JP2626087B2 - Parallel likelihood calculation device - Google Patents

Parallel likelihood calculation device

Info

Publication number
JP2626087B2
JP2626087B2 JP1237682A JP23768289A JP2626087B2 JP 2626087 B2 JP2626087 B2 JP 2626087B2 JP 1237682 A JP1237682 A JP 1237682A JP 23768289 A JP23768289 A JP 23768289A JP 2626087 B2 JP2626087 B2 JP 2626087B2
Authority
JP
Japan
Prior art keywords
likelihood
calculation
circuit
result
output
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.)
Expired - Fee Related
Application number
JP1237682A
Other languages
Japanese (ja)
Other versions
JPH03100600A (en
Inventor
知弘 岩▲崎▼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Mitsubishi Electric Corp
Original Assignee
Mitsubishi Electric Corp
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Mitsubishi Electric Corp filed Critical Mitsubishi Electric Corp
Priority to JP1237682A priority Critical patent/JP2626087B2/en
Publication of JPH03100600A publication Critical patent/JPH03100600A/en
Application granted granted Critical
Publication of JP2626087B2 publication Critical patent/JP2626087B2/en
Priority to HK98105530A priority patent/HK1006304A1/en
Anticipated expiration legal-status Critical
Expired - Fee Related legal-status Critical Current

Links

Description

【発明の詳細な説明】 〔産業上の利用分野〕 この発明は音声認識などを目的として行われる登録デ
ータ群に対する入力データの尤度演算を並列処理する並
列尤度演算装置に関するものである。
Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a parallel likelihood calculation device that performs a likelihood calculation of input data for a registration data group for speech recognition or the like in parallel.

〔従来の技術〕[Conventional technology]

従来、この種の並列演算回路として、第3図に示すよ
うなものがあった。この図は特許出願「連続音声認識装
置」(出願番号:特願昭63−104770、出願日:昭和63年
4月27日、発明者:岩崎知弘、石川泰、中島邦男)に示
された技術に基づいて作成された構成を示すブロック図
である。図において、Nを規定数として、1は入力デー
タに対する尤度演算を並列処理可能な小単位の尤度演算
に分割し、この小単位の尤度演算を実行させる尤度演算
命令を作成する尤度演算分割回路、2は前記尤度演算命
令に従い前記小単位の尤度演算を実行し尤度を含む演算
結果を出力する尤度演算回路、3は演算結果を記憶する
演算結果記憶回路、4は分割された小単位の尤度演算が
すべて終了した時点で演算結果記憶回路3に記憶されて
いる演算結果のうち尤度の高いN個の演算結果を出力す
る演算結果決定回路、5は入力データ、6は尤度演算命
令、7は演算結果、8は出力データである。このような
構成のシステムにおいては以下のようにして尤度演算が
並列処理される。
Conventionally, there has been a parallel operation circuit of this type as shown in FIG. This figure shows the technology shown in the patent application “Continuous speech recognition device” (application number: Japanese Patent Application No. 63-104770, application date: April 27, 1988, inventors: Tomohiro Iwasaki, Yasushi Ishikawa, Kunio Nakajima) FIG. 3 is a block diagram showing a configuration created based on the above. In the figure, with N being a specified number, 1 divides the likelihood calculation for input data into likelihood calculations in small units that can be processed in parallel, and creates a likelihood calculation instruction for executing the likelihood calculation in the small units. A degree calculation division circuit, a likelihood calculation circuit that executes the small-unit likelihood calculation according to the likelihood calculation instruction and outputs a calculation result including the likelihood, and a calculation result storage circuit that stores the calculation result, Is an operation result determination circuit that outputs N operation results with high likelihood among the operation results stored in the operation result storage circuit 3 when all the divided small unit likelihood operations are completed. Data, 6 is a likelihood calculation instruction, 7 is a calculation result, and 8 is output data. In the system having such a configuration, the likelihood calculation is performed in parallel as follows.

まず尤度演算分割回路1は外部より入力される入力デ
ータ5に対して行われる尤度演算を並列演算可能な演算
の単位である小単位の尤度演算に分割し、この小単位の
尤度演算を尤度演算回路2に行わせるための尤度演算命
令を作成する。尤度演算分割回路1は演算を行っていな
い尤度演算回路2に対して次々と尤度演算命令6を出力
することにより並列処理わ行わせる。尤度演算回路2で
はこの尤度演算命令6で指定されている尤度演算を行い
尤度を含む演算結果7aを出力する。演算結果7aは演算結
果記憶回路3に入力され記憶される。演算結果決定回路
4では尤度演算回路2における入力データ5に対する小
単位の尤度演算がすべて終了した時点において、演算結
果記憶回路3に記憶されている演算結果7bのうち尤度の
高いものからN個選択して出力データ8として出力す
る。
First, the likelihood calculation division circuit 1 divides a likelihood calculation performed on input data 5 input from the outside into a small unit likelihood calculation which is a unit of a calculation that can be performed in parallel. A likelihood calculation instruction for causing the likelihood calculation circuit 2 to perform a calculation is created. The likelihood calculation division circuit 1 outputs the likelihood calculation instructions 6 one after another to the likelihood calculation circuit 2 that has not performed the calculation, thereby performing the parallel processing. The likelihood calculation circuit 2 performs the likelihood calculation specified by the likelihood calculation instruction 6, and outputs a calculation result 7a including the likelihood. The operation result 7a is input to and stored in the operation result storage circuit 3. When all the likelihood calculations for the input data 5 in the likelihood calculation circuit 2 have been completed in the calculation result determination circuit 4, the calculation result 7 b stored in the calculation result storage circuit 3 starts with the one with the highest likelihood. N pieces are selected and output as output data 8.

