JP2003177912A - Method of calculating median of numerical data array using simd operation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To speed up the processing of a median filter which is one of noise removing means for an image by using a parallel operation function that a processor has for the operation of finding a median from the numerical data array. <P>SOLUTION: An SIMD operation function that the processor has is used for respective pieces of numerical data to be processed and one piece of data is compared with all of other pieces of numerical data in parallel. Then the number of pieces of 'data having larger attribute values than the one piece of data' is obtained. In a data array of 2n-1 pieces of data, the number of pieces of 'data having larger attribute values than the one piece of data' which is n-1 is the median in the numerical data array. When there are a plurality of pieces data having the same attribute value in the data array, data whose number of pieces of 'data having larger attribute values than the one piece of data' is n-1 is possibly absent. To process the exception, the number of pieces of 'data having larger attribute values than the one piece of data' is stored as a histogram and the median is obtained from the histogram. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of calculating a median value of numerical data (when they are arranged in the order of their attribute values), and SIMD calculation applied to a process such as a median filter for removing noise from an image. The present invention relates to a method of calculating a median value of a numerical data string using.

[0002]

2. Description of the Related Art In recent years, it has been required to process data such as images and sounds at high speed. In order to process data at high speed, there are a method of increasing the operation speed of the arithmetic device to speed up the data arithmetic processing, and a method of arithmetically processing a plurality of data in parallel by the arithmetic device to increase the number of data that can be simultaneously calculated. There is.

As a method of improving the operation speed, SIM
D (Single Instruction Multi)
There are operations of single data and multiple data). It is a form of parallel data processing and belongs to the latter. A conventional computer can be expanded and parallel data processing can be realized without changing its configuration.

A median filter is known as a method for removing noise in an image. With respect to the target pixel in the image, a group including its neighborhood is used as a numerical data string, and when the pixels are rearranged in the order of the density value of the pixel, the pixel located at the center thereof is replaced with the target pixel to remove noise. To carry out this filtering process, sorting work for rearranging the numerical data sequence of pixels in order of density value is required, but sorting using the simple comparison method requires many calculation steps.

[0005]

In order to obtain the median value of a numerical data string, sorting is carried out and comparison operations are frequently used. In order to shorten the processing time, it is necessary to speed up the processing of this comparison operation.

An object of the present invention is to provide a method of calculating a median value of a numerical data string using SIMD operation which uses a parallel operation function of a processor in an operation for obtaining a median value from a numerical data string and accelerates the processing. That is.

[0007]

[Means for Solving the Problems] In order to achieve the above,
The present invention performs parallel processing on numerical data to be processed by using the SIMD operation function of the processor.

First, for each of the numerical data, a parallel comparison operation of the self data and all the remaining numerical data is performed.

Next, the respective comparison results are obtained. In this way, the comparison result obtained for each data, that is, the number of “data having an attribute value larger than that of its own data” is obtained. 2n-1
In the (n is a natural number) number of data strings, the number of "data having a larger attribute value than its own data" is n-1, which is the median value in the numerical data string.

When there are a plurality of data having the same size attribute value in the data string, there may be no data in which the number of "data having a larger attribute value than its own data" is n-1. In order to handle this exception, the number of “data whose attribute value is larger than its own data” is stored as a histogram, and the median value is obtained from this histogram.

The present invention can be summarized in the following configurations if it is summarized and summarized above.

(1) A method for calculating a median value of a numerical data string using SIMD operation, characterized in that the median value of a data string is obtained at high speed by parallel data comparison using SIMD operation.

(2) SIM is characterized in that after the comparison data is arranged in the histogram table, the median value of the data string is obtained by searching the histogram table.
A method of calculating the median value of a numerical data string using the D operation.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (Embodiment 1) FIG. 1 is a drawing best showing the features of the present invention. The device is a memory in which 101 input data strings are stored, and SIMs 102 and 105.
D instruction execution register, 103 storing result output memory pointer, 104, 106 SIMD
Comparison operation circuit, bistogram creation circuit 108, 109
Of the histogram table and the median value output circuit of 110, and the median value is output as 111.

The number of data in the data string for obtaining the intermediate value is 2n
-1 (n is a natural number). Each register is divided into 2n-2 blocks, and one data is stored in each block.

