JP6227169B1 - Parallel processing apparatus and parallel processing method - Google Patents

Abstract

The data allocation unit (102a0) of the parent device (101a0) allocates a set of signal data and an imaging area to the child devices (101a1, 101a2). The master units (101b0, 101c0) in the slave units (101a1, 101a2) allocate the received signal data and imaging area pairs to the slave units (101b1-101b4, 101c1-101c4). The slave units (101b1 to 101b4, 101c1 to 101c4) independently perform image reproduction processing and transmit the processing results to the master units (101b0 and 101c0).

Description

  The present invention relates to a parallel processing apparatus and a parallel processing method capable of shortening the processing time for reconstructing observation results by, for example, a synthetic aperture radar (SAR) into image data by a Fast Factorized BackProjection method. Is.

  The SAR repeatedly radiates radio waves to an observation target along the traveling direction of an antenna platform such as a satellite or an aircraft, and collects reflected signals. As a result, if the sampling number of the reflected signal s is M and the number of reflected waves is N, signal data S represented by a complex matrix of N rows and M columns and a real number of N rows and 3 columns are obtained as observation data. Position information P represented by a matrix is obtained. From these observation data, for example, image data represented by a complex matrix of H rows and W columns is given to the image reproduction algorithm by providing imaging information represented by the image center position c, the number of pixels H × W, and the resolution. V can be reconstructed. In this image reproduction algorithm, there are various methods such as a BackProjection method that reproduces with high accuracy over a long time without approximation processing, and a RangeDoppler method that reproduces with low accuracy in a short time by approximation processing. The Fast Factorized BackProjection method is known as a method having a good balance between accuracy and processing speed.

When performing image reproduction using this Fast Factorized BackProjection method, first, a synthetic aperture length (full aperture) represented by connecting each position information of N points in time series order is recursively applied to a plurality of stages to perform sub-aperture division. To do. For example, if the division stage number K is set to 3 and the division numbers L ₁ ,..., L _K for each stage are set to 2, the position information P is obtained by dividing the N point into 8 parts as shown in the following formula (1). Sub-aperture position information P ^(sub) = [p ₁ , p ₂ ,..., P _{N / 8} ] ^T.

The signal data S is subjected to pulse compression processing, and similarly to the position information, eight sub-aperture signal data S ^(sub) = [s ₁ , S ₂ ,..., S _{N / 8} ] ^T.

Next, the calculation area of the image data V is divided into sub-images based on the imaging information. For example, if the image is divided so as to be divided into two equal parts, the four sub-image image regions V ^(sub) are obtained as calculation regions as represented by the following expression (3).

Here, pos is the operator to get a sampling position on the reflection signal corresponding to the pixel position, ref is corresponding to the signal S _n which is discretized in M points, the position pos given by a real number [...] It is an operator who acquires the sampling value to perform. When acquiring the real position, as represented by the following formula (5), each of the reflected signal S _n upsampling processing the sampling number again minced in a (signal S _n Fast Fourier Transform (FFT) After the conversion, the value 0 is set in the center area of the converted signal data (Padding) and the inverse fast Fourier transform (IFFT) is performed, and then implemented as an interpolation process (Interp) such as linear interpolation or spline interpolation. The

Finally, the sub-image image data D ^(1,1) corresponding to the same sub-image image region is synthesized for each recursive stage as shown in the following equation (6), and the final sub-image image data V ^(sub) is obtained.

  In this compositing process, every time the recursive stage progresses (K, K-1, K-2,..., 1), the sub-image images of different image grids become finer and the high resolution image is obtained. In order to perform composition, resampling processing “Resample” for the purpose of grid alignment and grid subdivision is performed. This resampling process is implemented as an interpolation process in which upsampling is performed for each dimension (vertical / horizontal) of the sub-image image data in the same manner as Expression (4). As described above, for example, the Fast Factorized BackProjection method disclosed in Non-Patent Document 1 reproduces high-resolution SAR image data from sub-aperture divided signal data.

  In general, SAR image reproduction processing takes a long time, and this Fast Factorized BackProjection method has a smaller amount of calculation in the SAR image reproduction algorithm, but still requires a long time. For this reason, for example, as disclosed in Non-Patent Document 2, parallel processing is performed in a plurality of processing units sharing a memory area, and image reproduction processing time is reduced.