以下、三つの尤度演算回路2を備え、最終的に演算結
果決定回路4において三つの演算結果を選択し出力する
並列尤度演算装置を例として説明を行う。第4図はこの
並列尤度演算装置において、第3図に示す入力データに
対する尤度演算が5つの小単位の尤度演算に分割できる
場合における、各尤度演算回路2における小単位の尤度
演算の処理のタイミングチャートと、各尤度演算回路2
から出力される演算結果を示している。分割された小単
位の尤度演算をそれぞれA,B,C,D,Eと呼び、3つの尤度
演算回路2をそれぞれL1,L2,L3と呼ぶ。はじめは尤度演
算回路L1,L2,L3では尤度演算は行われていないものとす
る。時間TOおいて尤度演算回路L1に対し小単位の尤度演
算Aを実行させる尤度演算命令が出力され、尤度演算回
路L2に対し小単位の尤度演算Bを実行させる尤度演算命
令が出力され、尤度演算回路L3に対し小単位の尤度演算
Cを処理する尤度演算命令が出力される。尤度演算回路
L1では時間TAにおいて小単位の尤度演算Aが終了し演算
結果A1,A2,A3が出力され演算結果記憶回路3に記憶され
る。演算結果A1,A2,A3は小単位の尤度演算Aにおける演
算結果のうち尤度が大きなものから三つ選択して出力し
たものでありそれぞれの尤度を7,4,2という値とする。
この例では最終的に演算結果決定回路4において三つの
演算結果を選択するため、各尤度演算回路L1,L2,L3にお
いても尤度の大きいものから三つの演算結果を出力すれ
ば十分である。処理の終了した尤度演算回路L1に対し次
の小単位の尤度演算Dを実行する演算命令が出力され尤
度演算が開始される。次に尤度演算回路L2において時間
TBに小単位の尤度演算Bが終了し小単位の尤度演算Aと
同様に演算結果B1,B2,B3が出力され演算結果記憶回路3
に記憶される。演算結果B1,B2,B3の尤度はそれぞれ9,5,
1であったとする。処理の終了した尤度演算回路L2に対
し次の小単位の尤度演算Eを実行する演算命令が出力さ
れ尤度演算が開始される。以下同様にして小単位の尤度
演算C,E,Dの順番で演算が終了し演算結果C1,C2,C3,E1,E
2,E3,D1,D2,D3が出力され演算結果記憶回路3に記憶さ
れるものとする。入力データ5に対するすべての小単位
の尤度演算が終了した時間TDにおいて、演算結果記憶回
路3には15個の演算結果が記憶されており、演算結果決
定回路4において、これらの演算結果のうち尤度の大き
いものから三つの演算結果B1,E1,A1が選択され、出力デ
ータとして出力される。
Hereinafter, a description will be given of an example of a parallel likelihood calculation device including three likelihood calculation circuits 2 and finally selecting and outputting three calculation results in the calculation result determination circuit 4. FIG. 4 shows a small unit likelihood in each likelihood calculation circuit 2 when the likelihood calculation for the input data shown in FIG. 3 can be divided into five small unit likelihood calculations in this parallel likelihood calculation device. Timing chart of calculation processing and each likelihood calculation circuit 2
5 shows the calculation result output from. The divided small unit likelihood calculations are called A, B, C, D, and E, respectively, and the three likelihood calculation circuits 2 are called L1, L2, and L3, respectively. At first, it is assumed that the likelihood calculation circuits L1, L2, L3 do not perform the likelihood calculation. At the time T O, a likelihood calculation instruction for executing the small-unit likelihood calculation A to the likelihood calculation circuit L1 is output, and the likelihood calculation for executing the small-unit likelihood calculation B to the likelihood calculation circuit L2. An instruction is output, and a likelihood operation instruction for processing the small unit likelihood operation C is output to the likelihood operation circuit L3. Likelihood calculation circuit
In the L1 time T A likelihood calculation A small unit is complete computation result A1, A2, A3 is stored is outputted to the arithmetic result storage circuit 3. The calculation results A1, A2, and A3 are obtained by selecting and outputting three of the calculation results in the small unit likelihood calculation A from the one having the highest likelihood, and setting each likelihood to a value of 7, 4, and 2. .
In this example, since three operation results are finally selected in the operation result determination circuit 4, it is sufficient for each of the likelihood operation circuits L1, L2, and L3 to output the three operation results in descending order of likelihood. . An operation instruction for executing the next small unit likelihood operation D is output to the likelihood operation circuit L1 that has completed the processing, and the likelihood operation is started. Next, in the likelihood calculation circuit L2, the time
The small-unit likelihood calculation B ends at T B , and the calculation results B1, B2, and B3 are output as in the small-unit likelihood calculation A, and the calculation result storage circuit 3
Is stored. The likelihood of the operation results B1, B2, B3 are 9, 5,
Suppose it was 1. An operation instruction for executing the next small-unit likelihood operation E is output to the likelihood operation circuit L2 that has completed the processing, and the likelihood operation is started. In the same manner, the operation is completed in the order of the small unit likelihood operations C, E, and D, and the operation results C1, C2, C3, E1, E
2, E3, D1, D2, and D3 are output and stored in the operation result storage circuit 3. At the time T D at which all the small-unit likelihood calculations on the input data 5 are completed, 15 calculation results are stored in the calculation result storage circuit 3. Among these, the three calculation results B1, E1, and A1 are selected from the one with the highest likelihood and output as output data.

