This invention is related to computer processors, and in particular to computer processors having operands substantially wider than the data path width between the processor and memory.
We have previously described a programmable processor and method for improving the performance of programmable processors by enabling the use of operands that are larger than the internal data path width of the processor. See, e.g. U.S. Pat. No. 8,269,784 entitled “Processor Architecture for Executing Wide Transform Slice Instructions,” which is incorporated by reference herein. This prior invention uses the contents of general-purpose registers to specify operands stored in memory that are wider than the processor internal data path. The operands are typically a multiple of times wider than the data path and are preferably stored in adjacent rows of a memory. The registers also specify the memory address at which the plurality of widths of data can be read (or written).
The first time this operation is executed, twiddle factors are loaded at the rate of the available processor memory bus into a “coefficient RAM” embedded in the execution unit (or, in an alternative embodiment, are already present in embedded ROM), and data are loaded into embedded “wide cache” memories. Successive operations reuse the twiddle factors to perform successive slices of FFT on data buffered in the embedded cache.
The exemplary hardware for performing this operation is also shown in
The invention described below provides additional functionality to enable expandably wide operations with improved efficiency on a broader range of algorithms than our prior technology.
The expandably wide operations disclosed here reduce the influence of the characteristics of the associated processor in the design of functional units performing the calculations, including (1) the width of the register file or files, (2) the clock rate of the processor, (3) the exception and interruption subsystem of the processor, and (4) the sequence of operations involving loading the wide cache operand and using the contents of the wide cache, e.g. streaming of operands and chaining of operations.
Further, the specification of expandably wide operations is designed such that the degree of parallelism by which the functional unit executes these operations can be varied from implementation to implementation without changing the definition of the instruction itself or the software employing these instructions. As such, a later implementation may use a functional unit with greater parallelism than the initial implementation without requiring alterations to the software designed for the initial implementation. In an alternative embodiment, functional units with a variable degree of parallelism may control the degree of parallelism to achieve active power and performance goals, again, without the need to change the application software. Such a functional unit with variable parallelism may vary the degree of parallelism by optionally powering-down (removing active power from) a portion of the functional unit and associated wide operand memory blocks.
In one embodiment a processor for executing instructions requiring expandably wide operands executes instructions that include an operation code field defining an operation to be performed by the processor in which the operation requires an operand which is wider than the data path between the processor and memory. The instruction also includes a data input field specifying an address of a data input region in a memory containing data to be input for the operation to be performed, and a function field specifying an address for a function region in the memory containing information about the function to be performed by the processor. The instruction also specifies a state field which can be used to store information about the state of execution of the instruction, e.g. if execution is interrupted. At least one of the operation code field, the data input field, and the function field provides an address in the memory for a block of data required by the operation to be performed, which block of data is wider than the widest data path between the processor and the memory.
In another embodiment, a processor for executing instructions requiring expandably wide operands includes a operation code field defining an operation to be performed by the processor, the operation requiring an operand which is wider than a widest data path between the processor and the memory. A field in the expandably wide instruction specifies an address of a block of data in the memory. The data in the block of data include a descriptor field that specifies within that block, locations of at least input data to be used in the operation to be performed by the processor, and function data to be used in performing the operation on the input data. At least one of the input data and the function data is wider than the widest data path between the processor and the memory. If desired locations for state data and output data within the memory block can also be included within the descriptor field.
Other objects, features, and advantages of the present invention will become apparent upon consideration of the following detailed description and the accompanying drawings, in which like reference designations represent like features throughout the figures.
This invention enables expandably wide operations in which instructions and operands substantially wider than a processor-memory data path can be used by a processor. The processor may be a main processor in a computer system, or it may be one of a plurality of processors within a computer system that share a single memory system in a directly connected or distributed fashion. In
The instruction uses a plurality of fields 22-26. Field 22 provides an operation code (‘opcode’). Examples of the operations implemented by the instruction are described below in conjunction with Table 1, and elsewhere. The other register fields 23, 24, 25, and 26 specify registers whose content provides addresses, which addresses can be virtual or real addresses, to the processor's memory system for the instruction operands described below. The memory system referenced here may be a processor's main memory system or one of a plurality of a processor's memory systems, e.g. a cache memory. This memory system, in various alternative embodiments, can include instruction caches and/or data caches, in various hierarchies to achieve memory density and average access time performance goals.
