Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
  • G06F9/30003—Arrangements for executing specific machine instructions
  • G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
  • G06F9/30036—Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
  • G06F9/30038—Instructions to perform operations on packed data, e.g. vector, tile or matrix operations using a mask
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
  • G06F9/30003—Arrangements for executing specific machine instructions
  • G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
  • G06F9/3001—Arithmetic instructions
  • G06F9/30014—Arithmetic instructions with variable precision
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
  • G06F17/10—Complex mathematical operations
  • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
  • G06F9/30003—Arrangements for executing specific machine instructions
  • G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
  • G06F9/3001—Arithmetic instructions
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
  • G06F9/30003—Arrangements for executing specific machine instructions
  • G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
  • G06F9/30036—Instructions to perform operations on packed data, e.g. vector, tile or matrix operations

Definitions

• This disclosure relates to the field of data processing systems. More particularly, this disclosure relates to data processing systems supporting vector arithmetic instructions.
• a vector arithmetic instruction may take two such vector input operands and perform an arithmetic operation specified by the instruction upon respective pairs of vector elements from within the two vector operands.
• Vector processing facilitates parallel computation.
• processing circuitry to perform processing operations; and decoder circuitry to decode program instructions to generate control signals to control said processing circuitry to perform said processing operations;
• said decoder circuitry is responsive to at least one mixed-element- sized vector arithmetic instruction to generate control signals to control said processing circuitry to perform an arithmetic processing operation upon a first vector of first source operand elements of a first bit size and a second vector of second source operand elements of a second bit size, said second bit size being greater than said first bit size.
• processing means for performing processing operations; and decoder means for decoding program instructions to generate control signals to control said processing circuitry to perform said processing operations;
• said decoder means is responsive to at least one mixed-element- sized vector arithmetic instruction to generate control signals to control said processing means to perform an arithmetic processing operation upon a first vector of first source operand elements of a first bit size and a second vector of second source operand elements of a second bit size, said second bit size being greater than said first bit size.
• decoding at least one mixed-element- sized vector arithmetic instruction to generate control signals to control processing circuitry to perform an arithmetic processing operation upon a first vector of first source operand elements of a first bit size and a second vector of second source operand elements of a second bit size, said second bit size being greater than said first bit size.
• Figure 1 schematically illustrates the data processing system supporting vector arithmetic instructions
• Figure 2 schematically illustrates the operation and syntax of a mixed-element-sized vector arithmetic instruction
• Figure 3 schematically illustrates how a given bit in a destination operand of an arithmetic instruction is dependent upon input operand bits taken from different bit positions
• Figure 4 is a flow diagram schematically illustrating use of a mixed-element-sized vector arithmetic instruction
• Figure 5 schematically illustrates a virtual machine implementation.
• Figure 1 schematically illustrates a data processing system 2 comprising a processor 4 coupled to a memory 6.
• the memory 6 stores data values 8 to the manipulated and program instructions 10 specifying data processing operations to be performed by the processor 4.
• Program instructions 10 are fetched by an instruction fetch unit 12 and passed to decoder circuitry 14.
• the decoder circuitry 14 generates control signals 16 which serve to control processing circuitry 18, 20, 22, 24 within the processor 4 to perform processing operations specified by the decoded
• the processor 4 supports both vector operations and scalar operations.
• Vector processing circuitry 18 acting upon vector operands stored within a vector register file 22 serves to perform vector processing operations.
• Scalar processing circuitry 20 acting upon scalar operand values stored within a scalar register file 24 serves to perform scalar processing operations.
• vector processing operations can take a variety of different forms. For example, SIMD (Single Instruction Multiple Data) processing operations are one form of vector processing operation. More generally, a vector processing operation is performed upon a plurality of vector elements which together formed a vector operand.
• a vector operand of 256 bits in length may be formed of 16 vector elements each of 16 bits of length.
• the processing operations performed upon the individual vector elements will typically be performed at least partially in parallel, but this need not always be the case.
• Scalar processing operations are performed with input operands comprising a single value, e.g. 64-bit scalar processing operations may be performed upon two 64-bit scalar input operands and generate a 64-bit scalar output operand.
• the vector processing circuitry 18 is formed in accordance with the present disclosure to support at least one mixed-element-sized vector arithmetic instruction performed under control of control signal 16 generated by decoder circuitry 14 when a mixed-element- sized vector arithmetic instruction is fetched, decoded and executed.
• Figure 2 schematically illustrates the operation and assembler syntax of a mixed-element- sized vector arithmetic instruction.
