perfaware/part2: Do RDTSC homework

build.bat

@@ -324,4 +324,8 @@ cl %part2_dir%\haversine_generator.c /W4 /WX /Z7 /nologo /O2 /Fe:haversine_gener
 
 cl %part2_dir%\haversine.c /W4 /WX /Z7 /nologo     /Fe:haversine_debug || exit /b 1
 cl %part2_dir%\haversine.c /W4 /WX /Z7 /nologo /O2 /Fe:haversine_release || exit /b 1
+
+cl %part2_dir%\listing_0071_os_timer_main.cpp /W4 /WX /Z7 /O2 /nologo /Fe:listing_0071_os_timer_main_release || exit /b 1
+cl %part2_dir%\listing_0072_cpu_timer_main.cpp /W4 /WX /Z7 /O2 /nologo /Fe:listing_0072_cpu_timer_main_release || exit /b 1
+cl %part2_dir%\listing_0073_cpu_timer_guessfreq_main.cpp /W4 /WX /Z7 /O2 /nologo /Fe:listing_0073_cpu_timer_guessfreq_release || exit /b 1
 popd

part2/haversine.c

@@ -6,7 +6,9 @@
 #include "haversine_stdlib.h"
 #include "haversine_stdlib.c"
 #include <math.h>
+
 #include "listing_0065_haversine_formula.cpp"
+#include "listing_0074_platform_metrics.cpp"
 
 typedef struct Str8FindResult {
     bool     found;
@@ -84,6 +86,8 @@ int main(int argc, char **argv)
         return -1;
     }
 
+    u64 cpu_start_time = ReadCPUTimer();
+
     HAV_Str8 arg_json    = {argv[1], strlen(argv[1])};
     HAV_Str8 arg_answers = {0};
     if (argc == 3)
@@ -93,9 +97,15 @@ int main(int argc, char **argv)
     if (!HAV_BufferIsValid(json_buffer))
         return 0;
 
-    f64 haversine_sum = 0;
+    u64 cpu_misc_setup_time = ReadCPUTimer();
-    size_t pair_count = 0;
+
-    HAV_Str8 json_it  = (HAV_Str8){.data = json_buffer.data, .size = json_buffer.size};
+    u64 cpu_elapsed_parse_time         = 0;
+    u64 cpu_elapsed_haversine_sum_time = 0;
+
+    f64 haversine_sum        = 0;
+    size_t pair_count        = 0;
+    HAV_Str8 json_it         = (HAV_Str8){.data = json_buffer.data, .size = json_buffer.size};
+    u64 cpu_begin_parse_time = ReadCPUTimer();
     for (;; pair_count++) {
         HAV_Str8BinarySplitResult x0_key = HAV_Str8_BinarySplit(json_it, HAV_STR8("x0"));
         if (!x0_key.rhs.size)
@@ -121,6 +131,9 @@ int main(int argc, char **argv)
         f64 x1 = StringToF64(x1_value.lhs);
         f64 y1 = StringToF64(y1_value.lhs);
 
+        u64 cpu_end_parse_time = ReadCPUTimer();
+        cpu_elapsed_parse_time += cpu_end_parse_time - cpu_begin_parse_time;
+
         #if 0
         HAV_PrintLnFmt("{x0: %.*s (%f), y0: %.*s (%f), x1 %.*s (%f), y1: %.*s (%f)}",
                        HAV_STR8_FMT(x0_value.lhs), x0,
@@ -129,12 +142,17 @@ int main(int argc, char **argv)
                        HAV_STR8_FMT(y1_value.lhs), y1);
         #endif
 
-        json_it = y1_value.rhs;
-
         f64 haversine_dist = ReferenceHaversine(x0, y0, x1, y1, /*EarthRadius*/ 6372.8);
         haversine_sum     += haversine_dist;
+        u64 cpu_end_sum_time = ReadCPUTimer();
+        cpu_elapsed_haversine_sum_time += cpu_end_sum_time - cpu_end_parse_time;
+
+        cpu_begin_parse_time = cpu_end_sum_time;
+        json_it              = y1_value.rhs;
     }
 
