Patent Grant
7619623
The disclosed technology relates to graphics applications, more specifically to graphics applications involving sparsely defined spatial data.
Many graphics applications involve sparsely defined spatial data. For example, image discontinuities such as sharp vector silhouettes are generally present at only a small fraction of pixels. Texture sprites often overlay high-resolution features at sparse locations. Image attributes such as alpha masks are mainly binary, requiring additional resolution at only a small subset of pixels. Surface texture or geometry can be represented as sparse 3D data.
Compressing such sparse data while retaining efficient random-access is a challenging problem. Current solutions include the following:
A perfect hash usually refers to a hash function that maps elements into a hash table without any collisions. Generally, all the elements map to distinct slots of the hash table. The probability that randomly assigning n elements in a table of size m results in a perfect hash is
When the table is large (i.e., m>>n) one can use the approximation ex≅1+x for small x to obtain
Thus, the presence of a hash collision is highly likely when the table size m is much less than n2. This is an instance of the well-known “birthday paradox” (e.g., a group of only 23 people have more than 50% chance of having at least one shared birthday).
The probability of finding a minimal perfect hash (e.g., where n=m) is
which uses Stirling's approximation log n!≅n log n−n. Therefore, the expected number of bits needed to describe these rare minimal perfect hash functions is intuitively
Several number-theoretical methods construct perfect hash functions by exploiting the Chinese remainder theorem. However, even for sets of a few dozen elements, these functions involve integer coefficients with hundreds of digits.
A more computer-amenable approach is to define the hash using one or more auxiliary tables. One approach uses three such tables and two nested hash functions to hash a sparse set of n integers taken from u={0 . . . , u−1}. Such a scheme takes constant time and 3n log n bits of memory. The hash is constructed with a deterministic algorithm that takes O(nu) time. Another approach reduces space complexity to the theoretically optimal Θ(n) bits, but the constant is large and the algorithm difficult.
Some schemes treat perfect hashing as an instance of sparse matrix compression. They map a bounded range of integers to a 2D matrix and compact the defined entries into a 1D array by translating the matrix rows. Sparse matrix compression is known to be NP-complete.
The most practical schemes achieve compact representations and scale to larger datasets by giving up guarantees of success. These probabilistic constructions may iterate over several random parameters until finding a solution. For example, one scheme defines a hash h(k)=h0(k)+g1[h1(k)]+g2[h2(k)]mod m, where functions h0,h1,h2 map string keys k to m,
r,
r respectively, and g1,g2 are two tables of size r. However, this algorithm takes expected time O(r4), and is practical only up to n=512 elements.
Another approach involves creating the first scheme with good average-case performance (˜11n bits) on large datasets. The insight is to assign values of auxiliary tables g1,g2 in decreasing order of number of dependencies. This approach also describes a second scheme that uses quadratic hashing and adds branching based on a table of binary values. This second scheme achieves ˜4n bits for datasets of size n˜106.
Spatial Hashing
Hashing is commonly used for point and region queries in multidimensional databases. Spatial hashing is also used in graphics for efficient collision detection among moving or deforming objects. However, these techniques employ imperfect hashing (e.g., traditional multi-probe hash tables implemented on the CPU).
These techniques do not transition to multidimensional tables. Also, they strive to make intermediate hashes as random as possible. As such, there exists a need for a perfect multidimensional hash function that preserves spatial coherence and thus improves runtime locality of reference.
The following Detailed Description presents a technique of implementing a perfect multidimensional hash function to densely pack sparse spatial data. The new technique described herein preserves spatial coherence and improves runtime locality of reference by creating a multidimensional hash table, creating a multidimensional offset table, and applying a perfect multidimensional hash function to map sparse multidimensional data into the hash table using the offset table. Implementations of this technique involve precomputing the perfect multidimensional hash function on static data to avoid hash collisions. Numerous graphics applications for this new technique include sharp image silhouettes, texture sprites, alpha channel compression, 3D-parameterized textures, 3D painting, simulation, and collision detection.
This Summary is provided to introduce a selection of concepts in a simplified form that is further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter. Additional features and advantages of the invention will become more apparent from the following detailed description, which proceeds with reference to the accompanying figures.
The file of this patent contains at least one drawing executed in color. Copies of this patent with color drawing(s) will be provided by the Patent and Trademark Office upon request and payment of the necessary fee.
