US20180005059A1 - Statistics Operations On Two Dimensional Image Processor - Google Patents

Abstract

A method is described that includes loading an array of content into a two-dimensional shift register. The two-dimensional shift register is coupled to an execution lane array. The method includes repeatedly performing a first sequence that includes: shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content. The method also includes repeatedly performing a second sequence that includes: shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column. The resultant row or column has values determined at least in part from the mathematical operations of the first sequence. The second sequence further includes performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.

Description

FIELD OF INVENTION
• The description pertains generally to image processing, and, more specifically, to statistics operations on a two dimensional image processor.
BACKGROUND
• Image processing typically involves the processing of pixel values that are organized into an array. Here, a spatially organized two dimensional array captures the two dimensional nature of images (additional dimensions may include time (e.g., a sequence of two dimensional images) and data type (e.g., colors)). In a typical scenario, the arrayed pixel values are provided by a camera that has generated a still image or a sequence of frames to capture images of motion. Traditional image processors typically fall on either side of two extremes.
• A first extreme performs image processing tasks as software programs executing on a general purpose processor or general purpose-like processor (e.g., a general purpose processor with vector instruction enhancements). Although the first extreme typically provides a highly versatile application software development platform, its use of finer grained data structures combined with the associated overhead (e.g., instruction fetch and decode, handling of on-chip and off-chip data, speculative execution) ultimately results in larger amounts of energy being consumed per unit of data during execution of the program code.
• A second, opposite extreme applies fixed function hardwired circuitry to much larger blocks of data. The use of larger (as opposed to finer grained) blocks of data applied directly to custom designed circuits greatly reduces power consumption per unit of data. However, the use of custom designed fixed function circuitry generally results in a limited set of tasks that the processor is able to perform. As such, the widely versatile programming environment (that is associated with the first extreme) is lacking in the second extreme.
• A technology platform that provides for both highly versatile application software development opportunities combined with improved power efficiency per unit of data remains a desirable yet missing solution.
SUMMARY
• A method is described that includes loading an array of content into a two-dimensional shift register. The two-dimensional shift register is coupled to an execution lane array. The method includes repeatedly performing a first sequence that includes: shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content. The method also includes repeatedly performing a second sequence that includes: shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column. The resultant row or column has values determined at least in part from the mathematical operations of the first sequence. The second sequence further includes performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.
• An apparatus is described having means for performing the first sequence above and for performing the second sequence above.
LIST OF FIGURES
• The following description and accompanying drawings are used to illustrate various embodiments. In the drawings:
• FIG. 1 shows an embodiment of an image processor hardware architecture;
• FIGS. 2a, 2b, 2c, 2dand 2edepict the parsing of image data into a line group, the parsing of a line group into a sheet and the operation performed on a sheet with overlapping stencils;
• FIG. 3 shows an embodiment of a stencil processor;
• FIG. 4 shows an embodiment of a data computation unit within a stencil processor;
• FIGS. 5a, 5b, 5c, 5d, 5e, 5f, 5g, 5h, 5i, 5jand 5kdepict an example of the use of a two-dimensional shift array and an execution lane array to determine a pair of neighboring output pixel values with overlapping stencils;
• FIG. 6 shows an embodiment of a unit cell for an integrated execution lane array and two-dimensional shift array;
• FIG. 7 depicts tiles within an image frame;
• FIG. 8 depicts raster processing across tile boundaries;
• FIG. 9 depicts processing across sheet boundaries;
• FIG. 10 depicts processing across sheet and tile boundaries with associated context switching;
• FIG. 11 shows accumulated values within a two dimensional shift register array;
• FIGS. 12athrough 12edepict a reduction process;
• FIG. 12fshows window statistics tracking;
• FIG. 13 shows random access memories of a stencil processor being allocated to groups of execution lanes;
• FIG. 14 shows a histogram;
• FIG. 15ashows a per execution lane allocation of memory space;
• FIG. 15bshows a merged per executional lane allocation of memory space;
• FIG. 16 shows first and second groups of execution lanes being allocated to first and second merged memory space regions;
• FIGS. 17athrough 17cshow a histogram reduction process;
• FIGS. 18aand 18bshow a mechanism for expanding the size of a histogram;
• FIG. 19 shows an integrated camera system and processor;
• FIG. 20 shows a statistics method performed by a two dimensional shift register;
• FIG. 21 shows a computing system.
DETAILED DESCRIPTIONa. Image Processor Hardware Architecture and Operation
• FIG. 1 shows an embodiment of anarchitecture100 for an image processor implemented in hardware. The image processor may be targeted, for example, by a compiler that converts program code written for a virtual processor within a simulated environment into program code that is actually executed by the hardware processor. As observed inFIG. 1, thearchitecture100 includes a plurality of line buffer units101_1 through101_M (hereinafter “line buffers”, “line buffer units” or the like) interconnected to a plurality of stencil processor units102_1 through102_N (hereinafter “stencil processors”, “stencil processor units” or the like) and corresponding sheet generator units103_1 through103_N (hereinafter “sheet generators”, “sheet generator units” or the like) through a network104 (e.g., a network on chip (NOC) including an on chip switch network, an on chip ring network or other kind of network). In an embodiment, any line buffer unit may connect to any sheet generator and corresponding stencil processor through thenetwork104.
• In an embodiment, program code is compiled and loaded onto acorresponding stencil processor102 to perform the image processing operations earlier defined by a software developer (program code may also be loaded onto the stencil processor's associated sheet generator103, e.g., depending on design and implementation). In at least some instances an image processing pipeline may be realized by loading a first kernel program for a first pipeline stage into a first stencil processor102_1, loading a second kernel program for a second pipeline stage into a second stencil processor102_2, etc. where the first kernel performs the functions of the first stage of the pipeline, the second kernel performs the functions of the second stage of the pipeline, etc. and additional control flow methods are installed to pass output image data from one stage of the pipeline to the next stage of the pipeline.
• In other configurations, the image processor may be realized as a parallel machine having two or more stencil processors102_1,102_2 operating the same kernel program code. For example, a highly dense and high data rate stream of image data may be processed by spreading frames across multiple stencil processors each of which perform the same function.
• In yet other configurations, essentially any directed acyclic graph (DAG) of kernels may be loaded onto the hardware processor by configuring respective stencil processors with their own respective kernel of program code and configuring appropriate control flow hooks into the hardware to direct output images from one kernel to the input of a next kernel in the DAG design.
• As a general flow, frames of image data are received by a macro I/O unit105 and passed to one or more of theline buffer units101 on a frame by frame basis. A particular line buffer unit parses its frame of image data into a smaller region of image data, referred to as a “line group”, and then passes the line group through thenetwork104 to a particular sheet generator. A complete or “full” singular line group may be composed, for example, with the data of multiple contiguous complete rows or columns of a frame (for brevity the present specification will mainly refer to contiguous rows). The sheet generator further parses the line group of image data into a smaller region of image data, referred to as a “sheet”, and presents the sheet to its corresponding stencil processor.
• In the case of an image processing pipeline or a DAG flow having a single input, generally, input frames are directed to the same line buffer unit101_1 which parses the image data into line groups and directs the line groups to the sheet generator103_1 whose corresponding stencil processor102_1 is executing the code of the first kernel in the pipeline/DAG. Upon completion of operations by the stencil processor102_1 on the line groups it processes, the sheet generator103_1 sends output line groups to a “downstream” line buffer unit101_2 (in some use cases the output line group may be sent_back to the same line buffer unit101_1 that earlier had sent the input line groups).
• One or more “consumer” kernels that represent the next stage/operation in the pipeline/DAG executing on their own respective other sheet generator and stencil processor (e.g., sheet generator103_2 and stencil processor102_2) then receive from the downstream line buffer unit101_2 the image data generated by the first stencil processor102_1. In this manner, a “producer” kernel operating on a first stencil processor has its output data forwarded to a “consumer” kernel operating on a second stencil processor where the consumer kernel performs the next set of tasks after the producer kernel consistent with the design of the overall pipeline or DAG.
• Astencil processor102 is designed to simultaneously operate on multiple overlapping stencils of image data. The multiple overlapping stencils and internal hardware processing capacity of the stencil processor effectively determines the size of a sheet. Here, within astencil processor102, arrays of execution lanes operate in unison to simultaneously process the image data surface area covered by the multiple overlapping stencils.
• As will be described in more detail below, in various embodiments, sheets of image data are loaded into a two-dimensional register array structure within thestencil processor units102. The use of sheets and the two-dimensional register array structure is believed to effectively provide for power consumption improvements by moving a large amount of data into a large amount of register space as, e.g., a single load operation with processing tasks performed directly on the data immediately thereafter by an execution lane array. Additionally, the use of an execution lane array and corresponding register array provide for different stencil sizes that are easily programmable/configurable.
• FIGS. 2athrough 2eillustrate at a high level embodiments of both the parsing activity of aline buffer unit101, the finer grained parsing activity of a sheet generator unit103, as well as the stencil processing activity of thestencil processor102 that is coupled to the sheet generator unit103.