+    u64 cpu_end_parse_and_sum_time = ReadCPUTimer();
+
     haversine_sum /= pair_count;
     size_t input_size = json_buffer.size;
     HAV_PrintLnFmt("Input size: %zu", input_size);
@@ -154,7 +172,21 @@ int main(int argc, char **argv)
             HAV_PrintLnFmt("Reference sum: %f", reference_haversine_sum);
             HAV_PrintLnFmt("Difference: %f", difference);
         }
-
     }
+
+    u64 cpu_end_time = ReadCPUTimer();
+    u64 cpu_elapsed_time = cpu_end_time - cpu_start_time;
+    u64 cpu_frequency  = EstimateCPUTimerFreq();
+    if (cpu_frequency)
+        printf("\nTotal time: %0.4fms (CPU freq %llu)\n", 1000.0 * (f64)cpu_elapsed_time / (f64)cpu_frequency, cpu_frequency);
+
+    u64 cpu_elapsed_setup_time = cpu_misc_setup_time - cpu_start_time;
+    u64 cpu_elapsed_verify_time = cpu_end_time - cpu_end_parse_and_sum_time;
+
+    printf("   Setup: %llu (%.2f%%)\n", cpu_elapsed_setup_time,         (f64)cpu_elapsed_setup_time         / (f64)cpu_elapsed_time * 100.0);
+    printf("   Parse: %llu (%.2f%%)\n", cpu_elapsed_parse_time,         (f64)cpu_elapsed_parse_time         / (f64)cpu_elapsed_time * 100.0);
+    printf("     Sum: %llu (%.2f%%)\n", cpu_elapsed_haversine_sum_time, (f64)cpu_elapsed_haversine_sum_time / (f64)cpu_elapsed_time * 100.0);
+    printf("  Verify: %llu (%.2f%%)\n", cpu_elapsed_verify_time,        (f64)cpu_elapsed_verify_time        / (f64)cpu_elapsed_time * 100.0);
+
     return 0;
 }

part2/haversine_stdlib.h

@@ -1,3 +1,5 @@
+#include <stdint.h>
+
 // NOTE: Macros
 // ============================================================================
 #define HAV_STRINGIFY2(token) #token
@@ -19,6 +21,7 @@
 
 typedef float f32;
 typedef double f64;
+typedef uint64_t u64;
 
 // NOTE: Globals
 // ============================================================================

part2/listing_0070_platform_metrics.cpp

@@ -0,0 +1,70 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 70
+   ======================================================================== */
+
+#if _WIN32
+
+#include <intrin.h>
+#include <windows.h>
+
+static u64 GetOSTimerFreq(void)
+{
+	LARGE_INTEGER Freq;
+	QueryPerformanceFrequency(&Freq);
+	return Freq.QuadPart;
+}
+
+static u64 ReadOSTimer(void)
+{
+	LARGE_INTEGER Value;
+	QueryPerformanceCounter(&Value);
+	return Value.QuadPart;
+}
+
+#else
+
+#include <x86intrin.h>
+#include <sys/time.h>
+
+static u64 GetOSTimerFreq(void)
+{
+	return 1000000;
+}
+
+static u64 ReadOSTimer(void)
+{
+	// NOTE(casey): The "struct" keyword is not necessary here when compiling in C++,
+	// but just in case anyone is using this file from C, I include it.
+	struct timeval Value;
+	gettimeofday(&Value, 0);
+	
+	u64 Result = GetOSTimerFreq()*(u64)Value.tv_sec + (u64)Value.tv_usec;
+	return Result;
+}
+
+#endif
+
+/* NOTE(casey): This does not need to be "inline", it could just be "static"
+   because compilers will inline it anyway. But compilers will warn about 
+   static functions that aren't used. So "inline" is just the simplest way 
+   to tell them to stop complaining about that. */
+inline u64 ReadCPUTimer(void)
+{
+	// NOTE(casey): If you were on ARM, you would need to replace __rdtsc
+	// with one of their performance counter read instructions, depending
+	// on which ones are available on your platform.
+	
+	return __rdtsc();
+}

part2/listing_0071_os_timer_main.cpp

@@ -0,0 +1,43 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 71
+   ======================================================================== */
+
+#include <stdint.h>
+#include <stdio.h>
+
+typedef uint64_t u64;
+typedef double f64;
+
+#include "listing_0070_platform_metrics.cpp"
+
+int main(void)
+{
+	u64 OSFreq = GetOSTimerFreq();
+	printf("    OS Freq: %llu\n", OSFreq);
+
+	u64 OSStart = ReadOSTimer();
+	u64 OSEnd = 0;
+	u64 OSElapsed = 0;
+	while(OSElapsed < OSFreq)
+	{
+		OSEnd = ReadOSTimer();
+		OSElapsed = OSEnd - OSStart;
+	}
+	
+	printf("   OS Timer: %llu -> %llu = %llu elapsed\n", OSStart, OSEnd, OSElapsed);
+	printf(" OS Seconds: %.4f\n", (f64)OSElapsed/(f64)OSFreq);
+	
+    return 0;
+}

