GDC Austin 2009 slides https://www.cmpevents.com/GDAU09/a.asp?option=C&V=11&SessID=9869 This session will describe the iPhone performance optimization lessons learned through many hours of tuning. We'll start with an overview of the performance analysis tools available on the iPhone SDK to help you narrow down your performance bottlenecks. Then we'll cover the best way to set up your render loops, rendering best practices, how to deal with the limited memory, or even how to drop down to assembly to use the forgotten vector floating point unit.

1. Squeezing Every Drop Of
Performance Out Of The
iPhone
Noel Llopis
Snappy Touch
http://twitter.com/snappytouch
noel@snappytouch.com
http://gamesfromwithin.com

8. Running at 40 FPS
on first attempt...

9. Running at 40 FPS
on first attempt...
... on the simulator!

10. Running at 40 FPS
on first attempt...
... on the simulator!
4 FPS!

12. • Performance analysis tools

13. • Performance analysis tools
• CPU

14. • Performance analysis tools
• CPU
• Game loop architecture

15. • Performance analysis tools
• CPU
• Game loop architecture
• Rendering

16. • Performance analysis tools
• CPU
• Game loop architecture
• Rendering
• Memory

17. Performance Analysis
Tools

18. Exploratory

19. Exploratory
• Turning features on and off in real time

20. Exploratory
• Turning features on and off in real time
• Works very well to get very rough
estimates

21. Exploratory
• Turning features on and off in real time
• Works very well to get very rough
estimates
• Simulation vs. geometry creation vs.
rendering

22. Instruments

23. Instruments

24. Instruments

25. Instruments

26. Instruments
• Versatile

27. Instruments
• Versatile
• Can display frame rate, object allocations,
memory usage, or even OpenGL calls.

28. Instruments
• Versatile
• Can display frame rate, object allocations,
memory usage, or even OpenGL calls.
• Rut it on the real device, not the simulator!

29. Instruments

30. Instruments

31. Instruments
• TODO: Screenshot digging deeper

32. Instruments

33. Instruments
• TODO: Screenshot with memory leaks

34. Instruments

35. Instruments
• First pass with Instruments confirmed that
Flower Garden was simulation bound.

36. Instruments
• First pass with Instruments confirmed that
Flower Garden was simulation bound.
• Especially the matrix multiplies and
geometry creation sections.

37. Shark

38. Shark
• CPU usage

39. Shark
• CPU usage
• More in depth than Instruments

40. Shark
• CPU usage
• More in depth than Instruments
• Can even report good info on cache misses

41. Shark

42. CPU

43. CPU

44. CPU
• 32-bit RISC ARM 11

45. CPU
• 32-bit RISC ARM 11
• 400-535Mhz

46. CPU
• 32-bit RISC ARM 11
• 400-535Mhz
• iPhone 2G/3G and iPod
Touch 1st and 2nd gen

47. CPU (iPhone 3GS)

48. CPU (iPhone 3GS)
• Cortex-A8 600MHz

49. CPU (iPhone 3GS)
• Cortex-A8 600MHz
• More advanced
architecture

50. Thumb Mode

51. Thumb Mode

52. Thumb Mode
• CPU has a special thumb
mode.

53. Thumb Mode
• CPU has a special thumb
mode.
• Less memory, maybe better
performance.

54. Thumb Mode
• CPU has a special thumb
mode.
• Less memory, maybe better
performance.
• No floating point support.

55. Thumb Mode
• CPU has a special thumb
mode.
• Less memory, maybe better
performance.
• No floating point support.
• It’s on by default!

56. Thumb Mode
• CPU has a special thumb
mode.
• Less memory, maybe better
performance.
• No floating point support.
• It’s on by default!
• Potentially HUGE wins turning
it off.

57. Thumb Mode
• CPU has a special thumb
mode.
• Less memory, maybe better
performance.
• No floating point support.
• It’s on by default!
• Potentially HUGE wins turning
it off.