• FIG. 2adepicts an embodiment of an input frame ofimage data201.FIG. 2aalso depicts an outline of three overlapping stencils202 (each stencil having a dimension of 3 pixels by 3 pixels) that a stencil processor is designed to operate over. The output pixel that each stencil respectively generates output image data for is highlighted in solid black. For brevity, the three overlappingstencils202 are depicted as overlapping only in the vertical direction. It is pertinent to recognize that in actuality a stencil processor may be designed to have overlapping stencils in both the vertical and horizontal directions.
• Because of the vertical overlappingstencils202 within the stencil processor, as observed inFIG. 2a, there exists a wide band of image data within the frame that a single stencil processor can operate over. As will be discussed in more detail below, in an embodiment, the stencil processors process data within their overlapping stencils in a left to right fashion across the image data (and then repeat for the next set of lines, in top to bottom order). Thus, as the stencil processors continue forward with their operation, the number of solid black output pixel blocks will grow right-wise horizontally. As discussed above, aline buffer unit101 is responsible for parsing a line group of input image data from an incoming frame that is sufficient for the stencil processors to operate over for an extended number of upcoming cycles. An exemplary depiction of a line group is illustrated as ashaded region203. In an embodiment, theline buffer unit101 can comprehend different dynamics for sending/receiving a line group to/from a sheet generator. For example, according to one mode, referred to as “full group”, the complete full width lines of image data are passed between a line buffer unit and a sheet generator. According to a second mode, referred to as “virtually tall”, a line group is passed initially with a subset of full width rows. The remaining rows are then passed sequentially in smaller (less than full width) pieces.
• With theline group203 of the input image data having been defined by the line buffer unit and passed to the sheet generator unit, the sheet generator unit further parses the line group into finer sheets that are more precisely fitted to the hardware limitations of the stencil processor. More specifically, as will be described in more detail further below, in an embodiment, each stencil processor consists of a two dimensional shift register array. The two dimensional shift register array essentially shifts image data “beneath” an array of execution lanes where the pattern of the shifting causes each execution lane to operate on data within its own respective stencil (that is, each execution lane processes on its own stencil of information to generate an output for that stencil). In an embodiment, sheets are surface areas of input image data that “fill” or are otherwise loaded into the two dimensional shift register array.
• As will be described in more detail below, in various embodiments, there are actually multiple layers of two dimensional register data that can be shifted on any cycle. For convenience, much of the present description will simply use the term “two-dimensional shift register” and the like to refer to structures that have one or more such layers of two-dimensional register data that can be shifted.
• Thus, as observed inFIG. 2b, the sheet generator parses aninitial sheet204 from theline group203 and provides it to the stencil processor (here, the sheet of data corresponds to the five by five shaded region that is generally identified by reference number204). As observed inFIGS. 2cand 2d, the stencil processor operates on the sheet of input image data by effectively moving the overlappingstencils202 in a left to right fashion over the sheet. As ofFIG. 2d, the number of pixels for which an output value could be calculated (nine in a darkened 3 by 3 array) from the data within the sheet is exhausted (no other pixel positions can have an output value determined from the information within the sheet). For simplicity the border regions of the image have been ignored.
• As observed inFIG. 2ethe sheet generator then provides anext sheet205 for the stencil processor to continue operations on. Note that the initial positions of the stencils as they begin operation on the next sheet is the next progression to the right from the point of exhaustion on the first sheet (as depicted previously inFIG. 2d). With thenew sheet205, the stencils will simply continue moving to the right as the stencil processor operates on the new sheet in the same manner as with the processing of the first sheet.
• Note that there is some overlap between the data of thefirst sheet204 and the data of thesecond sheet205 owing to the border regions of stencils that surround an output pixel location. The overlap could be handled simply by the sheet generator re-transmitting the overlapping data twice. In alternate implementations, to feed a next sheet to the stencil processor, the sheet generator may proceed to only send new data to the stencil processor and the stencil processor reuses the overlapping data from the previous sheet.
b. Stencil Processor Design and Operation
• FIG. 3 shows an embodiment of a stencilprocessor unit architecture300. As observed inFIG. 3, the stencil processor includes adata computation unit301, ascalar processor302 and associated memory303 and an I/O unit304. Thedata computation unit301 includes an array of execution lanes305, a two-dimensionalshift array structure306 and separate respectiverandom access memories307 associated with specific rows or columns of the array.
• The I/O unit304 is responsible for loading “input” sheets of data received from the sheet generator into thedata computation unit301 and storing “output” sheets of data from the stencil processor into the sheet generator. In an embodiment the loading of sheet data into thedata computation unit301 entails parsing a received sheet into rows/columns of image data and loading the rows/columns of image data into the two dimensionalshift register structure306 or respectiverandom access memories307 of the rows/columns of the execution lane array (described in more detail below). If the sheet is initially loaded intomemories307, the individual execution lanes within the execution lane array305 may then load sheet data into the two-dimensionalshift register structure306 from therandom access memories307 when appropriate (e.g., as a load instruction just prior to operation on the sheet's data). Upon completion of the loading of a sheet of data into the register structure306 (whether directly from a sheet generator or from memories307), the execution lanes of the execution lane array305 operate on the data and eventually “write back” finished data as a sheet directly back to the sheet generator, or, into therandom access memories307. If the execution lanes write back to random access memories907, the I/O unit304 fetches the data from therandom access memories307 to form an output sheet which is then forwarded to the sheet generator.
• Thescalar processor302 includes a program controller309 that reads the instructions of the stencil processor's program code from scalar memory303 and issues the instructions to the execution lanes in the execution lane array305. In an embodiment, a single same instruction is broadcast to all execution lanes within the array305 to effect a single instruction multiple data (SIMD)-like behavior from thedata computation unit301. In an embodiment, the instruction format of the instructions read from scalar memory303 and issued to the execution lanes of the execution lane array305 includes a very-long-instruction-word (VLIW) type format that includes more than one opcode per instruction. In a further embodiment, the VLIW format includes both an ALU opcode that directs a mathematical function performed by each execution lane's ALU (which, as described below, in an embodiment may specify more than one traditional ALU operation) and a memory opcode (that directs a memory operation for a specific execution lane or set of execution lanes).
• The term “execution lane” refers to a set of one or more execution units capable of executing an instruction (e.g., logic circuitry that can execute an instruction). An execution lane can, in various embodiments, include more processor-like functionality beyond just execution units, however. For example, besides one or more execution units, an execution lane may also include logic circuitry that decodes a received instruction, or, in the case of more multiple instruction multiple data (MIMD)-like designs, logic circuitry that fetches and decodes an instruction. With respect to MIMD-like approaches, although a centralized program control approach has largely been described herein, a more distributed approach may be implemented in various alternative embodiments (e.g., including program code and a program controller within each execution lane of the array305).
• The combination of an execution lane array305, program controller309 and two dimensionalshift register structure306 provides a widely adaptable/configurable hardware platform for a broad range of programmable functions. For example, application software developers are able to program kernels having a wide range of different functional capability as well as dimension (e.g., stencil size) given that the individual execution lanes are able to perform a wide variety of functions and are able to readily access input image data proximate to any output array location.
• Apart from acting as a data store for image data being operated on by the execution lane array305, therandom access memories307 may also keep one or more look-up tables. In various embodiments one or more scalar look-up tables may also be instantiated within the scalar memory303.
• A scalar look-up involves passing the same data value from the same look-up table from the same index to each of the execution lanes within the execution lane array305. In various embodiments, the VLIW instruction format described above is expanded to also include a scalar opcode that directs a look-up operation performed by the scalar processor into a scalar look-up table. The index that is specified for use with the opcode may be an immediate operand or fetched from some other data storage location. Regardless, in an embodiment, a look up from a scalar look-up table within scalar memory essentially involves broadcasting the same data value to all execution lanes within the execution lane array305 during a the same clock cycle. Additional details concerning use and operation of look-up tables is provided further below.
• FIG. 4 shows an embodiment of adata computation unit401. As observed inFIG. 4, thedata computation unit401 includes an array ofexecution lanes405 that are logically positioned “above” a two-dimensional shiftregister array structure406. As discussed above, in various embodiments, a sheet of image data provided by a sheet generator is loaded into the two-dimensional shift register406. The execution lanes then operate on the sheet data from theregister structure406.
• Theexecution lane array405 andshift register structure406 are fixed in position relative to one another. However, the data within theshift register array406 shifts in a strategic and coordinated fashion to cause each execution lane in the execution lane array to process a different stencil within the data. As such, each execution lane determines the output image value for a different pixel in the output sheet being generated. From the architecture ofFIG. 4 it should be clear that overlapping stencils are not only arranged vertically but also horizontally as theexecution lane array405 includes vertically adjacent execution lanes as well as horizontally adjacent execution lanes.