Ulander, LMH, Hellsten, H., and Stenstrom, G .: Synthetic-Aperture Radar Processing Using Fast Factorized Back-Projection, IEEE Trans.Aerospace and Electronic Systems, Vol. 39.No. 3, pp. 760-776, 2003 . Yu, M., Zhang, X., and Liu Z.:Acceleration of fast factorized back projection algorithm for bistatic SAR, in Proc. 2013 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp.2493-2496,2013

However, since the conventional parallel processing device is configured to be performed within the range where the memory area is shared as described above, it can cope with parallel processing that handles a plurality of memory areas by including communication processing. There was a problem that I could not.
In addition, if data is transmitted from the master unit to the slave unit so that communication processing is simply included, and parallelization is performed so that the results calculated by the slave unit are aggregated on the master unit side, the image playback time is shortened. However, since the time required for data communication and the synthesizing process on the parent side is required, there is a problem that the time for the entire image reproduction process cannot be reduced according to the number of parallel processes.
Furthermore, since the upsampling process is performed using all sampling points of each signal data, there is a problem that the processing time increases nonlinearly according to the number of samplings. That is, the conventional parallel processing apparatus has a problem that image data cannot be obtained in a short time from large-scale observation data even if parallel processing is used.

  The present invention has been made to solve the above problems, and provides a parallel processing device and a parallel processing method capable of reducing the processing time for forming image data based on signal data and an imaging region. The purpose is to obtain.

The parallel processing device according to the present invention performs processing for recursively dividing signal data, which is an observation signal including position information observed by a synthetic aperture radar, to construct an image, and is configured in a parent-child relationship of three generations or more. A plurality of arithmetic units , the arithmetic units having the lower generation receive the signal data and the imaging region from the upper generation, the image data from the lower generation, The image data or the image data processed by itself is transmitted to the upper generation, and the arithmetic unit having no lower generation receives the pair of the signal data and the imaging area from the upper generation, and the parent or the same belonging to the same parent-child relationship A data transmission / reception unit that transmits / receives signal data or an imaging region or both to / from a child of the same generation belonging to a parent-child relationship, and a signal data received by the data transmission / reception unit provided in a calculation unit having a lower generation image From the signal data received by the data transmission / reception unit provided in the data allocation unit that divides the region and assigns the divided signal data and imaging region set to the lower generation and the arithmetic unit that does not have the lower generation An image reproduction unit that creates image data corresponding to an imaging area and a calculation unit having a lower generation combine the image data received by the data transmission / reception unit, and a calculation unit having no lower generation uses a data transmission / reception unit. And an image composition unit for synthesizing image data from at least one of the image data received by the image reproduction unit and the image data from the image reproduction unit.

The parallel processing device according to the present invention allocates a set of signal data and an imaging area from the parent device side to the child device side in three generations or more, exchanges and changes data between the child devices, and generates a plurality of child devices The slave units perform the image data composition, the slave units perform image reproduction processing independently of each other, and the image reproduction processing, image synthesis processing, and data communication are processed asynchronously. Thereby, the image reproduction can be completed in a short time by parallel processing including communication processing.

It is a block diagram which shows the parallel processing apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the sub image division | segmentation of the image data area | region in the parallel processing apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the example of the sub-aperture division | segmentation with respect to the observation signal in the parallel processing apparatus of Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the parallel processing apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the change of the data size by the conventional upsampling process and interpolation process. It is explanatory drawing which shows the division | segmentation of the signal data in the parallel processing apparatus of Embodiment 2 of this invention.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
FIG. 1 is a configuration diagram of a parallel processing apparatus according to the present embodiment.
The parallel processing apparatus according to the present embodiment includes a plurality of generation units of three generations configured in a parent-child relationship. That is, the parallel processing device is divided into one parent device 101a0 and two child devices 101a1 and 101a2, and further divided into a parent device 101b0 and four child devices 101b1 to 101b4 in the child device 101a1. Similarly, the child device 101a2 is divided into a parent device 101c0 and four child devices 101c1 to 101c4. The parent device 101a0, the parent devices 101b0 and 101c0, and the child devices 101b1 to 101b4 and 101c1 to 101c4 correspond to a plurality of arithmetic units.