• the mixed-element-sized vector arithmetic instruction performs an arithmetic processing operation, such as a logical shift, a division, or a compare, upon a first vector of first source operand elements ai.
• an arithmetic processing operation such as a logical shift, a division, or a compare
• first vector of first source operand elements ai there are sixteen first source operand elements ao-a 15 .
• Each of these first source operand elements ai has a bit size of A.
• 16 first source operand elements ai are arranged into 4 disjoint subsets ao_a 3 , a4-a 7 , a 8 -an, and a 12 - a 15 .
• the mixed-element- sized vector arithmetic instruction has a second vector of second source operand elements bi.
• the second vector of second source operand elements is formed of four second source operand elements namely second source operand elements bo-b 3 , each of bit size B.
• first source operand elements and four second source operand elements there are sixteen first source operand elements and four second source operand elements, more generally there may be N first source operand elements and M and second source operand elements.
• the second source operand elements have a greater bit size B than the first source operand elements bit size A.
• the ratio of the bit size B of the second source operand elements to the bit size A of the first source operand elements may be 4: 1. This is the same as the ratio of the number of first source operand elements N to the number of second source operand elements M.
• each of the disjoint subsets of first source operand elements is subject to an arithmetic processing operation with a respective one of the second source operand elements being the second operand input. More particularly, first source operand elements ai-a 3 are subject to an arithmetic processing operation with the second source operand element bo being the second input to that arithmetic processing operation. For example, each of the first source operand elements ao-a 3 may be separately subjected to a logical shift right by a shift amount specified by the second source operand element bo. The resulting output operand has the same bit size A as the first source operand elements.
• arithmetic processing operations may be performed for the mixed-element-sized vector arithmetic instruction e.g. a first source operand element ai being divided by a corresponding second source operand element b j or a first source operand element ai being subjected to a compare (subtract) with a corresponding second source operand element b j .
• the arithmetic processing operation such as a logical shift, divide or compare is specified by a mnenonic such as LSR, DIV or CMP.
• the instruction specifies an element size for the first source operand elements, namely whether these are bytes B, half words. H, words W, or double words D. In the case of 64-bit double words D, a 512 bit vector operand would contain 8 such double words.
• the second source operand elements in such a case may be for example, 128-bit vector elements with two of the double word first source operand elements being associated with each of the 128-bit second source operand elements. It will be understood that many other different total vector sizes and vector element sizes may be employed depending upon the implementation and the requirements of the particular instruction set or the processing performed.
• the syntax of the mixed-element-sized vector arithmetic instructions continues by specifying the destination vector register Z D together with its element bit size A. This is followed by specifying the vector register of the first source operand namely Z sl together with its element bit size A. Finally, the vector register Zs 2 of the second source operand together with their element bit size B is specified. This is an example of a mixed-element-sized vector arithmetic instruction having two input operands and one output operand.
• two input operands may be specified with the result being written to a predicate register P D which contains a "true” or a "false” result corresponding to the result of comparing each element in Zsi with the wider elements in Zs 2 (subtraction).
• the processing operation is a shift operation that shifts a first source operand element by a shift amount specified by a corresponding second source operand element.
• the arithmetic processing operation is a division operation that divides a first source operand element by a divisor specified by the corresponding second source operand element.
• the arithmetic processing operation is a compare operation that compares a first source operand element with a corresponding second source operand element.
• the element bit size of the first source operand elements is specified by an element size field "x", as mentioned above, which specifies whether the first source operand elements have the size of 8 bits, 16 bits, 32 bits or 64 bits.
• the second source operand elements may in some example embodiments have a bit size specified by a field within the mixed-element- sized instruction.
• the second source operand elements may have a fixed size, such as 64 bits or 128 bits.
• the first source operand elements are smaller in bit size then the second source elements. If the second source operand elements have a bit size of 64, then the first source operand elements will have a bit size of one of 8, 16 or 32.
• Figure 3 schematically illustrates how an arithmetic instruction operates such that bit values
• 26 within a destination operand element are set in dependence upon one or more bit values of different bit significance within at least one of a corresponding first operand element 28 and second source operand element 30.
• respective bit positions 26 in the destination operand Di are dependent upon corresponding bits 29 having a higher order significance within the first source operand element 28 together with all of the bits 31 within the second source operand 30 which specify the right shift amount to be applied.
• the present disclosure teaches a system in which mixed-element-sized vector arithmetic instructions are supported. This is counter to the normal technical prejudice in this field. Normally all source operands for an arithmetic instruction all have a common element size. The present disclosure recognises that in certain circumstances the provision of mixed-element- sized vector arithmetic instructions provides advantages which justify the instruction bit space that such instructions consume within the instruction sets supported by the processor 4 and decoded by the decoder circuitry 14.