• Some notable architectural features of thedata computation unit401 include theshift register structure406 having wider dimensions than theexecution lane array405. That is, there is a “halo” ofregisters409 outside theexecution lane array405. Although thehalo409 is shown to exist on two sides of the execution lane array, depending on implementation, the halo may exist on less (one) or more (three or four) sides of theexecution lane array405. Thehalo405 serves to provide “spill-over” space for data that spills outside the bounds of theexecution lane array405 as the data is shifting “beneath” theexecution lanes405. As a simple case, a 5×5 stencil centered on the right edge of theexecution lane array405 will need four halo register locations further to the right when the stencil's leftmost pixels are processed. For ease of drawing,FIG. 4 shows the registers of the right side of the halo as only having horizontal shift connections and registers of the bottom side of the halo as only having vertical shift connections when, in a nominal embodiment, registers on either side (right, bottom) would have both horizontal and vertical connections.
• Additional spill-over room is provided by random access memories407 that are coupled to each row and/or each column in the array, or portions thereof (e.g., a random access memory may be assigned to a “region” of the execution lane array that spans 4 execution lanes row wise and 2 execution lanes column wise. For simplicity the remainder of the application will refer mainly to row and/or column based allocation schemes). Here, if an execution lane's kernel operations require it to process pixel values outside of the two-dimensional shift register array406 (which some image processing routines may require) the plane of image data is able to further spill-over, e.g., from thehalo region409 into random access memory407. For example, consider a 6×6 stencil where the hardware includes a halo region of only four storage elements to the right of an execution lane on the right edge of the execution lane array. In this case, the data would need to be shifted further to the right off the right edge of thehalo409 to fully process the stencil. Data that is shifted outside thehalo region409 would then spill-over to random access memory407. Other applications of the random access memories407 and the stencil processor ofFIG. 3 are provided further below.
• FIGS. 5athrough 5kdemonstrate a working example of the manner in which image data is shifted within the two dimensional shift register array “beneath” the execution lane array as alluded to above. As observed inFIG. 5a, the data contents of the two dimensional shift array are depicted in afirst array507 and the execution lane array is depicted by a frame505. Also, two neighboringexecution lanes510 within the execution lane array are simplistically depicted. In thissimplistic depiction510, each execution lane includes a register R1 that can accept data from the shift register, accept data from an ALU output (e.g., to behave as an accumulator across cycles), or write output data into an output destination.
• Each execution lane also has available, in a local register R2, the contents “beneath” it in the two dimensional shift array. Thus, R1 is a physical register of the execution lane while R2 is a physical register of the two dimensional shift register array. The execution lane includes an ALU that can operate on operands provided by R1 and/or R2. As will be described in more detail further below, in an embodiment the shift register is actually implemented with multiple (a “depth” of) storage/register elements per array location but the shifting activity is limited to one plane of storage elements (e.g., only one plane of storage elements can shift per cycle).FIGS. 5athrough 5kdepict one of these deeper register locations as being used to store the resultant X from the respective execution lanes. For illustrative ease the deeper resultant register is drawn alongside rather than beneath its counterpart register R2.
• FIGS. 5athrough 5kfocus on the calculation of two stencils whose central position is aligned with the pair of execution lane positions511 depicted within the execution lane array505. For ease of illustration, the pair ofexecution lanes510 are drawn as horizontal neighbors when in fact, according to the following example, they are vertical neighbors.
• As observed initially inFIG. 5a, the execution lanes511 are centered on their central stencil locations.FIG. 5bshows the object code executed by both execution lanes511. As observed inFIG. 5bthe program code of both execution lanes511 causes the data within theshift register array507 to shift down one position and shift right one position. This aligns both execution lanes511 to the upper left hand corner of their respective stencils. The program code then causes the data that is located (in R2) in their respective locations to be loaded into R1.
• As observed inFIG. 5cthe program code next causes the pair of execution lanes511 to shift the data within theshift register array507 one unit to the left which causes the value to the right of each execution lane's respective position to be shifted into each execution lane′ position. The value in R1 (previous value) is then added with the new value that has shifted into the execution lane's position (in R2). The resultant is written into R1. As observed inFIG. 5dthe same process as described above forFIG. 5cis repeated which causes the resultant R1 to now include the value A+B+C in the upper execution lane and F+G+H in the lower execution lane. At this point both execution lanes511 have processed the upper row of their respective stencils. Note the spill-over into a halo region on the left side of the execution lane array505 (if one exists on the left hand side) or into random access memory if a halo region does not exist on the left hand side of the execution lane array505.
• As observed inFIG. 5e, the program code next causes the data within the shift register array to shift one unit up which causes both execution lanes511 to be aligned with the right edge of the middle row of their respective stencils. Register R1 of both execution lanes511 currently includes the summation of the stencil's top row and the middle row's rightmost value.FIGS. 5fand 5gdemonstrate continued progress moving leftwise across the middle row of both execution lane's stencils. The accumulative addition continues such that at the end of processing ofFIG. 5gboth execution lanes511 include the summation of the values of the top row and the middle row of their respective stencils.
• FIG. 5hshows another shift to align each execution lane with its corresponding stencil's lowest row.FIGS. 5iand 5jshow continued shifting to complete processing over the course of both execution lanes' stencils.FIG. 5kshows additional shifting to align each execution lane with its correct position in the data array and write the resultant thereto.
• In the example ofFIGS. 5a-5knote that the object code for the shift operations may include an instruction format that identifies the direction and magnitude of the shift expressed in (X, Y) coordinates. For example, the object code for a shift up by one location may be expressed in object code asSHIFT 0, +1. As another example, a shift to the right by one location may expressed in object code as SHIFT +1, 0. In various embodiments shifts of larger magnitude may also be specified in object code (e.g.,SHIFT 0, +2). Here, if the 2D shift register hardware only supports shifts by one location per cycle, the instruction may be interpreted by the machine to require multiple cycle execution, or, the 2D shift register hardware may be designed to support shifts by more than one location per cycle. Embodiments of the later are described in more detail further below.
• FIG. 6ashows another, more detailed depiction of the unit cell for the array execution lane and shift register structure (registers in the halo region do not include a corresponding execution lane). The execution lane and the register space that is associated with each location in the execution lane array are, in an embodiment, implemented by instantiating the circuitry observed inFIG. 6aat each node of the execution lane array. As observed inFIG. 6a, the unit cell includes anexecution lane601 coupled to aregister file602 consisting of four registers R2 through R5. During any cycle, theexecution lane601 may read from or write to any of registers R1 through R5. For instructions requiring two input operands the execution lane may retrieve both of operands from any of R1 through R5.
• In an embodiment, the two dimensional shift register structure is implemented by permitting, during a single cycle, the contents of any of (only) one of registers R2 through R4 to be shifted “out” to one of its neighbor's register files throughoutput multiplexer603, and, having the contents of any of (only) one of registers R2 through R4 replaced with content that is shifted “in” from a corresponding one if its neighbors throughinput multiplexers604 such that shifts between neighbors are in a same direction (e.g., all execution lanes shift left, all execution lanes shift right, etc.). Although it may be common for a same register to have its contents shifted out and replaced with content that is shifted in on a same cycle, themultiplexer arrangement603,604 permits for different shift source and shift target registers within a same register file during a same cycle.
• As depicted inFIG. 6a, note that during a shift sequence an execution lane will shift content out from itsregister file602 to each of its left, right, top, and bottom neighbors. In conjunction with the same shift sequence, the execution lane will also shift content into its register file from a particular one of its left, right, top, and bottom neighbors. Again, the shift out target and shift in source should be consistent with a same shift direction for all execution lanes (e.g., if the shift out is to the right neighbor, the shift in should be from the left neighbor).
• Although in one embodiment the content of only one register is permitted to be shifted per execution lane per cycle, other embodiments may permit the content of more than one register to be shifted in/out. For example, the content of two registers may be shifted out/in during a same cycle if a second instance of themultiplexer circuitry603,604 observed inFIG. 6ais incorporated into the design ofFIG. 6a. Of course, in embodiments where the content of only one register is permitted to be shifted per cycle, shifts from multiple registers may take place between mathematical operations by consuming more clock cycles for shifts between mathematical operations (e.g., the contents of two registers may be shifted between math ops by consuming two shift ops between the math ops).
• If less than all the content of an execution lane's register files are shifted out during a shift sequence note that the content of the non shifted out registers of each execution lane remain in place (do not shift). As such, any non-shifted content that is not replaced with shifted-in content persists local to the execution lane across the shifting cycle. The memory unit (“M”) observed in each execution lane is used to load/store data from/to the random access memory space that is associated with the execution lane's row and/or column within the execution lane array. Here, the M unit acts as a standard M unit in that it is often used to load/store data that cannot be loaded/stored from/to the execution lane's own register space. In various embodiments, the primary operation of the M unit is to write data from a local register into memory, and, read data from memory and write it into a local register.
• With respect to the ISA opcodes supported by the ALU unit of thehardware execution lane601, in various embodiments, the mathematical opcodes supported by the hardware ALU are integrally tied with (e.g., substantially the same as) the mathematical opcodes supported by a virtual execution lane (e.g., ADD, SUB, MOV, MUL, MAD, ABS, DIV, SHL, SHR, MIN/MAX, SEL, AND, OR, XOR, NOT). As described just above, memory access instructions can be executed by theexecution lane601 to fetch/store data from/to their associated random access memory. Additionally thehardware execution lane601 supports shift operation instructions (right, left, up, down) to shift data within the two-dimensional shift register structure. As described above, program control instructions are largely executed by the scalar processor of the stencil processor.
c. Statistics Operations Performed on Stencil Processor