The disclosed technology proposes to losslessly pack sparse data into a dense table using a hash function h(p) on position p. Applying traditional hashing algorithms in the context of current graphics architecture presents several challenges:
A perfect hash function (e.g., one that is precomputed for a static set of elements to have no collisions) can make hashing more compatible with GPU parallelism. Desirably, the perfect hash function is a minimal perfect hash function (e.g., one in which the hash table contains no unused entries). The description of a minimal perfect hash function is expected to require a number of bits proportional to the number of data entries. Thus one cannot construct a perfect hash using an expression with a small number of machine-precision parameters. Rather, one must store additional data in auxiliary lookup tables.
The disclosed technology defines a perfect multidimensional hash function of the form
h(p)=h0(p)+Φ[h1(p)],
which combines two imperfect hash functions h0,h1 with an offset table Φ. Intuitively, a role of the offset table is to “jitter” the imperfect hash function h0 into a perfect one. Although the offset table uses additional memory, it can fortunately be made significantly smaller than the data itself (e.g., it typically has only 15-25% as many entries, each entry having just 8 bits per coordinate). Implemented on the GPU, this perfect hash allows data access using just one additional texture access plus approximately 4-6 more shader instructions depending on the application scenario.
Prior work on perfect hashing has focused on external storage of data records indexed by character strings or sparse integers. No known work has considered multidimensional data and its unique opportunities. Indeed, in computer graphics, 2D and 3D texture data is often accessed coherently by the parallel GPU, and is therefore swizzled, tiled, and cached. Ideally, hashed textures should similarly be designed to exploit access coherence.
Whereas prior work seeks to make intermediate hash functions like h0,h1 as random as possible, the disclosed technology instead designs them to be spatially coherent, resulting in efficient access into the offset table Φ. Second, the disclosed technology optimizes the offset values in Φ to maximize coherence of h itself. Creating a perfect hash is already a difficult combinatorial problem. Nonetheless, there remain enough degrees of freedom to improve coherence and thereby increase runtime hashing performance.
In addition to sparse data compaction, the disclosed technology also describes several schemes for encoding the spatial positions of these sparse samples. Specifically, it introduces domain bits, position tags, and parameterized position hashes.
For applications that require continuous local interpolation of the sparse data, two exemplary approaches can be considered. The first is to allow native filtering in the dedicated GPU hardware by grouping the data into sample-bordered blocks. In this setting, the disclosed technology replaces the traditional block indirection table by a compact spatial hash over the sparsely defined blocks. The limiting factor in using blocks is that data must be duplicated along block boundaries, thus discouraging small blocks sizes and leading to memory bloat.
A second exemplary solution attains a more compact representation by forgoing blocking and performing filtering explicitly as general-purpose computation. At present this incurs an appreciable loss in performance but can reduce memory by a factor 3 over blocking schemes.
In the 2D example 100, the 1282 image 102 contains a set of 1,381 pixels (8.4%) with supplemental information (e.g., vector silhouette data). This sparse pixel data is packed into a hash table 104 of size 382=1,444, which is much smaller than the original image. The perfect hash function is defined using an offset table 106 of size 182.
In the 3D example 200, a triangle mesh is colored by accessing a 3D texture 202 of size 1283. Only 41,132 voxels (2.0%) are accessed when rendering the surface using nearest-filtering. These sparse voxels are packed into a 3D table 204 of size 353=42,875 using a 203 offset table 206.
Exemplary Terminology
ūd=[0. . . (ū−1)]d, the sparse data consists of a subset s⊂u of n grid positions, where each position pεS has associated data record D(p). Thus, the data density is the fraction ρ=n/u. For datasets of codimension 1, such as curves in 2D or surfaces in 3D, one typically finds ρ˜1/ū.
One goal is to replace the sparsely defined data D(p) by a densely packed hashed texture H[h(p)] where:
One strategy is to let the hash table 304 be as compact as possible to contain the given data, and then to construct the offset table 306 to be as small as possible while still allowing a perfect hash function.
