很稀奇么？

http://research.scea.com/ps3_deferred_shading.pdf

  
Deferred Pixel Shading on the PLAYSTATION®3
1
  
Abstract— This paper studies a deferred pixel shading algorithm
implemented on a Cell/B.E.-based computer entertainment
system.
The pixel shader runs on the Synergistic Processing Elements
(SPEs) of the Cell/B.E. and works concurrently with the GPU to
render images.  The system's unified memory architecture allows
the Cell/B.E. and GPU to exchange data through shared textures.  
The SPEs use the Cell/B.E. DMA list capability to gather
irregular fine-grained fragments of texture data generated by the
GPU.  They return resultant shadow textures the same way.  The
shading computation ran at up to 85 Hz at HDTV 720p
resolution on 5 SPEs and generated 30.72 gigaops of
performance.  This is comparable to the performance of the
algorithm running on a state of the art high end GPU.  These
results indicate that the Cell/B.E. can effectively enhance the
throughput of a GPU in this hybrid system by alleviating the
pixel shading bottleneck.

Index Terms—Computer Graphics, HDTV, Parallel Algorithms,
Rendering
I.  INTRODUCTION
he current trend toward multi-core microprocessor
architectures has led to performance gains that exceed
the predictions of Moore's law.  Multiple cores first
became prevalent as fragment processors in graphics
processing units (GPUs).  More recently the CPUs for
computer entertainment systems and desktop systems have
embraced this trend.  In particular the Cell/B.E. processor
developed jointly by IBM, Sony and Toshiba contains up to
nine processor cores with a high concentration of floating
point performance per chip unit area.

We have explored the potential of the Cell/B.E. for
accelerating graphical operations in the PLAYSTATION®3
computer entertainment system.  This system combines the
Cell/B.E. with a state of the art GPU in a unified memory
architecture.  In this architecture both devices share access to
system memory and to graphics memory.  As a result they can
share data and processing tasks.

We explored moving pixel shader computations from the GPU
to the Cell/B.E. to create a hybrid real time rendering system.  

Alan Heirich is with the Research and Development department of Sony
Computer Entertainment America, Foster City, California.  
Louis Bavoil is with Sony Computer Entertainment America R&D and the
University of Utah, School of Computing, Salt Lake City, UT   (e-mail:
bavoil@sci.utah.edu).
Our initial results are encouraging and we find benefits from
the higher clock rate of the Cell/B.E. and the more flexible
programming model.  We chose an extreme test case that
stresses the memory subsystem and generates a significant
amount of DMA waiting.  Despite this waiting the algorithm
scaled efficiently with speedup of 4.33 on 5 SPEs.  This
indicates the Cell/B.E. can be effective in speeding up this sort
of irregular fine-grained shader.  These results would carry
over to less extreme shaders that have more regular data
access patterns.

The next two sections of this paper introduces the graphical
problems we are solving and describe related work.  We next
describe the architecture of the computer entertainment system
under study and performance measurements of the pixel
shader.  We study the performance of that shader on a test
image and compare it to the performance of a high-end state
of the art desktop GPU, the NVIDIA GeForce 7800 GTX.  
Our results show the delivered performance of the Cell/B.E.
and GPU were similar even though we were only using a
subset of the Cell/B.E. SPEs.  We finish with some
concluding remarks.
II.  PIXEL SHADING ALGORITHMS
We study variations of a Cone Culled Soft Shadow algorithm
[3].  This algorithm belongs to a class of algorithms known as  
shadow mapping algorithms [15].  We first review the basic
algorithm then describe some variations.
A.  Soft Shadows
Soft shadows are an integral part of computing global
illumination solutions.  Equation (1) describes an image with
soft shadows in which, for every pixel, the irradiance L
arriving at a visible surface point from an area light source is