part2/listing_0072_cpu_timer_main.cpp

@@ -0,0 +1,49 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 72
+   ======================================================================== */
+
+#include <stdint.h>
+#include <stdio.h>
+
+typedef uint64_t u64;
+typedef double f64;
+
+#include "listing_0070_platform_metrics.cpp"
+
+int main(void)
+{
+	u64 OSFreq = GetOSTimerFreq();
+	printf("    OS Freq: %llu\n", OSFreq);
+
+	u64 CPUStart = ReadCPUTimer();
+	u64 OSStart = ReadOSTimer();
+	u64 OSEnd = 0;
+	u64 OSElapsed = 0;
+	while(OSElapsed < OSFreq)
+	{
+		OSEnd = ReadOSTimer();
+		OSElapsed = OSEnd - OSStart;
+	}
+	
+	u64 CPUEnd = ReadCPUTimer();
+	u64 CPUElapsed = CPUEnd - CPUStart;
+	
+	printf("   OS Timer: %llu -> %llu = %llu elapsed\n", OSStart, OSEnd, OSElapsed);
+	printf(" OS Seconds: %.4f\n", (f64)OSElapsed/(f64)OSFreq);
+	
+	printf("  CPU Timer: %llu -> %llu = %llu elapsed\n", CPUStart, CPUEnd, CPUElapsed);
+	
+    return 0;
+}

part2/listing_0073_cpu_timer_guessfreq_main.cpp

@@ -0,0 +1,63 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 73
+   ======================================================================== */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+typedef uint64_t u64;
+typedef double f64;
+
+#include "listing_0070_platform_metrics.cpp"
+
+int main(int ArgCount, char **Args)
+{
+	u64 MillisecondsToWait = 1000;
+	if(ArgCount == 2)
+	{
+		MillisecondsToWait = atol(Args[1]);
+	}
+
+	u64 OSFreq = GetOSTimerFreq();
+	printf("    OS Freq: %llu (reported)\n", OSFreq);
+
+	u64 CPUStart = ReadCPUTimer();
+	u64 OSStart = ReadOSTimer();
+	u64 OSEnd = 0;
+	u64 OSElapsed = 0;
+	u64 OSWaitTime = OSFreq * MillisecondsToWait / 1000;
+	while(OSElapsed < OSWaitTime)
+	{
+		OSEnd = ReadOSTimer();
+		OSElapsed = OSEnd - OSStart;
+	}
+	
+	u64 CPUEnd = ReadCPUTimer();
+	u64 CPUElapsed = CPUEnd - CPUStart;
+	u64 CPUFreq = 0;
+	if(OSElapsed)
+	{
+		CPUFreq = OSFreq * CPUElapsed / OSElapsed;
+	}
+	
+	printf("   OS Timer: %llu -> %llu = %llu elapsed\n", OSStart, OSEnd, OSElapsed);
+	printf(" OS Seconds: %.4f\n", (f64)OSElapsed/(f64)OSFreq);
+	
+	printf("  CPU Timer: %llu -> %llu = %llu elapsed\n", CPUStart, CPUEnd, CPUElapsed);
+	printf("   CPU Freq: %llu (guessed)\n", CPUFreq);
+	
+    return 0;
+}

part2/listing_0074_platform_metrics.cpp

@@ -0,0 +1,98 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 74
+   ======================================================================== */
+
+#if _WIN32
+
+#include <intrin.h>
+#include <windows.h>
+
+static u64 GetOSTimerFreq(void)
+{
+	LARGE_INTEGER Freq;
+	QueryPerformanceFrequency(&Freq);
+	return Freq.QuadPart;
+}
+
+static u64 ReadOSTimer(void)
+{
+	LARGE_INTEGER Value;
+	QueryPerformanceCounter(&Value);
+	return Value.QuadPart;
+}
+
+#else
+
+#include <x86intrin.h>
+#include <sys/time.h>
+
+static u64 GetOSTimerFreq(void)
+{
+	return 1000000;
+}
+
+static u64 ReadOSTimer(void)
+{
+	// NOTE(casey): The "struct" keyword is not necessary here when compiling in C++,
+	// but just in case anyone is using this file from C, I include it.
+	struct timeval Value;
+	gettimeofday(&Value, 0);
+	
+	u64 Result = GetOSTimerFreq()*(u64)Value.tv_sec + (u64)Value.tv_usec;
+	return Result;
+}
+
+#endif
+
+/* NOTE(casey): This does not need to be "inline", it could just be "static"
+   because compilers will inline it anyway. But compilers will warn about 
+   static functions that aren't used. So "inline" is just the simplest way 
+   to tell them to stop complaining about that. */
+inline u64 ReadCPUTimer(void)
+{
+	// NOTE(casey): If you were on ARM, you would need to replace __rdtsc
+	// with one of their performance counter read instructions, depending
+	// on which ones are available on your platform.
+	
+	return __rdtsc();
+}
+
+static u64 EstimateCPUTimerFreq(void)
+{
+	u64 MillisecondsToWait = 100;
+	u64 OSFreq = GetOSTimerFreq();
+
+	u64 CPUStart = ReadCPUTimer();
+	u64 OSStart = ReadOSTimer();
+	u64 OSEnd = 0;
+	u64 OSElapsed = 0;
+	u64 OSWaitTime = OSFreq * MillisecondsToWait / 1000;
+	while(OSElapsed < OSWaitTime)
+	{
+		OSEnd = ReadOSTimer();
+		OSElapsed = OSEnd - OSStart;
+	}
+	
+	u64 CPUEnd = ReadCPUTimer();
+	u64 CPUElapsed = CPUEnd - CPUStart;
+	
+	u64 CPUFreq = 0;
+	if(OSElapsed)
+	{
+		CPUFreq = OSFreq * CPUElapsed / OSElapsed;
+	}
+	
+	return CPUFreq;
+}

