CN108984115A - Data parallel write-in, read method, apparatus and system - Google Patents
Data parallel write-in, read method, apparatus and system Download PDFInfo
- Publication number
- CN108984115A CN108984115A CN201810614178.3A CN201810614178A CN108984115A CN 108984115 A CN108984115 A CN 108984115A CN 201810614178 A CN201810614178 A CN 201810614178A CN 108984115 A CN108984115 A CN 108984115A
- Authority
- CN
- China
- Prior art keywords
- vector
- data
- written
- storage
- read
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
- G06F3/0601—Interfaces specially adapted for storage systems
- G06F3/0602—Interfaces specially adapted for storage systems specifically adapted to achieve a particular effect
- G06F3/062—Securing storage systems
- G06F3/0622—Securing storage systems in relation to access
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
- G06F3/0601—Interfaces specially adapted for storage systems
- G06F3/0602—Interfaces specially adapted for storage systems specifically adapted to achieve a particular effect
- G06F3/061—Improving I/O performance
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Complex Calculations (AREA)
Abstract
The present invention provides data parallel write-in, read method, apparatus and system, which comprises the write-in data directory for being written into data vector is transformed to one-dimensional writing address；According to the one-dimensional writing address and default write-in data amount check, the write-in data for obtaining the data vector to be written enable vector, the first storage index vector and the first storage address vector；According to the first storage index vector, said write data are enabled with vector, data vector to be written reorders described in the first storage address vector sum, the data vector to be written after enabling the first storage address vector described in vector sum according to the said write data after reordering and reordering is stored in the parallel storage.The present invention supports data to be written in parallel to from one or more dimensions, improves flexibility and the write efficiency of data write-in.
Description
Technical field
The invention belongs to field of data access technology, more particularly, to data parallel write-in, read method, device and
System.
Background technique
In recent years, artificial intelligence is widely used in every field.Intelligent algorithm be usually all data volume very
Big algorithm, therefore in order to accelerate the execution speed of intelligent algorithm, it not only needs to optimize arithmetic system, it is also necessary to which optimization is deposited
Storage system.
GPU is the hardware platform for the intelligent algorithm being widely used, and the storage of GPU has specifically for matrix meter
The multi-level buffer structure of calculation carrys out optimizing memory system.In Embedded Application field, due to the constraint of power consumption etc., usually using customization
The programmable chip of change rather than GPU realize intelligent algorithm.Having one kind in these embedded chips is that vector calculates
Machine core piece is well suited for carrying out intelligent algorithm acceleration, and this kind of computer chip is usually using vector memory as storage
System.
But for GPU, size, power consumption are all very big.Therefore, Embedded Application field use by very big
Limitation.Intelligent algorithm needs to carry out a large amount of matrix operations, therefore the data handled are often the data block of multidimensional, including
One peacekeeping multidimensional.Different algorithms needs to carry out parallel continuous read-write to data from one or more dimensions.And vector memory
The vector data of a specific length can only be fixedly accessed every time, and the flexibility ratio of data access is insufficient, and it is more to be unable to satisfy complexity
Demand of the intelligent algorithm of change to data access.
Summary of the invention
To overcome above-mentioned existing data access arrangement size and power consumption big, and the inflexible problem of data access or extremely
It partially solves the above problems, the present invention provides a kind of data parallel write-in, read method, apparatus and system.
According to the first aspect of the invention, a kind of data parallel wiring method is provided, comprising:
The write-in data directory for being written into data vector is transformed to one-dimensional writing address；Wherein, the data to be written
Vector is an one or more dimensions vector in multidimensional data matrix to be written, and said write data directory is described to be written
Index of first element to be written in the multidimensional data matrix to be written is to be written in all elements of data vector
Multidimensional data matrix in index；
According to the one-dimensional writing address and default write-in data amount check, the write-in number of the data vector to be written is obtained
According to enabled vector, the first storage index vector and the first storage address vector；Wherein, said write data enable every in vector
A element is for indicating whether the element of corresponding position in the data vector to be written is written；The first storage index vector
The vector being made of the index of each storing sub-units in the corresponding parallel storage of each element in the data vector to be written；
The first storage address vector is the address institute in the data vector to be written in the corresponding each storing sub-units of each element
The vector of composition；
According to the first storage index vector, vector, the first storage address vector are enabled to said write data
Reorder with the data vector to be written, according to the said write data after reordering enable vector sum described in first deposit
The data vector to be written after storage address vector will reorder is stored in the parallel storage.
A kind of data parallel read method is provided according to a second aspect of the present invention, comprising:
The reading data directory of data vector to be read is transformed to one-dimensional reading address；Wherein, the data to be read
Vector is an one or more dimensions vector in multidimensional data matrix to be read；The reading data directory is described to be read
Index of first element to be read in the multidimensional data matrix to be read in all elements of data vector；
According to the one-dimensional reading address and default reading data amount check, the reading number of the data vector to be read is obtained
According to enabled vector, the second storage index vector and the second storage address vector；Wherein, the reading data enable every in vector
A element is for indicating whether the element of corresponding position in the data vector to be read reads；The second storage index vector
By each element vector that the index of each storing sub-units is constituted in parallel storage in the data vector to be read；It is described
Second storage address vector is made of address of each element in each storing sub-units in the data vector to be read
Vector；
According to the second storage index vector, the second storage address vector described in vector sum is enabled to the reading data
Reorder, according to the reading data after reordering enable vector sum described in the second storage address vector from described parallel
Storing data vector is read in memory, and the storing data vector is reset according to the second storage index vector
Sequence obtains the data vector to be read.
A kind of data parallel writing station is provided according to a third aspect of the present invention, comprising:
First conversion module, the write-in data directory for being written into data vector are transformed to one-dimensional writing address；Its
In, the data vector to be written is an one or more dimensions vector in multidimensional data matrix to be written, said write number
According to first element to be written in all elements that index is the data vector to be written in the multidimensional data to be written
The index indexed in multidimensional data matrix to be written in matrix；
First obtains module, for obtaining described to be written according to the one-dimensional writing address and default write-in data amount check
The write-in data for entering data vector enable vector, the first storage index vector and the first storage address vector；Wherein, said write
Data enable each element in vector for indicating whether the element of corresponding position in the data vector to be written is written；Institute
The first storage index vector is stated as each storing sub-units in the corresponding parallel storage of each element in the data vector to be written
The vector that is constituted of index；The first storage address vector, which is that each element is corresponding in the data vector to be written, respectively to be deposited
The vector that address in storage subelement is constituted；
It is stored in module, for enabling vector, described first to said write data according to the first storage index vector
Data vector to be written described in storage address vector sum reorders, and enables vector according to the said write data after reordering
The data vector to be written after reordering with the first storage address vector is stored in the parallel storage.