Ω ⎥
⎦
⎤
⎢
⎣
⎡
= ∫
Ω
Vd
r
E L i l
light
light
2
cos cos
π
θ θ
(1)
In this equation Ωlight is the surface of the area light and dΩ is
the differential of surface area.  Elight is the light emissivity per
unit area, and θl , θi are the angles of exitance and incidence of
a ray of length r that connects the light to the surface point.  V
is the geometric visibility along this ray, either one or zero.
The distance term 1 / π r
2
reflects the reduction in subtended
solid angle that occurs with increasing distance.  This  
expression assumes that the material surface is diffuse
(Lambertian).
Deferred Pixel Shading on the
PLAYSTATION®3
Alan Heirich and Louis Bavoil
T   
Deferred Pixel Shading on the PLAYSTATION®3
2

When V=1 and Ωlight has area dΩ this equation describes
diffuse local illumination from a point light as is typically
computed by GPUs using rasterization.  When this equation is
expanded recursively in E (by treating each surface point as a
source of reflected light) the result is a restriction of the
Rendering Equation of global illumination  [12] to diffuse
surfaces.
B.  Cone Culled Soft Shadows

Equation (1) is traditionally solved by offline methods like ray
tracing.  Stochastic ray tracing samples the integrand at
various points on Ω and accumulates the result into L.  The
CCSS algorithm takes an analogous approach, rendering from
the light and gathering the radiance from the resulting
fragments into pixels.

The CCSS algorithm consists of fragment generation steps
and a pixel shading step.  We have implemented fragment
generation on the GPU and pixel shading on the Cell/B.E.  
The GPU is programmed in OpenGL-ES using Cg version 1.4
for shaders.  Fragments are rendering into OpenGL-ES
Framebuffer Object texture attachments using one or more
render targets.  These textures are then detached from the
Framebuffer Objects and used as input to the pixel shading
step.  The pixel shading step returns a shadow texture which is
then bound to the GPU for final rendering.

The algorithm is not physically correct and we accept many
approximations for the sake of real time performance.  Lights
are assumed to be spherical which simplifies the gathering
step.  Light fragments for each pixel are culled against conical
frusta rooted at the pixel centroid.  These frusta introduce
geometric distortions due to their mismatch with the actual
light frustum.

The culling step uses one square root and two divisions per
pixel.  No acceleration structure is used so the algorithm is
fully dynamic and requires no preprocessing.  The algorithm
produces high quality shadows.  It renders self-shadowed
objects more robustly than conventional shadow mapping
without requiring a depth bias or other parameters.

1)  Eye Render

The first fragment generation step captures the locations of
pixel centroids in world space.  This is done by rendering
from the eye view using a simple fragment shader that
captures transformed x, y and z for each pixel.  We capture z
rather than obtaining it from the Z buffer in order to avoid
imprecision problems that can produce artifacts.  We use the
depth buffer in the conventional way for fragment visibility
determination.

If this is used as a base renderer (in addition to rendering
shadows) then the first step also captures a shaded
unshadowed color image.  This unshadowed image will later
be combined with the shadow texture to produce a shadowed
final image.  For some shaders, such as approximate indirect
illumination, this step can also capture the surface normal
vectors at the pixel location.

2)  Light Render

The second fragment generation step captures the locations
and alpha values (transparency) of fragments seen from the
light.  For each light, for each shadow frustum, the scene is
rendered using the depth buffer to capture the first visible
fragments.   The positions and alphas of the fragments are
generated by letting the rasterizer interpolate the original
vertex attributes.   For some shaders, including colored
shadows and approximate indirect illumination, this step also
captures fragment colors.

3)  Pixel Shading

In the third step, performed on the Cell/B.E., light fragments
are gathered to pixels for shading.  Pixels are represented in
an HDTV resolution RGBA texture that holds (x,y,z) and a
background flag for each pixel.  Light fragments are contained
in one (or more) square textures.

Pixel shading proceeds in three steps:  gathering the kernel of
fragments for culling; culling these fragments against a
conical frustum; and finally computing a shadow value from
those fragments that survived culling.

4)  Fragment Gather

For each pixel, for each light, the pixel location (x,y,z) is
projected into the light view (x',y',z').  A kernel of fragments
surrounding location (x',y',0) in the light texture is gathered
for input to the culling step.  Figure 1 illustrates this projection
and the surrounding kernel.