58. Thumb Mode

59. Thumb Mode
• Turning off Thumb mode increased
performance in Flower Garden by over 2x

60. Thumb Mode
• Turning off Thumb mode increased
performance in Flower Garden by over 2x
• Heavy usage of floating point operations
though

61. Thumb Mode
• Turning off Thumb mode increased
performance in Flower Garden by over 2x
• Heavy usage of floating point operations
though
• Most games will probably benefit from
turning it off (especially 3D games)

62. Integer Divide

63. Integer Divide
There is no integer divide

64. Vector FP Unit

65. Vector FP Unit
• The main CPU has no
floating point support.

66. Vector FP Unit
• The main CPU has no
floating point support.
• Compiled C/C++/ObjC
code uses the vector
floating point unit for any
floating point operations.

67. Vector FP Unit
• The main CPU has no
floating point support.
• Compiled C/C++/ObjC
code uses the vector
floating point unit for any
floating point operations.
• Can program the VFP in
assembly for max
performance.

68. Vector FP Unit

69. Vector FP Unit
• Big win when many floating point
operations can be vectorized.

70. Vector FP Unit
• Big win when many floating point
operations can be vectorized.
• Matrix multiplies, skinning, etc

71. Vector FP Unit
• Big win when many floating point
operations can be vectorized.
• Matrix multiplies, skinning, etc
• vfpmath project in Google Code

72. Vector FP Unit
void Matrix4Mul(const float* src_mat_1, const float* src_mat_2,
float* dst_mat)
{
asm volatile (VFP_SWITCH_TO_ARM
VFP_VECTOR_LENGTH(3)
"fldmias %2, {s8-s23} nt"
"fldmias %1!, {s0-s3} nt"
"fmuls s24, s8, s0 nt"
"fmacs s24, s12, s1 nt"
//...

73. Game Loop
Architecture

74. Game loop

75. Game loop
Gather input

76. Game loop
Gather input
Update state

77. Game loop
Gather input
Update state
Render frame

78. Game loop
Gather input
Repeat 60 times per
second
Update state
Render frame

79. NSTimer

80. NSTimer
• Easiest way to drive the main loop

81. NSTimer
• Easiest way to drive the main loop
• All the samples from Apple

82. NSTimer
• Easiest way to drive the main loop
• All the samples from Apple
m_gameLoopTimer = [NSTimer
scheduledTimerWithTimeInterval:animationInterval
target:self
selector:@selector(doFrame)
userInfo:nil
repeats:YES];

83. NSTimer

84. NSTimer
• Not very accurate.
Frames can vary by as
much as 5-10 ms!

85. NSTimer
• Not very accurate.
Frames can vary by as
much as 5-10 ms!
Event

86. NSTimer
• Not very accurate.
Frames can vary by as
much as 5-10 ms!
Event
Event

87. NSTimer
• Not very accurate.
Frames can vary by as
much as 5-10 ms! NSTimer
Event
Event

88. NSTimer
• Not very accurate. Event
Frames can vary by as
much as 5-10 ms! NSTimer
Event
Event

89. NSTimer
NSTimer
• Not very accurate. Event
Frames can vary by as
much as 5-10 ms! NSTimer
Event
Event

90. NSTimer
Event
NSTimer
• Not very accurate. Event
Frames can vary by as
much as 5-10 ms! NSTimer
Event
Event

91. NSTimer

92. NSTimer
• Call it at a higher
frequency so main loop
is called right away.

93. NSTimer
• Call it at a higher
frequency so main loop
is called right away.
Event

94. NSTimer
• Call it at a higher
frequency so main loop
is called right away.
NSTimer
Event

95. NSTimer
• Call it at a higher
frequency so main loop
is called right away.
NSTimer
NSTimer
Event

96. NSTimer
• Call it at a higher
frequency so main loop
is called right away. NSTimer
NSTimer
NSTimer
Event

97. NSTimer
• Call it at a higher
frequency so main loop NSTimer
is called right away. NSTimer
NSTimer
NSTimer
Event

98. NSTimer
• Call it at a higher Event
frequency so main loop NSTimer
is called right away. NSTimer
NSTimer
NSTimer
Event

99. NSTimer
• Call it at a higher Event
frequency so main loop NSTimer
is called right away. NSTimer
• Unfortunately that NSTimer
floods the message NSTimer
queue and causes delay
Event
on touch events