In the expandably wide instruction, fields *in 23 and *out 26 specify addresses for a location in memory containing data to be input to the execution unit(s) 30 and a location for storing resulting output data from the execution unit(s) 30. (As notation herein, the asterisk ‘*’ is indicative of the declaration of an address variable in the C language.) The register field *function 24 specifies an address in memory which provides the function implemented by the instruction. The register field *state 25 specifies an address containing a description of the state of the operation. For example, in an exemplary embodiment of an expandably wide fast Fourier transform (EXWIDEFFT) operation, the function region contains the twiddle factor coefficients and/or a description of the type and manner in which the FFT is computed, such as specification of radix 2, radix 4, split radix 2/4, radix 8 or other butterfly schemes, and/or description of the precision, shifting, rounding and overflow strategies to perform. In this instruction, the state region contains control and status information, such as a specification of the intermediate progress of an ongoing FFT at the time that such an operation may be interrupted, as by a count of how many butterfly computations have been performed, or approximate values of the largest value encountered so far, for purposes of optimizing rounding and scaling of the intermediate FFT computation. The address of the memory region provided by the field is illustrated below the corresponding field by a block of memory, e.g. the *in field 23 provides the address of memory block 23a, the *out field the address of memory block 26a, etc.
In an alternative embodiment, a single *in/out field (rather than separate fields *in and *out) may specify a memory block used for both input and output. In addition, a single *function field may specify a memory block used for both function and status purposes. One or more of these fields may be duplicated, as when an instruction may have multiple input operands, or may be omitted. For example the *state field may be omitted if the instruction operation is short enough that it is not important if the operation may be interrupted. Another use of the state region 25a is to contain state information indicating that an operation may only be partially completed, as when such an instruction is interrupted to handle exception conditions arising within the instruction itself or arising from conditions from another source, such as an I/O event. Re-executing an instruction with state in this condition enables continuing the operation where it left off, enabling finishing the operation as if it was never interrupted. The memory block addressed by *state can be of arbitrary size, or absent, depending on the particular instruction. The *function field may be omitted if no further function specification is required beyond the one or more operation code fields of the instruction.
In an exemplary embodiment, the address of the memory block is used to control a transfer of memory data between the processor's memory system and these relatively smaller memory blocks indicated as 23a, 24a, 25a, and 26a. By managing these memory blocks as a cached copy of data in the processor's memory system, an initial execution of the instruction initiates copying of data into the memory blocks prior to, or at nearly the same time, as the initial execution of the instruction's operation. Subsequent executions of the same instruction or other instructions using the same memory address may be started without repeating the copying of the data, so long as the copy is still valid. Processor writes or writes initiated by other processors or other devices within the computer system may be tracked so that these writes cause the memory blocks to be kept up to date with respect to these operations, or in an alternative embodiment, to be marked as not valid or no longer valid, so that a subsequent instruction execution recopies the processor's memory data.
Cache coherency states commonly designated as M, O, E, S, I, or similar coherency schemes are employed by the exemplary embodiment in labeling these memory blocks (or portions of the memory blocks) to reduce the number of times data must be copied between the processor's memory system and these memory blocks. The physical design of these memory blocks is intended to accept such transfers at a rate governed by the design of the processor's memory system, while accesses by the functional unit may be at a substantially higher bandwidth, allowing the functional unit to perform at a greater bandwidth than if these operands were transferred directly from the processor's memory system.
In one embodiment, block 33 is a contiguous block of memory that contains the information from the separate fields. In alternative embodiments, the block can also contain virtual addresses or offsets to different areas of memory, or even different memories, that provide the needed values for the execution units 30 to implement the operation specified. The flexibility of the approach allows for future expansion, e.g. by adding information to the descriptor 34 to provide an offset to other areas of memory. The “other” field 36 indicates this. In an alternative embodiment, a small memory block or dedicated registers are designed to contain the descriptor information, and physically separate memory blocks, designed with appropriate bandwidth to the functional unit are employed to contain the *in, *out *function and *state data, to be filled via the cache coherent copying methods described herein. In this manner, the physical design continues to resemble that of
While memory block 33 preferably contains all of the information for the *in, *function, *state and *out, if one or more of those fields is not needed for a particular instruction, it need not be included within the block. Thus instead of using separate regions of memory 23a, 24a, 25a, and 26a as shown in
Once the location(s) in memory for the one or more fields *in, *function, *state and *out are identified, operand data from the processor's memory system is loaded into a cached memory block near or associated with the relevant execution unit of the processor. This cached memory block is sufficiently large to store the widest operand the processor can use in its execution units, and is not limited to the access bandwidth of the processor's memory system. Once the information is in the cache, the execution unit of processor can then retrieve (or store) the necessary *in, *function, *state and *out data for that operation. The wide cache memory is indicated in
Because the data to be transferred to or from the memory is substantially ‘wider’ than the data paths between the processor and memory, multiple cycles are used to move that data between the processor and memory. For example if the processor-memory data path is 128 bits wide, then moving an operand of 2048 bits requires 16 transfers over the processor-memory data path to place the 2048 bit operand in the cache memory associated with the functional unit used for execution of that operation.