It is not necessary to sample every location in the kernel, and
performance gains can be realized by subsampling strategies.
In our present work we are focused on system throughput and
so we use a brute-force computation over the entire kernel.

5)  Cone Culling

For each pixel, for each light, a conical frustum is constructed
tangent to the spherical light with its apex at the pixel centroid
as illustrated in figure 2.  The gathered fragments are tested
for inclusion in the frustum using an efficient point-in-cone
test.

The point-in-cone test performs these computations at each
pixel:

axis  =  light.centroid – pixel.centroid
alength
2

=  axis . axis
cos
2
θ  = alength2
/ (light.radius
2
+
alength2
)   
Deferred Pixel Shading on the PLAYSTATION®3
3
na = normalize(axis)



Figure 1 (kernel lookup). The pixel is projected from the
world into the light plane, which is equivalent to finding the
nearest fragment F in the light view to the ray from the pixel
to the light center.  In this example fragment F blocks the ray
from the light to the pixel, and we say F shadows the pixel.



Figure 2 (cone culling).  Computing the shadow intensity at a
pixel in a cone with the apex at the pixel and tangent to the
light sphere.   The fragments of the light view are fetched in a
kernel centered at the projection of the cone axis over the
light plane.  Fragments are tested for visibility using an
efficient point-in-cone test+.   

The point-in-cone then performs these computations for each
fragment:

fe = fragment.centroid  –
pixel.centroid
axisDotFe  =   na . fe
direction  =  (axisDotFe > 0)
flength2
  =  fe . fe
inside =  (cos
2
θ * flength2
<= axisDotFe
2
)
pointInCon
e
=  direction && inside

(An expression for cos
2
θ that more accurately reflects the
tangency between the cone and sphere is (alength2
–
light.radius
2
) / alength2
).

6)  Computing new shadow values

The final step is to compute shadow values from the
fragments that survived the culling step.  Here we describe
three such shading computations, and others are possible.  We
present detailed performance measurements of the
monochromatic shader in section 5.  We have implemented
substantial portions of the other shaders on the Cell/B.E. and
GPU to verify proof-of-concept.
a)  Monochromatic soft shadows

We can compute monochromatic soft shadows from
translucent surfaces by using a generalization of the
Percentage Closer Filtering algorithm  [14].  Among the
fragments that survived cone culling we compute the mean
alpha (transparency) value.   The resulting shadow factor is
one minus this mean.  At pixels where no fragments survived
culling the shadow factor is one. Test images for this shader
appear in figure 3.
b)  Colored soft shadows

We can obtain colored shadows by including the colors of the
translucent fragments and of the light source.  In addition to
computing the mean alpha value we also compute the mean
RGB for the fragments.  This requires gathering twice as
much fragment data for the shading computation.  We
multiply these quantities with the light source color to obtain a
colored shadow factor.  At pixels where no fragments
survived culling the shadow factor is one.
c)  Approximate indirect illumination

It is worth noting that an approximate indirect illumination
component can be computed similarly to Frisvad et. al.'s
Direct Radiance Map algorithm [5].  This requires accounting
for a transport path from light source to fragment to pixel.  
This estimate is approximate because it does not account for
occluding objects between the fragment and the pixel and also
because it only samples a limited kernel of fragments.

Assuming the fragment materials are diffuse (Lambertian), the
irradiance at the fragment can be estimated during the light
render step proportional to the cosine of the incident angle at
the fragment.  The subsequent reflected radiance at the pixel is
this irradiance times the cosine of the incident angle at the
pixel.  This radiance can be estimated during the pixel shading
step if we have the surface normal at the pixel.  This surface
normal can be generated during the eye render step.