part2/listing_0075_timed_haversine_main.cpp

@@ -0,0 +1,209 @@
+/* ========================================================================
+
+   (C) Copyright 2023 by Molly Rocket, Inc., All Rights Reserved.
+   
+   This software is provided 'as-is', without any express or implied
+   warranty. In no event will the authors be held liable for any damages
+   arising from the use of this software.
+   
+   Please see https://computerenhance.com for more information
+   
+   ======================================================================== */
+
+/* ========================================================================
+   LISTING 75
+   ======================================================================== */
+
+/* NOTE(casey): _CRT_SECURE_NO_WARNINGS is here because otherwise we cannot
+   call fopen(). If we replace fopen() with fopen_s() to avoid the warning,
+   then the code doesn't compile on Linux anymore, since fopen_s() does not
+   exist there.
+   
+   What exactly the CRT maintainers were thinking when they made this choice,
+   I have no idea. */
+#define _CRT_SECURE_NO_WARNINGS
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <math.h>
+#include <sys/stat.h>
+
+typedef uint8_t u8;
+typedef uint32_t u32;
+typedef uint64_t u64;
+
+typedef int32_t b32;
+
+typedef float f32;
+typedef double f64;
+
+struct haversine_pair
+{
+    f64 X0, Y0;
+    f64 X1, Y1;
+};
+
+#include "listing_0074_platform_metrics.cpp"
+#include "listing_0065_haversine_formula.cpp"
+#include "listing_0068_buffer.cpp"
+#include "listing_0069_lookup_json_parser.cpp"
+
+static buffer ReadEntireFile(char *FileName)
+{
+    buffer Result = {};
+        
+    FILE *File = fopen(FileName, "rb");
+    if(File)
+    {
+#if _WIN32
+        struct __stat64 Stat;
+        _stat64(FileName, &Stat);
+#else
+        struct stat Stat;
+        stat(FileName, &Stat);
+#endif
+        
+        Result = AllocateBuffer(Stat.st_size);
+        if(Result.Data)
+        {
+            if(fread(Result.Data, Result.Count, 1, File) != 1)
+            {
+                fprintf(stderr, "ERROR: Unable to read \"%s\".\n", FileName);
+                FreeBuffer(&Result);
+            }
+        }
+    }
+    else
+    {
+        fprintf(stderr, "ERROR: Unable to open \"%s\".\n", FileName);
+    }
+    
+    return Result;
+}
+
+static f64 SumHaversineDistances(u64 PairCount, haversine_pair *Pairs)
+{
+    f64 Sum = 0;
+    
+    f64 SumCoef = 1 / (f64)PairCount;
+    for(u64 PairIndex = 0; PairIndex < PairCount; ++PairIndex)
+    {
+        haversine_pair Pair = Pairs[PairIndex];
+        f64 EarthRadius = 6372.8;
+        f64 Dist = ReferenceHaversine(Pair.X0, Pair.Y0, Pair.X1, Pair.Y1, EarthRadius);
+        Sum += SumCoef*Dist;
+    }
+    
+    return Sum;
+}
+
+static void PrintTimeElapsed(char const *Label, u64 TotalTSCElapsed, u64 Begin, u64 End)
+{
+    u64 Elapsed = End - Begin;
+    f64 Percent = 100.0 * ((f64)Elapsed / (f64)TotalTSCElapsed);
+    printf("  %s: %llu (%.2f%%)\n", Label, Elapsed, Percent);
+}
+
+int main(int ArgCount, char **Args)
+{
+    u64 Prof_Begin = 0;
+    u64 Prof_Read = 0;
+    u64 Prof_MiscSetup = 0;
+    u64 Prof_Parse = 0;
+    u64 Prof_Sum = 0;
+    u64 Prof_MiscOutput = 0;
+    u64 Prof_End = 0;
+    
+    Prof_Begin = ReadCPUTimer();
+    
+    int Result = 1;
+    
+    if((ArgCount == 2) || (ArgCount == 3))
+    {
+        Prof_Read = ReadCPUTimer();
