Advanced Micro Devices Corporation

Nvidia Corporation

Norbert Juffa

Stuart F Oberman

Frederick Daniel Weber

Stephan G Meier

Radoslav Danilak

Fred Weber

Ravikrishna Cherukuri

Ming Siu

Brett Coon

Derrick R Meyer

Krishnan V Ramani

Scott A White

Michael David Achenbach

Jeffrey E Trull

Ravi Krishna Cherukuri

Ravi Krishna

Embodiments of the present invention set forth a technique for optimizing the performance and efficiency of complex, software-based computations, such as lighting computations. Data entering a graphics application programming interface (API) in a conventional arithmetic representation, such as floating-point or fixed-point, is converted to an internal logarithmic representation for greater computational efficiency. Lighting computations are then performed using logarithmic space arithmetic routines that, on average, execute more efficiently than similar routines performed in a native floating-point format. The lighting computation results, represented as logarithmic space numbers, are converted back to floating-point numbers before being transmitted to a graphics processing unit (GPU) for further processing. Because of efficiencies of logarithmic space arithmetic, performance improvements may be realized relative to prior art approaches to performing software-based floating-point operations.

Optimized 3D lighting computations using a logarithmic number system

In contrast to a conventional computing system in which the graphics processor (graphics processing unit or GPU) is treated as a slave to one or several CPUs, systems and methods are provided that allow the GPU to be treated as a central processing unit (CPU) from the perspective of the operating system. The GPU can access a memory space shared by other CPUs in the computing system. Caches utilized by the GPU may be coherent with caches utilized by other CPUs in the computing system. The GPU may share execution of general-purpose computations with other CPUs in the computing system.

Graphics processor with memory management unit and cache coherent link

The present invention enables efficient matrix multiplication operations on parallel processing devices. One embodiment is a method for mapping CTAs to result matrix tiles for matrix multiplication operations. Another embodiment is a second method for mapping CTAs to result tiles. Yet other embodiments are methods for mapping the individual threads of a CTA to the elements of a tile for result tile computations, source tile copy operations, and source tile copy and transpose operations. The present invention advantageously enables result matrix elements to be computed on a tile-by-tile basis using multiple CTAs executing concurrently on different streaming multiprocessors, enables source tiles to be copied to local memory to reduce the number accesses from the global memory when computing a result tile, and enables coalesced read operations from the global memory as well as write operations to the local memory without bank conflicts.

Efficient matrix multiplication on a parallel processing device

In parallel processing devices, for streaming computations, processing of each data element of the stream may not be computationally intensive and thus processing may take relatively small amounts of time to compute as compared to memory accesses times required to read the stream and write the results. Therefore, memory throughput often limits the performance of the streaming computation. Generally stated, provided are methods for achieving improved, optimized, or ultimately, maximized memory throughput in such memory-throughput-limited streaming computations. Streaming computation performance is maximized by improving the aggregate memory throughput across the plurality of processing elements and threads. High aggregate memory throughput is achieved by balancing processing loads between threads and groups of threads and a hardware memory interface coupled to the parallel processing devices.

Maximized memory throughput on parallel processing devices

One embodiment of the present invention sets forth a technique for performing fast integer division using commonly available arithmetic operations. The technique may be implemented in a two-stage process using a single-precision floating point reciprocal in conjunction with integer addition and multiplication. Furthermore, the technique may be fully pipelined on many conventional processors for performance that is comparable to the best available high-performance alternatives.

Pipelined integer division using floating-point reciprocal

Maximized memory throughput using cooperative thread arrays

A graphics processing unit is programmed to carry out cryptographic processing so that fast, effective cryptographic processing solutions can be provided without incurring additional hardware costs. The graphics processing unit can efficiently carry out cryptographic processing because it has an architecture that is configured to handle a large number of parallel processes. The cryptographic processing carried out on the graphics processing unit can be further improved by configuring the graphics processing unit to be capable of both floating point and integer operations.

Graphics processing unit used for cryptographic processing

Mapping the threads of a CTA to the elements of a tile for efficient matrix multiplication

Hardware/software-based mapping of CTAs to matrix tiles for efficient matrix multiplication

Hardware resource based mapping of cooperative thread arrays (CTA) to result matrix tiles for efficient matrix multiplication in computing system comprising plurality of multiprocessors

An execution unit is provided for executing a first instruction which includes an opcode field, a first operand field, and a second operand field. The execution unit includes a first input register for receiving a first operand specified by a value of the first operand field, and a second input register for receiving a second operand specified by a value of the second operand field. The execution unit further includes a comparator unit which is coupled to receive a value of the opcode field for the first instruction. The comparator unit is also coupled to receive the first and second operand values from the first and second input registers, respectively. The execution further includes a multiplexer which receives a plurality of inputs. These inputs include a first constant value, a second constant value, and the values of the first and second operand. If the decoded opcode value received by the comparator indicates that the first instruction is either a compare or extreme value function, the comparator conveys one or more control signals to the multiplexer for the purpose of selecting an output of the multiplexer as the result of the first instruction. If the first instruction is one of a plurality of extreme value instructions, the one or more control signals conveyed by the comparator unit select between the first operand and second operand to determine the result of the first instruction. If the first instruction is one of a plurality of compare instructions, the one or more control signals conveyed by the comparator unit select between the first and second constant value to determine the result of the first instruction. In another embodiment, a similar execution unit is provided which handles vector operands.

Microprocessor including an efficient implementation of extreme value instructions

An apparatus and method for superforwarding load operands in a microprocessor are provided. An execution unit in a microprocessor is configured to receive a load instruction and a subsequent instruction. If the load instruction corresponds to a simple load instruction, a destination operand of the load instruction can be superforwarded to a subsequent instruction if the subsequent instruction specifies a source operand that depends on the destination operand of the load instruction. The subsequent instruction is not required to wait until a load instruction executes or completes and can be scheduled and/or executed prior to or at the same time as the load instruction. Consequently, latencies associated with operand dependencies may be reduced.

Apparatus and method for superforwarding load operands in a microprocessor

A microprocessor configured to rapidly execute floating point store status word (FSTSW) type instructions that are immediately preceded by floating point compare (FCOM) type instructions is disclosed. FCOM-type instructions are modified to store their results to an architectural floating point status word and a temporary destination register. If an FSTSW-type instruction is detected immediately following an FCOM-type instruction, then the FSTSW-type instruction is transformed into a special fast floating point store status word (FSTSWEF) instruction. Unlike the FSTSW-type instruction, which is serializing and negatively impacts performance, the FSTSWEF instruction is not serializing and allows execution to continue without undue serialization. A computer system and method for rapidly executing FSTSW instructions immediately preceded by FCOM-type instructions are also disclosed.

Method and apparatus for rapid execution of FCOM and FSTSW

An apparatus and method for executing floating-point store instructions in a microprocessor is provided. If store data of a floating-point store instruction corresponds to a tiny number and an underflow exception is masked, then a trap routine can be executed to generate corrected store data and complete the store operation. In response to detecting that store data corresponds to a tiny number and the underflow exception is masked, the store data, store address information, and opcode information can be stored prior to initiating the trap routine. The trap routine can be configured to access the store data, store address information, and opcode information. The trap routine can be configured to generate corrected store data and complete the store operation using the store data, store address information, and opcode information.

Growing community of inventors

San Jose, CA, United States of America

Norbert Juffa