_perf_event_struct(struct perf_event_attr * pe,
const uint32_t type,
const uint64_t ev_config,
int lead) {
__builtin_memset(pe, 0, sizeof(struct perf_event_attr));
pe->type = type;
pe->size = sizeof(struct perf_event_attr);
pe->config = ev_config;
pe->disabled = !!lead;
pe->exclude_kernel = 1;
pe->exclude_hv = 1;
}
/* Fixed Counters */
static constexpr uint32_t core_instruction_ev = 0x003c;
static constexpr uint32_t core_instruction_idx = (1 << 30) + 0;
static constexpr uint32_t core_cycles_ev = 0x00c0;
static constexpr uint32_t core_cycles_idx = (1 << 30) + 1;
static constexpr uint32_t ref_cycles_ev = 0x0300;
static constexpr uint32_t ref_cycles_idx = (1 << 30) + 2;
/* programmable counters */
static constexpr uint32_t mem_load_retired_l1_hit = 0x01d1;
static constexpr uint32_t mem_load_retired_l1_miss = 0x08d1;
static constexpr uint32_t l1d_pending = 0x0148;
static constexpr uint32_t swpf_hit = 0xc824;
static constexpr uint32_t swpf_miss = 0x2824;
static constexpr uint32_t ev0 = l1d_pending;
#define NEVENTS 1
#if NEVENTS > 1
static constexpr uint32_t ev1 = swpf_miss;
#endif
int
init_perf_tracking() {
struct perf_event_attr pe;
init_perf_event_struct(&pe, PERF_TYPE_RAW, core_instruction_ev, 1);
int leadfd = perf_event_open(&pe, 0, -1, -1, 0);
err_assert(leadfd >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, core_cycles_ev, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, ref_cycles_ev, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
init_perf_event_struct(&pe, PERF_TYPE_RAW, ev0, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
#if NEVENTS > 1
init_perf_event_struct(&pe, PERF_TYPE_RAW, ev1, 0);
err_assert(perf_event_open(&pe, 0, -1, leadfd, 0) >= 0);
#endif
return leadfd;
}
void
start_perf_tracking(int leadfd) {
ioctl(leadfd, PERF_EVENT_IOC_RESET, 0);
ioctl(leadfd, PERF_EVENT_IOC_ENABLE, 0);
}
#define _V_TO_STR(X) #X
#define V_TO_STR(X) _V_TO_STR(X)
//#define LFENCE
#ifdef LFENCE
#define SERIALIZER() "lfence\n\t"
#else
#define SERIALIZER() \
"xorl %%ecx, %%ecx\n\t" \
"xorl %%eax, %%eax\n\t" \
"cpuid\n\t"
#endif
#define DO_PREFETCH
#ifdef DO_PREFETCH
#define DO_MEMORY_OP(addr) "prefetcht0 (%[" V_TO_STR(addr) "])\n\t"
#else
#define DO_MEMORY_OP(addr) "movl (%[" V_TO_STR(addr) "]), %%eax\n\t"
#endif
int
main() {
int fd = init_perf_tracking();
start_perf_tracking(fd);
uint64_t addr = get_addr();
// to ensure page in TLB
*((volatile uint64_t *)(addr + (PAGE_SIZE - 8))) = 0;
#if TODO == HIT
// loading from 0 offset to check cache miss / hit
*((volatile uint64_t *)addr) = 0;
#endif
uint32_t ecount0 = 0, ecount1 = 0, cycles_to_detect = 0;
asm volatile(
SERIALIZER()
"movl %[core_cycles_idx], %%ecx\n\t"
"rdpmc\n\t"
"movl %%eax, %[cycles_to_detect]\n\t"
"xorl %%ecx, %%ecx\n\t"
"rdpmc\n\t"
"movl %%eax, %[ecount0]\n\t"
#if NEVENTS > 1
"movl $1, %%ecx\n\t"
"rdpmc\n\t"
"movl %%eax, %[ecount1]\n\t"
#endif
SERIALIZER()
DO_MEMORY_OP(prefetch_addr)
SERIALIZER()
"xorl %%ecx, %%ecx\n\t"
"rdpmc\n\t"
"subl %[ecount0], %%eax\n\t"
"movl %%eax, %[ecount0]\n\t"
#if NEVENTS > 1
"movl $1, %%ecx\n\t"
"rdpmc\n\t"
"subl %[ecount1], %%eax\n\t"
"movl %%eax, %[ecount1]\n\t"
#endif
"movl %[core_cycles_idx], %%ecx\n\t"
"rdpmc\n\t"
"subl %[cycles_to_detect], %%eax\n\t"
"movl %%eax, %[cycles_to_detect]\n\t"
SERIALIZER()
: [ ecount0 ] "=&r"(ecount0),
#if NEVENTS > 1
[ ecount1 ] "=&r"(ecount1),
#endif
[ cycles_to_detect ] "=&r"(cycles_to_detect)
: [ prefetch_addr ] "r"(addr), [ core_cycles_idx ] "i"(core_cycles_idx)
: "eax", "edx", "ecx");
fprintf(stderr, "E0 : %d\n", ecount0);
fprintf(stderr, "E1 : %d\n", ecount1);
fprintf(stderr, "Cycles : %d\n", cycles_to_detect);
}