〔発明が解決しようとる課題〕[Problems to be solved by the invention]

並列処理には処理を複数の演算回路に分けることによ
り生じるデータ転送量の増大が問題となる。上記のよう
に従来の並列尤度演算装置では各尤度演算回路から出力
される演算結果をずて演算結果記憶回路に記憶する構造
であったため、演算結果の転送量が大きく、加えて演算
結果記憶回路の容量も大きなものが必要であるという問
題点があった。
A problem in parallel processing is an increase in the amount of data transfer caused by dividing the processing into a plurality of arithmetic circuits. As described above, the conventional parallel likelihood calculation device has a structure in which the calculation results output from the respective likelihood calculation circuits are stored in the calculation result storage circuit instead of the calculation results. There is a problem that a large capacity of the storage circuit is required.

本発明は上記のような問題点を解決するためになされ
たもので、演算結果記憶回路に転送される。演算結果の
うち最終的に演算結果決定回路で選択されないことが明
らかな尤度の小さな演算結果を予め除外することによ
り、演算結果記憶回路への演算結果の転送量を減少さ
せ、演算結果記憶回路も小さな容量にすることができる
並列尤度演算装置を得ることを目的とする。
The present invention has been made to solve the above problems, and is transferred to an operation result storage circuit. By previously excluding, from the operation results, operation results having a small likelihood that are not clearly selected by the operation result determination circuit in advance, the transfer amount of the operation results to the operation result storage circuit is reduced, and the operation result storage circuit is reduced. It is another object of the present invention to obtain a parallel likelihood calculation device capable of reducing the capacity.

〔課題を解決するための手段〕[Means for solving the problem]

この発明に係る並列尤度演算装置は、入力データに対
する尤度演算を並列処理可能な小単位の尤度演算に分割
する尤度演算分割回路1と、この尤度演算分割回路1で
分割された小単位の尤度演算を実行し尤度を含む演算結
果を出力する複数の尤度演算回路2と、この尤度演算回
路2より出力された演算結果のうち指定された閾値より
尤度の大きい演算結果のみを選択して出力する閾値外結
果除去回路9と、この閾値外結果除去回路9より出力さ
れる演算結果を記憶する演算結果記憶回路3と、この演
算結果記憶回路3に演算結果が新たに入力される度にこ
の演算結果記憶回路3に記憶されているすべての演算結
果の尤度のうち値が第N番目に高いものを尤度の閾値と
して上記尤度演算分割回路1に指定する閾値演算回路10
と、上記尤度演算分割回路1で分割された小単位の尤度
演算がすべて終了した時点で上記演算結果記憶回路3に
記憶されているすべての演算結果のうち尤度の高いもの
からN個の演算結果を選択して出力データとして出力す
る演算結果決定回路4とを備えたものである。
A parallel likelihood calculation device according to the present invention includes a likelihood calculation division circuit 1 that divides a likelihood calculation for input data into small-unit likelihood calculations that can be processed in parallel, and the likelihood calculation division circuit 1 A plurality of likelihood operation circuits 2 for executing a small unit of likelihood operation and outputting an operation result including the likelihood; and a calculation result output from the likelihood operation circuit 2 having a larger likelihood than a specified threshold value. An out-of-threshold result removal circuit 9 that selects and outputs only the operation result, an operation result storage circuit 3 that stores the operation result output from the out-of-threshold result removal circuit 9, and an operation result stored in the operation result storage circuit 3. Every time a new input is made, the likelihood calculation division circuit 1 designates, as a likelihood threshold, the likelihood threshold value of the Nth highest likelihood of all the calculation results stored in the calculation result storage circuit 3. Threshold calculation circuit 10
When all the likelihood calculations of the small units divided by the likelihood calculation division circuit 1 are completed, N of the calculation results having the highest likelihood among all the calculation results stored in the calculation result storage circuit 3 are obtained. And an operation result determination circuit 4 for selecting the operation result of (1) and outputting the result as output data.