100. Threads

101. Threads
• Put the game on a separate thread from
the UI one.

102. Threads
• Put the game on a separate thread from
the UI one.
• Runs as fast as possible without delays

103. Threads

104. Threads
• Careful what you do when parsing input
events from main thread

105. Threads
• Careful what you do when parsing input
events from main thread
• Can be more complex to debug

106. Threads
• Careful what you do when parsing input
events from main thread
• Can be more complex to debug
• Running as fast as possible might not always
be desirable (draining battery life)

107. Threads

108. Threads
• Can also split game update and rendering in
two different threads.

109. Threads
• Can also split game update and rendering in
two different threads.
• Syncing the two is much more difficult

110. Threads
• Can also split game update and rendering in
two different threads.
• Syncing the two is much more difficult
• Can only use one thread for each OpenGL
context

111. Threads
• Can also split game update and rendering in
two different threads.
• Syncing the two is much more difficult
• Can only use one thread for each OpenGL
context
• Theoretically best results, but maybe not
much difference

112. Thread Driving Loop

113. Thread Driving Loop
m_mainLoop = [[NSThread alloc] initWithTarget:self
selector:@selector(mainLoopTimer) object:nil];
[m_mainLoop start];

114. Thread Driving Loop
m_mainLoop = [[NSThread alloc] initWithTarget:self
selector:@selector(mainLoopTimer) object:nil];
[m_mainLoop start];
- (void)mainLoopTimer
{
while (![[NSThread currentThread] isCancelled])
{
[self performSelectorOnMainThread:@selector(doFrame)
withObject:nil waitUntilDone:YES];
usleep(2000);
}
}

115. Thread driving loop

116. Thread driving loop
• Can be best of both worlds

117. Thread driving loop
• Can be best of both worlds
• Consistent call to main loop

118. Thread driving loop
• Can be best of both worlds
• Consistent call to main loop
• No flooding of events (cuts in line)

119. CADisplayLink

120. CADisplayLink
• Triggered by display refresh

121. CADisplayLink
• Triggered by display refresh
• Can choose to get a call every X updates

122. CADisplayLink
• Triggered by display refresh
• Can choose to get a call every X updates
• Consistent update

123. CADisplayLink
• Triggered by display refresh
• Can choose to get a call every X updates
• Consistent update
• Perfect for games

124. CADisplayLink
• Triggered by display refresh
• Can choose to get a call every X updates
• Consistent update
• Perfect for games
m_displayLink = [CADisplayLink displayLinkWithTarget: selfselector:@selector(mainLoop:)];
m_displayLink.frameInterval = 2;
[m_displayLink addToRunLoop :[NSRunLoopcurrentRunLoop] forMode:NSDefaultRunLoopMode];

125. CADisplayLink

126. CADisplayLink
• Only available on SDK 3.1

127. CADisplayLink
• Only available on SDK 3.1
• Definitely the way to go for future games

128. Rendering

130. iPhone 2G
iPhone 3G
iPod Touch 1st, 2nd,
and some 3rd gen

131. iPhone 2G
iPhone 3G
iPod Touch 1st, 2nd,
and some 3rd gen
iPhone 3GS
High-end 3rd gen iPod
Touch

132. Game loop

133. Game loop

134. Game loop

135. Game loop

136. Game loop

137. Game loop

138. Game loop
• Avoid locking some resource while the
GPU is rendering.

139. Game loop
• Avoid locking some resource while the
GPU is rendering.
• Might get a small speed up by first
presenting the frame, then updating game
state, and finally rendering.

140. Game loop
• Avoid locking some resource while the
GPU is rendering.
• Might get a small speed up by first
presenting the frame, then updating game
state, and finally rendering.
• Almost no difference in Flower Garden

141. Rendering

142. Rendering
• A lot of the usual advice for GPUs:

143. Rendering
• A lot of the usual advice for GPUs:
• Minimize draw primitive calls

144. Rendering
• A lot of the usual advice for GPUs:
• Minimize draw primitive calls
• Minimize state changes

145. Rendering
• A lot of the usual advice for GPUs:
• Minimize draw primitive calls
• Minimize state changes
• ...

146. Vertices

147. Vertices
• Rendering vertices can become a
bottleneck

148. Vertices
• Rendering vertices can become a
bottleneck
• No true Vertex Buffer Objects on the
MBX, so CPU involved in every draw call

