dzaima · May 10, 2023 16:29 · dzaima · Mar 13, 2023
diff --git a/cachetest.c b/cachetest.c
 // WIDTH: number of bytes to load/store in one instruction for bandwidth tests
 //   default: 32
 //   WIDTH∊1 2 4 8: regular integer loads
 //   x86-64 WIDTH==16 - SSE; WIDTH==32 - AVX2
 //   aarch64: WIDTH==16 - NEON loads; WIDTH==32 uses an ldp
 // CLOCK_RATE: number of clock cycles in a second on the target processor
 // TEST_LAT: test latency instead of bandwidth; latency tests aligned 8-byte loads

 #ifndef CLOCK_RATE
  #define CLOCK_RATE 3.6e9 // 3.6GHz
 #endif
 #ifndef TEST_LAT
  #define TEST_LAT 0 // test latency
 #endif
 // #define TP_ONLY 1 // only load
 // #define TP_ONLY 2 // only store

 #if __x86_64__
  #include <immintrin.h>
 #elif __aarch64__
  #include <arm_neon.h>
 #endif
 #include <inttypes.h>
 #include <stdbool.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <stddef.h>
 #define NOINLINE     __attribute__((noinline))
 #define FORCE_INLINE __attribute__((always_inline)) static inline
 #define NORETURN     __attribute__((noreturn))
 #define KEEPG(X) __asm__("" :: "r"(X))
 #define KEEPV(X) __asm__("" :: "x"(X))
 #define IDENTG(X) ({ __auto_type x_ = (X); __asm__("" : "=r"(x_) : "0"(x_)); x_; })
 #define MEMM __asm__("" ::: "memory")
 //#define KEEP(X) ({ __auto_type x=(X); _Generic(x, void*: KEEPG(x), KEEPV(x)); 1; })
 typedef   int8_t i8;
 typedef  uint8_t u8;
 typedef  int16_t i16;
 typedef uint16_t u16;
 typedef  int32_t i32;
 typedef uint32_t u32;
 typedef  int64_t i64;
 typedef uint64_t u64;
 typedef double   f64;
 typedef size_t ux;


 static inline void _wymum(uint64_t *A, uint64_t *B){
  __uint128_t r=*A; r*=*B;
  *A=(uint64_t)r; *B=(uint64_t)(r>>64);
 }
 static inline uint64_t _wymix(uint64_t A, uint64_t B){ _wymum(&A,&B); return A^B; }
 //The wyrand PRNG that pass BigCrush and PractRand
 static inline uint64_t wyrand(uint64_t *seed){ *seed+=0xa0761d6478bd642full; return _wymix(*seed,*seed^0xe7037ed1a0b428dbull);}
 //fast range integer random number generation on [0,k) credit to Daniel Lemire. May not work when WYHASH_32BIT_MUM=1. It can be combined with wyrand, wyhash64 or wyhash.
 static inline uint64_t wy2u0k(uint64_t r, uint64_t k){ _wymum(&r,&k); return k; }