The use of memory blocks with descriptors is not limited to just one memory block per instruction. As shown in
For example, when the memory block is designated by an address that is aligned to a power-of-2 size, there will be one or more zero bits at the least-significant end of the address. Using this property, by adding one-half of the power-of-2 size of the memory block, the address can encode both an address and further specify a size of the memory block. By limiting the minimum size of the memory block to a specified size, one can ensure that there are one or more least significant address bits that will be zero. For example, if the minimum size of the expandably wide operand is four bytes, after adding one-half of the memory block size (in this case 2) to the address, the combined address and size will have a least significant bit that is zero. For example, setting the least significant bit can specify the presence of a descriptor. If the least significant bit is cleared, then the descriptor can be considered as absent. Alternatively, this convention can be reversed, so that the least significant bit is set to specify no descriptor and cleared to specify a descriptor. The contents of the descriptor field will generally contain offsets and sizes within the memory block at which the individual sub-memory blocks are themselves contained. These offsets and sizes may be specified as items of a certain number of bytes each, leaving the remainder of the memory block available for the various sub-memory blocks.
In an alternative embodiment, the overall size of the memory block, including the sub-blocks, may be encoded in a fixed field within the memory block itself, so that the overall size is not limited to a power of two as described above. Sub-block sizes may be designated by adding one-half of the size to the offset field, providing a compact representation, but requiring sub-block alignment and power-of-2 sub-block sizes, or alternatively the sub-block sizes and offsets may be separately specified items, removing limitations of alignment and sizes of the sub-blocks. To avoid security implications of specifying offsets that are larger than the overall size of the memory block, offsets and sizes may trigger an exception if they are impermissibly large. Alternatively, the offsets and sizes may be masked by a bit-field comprising a value of the overall size minus one (a value that has one bits in each location where a one bit is permissible, and a zero bit in each location where a one bit is impermissible). This avoids a reference to a sub-block that is not within the overall block.
Table 1 below lists some exemplary instructions using expandably wide operands.
The approach described in
Other expandably wide instructions can include instructions which implement scatter or gather functions. For example, operations invoked by EXWIDESCATTER or EXWIDEGATHER can scatter data to, or gather data from, a list of addresses specified by the function region. In addition, EXWIDE instructions can be implemented that convert input and/or output data from one type to another, e.g. from fixed point to floating point, or vice versa, as well as between desired precisions or sizes.
In each case, the state region describes the size and shape of the input, output and function operands, and the details of the operation to be performed, including internal state and status of progress. As mentioned above, in some operations not all of the *in, *out, *function, and *state parameters are necessary. For example, for an EWIDE-MATRIX-MULTIPLY instruction, there may be separate *in parameters for each of the two matrices or vectors to be multiplied, while the *function parameter may be omitted. The *in parameters may contain descriptors that specify the size and shape of the matrix operand and precision of the operands. Alternatively, the *function parameter may specify among other things, that a matrix multiplication is to be performed, rather than other matrix operations, as well as the precision of the operands and the shape of the matrices to multiply.
In earlier patents, for example, our prior patent cited above, we described a wide operand specifier to enable retrieval of the wide operand from memory. The specifier there contains a pointer to a power-of-two aligned region of memory, together with additional information that specifies the size and/or shape of the wide operand. The additional information is encoded into the low-order bits that are assured to be zero due to the alignment. As described above, additional low-order bits that can be assured to be otherwise zero by specifying the minimum size of the operand may also be deployed to, for example, specify whether a descriptor block is present for the wide operand.
The expandably wide operand field described here is more flexible, using additional formats, as well as different formats for the wide operand itself, providing additional options in operand size and shape. For example, the wide operand descriptor here may point to a region of memory that has reduced requirements for alignment, or no particular alignment at all. In such a case, there may be smaller set of low-order bits that may contain additional information, such as size and shape of the operand. This may occur, for example, if there is a plurality of wide operands, and the size and/or shape of one wide operand can be inferred from the size and/or shape of other wide operands, such as when a filter vector or matrix, in an aligned wide operand, is convolved against a not-aligned wide operand.
Additionally, information may be contained within the memory region that further delineates the portion of the wide operand containing relevant operands. For example, in various alternative embodiments, a descriptor block within the wide operand memory region may contain a starting offset and ending offset, or a starting offset and a length, or a length and an ending offset. The lengths and offsets may be measured in bytes, multiples of the element size, or otherwise. The descriptor block may be contained at a fixed offset, such as the beginning of the specified memory region, or at the end of the specified memory region, or may be located by an offset or length itself indicated at a fixed offset. By embedding the descriptor block within the memory region specified by the expandably wide operand descriptor, a wide operand source or destination is no longer limited to a power-of-two overall size.
In various alternative embodiments, e.g. as illustrated in
The shape data can identify a row or column size of a two-dimensional array of elements, and may be generalized to specification of the number of dimensions, not limited to two, but may be one, two, three, or any maximum number of dimensions, with the size in bytes, bits, or elements in each of the dimensions. In an alternative embodiment, dimensions may be implied by knowledge of the overall size of the relevant expandably wide operand, or the overall size may be inferred from the knowledge of the size of each dimension of the array.
Furthermore, the relevant operand may use only a portion of the expandably wide operand memory region. The remaining region may contain additional buffer space that the processor may write into to prepare the region for operations following, e.g. an EXWIDE instruction, without disrupting the operation of the EXWIDE instruction itself. For example, the EWIDEHUFFMAN instruction may decode a stream starting at a specified offset, and continuing until the result operand space is exhausted, the input stream is exhausted, or the operation is otherwise terminated.
