Performance benchmark for x86

6 years ago · df9180a30b
parent 1ca1046c57
commit df9180a30b
4 changed files with 312 additions and 5 deletions
--- a/tests/performance/benchmark.cpp
+++ b/tests/performance/benchmark.cpp
@ -0,0 +1,306 @@
+//RandomX performance test for x86
+//https://github.com/tevador/RandomX
+//License: GPL v3
+
+#include <cstdint> 
+#include <random>
+#include <iostream>
+#include <chrono>
+#include <sstream>
+#include <cmath>
+#include <cstring>
+
+#if defined(_WIN32) || defined(__MINGW32__) || defined(__CYGWIN__) || defined(__CYGWIN32__)
+  #define WINDOWS
+  #include <io.h>
+  #include <fcntl.h>
+#endif
+
+#if defined(__GNUC__) && defined(__x86_64__)
+  #include <x86intrin.h>
+  typedef unsigned __int128 uint128_t;
+  typedef __int128 int128_t;
+  static inline uint64_t umulhi64(uint64_t a, uint64_t b) {
+     return ((uint128_t)a * b) >> 64;
+  }
+  static inline uint64_t imulhi64(int64_t a, int64_t b) {
+     return ((int128_t)a * b) >> 64;
+  }
+  #define ror64 __rorq
+  #define rol64 __rolq
+  #define forceinline inline
+#elif defined(_MSC_VER) && defined(_M_X64)
+  #include <intrin.h>
+  #include <stdlib.h>
+  #define umulhi64 __umulh
+  static inline uint64_t imulhi64(int64_t a, int64_t b) {
+     int64_t hi;
+     _mul128(a, b, &hi);
+     return hi;
+  }
+  #define ror64 _rotr64
+  #define rol64 _rotl64
+  #define forceinline __forceinline
+#else
+	#error "Unsupported platform"
+#endif
+
+typedef union {
+	double f64;
+	int64_t i64;
+	uint64_t u64;
+	int32_t i32;
+	uint32_t u32;
+} convertible_t;
+
+forceinline void NOOP(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64;
+}
+
+forceinline void FNOOP(convertible_t& a, convertible_t& b, convertible_t& c) {
+	c.f64 = (double)a.i64;
+}
+
+forceinline void ADD_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 + b.u64;
+}
+
+forceinline void ADD_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u32 + b.u32;
+}
+
+forceinline void SUB_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 - b.u64;
+}
+
+forceinline void SUB_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u32 - b.u32;
+}
+
+forceinline void MUL_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 * b.u64;
+}
+
+forceinline void MULH_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = umulhi64(a.u64, b.u64);
+}
+
+forceinline void MUL_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = (uint64_t)a.u32 * b.u32;
+}
+
+forceinline void IMUL_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.i64 = (int64_t)a.i32 * b.i32;
+}
+
+forceinline void IMULH_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.i64 = imulhi64(a.i64, b.i64);
+}
+
+forceinline void DIV_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 / (b.u32 != 0 ? b.u32 : 1U);
+}
+
+forceinline void IDIV_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.i64 = a.i64 / (b.i32 != 0 ? b.i32 : 1);
+}
+
+forceinline void AND_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 & b.u64;
+}
+
+forceinline void AND_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u32 & b.u32;
+}
+
+forceinline void OR_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 | b.u64;
+}
+
+forceinline void OR_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u32 | b.u32;
+}
+
+forceinline void XOR_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 ^ b.u64;
+}
+
+forceinline void XOR_32(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u32 ^ b.u32;
+}
+
+forceinline void SHL_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 << (b.u64 & 63);
+}
+
+forceinline void SHR_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = a.u64 >> (b.u64 & 63);
+}
+
+forceinline void SAR_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.i64 = a.i64 >> (b.u64 & 63);
+}
+
+forceinline void ROL_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = rol64(a.u64, (b.u64 & 63));
+}
+
+forceinline void ROR_64(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.u64 = ror64(a.u64, (b.u64 & 63));
