chore: 初始化 ViperApps 仓库
纳入 11 个示例工程（不含 ViperOS，另行开专题）的源码：已完成新语法迁移（R1-R16 机械规则与 R1/R12 改写）。.gitignore 排除 ViperOS/、各工程编译产物（.tpv_build/.native_build/output/temp）、二进制与 ViperMNIST/data 数据集。
This commit is contained in:
TermiNexus committed
commit
df8e301720
54 files changed
+12112
No files matched your search
@@ -0,0 +1,237 @@
|
||||
/*
|
||||
* bench_cpp.cpp - 手写 C++ 基准实现 (进程内多次 run, 单进程输出多行)
|
||||
*
|
||||
* 与 bench_c.c 完全同口径:
|
||||
* 内核 out[i] = a[i]*b[i] + c[i], LEN = 16384, 每次 run = ITERS = 65536 趟,
|
||||
* 总 flops ≈ 2.147e9; 命令行第 1 参数 = runs(默认 1; 0 = 仅热身)。
|
||||
*
|
||||
* - 容器 std::vector; 内核经裸指针 + __restrict 传入, 便于自动向量化。
|
||||
* - 线程绑定: 入口 SetThreadAffinityMask 绑到首个可用逻辑核, 以 AFFINITY <core>
|
||||
* 行回报(-1 = 绑定失败)。与 bench_c.c / main.vp / bench_py.py 绑同一个核。
|
||||
* - 双计时: 每次 run 同时采集 QPC 墙钟微秒 与 QueryThreadCycleTime 线程 CPU
|
||||
* 微秒, 输出 RESULT <wall_us> <cpu_us> <flops>(flops 按墙钟秒换算)。
|
||||
* - 随机输入与 CHECK 口径与 C/Viper/Python 完全一致(固定种子 minstd LCG)。
|
||||
* - 编译口径与 bench_c.c 相同: -O2 -mavx2 -mfma -ffp-contract=fast,
|
||||
* 允许 FMA 收缩成单趟 vfmadd(单次舍入), 与 Viper 的 simd_fma_array 对齐。
|
||||
* - 防删: 每轮热身/每次 run 后升序读回 out 求 CHECK; noinline; __restrict。
|
||||
*/
|
||||
#ifndef NOMINMAX
|
||||
#define NOMINMAX
|
||||
#endif
|
||||
#include <cstdio>
|
||||
#include <cstdlib>
|
||||
#include <cmath>
|
||||
#include <vector>
|
||||
|
||||
/* QueryThreadCycleTime 是 Vista+ API: 显式抬到 0x0601 让 mingw 头文件声明它 */
|
||||
#ifndef _WIN32_WINNT
|
||||
#define _WIN32_WINNT 0x0601
|
||||
#endif
|
||||
#include <windows.h>
|
||||
|
||||
namespace {
|
||||
|
||||
/* 负载约束: N_ELEM × N_ITERS × 2 恒为 ≈2.147e9 flops/run(与其它实现可比),
|
||||
* 只调 N_ELEM 来改"工作集"大小。
|
||||
* N_ELEM=16384 -> a/b/c/out 四份共 512 KiB, 稳居本机每核 1 MiB 私有 L2 之内:
|
||||
* 访存不再打到共享 L3 / DRAM, 也就不会被其它进程抢带宽 —— 这正是之前 C++
|
||||
* 出现"零星高时延尖峰"的主因(与 std::vector/向量化无关, 计时区内核逐字相同)。 */
|
||||
constexpr long N_ELEM = 16384;
|
||||
constexpr long N_ITERS = 65536;
|
||||
constexpr int RUNS_MAX = 200;
|
||||
|
||||
constexpr double LCG_A = 48271.0;
|
||||
constexpr double LCG_M = 2147483647.0;
|
||||
constexpr double LCG_SEED = 12345.0;
|
||||
|
||||
/* ------------------------------------------------------------
|
||||
* 线程绑定: 绑到"第一个可用逻辑核", 返回 CPU 序号; 失败返回 -1。
|
||||
* 从 1<<0 逐个试, 第一个被接受的位就是进程亲和掩码里编号最小的核 ——
|
||||
* 四个实现绑到同一个核, 横向对比才有意义。
|
||||
* ------------------------------------------------------------ */
|
||||
int bind_first_available_core() {
|
||||
const int nbits = static_cast<int>(sizeof(DWORD_PTR) * 8);
|
||||
for (int i = 0; i < nbits; ++i) {
|
||||
if (SetThreadAffinityMask(GetCurrentThread(),
|
||||
static_cast<DWORD_PTR>(1) << i) != 0) {
|
||||
return i;
|
||||
}
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
|
||||
/* ------------------------------------------------------------
|
||||
* 线程 CPU 时间: QueryThreadCycleTime 返回本线程真正消耗过的 CPU 周期数
|
||||
* (现代 x86 的 TSC 是 invariant 的, 与核心实际频率无关)。
|
||||
*
|
||||
* 为什么不用 GetThreadTimes: 它的记账粒度跟随调度时钟节拍, 本机实测硬地板