〔作用〕[Action]

尤度演算分割回路1は入力データに対する尤度演算を
並列処理可能な小単位の尤度演算に分割する。尤度演算
回路2は上記分割された小単位の尤度演算を実行した尤
度を含む演算結果を出力する。閾値外結果除去回路9は
上記演算結果のうち指定された閾値より尤度の大きい演
算結果のみを選択して出力する。演算結果記憶回路3は
上記選択された演算結果を記憶する。閾値演算回路10は
演算結果記憶回路3に演算結果が新たに入力される度に
この演算結果記憶回路3に記憶されているすべての演算
結果の尤度のうち値が第N番目に高いものを尤度の閾値
として閾値外結果除去回路1に指定する。演算結果決定
回路4は尤度演算分割回路1で分割された小単位の尤度
演算がすべて終了した時点で演算結果記憶回路3に記憶
されているすべての演算結果のうち尤度の高いものから
N個の演算結果を選択して出力データとして出力する。
The likelihood calculation division circuit 1 divides the likelihood calculation on input data into likelihood calculations in small units that can be processed in parallel. The likelihood calculation circuit 2 outputs a calculation result including the likelihood of executing the above-mentioned divided small unit likelihood calculation. The out-of-threshold result removing circuit 9 selects and outputs only the operation result having a higher likelihood than the specified threshold value among the operation results. The operation result storage circuit 3 stores the selected operation result. Each time a calculation result is newly input to the calculation result storage circuit 3, the threshold value calculation circuit 10 determines the likelihood of all the calculation results stored in the calculation result storage circuit 3 having the Nth highest likelihood. It is designated to the out-of-threshold result removing circuit 1 as a likelihood threshold. The operation result determination circuit 4 starts from the operation result with the highest likelihood among all the operation results stored in the operation result storage circuit 3 at the time when all the small unit likelihood operations divided by the likelihood operation division circuit 1 are completed. N operation results are selected and output as output data.

〔実施例〕〔Example〕

第1図はこの発明の一実施例による並列尤度演算装置
の構成を示すブロック図である。第1図において、第3
図に示す構成要素に対応するものには同一の符号を付
し、その説明を省略する。第1図において、9は尤度演
算回路2より出力された演算結果7aのうち指定された閾
値より尤度の大きい演算結果のみを選択して出力する閾
値外結果除去回路、10は演算結果記憶回路3に新たな演
算結果7cが入力される度に演算結果記憶回路3に記憶さ
れているすべての演算結果の尤度のうち第N番目のもの
を尤度の閾値として閾値外結果除去回路9に指定する閾
値演算回路である。11は閾値演算回路10より出力される
閾値である。
FIG. 1 is a block diagram showing a configuration of a parallel likelihood calculation device according to an embodiment of the present invention. In FIG. 1, the third
Components corresponding to those shown in the drawings are denoted by the same reference numerals, and description thereof will be omitted. In FIG. 1, reference numeral 9 denotes an out-of-threshold result removing circuit for selecting and outputting only an operation result having a higher likelihood than a specified threshold value among operation results 7a output from the likelihood operation circuit 2, and 10 denotes an operation result storage Every time a new calculation result 7c is input to the circuit 3, the Nth one of the likelihoods of all the calculation results stored in the calculation result storage circuit 3 is set as a likelihood threshold, and the out-of-threshold result removal circuit 9 is used. Is a threshold value operation circuit specified in FIG. Reference numeral 11 denotes a threshold value output from the threshold value calculation circuit 10.

