US 20050060517 A1
In one embodiment, the present invention includes a method to determine whether execution of an instruction of a first thread may require a long latency and switch to a second thread if the instruction may require the long latency. In certain embodiments, at least one additional instruction may be executed in the first thread while preparing to switch threads.
1. A method comprising:
determining whether execution of an instruction of a first thread may require a long latency; and
switching to a second thread if the instruction may require the long latency.
2. The method of
3. The method of
4. The method of
5. The method of
6. The method of
7. The method of
8. A method comprising:
switching from a first thread to a second thread if a condition that may result in a stall of a processor pipeline occurs during execution of the first thread in the processor pipeline.
9. The method of
10. The method of
11. The method of
12. The method of
13. An article comprising a machine-readable storage medium containing instructions that if executed enable a system to:
switch from a first thread to a second thread if a condition that may result in a stall of a processor pipeline occurs during execution of the first thread in the processor pipeline.
14. The article of
15. The article of
16. The article of
17. An apparatus comprising:
a processor pipeline having a feedback loop to provide a feedback signal to cause the processor pipeline to switch from a first thread to a second thread, the feedback signal to originate from a location in the processor pipeline before instruction execution.
18. The apparatus of
19. The apparatus of
20. The apparatus of
21. The apparatus of
22. A system comprising:
a processor pipeline having a feedback loop to provide a feedback signal to cause the processor pipeline to switch from a first thread to a second thread, the feedback signal to originate from a location in the processor pipeline before instruction execution; and
a wireless interface coupled to the processor pipeline.
23. The system of
24. The system of
25. The system of
26. The system of
27. The system of
28. The system of
29. The system of
Processors, such as modern high-performance processors, are designed to execute a large number of instructions per clock cycle. Certain instructions produce a result only after a potentially large number of cycles. Such instructions may be known as “long latency” instructions, as a long time interval exists between the time an instruction is delivered and when it is executed. A long latency may occur, for example, when data required by an instruction needs to be loaded from a high level of memory. Such a load operation therefore may have a “load-use” penalty associated with it. That is, after a program issues such a load instruction, the data may not be available for multiple cycles, even if the data exists (i.e., “hits”) in a cache memory associated with the processor.
Processors typically allow execution to continue while a long latency instruction is outstanding. Often, however, data is needed relatively soon (e.g., within several clock cycles) because insufficient work remains to be done by the processor without the requested data. Accordingly, a need exists to improve processor performance in such situations.
While the type of processor which includes a pipeline in accordance with an embodiment of the present invention may vary, in one embodiment the processor may be a relatively simple in-order processor. In one embodiment, the processor may have a reduced instruction set computing (RISC) architecture, such as an architecture based on an Advanced RISC Machines (ARM) architecture. For example, in one embodiment a 32-bit version of an INTEL® XSCALE™ processor available from Intel Corporation, Santa Clara, Calif. may be used. However, in other embodiments the processor may be a different processor.
In one embodiment, I-cache 110 may be coupled to receive instructions from a bus interface unit of the processor. I-cache 110 may be used to store instructions, including instructions of multiple threads of a program. I-cache 110 may be coupled to provide instructions to instruction fetch unit 120. Alternately, instruction fetch unit 120 may receive instructions from a fill buffer (which may be within reservation and execution unit 150). Instruction fetch unit 120 may include, in certain embodiments, program counters for each thread to be executed on the processor, along with logic to sequence between the threads. In an embodiment in which out-of-order processing is implemented, instruction fetch unit 120 may include a branch target buffer that may be used to examine instructions fetched from I-cache 110 to determine whether any branching conditions exist.
As shown in
Still referring to
From register lookup 140, instructions may then proceed to reservation and execution unit 150 for scheduling execution of the instructions in the execution unit (or units) of the processor (e.g., an integer and/or floating point arithmetic logic unit (ALU)). In one embodiment, multiple execution units may be present. Such execution units may include a main execution pipeline, a memory pipeline and a multiply-accumulate (MAC) pipeline. In such an embodiment, the main execution pipeline may perform arithmetic and logic operations, as required for data processing instructions and load/store index calculations, and may further determine conditional instruction execution. The memory pipeline may include a data cache unit to handle load and store instructions. The MAC pipeline may be used to perform multiply and multiply-accumulate instructions.
The above-described flow through pipeline 100 may describe normal operational flow. Such flow may occur when instructions do not require long latencies prior to execution, or in the absence of other conditions that may lead to processor stalls.
However in accordance with various embodiments of the present invention, when certain predetermined events or conditions occur, such as an instruction that may require a long latency prior to execution, a feedback loop 125 from instruction decode unit 130 to instruction fetch unit 120 may be activated to cause instruction fetch unit 120 to prepare to switch threads and accordingly fetch instructions for the new thread.