+}
+
+forceinline void FADD(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.f64 = (double)a.i64 + (double)b.i64;
+}
+
+forceinline void FSUB(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.f64 = (double)a.i64 - (double)b.i64;
+}
+
+forceinline void FMUL(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.f64 = (double)a.i64 * (double)b.i64;
+}
+
+forceinline void FDIV(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.f64 = (double)a.i64 / (double)b.i64;
+}
+
+forceinline void FSQRT(convertible_t& a, convertible_t& b, convertible_t& c) {
+  double d = fabs((double)a.i64);
+  c.f64 = _mm_cvtsd_f64(_mm_sqrt_sd(_mm_setzero_pd(), _mm_load_pd(&d)));
+}
+
+static uint32_t mxcsr;
+
+forceinline void FROUND(convertible_t& a, convertible_t& b, convertible_t& c) {
+  c.f64 = (double)a.i64;
+  _mm_setcsr(mxcsr | ((uint32_t)(a.u64 << 13) & _MM_ROUND_MASK));
+}
+
+inline void init_FPU() {
+  mxcsr = (_mm_getcsr() | _MM_FLUSH_ZERO_ON) & ~_MM_ROUND_MASK;
+  _mm_setcsr(mxcsr);
+}
+
+template<typename T>
+bool tryParse(char* buffer, T& out) {
+	std::istringstream ss(buffer);
+	if (!(ss >> out)) {
+		std::cout << "Invalid value '" << buffer << "'" << std::endl;
+		return false;
+	}
+	return true;
+}
+
+//#define ITERATIONS 10000000
+#define SCRATCHPAD_SIZE (16 * 1024)
+#define SCRATCHPAD_LENGTH (SCRATCHPAD_SIZE / sizeof(convertible_t))
+#define SCRATCHPAD_MASK (SCRATCHPAD_SIZE / sizeof(convertible_t) - 1)
+#define SCRATCHPAD_16K(x) scratchpad[(x) & SCRATCHPAD_MASK]
+
+#define BENCHMARK(FUNC,TYPE) do { \
+	memcpy((void*)scratchpad, input, SCRATCHPAD_SIZE); \
+	tstart = std::chrono::high_resolution_clock::now(); \
+	for (uint64_t i = 0; i < iterations; ++i) { \
+		FUNC(SCRATCHPAD_16K(i + 8 + 0), r0, SCRATCHPAD_16K(i + 0)); \
+		SCRATCHPAD_16K(i + 0).u64 ^= r7.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 1), r1, SCRATCHPAD_16K(i + 1)); \
+		SCRATCHPAD_16K(i + 1).u64 ^= r6.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 2), r2, SCRATCHPAD_16K(i + 2)); \
+		SCRATCHPAD_16K(i + 2).u64 ^= r5.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 3), r3, SCRATCHPAD_16K(i + 3)); \
+		SCRATCHPAD_16K(i + 3).u64 ^= r4.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 4), r4, SCRATCHPAD_16K(i + 4)); \
+		SCRATCHPAD_16K(i + 4).u64 ^= r3.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 5), r5, SCRATCHPAD_16K(i + 5)); \
+		SCRATCHPAD_16K(i + 5).u64 ^= r2.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 6), r6, SCRATCHPAD_16K(i + 6)); \
+		SCRATCHPAD_16K(i + 6).u64 ^= r1.u64;\
+		FUNC(SCRATCHPAD_16K(i + 8 + 7), r7, SCRATCHPAD_16K(i + 7)); \
+		SCRATCHPAD_16K(i + 7).u64 ^= r0.u64;\
+	} \
+	tend = std::chrono::high_resolution_clock::now(); \
+	uint64_t acum = 0; \
+	for (int i = 0; i < SCRATCHPAD_LENGTH; ++i) \
+		acum += scratchpad[i].u64; \
+	std::cout << "| " << #FUNC << " | " << std::chrono::duration<double>(tend - tstart).count() << " | " << acum << " |" << std::endl; \
+  } while(false)
+
+
+int main(int argc, char** argv) {
+	uint64_t iterations;
+	if (argc > 1) {
+		if (!tryParse(argv[1], iterations))
+			return 1;
+	}
+	else {
+		iterations = 100000000;
+	}
+#ifdef WINDOWS
+  _setmode(_fileno(stdin), O_BINARY);
+#endif
+  convertible_t input[SCRATCHPAD_LENGTH];
+
+  std::cout << "Reading " << sizeof(input) << " bytes from STDIN..." << std::endl;
+  std::cin.read((char*)input, sizeof(input));
+
+  if (!std::cin) {
+	  std::cerr << "Insufficient input" << std::endl;
+	  return 1;
+  }
+
+  convertible_t scratchpad[SCRATCHPAD_LENGTH];
+  convertible_t r0, r1, r2, r3, r4, r5, r6, r7;
+  
+  r0.u64 = input[0].u64;
+  r1.u64 = input[1].u64;
+  r2.u64 = input[2].u64;
+  r3.u64 = input[3].u64;
+  r4.u64 = input[4].u64;
+  r5.u64 = input[5].u64;
+  r6.u64 = input[6].u64;
+  r7.u64 = input[7].u64;
+
+  std::chrono::high_resolution_clock::time_point tstart, tend;
+
+  std::cout << iterations << " iterations:" << std::endl << std::endl;
