Sony Playstation-2 VPU: A Study on the Feasibility of Utilizing Gaming Vector Hardware for Scientific Computing
By Pavan Tumati Advisors: Professor Sanjay Patel and Professor Todd Martinez ECE298/299 5/16/03
2 Abstract One of the driving forces pushing the demand for increased computational power in consumeroriented devices is the video game industry. Video games, in the past decade, have demonstrated an insatiable desire for computing hardware that accelerates a certain subset of mathematical operations related to efficient rendering of graphical objects on the computer screen. Quite conveniently, these same mathematical operations used to accelerate computer graphics also are capable of accelerating common operations in the realm of scientific computing. Sony Corporation has invested quite heavily in producing complete computing platforms to accelerate computer graphics. Its offering, the Sony Playstation-2, exposes to software developers a wide array of computational devices that specifically assist in accelerating common graphics related operations, such as vector mathematics. In addition, Sony has offered to the developer community the open-source Linux operating system. The combination of Linux, open source development tools, and detailed developer information allows for the investigation of the feasibility of future use of commodity video gaming hardware for the acceleration of scientific calculations. This document details our investigation and assessment of the potential of the Playstation-2 for scientific computing.

Table of Contents:

Chapter 1 Chapter 2 Chapter 3 Chapter 4 Chapter 5 Chapter 6 Chapter 7 Chapter 8 Bibliography/References System Overview, Resources, and Expectations Establishing A Performance Metric and Basic Tests Macromode Performance Analysis and Experiment Construction Micromode Analysis and Test Construction Micromode VPU + FPU Usage Analysis Alleviating Bottlenecks in Performance New Directions in Performance Analysis and Test Development Results and Feasibility of Usage Assessment
Chapter 1 - System Overview, Resources, Expectations and Results
The core of the Playstation-2 is popularly known as the Emotion Engine. The Emotion Engine is a collection of various subsystems which are, as described in Sonys manual, an effort to have the highest performance by adopting the latest technology and the most advanced manufacturing technology from the early stages, in order to secure a long product life with performance at the point of sale kept unchanged. Some of the advertised features are fast rendering, multi-path geometry, on-demand data decompression, application specific processors, and data path buffering in whats labeled a unified memory architecture, or simply UMA. A block diagram of the overall system substantiates some of Sonys claims:
Figure 1: Block Diagram of Emotion Engine Core (Adapted from EE Overview Manual, Sony Corporation)
The CPU, Vector Processing Unit (VPU), FPU, scratchpad ram, the instruction and data caches are all collectively labeled the EE Core. The CPU implements the superscalar 64-bit MIPS IV instruction set architecture. The instruction cache is 16Kbytes in size, and is two-way set associative. The data cache is 8Kbytes in size, two-way set associative, and supports a write-back protocol. The EE core features two vector operation processors contained within the VPU Unit for floating point vector operations. These vector units are intended to accelerate geometric calculations. The two vector operation processors within the VPU are known as VU0 and VU1. VU0 is connected to the CPU via a 128-bit wide coprocessor bus. Because of this connection, it is possible to issue to the VPU, from the CPU, what are known as macro instructions to do mathematical operations in VU0 hardware. (It is not possible, however, to issue macro instructions to VU1, because of the lack of a connecting bus between the CPU and VU1 registers.) Also, the VU0 and VU1 are connected to separate units, known as vector interface units (VIFs), VIF0 and VIF1. The VIF units can decompress packets of information transferred via DMA into the local memories of VU0 and VU1. It is important to mention that the registers of the vector units are 128-bits wide, and conveniently accommodate vectors containing four 32-

Chapter 3 Macromode Performance Analysis and Test Construction
In macromode, the principal components exercised within the Emotion Engine are the CPU, the COP2 bus, and VU0. In the diagram below, the exercised units are highlighted:
Figure 4: EE Core COP2 and VU0 Location (Adapted from EE Overview Manual, Sony Corporation)
In order to understand and exercise the VU0 unit, a small, crafty piece of code needed to be constructed or found. In this chapter, the objective of the code below, obtained from http://www.bd.wakwak.com/~yumimint/ps2linux/vu-tips.html , is to come up with a scalar result of the dot-product of two four-element vectors. For the purposes of demonstrating an example, supposing there were two vectors v1 and v2: v1 = < x1, y1, z1, w1 > v2 = < x2, y2, z2, w2 > To compute the dot product, code would be written that is functionally similar to what is listed below. (Note that the registers prefixed with VF to their name are macromode references to the floating point registers in VU0, the unit that gets exercised when requests are made over coprocessor bus 2.)
float COP2_DotProduct( void *src1, void *src2 ) { float result; __asm__ __volatile__ (" lqc2 vf16, 0x0( %1 ) lqc2 vf17, 0x0( %2 ) vaddw.x vf18, vf00, vf00 vmul.xyzw vf16, vf16, vf17 vmulax.x ACC, vf18, vf16x vmadday.x ACC, vf18, vf16y
vmaddaw.x vmaddz.x.set noat qmfc2 mtc1.set at ACC, vf18, vf16w vf16, vf18, vf16z $1, vf16 $1, %0
" : "=f" (result) : "r"(src1), "r"(src2) : "$1" ); return result; }
The code above doesnt make much sense until the notation for the various macro instructions is understood. Before reading the analysis below, it is important to note a few things. In the VPU of the PS2, the floating point registers have 4 32-bit parts. They are termed the w, z, y, and x components. The w component refers to bits 127-96. The z component refers to bits 95-64 The y component refers to bits 63-32 The x component refers to bits 31-0
Table 1: Vector Component Format

W (127-96)

z (95-64)

y (63-32)

x (31-0)
The VF00 register, however, has its components set to fixed values. The w value is set to 1.0, and all of the other parts of the register are set to 0.0. For the purpose of the understanding the detailed analysis below, assume that the contents of VF16 correspond to the parts of a vector v1 <x1, y1, z1, w1> and the contents of VF17 correspond to the parts of a vector v2 <x2, y2, z1, w2>. The dot after a register name in conjunction with a component letter is a reference to an element of a vector. For example, VF16.x refers to the x component of the 4 floating-point element vector contained within register VF16. The analysis of the code above is as follows:

Table 2: VU Macromode Dot Product Computation Code Breakdown
Instruction lqc2 vf16, 0x0( %1 ) lqc2 vf17, 0x0( %2 ) Vaddw.x vf18, vf00, vf00 Algebraic Expression VF16 = values VF17 = values VF18.x 4 floating point from parameter srcfloating point from src2 = VF00.x + VF00.w Important Things to Note
Vmul.xyzw vf16, vf16, vf17
VF16.x = VF16.x * VF17.x VF16.y = VF16.y * VF17.y VF16.z = VF16.z * VF17.z
VF18.x = 1 (Note VF00.x=0, VF00.w = 1, always ) VF16.x = x1*x2 VF16.y = y1*y2 VF16.z = z1*z2
Vmulax.x ACC, vf18, vf16x Vmadday.x ACC, vf18, vf16y Vmaddaw.x ACC, vf18, vf16w Vmaddz.x vf16, vf18, vf16z VF16.w = VF16.w * VF17.w ACC.x = VF18.x * VF16.x ACC.x = ACC.x + VF18.x * VF16.y ACC.x = ACC.x + VF18.x * VF16.w VF16.x = ACC.x + VF18.x * VF16.z VF16.w = w1*w2 ACC.x = x1 * x2 ACC.x = x1 * x2 + y1 * y2 ACC.x = x1 * x2 + y1 * y2 + w1 * w2
VF16.x = x1 * x2 + y1 *y2 + w1 * w2 + z1 * z2
$1 = VF16 Contents Transfer the floating point data to the value we want to return.

