Метод повышения параллелизма на уровне команд

Method of improving instruction level parallelism

+Basic five stage pipeline 1 2 3 4 5 6 71IFIDEX MEMWB2IFID EXMEMWB3IF IDEXMEMWB4 IFIDEXMEM5 IFIDEX(IF = Instruction Fetch, ID = Instruction Decode, EX = Execute, MEM = Memory access, WB = Register write back). In the fourth clock cycle (the green column), the earliest instruction is in MEM stage, and the latest instruction has not yet entered the pipeline. In computer engineering, instruction pipelining is a technique for implementing instruction level parallelism within a single processor. Pipelining attempts to keep every part of the processor busy with some instruction by dividing incoming instructions into a series of sequential steps (the eponymous "pipeline") performed by different processor units with different parts of instructions processed in parallel.

In a pipelined computer, instructions flow through the central processing unit (CPU) in stages. For example, it might have one stage for each step of the von Neumann cycle: Fetch the instruction, fetch the operands, do the instruction, write the results. A pipelined computer usually has "pipeline registers" after each stage. These store information from the instruction and calculations so that the logic gates of the next stage can do the next step. This arrangement lets the CPU complete an instruction on each clock cycle. It is common for even numbered stages to operate on one edge of the square wave clock, while odd numbered stages operate on the other edge. This allows more CPU throughput than a multicycle computer at a given clock rate, but may increase latency due to the added overhead of the pipelining process itself. Also, even though the electronic logic has a fixed maximum speed, a pipelined computer can be made faster or slower by varying the number of stages in the pipeline. With more stages, each stage does less work, and so the stage has fewer delays from the logic gates and could run at a higher clock rate. A pipelined model of computer is often the most economical, when cost is measured as logic gates per instruction per second. At each instant, an instruction is in only one pipeline stage, and on average, a pipeline stage is less costly than a multicycle computer. Also, when made well, most of the pipelined computer's logic is in use most of the time. In contrast, out of order computers usually have large amounts of idle logic at any given instant. Similar calculations usually show that a pipelined computer uses less energy per instruction. However, a pipelined computer is usually more complex and more costly than a comparable multicycle computer. It typically has more logic gates, registers and a more complex control unit. In a like way, it might use more total energy, while using less energy per instruction. Out of order CPUs can usually do more instructions per second because they can do several instructions at once. In a pipelined computer, the control unit arranges for the flow to start, continue, and stop as a program commands. The instruction data is usually passed in pipeline registers from one stage to the next, with a somewhat separated piece of control logic for each stage. The control unit also assures that the instruction in each stage does not harm the operation of instructions in other stages. For example, if two stages must use the same piece of data, the control logic assures that the uses are done in the correct sequence. When operating efficiently, a pipelined computer will have an instruction in each stage. It is then working on all of those instructions at the same time. It can finish about one instruction for each cycle of its clock. But when a program switches to a different sequence of instructions, the pipeline sometimes must discard the data in process and restart. This is called a "stall." Much of the design of a pipelined computer prevents interference between the stages and reduces stalls.

Seminal uses of pipelining were in the ILLIAC II project and the IBM Stretch project, though a simple version was used earlier in the Z1 in 1939 and the Z3 in 1941. Pipelining began in earnest in the late 1970s in supercomputers such as vector processors and array processors. One of the early supercomputers was the Cyber series built by Control Data Corporation. Its main architect, Seymour Cray, later headed Cray Research. Cray developed the XMP line of supercomputers, using pipelining for both multiply and add/subtract functions. Later, Star Technologies added parallelism (several pipelined functions working in parallel), developed by Roger Chen. In 1984, Star Technologies added the pipelined divide circuit developed by James Bradley. By the mid 1980s, pipelining was used by many different companies around the world. Pipelining was not limited to supercomputers. In 1976, the Amdahl Corporation's 470 series general purpose mainframe had a 7 step pipeline, and a patented branch prediction circuit.

In some early DSP and RISC processors, the documentation advises programmers to avoid such dependencies in adjacent and nearly adjacent instructions (called delay slots), or declares that the second instruction uses an old value rather than the desired value (in the example above, the processor might counter intuitively copy the unincremented value), or declares that the value it uses is undefined. The programmer may have unrelated work that the processor can do in the meantime; or, to ensure correct results, the programmer may insert NOPs into the code, partly negating the advantages of pipelining.

A branch out of the normal instruction sequence often involves a hazard. Unless the processor can give effect to the branch in a single time cycle, the pipeline will continue fetching instructions sequentially. Such instructions cannot be allowed to take effect because the programmer has diverted control to another part of the program. A conditional branch is even more problematic. The processor may or may not branch, depending on a calculation that has not yet occurred. Various processors may stall, may attempt branch prediction, and may be able to begin to execute two different program sequences (eager execution), each assuming the branch is or is not taken, discarding all work that pertains to the incorrect guess. A processor with an implementation of branch prediction that usually makes correct predictions can minimize the performance penalty from branching. However, if branches are predicted poorly, it may create more work for the processor, such as flushing from the pipeline the incorrect code path that has begun execution before resuming execution at the correct location. Programs written for a pipelined processor deliberately avoid branching to minimize possible loss of speed. For example, the programmer can handle the usual case with sequential execution and branch only on detecting unusual cases. Using programs such as gcov to analyze code coverage lets the programmer measure how often particular branches are actually executed and gain insight with which to optimize the code. In some cases, a programmer can handle both the usual case and unusual case with branch free code.

A pipelined processor may deal with hazards by stalling and creating a bubble in the pipeline, resulting in one or more cycles in which nothing useful happens. In the illustration at right, in cycle 3, the processor cannot decode the purple instruction, perhaps because the processor determines that decoding depends on results produced by the execution of the green instruction. The green instruction can proceed to the Execute stage and then to the Write back stage as scheduled, but the purple instruction is stalled for one cycle at the Fetch stage. The blue instruction, which was due to be fetched during cycle 3, is stalled for one cycle, as is the red instruction after it. Because of the bubble (the blue ovals in the illustration), the processor's Decode circuitry is idle during cycle 3. Its Execute circuitry is idle during cycle 4 and its Write back circuitry is idle during cycle 5. When the bubble moves out of the pipeline (at cycle 6), normal execution resumes. But everything now is one cycle late. It will take 8 cycles (cycle 1 through 8) rather than 7 to completely execute the four instructions shown in colors.