Abstract: According to embodiments of the invention, a step value and a step-interval cache coherency protocol may be used to update and invalidate data stored within cache memory. A step value may be an integer value and may be stored within a cache directory entry associated with data in the memory cache. Upon reception of a cache read request, along with the normal address comparison to determine if the data is located within the cache a current step value may be compared with the stored step value to determine if the data is current. If the step values match, the data may be current and a cache hit may occur. However, if the step values do not match, the requested data may be provided from another source. Furthermore, an application may update the current step value to invalidate old data stored within the cache and associated with a different step value.

Abstract: A circuit arrangement and method support a multithreaded rendering architecture capable of dynamically routing pixel fragments from a pixel fragment generator to any pixel shader from among a pool of pixel shaders. The pixel fragment generator is therefore not tied to a specific pixel shader, but is instead able to utilize multiple pixel shaders in a pool of pixel shaders to minimize bottlenecks and improve overall hardware utilization and performance during image processing.

Abstract: A multirate execution unit is capable of being operated in a plurality of modes, with the execution unit being capable of clocked at multiple different rates relative to a multithreaded issue unit such that, in applications where maximum performance is desired, the execution unit can be clocked at a rate that is faster than the clock rate for the multithreaded issue unit, and in applications where a lower power profile is desired, the execution unit can be throttled back to a slower rate to reduce the power consumption of the execution unit. When the execution unit is clocked at a faster rate than the multithreaded issue unit, the issue unit is permitted to issue more instructions per cycle than when the execution unit is throttled to the slower rate to increase overall instruction throughput.

Abstract: A processing unit includes multiple execution units and sequencer logic that is disposed downstream of instruction buffer logic, and that is responsive to a sequencer instruction present in an instruction stream. In response to such an instruction, the sequencer logic issues a plurality of instructions associated with a long latency operation to one execution unit, while blocking instructions from the instruction buffer logic from being issued to that execution unit. In addition, the blocking of instructions from being issued to the execution unit does not affect the issuance of instructions to any other execution unit, and as such, other instructions from the instruction buffer logic are still capable of being issued to and executed by other execution units even while the sequencer logic is issuing the plurality of instructions associated with the long latency operation.

Abstract: The present invention is generally related to integrated circuit devices, and more particularly, to methods, systems and design structures for the field of image processing, and more specifically to vector units for supporting image processing. A dual vector unit implementation is described wherein two vector units are configured receive data from a common register file. The vector units may independently and simultaneously process instructions. Furthermore, the vector units may be adapted to perform scalar operations thereby integrating the vector and scalar processing. The vector units may also be configured to share resources to perform an operation, for example, a cross product operation.

Abstract: A circuit arrangement, program product and circuit arrangement render stereoscopic images in a multithreaded rendering software pipeline using first and second rendering channels respectively configured to render left and right views for the stereoscopic image. Separate transformations are applied to received vertex data to generate transformed vertex data for use by each of the first and second rendering channels in rendering the left and right views for the stereoscopic image.

Abstract: A software-accessible special purpose register is architected into a processing unit in order to implement persistent vector multiplexer control of a vector-based execution unit. A persistent swizzle instruction is defined in an instruction set for the vector-based execution unit and is used to cause state information to be stored in the special purpose register such that the operand vectors processed by subsequent vector instructions executed by the vector-based execution unit will be selectively shuffled using the persisted state information. As a result, when multiple vector instructions require a common custom word ordering for one or more operand vectors, a single persistent swizzle instruction may be used to select the desired custom word ordering for all of the vector instructions.

Abstract: Persistent vector multiplexer control is used in a vector-based execution unit to control the shuffling of words in operand vectors processed by the execution unit. In addition, a persistent swizzle instruction is defined in an instruction set for the vector-based execution unit and is used to cause state information to be persisted such that the operand vectors processed by subsequent vector instructions executed by the vector-based execution unit will be selectively shuffled using the persisted state information. As a result, when multiple vector instructions require a common custom word ordering for one or more operand vectors, a single persistent swizzle instruction may be used to select the desired custom word ordering for all of the vector instructions.

Abstract: A circuit arrangement and method bypass the storage of requested data in a higher level cache of a multi-level memory architecture during the return of the requested data to a requester, while caching the requested data in a lower level cache. For certain types of data, e.g., data that is only used once and/or that is rarely modified or written back to memory, bypassing storage in a higher level cache reduces the likelihood of the requested data casting out frequently used data from the higher level cache. By caching the data in a lower level cache, however, the lower level cache can still snoop data requests and return requested data in the event the data is already cached in the lower level cache.

Abstract: The present invention is generally related to the field of image processing, and more specifically to an instruction set for processing images. Vector processing may involve rearranging vector operands in one or more source registers prior to performing vector operations. Typically, rearranging of operands in source registers is done by issuing a plurality of permute instructions that require excessive usage of temporary registers. Furthermore, the permute instructions may cause dependencies between instructions executing in a pipeline, thereby adversely affecting performance. Embodiments of the invention provide a level of muxing between a register file and a vector unit that allow for rearrangement of vector operands in source registers prior to providing the operands to the vector unit, thereby obviating the need for permute instructions.