A further benefit of this invention is that the processor may perform write operations that append additional symbols to the input stream without disrupting an operating instruction, e.g. an EWIDEHUFFMAN instruction. For example, the input stream can be represented as a circular buffer with offset values in the descriptor block marking the bounds of the circular buffer. Writing to the descriptor block in such a way as to add valid symbols to the input stream allows the EWIDEHUFFMAN to continue beyond the bounds that would have stopped it if the descriptor block of the input stream had not been changed. Similarly, in an alternative embodiment, writing to the descriptor block to remove output symbols from the result operand and thereby free-up space in the result operand may allow the EWIDEHUFFMAN to continue beyond the bounds that it would have stopped if the descriptor block of the input stream had not been changed. In addition, the descriptor block of both the input stream and of the output stream can be updated by the EXWIDEHUFFMAN instruction to represent the extent to which the input stream has been consumed, and the output stream has been filled at the completion of the instruction. Alternatively, the descriptor block may be updated at intermediate stages to permit simultaneous filling of the input stream or emptying of the output stream as the instruction progresses. In an alternative embodiment, the descriptor block of the input and/or output stream, or the descriptor block of a separate “*state” memory region controls the frequency of such intermediate updates. The EWIDEHUFFMAN instruction illustrates how other similarly stream-operating instructions can be employed.
By specifying these dimensions, sizes and element sizes in the descriptor block, the information within the expandably wide operand becomes essentially self-describing, so that this information no longer need be specified by the instruction, whether specified in the instruction itself, or within the value of register(s) specified by the instruction. Also, the values of dimensions, sizes and element sizes may imply the values of dimensions, sizes and element sizes of the result operand. The instruction using a wide operand may write these values into the memory region of the result operand, or in an alternative embodiment, may write these values into a descriptor region, or a region representing function or state.
In an alternative embodiment of an expandably wide instruction, e.g. the EWIDEHUFFMAN instruction, as a first cache coherence maintenance strategy, the descriptor blocks for the input and/or output streams are maintained in a cache coherence state that is exclusive to the functional unit while the functional unit is performing the instruction. If a processor, either the processor that initiated the instruction, or any other processor or I/O device attempts to read the descriptor block, the functional unit can provide a copy that reflects the current intermediate state of processing the input and output streams. Alternatively, an element of the descriptor block may indicate whether the functional unit is or is not actively processing the stream, or in other words, whether the instruction has completed. Reads of the output stream or writes to the input stream of values that the newly updated descriptor block indicates would be valid can proceed safely once the relevant descriptor block had been read.
Furthermore as an alternative cache coherence maintenance strategy, the handling of the descriptor blocks with respect to the cache coherence system is to suspend, block or delay reads of any portion of an output stream that may be changed by the instruction until a final value can be established. Reads of the output stream of portions of the output memory space for which the final values are present can proceed normally. Writes to the input stream are only permitted to the region of the input stream not actively in use by the instruction. In this manner, the architectural state of the memory is maintained, yet some operations to the input and output wide operands can proceed in parallel with the operation of the instruction. In other words, the state of the wide operands in memory is maintained as if the instruction is completed all at once, with none of the intermediate state visible. Nevertheless, after initiating such an operation, results from a previous operation can be read from the output stream, and new values for a subsequent invocation of the instruction can be written into the input stream, all while the operation is in progress. It is only when the descriptors are touched, that operations are serialized to maintain values as if the operation was performed all at once.
Programmers can choose appropriate cache coherence maintenance strategies, e.g., by a control field in one or more of the instructions, a register referenced by the instruction, or an input or output descriptor or the *function or *state wide operands. This approach can also be used in other operations that proceed in a highly serial fashion, for example, finite impulse response (FIR) and infinite impulse response (IIR) filters.
When multiplying matrices, mathematical convention is that the matrices are combined by a sum of products along specified dimensions of the left and right operand arrays, often a final dimension of the left and an initial dimension the right. It is also common, however, that matrices that are multiplied may need to be transposed or in some other way, reshaped to accommodate such convention. For example, the EXWIDEMATRIX instruction may include in the descriptor block of either operand, or a register operand or within the instruction, information that leads to an effective transposition or reshaping of the arrays to be multiplied, as they are multiplied. As another example, an EXWIDEMATRIX-MULTIPLY may be configured to transpose the left operand before combining the operands in a conventional matrix multiplication. Alternatively, the EXWIDEMATRIX instruction may specify dimensions of either or both of the left and/or right operand arrays to combine with a sum of products that may be different than the mathematical convention—this may have a similar effect to transposing either or both of the left and/or right operand arrays.
