Advanced Micro Devices Corporation

Microsoft Technology Licensing, LLC

Ati Technologies Ulc

Bradford Michael Beckmann

Gabriel H Loh

Sooraj Puthoor

James Michael O'Connor

David Allen Wood

Steven Karl Reinhardt

Lisa R Hsu

Marc S Orr

Yasuko Eckert

Alexandru Dutu

Nuwan S Jayasena

Shomit N Das

Shuai Che

Matthew David Sinclair

Michael Ignatowski

Steven Tony Tye

Brian Laird Sumner

Xianwei Zhang

Sergey Blagodurov

Anthony T Gutierrez

Johnathan R Alsop

Mithuna Shamabhat Thottethodi

Seyedmohammad SeyedzadehDelcheh

Bradley Michael Waters

Kishore Punniyamurthy

Blake A Hechtman

Mauricio Breternitz, Jr

Mark Donald Hill

Michael J Schulte

Zhe Wang

Benedict R Gaster

Ashwin Aji

Muhammad Amber Hassaan

Nicolai Hähnle

Ayse Yilmazer

Timothy G Rogers

Marcus Nathaniel Chow

Gokul Subramanian Ravi

Anirudh Mohan Kaushik

John Kalamatianos

Andrew Gregory Kegel

Jay E Fleischman

Jaewoong Chung

Michael Estlick

Arkaprava Basu

Onur Kayiran

Dmitri Yudanov

Yuan Xie

Khaled Hamidouche

Matthew J Tomei

Brandon K Potter

Junli Gu

Derek R Hower

Lena E Olson

Furkan Eris

Richard David Sodke

Nilay Vaish

Tushar Krishna

Subho Chatterjee

Pouya Fotouhi

Seyed Mohammad Seyedzadehdelcheh

Joseph Gross

Xulong Tang

Vadim Vadimovich Nikiforov

Yash Sanjeev Ukidave

An electronic device includes processing circuitry that executes a lookup table (LUT) vector instruction. Executing the lookup table vector instruction causes the processing circuitry to acquire a set of reference values by using each input value from an input vector as an index to acquire a reference value from a reference vector. The processing circuitry then provides the set of reference values for one or more subsequent operations. The processing circuitry can also use the set of reference values for controlling vector elements from among a set of vector elements for which a vector operation is performed.

Lookup table (LUT) vector instruction

A device includes a plurality of processing elements (PEs). A symmetric memory is allocated in each of the plurality of PEs. The device includes a switch connected to the plurality of PEs. The switch is to: receive, from a first processing element (PE) of the plurality of PEs, a message that includes a buffer offset, compute, based on the buffer offset, a first memory address of a first buffer in a first symmetric memory of the first PE and a second memory address of a second buffer in a second symmetric memory of a second PE of the plurality of PEs, and initiate, based on the first memory address and the second memory address, a direct memory access operation to access the first buffer and the second buffer.

Low latency offloading of collectives over a switch

Methods, devices, and systems for managing performance of a processor having multiple compute units. An effective number of the multiple compute units may be determined to designate as having priority. On a condition that the effective number is nonzero, the effective number of the multiple compute units may each be designated as a priority compute unit. Priority compute units may have access to a shared cache whereas non-priority compute units may not. Workgroups may be preferentially dispatched to priority compute units. Memory access requests from priority compute units may be served ahead of requests from non-priority compute units.

System performance management using prioritized compute units

An approach is provided for implementing register based single instruction, multiple data (SIMD) lookup table operations. According to the approach, an instruction set architecture (ISA) can support one or more SIMD instructions that enable vectors or multiple values in source data registers to be processed in parallel using a lookup table or truth table stored in one or more function registers. The SIMD instructions can be flexibly configured to support functions with inputs and outputs of various sizes and data formats. Various approaches are also described for supporting very large lookup tables that span multiple registers.

Register based SIMD lookup table operations

A technique for synchronizing workgroups is provided. Multiple workgroups execute a wait instruction that specifies a condition variable and a condition. A workgroup scheduler stops execution of a workgroup that executes a wait instruction and an advanced controller begins monitoring the condition variable. In response to the advanced controller detecting that the condition is met, the workgroup scheduler determines whether there is a high contention scenario, which occurs when the wait instruction is part of a mutual exclusion synchronization primitive and is detected by determining that there is a low number of updates to the condition variable prior to detecting that the condition has been met. In a high contention scenario, the workgroup scheduler wakes up one workgroup and schedules another workgroup to be woken up at a time in the future. In a non-contention scenario, more than one workgroup can be woken up at the same time.

Selective workgroup wake-up based on synchronization mechanism identification with high contention scenario

Systems, apparatuses, and methods for implementing a multi-tiered approach to cache compression are disclosed. A cache includes a cache controller, light compressor, and heavy compressor. The decision on which compressor to use for compressing cache lines is made based on certain resource availability such as cache capacity or memory bandwidth. This allows the cache to opportunistically use complex algorithms for compression while limiting the adverse effects of high decompression latency on system performance. To address the above issue, the proposed design takes advantage of the heavy compressors for effectively reducing memory bandwidth in high bandwidth memory (HBM) interfaces as long as they do not sacrifice system performance. Accordingly, the cache combines light and heavy compressors with a decision-making unit to achieve reduced off-chip memory traffic without sacrificing system performance.