|
||||
* 15.625ms —— 几十毫秒的 run 只能被记成 2~3 个节拍(误差 ±40%), 连
|
||||
* "CPU 时间 > 墙钟时间"这种物理上不可能的结果都会出现; 且
|
||||
* timeBeginPeriod(1) 在 Win10 2004+ 已无法改善该粒度(实测确认无效)。
|
||||
* ------------------------------------------------------------ */
|
||||
long long query_thread_cycles() {
|
||||
ULONG64 cycles = 0;
|
||||
QueryThreadCycleTime(GetCurrentThread(), &cycles);
|
||||
return static_cast<long long>(cycles);
|
||||
}
|
||||
|
||||
/* 用 QPC 标定 TSC 频率(周期/秒): 忙等 30ms 数周期, 重复 3 次取最大。0.0 = 失败。
|
||||
*
|
||||
* 为什么取最大: 标定窗口内本线程一旦被抢占, 消耗周期就少于墙钟流逝时间,
|
||||
* 算出 hz 偏小 -> 后续 cpu_us 系统性偏大(实测约 +0.13%, 表现为"cpu_us 偶尔
|
||||
* 比 wall_us 还大"这种看着像 bug 的伪影)。取三次最大值 = 挑最没被打断的一次。
|
||||
* 必须在 SetThreadAffinityMask 之后调用, 否则跨核的 TSC 可能不同步。 */
|
||||
double calibrate_tsc_hz(const LARGE_INTEGER &freq) {
|
||||
const double span = static_cast<double>(freq.QuadPart) * 0.030; /* 30ms 的 QPC 计次 */
|
||||
double best = 0.0;
|
||||
for (int t = 0; t < 3; ++t) {
|
||||
LARGE_INTEGER a{}, b{};
|
||||
const long long c0 = query_thread_cycles();
|
||||
QueryPerformanceCounter(&a);
|
||||
do {
|
||||
QueryPerformanceCounter(&b);
|
||||
} while (static_cast<double>(b.QuadPart - a.QuadPart) < span);
|
||||
const long long c1 = query_thread_cycles();
|
||||
const double dt = static_cast<double>(b.QuadPart - a.QuadPart);
|
||||
if (dt <= 0.0 || c1 <= c0) {
|
||||
continue;
|
||||
}
|
||||
const double hz =
|
||||
static_cast<double>(c1 - c0) * static_cast<double>(freq.QuadPart) / dt;
|
||||
if (hz > best) {
|
||||
best = hz;
|
||||
}
|
||||
}
|
||||
return best;
|
||||
}
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#define NOINLINE __declspec(noinline)
|
||||
#else
|
||||
#define NOINLINE __attribute__((noinline))
|
||||
#endif
|
||||
|
||||
/* 填充 a/b/c: 逐元素依次消耗同一个 LCG 流(a[i], b[i], c[i]) */
|
||||
void fill_abc(double *a, double *b, double *c, long n) {
|
||||
double state = LCG_SEED;
|
||||
const double m = LCG_M;
|
||||
const double mul = LCG_A;
|
||||
for (long i = 0; i < n; ++i) {
|
||||
state = std::fmod(state * mul, m);
|
||||
a[i] = state / m;
|
||||
state = std::fmod(state * mul, m);
|
||||
b[i] = state / m;
|
||||
state = std::fmod(state * mul, m);
|
||||
c[i] = state / m;
|
||||
}
|
||||
}
|
||||
|
||||
/* 内核: out[i] = a[i]*b[i] + c[i] (元素间独立, -O2 + AVX2 + FMA 自动向量化单趟融合乘加) */
|
||||
static NOINLINE void kernel_fma(double *__restrict out,
|
||||
const double *__restrict a,
|
||||
const double *__restrict b,
|
||||
const double *__restrict c,
|
||||
long n) {
|
||||
for (long i = 0; i < n; ++i) {
|
||||
out[i] = a[i] * b[i] + c[i];
|
||||
}
|
||||
}
|
||||
|
||||
/* 校验和: 升序逐元素累加 out(读回 out -> 保证内核的写未被删除) */
|
||||
double checksum_sum(const double *__restrict out, long n) {
|
||||
double s = 0.0;
|
||||
for (long i = 0; i < n; ++i) {
|
||||
s += out[i];
|
||||
}
|
||||
return s;
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
||||
int main(int argc, char **argv) {
|
||||
/* 命令行: bench_cpp.exe [runs] [len] [iters]
|
||||