Qmfc2 Mtc1

$1, vf16 $1, %0
$1 = VF16 (VU to EE data transfer) Store lower 32 bits of GPR into floating point register. (We constrained %0 to a floating point register via =f specification in gccs extended asm syntax.)
Its quite clear from the column showing the algebraic expression that the dot-product operation benefits immensely from the SIMD characteristic that allows for 4 single-precision floating point multiplications in 1 instruction. The gcc-inlined code above featuring accumulation operations, result retrieval (qmfc) code, and the actual dot-product computation code allow for us to then construct a test to measure the actual performance of the code in execution. In order to compute the effectiveness of the above macromode code and to compare overall throughput with that of other commodity systems, the function above was placed in competition with similarly functioning C-code compiled for other platforms. Notice that the assembly code above is encapsulated in a C-function, COP2_DotProduct(). In non-PS2 systems, this function was replaced with a more generic C function used to compute the dotproduct. The competition code then was run repeatedly to figure out how fast a given system could compute dot-products. All test programs were compiled using gcc for the respective architecture. It is important to note that in the utilized performance program, on non-PS2 test systems, the equivalent dot product calculation code did not utilize specialized SIMD instructions, such as IA32 MMX/SSE/SSE2 or G4 AltiVec instructions. It was decided that the performance tests would pit the PS2 facilities against average or generically compiled mathematical C programs. The general idea was to compare the PS2 code with code that was generically compiled by a C program on the IA32 and G4 architectures; the rationale behind this decision was that there would be specific advances in compilers and toolkits for providing a better, overall generic performance on the Playstation-2 for average mathematical programs relying on single-precision floating point arithmetic. The performance graph below (obtained through the use of gprof) was obtained. (Note that the x-axis represents the number of dot products performed, and the y-axis represents the time required to complete all the calculations.)

Figure 5: Playstation-2 Dot-Product Calculation Performace versus Macs and PCs (X axis = # of Dot Products, Y axis = time in seconds.)
It is noticeable that a generally linear trend in performance with the increase in the number of dot products being calculated. Also quite evident is the sharp performance drop off of the PlayStation-2 after a certain data-size threshold is reached. This performance drop off is considerable, and demonstrates that the 32MB integrated RAM will prove to be a significant bottleneck because of paging. Consequently, it can be stated: any utilization of the PlayStation2 for efficient mathematics must avoid paging penalties and disk I/O. Notice, however, that in the region where the PS2 was not paging, it substantially outperformed other processors in terms of OPs/cycle. Therefore, tests designed to measure the raw computational throughput of the PS2 hardware were refactored to avoid exercising the scenario in which the virtual memory and disk subsystem were exercised.
Chapter 4 Micromode Performance Analysis
Looking at the performance graph for macromode, it is evident that macromode performance is fairly decent in comparison to commodity hardware released at around the same time the PS2 was released (given the test conditions stated above with regard to optimizations), but does not allow us to take full advantage of the advertised hardware offerings of the PlayStation-2. It was speculated during experimentation that utilization of micromode would allow for a significant performance boost that would allow our performance graphs to reach the performance levels of more common (not necessarily top of the line) PC hardware. The primary difference in micromode code utilization is that the CPU no longer specifically issues instructions over the coprocessor bus; instead, the VUs operate completely independently and in parallel. In addition, in macromode, only 1 VUs FMACs are used. By switching to micromode, we gain the ability to exercise the FMACs of both VUs. Because of the utilization of 2 VUs, it was expected that the performance throughput would roughly double. This idealized expectation neglects any contributions from the CPU or the FPU. The diagram below shows the portions of the PS2 that were used in the VU performance tests:
Figure 6: EECore Components Used in Micromode Tests (Adapted from EE Overview Manual, Sony Corporation)
The primary downside, initially, of micromode performance analysis was that entirely new program code needed to be written and assembled so that data could be written into the VUs. This meant that control instructions would have to be separated from mathematical instructions, and that completely new code would have to be written and assembled using a separate assembler. No longer could instructions be issued to the VUs directly; instead, the CPU needed to take pre-assembled VU instructions and write those instructions into the micromem of each individual VU unit.

13 Because the operating system utilized was Linux, software executing in unprivileged usermode needed to execute a mmap() system call on Sony-issued device drivers that provided a convenient interface to user-mode applications. The mmap() function invoked on the /dev/ps2vpu0 and /dev/ps2vpu1 devices effectively adjusted the page tables associated with a process virtual address space to enable non-cached page-mappings corresponding to the physical address space memory corresponding to the VUmem and micromem of the individual VUs. Our test process had an address space, described by the maps entry in the /proc filesystem, somewhat similar to the following:
00300000-00301000 00400000-00409000 0fb60000-0fb79000 0fb88000-0fb89000 0fb89000-0fb8a000 10000000-10001000 10001000-1000b000 2aaab000-2aaac000 2aaac000-2aaae000 2aaaf000-2aaf3000 2aaf3000-2ab32000 2ab32000-2ab34000 2ab34000-2ab46000 2ab46000-2ab85000 2ab85000-2ab8c000 2ab8c000-2ab8d000 2ab8d000-2acbf000 2acbf000-2acfe000 2acfe000-2ad07000 2ad07000-2ad0c000 2ad0c000-2ad14000 2ad14000-2b763000 2b763000-2b764000 2b764000-2b765000 2b765000-2b766000 7ffee000-80000000 r-xs r-xp r-xp rw-p rwxp rw-p rwxp rw-p rw-s r-xp ---p rw-p r-xp ---p rw-p rw-p r-xp ---p rw-p rw-p rw-s rw-p rw-p rw-p rw-p rwxp 00002000 fffef000 03:02 03:02 03:02 03:02 00:00 03:02 00:00 00:00 03:02 03:02 03:02 03:02 03:02 03:02 03:02 00:00 03:02 03:02 03:02 00:00 03:02 00:00 03:02 03:02 03:02 00:0 /dev/tst ~/code/performance/performance.micromode.comb /lib/ld-2.2.2.so /lib/ld-2.2.2.so ~/code/performance/performance.micromode.comb
/dev/ps2vpu0 /lib/libm-2.2.2.so /lib/libm-2.2.2.so /lib/libm-2.2.2.so /lib/libpthread-0.9.so /lib/libpthread-0.9.so /lib/libpthread-0.9.so /lib/libc-2.2.2.so /lib/libc-2.2.2.so /lib/libc-2.2.2.so /dev/ps2vpu1 ~/code/performance/dp.elf /dev/ps2mem ~/code/performance/dp.elf
As mentioned in chapter 1, the VUs expect instructions 64-bits at a time. Each instruction is actually a pair of 32-bit instructions and these are fed individually to the floating point unit and the integer unit. Thus, the macromode code described above needed to be rewritten appropriately. The following code, logically similar to the macromode code, was constructed. In an attempt to optimize the code, the loop was unrolled manually such that 4 dot products were computed in one loop iteration. Notice also that branch-delay slots are supported in the VUs and are utilized in the following code:
.vu.data # # # # argument argument argument argument 1: 2: 3: 4: <size of dot product chunks in terms of vector count > <location of data chunk 1> <location of data chunk 2> 0 -> Status Location