• FIG. 7 shows an exemplary depiction of an image frame600. As observed inFIG. 7, the image700 can be viewed as being broken down into an array of tiles. The particular exemplary image ofFIG. 7 consists of an array of 16 tiles by 16 tiles. As observed inFIG. 8 each tile can be further broken down into an array of sheets. The exemplary depiction ofFIG. 8 shows, e.g., the first twotiles801,802 along the top row of tiles of the image700 ofFIG. 7. AsFIG. 8 shows, each tile can be broken down into an array of sheets. In the particular example ofFIG. 8, each tile is composed of a 16×16 array of sheets. Each sheet may be composed, for example, of a 16×16 array of pixel values.
• As described at length above, in various embodiments, a stencil processor is designed to process a sheet as its unit of data. Here, a sheet of image data is loaded into the stencil processor's two dimensional register space and the processor performs certain image processing tasks on the sheet of image data. As observed inFIG. 8, in various embodiments, the processing of an entire image includes, e.g., processing a row of sheets across a first tile801 (e.g., in a left to right fashion across the tile) and then continuing to process sheets in a same direction along the same row within a secondneighboring tile802.
• The process continues until all sheets along the particular row are processed across all tiles of the image. Then a next (e.g., lower) row of sheets is processed and the process repeats. For instance after the sixteenth sheet of the sixteenth tile along the first row of sheets of the image is processed, the first sheet of the first tile along the second row of sheets is next processed.
• The processing of statistics for an image often entails calculating a summation of pixel values within the image. For example, a common statistics operation is to calculate an average pixel value. As such, pixel values are summed across an image and then normalized by the number of pixels that were summed over. In various embodiments, image statistics are organized at tile granularity. As such, e.g., an average pixel value is maintained for each tile in an image. The tiles are then subsequently processed as needed to calculate other statistics for the image.
• FIG. 9 shows one approach for summing pixel values within a tile.FIG. 9 depicts the sheet arrangement for the upper left hand corner of a tile. As observed inFIG. 9, same positioned pixel values within a sheet array are summed over multiple sheets. For example, the upper leftmost pixel value of Sheet_0 is added with the upper leftmost pixel value of Sheet_1. The resulting summation is then added to the upper leftmost pixel value of Sheet_2. The process continues until all sheets in the tile have been accounted for.
• Recalling that the stencil processor processes data in units of sheets, the summation itself is performed on a sheet by sheet basis. As described above, in an embodiment, a stencil processor includes an execution lane and associated register space for each array location within a sheet. Initially, Sheet_0 is loaded into the stencil processor's two dimensional shift register structure and processed. A pixel value associated with each array location in the sheet is kept in the local register space of that array location within the two dimensional shift register. Then Sheet_1 is loaded into the stencil processor's two dimensional shift register structure and processed.
• A pixel value associated with each array location in Sheet_1 is added to the pixel value for the same array location in Sheet_0. The resultant for each array location is kept in the local register space of that location in the two-dimensional shift register structure and the process continues. Thus, with each new sheet that the stencil processor processes, a summation across all sheets processed so far can be accumulated for all array locations within the sheet. In an alternate approach, the above process is performed except that the accumulated summation data is kept in a stencil processor's internal memory space rather than register space (e.g., referring toFIG. 3, theRAM307 that is coupled to the memory execution units of the executional array305 and/or the scalar memory303 that is associated with the scalar processor303). In another alternate embodiment, the accumulated summation data is kept in memory that is external to the stencil processor such as memory that is coupled to aline buffer unit101.
• ComparingFIGS. 8 and 9, note that Sheet_16 oftile 0 will not be processed until after Sheet_15 oftile 15 has been processed. That is, a same row of sheets are processed across tiles. Thus, if statistics are tracked on a tile by tile basis, statistics context needs to switch each time the processing progresses across tile boundaries.FIG. 10 graphically depicts the context switching process. Here, a data structure that keeps statistics for a particular tile is assumed to be created and resident in the register space or internal memory space of a stencil processor when the stencil processor is processing that particular tile.
• For instance, whilesheets 0 through 15 are being processed for Tile_0, a statistics data structure for Tile_0 is kept in the internal memory and/or two-dimensional register space of the stencil processor. After Sheet_15 is processed for Tile_0, attime1001, the statistics data structure for Tile_0 is switched out of the stencil processor's two dimensional register space or internal memory space and the statistics data structure for Tile_1 is switched into the stencil processor's register space or internal memory space. Likewise, attime1002, in between the processing of Sheet_15 of Tile_1 and Sheet_0 of Tile_2, the statistics data structure for Tile_1 is switched out of the stencil processor's register space or internal memory and the statistics data structure for Tile_2 switched into the stencil processor's register or memory space.
• After Sheet_15 of Tile_15 is processed, attime1016, the statistics data structure of Tile_15 is switched out of the register or internal memory space of the stencil processor and the statistics data structure for Tile_0 is switched back into the register or internal memory space of the stencil processor as the process begins processing of the second row of sheets within the image. Likewise, attime1001, the statistics data structure for Tile_0 is switched out of the stencil processor's register space or internal memory space and the statistics data structure for Tile_1 is switched into the stencil processor's register space or internal memory space. Processing then continues in this fashion until all sheets oftiles 0 through 15 have been processed at which point the statistics data structures for each oftiles 0 through 15 are complete. The process then continues for the second row of sheets within the image.
• FIG. 11 shows an exemplary depiction of the aforementioned statistics data structure that may be kept for a particular tile. Here, as discussed above, for example, the statistics data structure may keep a summation of all pixel values at a same pixel location within a sheet of the tile. A “complete” data structure is created for the tile when all sheets within the tile have been summed over. That is, a complete data structure includes a summation at the same sheet pixel location across all sheets within the tile. For illustrative ease,FIG. 11 only graphically depicts the summation at the four corner pixel locations of the sheet.
• An additional statistical process is to reduce the entire accumulated array ofFIG. 11 into a scalar value that, e.g., sums over all the pixel locations of the array ofFIG. 11, which, in turn, corresponds to a summation of all of the pixel values within the entire tile. That is, withFIG. 11 being array of the summation of all pixel values within a tile at a same sheet pixel location, the summation across all elements in the array ofFIG. 11 corresponds to the summation of all pixel values within a tile on a sheet pixel location by sheet pixel location basis.FIGS. 12athrough 12eshow the reduction of the entire array into a scalar. As will be evident from the discussion ofFIGS. 12athrough 12e, the reduction process makes particular use of the two dimensional shift register array that is a component of the stencil processor.
• Here, the statistics array ofFIG. 11 is assumed to be initially loaded into the two dimensional shift array of the stencil processor.FIG. 12ashows a first shift action of the shift register in which every other column of data is shifted into its left neighbor. Each shift depicted inFIG. 12aincludes a pixel value being shifted and a pixel value in the location that the shifted pixel value is shifted into. These pair of pixel values are then added in the execution lane that is shifted into and kept in the register space of the same execution lane.
• As such, after the completion of the shift ofFIG. 12aand the following summation every other column of the array includes a summation of the neighboring pair of pixel values in the register space of each pixel location of the column. That is, e.g.,column1201 includes a summation of neighboring pixel values in every pixel location ofcolumn1201,column1202 includes a summation of neighboring pixel values in every pixel location ofcolumn1202, etc. Note that all the shifts observed inFIG. 12amay occur in a single cycle, and, the summation may occur in an immediately following cycle. As such, the reduction can reduce the array by 50% in only two cycles (cycles C1 and C2).
• As observed inFIG. 12b, nearest neighbor columns having a summation from the operation ofFIG. 12aare shifted and added. That is, after the operation ofFIG. 12a, onlycolumns1200,1202,1204,1206,1208,1210,1212, and1214 have summation content from the summation operation ofFIG. 12a(the other columns may also have summation content but their content is not utilized, in an embodiment, the shift and summations operations that determine the unused content is not performed to save power). As such, pixel values ofcolumn1202 are shifted to the left by two array locations intocolumn1200. Corresponding source to destination pairs also exist forcolumns1206 to1204,1210 to1208, and1214 to1212. As with the operation ofFIG. 12, the pixel value pairs are added by the execution lane logic at the destination. As such, after the operation ofFIG. 12b, summation values now exist incolumns1200,1204,1208, and1212. Note that in an embodiment, the two dimensional shift register is able to perform “two location hop” shifts in a single cycle. Thus, all of the shifts observed nFIG. 12bcan be performed in a single cycle (C3) and the summation operation can be performed in an immediately following cycle (C4). As such, after four cycles, the reduction operation can reduce the array ofFIG. 11 by 75%.
• Here, however, it is pertinent to point out that in various embodiments each execution lane has multiple registers. That is, referring briefly back toFIG. 6, each execution lane has multiple registers R1 through R5 for shifting data in/out and/or storing a resultant from an ALU operation. Additionally, many image processing algorithms have multiple “channels” of data. For example, an image composed of red (R), green (G), and blue (B) pixels may process a first channel composed of sheets/tiles/frames of R data separately from a second channel composed of sheets/tiles/frames of G data all of which are processed separately from a third channel of sheets/tiles/frames of B data.
• Although the aforementioned example ofFIGS. 12aand 12bindicated two cycles must be consumed to perform a shift and a summation, it is pertinent to recognize that this limitation only applies to operations within a same channel because of the summation's dependence on a prior shift. Importantly, in an embodiment, the execution lane array receives a VLIW type instruction word that can specify both a shift operation and an ALU operation within a single cycle. With this instruction format and multiple registers per array location, two different channels can be simultaneously reduced according to the core process ofFIGS. 12aand 12b. That is, in a software pipelined loop like fashion, e.g., in a first cycle R is shifted and G is summed, and, in a second following cycle the just shifted R is summed and the resultant from the just summed G is shifted. In this fashion, the overall processor can perform a shift and summation reduction in, e.g., every cycle across two channels.