* 省略则用上面 constexpr 的默认值; run_bench.py 多尺寸扫描时会显式传入。 */
|
||||
long n = N_ELEM;
|
||||
long iters = N_ITERS;
|
||||
int runs = 1;
|
||||
if (argc > 1) {
|
||||
runs = std::atoi(argv[1]);
|
||||
}
|
||||
if (argc > 2) {
|
||||
n = std::atol(argv[2]);
|
||||
}
|
||||
if (argc > 3) {
|
||||
iters = std::atol(argv[3]);
|
||||
}
|
||||
if (runs < 0) runs = 0;
|
||||
if (runs > RUNS_MAX) runs = RUNS_MAX;
|
||||
/* 内核按 4-lane ymm 分批, LEN 必须是 4 的倍数; 参数非法直接报错, 不静默修正 */
|
||||
if (n < 4 || n % 4 != 0 || iters < 1) {
|
||||
std::fprintf(stderr, "bad args: len=%ld(需 >=4 且为 4 的倍数) iters=%ld\n", n, iters);
|
||||
return 2;
|
||||
}
|
||||
const double total_flops = static_cast<double>(n) * static_cast<double>(iters) * 2.0;
|
||||
|
||||
/* 线程绑定(必做) + 提优先级: 消除跨核迁移, 并压低背景进程抢占造成的零星尖峰 */
|
||||
const int core = bind_first_available_core();
|
||||
SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_HIGHEST);
|
||||
|
||||
std::printf("CFG %ld %ld\n", n, iters);
|
||||
std::printf("AFFINITY %d\n", core);
|
||||
std::fflush(stdout);
|
||||
|
||||
std::vector<double> a(n), b(n), c(n), out(n);
|
||||
double *pa = a.data();
|
||||
double *pb = b.data();
|
||||
double *pc = c.data();
|
||||
double *pout = out.data();
|
||||
|
||||
LARGE_INTEGER freq, t0, t1;
|
||||
double last_chk = 0.0;
|
||||
|
||||
/* 输入生成一次(不计时) */
|
||||
fill_abc(pa, pb, pc, n);
|
||||
|
||||
QueryPerformanceFrequency(&freq);
|
||||
|
||||
/* TSC 频率标定(必须在绑核之后): 把 QueryThreadCycleTime 的周期数换算成微秒 */
|
||||
const double tsc_hz = calibrate_tsc_hz(freq);
|
||||
std::printf("TSC %.4e\n", tsc_hz);
|
||||
std::fflush(stdout);
|
||||
|
||||
/* 2 轮满负荷热身(每轮 = ITERS 趟内核 + 读回 out 求校验和) */
|
||||
for (int r = 0; r < 2; ++r) {
|
||||
for (long k = 0; k < iters; ++k) {
|
||||
kernel_fma(pout, pa, pb, pc, n);
|
||||
}
|
||||
last_chk = checksum_sum(pout, n);
|
||||
}
|
||||
|
||||
/* 连续计时 runs 次; 计时只包住 ITERS 趟内核
|
||||
* 双计时: 墙钟 QPC + 线程 CPU(QueryThreadCycleTime), 同时包住同一段内核 */
|
||||
for (int r = 0; r < runs; ++r) {
|
||||
const long long cy0 = query_thread_cycles();
|
||||
QueryPerformanceCounter(&t0);
|
||||
for (long k = 0; k < iters; ++k) {
|
||||
kernel_fma(pout, pa, pb, pc, n);
|
||||
}
|
||||
QueryPerformanceCounter(&t1);
|
||||
const long long cy1 = query_thread_cycles();
|
||||
|
||||
const double us =
|
||||
(static_cast<double>(t1.QuadPart - t0.QuadPart) * 1000000.0) /
|
||||
static_cast<double>(freq.QuadPart);
|
||||
/* 周期 -> 微秒; 标定失败(tsc_hz==0)时给 0, 表现为一眼可见的异常 */
|
||||
const double cpu_us = (tsc_hz > 0.0)
|
||||
? static_cast<double>(cy1 - cy0) * 1000000.0 / tsc_hz : 0.0;
|
||||
const double flops = total_flops / (us * 1e-6);
|
||||
std::printf("RESULT %.3f %.3f %.0f\n", us, cpu_us, flops);
|
||||
std::fflush(stdout);
|
||||
/* 计时区外读回 out: 阻止跨 run 的写折叠 */
|
||||
last_chk = checksum_sum(pout, n);
|
||||
}
|
||||
|
||||
std::printf("CHECK %.17g\n", last_chk);
|
||||
std::fflush(stdout);
|
||||
return 0;
|
||||
}
|
||||
Reference in new issue
Block a user