.text.globl uModeDotProduct uModeDotProduct: # First, the parameters need to be extracted: # # # # vi1 vi2 vi3 vi4 contains contains contains contains the the the the number of dot products we are calculating location of dot product array 1 location of dot product array 2 destination address of the results
# Low order address -> w, high is x sub.xyzw vf1, vf00, vf00 nop nop nop ilwr.x ilwr.y ilwr.z ilwr.w vi1, vi2, vi3, vi4, (vi0)x (vi0)y (vi0)z (vi0)w
##### Stage 1, Part of Stage 2 P1[16, 17, 18] -> P2[ 19, 20 ] # First attempt/pass adaptation of macromode computation code # for micromode code. nop lqi.xyzw vf16, (vi2++) nop lqi.xyzw vf17, (vi3++) OuterLoop1: addw.x vf18, vf00, vf00 iadd vi5, vi4, vi0 mul.xyzw vf16, vf16, vf17 sqi.xyzw vf1, (vi5++) mulax.x ACC, vf18, vf16x lqi.xyzw vf19, (vi2++) madday.x ACC, vf18, vf16y lqi.xyzw vf20, (vi3++) maddaw.x ACC, vf18, vf16w nop maddz.x vf16, vf18, vf16z nop addw.x vf21, vf00, vf00 mr32.w vf15, vf16 mul.xyzw vf19, vf19, vf20 mr32.z vf15, vf15 mulax.x ACC, vf21, vf19x lqi.xyzw vf22, (vi2++) madday.x ACC, vf21, vf19y lqi.xyzw vf23, (vi3++) maddaw.x ACC, vf21, vf19w lqi.xyzw vf25, (vi2++) maddz.x vf19, vf21, vf19z lqi.xyzw vf26, (vi3++) addw.x vf24, vf00, vf00 mr32.w vf15, vf19 mul.xyzw vf22, vf22, vf23 lqi.xyzw vf16, (vi2++) mulax.x ACC, vf24, vf22x lqi.xyzw vf17, (vi3++) madday.x ACC, vf24, vf22y mr32.yz vf15, vf15 maddaw.x ACC, vf24, vf22w nop maddz.x vf22, vf24, vf22z nop addw.x vf27, vf00, vf00 mr32.w vf15, vf22 mul.xyzw vf25, vf25, vf26 mr32.xyz vf15, vf15 mulax.x ACC, vf27, vf25x nop madday.x ACC, vf27, vf25y nop maddaw.x ACC, vf27, vf25w nop maddz.x vf25, vf27, vf25z nop nop mr32.w vf15, vf25 # Terminate execution and transfer control back to VU nop isubiu vi1, vi1, 4 nop ibgtz vi1, OuterLoop1 nop sqi.xyzw vf15, (vi4++) nop[t] nop[e] nop nop nop nop nop nop
For the sake of brevity, a more detailed instruction by instruction analysis is not provided here, but is provided in Appendix A. However, it can be witnessed that the same general computation code structure from the macromode example is presented in the left-hand column. The left-most instruction stream is, in fact, the stream of instructions that are issued to the floating point hardware. The right-most stream of instructions is issued to the integer/control unit. Notice that at the end of the instruction stream, there are No-Operation instructions (nops) with t and e flags specified. These instructions instruct the VU to raise an interrupt to signal to the CPU that computations have been completed. (The software used to conduct performance tests needed to wait on a queue until a secondary interrupt service routine awakened them.) In the initial test of the above code, only VU0 was used to measure performance. Eventually tests were performed utilizing VU1 alone, and also both VU0 and VU1. The following table relates performance of macromode with micromode in terms of time required to perform 200,000 dot products:

Table 3: VU Performance In Various Modes In Terms of Time for Calculating 200,000 dot products
Mode Macro Micro Micro Micro
Units used VU0 (via COP2) VU0 VU1 VU0+VU1*
Exec Time.144 seconds 14.8 seconds 3.1 seconds 3.5 seconds
*Data was transferred to the VUs only after both VUs completed processing their data. Macromode outperforms micromode! The significant performance difference is attributable to the amount of time spent transferring data across the main memory bus from RAM to the data memories of the VU units. (Note: It is important to note that the tests within this chapter relied on the CPU to perform data-transfers via the LQ and SQ opcodes, and did not rely on DMA and VIF hardware.) In addition, the number of system calls required for the user-mode code to control VU units contributed to an increased overhead. The amount of time required to transfer from user-mode to kernel-mode for system calls to VPU device drivers is significant. Latencies in the kernel control paths also contributed to performance degradation. Clearly, in order to avoid the performance penalties of data transfer and system-call overhead, another solution was required. Note, also, that the dot-product performance test lends itself to this behavior because the VU code does not systematically-reuse the data passed to it. It is speculated that if the ratio of transfer time to computation time were adjusted accordingly, significant performance increases would be realized. To get a feel for what sort of performance would be obtained without the transfer penalty, but with control-overhead a test was devised in which the VUs individually computed dot products over and over, without transferring data back and forth to RAM. The resulting time for 200000 dot products is:
Table 4: VU Micromode Performance Without Data Transfer Overhead

Mode Micro

Units used VU0+VU1*

Exec Time 3.0 seconds

For increased performance, significant modifications would be required to avoid kernel-control path and system call latencies. The assisting in proving the hypothesis that system-call overhead was the primary bottleneck, the code was restructured in a manner that allowed for calculation of 200,000 dot-products without any data transfers or raised-interrupts. In other words, to create the ideal situation without control-overhead and data-transfer overhead, the VU code was rewritten so the VU would repeat over 1 buffer and compute the dot-product repeatedly 200,000 times. This resulted in 200% improvement over macromode code, using only VU1! The results are shown in Table 5.

Table 5: VU Performance for 200,000 Dot Products without Control and Data Transfer Overhead

Units used VU1

Exec Time.07 seconds.
Chapter 5 Micromode VPU + FPU Usage Analysis
When the performance test software for determining the values in the previous section was designed, it was speculated that perhaps some benefit would arise from exploiting multiple threads of execution. The idea, simply, was to attach a control-thread, instantiated via the Linux p-threads interface, to each computational device. For example, in the previous section, if both VUs were utilized at the same time, VU0 always finished before VU1, and would sit idle until the software was finished. One thread would monitor and feed the information to VU0, one would monitor VU1, and yet another would be used to feed the FPU connected to the main processor. In the last chapter it was quite evident that the control-overhead was a primary bottleneck in achieving high performance out of the VPU. Therefore, it suffices to say that the added overhead of multiple threads did not alleviate this bottleneck, but instead exacerbated the problem at least, with regards to the dot-product test. However, it is suspected that, for software that requires a long compute time with respect to control-overhead time, a threaded software model would be beneficial in many cases, especially when dividing-and-conquering mathematical problems. The MIPSLinux kernel provided with the Sony Playstation-2 kit does not support utilizing the VPU with multiple-threads. That is, during user-mode process context switches, the registers of the VPU are not saved. Furthermore, only one user-mode thread of execution may have ownership of the VPU at any given time. To alleviate this problem and to perform tests on multi-threaded applications, an investigation was performed on the management of the FPU inside the kernel and a solution for the VPU was determined. In most user-mode processes in Linux, the kernel does not bother to save the floating point registers on a context switch. This is a performance enhancement in that the amount of time required for a context switch in user-mode decreases if the floating point registers do not have to be saved. For processes that require floating point instructions, what happens is that the usermode process' first FPU instruction generates an exception. The exception handler then appropriately sets the bits required in process-specific data structures in the kernel and enables access to the FPU. After the process-specific data structures are modified, the kernel then saves the floating point registers on a context switch. For adding multi-threaded support for the performance test software and other applications, two approaches could have been utilized: 1) Manage the VPU like the FPU and save the VPU registers visible over the COP2 bus on a context switch, or 2) Designate VPU-control code as critical sections, and guard them with mutual exclusion constructs. Option 1 was the more complex solution, and was avoided in the short term. Option 2 was studied in this document. In order to allow for multiple-threads to utilize the vector units at a time, a device driver (vpuperm) was written to facilitate access to the VUs. In Linux, whenever a user-mode to kernel-mode transition happens via the system call mechanism, the kernel saves all of the registers of the process to the stack. The vpuperm device driver simply modifies the permissions of threads which possess active vpuperm-file-descriptors by adjusting values on the stack. The system-call utilized was ioctl(), and the basic code is presented here:

19 microprogram transfer to code via the MPG command, or even a function call of sorts to force the VU to start executing code from a specific address in the code memories. Within the dot-product performance test, the dominant VIF operation was the UNPACK operation, which places information into the data memories of the VU. The dot-product performance test placed the microprogram into the code-memories in advanced, so no MPG operation ever needed to be issued. The VIFCode UNPACK packet is structured as follows (in C struct description):
typedef struct { unsigned int addr unsigned int reserved unsigned int usn unsigned int flg unsigned int numcount unsigned int cmdcode }VIFCodeStruct_unpack;
: 10; : 3; : 1; : 1; : 8; : 8;

/* /* /* /* /* /*

Bits 0-9, Start Address Bits 11-13, Reserved Bits 14, Signed Unsigned Decompression Bits 15, Address Mode - Add VIF1_Tops? Bits 16-23, Number of 128 bit chunks Bits 24-21, Command Code

*/ */ */ */ */ */

The VIFCodeStruct_unpack information needs to be placed ahead of the actual data in the buffer being transferred. The number of 128-bit chunks being transferred is specified within the packet information. By toggling bits in the actual command code field, the VIF unit can be instructed to raise an interrupt to signal the end of a data transfer. In the dot-product test performance application, this interrupt was used to signal that more information needed to be transmitted. Unfortunately, in current versions of the performance test, the utilization of the interrupt caused increased performance overhead, so no real performance benefit was gained in existing performance tests. The generalized code for performing a data transfer using Sony provided VPU device drivers and the UNPACK command is presented:
unsigned int vifstreamheader[] = { 0x00000000, 0x00000000, 0x00000000, 0x6C000000 }; [] dmapacket.ptr = (void *) fpacket; dmapacket.len = sizeof( VIF0FullTransferPacket ); vifevent = VIF0EventOpen(); if( vifevent < 0 ) { printf( "Unable to open VIF event object.\n" ); __RETURN( vifevent, "main()" ); } memset( (void *)( vu[0].datamem ), 0, PAGESIZE ); ioctl( vu[0].fildes, PS2IOC_SENDA, &dmapacket ); printf( "Transfer complete.\n" ); VIF0EventWait( vifevent ); VIF0EventClose( vifevent );

Chapter 7 Possible New Directions in Performance Analysis and Test Development
In the previous chapter, direct transfer of data to the VU data memories via the VIF units was discussed. Its evident that, even after data transfer and execution of dot-product code, information must still be reclaimed via memory copy cycles, or another DMA transfer. The user-mode code still has quite a bit of system-call overhead to initiate the DMA transfers. Clearly, to remove system-call overhead, more functionality needs to be integrated into the device drivers for the vector units, or computational code needs to be placed within kernel-space. With regards to the dot-product performance test, the ratio of control-overhead to computation time is still too high. Another alternative to the techniques previously used is to adjust the flow of data such that, instead of reclaiming the results from the vector units after computation, the data be transferred as texture data to the graphics synthesizer and its local memory via the GIF unit. The Playstation-2 graphics unit has 4MB of local memory in which textures can be stored, and through which data can be transferred back and forth via DMA and through KICK instructions through the vector units. The new region of the EE core that would be utilized is highlighted, in addition to other units used in micromode, in the block diagram below:
Figure 7: GIF and GS Unit location (Adapted from EE Overview Manual, Sony Corporation)
The graphics synthesizer introduces a great deal of complexity: textures must be managed, the local memory must be partitioned, and the hardware must be instructed not to perform any transformations on the incoming data. Future tests are expected to utilize the graphics synthesizer in the overall performance test. The development time for such a test is anticipated to be very high, due to the high configurability of the graphics synthesizer.
Chapter 8 Results and Feasibility of Usage Asssessment
While no dramatic performance improvement was shown in this particular study, the framework for performance improving tests has been created. Enough evidence has been presented that significant throughput can be expected if device drivers or kernel code is written to minimize VU control overhead. The dot-product performance test effectively allowed us to set expectations for the sort of throughput to be expected from the FMAC units in the VPU. This chapter summarizes the basic steps believed to be required to produce satisfactory throughput utilizing the VUs. If a device driver or more optimized kernel were introduced, then customized code tailored for scientific computing could take full advantage of the vector hardware in the Playstation-2. In short, the primary boundaries preventing exceptionally efficient code are: The overhead making a user-mode to kernel-mode transition Managing interrupts and threads within the kernel Data transfers to and from the VU units

The overhead involved with user-mode to kernel-mode transitions can be eliminated if a software architecture is introduced that allows for coordination of VU utilization entirely within kernel space. Currently, if a process wants to write information to the VUs using existing Sony Playstation-2 drivers, it needs to memory map the VUs code and data memories and perform a sequence of reads and writes to initialize the VUs, and then issue system calls to instruct the VU to begin operation and signal completion through an interrupt. The overhead of coordinating events is exceptionally high with regards to the amount of work the actual VU is performing. If the software is re-architected in such a manner that event coordination and management is performed in a very streamlined and quick manner entirely within the kernel, then a significant performance improvement could be achieved. When the VU or VIF signals an interrupt, the Sony device driver searches through a chain of registered event handlers to determine which threads need to be awoken. This is a very general purpose approach, but is entirely inappropriate for extracting optimal performance from the VUs. While the current approach may be appropriate if the VU spends quite a bit of time doing calculations without requiring more control operations, it is not appropriate for tests such as the dot-product performance test. Finally, the third improvement that would immensely increase overall throughput is if the DMAto-VIF transfer mechanism were used more effectively. If the information used in scientific calculations were packaged in such a manner that function calls were specified, and blocks of data were inlined with VIF codes, the overall CPU intervention required to manage the VU operation would decrease. This would enable for efficient utilization of the CPU and FPU while simultaneously allowing for intelligent use of the VUs. If this method were combined successfully with the utilization of the graphics synthesizer, it is believed a tremendous throughput would be achieved that would provide scientific applications with a significant, costeffective overall performance increase.

Bibliography

Sony Computer Entertainment, Inc. EE Core Instruction Set Manual, Sony Computer Entertainment, 2001. Sony Computer Entertainment, Inc. EE Users Manual, Sony Computer Entertainment, 2001. Sony Computer Entertainment, Inc. GS Users Manual, Sony Computer Entertainment, 2001. Sony Computer Entertainment, Inc. VU Users Manual, Sony Computer Entertainment, 2001. Daniel P. Bovet and Marco Cesati, Understanding the Linux Kernel. Sebastopol, CA: OReilly and Associates, 2001. Alessandro Rubini and Jonathan Corbet, Linux Device Drivers, Sebastopol, CA: OReilly and Associates, 2001.

CASE REPORT

Feasibility of Using the Sony PlayStation 2 Gaming Platform for an Individual Poststroke: A Case Report
Sheryl Flynn, PT, PhD, Phyllis Palma, PT, and Anneke Bender, PT
Abstract: Rationale: Many Americans live with physical functional limitations stemming from stroke. These functional limitations can be reduced by task-specic training that is repetitive, motivating, and augmented with feedback. Virtual reality (VR) is reported to offer an engaging environment that is repetitive, safe, motivating, and gives task-specic feedback. The purpose of this case report was to explore the use of a low-cost VR device [Sony PlayStation 2 (PS2) EyeToy] for an individual in the chronic phase of stroke recovery. Case: An individual two years poststroke with residual sensorimotor decits completed 20 one-hour sessions using the PS2 EyeToy. The games task requirements included target-based motion, dynamic balance, and motor planning. The feasibility of using the gaming platform was explored and a broad selection of outcomes was used to assess change in performance. Outcomes: Device use was feasible. Clinically relevant improvements were found on the Dynamic Gait Index and trends toward improvement on the Fugl-Meyer Assesment, Berg Balance Scale, UE Functional Index, Motor Activity Log, and Beck Depression Inventory. Conclusion: A low-cost VR system was easily used in the home. In the future it may be used to improve sensory/motor recovery following stroke as an adjunct to standard care physical therapy. Key words: virtual reality, rehabilitation, stroke (JNPT 2007;31: 180189)