  Base unit 101a0 includes data allocation unit 102a0, data transmission / reception unit 103a0, and image composition unit 104a0. Base units 101b0 and 101c0 include data allocation units 102b0 and 102c0, data transmission / reception units 103b0 and 103c0, and image composition units 104b0 and 104c0. I have. The slave units 101b1 to 101b4 and 101c1 to 101c4 include data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4, image reproduction units 102b1 to 102b4 and 102c1 to 102c4, and image composition units 104b1 to 104b4 and 104c1 to 104c4, respectively. ing.

Data transmission / reception unit 103a0 in base unit 101a0 receives signal data and an imaging area as input information. Further, the data transmission / reception unit 103a0 receives image data from the lower generation child devices 101a1 and 101a2, and transmits the image data or the image data processed by the parent device 101a0 as output information. The data allocation unit 102a0 divides the signal data received by the data transmission / reception unit 103a0 and the imaging area, and allocates the set of the divided signal data and the imaging area to the lower generation. The image composition unit 104a0 synthesizes the image data received by the data transmission / reception unit 103a0.
The data transmission / reception units 103b0 and 103c0 in the parent devices 101b0 and 101c0 receive the signal data and the imaging area from the parent device 101a0 which is the upper generation, and images from the child devices 101b1 to 101b4 and 101c1 to 101c4 which are the lower generations. Data is received, and this image data is transmitted to the upper generation 101a0. The data allocation units 102b0 and 102c0 divide the signal data received by the data transmission / reception units 103b0 and 103c0 and the imaging region, and set the set of the divided signal data and the imaging region as slave units 101b1 to 101b4 as lower generations. Assigned to 101c1 to 101c4. The image synthesis units 104b0 and 104c0 synthesize the image data received by the data transmission / reception units 103b0 and 103c0.

  The data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 in the slave units 101b1 to 101b4 and 101c1 to 101c4 receive a set of signal data and an imaging area from the parent units 101b0 and 101c0 which are upper generations. In addition, signal data or image area data, or both signal data and image area data are transmitted / received from the slaves 101b1 to 101b4 of the same generation. The image reproduction units 102b1 to 102b4 and 102c1 to 102c4 create image data corresponding to the imaging region from the signal data received by the data transmission / reception units 103b1 to 103b4. The image composition units 104b1 to 104b4 and 104c1 to 104c4 synthesize image data received by the data transmission / reception units 103b1 to 103b4 and image data from at least one of the image data from the image reproduction units 102b1 to 102b4 and 102c1 to 102c4. To do.

  In the first embodiment, for the sake of convenience of explanation, three generations of parent and child are set at a ratio of 1: 2: 4. However, any number of generations may be set as long as the parent-child relationship is two or more generations. . Further, the number of slave units may be set to an arbitrary number of 1 or more, and the slave units may not adopt an equal parent-child configuration.

  In FIG. 1, the data transmission / reception unit 103 a 0 of the parent device 101 a 0 receives information on the observation region P, the observation signal S, and the image area necessary for calculating the image data V as the calculation result as input data from the outside of the parallel processing device. To do. These input data are divided by the data allocation unit 102a0 and transferred to the child device 101a1 and the child device 101a2. As an explanation of the data division method of Fast Factorized BackProjection, FIG. 2 shows sub-image division of this image data area, and FIG. 3 shows an example of sub-aperture division for an observation signal.

  The data allocating unit 102a0 in FIG. 1 transmits the sub image image areas 202 to 205 and the position information necessary for imaging in order to each of the slave units 101a1 and 101a2 after performing the division setting described above, The final stage sub-aperture signal data 304aa to 304abb is sequentially transmitted to 101a1, and the final stage sub-aperture signal data 304baa to 304bbb is sequentially transmitted to the slave unit 101a2.

  The data transmitted from the parent device 101a0 is sequentially received by the data transmission / reception units 103b0 and 103c0 of the parent devices 101b0 and 101c0 in the child devices 101a1 and 101a2. Sub-aperture signal data 304aa-304abb, 304baa-304bbb received by master devices 101b0 and 101c0, position information necessary for imaging, and sub-image image areas 202-205 are divided into sub-aperture divided data and sub-image divided data. Are transmitted from the data allocation units 102b0 and 102c0 to the four slave units 101b1 to 101b4 and 101c1 to 101c4, respectively, and received by the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 of the respective slave units.