149. Vertices
• Rendering vertices can become a
bottleneck
• No true Vertex Buffer Objects on the
MBX, so CPU involved in every draw call
• Fixed on the 3GS though

150. Vertices

151. Vertices
• Make vertices as small as possible

152. Vertices
• Make vertices as small as possible
struct PetalVertex
{
#ifdef PETAL_SMALL_VERTEX
int16_t x, y, z;
int16_t u, v;
int16_t u1, v1;
#else
float x, y, z;
float u, v;
float u1, v1;
#endif
};

153. Vertices
• Make vertices as small as possible
struct PetalVertex
{
#ifdef PETAL_SMALL_VERTEX
int16_t x, y, z;
int16_t u, v;
int16_t u1, v1;
#else
float x, y, z;
float u, v;
float u1, v1;
#endif
};

154. Vertices

155. Vertices
• Align vertices on at least 4-byte boundaries

156. Vertices
• Align vertices on at least 4-byte boundaries
• Experiment with larger alignments

157. Vertices

158. Vertices
• Use indexed lists...

159. Vertices
• Use indexed lists...
• ... ordered as if they were strips.

160. Mixing OpenGL and
UIKit

161. Mixing OpenGL and
UIKit
• You can mix and match OpenGL
and UIKit

162. Mixing OpenGL and
UIKit
• You can mix and match OpenGL
and UIKit
• But you have to be very careful

163. Mixing OpenGL and
UIKit
• You can mix and match OpenGL
and UIKit
• But you have to be very careful
• Minimize UIKit elements on top of
EAGLView

164. Mixing OpenGL and
UIKit
• You can mix and match OpenGL
and UIKit
• But you have to be very careful
• Minimize UIKit elements on top of
EAGLView
• Keep them opaque

165. Mixing OpenGL and
UIKit
• You can mix and match OpenGL
and UIKit
• But you have to be very careful
• Minimize UIKit elements on top of
EAGLView
• Keep them opaque
• OK to have smaller EAGLView

166. Use iPhone Features

167. Use iPhone Features

168. Use iPhone Features

169. Memory

172. !=

173. !=

174. Memory rant

175. Memory rant
• No memory guarantees

176. Memory rant
• No memory guarantees
• No idea of how much memory will be
available

177. Memory rant
• No memory guarantees
• No idea of how much memory will be
available
• Not even how much memory at startup!

178. Memory rant
• No memory guarantees
• No idea of how much memory will be
available
• Not even how much memory at startup!
• Sometimes I’ve seen the game start with as
low as 3MB free memory!!!

179. Memory

180. Memory
• System notifies apps when memory is
running low

181. Memory
• System notifies apps when memory is
running low
• Apps are supposed to free as much
memory as they can

182. Memory
• System notifies apps when memory is
running low
• Apps are supposed to free as much
memory as they can
• Horrible situation for games!

183. Memory

184. Memory

185. Memory

186. Memory
WARNING!!

187. Memory
WARNING!!

188. Memory

189. Memory
WARNING!!

190. Memory
Wait...
WARNING!!

191. Memory
Wait...
... wait for it ...
WARNING!!

192. Memory
Wait...
... wait for it ...
WARNING!!

193. Memory
WARNING!!

194. Memory
WARNING!!

195. Memory
WARNING!!

196. Memory trick

197. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?

198. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?
• Allocate memory slowly.

199. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?
• Allocate memory slowly.
• When you get a warning, wait a little.

200. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?
• Allocate memory slowly.
• When you get a warning, wait a little.
• If free memory doesn’t increase, release
some.

201. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?
• Allocate memory slowly.
• When you get a warning, wait a little.
• If free memory doesn’t increase, release
some.
• Repeat until enough memory

202. Memory trick
• Do you prefer to allocate your memory all
at once at initialization time?
• Allocate memory slowly.
• When you get a warning, wait a little.
• If free memory doesn’t increase, release
some.
• Repeat until enough memory

203. Thank you!
Twitter: @snappytouch
Email: noel@snappytouch.com
Blog: http://gamesfromwithin.com