INTRODUCTION

ermanent neurological impairments and concomitant physical functional limitations occur in approximately 60%1 of the 700,000 people in the United States experiencing a new or recurrent stroke each year.2 Decits for this population have been well documented in the areas of balance and proprioception,3 strength,4 motor control,5 endurance, and aerobic capacity.6 Strong evidence points to the benets of rehabilitation and therapeutic exercise in all stages of stroke recovery,yet rehabilitation hospital stays are of an increasingly shorter duration. This presents the need for alternative, long-term, and economically feasible treatment options.
Division of Physical Therapy, College of Health and Human Sciences, Georgia State University, Atlanta, Georgia Address correspondence to: Sheryl Flynn, E-mail: sheryl@usc.edu Copyright 2007 Neurology Section, APTA ISSN: 1557-0576/07/3104-0180 DOI: 10.1097/NPT.0b013e31815d00d5
Virtual reality (VR) is a developing technology that has historically been used for the training of motor tasks involving highly complex activities such as surgical techniques,8 ight simulation,9 and military exercises.10 More recently, VR has been explored as a therapeutic tool to retrain faulty movement patterns and rehabilitation of function resulting from neural insult. While most current studies have used small sample sizes, promising trends have been established for improving hand function1113 and locomotor activityin individuals with chronic stroke, controlling and coordinating volitional movement for children with cerebral palsy,19 and decreasing akinesia for individuals with Parkinsons disease.20 Furthermore, gains achieved through VR practice have been shown to carry over in real-world activity, sometimes resulting in spontaneous functional improvement in activities of daily living.1113,21 Recent advances in the understanding of neural plasticity and motor function retraining suggest that greater quantity, duration, and intensity of practice are most effective in modifying neural (re)organization22 and inuencing motor learning.23 Also, increasing rehabilitative treatment duration is associated with a reduction in death/deterioration following stroke.24 Additionally, motor learning principles point to the value of augmented feedback regarding performance during the acquisition of a new motor skill. Research has further established the specic benet of visual and auditory feedback in the retraining of motor function, including such tasks as increasing stance symmetry, recovering activities of daily living (ADL) abilities, and improving balance in chronic stroke patients.2528 In fact, visual cues have been shown to provide a more potent motivational context for improving performance than verbal cues.26 As traditional physical rehabilitation is costly and not reimbursable long term, the development of novel interventions in which motor retraining can be performed without the constant guidance of a rehabilitative specialist is imperative. VR gaming using commonly available PC-based equipment is a self-directed activity not limited by the constraint of having a therapist present.29 As such, it is capable of providing unrestricted frequency of treatment for an indenite period of time in the individuals own home. Furthermore, in a VR system, augmented feedback occurs in the game environment, with visual and auditory cues informing the participant about their position in space and the success of their movement attempts. Moreover, the virtual environment (VE) provides an engaging and motivating framework for feedback,allowing the participant to become immersed in the virtual world and to experiJNPT Volume 31, December 2007

JNPT Volume 31, December 2007
Sony PlayStation Poststroke
ence the emotional sense of winning in a particular game.33 These attributes of VR training are important, as lack of motivation and adherence to exercise have been shown to be problematic in traditional treatment regimens, thus impacting therapeutic outcomes.19,31,Preliminary research in the area of VR has focused primarily on complex VR systems designed for specic therapeutic purposes, providing the participant with visual, sensory, and haptic feedback tailored to retrain a precise motor skill. What had not yet been explored is the therapeutic value of commonly available VR games not explicitly created for rehabilitation. Although less expensive and lower tech, these VR games do in fact share many tasks and behaviors similar to its higher cost predecessor, the IREX system.15 While off-the-shelf gaming technology lacks specicity, it has the advantage of mass accessibility, broad affordability, and the potential for home use. The use of VR games played with Sony Eyetoy: Play 2 gaming software may provide an engaging environment to challenge and retrain motor activity. As the nature of the technology allows for unlimited repetition with ongoing feedback, training in the VE may lead to neural reorganization and a decrease in the decits commonly found following stroke. The primary purpose of this case report was to explore the feasibility of using a VR gaming device (Sony Eyetoy: Play 2 on a PlayStation 2) to improve function two years after stroke for an individual who has exhausted all traditional rehabilitative interventions. Given this participants unique professional background as a retired professor of physical therapy, upon completion of the training program an in-depth interview was conducted to gather qualitative data regarding the participants perceptions about this device and its application to physical therapy. Outcomes for impairment, activity, and participation, before and after using the VR gaming device at home, are included.

CASE DESCRIPTION History

The participant, MRB, was a 76-year-old woman who sustained a right hemorrhagic stroke 17 months prior to the start of the study. Her past medical history was signicant for low blood pressure, myofascial pain syndrome, bromyalgia, and a hearing impediment for which she wore a hearing aid. Prior to her stroke, she led an active retirement lifestyle with no physical function or cognitive limitations. Before retirement, MRB worked as a physical therapist and a professor of physical therapy for many years. On the day of the stroke, MRB was taken to the emergency room, where an MRI was used to diagnose her condition. She remained in the ICU for two days, entered acute care for three days, and then transferred to inpatient rehabilitation for ve weeks. While in inpatient rehabilitation, she received physical, occupational, and speech therapy daily, with various adjuncts such as recreational and animal therapy. Upon leaving the hospital, MRB was ambulating on level surfaces using a quad cane and stand-by assistance. ADL could be performed independently, but often not at a functional speed; she required assistance with eating, groom 2007 Neurology Section, APTA
ing, and bathroom activities in order to complete them within a reasonable time frame. Cognitive decits were also noted, including problems with memory and word-nding. At the conclusion of inpatient rehabilitation, MRB began day therapy three times weekly for one month. During this period, she began to ambulate independently within the community. She also began to develop greater speed with ADL, but continued to require assistance with more complex ne motor tasks, such as buttoning, cutting food, and tying her shoes. Although MRB was able to communicate effectively, previously noted cognitive decits remained (wordnding and memory difculties), leading to a minor reduction in conversational speed. Following completion of available treatment opportunities, MRB engaged in a daily self-directed program of exercise and activity aimed at maximizing both motor and cognitive potential. She made minor modications to her home environment, including the addition of handrails in her bathroom and bedroom. Follow-up CAT and MRI imaging studies ruled out any additional cerebral vascular accidents (CVAs). She occasionally had low back pain with a rating of approximately 2 out of 10. MRB indicated that her hobbies were reading, studying history, and traveling, all of which had become difcult following her stroke. Being a physical therapist, she engaged in exercises aimed at improving her gait and arm and leg strength such as cone stacking coordination exercises, playing ping-pong, and kicking a ball. She had not changed her exercise program in the recent past nor planned to change her exercise program during the course of this study. She had no experience with computer games, video games, or VR programs. She was interested in trialing the VR technology to further her rehabilitation. MRB had the following characteristics necessary for the study and volunteered to participate. These characteristics included that the individual would (1) have had a previous stroke and completed available treatment options, (2) not currently be receiving physical therapy or be undergoing new treatment, such as surgery or a medication change, (3) not change their regular exercise regimen during the course of the study, and (4) be able to play VR games safely within their own home. The participant then signed an informed consent statement, in accordance with requirements of the Georgia State University Institutional Review Board.

Evaluation and Prognosis

In general, MRB had signicant motor and sensory recovery poststroke; however, she had difculty performing high-level coordination and balance tasks. From a motivation perspective, MRBs rehabilitation potential was excellent. However, due to her already high level of recovery, we were prepared for minimal change.