このような構成のシステムにおいては以下のようにし
て尤度演算が並列処理される。まず従来の並列尤度演算
装置の例と同様に、尤度演算分割回路1は外部より入力
される入力データ5に対して行われる尤度演算を小単位
の尤度演算に分割し、尤度演算命令を作成する。尤度演
算分割回路1は演算を行っていない尤度演算回路2に対
して次々と尤度演算命令6を出力することにより並列処
理を行わせる。尤度演算回路2ではこの尤度演算命令6
で指定されている尤度演算を行い尤度を含む演算結果7a
を出力する。閾値外結果除去回路9においては尤度演算
回路2から出力された演算結果7aのうち閾値演算回路10
により指定される閾値11より尤度の小さな演算結果7aを
除去し、残った演算結果7cのみ出力する。演算結果7cは
演算結果記憶回路3に入力され記憶される。演算結果決
定回路4では尤度演算回路2における入力データ5に対
する小単位の尤度演算がすべて終了した時点において、
演算記憶回路3に記憶されている演算結果7bのうち尤度
の高いものからN個選択して出力データ8として出力す
る。閾値演算回路10では演算結果記憶回路3に新たな演
算結果7cが入力される度に演算結果記憶回路3に記憶さ
れているすべての演算結果の尤度のうち値が第N番目に
高いものを尤度の閾値11として閾値外結果除去回路2に
指定する。
In the system having such a configuration, the likelihood calculation is performed in parallel as follows. First, as in the example of the conventional parallel likelihood calculation device, the likelihood calculation division circuit 1 divides the likelihood calculation performed on the input data 5 input from the outside into likelihood calculations in small units, and Create an operation instruction. The likelihood calculation division circuit 1 causes the likelihood calculation circuit 2 that has not performed the calculation to output the likelihood calculation instruction 6 one after another, thereby performing the parallel processing. In the likelihood calculation circuit 2, this likelihood calculation instruction 6
Performs the likelihood calculation specified in Step 7a and includes the likelihood
Is output. In the out-of-threshold result removing circuit 9, the threshold value calculating circuit 10 of the calculation results 7a output from the likelihood calculating circuit 2 is used.
Removes the calculation result 7a having a smaller likelihood than the threshold value 11 specified by, and outputs only the remaining calculation result 7c. The operation result 7c is input to and stored in the operation result storage circuit 3. In the operation result determination circuit 4, when all the likelihood operations of the input data 5 in the small unit in the likelihood operation circuit 2 are completed,
From the operation results 7b stored in the operation storage circuit 3, N items having the highest likelihood are selected and output as output data 8. Every time a new operation result 7c is input to the operation result storage circuit 3, the threshold value operation circuit 10 determines the likelihood of all the operation results stored in the operation result storage circuit 3 that has the Nth highest value. It is specified as the likelihood threshold value 11 to the out-of-threshold result removal circuit 2.

以下、第1図に示す三つの尤度演算回路2を備え、最
終的に演算結果決定回路4において三つの演算結果を選
択し出力する並列尤度演算装置を例として説明を行う。
第2図は、この並列尤度演算装置において入力データに
対する尤度演算が5つの小単位の尤度演算に分割できる
場合における、各尤度演算回路2における小単位の尤度
演算の処理タイミングチャートと、各尤度演算回路2か
ら出力される演算結果と、各閾値外結果除去回路9から
出力される演算結果と、閾値演算回路10により求められ
る閾値を示している。分割された小単位の尤度演算をそ
れぞれA,B,C,D,Eと呼び、3つの尤度演算回路2をそれ
ぞれL1,L2,L3と呼ぶ。はじめは尤度演算回路L1,L2,L3で
は尤度演算は行われておらず、尤度演算回路L1,L2,L3の
閾値は尤度の最低の値である0に設定されているものと
する。時間TOにおいて尤度演算回路L1に対し小単位の尤
度演算Aを実行させる尤度演算命令が出力され、尤度演
算回路L2に対し小単位の尤度演算Bを実行させる尤度演
算命令が出力され、尤度演算回路L3に対し小単位の尤度
演算Cを処理する尤度演算命令が出力される。尤度演算
回路L1では時間TAにおいて小単位の尤度演算Aが終了し
演算結果A1,A2,A3が出力され演算結果記憶回路3に記憶
される。演算結果A1,A2,A3は小単位の尤度演算Aにおけ
る演算結果のうち尤度が大きなものから三つ選択して出
力したものでありそれぞれの尤度を7,4,2という値とす
る。閾値外結果除去回路9ではこの時点での閾値は0で
あるため演算結果A1,A2,A3はすべて出力され演算結果記
憶回路3に記憶される。処理の終了した尤度演算回路L1
に対し次の小単位の尤度演算Dを実行する演算命令が出
力され尤度演算が開始される。同時に閾値演算回路10に
おいて演算結果A1,A2,A3より閾値が演算され演算結果記
憶回路3の尤度のうち値が第三番目の高い2という値が
閾値として閾値外結果除去回路9に指定される。
Hereinafter, a description will be given of a parallel likelihood calculation device having three likelihood calculation circuits 2 shown in FIG. 1 and finally selecting and outputting three calculation results in the calculation result determination circuit 4 as an example.
FIG. 2 is a processing timing chart of the small unit likelihood calculation in each likelihood calculation circuit 2 when the likelihood calculation for the input data can be divided into five small unit likelihood calculations in the parallel likelihood calculation device. , The operation result output from each likelihood operation circuit 2, the operation result output from each non-threshold result removal circuit 9, and the threshold value obtained by the threshold operation circuit 10. The divided small unit likelihood calculations are called A, B, C, D, and E, respectively, and the three likelihood calculation circuits 2 are called L1, L2, and L3, respectively. Initially, likelihood calculation circuits L1, L2, L3 do not perform likelihood calculation, and the thresholds of likelihood calculation circuits L1, L2, L3 are set to 0, which is the lowest value of likelihood. I do. At time T O , a likelihood operation instruction for executing the small unit likelihood operation A to the likelihood operation circuit L1 is output, and a likelihood operation instruction for executing the small unit likelihood operation B to the likelihood operation circuit L2. Is output to the likelihood calculation circuit L3 to output a likelihood calculation command for processing the likelihood calculation C in small units. The likelihood calculation circuit L1 terminates the small-unit likelihood calculation A at time T A , outputs the calculation results A1, A2, and A3 and stores them in the calculation result storage circuit 3. The calculation results A1, A2, and A3 are obtained by selecting and outputting three of the calculation results in the small unit likelihood calculation A from the one having the highest likelihood, and setting each likelihood to a value of 7, 4, and 2. . In the out-of-threshold result removing circuit 9, since the threshold value at this time is 0, all the operation results A1, A2, A3 are output and stored in the operation result storage circuit 3. Likelihood calculation circuit L1 for which processing has been completed
, An operation instruction for executing the next small unit likelihood operation D is output, and the likelihood operation is started. At the same time, the threshold value is calculated from the calculation results A1, A2, and A3 in the threshold value calculation circuit 10, and the value of the third highest likelihood 2 in the calculation result storage circuit 3 is designated as the threshold value in the out-of-threshold result removal circuit 9. You.