In this allocation method, if the images of the sub-aperture signal data 304aa to 304bbb and the sub-image image areas 202 to 205 are reproduced, the slave units 101b1 to 101b4 and 101c1 to 101c4 that perform image reproduction starting from the master unit 101a0. It does not matter how data is allocated to.
For example, the parent device 101a0 assigns all four sub-image image areas 202 to 205 to two child devices 101a1 and 101a2, and among these child devices 101a1 and 101a2, the parent devices 101b0 and 101c0 have four sub-image image regions. 202 to 205 are assigned to each of the four slave units 101b1 to 101b4 and 101c1 to 101c4, and all the sub-aperture signal data 304aa to 304bbb are sequentially transmitted to all four slave units 101b1 to 101b4 and 101c1 to 101c4. Assign to. In the slave units 101b1 to 101b4 and 101c1 to 101c4, the sub image images 305aa to 305abb and 305baa to 305bbb reproduced by the image reproducing units 102b1 to 102b4 and 102c1 to 102c4 are synthesized by the image synthesizing units 104b1 to 104b4 and 104c1 to 104c4, respectively. The composite images 307a and 307b are transmitted from the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 and received by the data transmission / reception units 103b0 and 103c0 of the master units 101b0 and 101c0. The sub-image image data 308 of 101b4, 101c1 to 101c4 may be integrated.

  Alternatively, four sub-image image areas 202 to 205 are assigned to the four slave units 101b1 to 101b4 and 101c1 to 101c4 one by one, and the four sub-aperture signal data 304aa to 304abb and 304baa to 304bbb are assigned to the four slave units 101b1. ˜101b4, 101c1 to 101c4 so that only one is assigned, and sub-aperture area data is exchanged between the slave units 101b1 to 101b4 and between the slave units 101c1 to 101c4 by the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4. All the sub image images 305aa to 305abb and 305baa to 305bbb are reproduced and synthesized by the image synthesis units 104b1 to 104b4 and 104c1 to 104c4, and the slave units 101b1 to 101b are configured. Between 4 may be configured to perform aggregation of sub-images data 308 between handset 101c1~101c4 the image synthesizing unit 104b0,104c0 of the master unit 101B0,101c0.

  Alternatively, the four sub image image areas 202 to 205 are all assigned to the four slave units 101b1 to 101b4 and 101c1 to 101c4, and the four sub aperture signal data 304aa to 304abb and 304baa to 304bbb are assigned to the four slave units 101b1 to 101b4. , 101c1 to 101c4 are transmitted in order, and four sub-image images 305aa to 305abb and 305baa to 305bbb are created on the side of the slave units 101b1 to 101b4 and 101c1 to 101c4, and between the slave units 101b1 to 101b4, Sub-aperture signal data 304aa-304abb, 304baa-304bbb are exchanged between 101c1-101c4, and a combination process is performed on the side of slave units 101b1-101b4, 101c1-101c4. Alternatively, the sub image images 305aa to 305abb and 305baa to 305bbb may be sequentially returned to the parent devices 101b0 and 101c0, and the image composition units 104b0 and 104c0 of the parent devices 101b0 and 101c0 may calculate the sub image image data 308. .

Alternatively, any combination of these methods may be used, for example, an arbitrary number of sub-aperture areas and sub-image image areas may be allocated to an arbitrary number of slave units, and composition processing may be performed on either the slave unit side, the master unit side, or both It may be configured as follows.
In addition, during image playback by the slave unit, synthesis processing by the master unit, data transmission / reception between the master unit and the slave unit, and data transmission / reception between the slave units are performed simultaneously and asynchronously for data in another area. You may comprise.

  When the parallel processing apparatus configured as described above is configured as a program, one or more computers are used. When two computers A and B are used, the computer A is in charge of the parent device 101a0, and the computers A and B are in charge of the child devices 101a1 and 101a2. Further, threads (or processes) for transmitting and receiving data are the parent devices 101b0 and 101c0 in the computers A and B, and one or more arithmetic units in the computers A and B are roles of the child devices 101b1 to 101b4 and 101c1 to 101c4. In charge.

  Inside the computer, for example, a master unit is set as a CPU (Central Processing Unit) and a slave unit is set as a processor such as a GPU (Graphics Processing Unit) or a MIC (Many Integrated Core), and images are processed by a number of arithmetic units provided in these processors. What is necessary is just to comprise so that reproduction | regeneration may be processed in parallel. In the case of this configuration, data communication is via the PCI Express bus.