 #include <time.h>
 static inline u64 nsTime() {
  struct timespec t;
  clock_gettime(CLOCK_REALTIME, &t);
  return (u64)(t.tv_sec*1000000000ll + t.tv_nsec);
 }
 #ifndef WIDTH
  #define WIDTH 32
 #endif
 #if WIDTH>8
  #define KEEPL(X) KEEPV(X)
  #if __x86_64__
    #if WIDTH==32
      #define INIT __m256 val = _mm256_set1_ps(123);
      #if UNALIGNED
        #define STORE _mm256_storeu_ps
        #define LOAD _mm256_loadu_ps
      #else
        #define STORE _mm256_store_ps
        #define LOAD _mm256_load_ps
      #endif
    #elif WIDTH==16
      #define INIT __m128 val = _mm_set1_ps(123);
      #if UNALIGNED
        #define STORE _mm_storeu_ps
        #define LOAD _mm_loadu_ps
      #else
        #define STORE _mm_store_ps
        #define LOAD _mm_load_ps
      #endif
    #endif
  #elif __aarch64__
    #define INIT float32x4_t val = vdupq_n_f32(123);
    #if WIDTH==16
      #define LOAD vld1q_f32
      #define STORE vst1q_f32
    #elif WIDTH==32
      #undef KEEPL
      #define KEEPL(X) ({ __auto_type x=X; KEEPV(x[0]); KEEPV(x[1]); })
      #define LOAD(P) (float32x4_t[2]){vld1q_f32(P), vld1q_f32(P+4)}
      #define STORE(P,V) vst1q_f32(P,V); vst1q_f32(P+4,V);
    #endif
  #endif
 #else
  #define INIT TY val = IDENTG((TY)123);
  #if WIDTH==8
    #define TY u64
  #elif WIDTH==4
    #define TY u32
  #elif WIDTH==2
    #define TY u16
  #elif WIDTH==1
    #define TY u8
  #endif
  #define KEEPL(X) KEEPG(X)
  #define STORE(P, V) *(TY*)(P) = (V)
  #define LOAD(P) *(TY*)(P)
 #endif

 NOINLINE u64 testThroughput(u8* data, u64 testsize, u64 rep, u64 breakonns, bool store) {
  u64 tns = 0;
  u64 minns = 1e18;
  while(tns < 1e8) {
    u64 sns = nsTime();
    INIT;
    if (store) {
      for (u64 k = 0; k < rep; k++) {
        for (u64 i = 0; i < testsize; i+= WIDTH*32) {
          u8* ptr = IDENTG(data+i); // IDENTG to avoid the addition propagating to complex addressing modes
          #pragma GCC unroll(32)
          for (u64 j = 0; j < 32; j++) { STORE((float*)(ptr + j*WIDTH), val); MEMM; }
        }
      }
    } else {
      for (u64 k = 0; k < rep; k++) {
        for (u64 i = 0; i < testsize; i+= WIDTH*32) {
          u8* ptr = IDENTG(data+i);
          #pragma GCC unroll(32)
          for (u64 j = 0; j < 32; j++) KEEPL(LOAD((float*)(ptr + j*WIDTH)));
        }
      }
    }
    
    u64 ens = nsTime();
    u64 cns = (ens-sns) / rep;
    if (cns < minns) {
      minns = cns;
      if (minns < breakonns) return minns;
    }
    tns+= ens-sns;
  }
  return minns;
 }
 NOINLINE u64 testLatency(u8* data, u64 bytes, u64 rep, u64 breakonns, bool unused) {
  u64 tns = 0;
  u64 minns = 1e18;
  while(tns < 1e8) {
    u64 sns = nsTime();
    for (u64 k = 0; k < rep; k++) {
      void** ptr = (void**)data;
      for (u64 i = 0; i < bytes/sizeof(void*); i+= 32) {
        #pragma GCC unroll(32)
        for (u64 j = 0; j < 32; j++) ptr = *ptr;
      }
      KEEPG(ptr);
    }
    u64 ens = nsTime();
    u64 cns = (ens-sns) / rep;
    if (cns < minns) {
      minns = cns;
      if (minns < breakonns) return minns;
    }
    tns+= ens-sns;
  }
  return minns;
 }

 #if TEST_LAT
  #define test testLatency
 #else
  #define test testThroughput
 #endif

 int main() {
  u64 max = 104857600; // 100MB
  u8* data = malloc(max+10240);
  data = (u8*) (((u64)data+4096) & ~(u64)4095);
  #if UNALIGNED
  data+= UNALIGNED;
  #endif
  #if TEST_LAT
    u64 seed = 12345;
    void** indexes = malloc(max+10240);
    for (u64 i = 0; i < max/sizeof(void*); i++) indexes[i] = (void*)i;
  #else
    for (u64 i = 0; i < max; i++) data[i] = i;
  #endif
  