• As observed inFIG. 12c, the reduction process continues but with (again) longer shifts. Here, pixel values ofcolumn1204 are shifted intocolumn1200 and added incolumn1200, and, pixel values ofcolumn1212 are shifted intocolumn1208 and added incolumn1208. As such, after the operation ofFIG. 12c(which again may only consume the next two cycles (C5 and C6) assuming the two dimensional shift register supports, e.g., shifts of 4 unit locations), summation values only exist incolumns1200 and1208.FIG. 12dshows complete reduction along the horizontal axis as the values incolumn1208 are shifted intocolumn1200 and added withincolumn1200. Again, assuming the two dimensional shift register can perform the shift observed inFIG. 12din a single cycle, the reduction ofFIG. 12dis reached by the eight cycle C8 (in an embodiment, the shift register does not generally support shifts of the magnitude observed inFIG. 12dbut does support, in a single cycle, a shift fromcolumn1208 to column1200).
• FIG. 12eshows the subsequent reductions that are performed alongcolumn1200 which ultimately reduce the entire array to a scalar summation value as of cycle C16. The reductions ofFIG. 12eare essentially the same reductions ofFIGS. 12athrough 12dbut in the vertical direction and only along one column. In an embodiment, as with the horizontal reductions ofFIGS. 12athrough 12d, each reduction sequence consumes one cycle for the shift register to perform its corresponding shift and another cycle is consumed adding a local value with a value that was just shifted into the same location.
• The above discussion of the reduction process ofFIGS. 12athrough 12ewas directed to an embodiment where, as shown inFIG. 11, a separate summation value was maintained for each pixel location in a sheet for all sheets within a tile. As discussed with respect toFIG. 10, however, before a “complete” statistics structure for a particular tile can be realized, all rows in the tile need to be processed. Because scanning across an image row will cross multiple tiles, data structures for the tiles are context switched in and out with each crossing of a tile boundary.
• Each per tile statistics data structure therefore includes a summation value or “counter” for each pixel location within a sheet for the sheets that have so far been processed. Thus, the dimensions of the statistics data structure for a tile correspond to the dimensions of a sheet within the tile (e.g., 16×16 summation values). If multiple tiles exist within an image (as suggested byFIGS. 9 and 10), the amount of statistics data that is kept for an image can be large (i.e., (the number of tiles in the image)×(the number of counters per tile)). Here, reduction of a tile into a scalar as soon as a complete data structure exists for a tile can help ease the statistics tracking burden.
• Alternatively, referring back toFIG. 10, the statistics array can be reduced to a scalar each time it is to be context switched out. For instance, the statistics array for tile_0 may be reduced to a scalar attime1001 just before it is stored and again attime1017 just before it is stored. Here, performing the reduction ofFIGS. 12athrough 12ebefore each context switch out may consume more processing (multiple scalar reductions are performed per tile rather than one scalar reduction per tile), but the data footprint size of the statistics being switched out is much smaller which makes it more likely the switched out statistics data can be kept internally within the stencil processor memory rather than being directed across the entire image processor to line buffer memory.
• Other statistical processes may be less burdensome. For example, according to one statistics keeping mode referred to as “global”, an entire frame/image is deemed to consume only a single tile. Therefore only one statistical data structure exists and context switching does not arise (i.e., no tile boundaries are crossed in the raster scan processing). The data structure is “complete” when all sheets within the image have been processed at which point a scalar for the entire image can be created by performing the reduction ofFIGS. 12athough12ediscussed just above.
• In another statistics tracking mode, depicted inFIG. 12f, statistics are kept for one or more “windows”1251,1252,1253 within animage1250 that may or may not overlap in various combinations. Sheets may be processed in a particular pattern (e.g., from left to right across the image before dropping down to process a next row). While processing within a particular window statistics are accumulated for each pixel location with the sheet dimensions consistent with the discussion ofFIG. 9 above. Upon leaving a window the accumulation activity stops. The statistical accumulation restarts upon the raster scan eventually crossing back into the window.
• For example, referring toFIG. 12f, statistics accumulations will stop forwindow1251 as the raster scan followspath1254 and crosses out ofwindow1251 but will recommence when the raster scan reachespath1255 and crosses back intowindow1251. Note that this behavior is followed for all windows which can result in statistics being concurrently accumulated for regions of overlapping windows. For example, referring toraster scan path1256, statistics tracking will restart forwindow1251 when point A is reached. Some time later statistics tracking will restart forwindow1252 when point B is reached. A further time later statistics tracking will cease forwindow1251 when point C is reached and then forwindow1252 when point D is reached. Once accumulations have completed for a window after the raster scan has covered the entire surface area of a window, the data that was accumulated for the window (again, one accumulated value per pixel location within a sheet) may be reduced according to the process discussed above with respect toFIGS. 12athrough12e.
• Recall from the discussion ofFIG. 4 that random access memories407 may be coupled to theexecution lane array405 and twodimensional shift register406 to give the stencil processor a local system memory function. Here, multiple execution lanes within theexecution lane array405 may share a same random access memory unit. For example, referring toFIG. 4, a first subset of execution lanes may be coupled to random access memory407_1, a second subset of execution lanes may be coupled to random access memory407_2, etc.FIG. 4 shows one approach where execution lanes that reside along a same row are coupled to a same random access memory.
• FIG. 13 shows another approach where two dimensional blocks of execution lanes are coupled to a same random access memory. In the particular example ofFIG. 13, blocks of four execution lanes are coupled to a same random access memory. Here, for instance, the upper leftmost block of four execution lanes are coupled to random access memory1307_1, the upper rightmost block of four execution lanes are coupled to random access memory1307_2, etc. Other separate blocks of four execution lanes are presumed to be coupled to their own respective random access memory but are not depicted for illustrative ease.
• Another common set of statistics that are kept for an image are a histogram. A histogram essentially divides a particular parameter into a number of bins. For example, an eight bit intensity value can be broken down into 256 separate bins (one bin for each unique eight bit value). The number of times a value is observed in the data of an image is kept track of each bin. As such, e.g., the number of times each unique eight bit value is observed is individually recorded with a count value for each bin. The spread of respective count values across the set of bins corresponds to a histogram.FIG. 14 shows an exemplary depiction of a histogram showing a separate count (along the vertical axis) for each of multiple bins (along the horizontal axis).
• FIGS. 15aand 15bpertain to a technique for extending the capabilities of the stencil processor ofFIGS. 4 and 13 to generate high resolution histograms. That is, a histogram with a large number of bins.FIG. 15ashows an embodiment of a nominal design point of the stencil processor ofFIGS. 4 and 13 in which each execution lane within a group of execution lanes that share a same random access memory is provided its own private memory space within the random access memory. Here, referring toFIG. 15a, the set of four execution lanes1501 are each coupled to random access memory1507. However, each execution lane is nominally design to only access its own private space within random access memory1507. As drawn inFIG. 15, for example, execution lane1501_1 is designed to only access addressable region1502_1 within random access memory. Other partitions of random access memory1507 are observed being allocated for the other execution lanes within the group1501.
• A problem with the per execution lane memory partitioning ofFIG. 15ais that the partitions effectively limit the number of histogram bins that the execution lanes can bin observed data to. That is, because each partition corresponds to a limited amount of memory space, each execution lane can only bin data items across a total number of bins that can be established within a single partition. This may have the undesirable effect of limiting the resolution of the histograms that the execution lanes can bin to (note also that in various embodiments only one of the execution lanes in a group may access the memory in a given cycle).
• As such,FIG. 15bshows a better approach in which, at least during a histogram mode, the partitions within a random access memory are effectively merged so that each execution lane that shares the random memory can freely access the total amount of merged space. Merging the partitions and permitting each of the execution lanes that share the random access memory to freely access the total amount of merged space effectively provides for a wider data field that accommodate more bins. With more bins being accommodated for, higher resolution histograms can result. As such, for example, if a single partition of the approach ofFIG. 15acan only hold count values for 256 bins, the merged space ofFIG. 15bcan hold count values for 1024 bins. Thus, the execution lanes of the approach ofFIG. 15acan only generate histograms having 256 bins while the same execution lanes in the approach ofFIG. 15bcan generate histograms having 1024 bins. Other embodiments can have other numbers of bins.
• FIGS. 16 and 17athrough17cpertain to an approach in which a histogram can be determined across an entire image by first having the execution lanes bin observed data to their local memories and then using the two-dimensional shift register structure and ALU capabilities of the execution lanes to tabulate respective counts over the bins.
• FIG. 16 shows a high level depiction in which neighboring execution lanes are grouped into neighboring blocks of four where each group is given access to a respective random access memory, where, the space of each random access memory is merged so that each execution lane within a same group can access the same merged memory space as described just above with respect to the description ofFIGS. 15aand 15b. For simplicity only two blocks ofgroups execution lanes1601,1602 are depicted along with their corresponding respective merged memory space1607_1,1607_2.
• To generate a histogram, a series of sheets for an image are loaded into the two-dimensional shift register. The bins of the histogram are allocated in the merged memory space for each execution lane group and the individual execution lanes increment a respective count value maintained for each bin within the merged memory space for the particular value that each execution lane observes. For example, ifexecution lanes 1, 2, 3, and 4 within a same group respectively observe values corresponding tobins 4, 9, 12, and 21, then, execution lane will update the count value for bin 4,execution lane 2 will update the count value forbin 9,execution lane 3 will update the count value for bin 12, and execution lane 4 will update the count value forbin 21. In various embodiments, because only one execution lane can access the memory that a group of execution lanes share, the execution lanes will perform their aforementioned updates in consecutive cycles.
