JPH0364904B2 - - Google Patents

Description

[Detailed description of the invention]

[Overview] When calculating vector data by giving a vector built-in function, the range of the input vector data is scaled (range reduction), and the calculation result is restored to its original level (scaling). be exposed. Since the standardization and restoration are related, a table is created by arranging the respective magnification data, and one index is used to read both data at the same time, thereby reducing the number of table accesses. [Industrial Application Field] The present invention relates to a vector processing device, and particularly to a processing method for speeding up vector intrinsic function processing. [Prior Art] Vector computers provide frequently used functions such as logarithms as vector built-in functions. Processing of vector intrinsic functions is performed by approximate calculation, but in order to have sufficient calculation accuracy,
Calculations are performed after standardizing the input data (generally aligning the effective digits by moving digits) so that the range of change in value falls within a certain frame. Also for this reason,
When returning a calculation result, the opposite operation of scaling is performed. (Here, this reverse operation is called restoration). Conventionally, these scaling and restoration are performed using separate constant tables. A constant table divides the range of input data values into multiple levels, and registers the multiplication factor by which the input data should be multiplied for each level.The multiplication factor for each level is created based on the input data value. It is referenced using the index. FIG. 3 shows an example of a conventional vector intrinsic function processing method. A in FIG. 3 shows the flow of the process or represents the steps of the process. Further, b in FIG. 3 shows the flow of data, 31 to 34 are vector registers VR-1 to VR-4, respectively, 35 is a memory, 36 is a standardization constant table, 3
7 represents a constant table for restoration. Note that a and b in the figure correspond in the horizontal direction. Vector registers VR-1 to VR-4 are
Each has a capacity capable of storing a plurality of vector elements (assumed to be n). Each VR−
Blocks shown with diagonal lines within VR-1 to VR-4 illustrate the positions of the corresponding vector elements. Further, the standardization constant table 36 and the restoration constant table 37 are
It has a plurality of (m) entries each associated with a plurality (m) of indexes (that is, level ranges), and each entry is set with a magnification of the level to which the entry corresponds. Next, the following processing steps will be explained in order. For each vector element of the input vector data stored in VR-1, its mantissa part (the data is in floating point format)
According to whether the value of is included in one of a plurality of different predetermined level ranges,
Create an index vector for standardization and store (load) it in VR-2. Using the index vector for each vector element stored in VR-2, the scaling constant table 36 is referred to one after another, and the corresponding magnification is read out and stored (loaded) in VR-3. Scale each vector element of VR-1 using the scaling factor of VR-3. Perform approximate calculation of the objective function using the scaled vector elements. Using the VR-2 index vector,
Refer to the restoration constant table 37, read out the magnification for each vector element, and
Store (load) in. The approximate calculation results using the VR-4 magnification are restored. In the above operations, the operation of reading data from the constant table in memory using the index vector and storing it in the vector register is called indexed vector load, and the instruction that instructs this operation is This is called a vector load instruction. [Problems to be solved by the invention] In vector built-in function processing in conventional vector computers, two constant tables are used for standardization and restoration of vector data, and indexed vector load instructions are issued to each of them. The magnification data was read out. Since the indexed vector load instruction is issued twice, there is a problem in that it takes time to execute and reduces the processing speed. [Means for Solving the Problems] The present invention aims to reduce the number of times indexed load instructions are used in vector intrinsic function processing. In the present invention, there is usually a correlation between the two operations of standardization of input data and restoration of the result of functional processing, and the magnification (constant) used in both operations is determined depending on the value of the input data. This is done with a focus on things that are uniquely determined. As a result, the standardization constant table and the restoration constant table are combined in memory into a single constant table, and the standardization magnification data and restoration magnification data corresponding to one index are stored in one entry. Both magnification data are read out simultaneously by one indexed vector load instruction, and then either one can be referenced at high speed as needed. FIG. 1 is a conceptual diagram for explaining the principle of the present invention. In the figure, 13 to 15 are vector registers VR-3 to VR-5, 16 is memory, 17 is a constant table, 18 is an entry, and 19
20 represents scaling factor data for standardization, 20 represents scaling factor data for restoration, 21 represents standardization processing, 22 represents function processing, and 23 represents restoration processing. Vector registers VR-3 to VR-5 are
Each has a capacity that can accommodate multiple vector elements. The constant table 17 has a plurality of entries 1 each having an index corresponding to a different level.
8, and any entry 18 includes:
The information includes scaling factor data 19 for standardization and scaling factor data 20 for restoration that are related to each other. In operation, an index vector is first created based on the values of vector elements of input vector data, and constant table 17 in memory 16 is accessed using an indexed vector load instruction. The entry with that index is read from the constant table 17, and the information is stored as a vector.
Stored (loaded) in register VR-3. In this way, vector register VR-3 contains
The related scaling factor and restoration scaling factor data for each vector element are sequentially set as a pair. Next, the data set in the vector register VR-3 is read out, and the scaling factor data therein is extracted and stored in the vector register VR-4. Extract the magnification data for restoration in the same way,
Store in vector register VR-5. Next, each vector element of the input vector data is standardized in a scaling process 21 using each scaling factor data of VR-4, and then a function calculation is performed in a function process 22. Then, the calculation result is used in the restoration process 23,
Restoration is performed using each restoration magnification data of VR-5 and output. [Operation] According to the present invention, each magnification used for standardization and restoration is combined for each index, stored in a single constant table, and read out together when the table is referenced. Therefore, it is sufficient to issue one indexed vector load instruction for one index vector, and the number of memory accesses can be reduced to one. Since each scaling factor for scaling and restoration is performed based on data transfer by bit manipulation between vector registers, faster processing is possible than constant table access in memory. In addition, in normal cases, comparatively simple numerical values are given to the scaling factor for standardization and the scaling factor for restoration, and each can usually be displayed in the upper double digits of the data. can be compressed into the upper half of each, thereby avoiding the need to increase the length of the entries in the constant table. For example, if the data for each magnification for standardization and restoration in the conventional method is 8 bytes long, each of these data is compressed to 4 bytes long and combined, and the entries in the constant table are made 8 bytes long. It can be suppressed. [Example] Next, the details of the present invention will be explained based on an example. FIG. 2 is a flowchart illustrating one embodiment of the present invention. A in FIG. 2 shows the flow of the process, and ' to ' represent the steps of the process. Also, b in Figure 2
indicate the data flow, and 11 to 15 are vector registers VR-1 to VR-5, respectively.
16 is a memory, 17 is a constant table in which scale data for standardization and restoration corresponding to the index is registered, 18 is an entry each associated with an index of a predetermined specific level, and 19 is a standard. 20 is magnification data for restoration. Note that FIG. 2 is shown in contrast to FIG. 3 for the conventional system. Next, the processing flow will be explained in order. ' Create a level index corresponding to the value of the mantissa for each vector element of vector register VR-1 in which input vector data is stored, and
Store in VR-2. ' Issue a load with index instruction using the index of vector register VR-2, refer to constant table 17 on memory 16, read the corresponding entry, and store the information in vector register VR-3 ( load). ' Extract scaling factor data from the information in vector register VR-3 and store it in vector register VR-4. Similarly, restoration magnification data is extracted from the information in vector register VR-3 and stored in vector register VR-5. ' Standardize the level of the input vector data of vector register VR-1 using the standardization magnification data of vector register VR-3. ′ Perform functional calculations using scaled vector data. ′ For the calculation result, vector register
VR-5 restoration magnification data is applied to restore and output the level. Next, as an example of vector intrinsic function processing,
The case of a logarithmic function will be explained. If p is an arbitrary exponent and q is a value of 1/16<q<1, and the variable x is expressed in hexadecimal in floating point format, x=16 ^p ·q (1). Here, if the natural logarithm of x (base = e) is expressed as y, then y=log _e X=log _e 2・log ₂ x=log _e 2・log ₂ (16 ^p・q
)=log _e 2・(log ₂ 16 ^p +log ₂ q)=log _e 2・(4p+
log ₂ q) ......(2). Add to this the standardization magnification: 1/4≦q・ω<
Applying the value ω of 1/2, y=log _e 2・(4p+log ₂ (q・ω)−log ₂ ω)
......(3) is obtained. In equation (3), the value of q·ω can be suppressed within a certain level range by appropriately setting the value of the scaling factor ω with respect to the level of the value of q. Further, log ₂ ω represents a restoration magnification, and is uniquely determined by the value of the standardization magnification ω. A specific example of a constant table is shown in the table below.