次に尤度演算回路L2において時間TBに小単位の尤度演
算Bが終了し小単位の尤度演算Aと同様に演算結果B1,B
2,B3が出力され演算結果記憶回路3に記憶される。演算
結果B1,B2,B3の尤度はそれぞれ9,5,1であったとする。
閾値外結果除去回路9でこの時点での閾値は2であるた
め、これより尤度の低い演算結果B3は除去され演算結果
B1とB2が出力される演算結果記憶回路3に記憶される。
処理の終了した尤度演算回路L2に対し次の小単位の尤度
演算Eを実行する演算命令が出力され尤度演算が開始さ
れる。同時に閾値演算回路10においては演算結果A1,A2,
A3,B1,B2より閾値が演算され5という値が閾値として閾
値外結果除去回路9に指定される。以下同様にして小単
位の尤度演算Cでは演算結果記憶回路3にC1のみが記憶
され閾値が6になり、小単位の尤度演算Eでは演算結果
記憶回路3にE1のみが記憶され閾値が7になり小単位の
尤度演算Dでは閾値外結果除去回路9よりすべての演算
結果D1,D2,D3が除去される。入力データに対するすべて
の小単位の尤度演算が終了した時間TDにおいて、演算結
果記憶回路3には7個の演算結果が記憶されており、演
算結果決定回路4において、これらの演算結果のうち尤
度の大きいものから三つの演算結果B1,E1,A1が選択さ
れ、出力結果として出力される。演算結果記憶回路3に
転送された演算結果は従来の並列尤度演算装置より少な
いが、同じ出力結果が得られる。
Then, the likelihood calculation circuit L2 terminates the small-unit likelihood calculation B at time T B , and calculates the calculation results B1, B in the same manner as the small-unit likelihood calculation A.
2 and B3 are output and stored in the operation result storage circuit 3. It is assumed that the likelihoods of the calculation results B1, B2, and B3 are 9, 5, and 1, respectively.
Since the threshold value at this point in the out-of-threshold result removing circuit 9 is 2, the operation result B3 having a lower likelihood is removed, and the operation result
B1 and B2 are stored in the output operation result storage circuit 3.
An operation instruction for executing the next small-unit likelihood operation E is output to the likelihood operation circuit L2 that has completed the processing, and the likelihood operation is started. At the same time, the operation results A1, A2,
The threshold value is calculated from A3, B1, and B2, and a value of 5 is designated as the threshold value to the out-of-threshold result removing circuit 9. In the same manner, in the small-unit likelihood calculation C, only C1 is stored in the calculation result storage circuit 3 and the threshold is 6, and in the small-unit likelihood calculation E, only E1 is stored in the calculation result storage circuit 3 and the threshold is set. In the small-unit likelihood calculation D, the out-of-threshold result removal circuit 9 removes all calculation results D1, D2, and D3. At the time T D at which all the small-unit likelihood calculations on the input data are completed, seven calculation results are stored in the calculation result storage circuit 3. Three calculation results B1, E1, and A1 are selected from those with the highest likelihood and output as output results. Although the calculation result transferred to the calculation result storage circuit 3 is smaller than that of the conventional parallel likelihood calculation device, the same output result can be obtained.