• After a number of sheets have been processed for the image, the merged memory space of each of the random access memories essentially contains a histogram for its constituent execution lanes. For example, referring toFIG. 16, the merged memory space within random access memory1607_1 contains the histogram data for execution lanes withingroup1601 and the merged memory space within random access memory1607_2 contains the histogram data for the execution lanes withingroup1602.
• A next operation then is to merge the content of each random access memory so that a histogram for all execution lanes can be realized.FIG. 17apertains to a simplistic example, in which the histogram only contains 32 bins (bins 0 to 31). Here, for each group of execution lanes and their corresponding merged memory space, the count values forbins 0 to 7 are loaded into the register space of the upper left execution lane, the count values forbins 8 to 15 are loaded into the register space of the upper right execution lane, the count values forbins 16 to 23 are loaded into the register space for the lower right execution lane, and the count values forbins 24 to 31 are loaded into the register space for the lower left execution lane.FIG. 17ashows the content of the respective register space of the execution lanes forgroups1701 and1702. Again, only two groups are shown for illustrative ease and in various embodiments the content pattern forgroups1701 are 1702 are repeated across all groups within the execution lane array. In embodiments, the loaded register content takes the form of a vector data structure that maintains respective count value for different bit positions in the data structure where the different bit positions correspond to the different bins being maintained by the data structure.
• FIG. 17bshows a first reduction sequence in which the histogram content of execution lanes from a first group is shifted into the register space of same relative positioned execution lanes from a second group. That is, e.g., the upper left execution lane of thesecond group1702 has its register content shifted into the register space of the upper left execution lane of thefirst group1701 to consolidate the histogram data forbins 0 through 7 collected across bothgroups1701,1702 into the register space of the upper left execution lane ofgroup1701. A corresponding operation is performed for the other three execution lane positions. Note that the shift can occur in a single cycle assuming the two dimensional shift register can shift data structures of appropriate size across two array lane locations in a single cycle. If the bin content for 8 bins consumes more data than can be shifted in a single cycle, then multiple cycles may be used.
• Regardless, after the bin content for thesecond group1702 has been shifted into the register space of thefirst group1701, the execution lanes perform a vector add of their respective pairs of bin data to accumulate total count values for their respective bins. That is, e.g., after the vector add operation, the upper execution lane ofgroup1701 will contain a data structure that has the total count for each ofbins 0 through 7 that was tabulated by bothgroups1701 and1702. A corresponding accumulation of count values will also exist in the register space of the other three execution lanes ofgroup1701. Again, like the shift operation, the vector summation can also be performed in a single cycle in various embodiments. Again, althoughFIGS. 17aand 17bonly show two groups, in various embodiments the operations ofFIGS. 17aand17 are simultaneously performed for all pairs of groups in the execution lane array.
• FIG. 17cshows a next iteration in which the accumulated counts that were just tabulated (e.g., as ofFIG. 17b) are again consolidated according to a similar process (except that the shift length is longer). For instance, referring briefly back toFIG. 17b, one can envision another pair ofgroups1703,1704 of four execution lanes each that physically reside directly beneathgroup1702 in the array.FIG. 17cshows both of thesegroups1703,1704. After the operation ofFIG. 17bis completed, the content ofgroups1703 and1704 will be accumulated ingroup1703. As such, the operation ofFIG. 17cshifts the accumulated content ofgroup1703 being shifted into the register space ofgroup1701. After the vector add operation ofFIG. 17c,group1701 will contain accumulated counts for all bins as originally tabulated by each ofgroups1701 through1704. In the simplistic 8×8 execution lane array ofFIG. 17c, all values will have been accumulated into the top row of groups after the vector add operations ofFIG. 17c.
• From here, a horizontal reduction amongst the top row of groups can commence similar to the vertical reductions ofFIG. 12e(with each different execution lane location within a group continuing to maintain its own set of bins). Eventually the process will consolidate to a single group where each array location in the group has the total count across the image for its corresponding bins.
• In various embodiments, histograms are keep track of at tile granularity. As such, context switching of histogram data similar to the context switching discussed above with respect toFIG. 10 can be performed during accumulation of the histogram. Here, for instance, the aforementioned merged data structures within the processor's random access memories are configured to store histogram data only for a specific tile. When the raster scanned processing of image data causes tile boundaries to be crossed over, the histogram data structure for the tile that processing just departed from is switched out (e.g., by being written out of the stencil processor memories into, e.g., a line buffer) and histogram data structures for the tile that processing just entered is switched in (e.g., be being written into the stencil processing memories). After the raster scanning fully processes a complete tile, the reductions ofFIGS. 17athrough 17cand thereafter can be performed to determine a complete histogram for the entire tile.
• Note that, conceivably, lower resolution histograms may not require merged storage space in the random access memories of the stencil processor. Here, for instance, the dedicated per execution lane memory space ofFIG. 15amay be sufficient to hold all the bins for a lower resolution histogram. As such, merging of the per execution lane memory spaces may not be necessary (and therefore may only be entertained for high resolution histograms). Further still, some histograms may be of sufficiently low resolution that the bin count values may be kept in the respective register space of the execution lanes (here, the register file may be dynamically indexed with a pixel value such that each execution lane could access a different register). In this case use of the random access memories is not necessary. For lower resolution histograms, such as those that do not require merging of per execution lane dedicated memory or register space, bin totals may be accumulated using the processes ofFIGS. 12athrough 12erather thanFIGS. 17athrough17c.
• Similarly, with respect to context switching, it is conceivable that only high resolution histogram data structures will need to be switched out of the stencil processor memory or register space (e.g., into a line buffer). Here, lower resolution histograms may be able to keep “switched out” histogram data context in the local memories and/or register space of the stencil processor.
• FIGS. 16 and 17a-17cdescribed a mechanism for a group of execution lanes to collectively share the memory resources that they are coupled to in order to allow for histograms with a larger number of bins. By contrast,FIGS. 18aand 18bdescribe a mechanism by which the bin count of a histogram can be extended across execution lanes of different groups.FIG. 18ashows, e.g., the upper left hand corner of the execution lane array. Here a first 4×2group1810 of execution lanes and a second 4×2 group ofexecution lanes1811 are observed. Bothgroups1810,1811 may share memory resources to expand bin size as described above with respect toFIGS. 16 and 17a-c.
• Here, if each ofgroups1810,1811 are able to individually support 1024 different histogram bins, both ofgroups1810,1811 could also be logically combined to support 2048 bins. Thus, histogram size in terms of number of bins can be expanded by combining groups of execution lanes. According to one approach referred to as “spatial”, the different execution lane groups within a same logical combination are responsible for binning into different bin ranges. For example, the execution lanes withingroup1810 are responsible for binning intobins 0 through 1023 and the execution lanes withingroup1811 are responsible for binning into bins 1024 through 2047. Thus, in order to fully process input pixel data so that it is guaranteed to be placed into the particular correct bin that it should be placed into, data is shifted as observed inFIG. 18b. That is, e.g., data that is to be binned is first loaded into the execution lanes of thefirst group1810 and if any the data items fall within bins 0-1023 they are binned by their respective execution lanes ingroup1810. After the initial binning processing, the data that was loaded into the first group is shifted into thesecond group1811 as observed inFIG. 18b. Any remaining items of data that were not binned by the first group because they fell with bin ranges 1024-2047 will be binned by respective execution lanes withingroup1811.
• Thus, in this example, data is initially loaded into every lane of every group in the execution lane array and the entire array performs processing on the data with shifts in between binning processes. For instance, data is initially loaded into both ofgroups1810 and1811. After the data is loaded,group1810 bins those data values that were loaded intogroup1810 that fall into bins 0-1023 andgroup1811 bins those data values that were loaded intogroup1811 that fall into bins 1024-2047. After this initially binning sequence, the data that was originally loaded intogroup1810 is shifted intogroup1811 and the data that was originally loaded intogroup1811 is shifted into another group (such as the group directly beneath group1811) that bins into bins 0-1023). For simplicityFIG. 18bdoes not show the shifting of the data that was originally loaded intogroup1811. Note that bin sizes can even be further extended by combining more than two groups. For instance, the 4×2 group of execution lanes directly beneathgroup1811 can be used to bin into bin range 2048-3072.
• In another alternative binning expansion approach referred to as temporal, rather than combine physical groups of execution lanes and use the two-dimensional shift register to shift between combined groups, instead, groups of execution lanes are not combined. Rather, the binning ranges are context switched in and out of the execution lanes. In this case, e.g., data is initially loaded into all execution lane groups. During a first iteration each group bins according to a first bin range (e.g., bins 0-1023). After binning into the first range is complete, the data is not shifted within the register array, but binning context information that determines which binning ranges apply are switched (e.g., context for binning into ranges 0-1023 are switched out of the execution lane groups and context for binning into ranges 1024-2047 are switched into the execution lane groups). The execution lanes then re-execute a binning process for the new, higher range. The process continues until after the final binning range context is switched in an binned to.
