added float SIMD version for testing

default to SSE2 on x64 updated x64 benchmarks with float version
erincatto · Aug 19, 2024 · 5634f4b · 5634f4b
1 parent 5bea931
commit 5634f4b
Show file tree

Hide file tree

Showing 21 changed files with 507 additions and 336 deletions.
diff --git a/.clang-format b/.clang-format
@@ -33,7 +33,7 @@ IncludeCategories:
 
 IndentExternBlock: NoIndent
 IndentCaseLabels: true
-IndentPPDirectives: BeforeHash
+#IndentPPDirectives: None
 IndentAccessModifiers: false
 AccessModifierOffset: -4
 

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -1,8 +1,9 @@
 cmake_minimum_required(VERSION 3.22)
 include(FetchContent)
+include(CMakeDependentOption)
 
 project(box2d
-	VERSION 3.0.0
+	VERSION 3.0.1
 	DESCRIPTION "A 2D physics engine for games"
 	HOMEPAGE_URL "https://box2d.org"
 	LANGUAGES C CXX
@@ -34,10 +35,13 @@ if (MSVC OR APPLE)
 	endif()
 endif()
 
+option(BOX2D_ENABLE_SIMD "Enable SIMD math (faster)" ON)
+
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64")
-	option(BOX2D_AVX2 "Enable AVX2 (faster)" ON)
+	cmake_dependent_option(BOX2D_AVX2 "Enable AVX2" OFF "BOX2D_ENABLE_SIMD" OFF)
 endif()
 
+
 if(PROJECT_IS_TOP_LEVEL)
 	# Needed for samples.exe to find box2d.dll
 	# set(CMAKE_LIBRARY_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/bin")