• mixed-element- sized vector arithmetic instructions may be used to store copies of that scalar operand not sharing the same bit size as the first source operand elements for respective operations of the loop.
• the provision of mixed- element- sized vector arithmetic instructions may avoid the need to copy a scalar operand from the scalar register file 24 to the vector processing circuitry 18 upon each loop iteration, which can be a relatively slow process compared to the vector processing circuitry 18 accessing the vector register file 22 to which it is more directly coupled.
• one use of the mixed-element-sized vector arithmetic instructions provided by the present disclosure is to copy a scalar operand value which contains more significant bits than the first source vector elements from the scalar register file 24 to each of the wider second source operand elements within a second vector operand.
• the second vector can then be used as one of the vector operand inputs to a mixed-element-sized vector arithmetic instruction which operates on a plurality of first source vector operand elements using the wider second source operand elements into which the scalar operand has been copied.
• FIG. 4 is a flow diagram schematically illustrating one example of the above type of operation.
• a 64-bit value from a scalar register Xi is directly copied to all vector register elements of vector operand Zs 2 .
• Step 34 then duplicates this 64-bit value from the vector register element to which is has been copied across all of the other vector register elements of the second vector.
• a processing loop which is to be performed is entered.
• Step 38 executes mixed- element-sized vector arithmetic instructions on a first input vector Zsi and a second input vector Z s2 to generate an output vector Z D .
• the multiple copies of the scalar register are present within appropriate elements of the second vector and accordingly do not need to be moved from the scalar register file 24 to the vector register file 22 as the processing proceeds.
• the looping executed is exited.
• first source operand elements ai and second source operand elements b j i.e. a disjoint subset arrangement whereby each disjoint subset of the first source operand elements has a single associated second source operand element. It will be appreciated that in some embodiments such a division of the first source operand elements and association with the second source operand elements need not be provided.
• the above example discusses arithmetic processing operations in the form of shift operations, division operations and compare operations. It will be appreciated that other forms of arithmetic operation may also be supported.
• the operands acted upon by the arithmetic operations may be binary number values where each bit has different positional significance such that the first source operand elements contain binary number values, the second source operand elements contain binary number values and the destination operand elements contain destination operand binary number values.
• the processing operation is performed on the binary number values as a whole rather than independent bitwise operations performed independently of other bits at different positions (such as OR, XOR or AND).
• Figure 5 illustrates a virtual machine implementation that may be used. Whilst the earlier described embodiments implement the present invention in terms of apparatus and methods for operating specific processing hardware supporting the techniques concerned, it is also possible to provide so-called virtual machine implementations of hardware devices. These virtual machine implementations run on a host processor 530 running a host operating system 520 supporting a virtual machine program 510. Typically, large powerful processors are required to provide virtual machine implementations which execute at a reasonable speed, but such an approach may be justified in certain circumstances, such as when there is a desire to run code native to another processor for compatibility or re-use reasons.
• the virtual machine program 510 provides an application program interface to an application program 500 which is the same as the application program interface which would be provided by the real hardware which is the device being modelled by the virtual machine program 510.
• the program instructions including the control of memory accesses described above, may be executed from within the application program 500 using the virtual machine program 510 to model their interaction with the virtual machine hardware.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Software Systems (AREA)
• General Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• Mathematical Physics (AREA)
• Mathematical Optimization (AREA)
• Pure & Applied Mathematics (AREA)
• Mathematical Analysis (AREA)
• Computational Mathematics (AREA)
• Data Mining & Analysis (AREA)
• Computing Systems (AREA)
• Algebra (AREA)
• Databases & Information Systems (AREA)
• Complex Calculations (AREA)
• Advance Control (AREA)
• Executing Machine-Instructions (AREA)

Priority Applications (6)

Applications Claiming Priority (2)

Publications (1)

Family

ID=54062956

Family Applications (1)

Country Status (9)

Cited By (2)

Families Citing this family (5)

Citations (3)

Family Cites Families (17)

• 2015
  • 2015-07-31 GB GB1513511.4A patent/GB2540943B/en active Active
• 2016
  • 2016-06-23 JP JP2018503593A patent/JP7071913B2/ja active Active
  • 2016-06-23 WO PCT/GB2016/051868 patent/WO2017021681A1/en not_active Ceased
  • 2016-06-23 KR KR1020187003580A patent/KR102584001B1/ko active Active
  • 2016-06-23 US US15/743,745 patent/US11003447B2/en active Active
  • 2016-06-23 EP EP16732707.1A patent/EP3329363B1/en active Active
  • 2016-06-23 CN CN201680043340.XA patent/CN107851016B/zh active Active
  • 2016-07-19 TW TW105122689A patent/TWI739754B/zh active
• 2017
  • 2017-12-31 IL IL256663A patent/IL256663B/en active IP Right Grant

Patent Citations (3)

Also Published As

Similar Documents

Legal Events