Multithreading in vector processors

• US 9,870,340 B2
• Filed: 03/30/2015
• Issued: 01/16/2018
• Est. Priority Date: 03/30/2015
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Evangelinos, Constantinos, Kim, Changhoan, Nair, Ravi
• Primary Examiner(s): Giroux, George
• Application Number:
  US14/672,568
• Publication Number:
  US 20160292128A1
• Time in Patent Office: 1,023 Days
• Field of Search: None
• CPC Class Codes: G06F 15/8053 Vector processors
• CPC Class Codes: G06F 15/8076 Details on data register access
• CPC Class Codes: G06F 15/8084 Special arrangements thereof, e.g. mask or switch
• CPC Class Codes: G06F 9/30036 Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
• CPC Class Codes: G06F 9/30189 according to execution mode, e.g. mode flag
• CPC Class Codes: G06F 9/321 Program or instruction counter, e.g. incrementing
• CPC Class Codes: G06F 9/3851 from multiple instruction streams, e.g. multistreaming

Abstract

In one embodiment, a system includes a processor having a vector processing mode and a multithreading mode. The processor is configured to operate on one thread per cycle in the multithreading mode. The processor includes a program counter register having a plurality of program counters, and the program counter register is vectorized. Each program counter in the program counter register represents a distinct corresponding thread of a plurality of threads. The processor is configured to execute the plurality of threads by activating the plurality of program counters in a round robin cycle.

Claims

1. A system comprising:

a processor having a vector processing mode and a multithreading mode, the processor being a vector processor and being configured to operate on one thread per cycle in the multithreading mode, and the processor comprising;

one or more program counter registers together comprising a plurality of program counters, each program counter register of the one or more program counter registers being vectorized into a corresponding subset of the plurality of program counters, and each program counter in the plurality of program counters of one or more program counter registers representing a distinct corresponding thread of a plurality of threads;

wherein the number of threads in the plurality of threads is limited by the number of program counters in the plurality of program counters of the one or more program counter registers;

the processor configured to execute the plurality of threads by activating the plurality of program counters in a round robin cycle;

an instruction buffer comprising a plurality of instructions;

wherein a first program counter in the plurality of program counters references a first instruction in the instruction buffer for execution by the processor in a first thread of the plurality of threads;

wherein a second program counter in the plurality of program counters references a second instruction in the instruction buffer for execution by the processor in a second thread of the plurality of threads, the first instruction being different from the second instruction;

wherein the processor is configured to skip, in the round robin cycle, a program counter lacking readiness based on a ready bit setting and determine a validity of each data element to be used in an operation for an active thread of the plurality of threads, said validity indicative of data being fetched from memory to be loaded into a data element, wherein responsive to a determination of an invalid operation, proceed to a next operation without incrementing the second program counter, and responsive to no invalid operation, process one or more output vectors and output a result.

2. The system of claim 1, the processor further comprising:

a branch operator configured to modify the plurality of program counters in the program counter registers; and

a mask register comprising a plurality of masks, wherein each mask in the mask register corresponds to a program counter in the one or more program counter registers;

wherein the branch operator is configured to modify, in the round robin cycle, each program counter in the program counter registers based on the value of the corresponding mask in the mask register.

3. The system of claim 1, the processor further comprising:

a readiness operator configured to determine validity of a plurality of data elements in a vector related to an operation;

wherein an active program counter is skipped, responsive to a data element in the vector being invalid.

4. The system of claim 1, the processor further comprising:

a synchronization register configured to indicate whether the plurality of threads have performed an initial task, wherein each of the plurality of threads is configured to set a corresponding bit in the synchronization register responsive to performing the initial task; and

wherein the plurality of threads are configured to pause their execution until the corresponding bits to the plurality of threads are set.

5. The system of claim 1, wherein the processor is configured to switch between the multithreading mode and the vector processing mode.