Another example of an expandably wide instruction is an EXWIDERESHAPE instruction that transposes or otherwise reshapes an array operand, copying the source operand data into a new result operand, or placing the result operand in place of the source operand into the same memory region. A separate EXWIDEEXTRACT instruction extracts a smaller elements size result from a wide operand using parameters specified by a parameter register or parameters specified in a wide operand or a descriptor block of a wide operand. The functions of the EXWIDERESHAPE instruction may be combined with the EXWIDEEXTRACT instruction to both transpose, or otherwise reshape an array operand, and extract elements from the source array. In an alternative embodiment, the functions of either or both of the EXWIDERESHAPE instruction and the EXWIDEEXTRACT instruction may be combined with other EXWIDE instructions, such as EXWIDEMATRIX-MULTIPLY, EXWIDEMATRIX-MULTIPLY-ADD or EXWIDEMATRIX-ADD.
The dimension, size, shape and/or element size parameters can be specified by a descriptor written into the result expandably wide operand prior to execution of the EXWIDE instruction that fills in the values. In such an embodiment, the prior value of the descriptor block influences the operation so as to produce values consistent with the parameter values. If desired, the descriptor block can control the operation of transposing or otherwise reshaping, extracting, or combining along a specified array dimension with a sum of products. Some of the descriptor block values influence the operand, while others are forced to values consistent with the input operands so as to produce a valid result.
The descriptor block values written by an EXWIDE instruction make the result self-describing, in that the combination of the descriptor block and the data that the descriptor block describes are collected together into a single expandably wide memory block that can be treated as a self-contained and self-described data structure. In addition, the value of the entire expandably wide operand memory region may be directly used as the source of a second EXWIDE instruction. In an alternative embodiment, operation of an EXWIDE instruction preferentially proceeds to fill in the descriptor block of the result operand as early as possible, but at least prior to completion of production of all the result values. This allows a second EXWIDE instruction that uses the prior wide operand result as a wide operand source to proceed and begin to make computational progress prior to completion of the first EXWIDE instruction. In an alternative embodiment, the use of MOESI or similar cache-coherence protocols are employed to allow the second operation to obtain the wide operand source values in a controlled manner.
Because of decoupling from register-related complications, an EXWIDEFFT operation can be extended to more powerful blocks of FFT computation—such as an entire 256-point complex FFT (rather than just a single slice), or a 1024-point or 4096-point FFT. These same EXWIDEFFT operations can be performed on implementations that scale down to a single multiplier, or up to the 16-multiplier or a 64-multiplier design or to larger designs capable to as much as an entire FFT stage or even the entire FFT at once. At large scale, multipliers may incorporate fixed twiddle factor values, so that multipliers may be implemented as multiply-by-constant hardware rather than a more general and generally larger two-operand multiplier. In an alternative embodiment, the same operation may be performed in a function unit that contains as little as a single integer multiplier-adder, to minimize power expenditure while maintaining compatibility with embodiments with large amount of parallelism.
One important advantage of expandably wide operations is that the size and shape of source and result operands can differ to accommodate greater precision in calculations. For example, in EXWIDEMATRIX the result operands could be double, or even quadruple, size compared to the source operands, and the source operands do not have to be same size, e.g. multiplying an 8 bit matrix by a 16 bit matrix can have a 32 bit result with no rounding and no overflows.
The EXWIDEMATRIX instruction may be implemented as a EXWIDEMATRIX-MULTIPLY instruction, which performs matrix-vector or matrix-matrix multiplication, using size and shape parameters taken from the wide operand descriptors or the *function or *state parameters, and using precision and rounding parameters specified from the *function or “*state parameters or from a sub-field of the instruction. In an alternative embodiment, the EXWIDEMATRIX instruction may be implemented as separate EXWIDEMATRIX-MULTIPLY-MATRIX-VECTOR and EXWIDEMATRIX-MULTIPLY-MATRIX-MATRIX instructions for the matrix-vector and matrix-matrix cases. In another alternative embodiment, the EXWIDEMATRIX instruction may be implemented as an EXWIDEMATRIX-MULTIPLY-VECTOR-VECTOR instruction that performs element-by-element vector multiplication, or the EXWIDEMATRIX instruction may be implemented as an EXWIDEMATRIX-ADD instruction, performing matrix-to-matrix addition or vector-to-vector addition using similar parameters as the EXWIDEMATRIX-MULTIPLY.
In further embodiments, the EXWIDEMATRIX instruction may be implemented as a EXWIDEMATRIX-MULTIPLY-ADD instruction, combining the matrix-vector or matrix-matrix multiplication with a matrix-to-matrix addition in a single instruction, without possibility of losing intermediate precision or range between the multiplication and addition operations. Still further, the EXWIDEMATRIX instruction may be implemented as a EXWIDEMATRIX-MULTIPLY-SUM instruction that performs a multiply-sum operation, combining an element-by-element multiplication with a horizontal summation of all the products, or detailed as an instruction that performs a combination of the above cases in a single instruction, using register or ex-wide-operand parameters to distinguish the cases.