• In various embodiments the image processor has configuration register space to establish any of the operating modes described above (e.g., context switching out of the stencil processor, merging of execution lane register space, global statistics tracking, per tile statistics tracking, window statistics tracking, etc.). The statistics operations described above can be codified in program code (e.g., object code executed by a stencil processor) and stored on a machine readable storage medium.
• Although various embodiments above have applied processes to rows in certain situations and columns in certain other situations, it should be readily apparent that processes performed on rows can instead be performed on columns and vice-versa.
• FIG. 19 shows an integrated camera1901, ISP/IPU, andprocessor1902. The camera1901 may include alens assembly1907 andimage sensor1906. Raw image data from the camera1901 may be passed from the camera1901 to an ISP subsystem that includes an image signal processor (ISP) and/or image processing unit (IPU)1904 and acamera statistics unit1905. In various embodiments, thecamera statistics unit1905 is a hardware component that calculates statistics from the raw image data that is generated by the camera1901 with dedicated logic circuitry. In still yet other embodiments, thecamera statistics unit1905 is a software component and camera statistics are executed onIPU hardware1904 which is capable of executing program code (here, image signal processing (ISP) hardware is traditionally not capable of executing program code). Here, such an IPU may configured to perform traditional ISP related tasks in software and therefore no traditional ISP is actually resident in the platform (i.e., all stats and image signal processing functions are performed in software by an IPU). In still other approaches, the ISP/IPU sub-system may be implemented as a combination of dedicated hardware statistics circuits, software executed on an IPU, and/or ISP hardware circuitry. Regardless, in implementations whereIPU hardware1904 at least partially executes statistics software (or where ISP/IPU hardware1904 acts as a co-processor or accelerator that receives more coarse grained statistics related commands from ahardware statistics engine1905 rather than execute fine grained program code), the ISP/IPU1904 may contain a two-dimensional shift register array structure and perform reductions and/or generate histograms consistently with any of the discussions and teachings provided at length above. In such embodiments, the ISP/IPU1904 may be designed to include any/all of the IPU features described at length above in preceding sections.
• Various hardware circuitry components of the ISP/IPU sub-system, may be integrated on a large system-on-chip that also includes theprocessor1902 and other electrical functions (e.g., a graphics processing unit (GPU)). Although in many embodiments theprocessor1902 may be a processing core of a computing system (such as a handheld device). In other embodiments, theprocessor1902 may be more tightly coupled to, or even be a component of, the camera1901.
• Thecamera statistics engine1905 calculates statistics from the pixel data generated by the camera1901. The statistics that are generated by thestatistics engine1905 may then be employed for higher order processes. For example, the statistics are forwarded to theprocessor1902 which performs higher level camera control functions using the statistics, such higher level functions including auto-focusing, auto-exposing, and/or auto white balancing. The camera control function may be implemented, e.g., assoftware1903 that executes on the processor1902 (such as3A camera control software associated with an Android operating system). Other types of statistics data for uses other than camera control may also be forwarded to theprocessor1902 or elsewhere for other processes (e.g., statistics information for computer vision software or hardware).
• FIG. 20 shows a methodology described by the preceding discussions. As observed inFIG. 20, the method includes loading an array of content into a two-dimensional shift register2001, where, the two-dimensional shift register is coupled to an execution lane array. The method further includes repeatedly performing afirst sequence2002 that comprises: shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides, and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content. The method further includes repeatedly performing asecond sequence2003 that comprises shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column. The resultant row or column have values determined at least in part from the mathematical operations of the first sequence. The second sequence further includes performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.
d. Implementation Embodiments
• It is pertinent to point out that the various image processor architecture features described above are not necessarily limited to image processing in the traditional sense and therefore may be applied to other applications that may (or may not) cause the image processor to be re-characterized. For example, if any of the various image processor architecture features described above were to be used in the creation and/or generation and/or rendering of animation as opposed to the processing of actual camera images, the image processor may be characterized as a graphics processing unit. Additionally, the image processor architectural features described above may be applied to other technical applications such as video processing, vision processing, image recognition and/or machine learning. Applied in this manner, the image processor may be integrated with (e.g., as a co-processor to) a more general purpose processor (e.g., that is or is part of a CPU of computing system), or, may be a stand alone processor within a computing system.
• The hardware design embodiments discussed above may be embodied within a semiconductor chip and/or as a description of a circuit design for eventual targeting toward a semiconductor manufacturing process. In the case of the later, such circuit descriptions may take of the form of a (e.g., VHDL or Verilog) register transfer level (RTL) circuit description, a gate level circuit description, a transistor level circuit description or mask description or various combinations thereof. Circuit descriptions are typically embodied on a computer readable storage medium (such as a CD-ROM or other type of storage technology).
• From the preceding sections is pertinent to recognize that an image processor as described above may be embodied in hardware on a computer system (e.g., as part of a handheld device's System on Chip (SOC) that processes data from the handheld device's camera). In cases where the image processor is embodied as a hardware circuit, note that the image data that is processed by the image processor may be received directly from a camera. Here, the image processor may be part of a discrete camera, or, part of a computing system having an integrated camera. In the case of the later the image data may be received directly from the camera or from the computing system's system memory (e.g., the camera sends its image data to system memory rather than the image processor). Note also that many of the features described in the preceding sections may be applicable to a graphics processor unit (which renders animation).
• FIG. 21 provides an exemplary depiction of a computing system. Many of the components of the computing system described below are applicable to a computing system having an integrated camera and associated image processor (e.g., a handheld device such as a smartphone or tablet computer). Those of ordinary skill will be able to easily delineate between the two.
• As observed inFIG. 21, the basic computing system may include a central processing unit2101 (which may include, e.g., a plurality of general purpose processing cores2115_1 through2015_N and amain memory controller2117 disposed on a multi-core processor or applications processor),system memory2102, a display2103 (e.g., touchscreen, flat-panel), a local wired point-to-point link (e.g., USB)interface2104, various network I/O functions2105 (such as an Ethernet interface and/or cellular modem subsystem), a wireless local area network (e.g., WiFi)interface2106, a wireless point-to-point link (e.g., Bluetooth)interface2107 and a GlobalPositioning System interface2108, various sensors2109_1 through2109_N, one ormore cameras2110, abattery2111, a powermanagement control unit2112, a speaker andmicrophone2113 and an audio coder/decoder2114.
• An applications processor ormulti-core processor2150 may include one or more generalpurpose processing cores2115 within itsCPU2101, one or moregraphical processing units2116, a memory management function2117 (e.g., a memory controller), an I/O control function2118 and animage processing unit2119. The generalpurpose processing cores2115 typically execute the operating system and application software of the computing system. Thegraphics processing units2116 typically execute graphics intensive functions to, e.g., generate graphics information that is presented on thedisplay2103. Thememory control function2117 interfaces with thesystem memory2102 to write/read data to/fromsystem memory2102. The powermanagement control unit2112 generally controls the power consumption of thesystem2100.
• Theimage processing unit2119 may be implemented according to any of the image processing unit embodiments described at length above in the preceding sections. Alternatively or in combination, theIPU2119 may be coupled to either or both of theGPU2116 andCPU2101 as a co-processor thereof. Additionally, in various embodiments, theGPU2116 may be implemented with any of the image processor features described at length above.
• Each of thetouchscreen display2103, the communication interfaces2104-2107, theGPS interface2108, thesensors2109, thecamera2110, and the speaker/microphone codec2113,2114 all can be viewed as various forms of I/O (input and/or output) relative to the overall computing system including, where appropriate, an integrated peripheral device as well (e.g., the one or more cameras2110). Depending on implementation, various ones of these I/O components may be integrated on the applications processor/multi-core processor2150 or may be located off the die or outside the package of the applications processor/multi-core processor2150.
• In an embodiment one ormore cameras2110 includes a depth camera capable of measuring depth between the camera and an object in its field of view. Application software, operating system software, device driver software and/or firmware executing on a general purpose CPU core (or other functional block having an instruction execution pipeline to execute program code) of an applications processor or other processor may perform any of the functions described above.
• Embodiments of the invention may include various processes as set forth above. The processes may be embodied in machine-executable instructions. The instructions can be used to cause a general-purpose or special-purpose processor to perform certain processes. Alternatively, these processes may be performed by specific hardware components that contain hardwired logic for performing the processes, or by any combination of programmed computer components and custom hardware components.
• Elements of the present invention may also be provided as a machine-readable medium for storing the machine-executable instructions. The machine-readable medium may include, but is not limited to, floppy diskettes, optical disks, CD-ROMs, and magneto-optical disks, FLASH memory, ROMs, RAMs, EPROMs, EEPROMs, magnetic or optical cards, propagation media or other type of media/machine-readable medium suitable for storing electronic instructions. For example, elements may be downloaded as a computer program transferred from a remote computer (e.g., a server) to a requesting computer (e.g., a client) by way of data signals embodied in a carrier wave or other propagation medium via a communication link (e.g., a modem or network connection).
• In the foregoing specification, specific example embodiments have been described. It will, however, be evident that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the appended claims. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.