6. A computer-implemented method, comprising:

storing, via a vector processor having a vector processing mode and a multithreading mode and configured to operate on one thread per cycle in the multithreading mode, in one or more program counter registers together comprising a plurality of program counters;

vectorizing one or more program counter registers that together comprise a plurality of program counters, into a corresponding subset of the plurality of program counters, wherein each program counter in the plurality of program counter registers represents a distinct corresponding thread of a plurality of threads;

wherein the number of threads in the plurality of threads is limited by the number of program counters in the plurality of program counters of the one or more program counter registers;

executing, by a processor, the plurality of threads by activating the plurality of program counters in a round robin cycle;

referencing, via a first program counter in the plurality of program counters, a first instruction in an instruction buffer for execution by the processor in a first thread of the plurality of threads, wherein the instruction buffer comprises a plurality of instructions;

referencing, via a second program counter in the plurality of program counters, a second instruction in an instruction buffer for execution by the processor in a second thread of the plurality of threads, the first instruction being different from the second instruction; and

skipping, via the processor, in the round robin cycle, a program counter lacking readiness based on a ready bit setting and determine a validity of each data element to be used in an operation for an active thread of the plurality of threads, said validity indicative of data being fetched from memory to be loaded into a data element, the skipping comprising;

responsive to a determination of an invalid operation, proceeding to a next operation without incrementing the second program counter, and

responsive to no invalid operation, processing one or more output vectors and output a result.

7. The method of claim 6, further comprising a mask register comprising a plurality of masks, wherein each mask in the mask register corresponds to a program counter in the plurality of program counter registers, the method further comprising:

modifying, in the round robin cycle, each program counter in the plurality of program counter registers based on a value of a corresponding mask in the mask register.

8. The method of claim 6, further comprising:

determining validity of a plurality of data elements in a vector related to an operation; and

skipping an active program counter, responsive to a data element in the vector being invalid.

9. The method of claim 6, further comprising:

setting, by each of the plurality of threads, an assigned bit in a synchronization register responsive to performing an initial task

pausing, by each of the plurality of threads, execution until a plurality of assigned bits for the plurality of threads are set in the synchronization register.

10. The method of claim 6, wherein the computer processor is configured to switch between the multithreading mode and the vector processing mode.

11. A computer program product for multithreading in a vector processor, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a vector processor having a vector processing mode and a multithreading mode and configured to operate on one thread per cycle in the multithreading mode, to cause the vector processor to perform a method comprising:

storing, in one or more program counter registers together comprising a plurality of program counters;

vectorizing one or more program counter registers that together comprise a plurality of program counters, into a corresponding subset of the plurality of program counters, wherein each program counter in the plurality of program counter registers represents a distinct corresponding thread of a plurality of threads;

wherein the number of threads in the plurality of threads is limited by the number of program counters in the plurality of program counters of the one or more program counter registers;

executing the plurality of threads by activating the plurality of program counters in a round robin cycle;

referencing, via a first program counter in the plurality of program counters, a first instruction in an instruction buffer for execution by the processor in a first thread of the plurality of threads, wherein the instruction buffer comprises a plurality of instructions;

referencing, via a second program counter in the plurality of program counters, a second instruction in an instruction buffer for execution by the processor in a second thread of the plurality of threads, the first instruction being different from the second instruction; and

skipping, in the round robin cycle, a program counter lacking readiness based on a ready bit setting and determine a validity of each data element to be used in an operation for an active thread of the plurality of threads, said validity indicative of data being fetched from memory to be loaded into a data element, the skipping comprising;

responsive to a determination of an invalid operation, proceeding to a next operation without incrementing the second program counter, and

responsive to no invalid operation, processing one or more output vectors and output a result.

12. The computer program product of claim 11, wherein a mask register comprises a plurality of masks, and wherein each of the plurality of masks in the mask register corresponds to a program counter in the plurality of program counter registers, the method further comprising:

modifying, in the round robin cycle, each program counter in the program counter registers based on a value of a corresponding mask in the mask register.

13. The computer program product of claim 11, the method further comprising:

determining validity of a plurality of data elements in a vector related to an operation; and

skipping an active program counter, responsive to a data element in the vector being invalid.

14. The computer program product of claim 11, the method further comprising:

setting, by each of the plurality of threads, an assigned bit in a synchronization register responsive to performing an initial task

pausing, by each of the plurality of threads, execution until a plurality of assigned bits for the plurality of threads are set in the synchronization register.

15. The computer program product of claim 11, wherein the processor is configured to switch between the multithreading mode and the vector processing mode.