Exemplary GPU Implementation
An exemplary scheme can be described mathematically using arrays with integer coordinates (e.g., 0, 1, . . . ,
Quantizing the offset vectors stored in Φ to 8 bits per coordinate provides enough flexibility for hashing even when the hash table size 256d are scaled by ┌
The following is high-level shading language (HLSL) pseudocode for the exemplary hashing function:
Several optimizations of this code are possible. For 2D domains, the two matrix multiplication mu1 (M[i], p) can be done in parallel within a float4 tuple. Also, the matrices M0,M1 are in fact scaled identity matrices, so the matrix multiplications reduce to a single multiply instruction.
In one example, the disclosed technology seeks to assign table sizes
Assuming a hash table to be a square (d=2) or cube (d=3), its size
In the example, the next step is to assign the offset table size
For fast construction, one can initially set
For compact construction, one can perform a binary search over
Exemplary Selection of Hash Coefficients
One approach is to fill the matrices M0,M1 (defining the intermediate hash functions h0,h1) with random prime coefficients. To improve hash coherence, these matrices are sought to be made more regular.
In the example, letting M0M1 just be identity matrices does not significantly hinder the construction of a perfect hash. The functions h0,h1 then simply wrap the spatial domain multiple times over the offset and hash tables, moving over domain points and table entries in lockstep. Thus, the offset table access Φ[h1(p)] is perfectly coherent. Although h0(p) is also coherent, the hash table access H[h(p)] is generally not because it is jittered by the offsets. However, if adjacent offset values in Φ are the same (e.g., if the offset table is locally constant), then h itself will also be coherent.
One necessary condition on h0,h1 is that they must map the defined data to distinct pairs. That is, pεS→(h0(p),h1(p)) must be injective. Indeed, if there were two points p1,p2εS with h0(p1)=h0(p2) and h1(p1)=h1(p2), then these points would always hash to the same slot h(p1)=h(p2) regardless of the offset stored in Φ[h1(p1)], making a perfect hash impossible.
The condition for injectivity is similar to a perfect hash of n elements into a table of size |H|×|Φ|=mr. One can derive a probability of success of
PrPH≅e−n
which seems ominously low—only 12% for σ=r/n=0.25. This could be the main reason that previous perfect hashing schemes resorted to additional tables and hash functions.
However, unlike prior work, the disclosed technology does not involve selecting h0,h1 to be random functions. Because the functions h0,h1 have periodicities
Exemplary Creation of Offset Table
The entries Φ[q] with the largest sets h1−1(q) of dependent data points are desirably processed first. The exemplary algorithm assigns offset values greedily according to this heuristic order (e.g., computed efficiently using a bucket sort). For each entry q, a search for an offset value Φ[q] is made such that the data entries h1−1(q) do not collide with any data previously assigned in the hash table 404, e.g.,
∀pεh1−1(q), H[h0(p)+Φ[q]]=undef.
The space of 8-bit-quantized offset values is min(
Exemplary Optimization of Hash Coherence
Because M1 is assigned to be the identity matrix in the example, accesses to the offset table Φ are coherent. How hash construction is modified to increase coherence of access to H is described next.
First, one considers the case that hash queries are constrained to the set of defined entries S⊂U. Let NS(p1,p2) be 1 if two defined points p1,p2εS are spatially adjacent in the domain (i.e. ∥p1−p2∥=1), or 0 otherwise. And, let NH(s1,s2) be similarly defined for slots in tables H. One seeks to maximize
It is this latter expression that is measured during construction. When assigning an offset value Φ[q], rather than selecting any value that is valid, one seeks one that maximizes coherence. Specifically, one examines the slots of H into which the points h1−1(q) map, and counts how many neighbors in H are also neighbors in the spatial domain:
The following heuristic candidates are considered:
As a postprocess, any undefined offset entries (i.e. for which h1−1(q)=Ø) are assigned values coherent with their neighbors. Table 2 (below) shows obtained values for the normalized coherence metric
The hash table generally stores data associated with a sparse subset of the domain. Depending on the application, it may be necessary to determine if an arbitrary query point lies in this defined subset.
Exemplary Constrained Access
Some scenarios such as 3D-parameterized surface textures guarantee that only the defined subset of the domain will ever be accessed.
Exemplary Domain Bit
For scenarios involving unconstrained access, one approach is to store a binary image over the domain, where each pixel (bit) indicates the presence of data (or blocks of data) in the hashed texture. One benefit is that a dynamic branch can be performed in the shader based on the stored bit, to completely bypass the hash function evaluations (h0,h1) and texture reads (Φ,H) on the undefined pixels.