〔Effect of the invention〕

According to the present invention, it is possible to shorten the processing time of a vector built-in function in a vector processing device and improve processing performance.

[Brief explanation of drawings]

FIG. 1 is a conceptual diagram showing the principle of the present invention, FIG. 2 is a flow chart of one embodiment of the present invention, and FIG. 3 is a flow chart of a conventional example. In Figure 1, 13: Vector register VR-3,
14: Vector register VR-4, 15: Vector register VR-5, 16: Memory, 17: Constant table, 18: Entry, 19: Magnification data for scaling, 20: Magnification data for restoration, 21: Standardization processing , 22: Function processing, 23: Restoration processing.

Claims (1)

[Claims] 1. In a vector processing device having a vector built-in function, scaling factor data 19 for normalizing the value of input vector data for vector built-in function processing within a predetermined level range; , and restoration magnification data 20 for restoring the result of vector built-in function processing to the level before standardization, are combined by a common index.
7 is provided on the memory 16, and when processing the vector built-in function, an index is created from the input vector data, the created index is used to refer to the constant table 17 on the memory 16, and the corresponding scaling factor data 19 is set.
and restoration magnification data 20 are simultaneously read out and stored in the first vector register 13. Next, the standardization magnification data and restoration magnification data are read out from the first vector register 13, and are stored in the second vector register 13. After performing standardization processing 21 of the input vector data by applying the standardization magnification data stored in the second vector register 14, The function processing 22 is performed, and the restoration processing 23 is performed to restore the level before standardization by applying the restoration magnification data stored in the third vector register 15 to the processing result. Vector built-in function processing method.