A kind of data parallel reading device is provided according to a fourth aspect of the present invention, comprising:
Second conversion module, for the reading data directory of data vector to be read to be transformed to one-dimensional reading address；Its
In, the data vector to be read is an one or more dimensions vector in multidimensional data matrix to be read, the reading number
According to first element to be read in all elements that index is the data vector to be read in the multidimensional data to be read
Index in matrix；
Second obtains module, for reading data amount check with default according to the one-dimensional reading address, continues described in acquisition
The reading data of data vector are taken to enable vector, the second storage index vector and the second storage address vector；Wherein, the reading
Data enable each element in vector for indicating whether the element of corresponding position in the data vector to be read reads；Institute
The second storage index vector is stated as each storing sub-units in the corresponding parallel storage of each element in the data vector to be read
The vector that is constituted of index；The second storage address vector, which is that each element is corresponding in the data vector to be read, respectively to be deposited
The vector that address in storage subelement is constituted；
Read module, for according to the second storage index vector, to described in the enabled vector sum of the readings data the
Two storage address vectors reorder, and enable the second storage address described in vector sum according to the reading data after reordering
Vector from the parallel storage read storing data vector, according to it is described second storage index vector to the storing data to
Amount reorders, and obtains the data vector to be read.
A kind of data parallel read-write system is provided according to a fifth aspect of the present invention, comprising:
Parallel storage and above-mentioned data parallel writing station and above-mentioned data parallel reading device.
The present invention provides a kind of data parallel write-in, read method, apparatus and system, and this method is by being written into data
The write-in data directory of vector is transformed to one-dimensional writing address, according to one-dimensional writing address and default write-in data amount check, obtains
The write-in data of data vector to be written enable vector, the first storage index vector and the first storage address vector, according to described
First storage index vector enables data to be written described in vector, the first storage address vector sum to said write data
Vector reorders, and enables the first storage address vector described in vector sum for rearrangement according to the said write data after reordering
The data vector to be written after sequence is stored in the parallel storage, so that data be supported to carry out simultaneously from one or more dimensions
Row write enters, and improves flexibility and the write efficiency of data write-in.
Detailed description of the invention
Fig. 1 is data parallel wiring method overall flow schematic diagram provided in an embodiment of the present invention；
Fig. 2 is to tie up to four-dimensional data vector to be written at two in data parallel wiring method provided in an embodiment of the present invention
Spend the schematic diagram being written in parallel to；
Fig. 3 is to tie up to four-dimensional data vector to be written at one in data parallel wiring method provided in an embodiment of the present invention
Spend the schematic diagram being written in parallel to；
Fig. 4 is that write-in data directory is transformed to one-dimensional write-in in data parallel wiring method provided in an embodiment of the present invention
The schematic diagram of address；
Fig. 5 is that write-in is reordered schematic network structure in data parallel wiring method provided in an embodiment of the present invention；
Fig. 6 is data parallel read method overall flow schematic diagram provided in an embodiment of the present invention；
Fig. 7 is data parallel writing station overall structure diagram provided in an embodiment of the present invention；
Fig. 8 is data parallel reading device overall structure diagram provided in an embodiment of the present invention.
Specific embodiment
With reference to the accompanying drawings and examples, specific embodiments of the present invention will be described in further detail.Implement below
Example is not intended to limit the scope of the invention for illustrating the present invention.
A kind of data parallel wiring method is provided in one embodiment of the invention, and Fig. 1 provides for the embodiment of the present invention
Data parallel wiring method overall flow schematic diagram, this method comprises: S101, is written into the write-in data rope of data vector
Draw and is transformed to one-dimensional writing address；Wherein, the data vector to be written is one one in multidimensional data matrix to be written
Dimension or multi-C vector, write-in data directory are first element to be written in all elements of data vector to be written to be written
Multidimensional data matrix in index multidimensional data matrix to be written in index；
Wherein, data vector to be written is the data vector being written in parallel to.It is to be written that data directory, which is written,
Index of first element to be written in multidimensional data matrix to be written in data vector.Wherein element to be written is to need
The element being written.Write-in data directory is become under control under the control of control signal w_ctrl0 with w_ctrl1
It changes, generates an one-dimensional writing address.For example, data vector to be written is four dimensional vectors [dim3, dim2, dim1, dim0],
In, dim3, dim2, dim1 and dim0 indicate different dimensions.Size of the data vector to be written in each dimension be respectively DIM3,
DIM2, DIM1 and DIM0.Maximum be written in parallel to data amount check be S, data vector W_DATA to be written be a length be N to
Amount.It is continuously written into from two dimensions of dim1 and dim0, as shown in Figure 2.When being continuously written into parallel from two dimensions, need
Define two parameters, i.e. dim1 dimension maximum concurrent reading and concurrent writing number K and dim0 dimension maximum concurrent reading and concurrent writing number L, K*L=S.
The form that is continuously written into parallel from mono- dimension of dim0 is as shown in figure 3, L=S, K=0 in this case.
S102 obtains the write-in data of data vector to be written according to one-dimensional writing address and default write-in data amount check
Enabled vector, the first storage index vector and the first storage address vector；Wherein, write-in data enable each element in vector
For indicating whether the element of corresponding position in data vector to be written is written；First storage index vector be data to be written to
The vector that the index of each storing sub-units is constituted in the corresponding parallel storage of each element in amount；First storage address vector is
The vector that address in data vector to be written in the corresponding each storing sub-units of each element is constituted；
Wherein, the number that write-in data amount check W_M is the element for needing to be written in data vector to be written is preset.According to one
Dimension writing address w_base and default write-in data amount check W_M are calculated, and the write-in data for obtaining data vector to be written make
It can vector, the first storage index vector and the first storage address vector.Wherein, write-in data enable vector, the first storage index
Vector sum the first storage address vector is respectively the vector that length is equal to N.It is 0 that data, which are written, and enable each element in vector W_BE
Or 1, for indicating whether the element of corresponding position in data vector W_DATA to be written is written, wherein 1 indicates write-in, 0 is indicated
It is not written into.First storage index vector W_BI be data vector W_DATA to be written in each element it is corresponding being deposited into and
The vector that the index of storing sub-units is constituted in line storage.First storage address vector W_BA is data vector W_ to be written
The vector that the address in each storing sub-units that DATA will be deposited into is constituted.
It is to be written to enable vector, the first storage address vector sum to write-in data according to the first storage index vector by S103
Data vector reorders, after being reordered according to enabled the first storage address of the vector sum vector of write-in data after reordering
Data vector to be written be stored in parallel storage.
