Nvidia Corporation

Ajay Sudarshan Tirumala

Jack H Choquette

Olivier Giroux

Shirish C Gadre

Peter Michael Nelson

Maciej Piotr Tyrlik

Manan Patel

Praveen Kaushik

Vinod Grover

Michael Alan Fetterman

Ronny Krashinsky

Gary M Tarolli

Gregory Diamos

Andrei Khodakovsky

Richard Craig Johnson

Konstantinos Kyriakopoulos

Rishkul Kulkarni

Frank Joseph Eaton

Peter Nelson

Amanpreet Grewal

Shekhar Vasant Divekar

Daniel Alan Stiffler

Ankita Upreti

Divya Shanmughan

Various embodiments include a parallel processing computer system that enables parallel instances of a program to synchronize at disparate addresses in memory. When the parallel program instances need to exchange data, the program instances synchronize based on a mask that identifies the program instances that are synchronizing. As each program instance reaches the point of synchronization, the program instance blocks and waits for all other program instances to reach the point of synchronization. When all program instances have reached the point of synchronization, at least one program instance executes a synchronous operation to exchange data. The program instances then continue execution at respective and disparate return addresses.

Techniques for efficiently synchronizing multiple program threads

Various techniques for accelerating dynamic programming algorithms are provided. For example, a fused addition and comparison instruction, a three-operand comparison instruction, and a two-operand comparison instruction are used to accelerate a Needleman-Wunsch algorithm that determines an optimized global alignment of subsequences over two entire sequences. In another example, the fused addition and comparison instruction is used in an innermost loop of a Floyd-Warshall algorithm to reduce the number of instructions required to determine shortest paths between pairs of vertices in a graph. In another example, a two-way single instruction multiple data (SIMD) floating point variant of the three-operand comparison instruction is used to reduce the number of instructions required to determine the median of an array of floating point values.

Implementing specialized instructions for accelerating dynamic programming algorithms

Various techniques for accelerating Smith-Waterman sequence alignments are provided. For example, threads in a group of threads are employed to use an interleaved cell layout to store relevant data in registers while computing sub-alignment data for one or more local alignment problems. In another example, specialized instructions that reduce the number of cycles required to compute each sub-alignment score are utilized. In another example, threads are employed to compute sub-alignment data for a subset of columns of one or more local alignment problems while other threads begin computing sub-alignment data based on partial result data received from the preceding threads. After computing a maximum sub-alignment score, a thread stores the maximum sub-alignment score and the corresponding position in global memory.

Techniques for storing sub-alignment data when accelerating Smith-Waterman sequence alignments

Implementing specialized instructions for accelerating Smith-Waterman sequence alignments

Techniques are disclosed for reducing the latency associated with performing data reductions in a multithreaded processor. In response to a single instruction associated with a set of threads executing in the multithreaded processor, a warp reduction unit acquires register values stored in source registers, where each register value is associated with a different thread included in the set of threads. The warp reduction unit performs operation(s) on the register values to compute an aggregate value. The warp reduction unit stores the aggregate value in a destination register that is accessible to at least one of the threads in the set of threads. Because the data reduction is performed via a single instruction using hardware specialized for data reductions, the number of cycles required to perform the data reduction is decreased relative to prior-art techniques that are performed via multiple instructions using hardware that is not specialized for data reductions.

Techniques for efficiently performing data reductions in parallel processing units

In one embodiment, a synchronization instruction causes a processor to ensure that specified threads included within a warp concurrently execute a single subsequent instruction. The specified threads include at least a first thread and a second thread. In operation, the first thread arrives at the synchronization instruction. The processor determines that the second thread has not yet arrived at the synchronization instruction and configures the first thread to stop executing instructions. After issuing at least one instruction for the second thread, the processor determines that all the specified threads have arrived at the synchronization instruction. The processor then causes all the specified threads to execute the subsequent instruction. Advantageously, unlike conventional approaches to synchronizing threads, the synchronization instruction enables the processor to reliably and properly execute code that includes complex control flows and/or instructions that presuppose that threads are converged.

Techniques for comprehensively synchronizing execution threads

A just-in-time (JIT) compiler binds constants to specific memory locations at runtime. The JIT compiler parses program code derived from a multithreaded application and identifies an instruction that references a uniform constant. The JIT compiler then determines a chain of pointers that originates within a root table specified in the multithreaded application and terminates at the uniform constant. The JIT compiler generates additional instructions for traversing the chain of pointers and inserts these instructions into the program code. A parallel processor executes this compiled code and, in doing so, causes a thread to traverse the chain of pointers and bind the uniform constant to a uniform register at runtime. Each thread in a group of threads executing on the parallel processor may then access the uniform constant.

Binding constants at runtime for improved resource utilization

A compiler parses a multithreaded application into cohesive blocks of instructions. Cohesive blocks include instructions that do not diverge or converge. Each cohesive block is associated with one or more uniform registers. When a set of threads executes the instructions in a given cohesive block, each thread in the set may access the uniform register independently of the other threads in the set. Accordingly, the uniform register may store a single copy of data on behalf of all threads in the set of threads, thereby conserving resources.

Uniform register file for improved resource utilization

A streaming multiprocessor (SM) includes a nanosleep (NS) unit configured to cause individual threads executing on the SM to sleep for a programmer-specified interval of time. For a given thread, the NS unit parses a NANOSLEEP instruction and extracts a sleep time. The NS unit then maps the sleep time to a single bit of a timer and causes the thread to sleep. When the timer bit changes, the sleep time expires, and the NS unit awakens the thread. The thread may then continue executing. The SM also includes a nanotrap (NT) unit configured to issue traps using a similar timing mechanism to that described above. For a given thread, the NT unit parses a NANOTRAP instruction and extracts a trap time. The NT unit then maps the trap time to a single bit of a timer. When the timer bit changes, the NT unit issues a trap.

Thread-level sleep in a multithreaded architecture

A synchronization instruction causes a processor to ensure that specified threads included within a warp concurrently execute a single subsequent instruction. The specified threads include at least a first thread and a second thread. In operation, the first thread arrives at the synchronization instruction. The processor determines that the second thread has not yet arrived at the synchronization instruction and configures the first thread to stop executing instructions. After issuing at least one instruction for the second thread, the processor determines that all the specified threads have arrived at the synchronization instruction. The processor then causes all the specified threads to execute the subsequent instruction. Advantageously, unlike conventional approaches to synchronizing threads, the synchronization instruction enables the processor to reliably and properly execute code that includes complex control flows and/or instructions that presuppose that threads are converged.

A method, system, and computer program product for executing divergent threads using a convergence barrier are disclosed. A first instruction in a program is executed by a plurality of threads, where the first instruction, when executed by a particular thread, indicates to a scheduler unit that the thread participates in a convergence barrier. A first path through the program is executed by a first divergent portion of the participating threads and a second path through the program is executed by a second divergent portion of the participating threads. The first divergent portion of the participating threads executes a second instruction in the program and transitions to a blocked state at the convergence barrier. The scheduler unit determines that all of the participating threads are synchronized at the convergence barrier and the convergence barrier is cleared.

Growing community of inventors

Cupertino, CA, United States of America

Ajay Sudarshan Tirumala