Intervention

The VR system used in this case was the Eyetoy: Play 2 (Sony, Park Ridge, New Jersey), a commercially available gaming system that uses a video capture interface to allow the user to interact directly with their own television screen. Components for this system include a color digital camera device with USB interface (an OmniVision OV519 Video Device with an OmniVision OV7648 sensor, manufactured by Logitech, Fremont, California), a PlayStation 2 (PS2) (model number SCPH-75001), a DUALSHOCK 2 Analog Controller with pressure sensitivity, and the Eyetoy: Play 2 disc. The total cost of the system is less than $200. The system uses motion and color-sensitive computer vision to
process images taken by the camera. As the user moves his/her body, the camera presents a real-time likeness on the television screen, with a graphic overlay of a virtual surrounding. Objects within the game environment move and react when contacted by the users image, creating an interactive experience between the two. Sound and visual feedback indicate the success or failure of movement relative to the game task. The Eyetoy: Play 2 provides the player with 23 different game experiences, each presenting similar movement challenges: accurate, target-based upper extremity motion, motor planning, dynamic sitting and standing balance, and eye-hand coordination (Table 1). The movement tasks are multiplanar and multidirectional, with rotational and diagonal components, mimicking essential aspects of functional movement. Games can be played from a sitting or standing position. Six additional multiplayer games are available for use with up to four people. MRB was provided with the system, which interfaced with her own television set via an AV cable. After the Sony Eyetoy was easily set up in her home, assistance was provided with lighting conditions and sound settings to allow for easy interaction with the screen. Setting up the system took less than one hour, and no modications were made to the games. During the set-up in her home, a place that was safe and free from objects that may cause injury if she were to fall was chosen. On initiation of training, she was asked to play each of the Eyetoy games at least twice, after which game choices were at her discretion. As she was able to safely maintain standing balance while playing the Eyetoy, she chose to complete the activity from a standing position. The participant completed a total of 20 one-hour sessions of play over four and a half weeks.The training sessions were completed in time intervals determined by the participants fatigue level. For example, sometimes MRB played for 30 minutes before resting, while on other days she played for shorter times before requiring a break. The variability was due to training intensity, other activities performed that day, or general fatigue possibly due to myofascial pain or bromyalgia. The research team did not supervise the training sessions; however, her companion monitored and encouraged her during the training sessions by being present in the room or home. Compliance was monitored via the use of daily play logs that listed games played, time of session, and comments regarding individual games, collected at the completion of 10 and 20 gaming sessions. The games played are summarized in Table 1. The research team contacted the participant weekly to answer questions and encourage her continued participation.

OUTCOMES Feasibility of Use
After conducting an extensive interview and reviewing the daily logs, we discovered that this device was quite feasible as well as quite enjoyable to use. Since this participant uniquely understood the benets of this research project, we queried her regarding her opinion of the device (Table 2). In general, she described the games as very motivating and
2007 Neurology Section, APTA
TABLE 1. Summary of Game Descriptions and Number of Times the Participant Played Each Game
Games Played Air Guitar Description The player uses her hands to strum along to music. Various symbols indicate movements ranging form single hits/picks, slides, strums, and windmills to rock along to some classic rock and roll riffs. The player attempts to pop all the bubbles as quickly as possible, leaving only the red bubbles. Bomb bubbles will detonate everything nearby, so player uses them wisely. Player uses her upper extremities in all planes to color various templates as she chooses. The player uses her arms to select the best techniques to nish a professional job, catch a runaway drill, chop wood, and shred trees while keeping other objects out of the shredder. The player must move her hands to beat out rhythms using various timed beats in a specied pattern. The player attempts to prevent the computer from scoring any goals on her. Her body is used to block as many shots as possible. Other parts of the game include an agility drill, reaction test, and a tness test. The player uses her hands to swing at pitches, run the bases, and try to score as many runs as possible with 10 baseballs. There are tools to increase the score such as a strike zone, mini-map of the ineld, and a direction gauge. The player uses her arms to punch out the opponent in three rounds or less. A musical tune indicates the opportunity to launch a combo attack to focus on the opponents weak spot. Other parts of this game include sparring, punching a heavy bag, and a speed bag. The player uses her strength and agility to defeat all enemies in the way using various karate techniques. The player swings her way across the monkey bars, grabbing as many bananas as possible while collecting various fruits for extra points and intermittently dodging opponents. The player grabs and drags all the ingredients needed to complete the orders to assigned specications followed by a speed competition against an opponent. The player has to clear all of the balls from the pool table utilizing various angles of movement with her upper extremities under time constraints. The player is to sneak around collecting hidden items, solve puzzles, and move slowly so no one detects her movements as she breaks out of prison. The player navigates through the solar system by body movements. The player uses his arms as paddles to play against a number of incrementally more challenging opponents # Times Played 4 Dynamic Balance X UE ROM Speed X Cognition X Reaction Time X TargetBased X

Bubblepop

Colors Do It Yourself (DIY)

X X X X

Drummin Goal Attack

Homerun

Knockout

Kung 2 Monkey Bars

Mr. Chef

Secret Agent

Solar System Table Tennis
Abbreviation: UE ROM, Upper Extremity Range of Motion
enjoyable to play. She also found some of the games to be too difcult or in need of modication for individuals with sensorimotor impairments. MRB commented on a number of possible benets of using this technology. She suggested that individuals who are playing the games might exercise for longer periods of time without realizing how much time had passed. She enjoyed the competitive nature of the games. She indicated during her midtest session that she had set a goal of wanting to knock down the wall before time ran out. She
arrived for her posttest smiling and declaring, I nally knocked down the whole wall before my time ran outI nally did it yesterday! Furthermore, she expressed enthusiasm for exercising at home. Shortly after her stroke, she noticed that her grandson was loosing interest in visiting her. During one of his visits, he observed her playing the games and asked if he could join her. Thereafter, he often asked to visit grandma again so that they could play together. Lastly, she found benet in the convenience of playing the games
TABLE 2. Players Comments
Game Players Comments Air Guitar It takes a little time to gure it out, but after a couple of times, it gets easier. Bubblepop Its fast and requires nesse with arm movement when trying not to pop the red bubbles. Colors An interesting way to move without constraints on the type of movement or the speed of movement. can do trunk movements too. Do It Yourself (DIY) Its a great game for a lot of various arm movements that is satisfying when youve succeeded in knocking down the wall. Drummin Its difcult to advance, but can get further with more repetitions. Goal Attack Its good for weight shifts. Homerun Hitting the ball is fun, but difcult to advance. Knockout Challenging, but good game for stamina and accuracy. Kung 2 Good arm reactions and breaking the ice is fun. Monkey Bars Difcult, but gets better with more repetitions. Mr. Chef Good for cognitive, fast-paced decision making. Pool The game requires difcult, varied arm movement in different planes. Secret Agent Good for arm movements, but gets more difcult as the game advances. Solar System Good for arm movements without the time constraints. Table Tennis Good for two ways of hitting/arm movements, but difcult due to timing.
throughout the day, instead of having to exercise all at one time. She concluded by suggesting that while its widespread application may be limited, this device shows promise.
The following section describes the outcomes of the standardized assessments used in this study. The order of the tests were randomly selected and performed by the research team who were not blinded as to the intervention (Table 3). The following tests were chosen either because they are valid and reliable for individuals post troke and/or because they measured across the ICF continuum.