Specifically, write-in data are enabled into vector, the first storage address vector sum data vector input write-in weight to be written
Sorting network, according to first storage index vector, to write-in data enable vector W_BE, the first storage address vector W_BA and to
Write-in data vector W_DATA reorders, and obtains the write-in data after reordering and enables vector W_BE_R, the first storage ground
Location vector W_BA_R and data vector W_DATA_R to be written.Wherein W_BA_R be each storing sub-units it is corresponding to be stored in it is to be written
Enter the address of each element in data vector, W_BE_R be each storing sub-units whether Shi Neng enabled vector, W_DATA_R is
Element in the corresponding data vector to be written to be stored in of each storing sub-units.
The present embodiment is transformed to one-dimensional writing address by being written into the write-in data directory of data vector, according to one-dimensional
Writing address and default write-in data amount check, the write-in data for obtaining data vector to be written enable vector, the first storage index
Vector sum the first storage address vector, according to first storage index vector, to write-in data enable vector, the first storage address to
Amount and data vector to be written reorder, and enable the first storage address of vector sum vector according to the write-in data after reordering
Data vector to be written after reordering is stored in parallel storage, so that it is parallel to support that data are carried out from one or more dimensions
Write-in improves flexibility and the write efficiency of data write-in.
On the basis of the above embodiments, step S101 is specifically included in the present embodiment: carrying out weight to write-in data directory
The corresponding index value of each default dimension being written in parallel in the write-in data directory to reorder is split as more by sequence respectively
A index value；
For example, as shown in figure 4, under the control of w_ctrl0 will write-in data directory [dim0, dim1, dim2, dim3] into
Rearrangement sequence, the write-in data directory [dimnp1, dimnp0, dimp1, dimp0] after being reordered.Wherein, dimp1 and
Dimp0 is two dimensions being written in parallel to, and dimnp1 and dimnp0 are two dimensions for not needing to be written in parallel to, to be written
Size on data vector dimnp1, dimnp0, dimp1 and dimp0 four dimensions be respectively DIMNP1, DIMNP0, DIMP1 and
DIMP0。
Defining two maximums for being written in parallel to dimension of dimp1 and dimp0 and being written in parallel to number is respectively K and L.K and L can be with
Performance according to parallel storage is configured.Dimp1 is split as two dimensions, i.e. dim1_p and dim1_b according to K.Wherein,
Dim1_p=dimp1%K, dim1_b=dimp1//K.Size of the data vector to be written on dimension dim1_p is DIM1_P
=K, the size on dimension dim1_b are DIM1_B=DIMP1//K.Dimp0 is split as two dimensions, i.e. dim0_p according to L
And dim0_b.Wherein, dim0_p=dimp0%L, dim0_b=dimp0//L.Data vector to be written is on dimension dim0_p
Size be DIM0_P=L, size on dimension dim0_b is DIM0_B=DIMP1//L.At this point, data vector to be written
6 DOF is split as by the four-dimension, the write-in data directory after fractionation be [dimnp1, dimnp0, dimp1_b, dimp1_p,
dimp0_b,dimp0_p].Size of the data vector to be written in each dimension be respectively DIMNP1, DIMNP0, DIMP1_B,
DIMP1_P, DIMP0_B and DIMP0_P.
It reorders again under the control of w_ctrl1 to the write-in data directory after fractionation, according to reordering again
Write-in data directory calculated, obtain one-dimensional writing address.
For example, by after fractionation write-in data directory [dimnp1, dimnp0, dimp1_b, dimp1_p, dimp0_b,
Dimp0_p] it is reordered again, the write-in data directory that is reordered again [dnp3, dnp2, dnp1, dnp0, dp1,
dp0].Wherein, dp1=dimp1_p, dp0=dimp0_p.Dnp3, dnp2, dnp1, dnp0 are then dimnp1, dimnp0,
What dimp1_b, dimp0_b reordered again.Data vector to be written after reordering again in corresponding dimension
Size is respectively DNP3, DNP2, DNP1, DNP0, DP1 and DP0, is by DIMNP1, DIMNP0, DIMP1_B, DIMP1_
P, DIMP0_B and DIMP0_P progress is similarly reordered again and is obtained.It is carried out according to the write-in data directory to reorder again
It calculates, obtains the formula of one-dimensional writing address w_base are as follows:
W_base=dp0+dp1*DP0+dnp0*DP0*DP1+dnp1*DP0*DP1*DNP0+dn p2*DP0*DP1*
DNP0*DNP1+dnp3*DP0*DP1*DNP0*DNP1*DNP1.The present embodiment is not limited to the dimension of data vector to be written, carries out
The dimension and the dimension split into that data are split, dimension when being also not necessarily limited to be sorted again without reordering again.
On the basis of the above embodiments, it is obtained in the present embodiment according to one-dimensional writing address and default write-in data amount check
The step of taking the write-in data of data vector to be written to enable vector, the first storage index vector and the first storage address vector tool
Body includes: to determine that write-in data enable the number for the element that vector intermediate value is 1, according to be written according to default write-in data amount check
Enter the difference between the length of data vector and default write-in data amount check, determines that write-in data enable the member that vector intermediate value is 0
The number of element；According to index of each element in data vector to be written in data vector to be written, one-dimensional writing address and to
The length of data vector is written, obtains the first storage index vector and the first storage address vector.
Specifically, according to the mathematical notation of the enabled vector W_BE of the available write-in data of W_M: W_BE=[W_M { 1 },
(N-W_M){0}].Preceding W_M element is written in data vector W_DATA i.e. to be written, and the other elements in W_DATA are not written into.
The calculation formula of first storage index vector W_BI are as follows: W_BI=(w_base+ [0,1,2 ..., N-1]) %N.First storage ground
Location vector W_BA calculation formula are as follows: W_BA=(w_base+ [0,1,2 ..., N-1]) //N.
It is enabled to write-in data according to the first storage index vector in the present embodiment on the basis of the various embodiments described above
The step of vector, the first storage address vector sum data vector to be written are reordered specifically includes: according to the first storage rope
The amount of guiding into obtains the write-in number that index of the corresponding element of index of each storing sub-units in data vector to be written is constituted
According to index vector；According to write-in data directory vector, vector, the first storage address vector sum number to be written are enabled to write-in data
It reorders according to vector.
Specifically, write-in data directory vector W_BI_R is vector corresponding with W_BI, indicates each in parallel storage
The index vector that index of the element that a storing sub-units will be written in data vector W_DATA to be written is constituted, the vector
It can be calculated by W_BI, formula are as follows: W_BI_R=(N-W_BI [0]+[0,1,2 ..., N-1]) %N.According to write-in number
Vector W_BE, the first storage address vector W_BA and data vector W_ to be written are enabled to write-in data according to index vector W_BI_R
DATA reorders.It is reset in sequence network in write-in, phase is independently carried out to W_BE, W_BA and W_DATA according to W_BI_R
Same reorders, and the structural schematic diagram that sequence network is reset in write-in is as shown in Figure 5.