FIG. 2 is a flow chart of the present invention. First 201
Then, 2n-1 (n is a natural number) elements from the memory 101 are stored in the register 102. As shown in FIG. 4, the storage method is as shown in 202 for the two registers 102, and each data of the input data sequence is sequentially input once to the target data 401.
And stores it in all 2n−2 blocks of the register 403. The remaining data is the register 402
Store in each block of. Further, the pointer of the data string is stored in the register 103 for outputting the final result.

At 204, comparison of each block between the two registers 402 and 403 is performed by SIMD operation. The comparison result (whether the data is large or small) is stored in the register 105 in the form of true (1) or false (0). In order to obtain the number of true values out of the values of 105, in 204, data (unit vector) in which the numerical value 1 (true) is stored in each block is prepared in the register 405, and 105 and 4
The number is obtained by performing inner product calculation for each block of 05 by SIMD calculation.

At 206, if the "number of items for which the comparison result is true" is n-1, that is, there are n items that are larger (smaller) than the own data. 401 is the median value of the data string, 2
At 09, the result is output and the processing ends. Otherwise, 20
In step 7, a histogram is created for exception processing (when there is no data whose “number of comparison results is true” is n).

In the histogram, as shown in FIG. 3, "number of items for which the comparison result is true" is taken as an item on the horizontal axis, the pointer value 302 of the target data is obtained from 103, and the target data is collected as 303 into one 301 Are arranged in the histogram table 109.

Then, returning to 202, all the data in the data string are selected and compared as target data (20
Also in 3), when there is no n−1 “number of items for which the comparison result is true”, the histogram table 10 is set in 208.
9 is counted from the end, and the data 301 corresponding to the (n-1) th data is output at 209.

(Embodiment 2) Here, a data string of nine numbers {d0, d1, 2, d3, d4, d5, d6, d7, d.
The case where the median of 8} = {2,5,0,8,5,7,5,8,0} is obtained will be described. When the target data is set to d0, the vector sequence of the result of the comparison operation is {1, 0, 1,
1,1,1,1,0}, and the "number of objects for which the comparison result is true" is the unit vector {1,1,1,1,1,
The inner product with 1, 1, 1} results in 6.

Since "the number of items for which the comparison result is true" is not 4, the next target data is selected and the comparison operation is similarly performed. Hereinafter, “the number of items for which the comparison result is true” when selecting target data d1d2, ..., d8 is {(d2, d8): 7, d0: 6, (d1d
4, d6): 3 pieces, d5: 2 pieces, (d3, d7): 0
There is no target data for which the result is four.

FIG. 3 shows this as a histogram, which is arranged in the order of "the number of objects for which the comparison result is true", that is, in the order of data size. When searching from the end of the horizontal axis, the value corresponding to the fourth value is “the number of objects for which the comparison result is true” is 3, and the corresponding value is (d1, d4, d6), that is, the center. The value is the value 5 of the data at this time.

[0024]

As described above, according to the method of the present invention, the calculation amount for obtaining the median by the conventional method of comparing the values one by one is on the order of the square of the number n of data. On the other hand, it is possible to obtain in the order of n by simultaneously executing n−1 numerical operations.

[Brief description of drawings]

FIG. 1 is a block diagram of an arithmetic unit embodying the present invention.

FIG. 2 is a flow chart when a median value is obtained according to the present invention.

FIG. 3 is a view showing a bistogram table in Example 2 in a multiplex manner.

FIG. 4 is a diagram showing a data storage method of a data string and a flow of operations in the present invention.

[Explanation of symbols]

101 Memory for reading target data string 102 Register for executing SIMD comparison operation 103 Pointer register 104 for data used when outputting calculation result 104 SIMD comparison operation circuit 105 SIMD comparison operation result output register 106 “Block whose result is true” by SIMD operation Circuit 107 for obtaining the number of "blocks for which the result is true" 108 histogram generation circuit 109 histogram table 110 circuit for obtaining the median value from the histogram 111 register 301 for outputting the median value data for histogram placement ( Number of “blocks for which the result is true” + data address) 302 Target data pointer value 303 Target data value 401 Target data 402 102 Registers 403 102 registers (pair Register data values to) 405 stored in the respective blocks "result is block true" stores the number calculation data of

Claims (2)

[Claims] 1. A method for calculating a median value of a numerical data string using SIMD operation, characterized in that the median value of a data string is obtained at high speed by parallel data comparison using SIMD operation. 2. A method of calculating a median value of a numerical data string using SIMD arithmetic, characterized in that after arranging comparison data in a histogram table, the median value of the data string is obtained by searching the histogram table.