  
  u64 bytes = 1024;
  while (bytes < max) {
    #if TEST_LAT
      // create random pointer chain
      u64 ptram = bytes/sizeof(void*);
      for (u64 i = 0; i < ptram; i++) { // shuffle ↕ptram pointers
        u64 j = wy2u0k(wyrand(&seed), i);
        void* t=indexes[i];
        indexes[i] = indexes[j];
        indexes[j] = t;
      }
      void** ptrs = (void**)data;
      u64 ci = indexes[ptram]-NULL;
      for (u64 i = 0; i < ptram+1; i++) { // create the chain
        ptrs[ci] = ptrs + (indexes[i]-NULL);
        ci = indexes[i]-NULL;
      }
    #endif
    
    u64 nses[2];
    for (int i = 0; i < (TEST_LAT? 1 : 2); i++) {
      #if TP_ONLY
        if (i!=TP_ONLY-1) continue;
      #endif
      u64 minns = test(data, bytes, 1, 10000, false);
      if (minns < 10000) {
        u64 rep = 10000 / minns;
        minns = test(data, bytes, rep, 0, i);
      }
      nses[i] = minns;
    }
    
    #if TEST_LAT
      printf("%9ldB:  %6.2fns, %6.2fc;  %ld ptrs, %ld total ns\n", bytes, nses[0]*1.0/ptram, (nses[0]/1e9*CLOCK_RATE)/ptram, ptram, nses[0]);
    #else
      printf("%9ldB: ", bytes);
      #if !TP_ONLY || TP_ONLY==1
        printf("load:%7.2fGB/s,%6.2fB/c", bytes/(nses[0]/1e9) / 1e9, bytes/(nses[0]/1e9) / CLOCK_RATE);
      #endif
      #ifndef TP_ONLY
        printf("      ");
      #endif
      #if !TP_ONLY || TP_ONLY==2
        printf("store:%7.2fGB/s,%6.2fB/c", bytes/(nses[1]/1e9) / 1e9, bytes/(nses[1]/1e9) / CLOCK_RATE);
      #endif
      printf("\n");
    #endif
    fflush(stdout);
    
    bytes*= 1.1;
    bytes = (bytes+1023) & ~(u64)1023;
  }
 }
	// WIDTH: number of bytes to load/store in one instruction for bandwidth tests
	// default: 32
	// WIDTH∊1 2 4 8: regular integer loads
	// x86-64 WIDTH==16 - SSE; WIDTH==32 - AVX2
	// aarch64: WIDTH==16 - NEON loads; WIDTH==32 uses an ldp
	// CLOCK_RATE: number of clock cycles in a second on the target processor
	// TEST_LAT: test latency instead of bandwidth; latency tests aligned 8-byte loads

	#ifndef CLOCK_RATE
	#define CLOCK_RATE 3.6e9 // 3.6GHz
	#endif
	#ifndef TEST_LAT
	#define TEST_LAT 0 // test latency
	#endif
	// #define TP_ONLY 1 // only load
	// #define TP_ONLY 2 // only store

	#if __x86_64__
	#include <immintrin.h>
	#elif __aarch64__
	#include <arm_neon.h>
	#endif
	#include <inttypes.h>
	#include <stdbool.h>
	#include <stdio.h>
	#include <stdlib.h>
	#include <string.h>
	#include <stddef.h>
	#define NOINLINE __attribute__((noinline))
	#define FORCE_INLINE __attribute__((always_inline)) static inline
	#define NORETURN __attribute__((noreturn))
	#define KEEPG(X) __asm__("" :: "r"(X))
	#define KEEPV(X) __asm__("" :: "x"(X))
	#define IDENTG(X) ({ __auto_type x_ = (X); __asm__("" : "=r"(x_) : "0"(x_)); x_; })
	#define MEMM __asm__("" ::: "memory")
	//#define KEEP(X) ({ __auto_type x=(X); _Generic(x, void*: KEEPG(x), KEEPV(x)); 1; })
	typedef int8_t i8;
	typedef uint8_t u8;
	typedef int16_t i16;
	typedef uint16_t u16;
	typedef int32_t i32;
	typedef uint32_t u32;
	typedef int64_t i64;
	typedef uint64_t u64;
	typedef double f64;
	typedef size_t ux;