For example, it is desired to which being written into data vector is that [dim3, dim2, dim1, dim0] is written by 4 storing sub-units structures
At parallel storage, i.e. N=4.Data vector to be written corresponds to DIM3=12, DIM2=10 in the size of each dimension,
DIM1=8, DIM0=6.Index value of the data directory in 4 dimensions is written to be interleaved under the control of w_ctrl0, exports
With the relationship of input are as follows: dimnp1=dim1, dimnp0=dim0, dimp1=dim3, dimp0=dim2.Data to be written to
Amount is DIMNP1=DIM1=8, DIMNP0=DIM0=6, DIMP1=DIM3=12, DIMP0=DIM2 in the size of each dimension
=10.In the two dimensions of dimp1 and dimp0, maximum concurrent reading and concurrent writing number is K=2, L=2 respectively.Foundation K and L respectively will
Dimp1 and dimp0 splits into two dimensions, obtains: dimp1_b=dimp1//K, dimp1_p=dimp1%K, dimp0_b=
Dimp0//L, dimp0_p=dimp0%L.The size in dimension that data vector to be written obtains after fractionation mutually should be
DIMP1_B=DIMP1//K=6, DIMP1_P=K=2, DIMP0_B=DIMP0//L=5, DIMP0_P=L=2.It will split
The dimp1_p in six dimension datas dimnp1, dimnp0, dimp1_b, dimp1_p, dimp0_b, dimp0_p obtained afterwards
With dimp0_p directly as the dp1 and dp0 in six final dimensions；Dimnp1, dimnp0, dimp1_b, dimp0_b will be
It is interleaved under the control of w_ctrl1, exports the relationship with input are as follows: dnp3=dimp1_b, dnp2=dimp0_b, dnp1=
Dimnp1, dnp0=dimnp0.Data vector to be written is DNP3=DIMP1_B=6, DNP2=in the size of each dimension
DIMP0_B=5, DNP1=DIMNP1=8, DNP0=DIMNP0=6, DP1=DIMP1_P=2, DP0=DIMP0_P=2.
When write-in data directory w_index=[0,2,0,0], preset data be written number W_M=4, data to be written to
When measuring [6,7,8,9] W_DATA=, data parallel writing process is as follows:
W_base is calculated according to write-in data directory w_index, it may be assumed that
W_base=dp0_p+dp1*DP0+dnp0*DP0*DP1+dnp1*DP0*DP1*DNP0+dnp2 * DP0*DP1*
DNP0*DNP1+dnp3*DP0*DP1*DNP0*DNP1*DNP1=2.
According to w_base and W_M, W_BE, W_BA and W_BI_R are calculated, it may be assumed that
[W_M { 1 }, (N-W_M) { 0 }]=[1,1,1,1] W_BE=；
W_BI=(w_base+ [0,1,2 ..., N-1]) %N=[2,3,0,1]；
W_BI_R=(N-W_BI [0]+[and 0,1,2 ..., N-1]) %N=[2,3,0,1]；
W_BA=(w_base+ [0,1,2 ..., N-1]) //N=[0,0,1,1]；
Sequence network is reset in W_BE, W_BA and W_DATA input write-in, under the control of W_BI_R, output are as follows:
W_BE_R=[1,1,1,1]；
W_BA_R=[1,1,0,0]；
W_DATA_R=[8,9,6,7].
Data vector to be written is written in parallel to according to W_BE_R, W_BA_R and W_BA_R.Due to W_BE_R vector
Middle each element is 1, therefore all storing sub-units are enabled, and W_BA_R gives the address of each storing sub-units, W_
DATA_R provides each storing sub-units data to be stored.
A kind of data parallel read method is provided in another embodiment of the present invention, and Fig. 6 mentions for the embodiment of the present invention
The data parallel read method overall flow schematic diagram of confession, this method comprises: S601, by the reading data of data vector to be read
Index is transformed to one-dimensional reading address；Wherein, the data vector to be read is one in multidimensional data matrix to be read
One or more dimensions vector reads first element to be read in all elements that data directory is data vector to be read and is continuing
The index in multidimensional data matrix taken；
Wherein, data vector to be read is the data vector for needing to be read parallel.It is to be read for reading data directory
Index of first element to be read in multidimensional data matrix to be read in data vector.Wherein element to be read is to need
The element being read out.Reading data directory is converted under the control of control signal r_ctrl0 and r_ctrl1, is generated
One one-dimensional reading address.One-dimensional reading address generating method is identical as one-dimensional writing address generation method.
S602 obtains the reading data of data vector to be read according to one-dimensional reading address and default reading data amount check
Enabled vector, the second storage index vector and the second storage address vector；Wherein, each element in the enabled vector of data is read
For indicating whether the element of corresponding position in data vector to be read reads；Second storage index vector be data to be read to
The each element vector that the index of each storing sub-units is constituted in parallel storage in amount；Second storage address vector is to continue
Take the vector that address of each element in each storing sub-units is constituted in data vector；
Wherein, the number for reading that data amount check is the element for needing to read in data vector to be read is preset.According to one-dimensional
It reads address R_base and the default data amount check R_M that reads is calculated, the reading data for obtaining data vector to be read are enabled
Vector, the second storage index vector and the second storage address vector.Wherein, read data enable vector, second storage index to
Amount and the second storage address vector are respectively the vector that length is equal to N.Read data enable vector R_BE in each element be 0 or
1, for indicating whether the element of corresponding position in data vector R_DATA to be read reads, wherein 1 indicates to read, 0 is indicated not
It reads.Second storage index vector R_BI is that each element stores son in parallel storage in data vector R_DATA to be read
The vector that the index of unit is constituted.Second storage address vector R_BA is to read each member in data vector R_DATA to be read
The vector that address of the element in each storing sub-units is constituted.
S603 carries out weight to enabled the second storage address of the vector sum vector of data is read according to the second storage index vector
Sequence enables vector sum the second storage address vector according to the reading data after reordering and reads storing data from parallel storage
Vector reorders to storing data vector according to the second storage index vector, obtains data vector to be read.
Specifically, data will be read and enable vector, the second storage address vector sum data vector input reading weight to be read
Sorting network enables vector R_BE and the second storage address vector R_BA to data are read according to the second storage index vector R_BI
It reorders, obtains the reading data after reordering and enable vector R_BE_R and the second storage address vector R_BA_R.Wherein
R_BA_R is the address of each element in the pre-stored data vector to be read to be read in each storing sub-units, and R_BE_R is
The enabled vector that whether each element reads in pre-stored data vector to be read in each storing sub-units.According to R_BE_R
Storing data vector R_DATA_R is read from parallel storage with R_BA_R.The sequence of each element is according to storage in R_DATA_R
What the sequence of subelement was arranged, but sequence identical as the value of element in data vector R_DATA to be read is different.According to
Two storage index vector R_BI reorder to R_DATA_R, obtain R_DATA.