This computation requires more DMA traffic to accommodate
the pixel normals.   Since this is not part of the gathered   
Deferred Pixel Shading on the PLAYSTATION®3
4
   

Figure 3: some test images of complex models rendered using the monochromatic shader.  (Left) the dandelion is a challenging
test for shadow algorithms.  The algorithm correctly reproduced the fine detail at the base of the plant as well as the internal
self-shadowing within the leaves.  (Right) a tree model with over 100,000 polygons rendered above a grass colored surface.

fragment data it can be accommodated efficiently using
predetermined transfers of large blocks of data.
III. RELATED WORK
There is an extensive existing literature on shadow
algorithms.  For a recent survey of real-time soft shadow
algorithms see [6].  For a broad review of traditional
shadow algorithms see [16].

The most efficient shadow algorithms work in image space
to compute the shading for each pixel with respect to a set
of point lights.  The original image-space algorithm for
point lights is shadow mapping [15].  In this algorithm the
visible surface of each pixel is transformed into the view of
the light and then compared against the first visible surface
as seen from the light.  If the first visible surface lies
between the transformed pixel and the light then the
transformed pixel is determined to be in shadow.

Traditional shadow mapping produces “hard'' shadows that
are solid black with jagged edges.  They suffer from many
artifacts including surface acne (false self-shadowing due to
Z imprecision) and aliasing from imprecision in sampling
the light view.

The Percentage Closer Filtering algorithm [14] is
implemented in current GPUs to reduce jagged shadow
edges.  This algorithm averages the results of multiple
depth tests within a pixel to produce fractional visibility for
pixels on shadow boundaries.  This has the effect of
softening shadow boundaries but since it is a point light
algorithm it does not produce the wide penumbrae that
characterize shadows from area lights.   


Adaptive Shadow Maps  [4,13] address the problem of
shadow map aliasing by computing the light view at
multiple scales of resolution.  The multiresolution map is
stored in the form of a hierarchical adaptive grid.  This
approach can be costly because the model must be rendered
multiple times from the light view, once for each scale of
resolution.

Layered Depth Interval maps [2] combine shadow maps
taken from multiple points on the light surface.  These are
resolved into a single map that represents fractional
visibility at multiple depths.  In practice four discrete
depths were sufficient to produce complex self-shadowing
in foliage models.  This method produces soft shadows at
interactive rates but is costly because it requires multiple
renders per light.  It does not address translucency.

The irregular Z-buffer [11] has been proposed for hardware
realization for real-time rendering.  It causes primitives to
be rasterized at points specified by a BSP tree rather than
on a regular grid.  As a result it can eliminate aliasing
artifacts due to undersampling.  This is similar to Alias-free
Shadow Maps [1].

Jensen and Christensen extended photon mapping [10] by
prolongating the rays shot from the lights and storing the
occluded hit points in a photon map which is typically a kd-
tree.   When rendering a pixel x the algorithm looks up the
nearest photons around x and counts the numbers of
shadow photons ns and  illumination photons ni in the
neighborhood.  The shadow intensity is then estimated as  V
= ni / (ns + ni).   Our algorithm uses similar concepts to
gather fragments and shade pixels, and in addition works
with translucent materials.   
Deferred Pixel Shading on the PLAYSTATION®3
5

Figure 4: the PLAYSTATION®3 architecture.  The 3.2 GHz Cell/B.E. contains a Power Architecture processor (the PPE) and
seven Synergistic Processing Elements (SPEs) each consisting of a Synergistic Processing Unit (SPU), 256 KB local store (LS),
and a Memory Flow Controller (MFC).  These processors are connected to each other and to the memory, GPU and peripherals
through a 153.6 GB/s Element Interconnect Bus (EIB).  The Cell/B.E. uses Extreme Data Rate (XDR) memory which has a peak
bandwidth of 25.6 GB/s.  The GPU interface (IOIF) to the EIB provides 20 GB/s in and 15 GB/s out.  Memory accesses by the
Cell/B.E. to GPU memory pass through the EIB, IOIF and GPU.  Access by the GPU to XDR pass through the IOIF, EIB and
MIC.