なお、上記実施例では尤度演算回路が三個の場合につ
いて説明したが、尤度演算回路の個数は三個以外の個数
にしてもよく、この場合も上記実施例と同様の効果があ
る。また同じく、上記実施例では演算結果決定回路にお
いて三個の演算結果を選択する場合について説明した
が、演算結果出力回路で選択する演算結果の個数も三個
以外でもよく、この場合も上記実施例と同様の効果があ
る。
In the above embodiment, the case where the number of likelihood calculation circuits is three has been described. However, the number of likelihood calculation circuits may be other than three, and the same effect as in the above embodiment can be obtained. Similarly, in the above embodiment, the case where three operation results are selected in the operation result determination circuit has been described. However, the number of operation results selected in the operation result output circuit may be other than three. Has the same effect as.

〔発明の効果〕〔The invention's effect〕

以上のようにこの発明によれば、演算結果記憶回路に
記憶されている演算結果のうち尤度の高いものから一定
数選択するための尤度の閾値を常時演算する閾値演算回
路と、尤度演算回路より出力された演算結果のうち閾値
演算回路で求められた閾値より尤度の大きい演算結果の
み選択して出力する閾値外結果除去回路とを備えて構成
したので、最終的に演算結果決定回路で選択されないこ
とが明らかな尤度の小さな演算結果を予め除外でき、こ
れにより演算結果記憶回路への演算結果の転送量を減少
でき、かつ演算結果記憶回路も小さな容量のものにする
ことができ、したがって、演算処理速度の向上及びハー
ドウェアの小規模化を図れるという効果が得られる。
As described above, according to the present invention, a threshold value operation circuit that constantly calculates a likelihood threshold value for selecting a certain number of operation results having a high likelihood from among operation results stored in an operation result storage circuit, A non-threshold result removing circuit that selects and outputs only the operation result having a greater likelihood than the threshold value obtained by the threshold value operation circuit among the operation results output from the operation circuit, and finally determines the operation result Calculation results with small likelihoods that are clearly not selected by the circuit can be excluded in advance, thereby reducing the amount of transfer of the calculation results to the calculation result storage circuit and reducing the size of the calculation result storage circuit. Therefore, the effect that the arithmetic processing speed can be improved and the hardware can be downsized can be obtained.

【図面の簡単な説明】[Brief description of the drawings]

第1図はこの発明の一実施例に係る並列尤度演算装置の
構成を示すブロック図、第2図は第1図に示す三つの尤
度演算回路を備え最終的に演算結果決定回路において三
つの演算結果を選択し出力する並列尤度演算装置におい
て、入力データに対する尤度演算が5つの小単位の尤度
演算に分割できる場合における、各尤度演算回路におけ
る小単位の尤度演算の処理のタイミングチャートと、各
尤度演算回路から出力される演算結果と、各演算結果除
去回路から出力される演算結果と、閾値演算回路により
求められる閾値とを示す図、第3図は従来の並列尤度演
算装置の構成を示すブロック図、第4図は第3図に示す
三つの尤度演算回路を備え最終的に演算結果決定回路に
おいて三つの演算結果を選択し出力する並列尤度演算装
置において、入力データに対する尤度演算が5つの小単
位の尤度演算に分割できる場合における、各尤度演算回
路における小単位の尤度演算の処理のタイミングチャー
トと、各尤度演算回路から出力される演算結果とを示す
図である。 1……尤度演算分割回路、2……尤度演算回路、3……
演算結果記憶回路、4……演算結果決定回路、9……閾
値外結果除去回路、10……閾値演算回路。
FIG. 1 is a block diagram showing a configuration of a parallel likelihood calculation device according to an embodiment of the present invention, and FIG. 2 is provided with three likelihood calculation circuits shown in FIG. In the parallel likelihood calculation device that selects and outputs two calculation results, processing of the small-unit likelihood calculation in each likelihood calculation circuit when the likelihood calculation for the input data can be divided into five small-unit likelihood calculations And FIG. 3 shows a timing chart of FIG. 5, a calculation result output from each likelihood calculation circuit, a calculation result output from each calculation result removal circuit, and a threshold value obtained by a threshold calculation circuit. FIG. 4 is a block diagram showing a configuration of the likelihood calculation device, and FIG. 4 is a parallel likelihood calculation device having three likelihood calculation circuits shown in FIG. 3 and finally selecting and outputting three calculation results in a calculation result determination circuit. At the input When the likelihood calculation for the data can be divided into five small-unit likelihood calculations, the timing chart of the small-unit likelihood calculation processing in each likelihood calculation circuit and the calculation output from each likelihood calculation circuit It is a figure which shows a result. 1 ... likelihood calculation division circuit, 2 ... likelihood calculation circuit, 3 ...
Calculation result storage circuit, 4... Calculation result determination circuit, 9... Out-of-threshold result removal circuit, 10.