Resource-aware compression

A processing system monitors and synchronizes parallel execution of workgroups (WGs). One or more of the WGs perform (e.g., periodically or in response to a trigger such as an indication of oversubscription) a waiting atomic instruction. In response to a comparison between an atomic value produced as a result of the waiting atomic instruction and an expected value, WGs that fail to produce a correct atomic value are identified as being in a waiting state (e.g., waiting for a synchronization variable). Execution of WGs in the waiting state is prevented (e.g., by a context switch) until corresponding synchronization variables are released.

Workgroup synchronization and processing

Systems, apparatuses, and methods for implementing a multi-kernel wavefront scheduler are disclosed. A system includes at least a parallel processor coupled to one or more memories, wherein the parallel processor includes a command processor and a plurality of compute units. The command processor launches multiple kernels for execution on the compute units. Each compute unit includes a multi-level scheduler for scheduling wavefronts from multiple kernels for execution on its execution units. A first level scheduler creates scheduling groups by grouping together wavefronts based on the priority of their kernels. Accordingly, wavefronts from kernels with the same priority are grouped together in the same scheduling group by the first level scheduler. Next, the first level scheduler selects, from a plurality of scheduling groups, the highest priority scheduling group for execution. Then, a second level scheduler schedules wavefronts for execution from the scheduling group selected by the first level scheduler.

Multi-kernel wavefront scheduler

Implementing heterogeneous wavefronts on a graphics processing unit (GPU) is disclosed. A scheduler assigns heterogeneous wavefronts for execution on a compute unit of a processing device. The heterogeneous wavefronts include different types of wavefronts such as vector compute wavefronts and service-level wavefronts that vary in resource requirements and instruction sets. As one example, heterogeneous wavefronts may include scalar wavefronts and vector compute wavefronts that execute on scalar units and vector units, respectively. Distinct sets of instructions are executed for the heterogeneous wavefronts on the compute unit. Heterogeneous wavefronts are processed in the same pipeline of the processing device.

Implementing heterogeneous wavefronts on a graphics processing unit (GPU)

Systems, apparatuses, and methods for managing a number of wavefronts permitted to concurrently execute in a processing system. An apparatus includes a register file with a plurality of registers and a plurality of compute units configured to execute wavefronts. A control unit of the apparatus is configured to allow a first number of wavefronts to execute concurrently on the plurality of compute units. The control unit is configured to allow no more than a second number of wavefronts to execute concurrently on the plurality of compute units, wherein the second number is less than the first number, in response to detection that thrashing of the register file is above a threshold. The control unit is configured to detect said thrashing based at least in part on a number of registers in use by executing wavefronts that spill to memory.

Management of thrashing in a GPU

A processor includes a task scheduling unit and a compute unit coupled to the task scheduling unit. The task scheduling unit performs a task dependency assessment of a task dependency graph and task data requirements that correspond to each task of the plurality of tasks. Based on the task dependency assessment, the task scheduling unit schedules a first task of the plurality of tasks and a second proxy object of a plurality of proxy objects specified by the task data requirements such that a memory transfer of the second proxy object of the plurality of proxy objects occurs while the first task is being executed.

Hardware assisted fine-grained data movement

Techniques for executing workgroups are provided. The techniques include executing, for a first workgroup of a first kernel dispatch, a workgroup dependency instruction that includes an indication to prioritize execution of a second workgroup of a second kernel dispatch, and in response to the workgroup dependency instruction, dispatching the second workgroup of the second kernel dispatch prior to dispatching a third workgroup of the second kernel dispatch, wherein no workgroup dependency instruction including an indication to prioritize execution of the third workgroup has been executed.

Fine-grained conditional dispatching

In some embodiments, a memory controller in a processor includes a base value cache, a compressor, and a metadata cache. The compressor is coupled to the base value cache and the metadata cache. The compressor compresses a data block using at least a base value and delta values. The compressor determines whether the size of the data block exceeds a data block threshold value. Based on the determination of whether the size of the compressed data block generated by the compressor exceeds the data block threshold value, the memory controller transfers only a set of the compressed delta values to memory for storage. A decompressor located in the lower level cache of the processor decompresses the compressed data block using the base value stored in the base value cache, metadata stored in the metadata cache and the delta values stored in memory.

Data compression system using base values and methods thereof

Systems, apparatuses, and methods for dynamic graphics processing unit (GPU) register allocation are disclosed. A GPU includes at least a plurality of compute units (CUs), a control unit, and a plurality of registers for each CU. If a new wavefront requests more registers than are currently available on the CU, the control unit spills registers associated with stack frames at the bottom of a stack since they will not likely be used in the near future. The control unit has complete flexibility determining how many registers to spill based on dynamic demands and can prefetch the upcoming necessary fills without software involvement. Effectively, the control unit manages the physical register file as a cache. This allows younger workgroups to be dynamically descheduled so that older workgroups can allocate additional registers when needed to ensure improved fairness and better forward progress guarantees.

Dynamic graphical processing unit register allocation

Growing community of inventors

Redmond, WA, United States of America

Bradford Michael Beckmann