+
+  std::cout << "| operation | time [s] | (result) |" << std::endl;
+  std::cout << "|-----------|----------|----------|" << std::endl;
+
+  BENCHMARK(NOOP, u64);
+  BENCHMARK(ADD_64, u64);
+  BENCHMARK(ADD_32, u64);
+  BENCHMARK(SUB_64, u64);
+  BENCHMARK(SUB_32, u64);
+  BENCHMARK(MUL_64, u64);
+  BENCHMARK(MULH_64, u64);
+  BENCHMARK(MUL_32, u64);
+  BENCHMARK(IMUL_32, u64);
+  BENCHMARK(IMULH_64, u64);
+  BENCHMARK(DIV_64, u64);
+  BENCHMARK(IDIV_64, u64);
+  BENCHMARK(AND_64, u64);
+  BENCHMARK(AND_32, u64);
+  BENCHMARK(OR_64, u64);
+  BENCHMARK(OR_32, u64);
+  BENCHMARK(XOR_64, u64);
+  BENCHMARK(XOR_32, u64);
+  BENCHMARK(SHL_64, u64);
+  BENCHMARK(SHR_64, u64);
+  BENCHMARK(SAR_64, u64);
+  BENCHMARK(ROR_64, u64);
+  BENCHMARK(ROL_64, u64);
+  
+  init_FPU();
+
+  BENCHMARK(FNOOP, f64);
+  BENCHMARK(FADD, f64);
+  BENCHMARK(FSUB, f64);
+  BENCHMARK(FMUL, f64);
+  BENCHMARK(FDIV, f64);
+  BENCHMARK(FSQRT, f64);
+  BENCHMARK(FROUND, f64);
+
+  return 0;
+}
--- a/tests/performance/test1.data
+++ b/tests/performance/test1.data
--- a/tests/performance/test2.data
+++ b/tests/performance/test2.data
--- a/tests/rx2c.py
+++ b/tests/rx2c.py
@ -126,12 +126,12 @@ def writeEpilog(file):
    file.write("\tend:\n")
    file.write("\t\tclockEnd = clock();\n")
    for i in range(8):
-        file.write('\t\tprintf("r{0} = %-36lu f{0} = %g\\n", r{0}, f{0});\n'.format(i))
+        file.write('\t\tprintf("r{0} = %-36" PRIu64 " f{0} = %g\\n", r{0}, f{0});\n'.format(i))
    file.write(("\t\tuint64_t spadsum = 0;\n"
        "\t\tfor(int i = 0; i < SCRATCHPAD_LENGTH; ++i) {\n"
        "\t\t	spadsum += scratchpad[i].u64;\n"
        "\t\t}\n"
-        '\t\tprintf("scratchpad sum = %lu\\n", spadsum);\n'
+        '\t\tprintf("scratchpad sum = %" PRIu64 "\\n", spadsum);\n'
        '\t\tprintf("runtime: %f\\n", (clockEnd - clockStart) / (double)CLOCKS_PER_SEC);\n'
        "#ifdef RAM\n"
        "\t\t_mm_free((void*)mmu.buffer);\n"
@ -632,8 +632,8 @@ def writeMain(file):
                "	register uint64_t r0, r1, r2, r3, r4, r5, r6, r7;\n"
                "	register double f0, f1, f2, f3, f4, f5, f6, f7;\n"
                "	register uint64_t ic, sp;\n"
-                "	convertible_t scratchpad[SCRATCHPAD_LENGTH] __attribute__ ((aligned (16)));\n"
                "	stack_t stack[STACK_LENGTH];\n"
+                "	convertible_t scratchpad[SCRATCHPAD_LENGTH] __attribute__ ((aligned (16)));\n"
                "	mmu_t mmu;\n"
                "	uint32_t mxcsr;\n"
                ))
@ -646,6 +646,7 @@ def writeProlog(file):
                "#include <emmintrin.h>\n"
                "#include <wmmintrin.h>\n"
                "#include <math.h>\n"
+                "#include <inttypes.h>\n"
                "typedef uint32_t addr_t;\n"
                "typedef unsigned __int128 uint128_t;\n"
                "typedef __int128 int128_t;\n"
@ -669,14 +670,14 @@ def writeProlog(file):
                "	const char* buffer;\n"
                "#endif\n"
                "} mmu_t;\n"
-                "#define DRAM_SIZE (1UL << 32)\n"
+                "#define DRAM_SIZE (1ULL << 32)\n"
                "#define SCRATCHPAD_SIZE (256 * 1024)\n"
                "#define SCRATCHPAD_LENGTH (SCRATCHPAD_SIZE / sizeof(convertible_t))\n"
                "#define SCRATCHPAD_MASK14 (16 * 1024 / sizeof(convertible_t) - 1)\n"
                "#define SCRATCHPAD_MASK18 (SCRATCHPAD_LENGTH - 1)\n"
                "#define SCRATCHPAD_16K(x) scratchpad[(x) & SCRATCHPAD_MASK14]\n"
                "#define SCRATCHPAD_256K(x) scratchpad[(x) & SCRATCHPAD_MASK18]\n"
-                "#define STACK_LENGTH (32 * 1024)\n"
+                "#define STACK_LENGTH (128 * 1024)\n"
                "#ifdef RAM\n"
                "#define DRAM_READ(mmu) (convertible_t)*(uint64_t*)((mmu)->buffer + (mmu)->m0)\n"
                "#define PREFETCH(mmu) _mm_prefetch(((mmu)->buffer + (mmu)->m1), _MM_HINT_T0)\n"