inline atomics are faster

benchmark/amd7950x_avx2/joint_grid.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,2664.01
-2,1567.06
-3,1080.91
-4,872.047
-5,761.89
-6,675.953
-7,621.694
-8,574.929
+1,2565.3
+2,1533.85
+3,1053.01
+4,865.702
+5,748.989
+6,658.623
+7,613.633
+8,578.973

benchmark/amd7950x_avx2/large_pyramid.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,1570.29
-2,871.185
-3,613.11
-4,471.278
-5,401.635
-6,353.07
-7,311.721
-8,307.579
+1,1584.03
+2,841.862
+3,598.747
+4,471.697
+5,396.306
+6,347.983
+7,309.086
+8,308.832

benchmark/amd7950x_avx2/many_pyramids.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,2682.25
-2,1481.3
-3,977.918
-4,767.89
-5,605.041
-6,520.386
-7,446.211
-8,419.113
+1,2663.34
+2,1407.85
+3,934.506
+4,725.271
+5,590.288
+6,502.513
+7,422.693
+8,395.369

benchmark/amd7950x_avx2/rain.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,6819.14
-2,4114.75
-3,3104.05
-4,2562.53
-5,2182.8
-6,1946.25
-7,1759.5
-8,1628.84
+1,6526.9
+2,4017.74
+3,3052.28
+4,2493.5
+5,2149.12
+6,1911.45
+7,1735.97
+8,1625.71

benchmark/amd7950x_avx2/smash.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,1722.58
-2,1081.59
-3,812.772
-4,673.528
-5,591.922
-6,536.279
-7,502.897
-8,479.575
+1,1562.06
+2,1020.55
+3,781.193
+4,661.318
+5,580.115
+6,530.502
+7,489.672
+8,467.998

benchmark/amd7950x_avx2/spinner.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,4341.12
-2,2694.07
-3,2050.24
-4,1641.69
-5,1442.97
-6,1286.99
-7,1175.04
-8,1121.46
+1,4106.89
+2,2635.21
+3,1986.32
+4,1614.21
+5,1420.63
+6,1272.06
+7,1157.16
+8,1091.25

benchmark/amd7950x_avx2/tumbler.csv

@@ -1,9 +1,9 @@
 threads,ms
-1,1743.49
-2,1096.53
-3,824.097
-4,675.961
-5,592.045
-6,526.278
-7,489.226
-8,467.981
+1,1613.69
+2,1061.88
+3,810.738
+4,670.874
+5,578.959
+6,519.473
+7,479.842
+8,437.851

src/atomic.h

@@ -3,82 +3,77 @@
 
 #pragma once
 
-#include <stdint.h>
-
-// Compare to SDL_AtomicInt
-
-struct b3AtomicInt
-{
-	int value;
-};
+#include "core.h"
 
-struct b3AtomicU32
-{
-	uint32_t value;
-};
+#include <stdint.h>
 
 #if defined( _MSC_VER )
-	#include <intrin.h>
-
-static inline void b3AtomicStoreInt( b2AtomicInt* obj, int desired )
-{
-	_InterlockedExchange( (volatile long*)obj, (long)desired );
-}
-
-static inline int b3AtomicLoadInt( volatile int* obj )
-{
-	return _InterlockedOr( (volatile long*)obj, 0 );
-}
-
-static inline void b3AtomicStoreUInt( volatile uint32_t* obj, uint32_t desired )
-{
-	_InterlockedExchange( (volatile long*)obj, (long)desired );
-}
+#include <intrin.h>
+#endif
 
-static inline int b3AtomicFetchAddInt( volatile int* obj, int increment )
+static inline void b2AtomicStoreInt( b2AtomicInt* a, int value )
 {
-	return _InterlockedExchangeAdd( (volatile long*)obj, (long)increment );
+#if defined( _MSC_VER )
+	(void)_InterlockedExchange( (long*)&a->value, value );
+#elif defined( __GNUC__ ) || defined( __clang__ )
+	__atomic_store_n( &a->value, value, __ATOMIC_SEQ_CST );
+#else
+#error "Unsupported platform"
+#endif
 }
 
-static inline bool b3AtomicCompareExchangeInt( volatile int* obj, int* expected, int desired )
+static inline int b2AtomicLoadInt( b2AtomicInt* a )
 {
-	int original = _InterlockedCompareExchange( (volatile long*)obj, (long)desired, (long)*expected );
-	if ( original == *expected )
-	{
-		return true;
-	}
-
-	*expected = original;
-	return false;
-}
-
+#if defined( _MSC_VER )
+	return _InterlockedOr( (long*)&a->value, 0 );
 #elif defined( __GNUC__ ) || defined( __clang__ )
-
-static inline void b3AtomicStoreInt( volatile int* obj, int desired )
-{
-	__atomic_store_n( obj, desired, __ATOMIC_SEQ_CST );
+	return __atomic_load_n( &a->value, __ATOMIC_SEQ_CST );
+#else
+#error "Unsupported platform"
+#endif
 }
 