The memory regions addressed by an EXWIDE operation are transferred into wide caches embedded in the corresponding EXWIDE execution unit. The input and output caches benefit when there is a high degree of locality: that is, once a single data item in a cache line is touched, the other items will be touched as well. The function and state caches benefit from a high degree of reuse: that is, once function and state data are fetched from the memory hierarchy into cache they are reused many times. Typical caches comprise single or dual ported memory cells (or even ROM cells, as in the case of FFT roots of unity, so called twiddle factors), and the number of cache entries needed are modest (e.g., 1-8 entries), so area and power dissipation are lower than multi-ported register file access would be for the same operand bandwidth.
Expandably wide operation hardware can approach the low area and power of ASICs, while streamlining integration with a general-purpose host processor. For example, an EWIDEHUFFMAN execution unit can have a narrow (e.g., 32-bit integer or smaller) data path specialized for Huffman decoding, with power gated off when idle. By issuing a single EWIDEHUFFMAN instruction, the host processor launches coding of a long stream (e.g. thousands) of symbols within the Huffman unit autonomously, while the host proceeds with other instructions. The host can query Huffman status by reading the state memory region or a memory-mapped event register. Alternatively, the event register can have an associated memory-mapped mask that enables interrupts to signal the host.
A multiplicity of execution units can implement a particular expandably wide operation. For example for server applications that require high (de)compression performance, multiple small serial Huffman units could each be initiated, each with a single EWIDEHUFFMAN instruction, and then be capable of independently processing multiple long streams of symbols in parallel.
Expandably wide execution units can operate asynchronously at a different clock rate than the host processor, using the same mechanisms that accommodate asynchrony in the memory hierarchy. For example, an EXWIDEFILTER execution unit can perform an IIR or FIR filter on data deposited into a circular buffer region of memory at a clock rate synchronized to an external mixed-signal interface, to provide a precise real-time feedback signal for reliable locking of the interface.
The memory contents fetched from the memory system are combined with execute instructions not performed by the access units, and entered into four execute instruction units (labeled “E-Unit”) 105. For expandably wide instructions, memory contents fetched from the memory system are also provided to the expandably wide operand caches 116 by bus 120. Within the E-Units 115, instructions and memory data are presented to execution register files that fetch execution register file source operands. The instructions are coupled to the arbitration unit 124 that selects which instructions from the four threads are to be routed to the available execution functional units 130, 132, 134, 136, and 138. Each execution functional unit includes a cache memory 116 for holding an expandably wide operand. The execution functional units 130-138 are arithmetic and logic units that are optimized to perform the various operations using the expandably wide operands. For example, the matrix execution functional unit is a unit designed to perform operations on matrices. A field in the opcode can specify the execution functional unit used for a particular instruction.
The execution functional units 130-138 each contain dedicated storage 116 to permit storage of source operands, including the expandably wide operands discussed above. These cache memories typically have a width that is a multiple of the width of data path 140 between the processor and memory. Thus, if the width of the data path 140 is 128 bits, the dedicated storage 116 preferably has a width which is a multiple of that 128 but width, e.g., a width of 256, 512, 1024 or 2048 bits. Of course depending on the particular function to be performed and the data associated therewith, some operands will utilize the full width of the dedicated storage, while others will not. As mentioned above, transfer of these expandably wide operands from the memory 108 to the storage 116 associated with the individual functional units will require multiple cycles in which portions of expandably wide operands are loaded sequentially into the dedicated storage 116. Once loaded, however, the expandably wide operands may then be used at substantially the same time.
This description of the invention has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form described, and many modifications and variations are possible in light of the teaching above. The embodiments were chosen and described to explain the principles of the invention and its practical applications. This description will enable others skilled in the art to best utilize and practice the invention in various embodiments and with various modifications as are suited to a particular use. The scope of the invention is defined by the following claims.
This application is a continuation of U.S. patent application Ser. No. 14/749,955, filed on Jun. 25, 2015; which claims priority from U.S. Provisional Patent Application No. 62/019,101, filed Jun. 30, 2014, and entitled, “Processor Architecture and Methods for Expandably Wide Operations with Improved Efficiency”. Each of these references is hereby incorporated in its entirety for all purposes.