Pretest 84.96% 24 12.3.50 5.11.33
Midtest 87.00% 26 10.3.92 4.10.25
Posttest 94.70% 29 11.4.70 5.11.50
Six-Month 98.00% 26 9.5.00 5.12.00
on the accuracy of subjective recall and a consistent reading of test questions.46

Six-Minute Walk Test

The distance MRB was able to ambulate during the Six-Minute Walk Test (6MWT) varied and did not demonstrate identiable changes over time. Her pretest distance was 1282 feet, midtest distance was 1250 feet, and posttest distance was 1337 feet. Each distance was less than expected norms for women with a mean age of 62 , but this norm was established in the healthy elderly population and may not be relevant for an individual post-CVA.60,61
Beck Depression Inventory
A minor decline in self-reported depressive symptoms was evident in this participant, although no scores (pretest, midtest, or posttest) were outside the normal range. She identied six items as originally problematic, dropping to two items by the posttest. Further decline was evident by the six-month follow-up, with a score of 0 at that time. Generally, a score of nine or greater indicates the presence of depression according to Beck Depression Inventory (BDI) criteria.4750

Motor Activity Log

The Motor Activity Log (MAL) is a 26-item interview that measures upper extremity functional poststroke. MRB scored a 5/5 per self-report on the Actual Amount of Use (MAL-AAU) portion of the MAL at the pretest, indicating that her affected extremity participated in all ADL.62 The Quality of Movement (MAL-QOM), however, was rated a 3.5/5 at the pretest, a score that reveals a deciency in the quality of motion. By the posttest, the QOM was rated a 4.7/5, signifying the self-perception that quality of motion had improved following training.

Berg Balance Scale5153

The Berg Balance Scale (BBS), a 14-item multitask balance test, showed consistent improvement over the pretest to posttest period. The participant scored an initial 51/54 and a nal 54/54, a score that was maintained at the six-month follow-up. As only three points were deducted at the pretest and she eventually achieved a perfect score on this measure, a ceiling effect may have contributed to a lack of signicant change for this individual.

Modied Ashworth Scale

We used the Modied Ashworth Scale (MAS) to assess spasticity. The MAS revealed variable and inconclusive scoring: an initial score of 2, a 0 at midtest, a 1 at posttest, and a 0 at the six-month follow-up.63 This participant generally experienced only mild and sporadic spasticity that did not signicantly impact her life. This remained true in all stages prior to, during, and following training.

Dynamic Gait Index

The participant in this is case demonstrated some improvements across the levels of impairment, activity, and participation in chronic stroke following daily play with the Sony Eyetoy. Improvements at the level of impairment measured by the FMA were mirrored by improvement in functional skill, as assessed by the DGI, UEFI, and MAL. This indicates that gains attained had real world application and that training completed in a VE may create functionally relevant change, even two years poststroke. MRB also showed an improvement in tests related to cognitive and emotional function, specically the MMSE and BDI. Moreover, testing at the six-month follow-up indicates that improvements were maintained in the absence of consistent training (Table 3). Measures on which the subject did not demonstrate change were the BBS, TUG, 6MWT, MAL-AAU, MAS, and FRT. For the most part, this lack of change may have been due to a ceiling effect; the participant was able to achieve perfect or near perfect scores on the BBS and MAL-AAU. Additionally, the MAS measured very little spasticity at all testing periods and the participant conrmed this was not a signicant symptom in her life. Considering that most of the games require very little movement over ground, lack of improvement on the TUG and 6MWT may indicate that VR training may not specically impact walking speed. The FRT remained relatively unchanged throughout the training, although other balance measures did show improvement. The FRT is a measure of balance as the center of gravity is maximally extended over the base of support. It may be that VR training has a greater impact on dynamic balance, such as the ability to balance while attending to multiple stimuli or while performing various movement tasks. Perhaps the most striking change following training was evident in the sensory component of the FMA, where the participant demonstrated consistent gains in her score, from an initial 10/24 during the pretest to a nal 21/24 at the posttest. The increase in score was primarily due to an increase in proprioceptive ability, an improvement at the level of impairment that has important functional implications. As discussed earlier, VR training with the Sony Eyetoy provides concurrent visual feedback regarding body position in space, as every movement a player performs is mirrored by their virtual counterpart. This mirroring occurs as the player engages in target-specic action, providing ongoing feedback about actual vs intended movement. An example of how this occurs is ping-pong, a game in which the player faces a virtual opponent with their hand functioning as the paddle. During this game, the player adjusts movements to hit the ball while viewing these adjustments in real time on the television screen. The attributes of the training may have contributed to the participants increase in proprioception, as an awareness of body position is specically encouraged. A return to full proprioceptive sense has great functional signicance, as it plays a role in all movements; the participant in this study reported being able to walk over uneven surfaces to her mailbox for the rst time after completing the training. The outcome on the DGI is worth noting relative to injury prevention. The participant achieved a score of 21/24

If I were running a PT department, Id want to have [a Sony Eyetoy] for people who dont like exercise, [who] dont want to keep doing something that they dislike, not understanding what its doing for them. I just lose time. And I like to lose time. doing exercise. And it has such a wide variety. everybody likes playing games.
is an inherent motivation to participate in the activity and, perhaps, to creatively explore new movement choices. VR is a highly adaptable technology that, in contrast with other more expensive VR or biofeedback technologies designed for use in the clinic or laboratory, can be easily set up in the individuals own environment. Additionally, the use of the Sony Eyetoy can be adjusted to accommodate a wide variety of stages in recovery, as it can be used with those who are ambulatory, those beginning to work on standing balance, or those who use a wheelchair as a primary means of mobility. The potential convenience of this system makes it appealing. Even during the early stages of stroke recovery, therapy is typically performed for three hours a day, resulting in long periods of time spent immobile within the conned space of a hospital room. VR systems may provide a way for general movement and exercise to be performed safely with minimal supervision in this context. Individuals with residual decits following a CVA have demonstrated improvement in function long after acute stages of stroke recovery.Given that stroke survivors typically exhaust traditional options for physical therapy within approximately two to four months, available treatment may be insufcient to realize full potential for recovery. Given the implications for long-term improvement in quality of life, as well as a possible decline in costs associated with long-term stroke care, the development of rehabilitation opportunities in chronic stroke seems essential. Perhaps the best evidence that the participant in this case report found the Sony Eyetoy a relevant form of training is that she reported purchasing her own upon the completion of the study. She continues to play on a regular basis and has reported that it is a fun activity both she and her grandson participate in together.

SUMMARY

The improvements demonstrated by the participant in this case study present preliminary evidence for the effectiveness of low-cost VR sensory and motor retraining. Challenges to balance, proprioception, coordination, and motor planning with concurrent visual and auditory feedback may provide an environment conducive to sensorimotor recovery. As no adverse effects were noted during the gaming sessions, self-directed training in the home environment can be assumed to be relatively safe, even for the older adult. Furthermore, the training was found to be engaging and motivating for this individual, in a way that other forms of exercise were not. The objective of this research was not to look at replacing traditional physical therapy options, but to explore lowcost VR gaming as an adjunct to these options. Commonly available VR systems may have a role to play in a treatment gap, creating therapeutic alternatives that are effective, inexpensive, and result in a decrease in hospital expenses over a lifetime. REFERENCES