IV. PLAYSTATION®3 SYSTEM

Figure 4 shows a diagram of the PLAYSTATION®3
computer entertainment system and its 3.2 GHz Cell/B.E.
multiprocessor CPU.  The Cell/B.E. consists of an IBM
Power Architecture core called the PPE and seven SPEs.  
(While the Cell/B.E. architecture specifies eight SPEs our
system uses Cell/B.E.s with seven functioning SPEs in
order to increase manufacturing yield.)  The processors are
connected to each other and to system memory through a
high speed Element Interconnect Bus (EIB).  This bus is
also connected to an interface (IOIF) to the GPU and
graphics memory.  This interface translates memory
accesses in both directions, allowing the PPE and SPEs
access to graphics memory and providing the GPU with
access to system memory.  This feature makes the system a
unified memory architecture since graphics memory and
system memory both are visible to all processors within a
single 64-bit address space.

The PPE is a two way in order super-scalar Power
Architecture core with a 512 KB level 2 cache.  The SPEs
are excellent stream processors with a SIMD (single
instruction, multiple data) instruction set and with 256 KB
local memory each.  SIMD instructions operate on 16-byte
registers and load from and store to the local memory.   The
registers may be used as four 32-bit integers or floats, eight
halfwords, or sixteen individual bytes.  DMA (direct
memory access) operations explicitly control data transfer
among SPE local memories, the PPE level 2 cache, system
memory, and graphics memory.  DMA operations can chain
up to 2048 individual transfers in size multiples of eight
bytes.

The system runs a specialized multitasking operating
system.  The Cell/B.E. processors are programmed in C++
and C with special extensions for SIMD operations.  We
used the GNU toolchain g++, gcc and gdb.  The GPU is
programmed using the OpenGL-ES graphics API and the
Cg shader language.

The Cell/B.E. supports a rich variety of communication and
synchronization primitives and programming constructs.
Rather than describe these here we refer the interested
reader to the publicly available Cell/B.E. documentation
[7]-[9].
V. RESULTS
We implemented the CCSS algorithm as described in
section 2 using the monochromatic pixel shader described
in II.5 and II.6.a.  We implemented it in hybrid form on the
computer entertainment system using the Cell/B.E. and
GPU, and also on a standalone high end GPU for
comparison.

On the Cell/B.E. we measured performance in three stages:
fragment rendering, shadow generation, and final draw.  
Times and performance measurements are shown in tables
1 through 4.

Eye
render
Light
render
1-SPE 5-SPEs Draw
time
10.11  3.29  50.47 11.65 5.6

Table 1: Performance of stages of the algorithm.  All times
are in milliseconds.  The eye and light render stages are
performed on the GPU as is the final draw.  Pixel shading
is performed on the SPEs.  We measured the time for pixel
shading using from 1 to 5 SPEs.  The results showed good
parallel speedup.  Detailed measurements of pixel shading
are given in tables 2 and 3.
A.  Cell/B.E. Software Implementation
Eye and light fragments are rendered to OpenGL-ES
Framebuffer Object texture attachments.  We used 32 bit
float RGBA textures for all data.  The textures for these
attachments may be allocated in linear, swizzled or tiled   
Deferred Pixel Shading on the PLAYSTATION®3
6
formats in either GPU or system memory.  We
experimented with all combinations of texture format and
location in order to find the combination that gave the best
performance.

GPU performance is highest rendering to native tiled
format in GPU memory.  The performance advantage is
high enough that it is worth rendering in tiled format and
then reformatting the data to linear allocation for processing
by the Cell/B.E.  In order to minimize the latencies incurred
by the SPEs in accessing this data we reformat the data into
system memory rather than GPU memory.

The key to running any algorithm on the SPEs is to develop
a streaming formulation in which data can be moved
through the processor in blocks.  We move eye data in
scanline order and double buffer the scanline input.  While
one scanline of pixels is being processed we prefetch the
next scanline.  As each scanline is completed it is written to
the shadow texture.  We have measured the DMA waiting
for the scanline data and it was negligible.

For every pixel of input we generate a series of DMA
transactions to gather the necessary light fragments.  The
source address for each transaction is a location inside the
light fragment buffer.  We compute this address by
applying a linear transform (matrix multiplication) to the
eye data (x,y,z) to obtain a light coordinate (x',y',z').