	static inline void _wymum(uint64_t A, uint64_t B){
	__uint128_t r=A; r=*B;
	A=(uint64_t)r; B=(uint64_t)(r>>64);
	}
	static inline uint64_t _wymix(uint64_t A, uint64_t B){ _wymum(&A,&B); return A^B; }
	//The wyrand PRNG that pass BigCrush and PractRand
	static inline uint64_t wyrand(uint64_t seed){ seed+=0xa0761d6478bd642full; return _wymix(seed,seed^0xe7037ed1a0b428dbull);}
	//fast range integer random number generation on [0,k) credit to Daniel Lemire. May not work when WYHASH_32BIT_MUM=1. It can be combined with wyrand, wyhash64 or wyhash.
	static inline uint64_t wy2u0k(uint64_t r, uint64_t k){ _wymum(&r,&k); return k; }

	#include <time.h>
	static inline u64 nsTime() {
	struct timespec t;
	clock_gettime(CLOCK_REALTIME, &t);
	return (u64)(t.tv_sec*1000000000ll + t.tv_nsec);
	}
	#ifndef WIDTH
	#define WIDTH 32
	#endif
	#if WIDTH>8
	#define KEEPL(X) KEEPV(X)
	#if __x86_64__
	#if WIDTH==32
	#define INIT __m256 val = _mm256_set1_ps(123);
	#if UNALIGNED
	#define STORE _mm256_storeu_ps
	#define LOAD _mm256_loadu_ps
	#else
	#define STORE _mm256_store_ps
	#define LOAD _mm256_load_ps
	#endif
	#elif WIDTH==16
	#define INIT __m128 val = _mm_set1_ps(123);
	#if UNALIGNED
	#define STORE _mm_storeu_ps
	#define LOAD _mm_loadu_ps
	#else
	#define STORE _mm_store_ps
	#define LOAD _mm_load_ps
	#endif
	#endif
	#elif __aarch64__
	#define INIT float32x4_t val = vdupq_n_f32(123);
	#if WIDTH==16
	#define LOAD vld1q_f32
	#define STORE vst1q_f32
	#elif WIDTH==32
	#undef KEEPL
	#define KEEPL(X) ({ __auto_type x=X; KEEPV(x[0]); KEEPV(x[1]); })
	#define LOAD(P) (float32x4_t[2]){vld1q_f32(P), vld1q_f32(P+4)}
	#define STORE(P,V) vst1q_f32(P,V); vst1q_f32(P+4,V);
	#endif
	#endif
	#else
	#define INIT TY val = IDENTG((TY)123);
	#if WIDTH==8
	#define TY u64
	#elif WIDTH==4
	#define TY u32
	#elif WIDTH==2
	#define TY u16
	#elif WIDTH==1
	#define TY u8
	#endif
	#define KEEPL(X) KEEPG(X)
	#define STORE(P, V) (TY)(P) = (V)
	#define LOAD(P) (TY)(P)
	#endif

	NOINLINE u64 testThroughput(u8* data, u64 testsize, u64 rep, u64 breakonns, bool store) {
	u64 tns = 0;
	u64 minns = 1e18;
	while(tns < 1e8) {
	u64 sns = nsTime();
	INIT;
	if (store) {
	for (u64 k = 0; k < rep; k++) {
	for (u64 i = 0; i < testsize; i+= WIDTH*32) {
	u8* ptr = IDENTG(data+i); // IDENTG to avoid the addition propagating to complex addressing modes
	#pragma GCC unroll(32)
	for (u64 j = 0; j < 32; j++) { STORE((float)(ptr + jWIDTH), val); MEMM; }
	}
	}
	} else {
	for (u64 k = 0; k < rep; k++) {
	for (u64 i = 0; i < testsize; i+= WIDTH*32) {
	u8* ptr = IDENTG(data+i);
	#pragma GCC unroll(32)
	for (u64 j = 0; j < 32; j++) KEEPL(LOAD((float)(ptr + jWIDTH)));
	}
	}
	}