Since current graphics hardware lacks support for single-bit textures, one approach involves packing each 4×2 block of domain bits into one pixel of an 8-bit luminance image. To dereference the bit, we perform a lookup in a 4×2×256 texture.
If a non-sparse image is already defined over the domain, another exemplary strategy is to hide the domain bit within this image, such as in the least-significant bit of a color or alpha channel. Such a hidden bit is convenient to indicate the presence of sparse supplemental data beyond that in the normal image.
Exemplary Position Tag
When the data is very sparse, storing even a single bit per domain point may take significant memory. Another approach is to let each slot of the hash table include a tag identifying the domain position {circumflex over (p)} of the stored data. Given a query point, one can then simply compare it with the stored tag.
Encoding a position {circumflex over (p)}εU requires a minimum of log2u bits. The position tags can be stored in an image with d channels of 16 bits, thus allowing a domain grid resolution of ū=64K. Such position tags are more concise than a domain bit image if d·16·m<1·u or equivalently if the data density n/u<1/(16d).
Exemplary Parameterized Position Hash
The set h−1(s)⊂U of domain points mapping to a slot s of the hash table has average size u/n. One goal is to encode which one is the defined point {circumflex over (p)}εS. One solution is to store in each slot a tuple (k,hk({circumflex over (p)})), where the integer kε{1, . . . , K} locally selects a parameterized hash function hk(p) such that the defined point {circumflex over (p)} has a hash value hk({circumflex over (p)})ε{1, . . . , R} different from that of all other all domain points mapping to that slot. More precisely,
∀pεh−1(s)\{circumflex over (p)}, hk(p)≠hk({circumflex over (p)}) (1)
The assignment of tuples (k,hk({circumflex over (p)})) can proceed after hash construction as follows. One first assigns k=1 and computes hk({circumflex over (p)}) at all slots. For the few slots without a defined point {circumflex over (p)}, one can assign hk({circumflex over (p)})=1. One can then sweep through the full domain U to find the undefined points whose parameterized hash values (under k=1) conflict with hk({circumflex over (p)}), and mark those slots. One can then make a second sweep through the domain to accumulate the sets h−1(s) for those slots with conflicts. Finally, for each such slot, one can try all values of k to satisfy (1).
Unlike the exemplary perfect hash function h, one can let hk(p) be as random as possible since it is not used to dereference memory. For fast evaluation, one can use hk=frac(dot(p, rsqrt(p+k*cl))), set K=R=256, and store (k,hk({circumflex over (p)})) as a 2-channel 8-bit image.
Exemplary Hybrid Scheme
Another approach is to combine the domain bit and position hash techniques. One can store a domain bit image at coarse resolution to efficiently discard blocks of undefined pixels. When this test fails, one can access the table and use the stored position hash to determine if the queried entry is defined.
Sparse data Dip) can represent either constant attributes over discrete grid cells (e.g., sprite pointers, line coefficients, or voxel occupancy) or samples of a smooth underlying function (e.g., color or transparency). In this second case, one seeks to evaluate a continuous reconstruction filter.
Exemplary Filtering With Blocking
One approach is to enable native hardware bilinear/trilinear filtering by grouping pixels into blocks. An original domain of size w=
The disclosed technology replaces the indirection table with a hash function, which needs only ˜4 bits per defined block. In addition, for the case of 2D unconstrained access, one must encode the defined blocks using either a domain bit or position hash, for a total of 4n+u or 20n bits respectively.
Table 1 shows quantitative comparisons of indirection tables and domain-bit hashes for various block sizes. As can be seen from the table, the hash offset table is much more compact than the indirection table and therefore encourages smaller block sizes.
The following is exemplary HLSL pseudocode for block-based spatial hashing:
Exemplary Filtering Without Blocking
To remove the overhead of sample replication, another exemplary approach performs explicit (e.g., non-native) filtering on an unblocked representation. The shader retrieves the nearest 2d samples from the hashed data and blends them.
As shown in the first row of Table 1, in 3D where access is constrained, memory use is reduced by an impressive factor of 3. The unblocked hashed texture is only 16% larger than the defined data values.