Number | Name | Date | Kind |
---|---|---|---|
3581279 | Arbuckle | May 1971 | A |
3833889 | Cray | Sep 1974 | A |
4068300 | Bachman | Jan 1978 | A |
4251875 | Marver et al. | Feb 1981 | A |
4353119 | Daniel et al. | Oct 1982 | A |
4393468 | New et al. | Jul 1983 | A |
4658349 | Tabata et al. | Apr 1987 | A |
4658908 | Hannukainen | Apr 1987 | A |
4785393 | Chu et al. | Nov 1988 | A |
4823259 | Aichelmann et al. | Apr 1989 | A |
4930106 | Danilenko et al. | May 1990 | A |
5031135 | Patel et al. | Jul 1991 | A |
5170399 | Cameron et al. | Dec 1992 | A |
5185861 | Valencia et al. | Feb 1993 | A |
5280598 | Osaki et al. | Jan 1994 | A |
5283886 | Nishii et al. | Feb 1994 | A |
5325493 | Herrell et al. | Jun 1994 | A |
5333280 | Ishikawa et al. | Jul 1994 | A |
5375215 | Hanawa et al. | Dec 1994 | A |
5426379 | Trimberger | Jun 1995 | A |
5430556 | Ito et al. | Jul 1995 | A |
5471593 | Branigin et al. | Nov 1995 | A |
5481686 | Dockser | Jan 1996 | A |
5487024 | Girardeau et al. | Jan 1996 | A |
5509137 | Itomitsu et al. | Apr 1996 | A |
5535225 | Mayhew et al. | Jul 1996 | A |
5550988 | Sarangdhar et al. | Aug 1996 | A |
5551005 | Sarangdhar et al. | Aug 1996 | A |
5574939 | Keckler et al. | Nov 1996 | A |
5579253 | Lee et al. | Nov 1996 | A |
5598362 | Adelman et al. | Jan 1997 | A |
5600814 | Gahan et al. | Feb 1997 | A |
5604864 | Noda | Feb 1997 | A |
5636363 | Bourekas et al. | Jun 1997 | A |
5646626 | Willis et al. | Jul 1997 | A |
5669012 | Shimizu et al. | Sep 1997 | A |
5671170 | Markstein et al. | Sep 1997 | A |
5675526 | Peleg et al. | Oct 1997 | A |
5717946 | Satou et al. | Feb 1998 | A |
5721892 | Peleg et al. | Feb 1998 | A |
5740093 | Sharangpani et al. | Apr 1998 | A |
5742840 | Moussouris et al. | Apr 1998 | A |
5745729 | Greenley et al. | Apr 1998 | A |
5745778 | Alfieri et al. | Apr 1998 | A |
5752001 | Dulong et al. | May 1998 | A |
5752264 | Blake et al. | May 1998 | A |
5765216 | Weng et al. | Jun 1998 | A |
5768546 | Kwon et al. | Jun 1998 | A |
5778412 | Gafken et al. | Jul 1998 | A |
5799165 | Favor et al. | Aug 1998 | A |
5802336 | Mittal et al. | Sep 1998 | A |
5826079 | Boland et al. | Oct 1998 | A |
5826081 | Zolnowsky et al. | Oct 1998 | A |
5835744 | Tran et al. | Nov 1998 | A |
5835782 | Lin et al. | Nov 1998 | A |
5835968 | Mahalingaiah et al. | Nov 1998 | A |
5872972 | Boland et al. | Feb 1999 | A |
5889983 | Mittal et al. | Mar 1999 | A |
5933627 | Parady | Aug 1999 | A |
5933650 | van Hook et al. | Aug 1999 | A |
5935240 | Mennemeier et al. | Aug 1999 | A |
5940859 | Bistry et al. | Aug 1999 | A |
5991531 | Song et al. | Nov 1999 | A |
5999959 | Weng et al. | Dec 1999 | A |
6006299 | Wang et al. | Dec 1999 | A |
6038675 | Gabzdyl et al. | Mar 2000 | A |
6041404 | Abdallah et al. | Mar 2000 | A |
6058408 | Fischer et al. | May 2000 | A |
6061780 | Shippy et al. | May 2000 | A |
6105053 | Kimmel et al. | Aug 2000 | A |
6131145 | Matsubara et al. | Oct 2000 | A |
6134635 | Reams et al. | Oct 2000 | A |
6141384 | Wittig et al. | Oct 2000 | A |
6141675 | Slavenburg et al. | Oct 2000 | A |
6170051 | Dowling | Jan 2001 | B1 |
6211892 | Huff et al. | Apr 2001 | B1 |
6212618 | Roussel et al. | Apr 2001 | B1 |
6237016 | Fischer et al. | May 2001 | B1 |
6243803 | Abdallah et al. | Jun 2001 | B1 |
6260135 | Yoshida | Jul 2001 | B1 |
6263428 | Nonomura et al. | Jul 2001 | B1 |
6266758 | van Hook et al. | Jul 2001 | B1 |
6269390 | Boland | Jul 2001 | B1 |
6292815 | Abdallah et al. | Sep 2001 | B1 |
6292879 | Fong | Sep 2001 | B1 |
6295599 | Hansen et al. | Sep 2001 | B1 |
6317824 | Thakkar et al. | Nov 2001 | B1 |
6351801 | Christie et al. | Feb 2002 | B1 |
6370559 | Hoffman et al. | Apr 2002 | B1 |
6377970 | Abdallah et al. | Apr 2002 | B1 |
6378060 | Moussouris et al. | Apr 2002 | B1 |
6385634 | Peleg et al. | May 2002 | B1 |
6408325 | Shaylor et al. | Jun 2002 | B1 |
6418529 | Roussel et al. | Jul 2002 | B1 |