The present embodiment is by being transformed to one-dimensional reading address for the reading data directory of data vector to be read, according to one-dimensional
Address and default reading data amount check are read, the write-in data for obtaining data vector to be read enable vector, the second storage index
Vector sum the second storage address vector enables the first storage address of vector sum to write-in data according to the second storage index vector
Vector reorders, and enables the second storage address of vector sum vector from parallel storage according to the reading data after reordering
Storing data vector is read, is reordered according to the second storage index vector to storing data vector, obtains data to be read
Vector improves the flexibility and write-in effect of reading data so that data be supported to be read parallel from one or more dimensions
Rate.
On the basis of the above embodiments, step S601 is specifically included in the present embodiment: by the reading of data vector to be read
The step of taking data directory to be transformed to one-dimensional reading address specifically includes: reordering, will reorder to data directory is read
Reading data directory in the corresponding index value of each default dimension that is read parallel be split as multiple index values respectively；To tearing open
Reading data directory after point reorders again, is calculated according to the reading data directory to reorder again, obtains one
Dimension reads address.
For example, reorder data directory [dim0, dim1, dim2, dim3] is read under the control of r_ctrl0,
Reading data directory [dimnp1, dimnp0, dimp1, dimp0] after being reordered.Wherein, dimp1 and dimp0 is parallel
Two dimensions read, dimnp1 and dimnp0 are two dimensions for not needing to be read parallel, data vector to be read
Size on dimnp1, dimnp0, dimp1 and dimp0 four dimensions is respectively DIMNP1, DIMNP0, DIMP1 and DIMP0.
Defining dimp1 and the dimp0 two parallel parallel numbers that read of the maximum for reading dimension is respectively K and L.K and L can be with
Performance according to parallel storage is configured.Dimp1 is split as two dimensions, i.e. dim1_p and dim1_b according to K.Wherein,
Dim1_p=dimp1%K, dim1_b=dimp1//K.Size of the data vector to be read on dimension dim1_p is DIM1_P
=K, the size on dimension dim1_b are DIM1_B=DIMP1//K.Dimp0 is split as two dimensions, i.e. dim0_p according to L
And dim0_b.Wherein, dim0_p=dimp0%L, dim0_b=dimp0//L.Data vector to be read is on dimension dim0_p
Size be DIM0_P=L, size on dimension dim0_b is DIM0_B=DIMP1//L.At this point, data vector to be read
6 DOF is split as by the four-dimension, the reading data directory after fractionation be [dimnp1, dimnp0, dimp1_b, dimp1_p,
dimp0_b,dimp0_p].Size of the data vector to be read in each dimension be respectively DIMNP1, DIMNP0, DIMP1_B,
DIMP1_P, DIMP0_B and DIMP0_P.
It reorders to the reading data directory after fractionation, is carried out according to the reading data directory to reorder again again
It calculates, obtains one-dimensional reading address.
For example, by after fractionation reading data directory [dimnp1, dimnp0, dimp1_b, dimp1_p, dimp0_b,
Dimp0_p] it is reordered again under the control of r_ctrl1, the reading data directory that is reordered again [dnp3,
dnp2,dnp1,dnp0,dp1,dp0].Wherein, dp1=dimp1_p, dp0=dimp0_p.Dnp3, dnp2, dnp1, dnp0 are then
It is dimnp1, dimnp0, dimp1_b, what dimp0_b reordered again.Data vector to be read after reordering again
The size in corresponding dimension be respectively DNP3, DNP2, DNP1, DNP0, DP1 and DP0, be by DIMNP1,
DIMNP0, DIMP1_B, DIMP1_P, DIMP0_B and DIMP0_P progress are similarly reordered again and are obtained.According to weighing again
The reading data directory of sequence is calculated, and the one-dimensional formula for reading address w_base is obtained are as follows:
W_base=dp0+dp1*DP0+dnp0*DP0*DP1+dnp1*DP0*DP1*DNP0+dn p2*DP0*DP1*
DNP0*DNP1+dnp3*DP0*DP1*DNP0*DNP1*DNP1.The present embodiment is not limited to the dimension of data vector to be read, carries out
The dimension and the dimension split into that data are split, dimension when being also not necessarily limited to be sorted again without reordering again.
On the basis of the above embodiments, it is obtained in the present embodiment according to one-dimensional reading address and default reading data amount check
The step of taking the reading data of data vector to be read to enable vector, the second storage index vector and the second storage address vector tool
Body includes:
It according to default reading data amount check, determines and reads the number that data enable the element that vector intermediate value is 1, according to continuing
The length and the default difference read between data amount check for taking data vector determine that reading data enables the member that vector intermediate value is 0
The number of element；According to index of each element in data vector to be read in data vector to be read, one-dimensional reading address and to
The length of data vector is read, the second storage index vector and the second storage address vector are obtained.
Specifically, according to the available mathematical notation for reading data and enabling vector R_BE of R_M: R_BE=[R_M { 1 },
(N-R_M){0}].Preceding R_M element is written in data vector R_DATA i.e. to be read, and the other elements in R_DATA are not read.
The calculation formula of second storage index vector R_BI are as follows: R_BI=(r_base+ [0,1,2 ..., N-1]) %N.Second storage ground
Location vector R_BA calculation formula are as follows: R_BA=(r_base+ [0,1,2 ..., N-1]) //N.
It is enabled to data are read according to the second storage index vector in the present embodiment on the basis of the various embodiments described above
The step of vector sum the second storage address vector is reordered specifically includes: according to the second storage index vector, acquisition is respectively deposited
The reading data directory vector that index of the corresponding element of index of storage subelement in data vector to be read is constituted；According to
Data directory vector is read, is reordered to enabled the second storage address of the vector sum vector of data is read.
Specifically, reading data directory vector R_BI_R is vector corresponding with R_BI, indicates each in parallel storage
The index vector that index of the element that a storing sub-units will be read in data vector R_DATA to be read is constituted, the vector
It can be calculated by R_BI, formula are as follows: R_BI_R=(N-R_BI [0]+[0,1,2 ..., N-1]) %N.According to reading number
Vector R_BE and the second storage address vector R_BA is enabled to write-in data according to index vector R_BI_R to reorder.It is reading
Reset in sequence network, according to R_BI_R R_BE and R_BA is independently carried out it is identical reorder, read and reset sequence network
Structure is identical with the structure of sorting network in write-in.Storing data vector is read from parallel storage according to R_BE_R and R_BA_R
R_DATA_R.The sequence of each element is arranged according to the sequence of storing sub-units in R_DATA_R, with data to be read
The value of element is identical in vector R_DATA but sequence is different.It is reordered according to R_BI to R_DATA_R, obtains R_DATA.