These transactions are bundled into long DMA lists.  By
having multiple DMA lists in flight concurrently we buffer
fragment data in order to minimize DMA waiting.  We
experimented with the number and size of the DMA lists in
order to minimize runtime.  We found that having four
DMA lists was optimal and that larger numbers did not
reduce the runtime.  We found similarly that fetching 128
pixels per DMA list was optimal and that longer DMA lists
did not reduce runtime.

We parallelized the computation across multiple SPEs by
distributing scanlines to processors.  This is straightforward
and provides balanced workloads.  We scheduled tasks
using an event queue abstraction provided by the operating
system that is based on one of the Cell/B.E.
synchronization primitives, the mailbox.  We measured the
cost of this abstraction at less than 100 microseconds per
frame.  When running in parallel on multiple SPEs the
individual   processors completed their work within 100
microseconds of each other.

Each SPE computes a set of scanlines for the shadow
texture.  They deliver their result directly into GPU
memory in order to minimize the final render time.   
B.  Measurements
We validated the correctness of the implementation by
rendering a variety of models under different conditions.
We then made detailed measurements of performance and
scaling of the tree model in figure 3.  These measurements
appear in tables 2 and 3.  All of our measurements used a
single light source.  The tree model contains over 100,000
polygons.  The performance of the shading computation is
independent of the time required to generate the fragments,
and thus is independent of the geometric complexity of the
model.

  1-
SPE
2-
SPEs
3-
SPEs
4-
SPEs
5-
SPEs
Full  50.47 28.86 16.78 13.25 11.65
Hz  19 34 59 75 85
Speedup  1  1.75 3.01 3.81 4.33
Scaling  1  0.87 1.00 0.95 0.87
No
waiting
41.97 21.05 14.09 10.63 8.56
Speedup  1  1.99 2.98 3/95 4.90
Scaling  1  1.00 0.99 0.99 0.98

Table 2: Parallel performance of the pixel shading
computation.  All times are in milliseconds.  Images were
rendered at HDTV 720p resolution (1280x720 pixels).  The
tree was rendered with data-dependent optimizations
disabled in order to obtain worst-case times.  The image
was rendered using the full algorithm (“full'') and with the
DMA fragment gather operation disabled (“no waiting'').  
The computation was exactly the same in both cases, but in
the “no waiting'' case the shader processed uninitialized
fragment data.  The speedup and scaling efficiency was
evaluated in all cases.  These results show that the
computation speeds up almost perfectly but that substantial
time is lost waiting for the gather operation.  Further
information about the DMA costs appears in table 3.

  1-
SPE
2-
SPEs
3-
SPEs
4-
SPEs
5-
SPEs
Wait
time
8.50 7.81  2.69  2.62  3.09
%
waiting
17 27  16  20  27
DMA
GB/s
2.53 4.43  7.62  9.66  10.98
DMA
per
second
42.47
M
74.27
M
127.73
M
161.76
M
183.97
M

Table 3: DMA costs on different numbers of SPEs. All
times are in milliseconds.  The algorithm spent
considerable time waiting for the results of the DMA
fragment gather operation (“wait time'').  Expressed as a
percentage of the pixel shading computation, the
monochromatic shader spent between 17 and 27 percent
waiting for fragment DMA.  This explains the deviation
from ideal scaling in table 2.  The Cell/B.E. sustained 10.98
GB/s of DMA traffic using packet sizes that were
predominantly 48 bytes in length, and over 183 mega-
transactions (M=10242
) per second.
   
Deferred Pixel Shading on the PLAYSTATION®3
7
All images were rendered at HDTV 720p resolution,  
1280x720 pixels.  We used lightmap resolution of  
1024x1024 in our experiments and a 3x3 fragment kernel.  
In order to ensure that we measured worst-case
performance we disabled optimizations that skipped
background pixels and transparent fragments.  We
measured performance on one to five SPEs.  In our tests the
other two SPEs were in use by graphics and operating
system services.
C.  Data Analysis
Tables 1 and 2 show that the shading calculation can be
sped up to meet any realistic performance requirement.  
The monochromatic shader ran at 85 Hz using 5 SPEs and
at 34 Hz using 2 SPEs.  Videogames are typically rendered
at 30 or 60 frames per second.  Shading calculations should
generally run at these rates, but for shadow generation it is
possible to use lower frame rates without affecting image
quality.  It would also be possible to use shadows generated
at 720p resolution with a base image rendered at a higher
1080p resolution (1920x1080 pixels).

