JPS6267676A - Vector data processor - Google Patents

Abstract

PURPOSE:To perform interelement arithmetic among vector data such as the inner product and total sum of the vector data without any deterioration in performance of a device by distributing and storing requests for the vector data in plural vector registers. CONSTITUTION:A control means 8 controls registers 1-4 and an input crossbar 7 with a normal vector instruction so that vector data inputted to the input crossbar is stored in one specified vector register. When the register is specified, on the other hand, the registers 1-4 and crossbar 7 are so controlled that elements of the vector data inputted to the crossbar 7 are inputted to and stored in a vector register (combination of registers 1 and 2 or 3 and 4) corresponding to the vector register specified with a vector instruction alternately. Consequently, processing including interelement arithmetic among the vector data such as the total sum and internal product of vectors is carried out by the combination of several instructions without any deterioration in the performance of a vector data processor.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a vector data processing device.

[Conventional technology]

Conventionally, vector data processing devices have been configured to include a plurality of vector registers and an input crossbar for selecting an arbitrary vector register to store vector data, and store one vector data in one vector register.

Since vector data is stored in one vector register, there is a drawback that inter-element operations within one vector data cannot be performed.

[Means for solving problems]

The vector data processing device of the present invention usually performs the following operations based on a specification made in an instruction issued before issuing a vector instruction that involves storing vector data in a vector register, or a specification made accompanying the vector instruction. The present invention is characterized by comprising a control means for distributing and storing elements of vector data stored in one vector register in a plurality of vector registers.Therefore, inter-element operations in vector data such as vector summation and inner product can be performed. By combining several vector instructions, it is possible to perform processing including the following without deteriorating the performance of the vector data processing device.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a main part of an embodiment of a vector data processing apparatus of the present invention, and FIG. 2 is a diagram for explaining the operation of the embodiment of FIG. 1.

This embodiment has four vector registers 1.2.8°4. The vector data output of vector register 1.2 is input to vector adder 5 and added, and the vector data resulting from the addition is selected by input crossbar 7 and stored in either vector register 1.2 or 8.4. It is configured like this. The vector data output of the vector register 3.4 is similarly input to the vector multiplier 6 and multiplied.
Vector data resulting from the multiplication is selected by the input crossbar 7 and stored in one of the vector registers 1.2.3.4. Furthermore, vector data from sources other than the vector calculator 5.6, such as the main memory, is also input to the input crossbar 7 and is selected as input 0, and stored in either of the vector registers 1.2 and 3.4. In a normal vector instruction, the control means 8 controls the input crossbar 7 for one specified vector register.
Vector registers 1.2.8.4. The input crossbar 7 is controlled and vector data is distributed and stored in vector registers, but if a designation is made, the vector register (vector register 1 and vector register 2 or vector register The vector registers 1.2.8.4 and the input crossbar 7 are controlled so that elements of the vector data input to the input crossbar 7 are alternately input and stored in the combination of the register 3 and the vector register 4.

Next, the operation of this embodiment will be described with reference to FIGS. 1 and 2, taking as an example the case of calculating the inner product of vectors.

Now, calculate the inner product of the two vectors with 8 elements.
= (aOr 'Im ''' * a5. a7 ) and B = (bOn ble..., b6. b7). First, vector data A.

B is stored in the vector register 8.4 as shown in FIG. 2(1). Next, if you issue a vector multiplication instruction that distributes and stores the operation results in vector registers 1.2,
The vector register 1.2 stores vector data as a result of the calculation, as shown in FIG. 2 (2). Next, when a vector addition instruction is issued twice to distribute and store the operation results in vector registers 1.2, vector registers 1.2
will be as shown in Figure 2 ■, (4J reference F).Finally, when a vector addition instruction is issued to store the operation result in vector register 1, the vector A will be added as shown in Figure 2 (5).
, find the dot product of B.

Kokote, obviously Kanayo Uni Figure 2 (21, (31, (41)
, (51 is an operation for calculating the sum of the elements of vector data. First, the vector for which the sum is to be calculated is separately stored in vector registers 1.2, and then the vectors are stored separately in vector registers 1.
Issue vector addition instructions that distribute and store the operation results in vector registers 1 and 2 until the number of elements of the operation results stored in vector registers 1 and 2 reaches 1, and finally issue a normal vector addition instruction with 1 element. The sum can be found. Furthermore, by replacing this operation with vector register 8.4 and a vector multiplication instruction, the line type of vector data can be determined.

In addition, in the operation of calculating the sum of these vector data and the line type, the number of elements of the vector data is 9m (the milk is a positive integer).
Otherwise, it is necessary to use values that do not affect the calculation results as elements of the vector data so that the number of additional elements becomes 2m, but the details thereof will be omitted. Although detailed explanations are omitted, as another example, a vector calculator that compares multiple input data and outputs a larger element or smaller element depending on the specification is used, and when calculating the sum as described above, The maximum value element or minimum value element of vector data can be determined by combining similar vector instructions.

〔Effect of the invention〕

As explained above, the present invention distributes and stores elements of vector data in a plurality of vector registers when storing vector data in a vector register, thereby performing inter-element operations in vector data such as the sum of inner products of vector data. There are some effects that can be done.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of a vector data processing apparatus of the present invention, and FIG. 2 is a diagram for explaining the operation of the embodiment of FIG. 1. 1.2. f3.4... Vector register, 5... Vector adder, 6... Vector multiplier, 7... Input crossbar, 8... Control means. Patent applicant: NEC Corporation Figure 2

Claims (1)

[Scope of Claim] A vector data processing device including a plurality of vector registers and an input crossbar for selecting and storing vector data in any of the vector registers, An element of vector data normally stored in one vector register can be transferred to a plurality of vector registers by a specification made in an instruction issued before a vector instruction is issued, or a specification made accompanying the vector instruction. 1. A vector data processing device characterized by comprising a control means for distributing and storing data into vector data.