Claims (27)

1. A method, comprising:

loading an array of content into a two-dimensional shift register, the two-dimensional shift register coupled to an execution lane array;

repeatedly performing a first sequence comprising:

shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content; and,

repeatedly performing a second sequence comprising:

shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column, the resultant row or column having values determined at least in part from the mathematical operations of the first sequence, and performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.

2. The method ofclaim 1 wherein the mathematical operations comprise addition.

3. The method ofclaim 2 wherein the content comprises image intensity values.

4. The method ofclaim 2 wherein the content comprises per bin counts of a histogram.

5. The method ofclaim 4 wherein content containing counts for a set of bins of the histogram are shifted into a location having content containing counts for the same set of bins of the histogram.

6. The method ofclaim 1 wherein the content is image content.

7. The method ofclaim 1 wherein the second sequence further comprises:

shifting with the shift register content from a second set of first locations along a second resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding second set of second locations along the second resultant row or column, the second resultant row or column having values determined at least in part from the mathematical operations of the first sequence, and performing second mathematical operations on items of content from the second set of first locations and respective items of content from the second set of second locations.

8. The method ofclaim 6 wherein the content comprises per bin counts of a histogram wherein the set of first locations and the second set of first locations include per bin counts of different bins of the histogram.

9. The method ofclaim 8 further comprising, prior to the first and second sequences, merging storage space nominally reserved for individual execution lanes and having groups of execution lanes write to merged storage space to update bin counts.

10. The method ofclaim 1 further comprising repeatedly performing the loading, the first sequence and the second sequence over any of:

an entire image globally;

tiles of an image;

a window of an image.

11. The method ofclaim 10 wherein tile of an image are operated over and further performing a context switch of accumulated content across processing of tile boundaries.

12. A machine readable storage medium containing program code that when processed by a processor comprising a two-dimensional shift register coupled to an execution lane array causes a method to be performed, comprising:

loading an array of content into the two-dimensional shift register;

repeatedly performing a first sequence comprising:

shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content; and,

repeatedly performing a second sequence comprising:

shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column, the resultant row or column having values determined at least in part from the mathematical operations of the first sequence, and performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.

13. The machine readable storage medium ofclaim 12 wherein the mathematical operations comprise addition.

14. The machine readable storage medium ofclaim 13 wherein the content comprises image intensity values.

15. The machine readable storage medium ofclaim 13 wherein the content comprises per bin counts of a histogram.

16. The machine readable storage medium ofclaim 15 wherein content containing counts for a set of bins of the histogram are shifted into a location having content containing counts for the same set of bins of the histogram.

17. The machine readable storage medium ofclaim 12 wherein the content is image content.

18. The machine readable storage medium ofclaim 12 wherein the second sequence further comprises:

shifting with the shift register content from a second set of first locations along a second resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding second set of second locations along the second resultant row or column, the second resultant row or column having values determined at least in part from the mathematical operations of the first sequence, and performing second mathematical operations on items of content from the second set of first locations and respective items of content from the second set of second locations.

19. The machine readable storage medium ofclaim 17 wherein the content comprises per bin counts of a histogram wherein the set of first locations and the second set of first locations include per bin counts of different bins of the histogram.

20. The machine readable storage medium ofclaim 19 further comprising, prior to the first and second sequences, merging storage space nominally reserved for individual execution lanes and having groups of execution lanes write to merged storage space to update bin counts.

21. The machine readable storage medium ofclaim 12 further comprising repeatedly performing the loading, the first sequence and the second sequence over any of:

an entire image globally;

tiles of an image;

a window of an image.

22. The machine readable storage medium ofclaim 21 wherein tile of an image are operated over and further performing a context switch of accumulated content across processing of tile boundaries.

23. A computing system, comprising:

a plurality of processing cores;

a memory controller coupled to a system memory, the memory controller coupled to the plurality of processing cores;

an image processor, the image processor having a two-dimensional shift register coupled to an execution lane array;

a storage medium containing program code that when processed by the image processor causes a method to be performed, comprising:

loading an array of content into the two-dimensional shift register;

repeatedly performing a first sequence comprising:

shifting with the shift register first content residing along a particular row or column into another parallel row or column where second content resides and performing mathematical operations with a particular corresponding row or column of the execution lane array on the first and second content; and,

repeatedly performing a second sequence comprising:

shifting with the shift register content from a set of first locations along a resultant row or column that is parallel with the rows or columns of the first sequence into a corresponding set of second locations along the resultant row or column, the resultant row or column having values determined at least in part from the mathematical operations of the first sequence, and performing mathematical operations on items of content from the set of first locations and respective items of content from the set of second locations with the execution lane array.

24. The computing system ofclaim 23 wherein the mathematical operations comprise addition.

25. The computing system ofclaim 24 wherein the content comprises image intensity values.

26. The computing system ofclaim 23 wherein the content comprises per bin counts of a histogram.

27. The computing system ofclaim 26 wherein content containing counts for a set of bins of the histogram are shifted into a location having content containing counts for the same set of bins of the histogram.

Images