+        buffer InputJSON = ReadEntireFile(Args[1]);
+        Prof_MiscSetup = ReadCPUTimer();
+        
+        u32 MinimumJSONPairEncoding = 6*4;
+        u64 MaxPairCount = InputJSON.Count / MinimumJSONPairEncoding;
+        if(MaxPairCount)
+        {
+            buffer ParsedValues = AllocateBuffer(MaxPairCount * sizeof(haversine_pair));
+            if(ParsedValues.Count)
+            {
+                haversine_pair *Pairs = (haversine_pair *)ParsedValues.Data;
+                
+                Prof_Parse = ReadCPUTimer();
+                u64 PairCount = ParseHaversinePairs(InputJSON, MaxPairCount, Pairs);
+                Prof_Sum = ReadCPUTimer();
+                f64 Sum = SumHaversineDistances(PairCount, Pairs);
+                Prof_MiscOutput = ReadCPUTimer();
+                
+                Result = 0;
+                
+                fprintf(stdout, "Input size: %llu\n", InputJSON.Count);
+                fprintf(stdout, "Pair count: %llu\n", PairCount);
+                fprintf(stdout, "Haversine sum: %.16f\n", Sum);
+                
+                if(ArgCount == 3)
+                {
+                    buffer AnswersF64 = ReadEntireFile(Args[2]);
+                    if(AnswersF64.Count >= sizeof(f64))
+                    {
+                        f64 *AnswerValues = (f64 *)AnswersF64.Data;
+                        
+                        fprintf(stdout, "\nValidation:\n");
+                        
+                        u64 RefAnswerCount = (AnswersF64.Count - sizeof(f64)) / sizeof(f64);
+                        if(PairCount != RefAnswerCount)
+                        {
+                            fprintf(stdout, "FAILED - pair count doesn't match %llu.\n", RefAnswerCount);
+                        }
+                        
+                        f64 RefSum = AnswerValues[RefAnswerCount];
+                        fprintf(stdout, "Reference sum: %.16f\n", RefSum);
+                        fprintf(stdout, "Difference: %.16f\n", Sum - RefSum);
+                        
+                        fprintf(stdout, "\n");
+                    }
+                }
+            }
+            
+            FreeBuffer(&ParsedValues);
+        }
+        else
+        {
+            fprintf(stderr, "ERROR: Malformed input JSON\n");
+        }
+        
+        FreeBuffer(&InputJSON);
+    }
+    else
+    {
+        fprintf(stderr, "Usage: %s [haversine_input.json]\n", Args[0]);
+        fprintf(stderr, "       %s [haversine_input.json] [answers.f64]\n", Args[0]);
+    }
+    
+    Prof_End = ReadCPUTimer();
+    
+    if(Result == 0)
+    {
+        u64 TotalCPUElapsed = Prof_End - Prof_Begin;
+        
+        u64 CPUFreq = EstimateCPUTimerFreq();
+        if(CPUFreq)
+        {
+            printf("\nTotal time: %0.4fms (CPU freq %llu)\n", 1000.0 * (f64)TotalCPUElapsed / (f64)CPUFreq, CPUFreq);
+        }
+        
+        PrintTimeElapsed("Startup", TotalCPUElapsed, Prof_Begin, Prof_Read);
+        PrintTimeElapsed("Read", TotalCPUElapsed, Prof_Read, Prof_MiscSetup);
+        PrintTimeElapsed("MiscSetup", TotalCPUElapsed, Prof_MiscSetup, Prof_Parse);
+        PrintTimeElapsed("Parse", TotalCPUElapsed, Prof_Parse, Prof_Sum);
+        PrintTimeElapsed("Sum", TotalCPUElapsed, Prof_Sum, Prof_MiscOutput);
+        PrintTimeElapsed("MiscOutput", TotalCPUElapsed, Prof_MiscOutput, Prof_End);
+    }
+    
+    return Result;
+}