6426746 | Hsieh et al. | Jul 2002 | B2 |
6438660 | Reams et al. | Aug 2002 | B1 |
6453368 | Yamamoto et al. | Sep 2002 | B2 |
6463525 | Prabhu et al. | Oct 2002 | B1 |
6470370 | Fischer et al. | Oct 2002 | B2 |
6516406 | Eitan et al. | Feb 2003 | B1 |
6567908 | Furuhashi et al. | May 2003 | B1 |
6631389 | Lin et al. | Oct 2003 | B2 |
6633897 | Peek et al. | Oct 2003 | B1 |
6725356 | Hansen et al. | Apr 2004 | B2 |
6766515 | Bitar et al. | Jul 2004 | B1 |
6804766 | Noel et al. | Oct 2004 | B1 |
8269784 | Hansen et al. | Sep 2012 | B2 |
9582281 | Kruecken | Feb 2017 | B2 |
20010021971 | Gibson et al. | Sep 2001 | A1 |
20090089540 | Hansen et al. | Apr 2009 | A1 |
20100325188 | Olson et al. | Dec 2010 | A1 |
20130173888 | Hansen et al. | Jul 2013 | A1 |
20130283019 | Ould-ahmed-vall et al. | Oct 2013 | A1 |
20150378734 | Hansen et al. | Dec 2015 | A1 |
Number | Date | Country |
---|---|---|
1653446 | Aug 2005 | CN |
212571 | Mar 1987 | EP |
651514 | May 1995 | EP |
800280 | Oct 1997 | EP |
1024603 | Aug 2000 | EP |
1102161 | May 2001 | EP |
03098145 | Apr 1991 | JP |
06149723 | May 1994 | JP |
07114496 | May 1995 | JP |
Entry |
---|
U.S. Appl. No. 14/749,955 , “First Action Interview Pilot Program Pre-Interview Communication”, dated Mar. 9, 2017, 18 pages. |
U.S. Appl. No. 14/749,955 , “Notice of Allowance”, dated Jun. 13, 2017, 20 pages. |
U.S. Control No. 95/000,100 , “Office Action in inter partes Reexamination”, dated Mar. 19, 2009. |
U.S Control No. 95/000,100 , “Right of Appeal Notice (37 CFR 1.953) in inter partes Reexamination”, Jul. 11, 2009. |
European Application No. EP10191073 , “Extended European Search Report”, dated Jan. 17, 2011. |
European Application No. EP10191073 , “Summons to Attend Oral Proceedings for European Patent Office Application”, Dec. 4, 2012. |
Gwennap , “UltraSparc Adds Multimedia Instructions”, Microprocessor Report, vol. 8, No. 6, Dec. 5, 1994, pp. 1-3. |
Hansen , “Architecture of a Broadband Mediaprocessor”, Proceedings of Compcon, IEEE Computer Science, 1996, pp. 334-340. |
Hansen , “MicroUnity's MediaProcessor architecture”, IEEE Micro, vol. 16, No. 4, Aug. 1996, pp. 34-41. |
Hendrix , “Viterbi Decoding Techniques for the TMS320C54x Family”, Texas Instruments, Jan. 2002. |
Lee et al., “An efficient recursive cell architecture of modified Euclid's algorithm for decoding Reed-Solomon codes”, IEEE Transactions on Consumer Electronics, vol. 48, No. 4, Nov. 2002, pp. 845-849. |
Lee , “High-speed VLSI architecture for parallel Reed-Solomon decoder”, IEEE Transactions on Very Large Scale Integration (VLSI) Systems vol. 11, No. 2, Apr. 2003, pp. 288-294. |
Leijten-Nowak et al., “An FPGA architecture with enhanced datapath functionality”, Proceedings of the 2003 ACM/SIGDA eleventh international symposium on Field Programmable Gate Arrays, Feb. 2003, pp. 195-204. |
International Application No. PCT/US2015/038078 , “International Preliminary Report on Patentability”, dated Jan. 12, 2017, 7 pages. |
International Application No. PCT/US2015/038078 , “International Search Report and Written Opinion”, dated Sep. 29, 2015, 9 pages. |
Rice , “Multiprecision division on an 8-bit processor”, Proceedings of the 13th IEEE Symposium on Computer Arithmetic,, dated Jul. 1997, pp. 74-81. |
Welch , “A fixed-point fast Fourier transform error analysis”, IEEE Transactions on Audio and Electroacoustics, vol. 17 , No. 2, Jun. 1969, pp. 151-157. |
European Patent Application No. EP15815691.9 , “Extended European Search Report Received”, dated Nov. 27, 2017, 10 pages. |
Chinese Patent Application No. 201580036141.1, “Office Action”, dated Sep. 20, 2018, 12 pages. |
Number | Date | Country | |
---|---|---|---|
20180173635 A1 | Jun 2018 | US |
Number | Date | Country | |
---|---|---|---|
62019101 | Jun 2014 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 14749955 | Jun 2015 | US |
Child | 15698180 | US |