Exemplary Mipmapping
Defining a traditional mipmap pyramid over the packed data creates filtering artifacts even in the presence of blocking because the coarser mipmap levels incorrectly blend data across blocks. One solution is as follows. One computes a correct mipmap over the domain and arranges all mipmap levels into a flattened, broader domain using a simple function. Then, one constructs a spatial hash on this new flattened domain (either with or without blocking). At runtime, one determines the mipmap LOD using one texture lookup, performs separate hash queries on the two nearest mipmap levels, and blends the retrieved colors.
Native hardware mipmap filtering would be possible by assigning two mipmap levels to the packed texture. However, correct filtering would require allocating (b+3)d samples to each block (where b is odd) so it would incur a significant overhead. For instance, blocks of size b=5 in 2D would need (5+3)2+(3+1)2=80 samples rather than (5+1)2=36 samples.
Exemplary Images with Discontinuities
Several schemes embed discontinuities in an image by storing vector information at its pixels. These schemes allocate vector data at all pixels even though discontinuities are usually sparse. They reduce memory use through coarse quantization and somewhat intricate encodings. Spatial hashing offers a simple, compact solution useful in conjunction with any such scheme—whether vector data is implicit or parametric, linear or higher-order, and with or without comers.
To demonstrate the feasibility and performance of the hash approach, one can implement a representation of binary images with piecewise linear boundaries. For each square cell of the domain image, one can store two bits b1,b2. Bit b1 is the primary color of the cell, and bit b2 indicates if any boundary lines pass through the cell. If b2=1, the shader accesses a hashed texture to retrieve the coefficients ai,bi,ci, i=1,2 of two oriented lines passing through the cell, li(x,y)=aix+biy+c, where x,y are cell-local coordinates. The binary color at (x,y) is simply defined as
b1x or (l1(x,y)>0^l2(x,y)>0).
One can pack the 2 bits per pixel of the domain image as 2×2 blocks into individual pixels of an 8-bit image, and pack the hashed set of line coefficients as two RGB 8-bit images.
One can implement antialiasing. The complete shader, including hashing, takes 40 instructions. One benefit of is that dynamic branching on the domain bit b2 lets the shader run extremely quickly on pixels away from the boundaries. For those pixels near discontinuities, the shader makes a total of 5 texture reads: the packed domain bit, an unpacking decode table, the hash offset value, and two triples of line coefficients. The image in
Exemplary Texture Sprites
Sprites are generally high-resolution decals instanced over a domain using texture indirection. For example, they can be used to place character glyphs on a page. Spatial hashing can be used to compactly store such sprite maps.
Exemplary Alpha Channel Compression
In images with alpha masks, most alpha values are either 0 or 1, and only a small remaining subset is fractional. One may pack this sparse subset in a hashed texture, which is blocked to support native bilinear filtering.
In one example, an R5G5B5A1 image is used where the one-bit alpha channel is 1 to indicate full opacity, the color (0,0,0,0) is reserved for full transparency, or else the fractional alpha value is placed in the spatial hash. Storage for the alpha channel can be reduced from 8 to 1.7 bits per pixel (including the 1 bit alpha channel). Rendering rate is about 1170 frames/sec. The following is exemplary HLSL pseudocode:
For traditional R8G8B8 images, an alternative is to use a coarse 2-bit domain image at the same resolution as the hashed alpha blocks, yielding overall alpha storage of 0.92 bits per pixel and a rendering rate of 830 frames/sec.
Exemplary 3D-Parameterized Surface Texture.
Octree textures generally store surface color as sparse volumetric data parameterized by the intrinsic surface geometry. Such volumetric textures offer a simple solution for seamless texturing with nicely distributed spatial samples.
Perfect hashing provides an efficient packed representation. One can use a block-based hash for native trilinear filtering.
The disclosed technology involves storing data at a spatially uniform resolution. One could support adaptive resolution by introducing a mipmapped indirection table with sharing of blocks between levels as in. Spatial hashing could then be used to compress this mipmapped indirection table.
Exemplary 3D Painting
A 3D hashed texture is usually well-suited for interactive painting, because it is compact enough to uniformly sample a surface at high resolution yet efficient enough for real-time display and modification. One advantage over adaptive schemes such as octrees is that, just as in traditional 2D painting, one need not update any pointer-based structures during interaction.
