Abstract: Methods and apparatus for source operand collector caching. In one embodiment, a processor includes a register file that may be coupled to storage elements (i.e., an operand collector) that provide inputs to the datapath of the processor core for executing an instruction. In order to reduce bandwidth between the register file and the operand collector, operands may be cached and reused in subsequent instructions. A scheduling unit maintains a cache table for monitoring which register values are currently stored in the operand collector. The scheduling unit may also configure the operand collector to select the particular storage elements that are coupled to the inputs to the datapath for a given instruction.

Abstract: Systems and methods for throttling GPU execution performance to avoid surges in DI/DT. A processor includes one or more execution units coupled to a scheduling unit configured to select instructions for execution by the one or more execution units. The execution units may be connected to one or more decoupling capacitors that store power for the circuits of the execution units. The scheduling unit is configured to throttle the instruction issue rate of the execution units based on a moving average issue rate over a large number of scheduling periods. The number of instructions issued during the current scheduling period is less than or equal to a throttling rate maintained by the scheduling unit that is greater than or equal to a minimum throttling issue rate. The throttling rate is set equal to the moving average plus an offset value at the end of each scheduling period.

Abstract: One embodiment of the present disclosure sets forth an optimized way to execute pre-scheduled replay operations for divergent operations in a parallel processing subsystem. Specifically, a streaming multiprocessor (SM) includes a multi-stage pipeline configured to insert pre-scheduled replay operations into a multi-stage pipeline. A pre-scheduled replay unit detects whether the operation associated with the current instruction is accessing a common resource. If the threads are accessing data which are distributed across multiple cache lines, then the pre-scheduled replay unit inserts pre-scheduled replay operations behind the current instruction. The multi-stage pipeline executes the instruction and the associated pre-scheduled replay operations sequentially. If additional threads remain unserviced after execution of the instruction and the pre-scheduled replay operations, then additional replay operations are inserted via the replay loop, until all threads are serviced.

Abstract: One embodiment of the present invention sets forth a technique for performing a shaped access of a register file that includes a set of N registers, wherein N is greater than or equal to two. The technique involves, for at least one thread included in a group of threads, receiving a request to access a first amount of data from each register in the set of N registers, and configuring a crossbar to allow the at least one thread to access the first amount of data from each register in the set of N registers.

Abstract: Systems and methods for scheduling instructions using pre-decode data corresponding to each instruction. In one embodiment, a multi-core processor includes a scheduling unit in each core for selecting instructions from two or more threads each scheduling cycle for execution on that particular core. As threads are scheduled for execution on the core, instructions from the threads are fetched into a buffer without being decoded. The pre-decode data is determined by a compiler and is extracted by the scheduling unit during runtime and used to control selection of threads for execution. The pre-decode data may specify a number of scheduling cycles to wait before scheduling the instruction. The pre-decode data may also specify a scheduling priority for the instruction. Once the scheduling unit selects an instruction to issue for execution, a decode unit fully decodes the instruction.

Abstract: Systems and methods for scheduling instructions without instruction decode. In one embodiment, a multi-core processor includes a scheduling unit in each core for scheduling instructions from two or more threads scheduled for execution on that particular core. As threads are scheduled for execution on the core, instructions from the threads are fetched into a buffer without being decoded. The scheduling unit includes a macro-scheduler unit for performing a priority sort of the two or more threads and a micro-scheduler arbiter for determining the highest order thread that is ready to execute. The macro-scheduler unit and the micro-scheduler arbiter use pre-decode data to implement the scheduling algorithm. The pre-decode data may be generated by decoding only a small portion of the instruction or received along with the instruction. Once the micro-scheduler arbiter has selected an instruction to dispatch to the execution unit, a decode unit fully decodes the instruction.

Abstract: Methods and apparatus for source operand collector caching. In one embodiment, a processor includes a register file that may be coupled to storage elements (i.e., an operand collector) that provide inputs to the datapath of the processor core for executing an instruction. In order to reduce bandwidth between the register file and the operand collector, operands may be cached and reused in subsequent instructions. A scheduling unit maintains a cache table for monitoring which register values are currently stored in the operand collector. The scheduling unit may also configure the operand collector to select the particular storage elements that are coupled to the inputs to the datapath for a given instruction.

Abstract: One embodiment of the present invention sets forth an optimized way to execute replay operations for divergent operations in a parallel processing subsystem. Specifically, the streaming multiprocessor (SM) includes a multistage pipeline configured to batch two or more replay operations for processing via replay loop. A logic element within the multistage pipeline detects whether the current pipeline stage is accessing a shared resource, such as loading data from a shared memory. If the threads are accessing data which are distributed across multiple cache lines, then the multistage pipeline batches two or more replay operations, where the replay operations are inserted into the pipeline back-to-back. Advantageously, divergent operations requiring two or more replay operations operate with reduced latency. Where memory access operations require transfer of more than two cache lines to service all threads, the number of clock cycles required to complete all replay operations is reduced.

Abstract: One embodiment of the present invention sets forth an improved way to prefetch instructions in a multi-level cache. Fetch unit initiates a prefetch operation to transfer one of a set of multiple cache lines, based on a function of a pseudorandom number generator and the sector corresponding to the current instruction L1 cache line. The fetch unit selects a prefetch target from the set of multiple cache lines according to some probability function. If the current instruction L1 cache 370 is located within the first sector of the corresponding L1.5 cache line, then the selected prefetch target is located at a sector within the next L1.5 cache line. The result is that the instruction L1 cache hit rate is improved and instruction fetch latency is reduced, even where the processor consumes instructions in the instruction L1 cache at a fast rate.

Abstract: One embodiment of the present invention sets forth a technique that provides an efficient way to retrieve operands from a register file. Specifically, the instruction dispatch unit receives one or more instructions, each of which includes one or more operands. Collectively, the operands are organized into one or more operand groups from which a shaped access may be formed. The operands are retrieved from the register file and stored in a collector. Once all operands are read and collected in the collector, the instruction dispatch unit transmits the instructions and corresponding operands to functional units within the streaming multiprocessor for execution. One advantage of the present invention is that multiple operands are retrieved from the register file in a single register access operation without resource conflict. Performance in retrieving operands from the register file is improved by forming shaped accesses that efficiently retrieve operands exhibiting recognized memory access patterns.

Abstract: One embodiment of the present invention sets forth a technique for speculatively issuing instructions to allow a processing pipeline to continue to process some instructions during rollback of other instructions. A scheduler circuit issues instructions for execution assuming that, several cycles later, when the instructions reach multithreaded execution units, that dependencies between the instructions will be resolved, resources will be available, operand data will be available, and other conditions will not prevent execution of the instructions. When a rollback condition exists at the point of execution for an instruction for a particular thread group, the instruction is not dispatched to the multithreaded execution units. However, other instructions issued by the scheduler circuit for execution by different thread groups, and for which a rollback condition does not exist, are executed by the multithreaded execution units.

Abstract: A parallel thread processor executes thread groups belonging to multiple cooperative thread arrays (CTAs). At each cycle of the parallel thread processor, an instruction scheduler selects a thread group to be issued for execution during a subsequent cycle. The instruction scheduler selects a thread group to issue for execution by (i) identifying a pool of available thread groups, (ii) identifying a CTA that has the greatest seniority value, and (iii) selecting the thread group that has the greatest credit value from within the CTA with the greatest seniority value.