	u64 ens = nsTime();
	u64 cns = (ens-sns) / rep;
	if (cns < minns) {
	minns = cns;
	if (minns < breakonns) return minns;
	}
	tns+= ens-sns;
	}
	return minns;
	}
	NOINLINE u64 testLatency(u8* data, u64 bytes, u64 rep, u64 breakonns, bool unused) {
	u64 tns = 0;
	u64 minns = 1e18;
	while(tns < 1e8) {
	u64 sns = nsTime();
	for (u64 k = 0; k < rep; k++) {
	void ptr = (void)data;
	for (u64 i = 0; i < bytes/sizeof(void*); i+= 32) {
	#pragma GCC unroll(32)
	for (u64 j = 0; j < 32; j++) ptr = *ptr;
	}
	KEEPG(ptr);
	}
	u64 ens = nsTime();
	u64 cns = (ens-sns) / rep;
	if (cns < minns) {
	minns = cns;
	if (minns < breakonns) return minns;
	}
	tns+= ens-sns;
	}
	return minns;
	}

	#if TEST_LAT
	#define test testLatency
	#else
	#define test testThroughput
	#endif

	int main() {
	u64 max = 104857600; // 100MB
	u8* data = malloc(max+10240);
	data = (u8*) (((u64)data+4096) & ~(u64)4095);
	#if UNALIGNED
	data+= UNALIGNED;
	#endif
	#if TEST_LAT
	u64 seed = 12345;
	void** indexes = malloc(max+10240);
	for (u64 i = 0; i < max/sizeof(void); i++) indexes[i] = (void)i;
	#else
	for (u64 i = 0; i < max; i++) data[i] = i;
	#endif


	u64 bytes = 1024;
	while (bytes < max) {
	#if TEST_LAT
	// create random pointer chain
	u64 ptram = bytes/sizeof(void*);
	for (u64 i = 0; i < ptram; i++) { // shuffle ↕ptram pointers
	u64 j = wy2u0k(wyrand(&seed), i);
	void* t=indexes[i];
	indexes[i] = indexes[j];
	indexes[j] = t;
	}
	void ptrs = (void)data;
	u64 ci = indexes[ptram]-NULL;
	for (u64 i = 0; i < ptram+1; i++) { // create the chain
	ptrs[ci] = ptrs + (indexes[i]-NULL);
	ci = indexes[i]-NULL;
	}
	#endif

	u64 nses[2];
	for (int i = 0; i < (TEST_LAT? 1 : 2); i++) {
	#if TP_ONLY
	if (i!=TP_ONLY-1) continue;
	#endif
	u64 minns = test(data, bytes, 1, 10000, false);
	if (minns < 10000) {
	u64 rep = 10000 / minns;
	minns = test(data, bytes, rep, 0, i);
	}
	nses[i] = minns;
	}

	#if TEST_LAT
	printf("%9ldB: %6.2fns, %6.2fc; %ld ptrs, %ld total ns\n", bytes, nses[0]1.0/ptram, (nses[0]/1e9CLOCK_RATE)/ptram, ptram, nses[0]);
	#else
	printf("%9ldB: ", bytes);
	#if !TP_ONLY \|\| TP_ONLY==1
	printf("load:%7.2fGB/s,%6.2fB/c", bytes/(nses[0]/1e9) / 1e9, bytes/(nses[0]/1e9) / CLOCK_RATE);
	#endif
	#ifndef TP_ONLY
	printf(" ");
	#endif
	#if !TP_ONLY \|\| TP_ONLY==2
	printf("store:%7.2fGB/s,%6.2fB/c", bytes/(nses[1]/1e9) / 1e9, bytes/(nses[1]/1e9) / CLOCK_RATE);
	#endif
	printf("\n");
	#endif
	fflush(stdout);

	bytes*= 1.1;
	bytes = (bytes+1023) & ~(u64)1023;
	}
	}