In one embodiment, identifications of predetermined conditions may be stored in instruction decode unit 130, although the scope of the present invention is not limited in this regard. For example, the identifications may be stored in a lookup table or other storage within instruction decode unit 130. In this manner, when an instruction is received by instruction decode unit 130, it may be analyzed against entries in the lookup table to determine whether the instruction corresponds to one of the predetermined conditions. If so, a feedback signal on feedback loop 125 may be activated. Alternately, logic in instruction decode unit 130 may be used to detect for the presence or occurrence of a predetermined condition. In still other embodiments, microcode in instruction decode unit 130 may determine the presence or occurrence of a predetermined condition.
While the predetermined conditions may vary in different embodiments, an instruction that may require a long latency prior to execution may be considered to be a predetermined condition. As used herein, the term “long latency” means a time period between receipt of an instruction and execution of the instruction that causes the processor to suffer one or more stalls. Thus a long latency period may be several cycles or may be hundreds of cycles, depending on the ability of the processor to perform other instructions in the latency period. For example, a load instruction that requires the obtaining of information from system memory (e.g., as on a cache miss) may require hundreds of cycles, while a load instruction that obtains data from a cache (such as a level 1 (L1) cache) closely associated with the processor may require fewer than ten cycles. In certain embodiments, both load instructions may be considered long latency instructions, as a processor may suffer stall cycles before data is ready for processing.
Thus a load instruction is an example instruction that may cause a long latency and thus may be considered a predetermined condition. While the latency caused by a load instruction may vary depending on the level of memory at which the data is obtained, the presence of a load instruction itself, regardless of actual latency, may be sufficient to cause a feedback signal in accordance with an embodiment of the present invention. In other words, a feedback signal in accordance with an embodiment of the present invention may be based on stochastic models. That is, the predetermined conditions may be selected based on a knowledge that the conditions may, but need not necessarily, cause a latency that leads to pipeline stalls.
Other examples of instructions that may be considered to be a predetermined condition may include store instructions and certain arithmetic instructions. For example, a floating point divide operation may be a condition that causes a feedback signal. In addition, other operations which require accessing a memory subsystem may be a condition causing a feedback signal to be initiated.
When a predetermined condition is detected, instructions of another thread may be fetched and executed so that few or no stall cycles occur. Thus in certain embodiments, performance may be significantly increased in multi-thread contexts. At the same time, no performance difference occurs in a single thread context, because during single thread operation, instructions pass directly through pipeline 100, as discussed above.
Referring now to
If it is determined that the instruction may not have a long latency (i.e., a predetermined condition has not occurred), the instruction may be executed (block 240) and a next instruction of the current thread may be fetched (block 210).
If instead it is determined that a long latency may result (i.e., a predetermined condition has occurred), the pipeline may be prepared to switch threads (block 250). In one embodiment, preparation to switch threads may include executing a remaining one or several instructions of the first thread prior to the thread switch. In such manner, additional instructions may be performed without a processor stall. More so, because such instructions are already present in the processor pipeline, they need not be flushed. Thus instructions of a first thread may continue to be processed while the second thread is prepared for execution. For example, an embodiment executed in the processor pipeline 100 of
Preparing to switch threads may further include setting instruction fetch unit 120 to fetch an instruction of the new thread. For example, a program counter for the second thread within instruction fetch unit 120 may be selected and used to fetch the next instruction of the thread.
Next, the threads may be switched (block 260). In the embodiment of
Referring now to Table 1, shown is an example of execution of a code portion in accordance with an embodiment of the present invention:
As shown in Table 1, a first thread (Thread 0) may be executing in a processor pipeline. When a long latency instruction is detected by instruction decode unit 130 (i.e., the Load to register 0 (LDR to R0) instruction), a signal may be sent on feedback line 125 to instruction fetch unit 120 to prepare for a new thread. At the same time, one or more additional instructions of the first thread may be executed in the processor pipeline (e.g., SUB R3 and XOR R4).
Then as shown in Table 1, a second thread (i.e., Thread 1) may be switched to and begun activation. As with the first thread, the second thread may begin executing instructions and continue until a long latency instruction is encountered (i.e., the Load to register 7 (LDR to R7) instruction). As above, at the detection of a long latency instruction, preparation may be made to switch threads again, while at the same time processing one or more additional instructions of the current thread.
Finally as shown in Table 1, the original thread (i.e., Thread 0) may again be switched to and begun execution. While shown and discussed in Table 1 as having two threads of operation, it is to be understood that in other embodiments multithread operations may encompass more than two threads.
Embodiments of the present invention may be implemented in code and may be stored on a storage medium having stored thereon instructions which can be used to program a system, such as a wireless device to perform the instructions. The storage medium may include, but is not limited to, any type of disk including floppy disks, optical disks, compact disk read-only memories (CD-ROMs), compact disk rewritables (CD-RWs), and magneto-optical disks, semiconductor devices such as read-only memories (ROMs), random access memories (RAMs), erasable programmable read-only memories (EPROMs), flash memories, electrically erasable programmable read-only memories (EEPROMs), magnetic or optical cards, or any type of media suitable for storing electronic instructions, including programmable storage devices.
As shown in
Although the description makes reference to specific components of device 500, it is contemplated that numerous modifications and variations of the described and illustrated embodiments may be possible.
While the present invention has been described with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the true spirit and scope of this present invention.