  Alternatively, the master unit may be set to a CPU 1 thread and the slave unit may be set to a plurality of threads corresponding to the number of CPU cores, and the master unit and the slave unit may share the memory space on the same computer. In this case, data communication is not performed explicitly, but is performed as data input / output to / from the shared memory area. Further, when the parent machine and the child machine are different computers, communication is performed via a network. In addition, for example, an unevenly structured parent-child relationship such as a relationship between a child device having 10 grandchild devices and a child device having one grandchild device from one parent device has different performance. A plurality of GPUs and computers may be used at the same time.

  In addition to being installed as ground equipment, this parallel processing device may be configured as part of the hardware of the radar system and mounted on a mobile body such as a car, an aircraft, or a satellite.

FIG. 4 is a flowchart showing a parallel processing method using the parallel processing device.
The arithmetic units (child devices 101a1 and 101a2) given the roles of the parent device and the child device select received data that has not been processed by the data transmitting / receiving units 103b0 and 103c0 in the parent devices 101b0 and 101c0 (step ST1). If there is no received data, the process proceeds to confirmation of whether or not a process end notification has been received (step ST10). When the acquired received data is input data (sub-aperture signal data and sub-image image area), one or more unprocessed data is selected, and arithmetic units (child devices 101b1 to 101b4 and 101c1) having no child devices. 101c4), image data is calculated by image reproduction processing (step ST2) by the image reproduction units 102b1 to 102b4 and 102c1 to 102c4. Further, the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 of the slave units 101b1 to 101b4 and 101c1 to 101c4 determine whether or not to exchange input data and image data with other slave units (step ST3). ) After replacement (step ST4) if necessary, the image reproduction process of step ST2 is performed using the input data and image data after the replacement.

In step ST1, after the input data is selected, in the arithmetic units (master units 101b0 and 101c0) having the slave units, the data transmitting / receiving units 103b0 and 103c0 transmit the input data to the slave units 101b1 to 101b4 and 101c1 to 101c4 (step ST1). ST5) Waits for the return of image data from the slave units 101b1 to 101b4 and 101c1 to 101c4, and acquires the image data (step ST6). That is, the divided image data of the slave units are collected.
When the acquisition of the image data is completed, the data transmission / reception units 103b0 and 103c0 check whether there is unprocessed input data (step ST7). In step ST3, when input data is not exchanged with other slave units, the process proceeds to step ST7. If unprocessed input data exists in step ST7, the process returns to the data selection process in step ST1. If there is no unprocessed input data in step ST7, the image data reproduced by the image synthesis units 104b0 and 104c0 or the image synthesis units 104b1 to 104b4 and 104c1 to 104c4 is synthesized (step ST8). Is transmitted to the parent device 101a0 or 101b0, 101c0 (step ST9).

  On the other hand, in step ST1, the unprocessed received data selection result in the data transmission / reception units 103b0 and 103c0 of the parent devices 101b0 and 101c0 or the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 of the child devices 101b1 to 101b4 and 101c1 to 101c4 are not processed. If the received data selection result of the process is image data, the process proceeds to image composition processing in step ST8.

  After step ST9, the data transmission / reception units 103b0 and 103c0 of the parent devices 101b0 and 101c0 and the data transmission / reception units 103b1 to 103b4 and 103c1 to 103c4 of the child devices 101b1 to 101b4 and 101c1 to 101c4 confirm the termination notification (step ST10). If there is no notification, the process returns to the data selection process in step ST1 again, and if there is an end notification, the process ends.

  In the flowchart of FIG. 4, for convenience of explanation, a procedure for sequentially performing image reproduction processing, data exchange processing, data transmission processing, and image composition processing is shown. It may be configured to move to the next block before the above process is completed. For example, the input data A is transmitted to the child device (step ST5), the image data generated from the other input data B is received from the child device (step ST6), and the image composition process for the other input data C (step ST8) is performed. May be processed simultaneously. Alternatively, while receiving data and selecting unprocessed data A (step ST1), image reproduction processing for another data B (step ST2) and data exchange for another data C (step ST4), You may process simultaneously the image-synthesis process (step ST8) with respect to another data D, and transmission (step ST9) of another data E to the main | base station.