Claims (1)

(57)【特許請求の範囲】(57) [Claims] 【請求項1】入力データに対する尤度演算を並列処理可
能な小単位の尤度演算に分割する尤度演算分割回路と、
この尤度演算分割回路で分割された小単位の尤度演算を
実行し尤度を含む演算結果を出力する複数の尤度演算回
路と、この尤度演算回路より出力された演算結果のうち
指定された閾値より尤度の大きい演算結果のみを選択し
て出力する閾値外結果除去回路と、この閾値外結果除去
回路より出力される演算結果を記憶する演算結果記憶回
路と、この演算結果記憶回路に演算結果が新たに入力さ
れる度にこの演算結果記憶回路に記憶されているすべて
の演算結果の尤度のうち値が第N番目に高いものを尤度
の閾値として上記閾値外結果除去回路に指定する閾値演
算回路と、上記尤度演算分割回路で分割された小単位の
尤度演算がすべて終了した時点で上記演算結果記憶回路
に記憶されているすべての演算結果のうち尤度の高いも
のからN個の演算結果を選択して出力データとして出力
する演算結果決定回路とを備えたことを特徴とする並列
尤度演算装置。
1. A likelihood operation division circuit for dividing a likelihood operation on input data into small unit likelihood operations that can be processed in parallel,
A plurality of likelihood calculation circuits that execute the likelihood calculation of the small unit divided by the likelihood calculation division circuit and output a calculation result including the likelihood, and a designation among the calculation results output from the likelihood calculation circuit An out-of-threshold result removing circuit that selects and outputs only an operation result having a higher likelihood than the threshold value, an operation result storage circuit that stores the operation result output from the out-of-threshold result removing circuit, and an operation result storage circuit Each time a calculation result is newly input to the calculation result storage circuit, the Nth highest likelihood of all the calculation results stored in the calculation result storage circuit is set as the likelihood threshold, and the out-of-threshold result removal circuit is used. , And when all the likelihood calculations of the small units divided by the likelihood calculation division circuit are completed, all the calculation results stored in the calculation result storage circuit have a high likelihood. N operations from things Parallel likelihood computation device is characterized in that a calculation result determination circuit for selecting and outputting result as output data.
JP1237682A 1989-09-08 1989-09-13 Parallel likelihood calculation device Expired - Fee Related JP2626087B2 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP1237682A JP2626087B2 (en) 1989-09-13 1989-09-13 Parallel likelihood calculation device
HK98105530A HK1006304A1 (en) 1989-09-08 1998-06-17 Load detector device for an elevator

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP1237682A JP2626087B2 (en) 1989-09-13 1989-09-13 Parallel likelihood calculation device

Publications (2)

Publication Number Publication Date
JPH03100600A JPH03100600A (en) 1991-04-25
JP2626087B2 true JP2626087B2 (en) 1997-07-02

Family

ID=17018947

Family Applications (1)

Application Number Title Priority Date Filing Date
JP1237682A Expired - Fee Related JP2626087B2 (en) 1989-09-08 1989-09-13 Parallel likelihood calculation device

Country Status (1)

Country Link
JP (1) JP2626087B2 (en)

Also Published As

Publication number Publication date
JPH03100600A (en) 1991-04-25

Similar Documents

Publication Publication Date Title
JPS6077265A (en) Vector processor
EP0478006B1 (en) Method and apparatus for searching data
CN112015366B (en) Data sorting method, data sorting device and database system
JP2626087B2 (en) Parallel likelihood calculation device
JP2503984B2 (en) Information processing device
JPS6310263A (en) Vector processor
US20050097140A1 (en) Method for processing data streams divided into a plurality of process steps
JPH0652215A (en) Matrix arithmetic processor
JP3347592B2 (en) Merge sort processor
JPH04364525A (en) Parallel arithmetic unit
JP3447180B2 (en) Data operation circuit
JPH05158895A (en) System for improving loop calculation efficiency in parallel computer system
JPH04184535A (en) Parallel arithmetic units
JP2001357395A (en) Device and method for processing picture
CN116225532A (en) General processor supporting acceleration vector operation
JPS6277651A (en) Branch processing system for data flow type computer
KR0157337B1 (en) Multi-bit adder for digital signal process
JP2503983B2 (en) Information processing device
JPH0810432B2 (en) Information processing device
JPS61193201A (en) Optimization and extraction processing method for characteristic function
JPS59197920A (en) Address controlling device
JPH0810431B2 (en) Information processing device
JPH02130629A (en) External sort calculating method
JPH04125733A (en) Pipeline information processing system
JPH05738B2 (en)

Legal Events

Date Code Title Description
LAPS Cancellation because of no payment of annual fees