medfall

commit fdb946a5de01a3bbf38d4779baebf7756ae51e7b
parent 2124510c91023c4b1afdf85cab96d3b5bcc10385
Author: Michael Savage <mikejsavage@gmail.com>
Date:   Sat Jun  3 01:18:41 +0300

Use rdtsc in the profiler

Diffstat:
platform_rdtsc.h
 | 
16
++++++++++++++++
profiler.cc
 | 
32
++++++++++++++++----------------
stats.cc
 | 
10
++++++----
3 files changed, 38 insertions(+), 20 deletions(-)
diff --git a/platform_rdtsc.h b/platform_rdtsc.h
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "platform.h"
+#include "platform_inline.h"
+
+#if COMPILER_MSVC
+#include <intrin.h>
+#elif COMPILER_GCCORCLANG
+#include <x86intrin.h>
+#else
+#error new compiler
+#endif
+
+forceinline u64 rdtsc() {
+	return __rdtsc();
+}
diff --git a/profiler.cc b/profiler.cc
@@ -8,7 +8,7 @@
 #include "strlcat.h"
 #include "ggformat.h"
 #include "log.h"
-#include "platform_time.h"
+#include "platform_rdtsc.h"
 
 struct Timer {
 	const char * name;
@@ -19,12 +19,12 @@ struct Timer {
 
 struct StackTimer {
 	u32 timer_idx;
-	double pushed_at;
+	u64 pushed_at;
 };
 
 struct LogEntry {
 	str< 256 > message;
-	double dt;
+	u64 clocks;
 	u32 first_child;
 	u32 prev_sibling;
 };
@@ -59,7 +59,7 @@ void profiler_push( u32 timer_idx ) {
 	ASSERT( timer_stack_top < TIMER_STACK_SIZE );
 
 	timer_stack[ timer_stack_top ].timer_idx = timer_idx;
-	timer_stack[ timer_stack_top ].pushed_at = get_time();
+	timer_stack[ timer_stack_top ].pushed_at = rdtsc();
 	timer_stack_top++;
 }
 
@@ -72,22 +72,22 @@ static void profiler_logger_init() {
 	log_file = fopen( log_path.c_str(), "w" );
 }
 
-static void write_node( u32 idx, double frame_dt ) {
+static void write_node( u32 idx, u64 frame_clocks ) {
 	const LogEntry & entry = logs[ idx ];
 	if( entry.prev_sibling != 0 ) {
-		write_node( entry.prev_sibling, frame_dt );
+		write_node( entry.prev_sibling, frame_clocks );
 	}
 
-	ggprint_to_file( log_file, "{} ({.2}%)\n", entry.message.c_str(), entry.dt * 100.0 / frame_dt );
+	ggprint_to_file( log_file, "{} ({.2}%)\n", entry.message.c_str(), entry.clocks * 100.0 / frame_clocks );
 
 	if( entry.first_child != 0 ) {
-		write_node( entry.first_child, frame_dt );
+		write_node( entry.first_child, frame_clocks );
 	}
 }
 
-static void write_frame( double frame_dt ) {
+static void write_frame( u64 frame_clocks ) {
 	profiler_logger_init();
-	write_node( num_log_entries - 1, frame_dt );
+	write_node( num_log_entries - 1, frame_clocks );
 	ggprint_to_file( log_file, "\n" );
 	fflush( log_file );
 }
@@ -95,14 +95,14 @@ static void write_frame( double frame_dt ) {
 void profiler_pop( const char * name ) {
 	ASSERT( timer_stack_top > 0 );
 
-	double now = get_time();
+	u64 now = rdtsc();
 
 	timer_stack_top--;
 	u32 idx = timer_stack[ timer_stack_top ].timer_idx;
-	double dt = now - timer_stack[ timer_stack_top ].pushed_at;
+	u64 clocks = now - timer_stack[ timer_stack_top ].pushed_at;
 
 	Timer * timer = &all_timers[ idx ];
-	stats_record( &timer->stats, dt );
+	stats_record( &timer->stats, clocks );
 
 	if( name == NULL ) {
 		name = timer->name;
@@ -115,8 +115,8 @@ void profiler_pop( const char * name ) {
 
 	ASSERT( num_log_entries < TIMER_LOG_SIZE );
 	LogEntry * entry = &logs[ num_log_entries ];
-	entry->message.sprintf( "{}{} ({}:{}) - {.3}ms", spaces, name, timer->file, timer->line, dt * 1000.0 );
-	entry->dt = dt;
+	entry->message.sprintf( "{}{} ({}:{}) - {.3} Kclocks", spaces, name, timer->file, timer->line, clocks / 1024 );
+	entry->clocks = clocks;
 	entry->first_child = last_at_depth[ timer_stack_top + 1 ];
 	entry->prev_sibling = last_at_depth[ timer_stack_top ];
 
@@ -126,7 +126,7 @@ void profiler_pop( const char * name ) {
 	num_log_entries++;
 
 	if( timer_stack_top == 0 ) {
-		write_frame( dt );
+		write_frame( clocks );
 		num_log_entries = 1;
 		memset( last_at_depth, 0, sizeof( last_at_depth ) );
 	}
diff --git a/stats.cc b/stats.cc
@@ -53,10 +53,12 @@ void format( FormatBuffer * fb, const Stats & stats, const FormatOpts & opts ) {
 	u32 num_samples = checked_cast< u32 >( min( stats.num_records, u64( STATS_NUM_QUART_SAMPLES ) ) );
 	std::sort( stats.samples, stats.samples + num_samples );
 
-	ggformat_impl( fb, "n = {}, mean = {}, stddev = {}, range = {} to {}, 25% = {}, 75% = {}",
-		stats.num_records,
+	ggformat_impl( fb, "total = {}, n = {}, mean = {}, stddev = {}, min = {}, 25% = {}, 75% = {}, max = {}",
+		stats.sum, stats.num_records,
 		stats_mean( &stats ), stats_stddev( &stats ),
-		stats.min, stats.max,
-		stats.samples[ num_samples / 4 ], stats.samples[ num_samples * 3 / 4 ]
+		stats.min,
+		stats.samples[ num_samples / 4 ],
+		stats.samples[ num_samples * 3 / 4 ],
+		stats.max
 	);
 }