  As described above, according to the parallel processing device of the first embodiment, the configuration is a parent-child relationship, and each of the two or more generations includes a plurality of arithmetic units each operating in parallel. In an arithmetic unit having signal data and an imaging area as input information or from the upper generation, image data from the lower generation, and image data or image data processed by itself as output information or to the upper generation A data transmission / reception unit that receives a set of signal data and an imaging region from the upper generation, and transmits / receives the signal data or the imaging region or both from the same generation; A data allocation unit that is provided in a computation unit having a generation, divides the signal data received by the data transmission / reception unit and the imaging region, and assigns the divided signal data and imaging region pair to the lower generation; Have generations of An image reproducing unit that creates image data corresponding to an imaging area from signal data received by the data transmission / reception unit, and a computation unit having a lower generation, receives image data received by the data transmission / reception unit. The arithmetic unit that combines and has no lower generation includes an image composition unit that synthesizes image data received from the data transmission / reception unit and image data from at least one of the image data from the image reproduction unit, The processing time for forming image data based on the signal data and the imaging area can be reduced.

  Further, according to the parallel processing device of the first embodiment, the signal data is an observation signal including position information observed by the synthetic aperture radar, and the data allocation unit converts the observation signal into one or more layers according to the synthetic aperture length. Since the sub-aperture division is performed recursively above the division and the imaging region is sub-image divided into a plurality of small regions, and the image reproduction unit reproduces the SAR image by the SAR image reproduction algorithm. It is possible to shorten the processing time for reconstructing the observation results by the image data.

  In addition, according to the parallel processing device of the first embodiment, the composition processing is performed by at least one of the slave unit side and the master unit side in the parent-child relationship and the image is being reproduced by the slave unit side computation unit. In addition, all the compositing processing by the arithmetic unit on the base unit side for data in another area in the small area, data transmission / reception between the arithmetic unit on the base unit side and the slave unit side, and data transmission / reception between the arithmetic unit on the slave unit side Since the simultaneous execution is performed asynchronously, the processing time for reconstructing the observation result by the synthetic aperture radar into image data can be shortened.

  Further, according to the parallel processing device of the first embodiment, the arithmetic unit on the parent device side and the child device side in the parent-child relationship is configured by a plurality of computers, and a plurality of computers are connected by a network, When data transmission / reception between the computation unit and the computation unit on the slave unit and data transmission / reception between the computation units on the slave unit are performed by different computers, they are processed by network communication, and when performed by the same computer, on the shared memory Since data is processed in a shared form, processing to form image data based on signal data and imaging area even when the computing unit on the master unit side and slave unit side is composed of multiple computers The time can be shortened.

  Further, according to the parallel processing device of the first embodiment, in the computer, the arithmetic unit on the master unit side is a CPU, and the arithmetic unit on the slave unit side is a GPU or MIC connected by a bus. Data transmission / reception between the computing unit of the slave unit and the computing unit of the slave unit and data transmission / reception between the computation units of the slave unit are processed by communication when crossing the bus, and on the shared memory in the computer when the bus is not crossed Therefore, even when the calculation unit is configured inside the computer, the processing time for forming the image data based on the signal data and the imaging region can be shortened.

  In addition, according to the parallel processing device of the first embodiment, the number of child devices that a specific parent device has in the parent-child relationship is different from the number of child devices that other parent devices have. A parallel processor can be configured regardless of the number.

  Further, according to the parallel processing device of the first embodiment, the plurality of arithmetic units are arithmetic units of the radar system, and the radar system is mounted on the platform of the mobile unit. As a system, the processing time for forming image data based on the signal data and the imaging region can be reduced.

  Also, according to the parallel processing method of the first embodiment, the parallel processing device described in the first embodiment is used, and the received data selection step for selecting unprocessed data from the received data and the received data selection step are selected. At least one of the unprocessed input data selected in the received data selection step and the unprocessed input data selected in the received data selection step and the image data imaged in the image playback processing step is generated in the same generation. A data exchange step for exchanging with the computation unit, an unprocessed input data selected in the received data selection step, an input data transmission step for transmitting to the lower generation computation unit, and an input data transmission step An image data aggregating step for receiving image data processed by a lower generation arithmetic unit according to the transmitted input data, and image reproduction An image synthesis step for synthesizing the image data in the processing step or the image data aggregation step, and an image data transmission step for transmitting the image data processed in the image synthesis step as output information or to an arithmetic unit of an upper generation Since it is provided, it is possible to realize a parallel processing device capable of reducing the processing time for forming the image data based on the signal data and the imaging region.