1. Gresham GE, Fitzpatrick TE, Wolf PA, et al. Residual disability in survivors of strokethe Framingham study. N Engl J Med. 1975;293: 954 956. 2. American Heart Association. Heart disease and stroke statistics-2006 update. Circulation. 2006;(113):e85 e151.
As the participant is a former physical therapist, this opinion is offered from an experienced perspective. For the individual that feels an emotional drive to win a game, there
3. Tyson SF, Hanley M, Chillala J, et al. Balance disability after stroke. Phys Ther. 2006;86:30-38. 4. Duncan P, Richards L, Wallace D, et al. A randomized, controlled pilot study of a home-based exercise program for individuals with mild and moderate stroke. Stroke. 1998;29:20552060. 5. Duncan P, Studenski S, Richards L, et al. Randomized clinical trial of therapeutic exercise in subacute stroke. Stroke. 2003;34:21732180. 6. Potempa K, Lopez M, Braun LT, et al. Physiological outcomes of aerobic exercise training in hemiparetic stroke patients. Stroke. 1995; 26:101105. 7. Dean CM, Shepherd RB. Task-related training improves performance of seated reaching tasks after stroke. A randomized controlled trial. Stroke. 1997;28:722728. 8. McCloy R, Stone R. Science, medicine, and the future: virtual reality in surgery. BMJ. 2001;323:912915. 9. Ungs TJ. Simulator induced syndrome: evidence for long-term aftereffects. Aviat Space Environ Med. 1989;60:252-255. 10. Rizzo A, Pair J, McNerney PJ, et al. Development of a VR therapy application for Iraq war military personnel with PTSD. Stud Health Technol Inform. 2005;111:407 413. 11. Merians AS, Jack D, Boian R, et al. Virtual reality-augmented rehabilitation for patients following stroke. Phys Ther. 2002;82:898 915. 12. Kuttuva M, Boian R, Merians A, et al. The Rutgers Arm, a rehabilitation system in virtual reality: a pilot study. Cyberpsychol Behav. 2006;9: 148 151. 13. Henderson A, Korner-Bitensky N, Levin M. Virtual reality in stroke rehabilitation: a systematic review of its effectiveness for upper limb motor recovery. Top Stroke Rehabil. 2007;14:52 61. 14. Jaffee DL, Brown DA, Pierson-Crey C, et al. Stepping over obstacles to improve walking in individuals with poststroke hemiplegia. J Rehabil Res Dev. 2004;41:283292. 15. Kwon Y, Joong HK, Mi YL, et al. Recovery in chronic stroke: an experimenter-blind randomized study virtual reality-induced cortical reorganization and associated locomotor. Stroke. 2005;36:1166 1171. 16. Deutsch JE, Merians AS, Adamovich S, et al. Development and application of virtual reality technology to improve hand use and gait of individuals post-stroke. Restor Neurol Neurosci. 2004;22:371386. 17. Deutsch JE, Paserchia C, Vecchione C, et al. Improved gait and elevation speed in individuals post-stroke after lower extremity training in virtual environments. J Neurol Phys Ther. 2004;28:185186. 18. Fung J, Richards CL, Malouin F, et al. A treadmill and motion coupled virtual reality system for gait training post-stroke. Cyberpsychol Behav. 2006;9:157162. 19. Bryanton C, Bosse J, Brien M, et al. Feasibility, motivation, and selective motor control: virtual reality compared to conventional home exercise in children with cerebral palsy. Cyberpsychol Behav. 2006;9: 123128. 20. Ferrarin M, Brambilla M, Garavello L, et al. Microprocessor-controlled optical stimulating device to improve the gait of patients with Parkinsons disease. Med Biol Eng Comput. 2004;42:328 332. 21. Jack D, Boian R, Merians AS, et al. Virtual reality-enhanced stroke rehabilitation. IEEE Trans Neural Syst Rehabil Eng. 2001;9:308 318. 22. Krakauer JW. Motor learning: its relevance to stroke recovery and neurorehabilitation. Curr Opin Neurol. 2006;19:84 90. 23. Winstein CJ, Merians AS, Sullivan KJ. Motor learning after unilateral brain damage. Neuropsychologia. 1999;37:975987. 24. Langhorne P, Wagenaar R, Partridge C. Physiotherapy after stroke: more is better? Physiother Res Int. 1996;1:75 88. 25. Shumway-Cook AA, Anson DD, Haller SS. Postural sway biofeedback: its effect on reestablishing stance stability in hemiplegic patients. Arch Phys Med Rehabil. 1988;69:395 400. 26. Sackey CCM, Lincoln NNB. Single blind randomized controlled trial of visual feedback after stroke: effects on stance symmetry and function. Disabil Rehabil. 1997;19:536 546. 27. Cheng PT, Wang CM, Chung CY, et al. Effects of visual feedback rhythmic weight-shift training on hemiplegic stroke patients. Clin Rehabil. 2004;18:747753. 28. Engardt M, Knutsson E. Dynamic thigh muscle strength after auditory feedback training of body weight distribution in stroke patients. Physiother Theory Pract. 1994;10:103112. 29. Rizzo AA, Wiederhold M, Buckwalter JG. Basic issues in the use of

55. 56. 57. 58. 59. 60. 61. 62.
measured by dynamic gait index in individuals with unilateral vestibular hypofunction. Otol Neurotol. 2004;25(5):746 751. Whitney SL, Hudak MT, Marchetti GF. The dynamic gait index relates to self-reported fall history in individuals with vestibular dysfunction. J Vestib Res. 2000;10(2):99 105. Folstein MF, Folstein SE, McHugh PR. Mini-mental state. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res. 1975;12(3):189 198. Appelros P. Characteristics of Mini-Mental State Examination 1 year after stroke. Acta Neurologica Scandinavica. 2005;112(2):88 92. Podsiadlo D, Richardson S. The timed Up & Go: a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc. 1991; 39(2):142148. Shumway-Cook A, Brauer S, Woollacott M. Predicting the probability for falls in community-dwelling older adults using the Timed Up & Go Test. Phys Ther. 2000;80(9):896 903. Harada N, Chiu V, Damron-Rodriguez J, et al. Screening for balance and mobility impairment in elderly individuals living in residential care facilities. Phys Ther. 1995;75(6):462 469. Enright PL, McBurnie MA, Bittner V, et al. The 6-min walk test: a quick measure of functional status in elderly adults. Chest. 2003;123:387398. Uswatte G, Taub E, Morris D, et al. Reliability and validity of the upper-extremity Motor Activity Log-14 for measuring real-world arm use. Stroke. 2005;36(11):24932496.
63. Gregson JM, Leathley M, Moore AP, et al. Reliability of the Tone Assessment Scale and the modied Ashworth scale as clinical tools for assessing poststroke spasticity. Arch Phys Med Rehabil. 1999;80(9): 10131016. 64. Duncan PW, Weiner DK, Chandler J, et al. Functional reach: a new clinical measure of balance. J Gerontol. 1990;45(6):M192197. 65. Hyndman D, Asburn A, Stack E. Fall events among people with stroke living in the community: circumstances of falls and characteristics of fallers. Arch Phys Med Rehabil. 2002;83(2):165. 66. Whitney SL, Poole JL, Cass SP. A review of balance instruments for older adults. Am J Occup Ther. 1998;52(8):666 671. 67. Lennon SS, Baxter DD, Ashburn AA. Physiotherapy based on the Bobath concept in stroke rehabilitation: a survey within the UK. Disabil Rehabil. 2001;23(6):254 262. 68. Wolf SL, Winstein CJ, Miller JP, et al. Effect of constraint-induced movement therapy on upper extremity function 3 to 9 months after stroke: the EXCITE randomized clinical trial. JAMA. 2006;296(17): 20952104. 69. Wolf SL, Lecraw DE, Barton LA, et al. Forced use of hemiplegic upper extremities to reverse the effect of learned nonuse among chronic stroke and head-injured patients. Exp Neurol. 1989;104(2):125132. 70. Kunkel A, Kopp B, Muller G, et al. Constraint-induced movement therapy for motor recovery in chronic stroke patients. Arch Phys Med Rehabil. 1999;80 (6):624 628.

A m e r i c a n

P h y s i c a l

T h e r a p y

A s s o c i a t i o n
Preconference Course: Section on Geriatrics
Clinical Residency 101: Getting Started and Doing It Well
Wednesday, February 6, Contact Hours
Presented by: Carol M Davis, PT, EdD, MS, FAPTA; Greg W Hartley, PT, MSPT, GCS; Teresa Schuemann, PT, SCS, ATC, CSCS; Patricia McCord, PT, FAAOMPT, OCS
This workshop is ideal for individuals and organizations interested in developing a credentialed clinical residency. Learn about the process from individuals who have guided their clinical residency through a successful credentialing outcome and from representatives of APTAs Committee on Residency Credentialing. Innovative ways to address the credentialing criteria will be explored to make a clinical residency t your unique situation. Cosponsored by the following APTA sections: Sports, Womens Health, Neurology, Orthopaedics, and Federal Physical Therapy. Members of the Section on Geriatrics and of all cosponsoring sections register at a discount.
Course space is limited: Visit www.apta.org and click events to register today!