Abstract: A method and circuit arrangement process a workload in a multithreaded processor that includes a plurality of hardware threads. Each thread receives at least one message carrying data to process the workload through a respective inbox from among a plurality of inboxes. A plurality of messages are received at a first inbox among the plurality of inboxes, wherein the first inbox is associated with a first thread among the plurality of hardware threads, and wherein each message is associated with a priority. From the plurality of received messages, a first message is selected to process in the first thread based on that first message being associated with the highest priority among the received messages. A second message is selected to process in the first thread based on that second message being associated with the earliest time stamp among the received messages and in response to processing the first message.

Abstract: By mapping leaf nodes of a spatial index to processing elements, efficient distribution of workload in an image processing system may be achieved. In addition, processing elements may use a thread table to redistribute workload from processing elements which are experiencing an increased workload to processing elements which may be idle. Furthermore, the workload experienced by processing elements may be monitored in order to determine if workload is balanced. Periodically the leaf nodes for which processing elements are responsible may be remapped in response to a detected imbalance in workload. By monitoring the workload experienced by the processing elements and remapping leaf nodes to different processing elements in response to unbalanced workload, efficient distribution of workload may be maintained. Efficient distribution of workload may improve the performance of the image processing system.

Abstract: According to embodiments of the invention, a distributed time base signal may be coupled to a memory directory which provides address translation for data located within a memory cache. The memory directory may have attribute bits which indicate whether or not the memory entries have been accessed by the distributed time base signal. Furthermore, the memory directory may have attribute bits which indicate whether or not a memory directory entry should be considered invalid after an access to the memory entry by the distributed time base signal. If the memory directory entry has been accessed by the distributed time base signal and the memory directory entry should be considered invalid after the access by the time base signal, any attempted address translation using the memory directory entry may cause a cache miss. The cache miss may initiate the retrieval of valid data from memory.

Abstract: One embodiment of the invention provides a method and apparatus for utilizing memory. The method includes reserving a first portion of a cache in a processor for an inbox. The inbox is associated with a first thread being executed by the processor. The method also includes receiving a packet from a second thread, wherein the packet includes an access request. The method further includes using inbox control circuitry for the inbox to process the received packet and determine whether to grant the access request included in the packet.

Abstract: A vectorizable execution unit is capable of being operated in a plurality of modes, with the processing lanes in the vectorizable execution unit grouped into different combinations of logical execution units in different modes. By doing so, processing lanes can be selectively grouped together to operate as different types of vector execution units and/or scalar execution units, and if desired, dynamically switched during runtime to process various types of instruction streams in a manner that is best suited for each type of instruction stream. As a consequence, a single vectorizable execution unit may be configurable, e.g., via software control, to operate either as a vector execution or a plurality of scalar execution units.

Abstract: A circuit arrangement, program product and method are provided for resetting a dynamically grown Accelerated Data Structure (ADS) used in image processing in which an ADS is initialized by reusing the root node of a prior ADS and resetting at least one node in the prior ADS to break a link between the reset node and a linked-to node in the prior ADS. By doing so, the memory allocated to the prior ADS may be reused for the new ADS, without having to clear or wipe out all of the allocated memory. In addition, in some instances, given the similarity of many image frames, often some or all of the node structure of a prior ADS may be reused for a new ADS, requiring only the contents of nodes to be cleared, instead of having to clear out all of the nodes in the prior ADS. As a result, the processing overhead associated with initializing a new ADS can be significantly reduced.

Abstract: Embodiments of the invention provide logic within the store data path between a processor and a memory array. The logic may be configured to misalign vector data as it is stored to memory. By misaligning vector data as it is stored to memory, memory bandwidth may be maximized while processing bandwidth required to store vector data misaligned is minimized. Furthermore, embodiments of the invention provide logic within the load data path which allows vector data which is stored misaligned to be aligned as it is loaded into a vector register. By aligning misaligned vector data as it is loaded into a vector register, memory bandwidth may be maximized while processing bandwidth required to align misaligned vector data may be minimized.

Abstract: A method, program product and system for conducting a ray tracing operation where the rendering compute requirement is reduced or otherwise adjusted in response to a changing vantage point. Aspects may update or reuse an acceleration data structure between frames in response to the changing vantage point. Tree and image construction quality may be adjusted in response to rapid changes in the camera perspective. Alternatively or additionally, tree building cycles may be skipped. All or some of the tree structure may be built in intervals, e.g., after a preset number of frames. More geometric image data may be added per leaf node in the tree in response to an increase in the rate of change. The quality of the rendering algorithm may additionally be reduced. A ray tracing algorithm may decrease the depth of recursion, and generate fewer cast and secondary rays. The ray tracer may further reduce the quality of soft shadows, resolution and global illumination samples, among other quality parameters.

Abstract: A method, program product and system for conducting a ray tracing operation where the rendering compute requirement is reduced by varying the size of bounding volumes into which image data is divided and/or by varying a number of primitives included within nodes of an acceleration data structure that correspond to the bounding volumes.

Abstract: A bus bridge between a high speed computer processor bus and a high speed output bus. The preferred embodiment is a bus bridge between a GPUL bus for a GPUL PowerPC microprocessor from International Business Machines Corporation (IBM) and an output high speed interface (MPI). Another preferred embodiment is a bus bridge in a bus transceiver on a multi-chip module.