  Further, according to the parallel processing method of the first embodiment, the image reproduction processing step, the data exchange step, the input data transmission step, and the image synthesis step are performed using independent work buffer areas in each step. Since the transition to the next step is performed asynchronously before the processing of the step is completed, the processing time for forming the image data based on the signal data and the imaging region can be further shortened.

Embodiment 2. FIG.
In the image reproduction process according to the first embodiment, the necessary data upsampling process may be performed on a small area by dividing the data into a set of small areas.

FIG. 5 shows changes in data size due to conventional upsampling processing and interpolation processing. In FIG. 5, the N-point signal data 401 is up-sampled to the RN-point signal data 405 obtained by finely dividing the sampling number by R times, and the T-point signal data is acquired using this as the interpolation source data.
In the processing shown in FIG. 5, first, signal data 402 in which the values of all signal elements are changed by Fourier transform is obtained. The Fourier transformed signal data 402 is equally divided into two to obtain divided signal data 403a and 403b. The signal data 404 of the RN point is created so that the value 0 is filled between the divided signal data. By performing inverse Fourier transform on this signal data 404, signal data 405 after upsampling is obtained. The interpolated data 406 is calculated for each point by referring to the data after the upsampling. For this reason, the calculation amount requires a non-linear time with respect to the length N.

  FIG. 6 is an explanatory diagram showing division of signal data for reducing the amount of calculation of the upsampling process in the second embodiment. First, the area of the interpolated data 406 is divided into four small areas 502a to 502d. On the other hand, the reference destination signal data 401 is divided into four small areas 501a to 501d so that the areas necessary for the interpolation processing of each of the small areas 502a to 502d are included in the divided areas. At this time, the small areas 501a to 501d have areas that overlap each other. Here, for convenience of explanation, the number of divisions is set to four and handled as being divided into equal sizes, but the number of divisions may be any number of two or more. Further, the division size and the overlap area size B may be unequal. Thereby, the interpolation process using the upsampling process shown in FIG. 5 is processed in four steps so as to calculate each of the small areas 502a to 502d from each of the small areas 501a to 501d of the interpolation source data. The As a result, the number of data points in the upsampling process that increases the processing time in a non-linear manner with the increase in the number of points is reduced, so that the effect of reducing the processing time can be obtained.

  As described above, according to the parallel processing device of the second embodiment, the image reproduction unit performs signal data upsampling processing, divides the interpolated data into a plurality of small regions, and calculates each small region. The original signal data is divided into a plurality of areas so that the required area is included, and the upsampling process and the interpolation process are performed independently for each of the divided original signal data. The number of processing data points can be reduced, and the reproduction processing time of the SAR image data can be further reduced.

  In the present invention, within the scope of the invention, any combination of the embodiments, or any modification of any component in each embodiment, or omission of any component in each embodiment is possible. .

  As described above, the parallel processing device according to the present invention relates to a configuration for reducing the processing time for forming image data based on signal data and an imaging region. It is suitable for use in a radar system that is an observation signal including observed position information.

  101a0, 101b0, 101c0 Master unit, 101a1, 101a2, 101b1, 101b2, 101b3, 101b4, 101c1, 101c2, 101c3, 101c4 Slave unit, 102a0, 102b0, 102c0 Data allocation unit, 102b1, 102b2, 102b3, 102b4, 102c1, 102c2 , 102c3, 102c4 Image playback unit, 103a0, 103b0, 103c0, 103b1, 103b2, 103b3, 103b4, 103c1, 103c2, 103c3, 103c4 Data transmission / reception unit, 104a0, 104b0, 104c0, 104b1, 104b2, 104b3, 104b4, 104c1, 104c2 104c3, 104c4 Image composition unit.