Table 3 analyzes the time spent waiting for DMA
transactions to complete.  This was as much as 27% of the
total time.  Note that if we were able to remove all of this
DMA waiting the performance on 5 SPEs would reach 116
frames per second as indicated by the”no waiting'' data in
table 1.

While it is difficult to observe the DMA behavior directly
we can reason about the bottlenecks in our computation.  
Every DMA transaction costs the memory system at least
eight cycles of bandwidth no matter how small the
transaction.   Thus 400 M transactions per second is an
upper limit of the system memory performance.  The shader
generated 183.97 M DMA transactions per second which
does not approach the limits of the memory system.  Most
of these were 48-byte gathers of light view fragments,
while the rest were block transfers of entire scanlines 20
KB in size.

We profiled the runtime code to measure the number of
SIMD operations that were spent in DMA address
calculations.  The results appear in table 4.  We found that
we were spending between 14% and 17% of operations
supporting the DMA gather operation.

DMA
addressing
Shading Total  DMA
percentage
16,358,400 79,718,400 96,076,800 17

Table 4: Results of run-time profiling.  These figures count
the number of SIMD instructions executed per frame for
both shaders in the inner loop and DMA addressing
calculations.  It does not include the cost of scalar code
that controls the outer loop.  The number of operations is
four times the number of instructions.  The last column
shows the percentage of SIMD operations that were spent
computing addresses for the DMA gather.

We also measured the time to execute the scalar control
logic and perform the DMA for the eye render fragments in
order to better estimate the cost of shaders with scanline
order data access.  These DMA operations are for an entire
scanline at a time, 20 K bytes in size.    Each frame reads
and writes each scanline once for a total of 28.125
megabytes of DMA activity using two transactions.  On one
SPE this required 2.13 ms of time yielding an effective
transfer rate of over 12.89 GB/s.  For shaders with scanline
order access, it should be possible to read as much as five
times as much scanline data without exhausting the overall
DMA bandwidth or the number of DMA transactions.
D.  Comparison to GeForce 7800 GTX GPU
We implemented the same algorithm on a high end state of
the art GPU, the NVIDIA GeForce 7800 GTX running in a
Linux workstation.  This GPU has 24 fragment shader
pipelines running at 430 Mhz and processes 24 fragments
in parallel.  By comparison the 5 SPEs that we used process
20 pixels in parallel in quad-SIMD form.

The GeForce required 11.1 ms to complete the shading
operation.  In comparison the Cell/B.E. required 11.65 ms
including the DMA waiting time, and would require only
8.56 ms if the DMA waiting were eliminated.  The
performance of the Cell/B.E. with 5 SPEs was thus
comparable to one of the fastest GPUs currently available,
even though our implementation spent 27% of its time
waiting for DMA.  Results would presumably be even
better on 7 SPEs, or on fewer SPEs if we could reduce or
eliminate the DMA waiting.
VI.  REMARKS
We have explored moving pixel shaders from the GPU to
the Cell/B.E. processor of the PLAYSTATION®3
computer entertainment system.  Our initial results are
encouraging as they show it is feasible to attain scalable
speedup and high performance even for shaders with
irregular fine-grained data access patterns.  Removing the
computation from the GPU effectively increases the frame
rate, or more likely, the geometric complexity of the models
that can be rendered in real time.