For example, by the data vector W_DATA=to be written being written in parallel storage cited in above-described embodiment
[6,7,8,9] it is read out.I.e. [6,7,8,9] data vector R_DATA=to be read when, read data directory r_dex=[0,
2,0,0], preset and read data amount check R_M=4, data parallel reading process is as follows:
Index r_index calculates r_base according to read data, it may be assumed that
R_base=dp0_p+dp1*DP0+dnp0*DP0*DP1+dnp1*DP0*DP1*DNP0+dnp2 * DP0*DP1*
DNP0*DNP1+dnp3*DP0*DP1*DNP0*DNP1*DNP1=2.
According to r_base and R_M, R_BE, R_BA and R_BI_R are calculated, it may be assumed that
[R_M { 1 }, (N-R_M) { 0 }]=[1,1,1,1] R_BE=；
R_BI=(r_base+ [0,1,2 ..., N-1]) %N=[2,3,0,1]；
R_BI_R=(N-R_BI [0]+[and 0,1,2 ..., N-1]) %N=[2,3,0,1]；
R_BA=(r_base+ [0,1,2 ..., N-1]) //N=[0,0,1,1]；
R_BE and r_BA input, which is read, resets sequence network, under the control of R_BI_R, output are as follows:
R_BE_R=[1,1,1,1]；
R_BA_R=[1,1,0,0]；
Storing data vector R_DATA_R is read from parallel storage according to R_BE_R and R_BA_R and W_BA_R.Due to
Each element is 1 in R_BE_R vector, therefore all storing sub-units are enabled, and R_BA_R gives each storing sub-units
Address, the data of reading are R_DATA_R=[8,9,6,7].R_DATA_R is identical as the value of element in R_DATA, but sequence is not
Together, it reorders.It reorders under the control of R_BI to R_DATA_R, obtains R_DATA.
A kind of data parallel writing station is provided in another embodiment of the present invention, with reference to Fig. 7.The device is for real
Existing above-mentioned each data parallel wiring method embodiment.Therefore, retouching in the data parallel wiring method in foregoing embodiments
It states and defines, can be used for the understanding of each execution module in the embodiment of the present invention.
Data parallel writing station includes: the write-in data rope that the first conversion module 701 is used to be written into data vector
Draw and is transformed to one-dimensional writing address；Wherein, the data vector to be written is one one in multidimensional data matrix to be written
Dimension or multi-C vector, write-in data directory are first element to be written in all elements of data vector to be written to be written
Multidimensional data matrix in index；First, which obtains module 702, is used for according to one-dimensional writing address and default write-in data amount check,
The write-in data for obtaining data vector to be written enable vector, the first storage index vector and the first storage address vector；Wherein,
Write-in data enable each element in vector for indicating whether the element of corresponding position in data vector to be written is written；The
One storage index vector in data vector to be written in the corresponding parallel storage of each element each storing sub-units index institute
The vector of composition；First storage address vector is the address in data vector to be written in the corresponding each storing sub-units of each element
The vector constituted；It is stored in module 703 to be used for according to the first storage index vector, vector, the first storage is enabled to write-in data
Address vector and data vector to be written reorder, and enable the storage of vector sum first ground according to the write-in data after reordering
Data vector to be written after location vector will reorder is stored in parallel storage.
On the basis of the above embodiments, the first conversion module is specifically used in the present embodiment: to write-in data directory into
Rearrangement sequence splits the corresponding index value of each default dimension being written in parallel in the write-in data directory to reorder respectively
For multiple index values；It reorders again to the write-in data directory after fractionation, according to the write-in data rope to reorder again
It introduces row to calculate, obtains one-dimensional writing address.
On the basis of the above embodiments, the first acquisition module is specifically used in the present embodiment: according to default write-in data
Number determines that write-in data enable the number for the element that vector intermediate value is 1, is write according to the length of data vector to be written with default
Enter the difference between data amount check, determines that write-in data enable the number for the element that vector intermediate value is 0；According to data to be written to
The length of index of each element in data vector to be written, one-dimensional writing address and data vector to be written in amount obtains the
One storage index vector and the first storage address vector.
On the basis of the various embodiments described above, writing module is specifically used in the present embodiment: according to first storage index to
Amount obtains the write-in data rope that index of the corresponding element of index of each storing sub-units in data vector to be written is constituted
The amount of guiding into；According to write-in data directory vector, to write-in data enable vector, the first storage address vector sum data to be written to
Amount reorders.
Write-in data directory by being written into data vector is transformed to one-dimensional writing address, according to one-dimensional writing address
With default write-in data amount check, the write-in data for obtaining data vector to be written enable vector, the first storage index vector and the
It is to be written to enable vector, the first storage address vector sum to write-in data according to the first storage index vector for one storage address vector
Enter data vector to reorder, enabling the first storage address of vector sum vector according to the write-in data after reordering will reorder
Data vector to be written deposit parallel storage afterwards mentions so that data be supported to be written in parallel to from one or more dimensions
The flexibility of high data write-ins and write efficiency.
A kind of data parallel reading device is provided in another embodiment of the present invention, with reference to Fig. 8.The device is for real
Existing above-mentioned each data parallel read method embodiment.Therefore, retouching in the data parallel read method in foregoing embodiments
It states and defines, can be used for the understanding of each execution module in the embodiment of the present invention.
Data parallel writing station includes: that the second conversion module 801 is used for the reading data rope of data vector to be read
Draw and is transformed to one-dimensional reading address；Wherein, the data vector to be read is one one in multidimensional data matrix to be read
Dimension or multi-C vector, reading data directory is first element to be read in all elements of data vector to be read to be read
Multidimensional data matrix in index；Second, which obtains module 802, is used for according to one-dimensional reading address and default reading data amount check,
The reading data for obtaining data vector to be read enable vector, the second storage index vector and the second storage address vector；Wherein,
Each element that data enable in vector is read to be used to indicate whether the element of corresponding position in data vector to be read reads；The
Two storage index vectors in data vector to be read in the corresponding parallel storage of each element each storing sub-units index institute
The vector of composition；Second storage address vector is the address in data vector to be read in the corresponding each storing sub-units of each element
The vector constituted；Read module 803 is used for according to the second storage index vector, enables the storage of vector sum second to data are read
Address vector reorders, and enables vector sum the second storage address vector from parallel memorizing according to the reading data after reordering
Device reads storing data vector, is reordered according to the second storage index vector to storing data vector, obtains access of continuing
According to vector.
On the basis of the above embodiments, the second conversion module is specifically used in the present embodiment: to read data directory into
Rearrangement sequence splits the corresponding index value of each default dimension read parallel in the reading data directory to reorder respectively
For multiple index values；It reorders again to the reading data directory after fractionation, according to the reading data rope to reorder again
It introduces row to calculate, obtains one-dimensional reading address.