Claims (10)

A process for constructing an image by recursively dividing signal data, which is an observation signal including position information observed by a synthetic aperture radar, is constructed, and a plurality of arithmetic units configured in a parent-child relationship of three generations or more and operating in parallel Prepared,
The plurality of arithmetic units are:
An arithmetic unit having a lower generation receives signal data and an imaging area from the upper generation, receives image data from the lower generation, transmits the image data or image data processed by itself to the upper generation, An arithmetic unit having no generation receives a set of signal data and an imaging region from the upper generation, and generates signal data or imaging between a parent belonging to the same parent-child relationship or a child of the same generation belonging to the same parent-child relationship. A data transmission / reception unit that transmits / receives an area or both;
A data allocation unit that is provided in a calculation unit having a lower generation, divides the signal data received by the data transmission / reception unit and the imaging region, and allocates the divided signal data and imaging region group to the lower generation When,
An image reproduction unit that is provided in a calculation unit that does not have a lower generation and creates image data corresponding to an imaging region from signal data received by the data transmission / reception unit;
In the calculation unit having a lower generation, the image data received by the data transmission / reception unit is combined, and in the calculation unit having no lower generation, the image data received by the data transmission / reception unit and the image data from the image reproduction unit A parallel processing apparatus comprising: an image synthesis unit that synthesizes image data from at least one of the above. The data allocating unit recursively sub-aperture-divides the observation signal into three or more hierarchies and two or more divisions per hierarchy according to a synthetic aperture length, and sub-images the imaging region into a plurality of sub-regions,
The parallel processing apparatus according to claim 1, wherein the image reproduction unit reproduces a SAR image by a SAR (Synthetic Aperture Radar) image reproduction algorithm.   In the parent-child relationship, at least one of the slave unit side and the master unit side performs synthesis processing, and the image of the other region in the small region is targeted during image reproduction by the slave unit side computation unit 3. The synthesizing process by the arithmetic unit on the base unit side, the data transmission / reception between the arithmetic unit on the base unit side and the slave unit side, and the data transmission / reception between the arithmetic units on the slave unit side are all asynchronously executed simultaneously. The parallel processing device described.   The image reproduction unit performs upsampling processing of the signal data, divides the data after interpolation processing into a plurality of small regions, and converts the original signal data into a plurality of regions so that a region necessary for calculation of each small region is included. 3. The parallel processing according to claim 2, wherein the adjacent areas are divided so as to overlap each other, and the upsampling process and the interpolation process are independently performed on each of the divided original signal data. Processing equipment.   The computing unit on the parent device side and the child device side in the parent-child relationship is configured by a plurality of computers, the plurality of computers are connected by a network, and between the computing unit on the parent device side and the computing unit on the child device side When data transmission / reception and data transmission / reception between the computing units on the slave unit are performed by different computers, they are processed by network communication. When they are performed by the same computer, they are processed by sharing data on a shared memory. The parallel processing apparatus according to claim 1.   Inside the computer, the arithmetic unit on the master unit side is a CPU (Central Processing Unit), and the arithmetic unit on the slave unit side is a GPU (Graphics Processing Unit) or MIC (Many Integrated Core) connected by a bus. Data transmission / reception between the computation unit on the machine side and the computation unit on the slave unit side and data transmission / reception between the computation units on the slave unit side are processed by communication when straddling the bus, and the computer if not across the bus 6. The parallel processing apparatus according to claim 5, wherein the processing is performed in such a manner that the data is shared on the shared memory.   2. The parallel processing apparatus according to claim 1, wherein the number of child devices possessed by a specific parent device in a parent-child relationship is different from the number of child devices possessed by another parent device.   The parallel processing apparatus according to claim 1, wherein the plurality of calculation units are calculation units of a radar system, and the radar system is mounted on a platform of a moving body. Using the parallel processing device according to claim 1,
A received data selection step for selecting unprocessed data from the received data;
An image reproduction processing step for imaging the unprocessed input data selected in the received data selection step;
A data exchange step for exchanging at least one of the unprocessed input data selected in the received data selection step and the image data imaged in the image reproduction processing step with the same generation operation unit;
An input data transmission step of transmitting the unprocessed input data selected in the reception data selection step to the lower generation computing unit;
An image data aggregating step for receiving image data processed by the lower generation calculation unit by the input data transmitted in the input data transmitting step;
An image synthesis step of synthesizing the image data of the image reproduction processing step or the image data aggregation step;
A parallel processing method comprising: an image data transmission step of transmitting the image data processed in the image synthesis step as output information or to an upper generation arithmetic unit.   The image reproduction processing step, the data exchange step, the input data transmission step, and the image synthesis step are performed using the independent work buffer area in each step, and before the processing of each step is completed, The parallel processing method according to claim 9, wherein the processing shifts asynchronously to steps.

Images