Current graphics systems do not allow efficient rendering into a 3D texture. Thus, to enable fast modification of the hashed data on current systems, the disclosed technology extends its hash function to map 3D domains to 2D textures. This involves redefining M0,M1 as 2×3 matrices of the form
where C1,C2 are coprime with both
Position tags are stored along with the hashed data. Then, during painting, rasterization passes are performed over the 2D hashed texture. For each pixel, the shader compares the paintbrush position with the stored position tag and updates the hashed color appropriately. After painting is complete, the hashed 2D data could be transferred to a block-based 3D hash or to a conventional texture atlas.
Exemplary 3D Simulation
One can also let a finite-element simulation modify the surface data, again as a rasterization pass over the 2D hash table. Here the elements are voxels intersecting the surface. For each element one stores 2D pointers to the 3D-adjacent elements (some of which may be undefined).
Exemplary Surface Collision Detection
A spatial hash enables an efficient object-space framework for conservative collision detection—one can discretize two surfaces SA,SB into voxels and intersect these. Let Voxg(S) be the sparse voxels of size g that intersect surface S. Rather than directly computing Voxg(SA)∩Voxg(SB), the voxel centers of one surface are tested against a dilated version of the voxels from the other surface. That is, Voxg(SA+Se)∩Centers(Voxg(SB)) is computed where “+” denotes Minkowski sum, Se is a sphere of voxel circumradius e=g√{square root over (3)}/2, and Centers returns the voxel centers.
The sparse voxels Voxg(SA+Se) are stored as a blocked spatial hash, and the points Centers(Voxg(SB)) as a 2D image. At runtime, given a rigid motion of SB, a rasterization pass is applied over its stored voxel centers. The shader transforms each center and tests if it lies within a defined voxel of the spatial hash of SA. The intersecting voxels of SB provide a tight conservative approximation of the intersection curve. In addition to using a traditional occlusion query, one can render the intersecting voxels by letting the 2D image of SB be defined as a second (3D→2D) spatial hash.
Table 2 summarizes quantitative results for spatial hashing in various applications. The bits per entry required by the offset table assume 8-bit offset coordinates. Also included is the theoretical bit rate if one were allowed coordinates with ┌log2
H
The two preprocess times are for fast construction and for binary search optimization over the offset table size. (All other results assume optimized table sizes.) Runtime rendering performance is compared on perfect hashes constructed using (1) random matrices M0,M1 to simulate pseudorandom noncoherent hash functions, (2) identity matrices M0,M1 but no coherence optimization, and finally (3) coherence optimization. The optimization finds coherence as reflected in the metric
With reference to
The storage 740 may be removable or non-removable, and includes magnetic disks, magnetic tapes or cassettes, CD-ROMs, CD-RWs, DVDs, or any other medium which can be used to store information and which can be accessed within the computing environment 700. The storage 740 stores instructions for the software 780 implementing the described methods for perfect multidimensional spatial hashing.
The input device(s) 750 may be a touch input device, such as a keyboard, mouse, pen, or trackball, a voice input device, a scanning device, or another device, that provides input to the computing environment 700. For audio, the input device(s) 750 may be a sound card or similar device that accepts audio input in analog or digital form, or a CD-ROM reader that provides audio samples to the computing environment 700. The output device(s) 760 may be a display, printer, speaker, CD-writer, or another device that provides output from the computing environment 700.
The communication connection(s) 770 enable communication over a communication medium to another computing entity. The communication medium conveys information such as computer-executable instructions, compressed graphics information, or other data in a modulated data signal.
Computer-readable media are any available media that can be accessed within a computing environment 700. By way of example, and not limitation, with the computing environment 700, computer-readable media include memory 720, storage 740, and combinations of any of the above.
In view of the many possible embodiments to which the principles of the disclosed technology may be applied, it should be recognized that the illustrated embodiments are only preferred examples of the technology and should not be taken as limiting the scope of the following claims. We, therefore, claim as our invention all that comes within the scope and spirit of these claims.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6014733 | Bennett | Jan 2000 | A |
| 6513041 | Tarin | Jan 2003 | B2 |
| 6633297 | McCormack et al. | Oct 2003 | B2 |
| 20030128876 | Yamaguchi | Jul 2003 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20070244850 A1 | Oct 2007 | US |