On the basis of the above embodiments, the second acquisition module is specifically used in the present embodiment: according to default reading data
Number determines and reads the number that data enable the element that vector intermediate value is 1, is read according to the length of data vector to be read with default
The difference between data amount check is taken, determines and reads the number that data enable the element that vector intermediate value is 0；According to data to be read to
Index of each element in data vector to be read, the one-dimensional length for reading address and data vector to be read in amount obtain the
Two storage index vectors and the second storage address vector.
On the basis of the various embodiments described above, read module is specifically used in the present embodiment: according to second storage index to
Amount obtains the reading data rope that index of the corresponding element of index of each storing sub-units in data vector to be read is constituted
The amount of guiding into；Index vector according to read data reorders to enabled the second storage address of the vector sum vector of data is read.
The present embodiment is by being transformed to one-dimensional reading address for the reading data directory of data vector to be read, according to one-dimensional
Address and default reading data amount check are read, the write-in data for obtaining data vector to be read enable vector, the second storage index
Vector sum the second storage address vector enables the first storage address of vector sum to write-in data according to the second storage index vector
Vector reorders, and enables the second storage address of vector sum vector from parallel storage according to the reading data after reordering
Storing data vector is read, is reordered according to the second storage index vector to storing data vector, obtains data to be read
Vector improves the flexibility and write-in effect of reading data so that data be supported to be read parallel from one or more dimensions
Rate.
A kind of data parallel read-write system is provided in another embodiment of the present invention, the data parallel read-write system packet
Include parallel storage, any data in above-mentioned each data parallel writing station embodiment is written in parallel to device and above-mentioned each data
The parallel reading device of any data in parallel reading device embodiment.
Finally, the present processes are only preferable embodiment, it is not intended to limit the scope of the present invention.It is all
Within the spirit and principles in the present invention, any modification, equivalent replacement, improvement and so on should be included in protection of the invention
Within the scope of.
Claims (11)
1. a kind of data parallel wiring method characterized by comprising
The write-in data directory for being written into data vector is transformed to one-dimensional writing address；Wherein, the data vector to be written
For an one or more dimensions vector in multidimensional data matrix to be written, said write data directory is the data to be written
Index of first element to be written in the multidimensional data matrix to be written in all elements of vector；
According to the one-dimensional writing address and default write-in data amount check, the write-in data for obtaining the data vector to be written make
It can vector, the first storage index vector and the first storage address vector；Wherein, said write data enable each member in vector
Element is for indicating whether the element of corresponding position in the data vector to be written is written；The first storage index vector is institute
State the vector that the index of each storing sub-units in the corresponding parallel storage of each element in data vector to be written is constituted；It is described
First storage address vector is made of the address in the data vector to be written in the corresponding each storing sub-units of each element
Vector；
According to the first storage index vector, vector, the first storage address vector sum institute are enabled to said write data
It states data vector to be written to reorder, enables the first storage ground described in vector sum according to the said write data after reordering
The data vector to be written after location vector will reorder is stored in the parallel storage.
2. the method according to claim 1, wherein the write-in data directory for being written into data vector is transformed to
The step of one-dimensional writing address, specifically includes:
It reorders to said write data directory, it is each by being written in parallel in the said write data directory to reorder
The default corresponding index value of dimension is split as multiple index values respectively；
It reorders again to the said write data directory after fractionation, according to the said write data directory to reorder again
It is calculated, obtains the one-dimensional writing address.
3. the method according to claim 1, wherein according to the one-dimensional writing address and default write-in data
Number, obtain the data vector to be written write-in data enable vector, the first storage index vector and the first storage address to
The step of amount, specifically includes:
According to the default write-in data amount check, determine that said write data enable the number for the element that vector intermediate value is 1, according to
Difference between the length of the data vector to be written and the default write-in data amount check, determines that said write data are enabled
The number for the element that vector intermediate value is 0；
According to index, the one-dimensional write-in ground of each element in the data vector to be written in the data vector to be written
The length of location and the data vector to be written obtains the first storage index vector and the first storage address vector.
4. method according to claim 1 to 3, which is characterized in that according to the first storage index vector, to institute
State the step of write-in data enable vector, data vector to be written described in the first storage address vector sum is reordered tool
Body includes:
According to the first storage index vector, the corresponding element of index of each storing sub-units is obtained described to be written
The write-in data directory vector that index in data vector is constituted；
According to said write data directory vector, vector, the first storage address vector sum institute are enabled to said write data
Data vector to be written is stated to reorder.
5. a kind of data parallel read method characterized by comprising
The reading data directory of data vector to be read is transformed to one-dimensional reading address；Wherein, the data vector to be read
For an one or more dimensions vector in multidimensional data matrix to be read, the reading data directory is the data to be read
Index of first element to be read in the multidimensional data matrix to be read in all elements of vector；
According to the one-dimensional reading address and default reading data amount check, the reading data for obtaining the data vector to be read make
It can vector, the second storage index vector and the second storage address vector；Wherein, each member read in the enabled vector of data
Element is for indicating whether the element of corresponding position in the data vector to be read reads；The second storage index vector is institute
State each element vector that the index of each storing sub-units is constituted in parallel storage in data vector to be read；Described second
Storage address vector by address of each element in each storing sub-units in the data vector to be read constitute to
Amount；
According to the second storage index vector, the second storage address vector described in vector sum is enabled to the reading data and is carried out
It reorders, enables the second storage address vector described in vector sum from the parallel memorizing according to the reading data after reordering
Storing data vector is read in device, is reordered, is obtained to the storing data vector according to the second storage index vector
Take the data vector to be read.
6. according to the method described in claim 5, it is characterized in that, the reading data directory of data vector to be read is transformed to
The step of one-dimensional reading address, specifically includes:
It reorders to the reading data directory, it is each by being read parallel in the reading data directory to reorder
The default corresponding index value of dimension is split as multiple index values respectively；
It reorders again to the reading data directory after fractionation, according to the reading data directory to reorder again
It is calculated, obtains the one-dimensional reading address.
7. according to the method described in claim 5, it is characterized in that, according to the one-dimensional reading address and default reading data
Number, obtain the data vector to be read reading data enable vector, the second storage index vector and the second storage address to
The step of amount, specifically includes:
According to the default reading data amount check, determine that the reading data enable the number for the element that vector intermediate value is 1, according to
The length of the data vector to be read and the default difference read between data amount check, determine that the reading data are enabled
The number for the element that vector intermediate value is 0；
According to index, the one-dimensional reading ground of each element in the data vector to be read in the data vector to be read
The length of location and the data vector to be read obtains the second storage index vector and the second storage address vector.
