[a1b9bc3] | 1 | #include <atomic> |
---|
| 2 | #include <iostream> |
---|
| 3 | #include <locale> |
---|
| 4 | #include <thread> |
---|
| 5 | |
---|
| 6 | #include <x86intrin.h> |
---|
| 7 | |
---|
| 8 | struct __attribute__((aligned(128))) Global_t { |
---|
| 9 | volatile size_t value; |
---|
| 10 | } global; |
---|
| 11 | |
---|
| 12 | static const size_t iterations = 1'000'000'000; |
---|
| 13 | |
---|
| 14 | size_t read() { |
---|
| 15 | // size_t r = __atomic_load_n(&global.value, __ATOMIC_RELAXED); |
---|
| 16 | // _mm_stream_si64((long long int*)&global.value, r); |
---|
| 17 | // // _mm_clflush( (void*)&global.value ); |
---|
| 18 | // // __builtin_prefetch((void*)&global.value); |
---|
| 19 | // asm volatile( |
---|
| 20 | // "PREFETCHNTA %[target]" |
---|
| 21 | // : |
---|
| 22 | // : [target] "m" (global.value) |
---|
| 23 | // ); |
---|
| 24 | // return r; |
---|
| 25 | return __atomic_load_n(&global.value, __ATOMIC_SEQ_CST); |
---|
| 26 | |
---|
| 27 | // __m128i r = _mm_stream_load_si128((__m128i*)&global.value); |
---|
| 28 | // asm volatile( |
---|
| 29 | // "PREFETCHNTA %[target]" |
---|
| 30 | // : |
---|
| 31 | // : [target] "m" (global.value) |
---|
| 32 | // ); |
---|
| 33 | // return ((Global_t*)&r)->value; |
---|
| 34 | // size_t r; |
---|
| 35 | // asm volatile( |
---|
| 36 | // "MOVNTI %[target], %[r]\n\t" |
---|
| 37 | // : [r] "=r" (r) |
---|
| 38 | // : [target] "m" (global.value) |
---|
| 39 | // ); |
---|
| 40 | // return r; |
---|
| 41 | } |
---|
| 42 | |
---|
| 43 | void write(size_t v) { |
---|
| 44 | // __atomic_store_n(&global.value, v, __ATOMIC_SEQ_CST); |
---|
| 45 | // __atomic_store_n(&global.value, v, __ATOMIC_RELAXED); |
---|
| 46 | // asm volatile( |
---|
| 47 | // "MOVNTI %[v], %[target]\n\t" |
---|
| 48 | // : |
---|
| 49 | // : [target] "m" (global.value), [v] "r" (v) |
---|
| 50 | // ); |
---|
| 51 | _mm_stream_si64((long long int*)&global.value, v); |
---|
| 52 | } |
---|
| 53 | |
---|
| 54 | void reader(size_t * reads, size_t * diffs, size_t * m) { |
---|
| 55 | size_t last = read(); |
---|
| 56 | for(size_t i = 0; i < iterations; i++) { |
---|
| 57 | size_t val = read(); |
---|
| 58 | if(last != val) (*diffs)++; |
---|
| 59 | last = val; |
---|
| 60 | if(last > *m) *m = last; |
---|
| 61 | (*reads)++; |
---|
| 62 | } |
---|
| 63 | } |
---|
| 64 | |
---|
| 65 | std::atomic<bool> done = { false }; |
---|
| 66 | |
---|
| 67 | void writer() { |
---|
| 68 | size_t v = 0; |
---|
| 69 | while(!done) { |
---|
| 70 | v++; |
---|
| 71 | write(v); |
---|
| 72 | __atomic_thread_fence(__ATOMIC_SEQ_CST); |
---|
| 73 | } |
---|
| 74 | } |
---|
| 75 | |
---|
| 76 | int main() { |
---|
| 77 | std::cout.imbue(std::locale("")); |
---|
| 78 | size_t reads = 0; |
---|
| 79 | size_t diffs = 0; |
---|
| 80 | size_t max = 0; |
---|
| 81 | auto w = std::thread(writer); |
---|
| 82 | auto r = std::thread(reader, &reads, &diffs, &max); |
---|
| 83 | r.join(); |
---|
| 84 | done = true; |
---|
| 85 | w.join(); |
---|
| 86 | std::cout << reads << " " << diffs << " " << max << std::endl; |
---|
| 87 | } |
---|