diff --git a/benchmark/amd7950x/joint_grid.csv b/benchmark/amd7950x/joint_grid.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,331.343
-2,638.04
-3,932.731
-4,1200.15
-5,1480.23
-6,1718.79
-7,1930.12
-8,2133.65
+1,333.121
+2,638.057
+3,928.95
+4,1205.85
+5,1479.54
+6,1699.99
+7,1974.84
+8,2043.64
diff --git a/benchmark/amd7950x/large_pyramid.csv b/benchmark/amd7950x/large_pyramid.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,325.705
-2,616.127
-3,886.575
-4,1118.85
-5,1331.22
-6,1498.6
-7,1685.28
-8,1728.1
+1,336.895
+2,602.665
+3,878.207
+4,1117.02
+5,1304.79
+6,1482.92
+7,1663.97
+8,1661.13
diff --git a/benchmark/amd7950x/many_pyramids.csv b/benchmark/amd7950x/many_pyramids.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,82.8619
-2,160.906
-3,236.027
-4,300.688
-5,368.315
-6,429.822
-7,498.81
-8,549.271
+1,84.8025
+2,163.264
+3,234.388
+4,305.216
+5,369.85
+6,434.45
+7,497.573
+8,525.427
diff --git a/benchmark/amd7950x/smash.csv b/benchmark/amd7950x/smash.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,173.898
-2,277.19
-3,357.566
-4,430.528
-5,483.446
-6,525.652
-7,566.859
-8,598.553
+1,174.051
+2,276.742
+3,352.751
+4,421.773
+5,479.049
+6,522.318
+7,556.193
+8,586.672
diff --git a/benchmark/amd7950x/tumbler.csv b/benchmark/amd7950x/tumbler.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,373.066
-2,581.852
-3,764.444
-4,902.898
-5,1044.99
-6,1143.44
-7,1229.87
-8,1299.61
+1,376.3
+2,576.749
+3,737.749
+4,883.315
+5,1024.69
+6,1120.48
+7,1197.2
+8,1212.85
diff --git a/benchmark/amd7950x_float/joint_grid.csv b/benchmark/amd7950x_float/joint_grid.csv
@@ -0,0 +1,9 @@
+threads,fps
+1,362.12
+2,685.873
+3,998.169
+4,1274.09
+5,1590.5
+6,1841.48
+7,2036.83
+8,2152.76
diff --git a/benchmark/amd7950x_float/large_pyramid.csv b/benchmark/amd7950x_float/large_pyramid.csv
@@ -0,0 +1,9 @@
+threads,fps
+1,148.238
+2,279.403
+3,407.797
+4,524.174
+5,635.423
+6,716.434
+7,799.394
+8,880.242
diff --git a/benchmark/amd7950x_float/many_pyramids.csv b/benchmark/amd7950x_float/many_pyramids.csv
@@ -0,0 +1,9 @@
+threads,fps
+1,38.1845
+2,73.9263
+3,108.337
+4,139.456
+5,171.725
+6,198.861
+7,229.515
+8,253.222
diff --git a/benchmark/amd7950x_float/smash.csv b/benchmark/amd7950x_float/smash.csv
@@ -0,0 +1,9 @@
+threads,fps
+1,130.637
+2,210.938
+3,275.828
+4,341.204
+5,386.281
+6,426.426
+7,452.909
+8,467.611
diff --git a/benchmark/amd7950x_float/tumbler.csv b/benchmark/amd7950x_float/tumbler.csv
@@ -0,0 +1,9 @@
+threads,fps
+1,247.067
+2,403.606
+3,523.705
+4,629.426
+5,734.511
+6,800.338
+7,857.235
+8,898.919
diff --git a/benchmark/amd7950x_sse2/joint_grid.csv b/benchmark/amd7950x_sse2/joint_grid.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,357.551
-2,691.193
-3,1010.45
-4,1317.42
-5,1590.65
-6,1858.78
-7,2074.2
-8,2261.67
+1,360.077
+2,687.48
+3,998.479
+4,1261.45
+5,1581.53
+6,1825.64
+7,2067.46
+8,2216.48
diff --git a/benchmark/amd7950x_sse2/large_pyramid.csv b/benchmark/amd7950x_sse2/large_pyramid.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,186.185
-2,351.045
-3,511.316
-4,636.035
-5,765.404
-6,875.296
-7,991.353
-8,961.402
+1,288.876
+2,527.399
+3,769.81
+4,982.428
+5,1151.91
+6,1323.49
+7,1474.09
+8,1552.6
diff --git a/benchmark/amd7950x_sse2/many_pyramids.csv b/benchmark/amd7950x_sse2/many_pyramids.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,48.5561
-2,92.6231
-3,137.175
-4,176.644
-5,214.941
-6,253.39
-7,288.631
-8,312.527
+1,75.3333
+2,141.977
+3,205.225
+4,266.523
+5,330.244
+6,380.809
+7,433.287
+8,482.241
diff --git a/benchmark/amd7950x_sse2/smash.csv b/benchmark/amd7950x_sse2/smash.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,142.532
-2,228.987
-3,299.951
-4,364.679
-5,413.564
-6,453.351
-7,489.239
-8,519.379
+1,165.538
+2,263.517
+3,338.066
+4,405.629
+5,461.45
+6,506.119
+7,540.182
+8,563.682
diff --git a/benchmark/amd7950x_sse2/tumbler.csv b/benchmark/amd7950x_sse2/tumbler.csv
@@ -1,9 +1,9 @@
 threads,fps
-1,276.905
-2,453.522
-3,592.946
-4,702.383
-5,826.52
-6,919.179
-7,1009.05
-8,1062.61
+1,326.657
+2,521.743
+3,671.396
+4,805.81
+5,928.274
+6,1019.45
+7,1082.76
+8,1109.95
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -116,6 +116,10 @@ if (BOX2D_VALIDATE)
 	target_compile_definitions(box2d PRIVATE BOX2D_VALIDATE)
 endif()
 
+if (BOX2D_ENABLE_SIMD)
+	target_compile_definitions(box2d PRIVATE BOX2D_ENABLE_SIMD)
+endif()
+
 if (MSVC)
 	message(STATUS "Box2D on MSVC")	
 	if (BUILD_SHARED_LIBS)