-static inline int b3AtomicLoadInt( volatile int* obj )
+static inline int b2AtomicFetchAddInt( b2AtomicInt* a, int increment )
 {
-	return __atomic_load_n( obj, __ATOMIC_SEQ_CST );
+#if defined( _MSC_VER )
+	return _InterlockedExchangeAdd( (long*)&a->value, (long)increment );
+#elif defined( __GNUC__ ) || defined( __clang__ )
+	return __atomic_fetch_add( &a->value, increment, __ATOMIC_SEQ_CST );
+#else
+#error "Unsupported platform"
+#endif
 }
 
-static inline void b3AtomicStoreUInt( volatile uint32_t* obj, uint32_t desired )
+static inline bool b2AtomicCompareExchangeInt( b2AtomicInt* a, int expected, int desired )
 {
-	__atomic_store_n( obj, desired, __ATOMIC_SEQ_CST );
+#if defined( _MSC_VER )
+	return _InterlockedCompareExchange( (long*)&a->value, (long)desired, (long)expected ) == expected;
+#elif defined( __GNUC__ ) || defined( __clang__ )
+	// The value written to expected is ignored
+	return __atomic_compare_exchange_n( &a->value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST );
+#else
+#error "Unsupported platform"
+#endif
 }
 
-static inline int b3AtomicFetchAddInt( volatile int* obj, int arg )
+static inline void b2AtomicStoreU32( b2AtomicU32* a, uint32_t value )
 {
-	return __atomic_fetch_add( obj, arg, __ATOMIC_SEQ_CST );
+#if defined( _MSC_VER )
+	(void)_InterlockedExchange( (long*)&a->value, value );
+#elif defined( __GNUC__ ) || defined( __clang__ )
+	__atomic_store_n( &a->value, value, __ATOMIC_SEQ_CST );
+#else
+#error "Unsupported platform"
+#endif
 }
 
-static inline bool b3AtomicCompareExchangeInt( volatile int* obj, int* expected, int desired )
+static inline uint32_t b2AtomicLoadU32( b2AtomicU32* a )
 {
-	return __atomic_compare_exchange_n( obj, expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST );
-}
-
+#if defined( _MSC_VER )
+	return (uint32_t)_InterlockedOr( (long*)&a->value, 0 );
+#elif defined( __GNUC__ ) || defined( __clang__ )
+	return __atomic_load_n( &a->value, __ATOMIC_SEQ_CST );
 #else
-	#error "Unsupported platform"
+#error "Unsupported platform"
 #endif
+}

src/broad_phase.c

@@ -9,6 +9,7 @@
 
 #include "aabb.h"
 #include "array.h"
+#include "atomic.h"
 #include "body.h"
 #include "contact.h"
 #include "core.h"

src/core.c

@@ -27,7 +27,7 @@
 
 #endif
 
-#include "box2d/math_functions.h"
+#include "atomic.h"
 
 // This allows the user to change the length units at runtime
 float b2_lengthUnitsPerMeter = 1.0f;
@@ -71,6 +71,7 @@ b2Version b2GetVersion( void )
 	return ( b2Version ){ 3, 1, 0 };
 }
 
+#if 0
 #if defined( _MSC_VER )
 #include <intrin.h>
 #endif
@@ -141,6 +142,7 @@ uint32_t b2AtomicLoadU32( b2AtomicU32* a )
 #error "Unsupported platform"
 #endif
 }
+#endif
 
 static b2AllocFcn* b2_allocFcn = NULL;
 static b2FreeFcn* b2_freeFcn = NULL;

src/core.h

@@ -142,10 +142,12 @@ typedef struct b2AtomicU32
 	uint32_t value;
 } b2AtomicU32;
 
+#if 0
 void b2AtomicStoreInt( b2AtomicInt* a, int value );
 int b2AtomicLoadInt( b2AtomicInt* a );
 int b2AtomicFetchAddInt( b2AtomicInt* a, int increment );
 bool b2AtomicCompareExchangeInt( b2AtomicInt* obj, int expected, int desired );
 
 void b2AtomicStoreU32( b2AtomicU32* a, uint32_t value );
 uint32_t b2AtomicLoadU32( b2AtomicU32* a );
+#endif

src/solver.c

@@ -4,6 +4,7 @@
 #include "solver.h"
 
 #include "array.h"
+#include "atomic.h"
 #include "bitset.h"
 #include "body.h"
 #include "contact.h"

src/table.c

@@ -3,6 +3,7 @@
 
 #include "table.h"
 
+#include "atomic.h"
 #include "core.h"
 #include "ctz.h"
 

test/test_table.c

@@ -1,6 +1,7 @@
 // SPDX-FileCopyrightText: 2023 Erin Catto
 // SPDX-License-Identifier: MIT
 
+#include "atomic.h"
 #include "core.h"
 #include "ctz.h"
 #include "table.h"
@@ -64,7 +65,7 @@ int TableTest( void )
 
 #if B2_SNOOP_TABLE_COUNTERS
 		extern b2AtomicInt b2_probeCount;
-		b2AtomicStoreInt(&b2_probeCount, 0);
+		b2AtomicStoreInt( &b2_probeCount, 0 );
 #endif
 
 		// Test key search