We can also conclude that the performance of the Cell/B.E.
is superior to a current state of the art high end GPU in that
we achieved comparable performance despite performance
limitations and despite using only part of the available
processing power.  Our current implementation loses
substantial performance due to DMA waiting.  This results
from the fine-grained irregular access to memory and is
specific to the type of shaders we have chosen to
implement.  We have explored shaders based on shadow
mapping [15] which require evaluating GPU fragments
generated from multiple viewpoints.  These multiple
viewpoints are related to each other by a linear viewing
transformation.  Gathering the data from these multiple
viewpoints requires fine-grained irregular memory access.   
Deferred Pixel Shading on the PLAYSTATION®3
8
This represents worst-case behavior for any memory
system.
REFERENCES
[1]  Timo Aila and Samuli Laine, “Alias-Free Shadow Maps,”  in Proc.
Rendering Techniques 2004: 15th  Eurographics  Workshop on
Rendering, 2004, pp. 161-166.
[2]  Maneesh Agrawala, Ravi Ramamoorthi,  Alan Heirich and Laurent
Moll, “Efficient Image-Based Methods for Rendering Soft
Shadows,” in Proc. ACM SIGGRAPH, 2000, pp. 375-384.
[3]  Louis Bavoil and Claudio T. Silva,. “Real-Time Soft Shadows with
Cone Culling,” ACM SIGGRAPH Sketches and Applications, 2006.
[4]  Randima Fernando, Sebastian Fernandez, Kavita Bala and Donald P.
Greenberg, “Adaptive Shadow Maps”, in  Proc. ACM SIGGRAPH,
2001, pp. 387-390.
[5]  J. R. Frisvad and R. R. Frisvad and N. J. Christensen and P. Falster,
“Scene independent real-time  indirect illumination,”, in  Proc.
Computer Graphics International, 2005, pp. 185-190.
[6]  Jean-Marc Hasenfratz, Marc Lapierre, Nicolas Holzschuch and
Francois Sillion, “A survey of Real-Time Soft Shadows Algorithms,”
Computer Graphics Forum, vol. 22, no. 4, 2003, pp. 753-774.
[7]  IBM, Sony and Toshiba, “Cell Broadband Engine Architecture
version 1.0,” August 8, 2005.
[8]  IBM, Sony and Toshiba, “SPU  Assembly Language Specification
version 1.3,” October 20, 2005.
[9] IBM, Sony and Toshiba, “SPU C/C++ Language Extensions version
2.1,” October 20, 2005.
[10]  Henrik Wann Jensen and Per H. Christensen, “Efficient Simulation
of Light Transport in Scenes with Participating Media Using Photon
Maps,”, in Proc. ACM SIGGRAPH, 1998, pp. 311-320.
[11]  Gregory S. Johnson, Juhyun Lee, Christopher A. Burns and William
R. Mark, “The irregular Z-buffer: Hardware acceleration for irregular
data structures,”  ACM Transactions on Graphics, vol. 24, no. 4,
2005, pp. 1462-1482.
[12]  James T. Kajiya, “The Rendering Equation,” in  Proc. ACM
SIGGRAPH, 1986, pp. 143-150.
[13]  Aaron Lefohn, Shubhabrata Sengupta, Joe M. Kniss, Robert Strzodka
and John D. Owens, “Dynamic Adaptive Shadow Maps on Graphics
Hardware,”  ACM SIGGRAPH Conference Abstracts and
Applications, 2005.
[14]  William T. Reeves, David H. Salesin and Robert L. Cook,
“Rendering Antialiased Shadows with Depth Maps,” in Proc. ACM
SIGGRAPH, 1987, pp. 283-291.
[15]  Lance Williams, “Casting Curved Shadows on Curved Surfaces,” in
Proc. ACM SIGGRAPH, 1978, pp.  270-274.
[16]  Andrew Woo, Pierre Poulin and  Alain Fournier, “A Survey of
Shadow Algorithms,” IEEE Computer Graphics & Applications, vol.
10, no. 6, pp. 13-32.

引用:

原帖由 倍舒爽 于 2009-2-12 04:38 发表



只看到The system's unified memory architecture allows
the Cell/B.E. and GPU to exchange data through shared textures.  

这不是xo么？？

另外p3版的什么“幽灵行动：尖峰战士”也是这种渲染方式 ...

SCEA的网站宣传XO不成？