8. according to any method of claim 5-7, which is characterized in that according to the second storage index vector, to institute
The step of the second storage address vector described in the enabled vector sum of reading data is reordered is stated to specifically include:
According to the second storage index vector, the corresponding element of index of each storing sub-units is obtained described to be read
The reading data directory vector that index in data vector is constituted；
According to the reading data directory vector, the second storage address vector described in vector sum is enabled to the reading data and is carried out
It reorders.
9. a kind of data parallel writing station characterized by comprising
First conversion module, the write-in data directory for being written into data vector are transformed to one-dimensional writing address；Wherein, institute
Stating data vector to be written is an one or more dimensions vector in multidimensional data matrix to be written, said write data directory
It is first element to be written in all elements of the data vector to be written in the multidimensional data matrix to be written
Index；
First obtains module, for obtaining the number to be written according to the one-dimensional writing address and default write-in data amount check
Vector, the first storage index vector and the first storage address vector are enabled according to the write-in data of vector；Wherein, said write data
Each element in enabled vector is for indicating whether the element of corresponding position in the data vector to be written is written；Described
One storage index vector is the rope of each storing sub-units in the corresponding parallel storage of each element in the data vector to be written
Draw constituted vector；The first storage address vector is corresponding each storage of each element in the data vector to be written
The vector that address in unit is constituted；
It is stored in module, for enabling vector, first storage to said write data according to the first storage index vector
Address vector and the data vector to be written reorder, and enable vector sum institute according to the said write data after reordering
The data vector to be written after stating the first storage address vector and reordering is stored in the parallel storage.
10. a kind of data parallel reading device characterized by comprising
Second conversion module, for the reading data directory of data vector to be read to be transformed to one-dimensional reading address；Wherein, institute
Stating data vector to be read is an one or more dimensions vector in multidimensional data matrix to be read, the reading data directory
It is first element to be read in all elements of the data vector to be read in the multidimensional data matrix to be read
Index；
Second obtains module, for reading data amount check with default according to the one-dimensional reading address, access of continuing described in acquisition
Vector, the second storage index vector and the second storage address vector are enabled according to the reading data of vector；Wherein, the reading data
Each element in enabled vector is for indicating whether the element of corresponding position in the data vector to be read reads；Described
Two storage index vectors are the rope of each storing sub-units in the corresponding parallel storage of each element in the data vector to be read
Draw constituted vector；The second storage address vector is corresponding each storage of each element in the data vector to be read
The vector that address in unit is constituted；
Read module, for being deposited to described in the enabled vector sum of the reading data second according to the second storage index vector
Storage address vector reorders, and enables the second storage address vector described in vector sum according to the reading data after reordering
From the parallel storage read storing data vector, according to it is described second storage index vector to the storing data vector into
Rearrangement sequence obtains the data vector to be read.
11. a kind of data parallel read-write system, which is characterized in that including parallel storage, and as described in claim 9 and 10
Device.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810614178.3A CN108984115B (en) | 2018-06-14 | 2018-06-14 | Data parallel writing and reading method, device and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810614178.3A CN108984115B (en) | 2018-06-14 | 2018-06-14 | Data parallel writing and reading method, device and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108984115A true CN108984115A (en) | 2018-12-11 |
| CN108984115B CN108984115B (en) | 2020-07-28 |
Family
ID=64540446
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810614178.3A Expired - Fee Related CN108984115B (en) | 2018-06-14 | 2018-06-14 | Data parallel writing and reading method, device and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108984115B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1971537A (en) * | 2005-11-25 | 2007-05-30 | 杭州中天微系统有限公司 | Access method of matrix data and storage device of the matrix data |
| CN101478608A (en) * | 2009-01-09 | 2009-07-08 | 南京联创科技股份有限公司 | Fast operating method for mass data based on two-dimensional hash |
| US20150234662A1 (en) * | 2014-02-19 | 2015-08-20 | Mediatek Inc. | Apparatus for mutual-transposition of scalar and vector data sets and related method |
| CN104978148A (en) * | 2014-04-09 | 2015-10-14 | 瑞萨电子(中国)有限公司 | Data writing method and device and data reading method and device |
-
2018
- 2018-06-14 CN CN201810614178.3A patent/CN108984115B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1971537A (en) * | 2005-11-25 | 2007-05-30 | 杭州中天微系统有限公司 | Access method of matrix data and storage device of the matrix data |
| CN101478608A (en) * | 2009-01-09 | 2009-07-08 | 南京联创科技股份有限公司 | Fast operating method for mass data based on two-dimensional hash |
| US20150234662A1 (en) * | 2014-02-19 | 2015-08-20 | Mediatek Inc. | Apparatus for mutual-transposition of scalar and vector data sets and related method |
| CN104978148A (en) * | 2014-04-09 | 2015-10-14 | 瑞萨电子(中国)有限公司 | Data writing method and device and data reading method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108984115B (en) | 2020-07-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11403025B2 (en) | Matrix transfer accelerator system and method | |
| JP7358382B2 (en) | Accelerators and systems for accelerating calculations | |
| CN110390384A (en) | A kind of configurable general convolutional neural networks accelerator | |
| CN107992943A (en) | Addressed for convolutional neural networks | |
| CN108171317A (en) | A kind of data-reusing convolutional neural networks accelerator based on SOC | |
| WO2020046859A1 (en) | Systems and methods for neural network convolutional layer matrix multiplication using cache memory | |
| CN103810111A (en) | Address Generation In An Active Memory Device | |
| CN103745008A (en) | Sorting method for big data indexing | |
| CN103345430B (en) | The fuzzy balance method in distributed storage pond and system | |
| CN104778025B (en) | The circuit structure of pushup storage based on random access storage device | |
| CN101655821B (en) | Method and apparatus for settling Hash address conflict when mapping address space | |
| CN109952559A (en) | The stream transmission engine replicated with independent optional element and in groups | |
| CN103677655B (en) | A kind of two-dimemsional number group data stream reading/writing method on a memory and device | |
| CN109472361A (en) | Neural network optimization | |
| CN103064749B (en) | Inter-process communication method | |
| CN104572519B (en) | A kind of multiport memory access controller and its control method for multiprocessor | |
| CN116414765B (en) | FPGA chip, transparent transmission method, logic test module and method | |
| CN108984115A (en) | Data parallel write-in, read method, apparatus and system | |
| CN104850391B (en) | Handle the devices and methods therefor of multiple data sets | |
| JPWO2005098612A1 (en) | Important component priority calculation method and device | |
| CN102411557B (en) | Multi-granularity parallel FFT (Fast Fourier Transform) computing device | |
| CN108701102A (en) | Direct memory access controller, method for reading data and method for writing data | |
| CN109800867A (en) | A kind of data calling method based on FPGA chip external memory | |
| CN107315565A (en) | It is a kind of to be used to generate the random vector apparatus and method obeyed and be necessarily distributed | |
| CN110018847A (en) | Configurable register and the data access method based on configurable register |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200728 Termination date: 20210614 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |