US20080077769A1 - Apparatus for efficient lfsr in a simd processor - Google Patents

Apparatus for efficient lfsr in a simd processor Download PDF

Info

Publication number
US20080077769A1
US20080077769A1 US11923576 US92357607A US2008077769A1 US 20080077769 A1 US20080077769 A1 US 20080077769A1 US 11923576 US11923576 US 11923576 US 92357607 A US92357607 A US 92357607A US 2008077769 A1 US2008077769 A1 US 2008077769A1
Authority
US
Grant status
Application
Patent type
Prior art keywords
vector
vrs
vrd
element
lfsr
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Abandoned
Application number
US11923576
Inventor
Tibet MIMAR
Original Assignee
Tibet MIMAR
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00Error detection; Error correction; Monitoring
    • G06F11/22Detection or location of defective computer hardware by testing during standby operation or during idle time, e.g. start-up testing
    • G06F11/26Functional testing
    • G06F11/27Built-in tests

Abstract

The apparatus provides for efficient implementation of multiple-bit leap-forward LFSR calculation in a SIMD processor. This provides an accelerated and programmable way to implement LFSR calculations in a SIMD processor. Conditional vector exclusive-OR accumulation is used by manipulating the leap-forward matrix, whereby one conditional vector exclusive-OR operation is performed for each column and partial results are accumulated. For an N-wide SIMD this results in close to N times acceleration of leap-forward LFSR calculation without additional resources or dedicated logic.

Description

    BACKGROUND OF THE INVENTION
  • 1. Field of the Invention
  • The invention relates generally to the field of processor chips and specifically to the field of Single-Instruction Multiple-Data (SIMD) processors. More particularly, the present invention relates to Linear Feedback Shift Register (LFSR) implementation in a SIMD processing system.
  • 2. Background
  • Linear Feedback Shift Registers (LFSRs) are commonly used in applications where pseudo-random bit streams are required. LFSRs are the functional building blocks of circuits like the pseudo-random noise (PN) code generator and Gold code generators commonly used in Code Division Multiple Access (CDMA) systems. These random numbers are used in a wide variety of applications, including data encryption, circuit testing, system simulation and Monte Carlo method. In the past, random number generation is done either in software on a scalar processor, or in hardware using shift-registers and exclusive-or gates. These generate one bit of output at a time. The data generation rate of these approaches is nowhere near adequate of what is needed by the latest generation of systems. SIMD and other processors like to consume large amounts of data in parallel because of their inherent parallelism. Several hardware approaches have been used to generate LFSR output at a much higher data rate. Accumulation method is a straightforward extension of previous one-bit methods. In this method, we can obtain an N-bit value by accumulating the one-bit N times. This can be done either by utilizing N copies of the identical hardware or by repeating the one-bit generator for N clocks. Leap-Forward LFSR technique advances the LFSR N steps in one clock cycle. This is based on the observation that LFSR is a linear system and can be written in vector format. Lagged Fibonacci method processes an N-bit word directly using exclusive-OR operator, which can be bit wise XOR, addition, or multiplication. This approach requires L previous values to be memorized, i.e., kept in FIFO register file memory.
  • All these hardware approaches require considerable number of gates to implement this as a fixed-function. This means these gates cannot be used for other functions, or even for a different type of random number. Applications nowadays may require several different type of LFSRs and each of these has to be implemented separately.
  • Multiple-Bit Leap-Forward LFSR
  • Leap-forward LFSR method utilizes only one LFSR and shifts out several bits. This method is based on the observation that an LFSR is a linear system and the register state can be written in vector format:
    Q(i+1)=A·q(i)
  • In this equation, q(i+1) and q(i) are the contents of shift register at (i+1)th and steps, and A is the transition matrix. After the LFSR advances k steps, the equation becomes Q ( i + k ) = A · q ( i + k - 1 ) = A · ( A · q ( i + k - 2 ) = A 2 · q ( i + k - 2 ) = Q ( i + k ) = Ak · q ( i )
  • The matrix calculation is such that logical AND operation is used instead of the traditional multiply and exclusive-OR operation is used instead of the traditional summation in matrix multiplication. The symbol “·” represents the multiply which is implemented as binary AND operation. Thus, we can calculate Ak from A. Such an LFSR could leap k steps in one clock cycle.
  • Let us use the 4-bit LFSR as an example of how matrix operations are carried out. FIG. 1 illustrates the matrix operations.
  • SUMMARY OF THE INVENTION
  • The present invention provides a method by which SIMD processor could implement the leap-forward LFSR technique with minor Instruction Set Architecture (ISA) changes, and could generate a large number of LFSR bits in parallel. For example, an 8-wide SIMD with 16-bits per element could generate 128-bits using only the same number of SIMD instructions as the number of LFSR delay stages. This provides a software programmable and flexible approach for implementing multiple LFSR and with little or no hardware overhead. The two features of SIMD ISA required are conditional execution of any SIMD instruction based on a scalar value and the ability to accumulate XOR values. The latter refers to performing a vector operation implementing bit wise exclusive OR (XOR) between a vector register and the vector accumulator.
  • BRIEF DESCRIPTION OF THE DRAWINGS
  • The accompanying drawings, which are incorporated and form a part of this specification, illustrate prior art and embodiments of the invention, and together with the description, serve to explain the principles of the invention:
  • Prior Art FIG. 1 illustrates an example of multiplying multiple LFSR matrices together using logical-AND and logical-XOR operations.
  • FIG. 2 shows the 27-tap LFSR used for a PN sequence generation.
  • FIG. 3 illustrates the leap-forward LFSR calculation for an 8-wide SIMD.
  • FIG. 4 illustrates per-vector-element Condition Code and Mask Control of SIMD Operations, that is, the operation of enable/disable bit control and condition code control of vector operations
  • FIG. 5 shows the programming model of vector registers for the preferred embodiment.
  • FIG. 6 shows the VXOR vector instruction definition for the preferred embodiment.
  • FIG. 7 shows the LDVCR Load Vector Condition Code Register scalar instruction definition for the preferred embodiment.
  • FIG. 8 shows the LDVRNL Load Vector Register with a scalar register definition for the preferred embodiment. This instruction vectorizes a scalar value.
  • FIG. 9 shows the block diagram of RISC plus SIMD processors working as dual-issue and having a common shared memory for the preferred embodiment.
  • FIG. 10 shows the different types of combining dual-input operands for vector operations. For the LFSR calculation, only the standard vector element-to-element mapping is used.
  • DETAILED DESCRIPTION
  • In the preferred embodiment of present invention OR1200 processor core with OpenRISC instruction set architecture is used as the RISC processor. The overall architecture of processor could be considered a combination of Long Instruction Word (LIW) and Single Instruction Multiple Data Stream (SIMD). This is because it issues two instructions every clock cycle, one RISC instruction and one SIMD instruction. SIMD processor can have any number of processing elements. OR1200 instruction is scalar working on a 32-bit data unit, and SIMD processor is a vector unit working on 8 16-bit data units in parallel.
  • The data memory is 128-bits wide to support 8 wide SIMD operations. The scalar OR1200 and the vector unit share the data memory. A cross bar is used to handle memory alignment transparent to the software, and also to select a portion of memory to access by OR1200 processor. The data memory is dual-port SRAM that is concurrently accessed by the SIMD processor and DMA engine. The data memory is also used to store constants and history information as well input as input and output video data. This data memory is shared between the RISC and SIMD processor.
  • While the DMA engine is transferring the processed data block out or bringing in the next 2-D block of video data, the vector processor concurrently processes the other data memory module contents. Successively, small 2-D blocks of video frame such as 64 by 64 pixels are DMA transferred, where these blocks could be overlapping on the input for processes that require neighborhood data such as 2-D convolution.
  • Vector processor or SIMD simply performs data processing, i.e., it has no program flow control instructions. OR1200 scalar processor is used for all program flow control. MIPS processor also additional instructions to load and store vector registers.
  • Each instruction word is 64 bits wide, and contains one scalar and one vector instruction. The OR1200 processor executes the scalar instruction, and the SIMD vector processor executes vector instruction. In assembly code, one scalar instruction and one vector instruction are written together on one line, separated by a semicolon. Comments could follow using double forward slashes as in C++. In this example, scalar processor is acting as the I/O processor loading the vector registers, and vector unit is performing vector-multiply
  • (VMUL) and vector-multiply-accumulate (VMAC) operations. These vector operations are performed on 8 input element pairs, where each element is 16-bits.
    Figure US20080077769A1-20080327-C00001
  • If a line of assembly code does not contain a scalar and vector instruction pair, the assembler will infer a NOP for the missing instruction. This NOP could be explicitly written or simply omitted.
  • In general, RISC processor has the standard OpenRISC instruction set plus vector load and store instructions. Both RISC and SIMD has register-to-register model, i.e., operate only on data in registers. RISC has the standard 32 32-bit data registers. SIMD vector processor has its own set of vector register, but depends on the RISC processor to load and store these registers between the data memory and vector register file.
  • Using 16-bits of interim resolution between pipeline stages of video processing, and 48-bit accumulation within a stage produces high quality video results, as opposed to using 12-bits and smaller accumulators.
  • SIMD Vector Unit and Basic Modes of Operation
  • The vector unit consists of multiple vector register files and a vector execution unit. The vector execution unit consists of multiple identical execution units, where each processing element operates on its slice of the vector register file. Each processing unit has its own 48-bit wide accumulator register for holding the exact results of multiply, accumulate, and multiply-accumulate operations.
  • The vector unit uses a load-store model, i.e., all vector operations uses operands sourced from vector registers, and the results of these operations are stored back to the register file. For example, the instruction “VMUL VR4, VRO, VR31” multiplies eight pairs of corresponding elements from vector registers VRO and VR31, and stores the results into vector register VR4. The results of the multiplication for each element results in a 32-bit result, which is stored into the accumulator for that element position. Then this 32-bit result for element is clamped and mapped to 16-bits before storing into elements of destination register.
  • There are three basic forms of vector operations, as shown in FIG. 9. In its most basic form, all corresponding elements from two source vector operands are operated pair-wise. This is the form most commonly used for FIR filters, 2-D convolution and other video operations. Second form is the broadcast mode, where one element of a vector register is used as the source operand to pair with all elements of a second vector register. The third form is where any element of one source vector register is paired with any element of a second source vector register. The element selection is determined by a third source vector register, which also acts as a mask to disable writing into certain output elements optionally. The LFSR implementation discussed here is limited to standard vector mapping format.
  • Please note that vector elements are numbered from 0 to 15, with element 0 corresponding to the “leftmost” or most significant bits of the 256-bit vector element. This is consistent with big endian addressing, where the bytes 0 and 1 align with element 0 and bytes 31 and 32 align with element 31. Even though we used 16-bit short words here, all elements are in two bytes, but nonetheless all addresses are specified in terms of bytes to be consistent with general RISC notation.
  • Programmers' Model
  • The programmers' model is shown in FIG. 4-5. All basic OR1200 programmer's model registers are included, which includes thirty-two 32-bit registers. The vector unit model has three sets vector registers: primary, alternate, and motion estimation registers, as the following will describe.
  • Primary Vector Registers (VR31-VRO)
  • These 32 256-bit register file is the primary workhorse of data crunching. These registers contain 8 16-bit elements. These registers can be used as source and destination of vector operations. In parallel with vector operations, these registers could be loaded or stored from/to data memory by the scalar unit.
  • Motion Estimation Vector Registers
  • These registers are not relevant for the LFSR calculations here.
  • Vector Accumulators
  • The accumulator registers are shown in three parts: high, middle, and low 16-bits for each element. These three portions make up the 48-bit accumulator register corresponding to each element position.
  • Vector Condition Codes
  • There are sixteen condition code flags for each vector element. Two of these are permanently wired as true and false. The other 14 condition flags are set by the vector compare instruction (VCMP), or loaded by LDVCR scalar instruction. All vector instructions are conditional in nature and use these flags. We will revisit the condition codes later in detail.
  • Instruction Set
  • We could categorize instructions into three groups:
  • A. OR1200 Instructions;
  • B. Load Vector, Store Vector Instructions, and vector LUT instructions;
  • C. SIMD Instructions.
  • The first two, i.e., OR1200 and load/store vector instruction are executed by the OR1200 core, SIMD instructions are executed by the vector unit. We will refer to these as scalar and vector instructions, respectively. Vision processor is dual-issue and can execute one scalar and one vector instruction every clock cycle. The scalar processor handles program flow control for both scalar and vector unit at the same time.
  • There are six vector processor instruction formats that are categorized into four groups in general, as shown in Table 1. VRd refers to the destination vector register and VRs-1, VRs-2, and VRs-3 refer to source-1, source-2, and source-3 vector registers, respectively. Not all instructions require two or three source vector registers; some instructions require only one source vector register. VRs-2 is typically used for mapping control. VRs-1, VRs-2, VRs-3 and VRd are part of the main vector register file. All SIMD vector instructions are conditional, i.e., their execution is based on a selected condition code flag. Optional CC represents the condition code selection, and it could be omitted if “always true” is to be selected.
    TABLE 1
    Vector Instruction Format Groups.
    For-
    mat
    # Instruction Syntax Mode
    0 <Instruction>.<CC> VRd, VRs-1, VRs-2 Standard
    <Instruction>.<CC> VRd, VRs-1, VRs-2, Rn
    <Instruction>.<CC> VRd, VRs-1, VRs-2, #imm5
    1 <Instruction>.<CC> VRd, VRs-1, V-2[element] Broadcast
    2 <Instruction>.<CC> VRd, VRs-1, VRs-2, VRs-3 Full Mapping
    3 (Reserved)
  • Format 0 refers to standard element-to-element mapping. This format could be used with or without a scalar register.
  • Format 1 uses the broadcast form, where one vector element of source-2 vector register is operated across all elements of source-1 vector register.
  • Format 2 refers to full mapping of vector elements. This form uses another vector register selected to determine the mapping, where any two elements of two source vector registers could be paired for vector operations.
  • The control fields of the vector control register, VRs-3, for each element are defined as follows:
      • Bits 3-0: Select source element from VRs-1 vector register;
      • Bits 5-4: Reserved. Set these as zero;
      • Bit 6: When set to one, selects zero for VRs-1
      • Bit 7: When set to one, scales down VRs-1 by one bit arithmetic shift, prior to operation.
      • Bits 11-8: Select source element from VRs-2 vector register;
      • Bits 13-12: Reserved. Set these as zero;
      • Bit 14: When set to one, selects zero for VRs-2
      • Bit 15: Mask bit, when set to one disables writing output for that element.
        The condition code select field is common to all vector elements, and is defined as part of condition field.
        Vector Load and Store Instructions
  • The vector load and store instructions are used to transfer contents of primary, alternate and motion estimation vector registers to and from on-chip data memory of Vision Processor. The most commonly used one is LDV, which loads eight 16-bit vector elements to a specified primary vector register. The effective address (EA) of load from memory is determined using indirect-with-offset addressing mode. This effective-address point to the element-O most-significant byte in data memory. Vision Processor uses big-endian addressing. Furthermore, all addresses are specified in terms of byte addresses, even though the granularity of operations is in terms of 16-bit elements. LDV instruction loads byte pointed by EA and 16 following bytes into specified vector register, where byte zero will go to high byte of element 0 and byte 1 will be stored to low-byte of element 0. Store vector instruction does the opposite, i.e., moves contents of a vector register to data memory.
  • The Vision Processor load and store hardware will automatically handle any memory alignment issues due to large memory width, as we described in the previous section. The programmer specifies addresses in terms of byte, but the effective address has to be aligned to 16-bit boundaries, because all vector elements are 16-bits wide. Both the offset and base register is required to be even byte addresses. If the offset is odd, then the least significant is discarded and a warning message is issued. If the base address' least significant bit is not zero, then it is simply discarded without warning while executing the instruction. The opcode mapping of offset values will only store bits one and higher, since address bit must be zero, and thus there is no reason to waste opcode bit space for this.
    Vector Load and Store Instructions (Part of Scalar Processor)
    Assembly Syntax Description
    LDV VRd, offset-17 (Rn) Load all elements of vector register.
    EA = Rn + (offset-17);
    VRd
    Figure US20080077769A1-20080327-P00801
    Elements Starting @ EA;
    Offset-17 is unsigned in bytes.
    LDV. <M> VRd [element], offset-9 (Rn) M is the coded load type value: 1, 2, 4 elements
    EA = Rn + (offset-6);
    VRd [element . . . element + N]
    Figure US20080077769A1-20080327-P00801
    M Elements @EA;
    Offset-9 is unsigned in bytes.
    LDVME offset-11 (Rn) Load 8 elements of Motion Estimation vector register file into
    register O. All ME registers are shifted from 8 toward 1 as
    loading of VRME16 occurs, i.e., we have:
    VRMEI
    Figure US20080077769A1-20080327-P00801
    VRME2;
    VRME<n>F VRME<n + 1>;
    EA = Rn + (offset-11 * 2);
    New 128-bit value @ EA-) → VRME16;
    Offset-11 is unsigned in I6-bit half-words.
    LDVBS VRd, offset-14 (Rn) Load 8 Bytes into Low Bytes of Vector Register Elements with
    Sign-Extension:
    EA = Rn + (offset-14);
    For n = 0; n < 32; n++
    VRd[n]
    Figure US20080077769A1-20080327-P00801
    SE (byte @ EA + n);
    Offset-14 is unsigned in bytes.
    LDVB VRd, offset-14 (Rn) Load 8 Bytes into Low Bytes of Vector Register Elements:
    EA = Rn + (offset-14); Offset-14 is unsigned in bytes.
    For n = 0; n, 3 2; n++
    VRd[n]
    Figure US20080077769A1-20080327-P00801
    18 II byte @ EA + n);
    LVRNH VRd, Rn Vectorize High Word of Scalar Register:
    VRd Rn[31:16]
    LVRNL VRd, Rn Vectorize Low Word of Scalar Register:
    VRd
    Figure US20080077769A1-20080327-P00801
    Rn[15:0]
    LDVCR offset-14 (Rn) Load Vector Condition Register: Bits 14 and 15 of each element
    remain unchanged.
    EA = Rn + (offset-14); Offset-14 is unsigned in bytes.
    VCR
    Figure US20080077769A1-20080327-P00801
    8words@EA;
    STVCR offset-14 (Rn) Store Vector Condition Register:
    EA = Rn + (offset-14); Offset-14 is unsigned in bytes.
    8 words @ EA
    Figure US20080077769A1-20080327-P00801
    VCR;
    STV VRs, offset-14 (Rn) Store all elements of vector register.
    EA = Rn + (offset-14); Offset-14 is unsigned in bytes.
    8 words @ EA E
    Figure US20080077769A1-20080327-P00801
    VRs;
    STVB VRs, offset-14 (Rn) Store High Bytes of Vector Register Elements
    EA = Rn + (offset-14); Offset-14 is unsigned in bytes.
    For n = 0; n < 8; n++
    (byte @ EA + n)
    Figure US20080077769A1-20080327-P00801
    VRs[n]15:8;

    Notes:

    No vector or scalar condition codes are affected.
  • Arithmetic Instructions
    Assembly Syntax Description
    Arithmetic Instructions
    VINC.[cond] VRd, VRs-1, #imm5, VRs-3 Increment by One:
    VINC.[cond] VRd, VRs, #imm5 VRd
    Figure US20080077769A1-20080327-P00801
    VRs + imm5
    imm5: 5-bit unsigned number.
    VABS.[cond] VRd, VRs, VRs-3 Absolute Value:
    VABS.[cond] VRd, VRs VACC 4-abs (VRs)
    VRd 4-Signed-Clamp (VACC)
    VADD.[cond] VRd, VRs-1, VRs-2, VRs-3 Addition:
    VADD.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VRs-1 + VRs-2
    VADD.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VADDS.[cond] VRd, VRs-1, VRs-2, VRs-3 Addition Scaled:
    VADDS.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    (VRs-1 + VRs-2)/2
    VADDS.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VSUB.[cond] VRd, VRs-1, VRs-2, VRs-3 Subtraction:
    VSUB.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VRs1 − VRs-2
    VSUB.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VMUL.[cond] VRd, VRs-1, VRs-2, VRs-3 Multiply:
    VMUL.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VRs1 * VRs-2
    VMUL.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VABSD.[cond] VRd, VRs-1, VRs-2, VRs-3 Absolute Difference:
    VABSD.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    abs (VRs1 − VRs-2)
    VABSD.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VABSDS.[cond] VRd, VRs-1, VRs-2, VRs-3 Absolute Difference Scaled:
    VABSDS.[cond] VRd, VRs-I, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    abs (VRs1 − VRs-2)/2
    VABSDS.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    Vector-Accumulate Instructions: Results Affect Accumulator and Destination Vector Register.
    VSAD.[cond] VRd, VRs-1, VRs-2, VRs-3 Sum-of-Absolute-Differences:
    VSAD.[cond] VRd, VRs-1, VRs-2 VACC
    Figure US20080077769A1-20080327-P00801
    VACC + abs (VRs-1 − VRs-2)
    VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VACC)
    VADDA.[cond] VRd, VRs-1, VRs-2, VRs-3 Add-Accumulate:
    VADDA.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VACC + (VRs-1 + VRs-2)
    VADDA.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
    VADDSA.[cond] VRd, VRs-1, VRs-2, VRs-3 Add-Subtract-From-Accumulator:
    VADDSA.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VACC − (VRs-1 + VRs-2)
    VADDSA.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
    VSUBA.[cond] VRd, VRs-1, VRs-2, VRs-3 Subtract-Accumulate:
    VSUBA.[cond] VRd, VRs-1, VRs-2 VACC
    Figure US20080077769A1-20080327-P00801
    VACC + (VRs-1 − VRs-2)
    VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
    VSUBSA.[cond] VRd, VRs-1, VRs-2, VRs-3 Subtract-Subtract-From-Accumulator::
    VSUBSA.[cond] VRd, VRs-1, VRs-2 VACC
    Figure US20080077769A1-20080327-P00801
    VACC − (VRs-1 − VRs-2)
    VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
    VMAC.[cond] VRd, VRs-1, VRs-2, VRs-3 Multiply-Accumulate:
    VMAC.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VACC + (VRs-1 * VRs-2)
    VMAC.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
    VMAS.[cond] VRd, VRs-1, VRs-2, VRs-3 Multiply-Subtract-Accumulate:
    VMAS.[cond] VRd, VRs-1, VRs-2 [element] VACC
    Figure US20080077769A1-20080327-P00801
    VACC − (VRs-1 * VRs-2)
    VMAS.[cond] VRd, VRs-1, VRs-2 VRd
    Figure US20080077769A1-20080327-P00801
    Signed-Clamp (VCC)
  • Vector Logical Instructions
    Assembly Syntax Description
    VOR.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical OR Operation:
    VOR.[cond] VRd, VRs-1, VRs-2 [element] VRd
    Figure US20080077769A1-20080327-P00801
    VRs-1 OR VRs-2
    VOR.[cond] VRd, VRs-1, VRs-2
    VNOR.[cond] VRd, VRs 1, VRs-2, VRs-3 Logical NOR Operation:
    VNOR.[cond] VRd, VRs-1, VRs-2 [element] VRd
    Figure US20080077769A1-20080327-P00801
    VRs-1 NOR VRs-2
    VNOR.[cond] VRd, VRs-1, VRs-2
    VXOR.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical XOR Operation:
    VXOR.[cond] VRd, VRs-1, VRs-2 [element] VRd
    Figure US20080077769A1-20080327-P00801
    VRs-1 XOR VRs-2
    VXOR.[cond] VRd, VRs-1, VRs-2
    VXORACC.[cond] VRd, VRs-1, VRs-2, VRs-3 Accumulate Logical XOR
    VXORACC.[cond] VRd, VRs-1, VRs-2 [element] Operation: VACC31 . . . 16
    Figure US20080077769A1-20080327-P00801
    VXORACC.[cond] VRd, VRs-1, VRs-2 VACC31 . . . 16 XOR
    (VRs-1 XOR VRs-2);
    VRd
    Figure US20080077769A1-20080327-P00801
    VACC31 . . . 16;
    VAND.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical AND Operation:
    VAND.[cond] VRd, VRs-1, VRs-2 [element] VRd
    Figure US20080077769A1-20080327-P00801
    VRs-1 AND VRs-2
    VAND.[cond] VRd, VRs-1, VRs-2
    VANDNOT.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical AND with a
    VANDNOT.[cond] VRd, VRs-1, VRs-2 [element] complement Operation: VRd
    Figure US20080077769A1-20080327-P00801
    VANDNOT.[cond] VRd, VRs-1, VRs-2 VRs-1 AND NOT (VRs-2)
    VNAND.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical NAND Operation:
    VNAND.[cond] VRd, VRs-1, VRs-2 [element] VRd
    Figure US20080077769A1-20080327-P00801
    VRs-1 NAND VRs-2
    VNAND.[cond] VRd, VRs-1, VRs-2
    VNOT.[cond] VRd, VRs, VRs-3 Logical NOT Operation:
    VNOT.[cond] VRd, VRs [element] VRd
    Figure US20080077769A1-20080327-P00801
    NOT (VRs)
    VNOT.[cond] VRd, VRs
    VLSL.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical Shift Left by N Bits
    VLSL.[cond] VRd, VRs-1, VRs2 [element] Operation: VRd
    Figure US20080077769A1-20080327-P00801
    VLSL.[cond] VRd, VRs-2, VRs-2 (VRs-1) << VRs-2
    VLSR.[cond] VRd, VRs-1, VRs-2, VRs-3 Logical Shift Right by N Bits
    VLSR.[cond] VRd, VRs-1, VRs2 [element] Operation: VRd
    Figure US20080077769A1-20080327-P00801
    VLSR.[cond] VRd, VRs-2, VRs-2 (VRs-1) >> VRs-2

    Leap-Forward LFSR Implementation
  • The present invention uses existing resources of a SIMD processor with small changes to accommodate the operation leap-forward LFSR calculation. As an example, let us assume we have a 8-wide SIMD and we are to calculate the code values for LFSR shown in FIG. 2. Typically, such an LFSR could leap k steps in one clock cycle. For example, we could leap 27 clock cycles, and then use the 27 bits of Q values are the 27-bits of generated code. We actually need to generate 128-bits of code to match the SIMD width. This would require multiple leap forward matrices for a combined matrix size of 128×27, as shown in FIG. 3. In this combined matrix, we would have:
    J=128;
    K=J+27=155;
    L=K+27=J+2*27=182;
    M=L+27=J+3*27=209;
    N=M+27=J+4*27=236;
  • We could implement the matrix operation in sequence of rows, but this would require AND followed by XOR gates. Alternatively, we could operate column-by-column and XOR accumulate partial results as we proceed. This requires looking the first value of q (vector of 27 ones or zeros for our example), and if this value is one, exclusive OR operation of first column of A′ matrix with the vector accumulator. The vector accumulator is initialized as zeros. Next, we would look at the second value of q, and we would conditionally exclusive-OR the second column of A′ with vector accumulator, if the second value of q is a one, and so forth. We could store the A′ matrix in 27 alternate vector registers.
  • We would first load the 27-bits of input vector q, into a scalar 32-bit register. We would then load the Vector Condition Register (VCR) from this register, by “splatting” or broadcasting 14-bits of this to all elements of VCR. In the preferred embodiment, we could only use 14-bits of each VCR vector element, because the other two bits are hardwired as true and false conditions. These fourteen condition bits could then be used as the flags to conditionally execute the vector XOR-accumulate instructions.
  • The LFSR calculation would be implemented with SIMD vector XOR instruction given by the following (using the standard mapping option):
  • VXORACC.n VRd, VRs-1, VRs-2
  • This would perform the following operation using C-type pseudo language:
    If (VCR,, = =1)
    {
    VACC31..16F VACC31..16XOR (VRs-1 XOR VRs-2);
    VRd F VACC31..16;
    }

    In other words, the conditional XOR vector operation is executed only if the selected bit “n” of Vector Condition Register (VCR) is one. In this case we expect, all elements to have the same “n” value due to how they were loaded.
  • We could use the VRs-1 and VRs-2 as the input vector and vector accumulator, but this has the disadvantage that multiple instructions could not be pipelines, unless the SIMD instruction pipeline bypasses intermediate results. In the preferred embodiment, the pipeline is exposed (i.e., not bypassed). However, if the use the vector accumulator, then this issue will not occur, and we could use back-to-back instructions. Thus, we will use VRs-2 as the input vector and VRs-1 as all zeros for this example. We assume vector accumulator is all zeros before we start. The following code segment illustrates how we could generate 128-bits of Q from q. The details of the LDVCR, VXORACC are given in FIGS. 6 and 7, respectively. The preferred embodiment could execute one scalar and one vector instruction for each clock cycle. We have not shown this explicitly by having one scalar and one vector instruction in this example for the sake of readability, but the assembler will still one scalar and one vector instruction into a single 64-bit Opcode for execution.
    //Generate 128-Bit Code
    //r0 = Input value in bits 26:0
    /VR 1 = Transformation Matrix A, left most column
    /VR2 = Transformation Matrix A, second left most column
    /VR3-27 = Other Columns of Transformation Matrix A, left to right
    /VR31 = Will contain result at the end of this code segment.
    ANDI rl, r0, Oxbff; //Get bits 13:0 of
    LVRNL VR0, rl; //Load rl into all elements of VRO;
    STV VRO, O(r2); //Store in memory to a temporary location
    LDVCR 0(r2); // Load VCR from the temporary location.
    VXOR VR0, VR0, VR0; //Zero out VRO elements
    VMUL VR0, VR0, VR0; //Zero Out Vector Accumulator
    VXORACC.acc.0 VR31, VR0, VRI;
    VXORACC.acc.1 VR31, VR0, VR2;
    VXORACC.acc.2 VR31, VR0, VR3;
    VXORACC.acc.3 VR31, VR0, VR4;
    VXORACC.acc.4 VR31, VR0, VR5;
    VXORACC.acc.5 VR31, VR0, VR6;
    VXORACC.acc.6 VR31, VR0, VR7;
    VXORACC.acc.7 VR31, VR0, VR8;
    VXORACC.acc.8 VR31, VR0, VR9;
    VXORACC.acc.9 VR31, VR0, VRIO;
    VXORACC.acc.I0 VR31, VR0, VR11;
    VXORACC.acc.l1 VR31, VR0, VR12;
    VXORACC.acc.12 VR31, VR0, VR13;
    VXORACC.acc.13 VR31, VR0, VR14;
    SRL r0, r0, 14;
    ANDI r1, r0, Oxbff; //Get bits 13:0 of
    LVRNL VR0, rl; //Load rl into all elements of VR0; //Store
    STV VR0, 0(r2); in memory to a temporary location // Load
    LDVCR 0(r2); VCR from the temporary location.
    VXORACC.acc.0 VR31, VR0, VR15;
    VXORACC.acc.1 VR31, VR0, VR16;
    VXORACC.acc.2 VR31, VR0, VR17;
    VXORACC.acc.3 VR31, VR0, VR18;
    VXORACC.acc.4 VR31, VR0, VR19;
    VXORACC.acc.5 VR31, VR0, VR20;
    VXORACC.acc.6 VR31, VR0, VR21;
    VXORACC.acc.7 VR31, VR0, VR22;
    VXORACC.acc.8 VR31, VR0, VR23;
    VXORACC.acc.9 VR31, VR0, VR24;
    VXORACC.acc.10 VR31, VR0, VR25;
    VXORACC.acc.11 VR31, VR0, VR26;
    VXORACC.acc.12 VR31, VR0, VR27;
  • As shown, we would need about 38 vector instructions to generate 128 bits of code leap-forward LFSR code.

Claims (9)

    I claim:
  1. 1. A SIMD processor core for performing vector operations comprising:
    a) a set of vector registers wherein each vector register comprises N elements wherein the N elements comprising parts of one-dimensional vector and a two-dimensional array and said set of vector registers are grouped together and are operably coupled to a plurality of read ports and plurality of write ports for accessing said set of vector registers at substantially the same time; and
    b) a plurality of arithmetic and logic processing units, each having one or two data inputs, wherein the inputs of each processing unit is operably coupled to read ports of said vector register file, and the output of these processing units are coupled to a write port of the vector register file; and
    c) Plurality of accumulator registers, the input of which are operably coupled to the output of processing units, and wherein the output of the accumulator registers are operably coupled to a write port of the vector register file.
  2. 2. The apparatus of claim 1, where LFSR leap-forward matrix is preformed or precalculated and stored for use in runtime. The values of this matrix are loaded into a plurality of vector registers and are used for LFSR leap-forward calculations.
  3. 3. The apparatus of claim 2, where the matrix multiplication with a vector input data is performed using column-by-column conditional exclusive-OR operation which is accumulated over all the columns of the matrix using exclusive-OR accumulation of interim result with the vector accumulator. A given column of LFSR matrix is exclusive-OR'ed with the vector accumulator, only if the corresponding bit of the input data vector is one, and otherwise that column is skipped.
  4. 4. The apparatus of claim 3, where a RISC processor is combined with a SIMD processor, whereby the RISC processor performs I/O processing and flow control for both itself and the SIMD processor, and one RISC and one SIMD instruction is executed every clock cycle.
  5. 5. The apparatus of claim 4, wherein the value of N is selected from a group of values consisting of integer values between 4 and 256.
  6. 6. The apparatus of claim 5, wherein each vector element is a byte.
  7. 7. The apparatus of claim 5, wherein each vector element is a 16-bit fixed-point word.
  8. 8. The apparatus of claim 5, wherein each vector element is a 32-bit fixed point word.
  9. 9. The apparatus of claim 5, wherein each vector element is a floating-point value.
US11923576 2004-04-05 2007-10-24 Apparatus for efficient lfsr in a simd processor Abandoned US20080077769A1 (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
US55987404 true 2004-04-05 2004-04-05
US11095435 US7302627B1 (en) 2004-04-05 2005-03-31 Apparatus for efficient LFSR calculation in a SIMD processor
US11923576 US20080077769A1 (en) 2004-04-05 2007-10-24 Apparatus for efficient lfsr in a simd processor

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US11923576 US20080077769A1 (en) 2004-04-05 2007-10-24 Apparatus for efficient lfsr in a simd processor

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
US11095435 Continuation US7302627B1 (en) 2004-04-05 2005-03-31 Apparatus for efficient LFSR calculation in a SIMD processor

Publications (1)

Publication Number Publication Date
US20080077769A1 true true US20080077769A1 (en) 2008-03-27

Family

ID=38722079

Family Applications (2)

Application Number Title Priority Date Filing Date
US11095435 Expired - Fee Related US7302627B1 (en) 2004-04-05 2005-03-31 Apparatus for efficient LFSR calculation in a SIMD processor
US11923576 Abandoned US20080077769A1 (en) 2004-04-05 2007-10-24 Apparatus for efficient lfsr in a simd processor

Family Applications Before (1)

Application Number Title Priority Date Filing Date
US11095435 Expired - Fee Related US7302627B1 (en) 2004-04-05 2005-03-31 Apparatus for efficient LFSR calculation in a SIMD processor

Country Status (1)

Country Link
US (2) US7302627B1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060227966A1 (en) * 2005-04-08 2006-10-12 Icera Inc. (Delaware Corporation) Data access and permute unit
US8719323B2 (en) 2010-06-16 2014-05-06 Lsi Corporation Methods for efficient state transition matrix based LFSR computations
US9281846B2 (en) 2013-07-31 2016-03-08 Globalfoundries Inc Turbo encoding on a parallel processor

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8650240B2 (en) * 2009-08-17 2014-02-11 International Business Machines Corporation Complex matrix multiplication operations with data pre-conditioning in a high performance computing architecture
US8577950B2 (en) * 2009-08-17 2013-11-05 International Business Machines Corporation Matrix multiplication operations with data pre-conditioning in a high performance computing architecture
US9600281B2 (en) 2010-07-12 2017-03-21 International Business Machines Corporation Matrix multiplication operations using pair-wise load and splat operations
US8504603B2 (en) 2010-07-28 2013-08-06 Lsi Corporation Method and system for parallel computation of linear sequential circuits
CN104011675B (en) * 2011-12-20 2017-07-07 联发科技瑞典有限公司 A vector execution unit for a digital signal processor
CN103543983B (en) * 2012-07-11 2016-08-24 世意法(北京)半导体研发有限责任公司 A novel method for improving the data access on fir throughput data path architecture balanced performance properties
US9361116B2 (en) * 2012-12-28 2016-06-07 Intel Corporation Apparatus and method for low-latency invocation of accelerators
US9417873B2 (en) 2012-12-28 2016-08-16 Intel Corporation Apparatus and method for a hybrid latency-throughput processor
US9542193B2 (en) 2012-12-28 2017-01-10 Intel Corporation Memory address collision detection of ordered parallel threads with bloom filters

Citations (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5513366A (en) * 1994-09-28 1996-04-30 International Business Machines Corporation Method and system for dynamically reconfiguring a register file in a vector processor
US5555428A (en) * 1992-12-11 1996-09-10 Hughes Aircraft Company Activity masking with mask context of SIMD processors
US5838984A (en) * 1996-08-19 1998-11-17 Samsung Electronics Co., Ltd. Single-instruction-multiple-data processing using multiple banks of vector registers
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US5903769A (en) * 1997-03-31 1999-05-11 Sun Microsystems, Inc. Conditional vector processing
US5973705A (en) * 1997-04-24 1999-10-26 International Business Machines Corporation Geometry pipeline implemented on a SIMD machine
US5991531A (en) * 1997-02-24 1999-11-23 Samsung Electronics Co., Ltd. Scalable width vector processor architecture for efficient emulation
US6058465A (en) * 1996-08-19 2000-05-02 Nguyen; Le Trong Single-instruction-multiple-data processing in a multimedia signal processor
US20030014457A1 (en) * 2001-07-13 2003-01-16 Motorola, Inc. Method and apparatus for vector processing
US6665790B1 (en) * 2000-02-29 2003-12-16 International Business Machines Corporation Vector register file with arbitrary vector addressing
US6959378B2 (en) * 2000-11-06 2005-10-25 Broadcom Corporation Reconfigurable processing system and method
US6963341B1 (en) * 2002-06-03 2005-11-08 Tibet MIMAR Fast and flexible scan conversion and matrix transpose in a SIMD processor
US7191317B1 (en) * 1999-07-21 2007-03-13 Broadcom Corporation System and method for selectively controlling operations in lanes
US7308559B2 (en) * 2000-02-29 2007-12-11 International Business Machines Corporation Digital signal processor with cascaded SIMD organization
US7376812B1 (en) * 2002-05-13 2008-05-20 Tensilica, Inc. Vector co-processor for configurable and extensible processor architecture

Patent Citations (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5555428A (en) * 1992-12-11 1996-09-10 Hughes Aircraft Company Activity masking with mask context of SIMD processors
US5513366A (en) * 1994-09-28 1996-04-30 International Business Machines Corporation Method and system for dynamically reconfiguring a register file in a vector processor
US5838984A (en) * 1996-08-19 1998-11-17 Samsung Electronics Co., Ltd. Single-instruction-multiple-data processing using multiple banks of vector registers
US6058465A (en) * 1996-08-19 2000-05-02 Nguyen; Le Trong Single-instruction-multiple-data processing in a multimedia signal processor
US5991531A (en) * 1997-02-24 1999-11-23 Samsung Electronics Co., Ltd. Scalable width vector processor architecture for efficient emulation
US5903769A (en) * 1997-03-31 1999-05-11 Sun Microsystems, Inc. Conditional vector processing
US5973705A (en) * 1997-04-24 1999-10-26 International Business Machines Corporation Geometry pipeline implemented on a SIMD machine
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US7159100B2 (en) * 1997-10-09 2007-01-02 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US7546443B2 (en) * 1997-10-09 2009-06-09 Mips Technologies, Inc. Providing extended precision in SIMD vector arithmetic operations
US7191317B1 (en) * 1999-07-21 2007-03-13 Broadcom Corporation System and method for selectively controlling operations in lanes
US7467288B2 (en) * 2000-02-29 2008-12-16 International Business Machines Corporation Vector register file with arbitrary vector addressing
US6665790B1 (en) * 2000-02-29 2003-12-16 International Business Machines Corporation Vector register file with arbitrary vector addressing
US7308559B2 (en) * 2000-02-29 2007-12-11 International Business Machines Corporation Digital signal processor with cascaded SIMD organization
US6959378B2 (en) * 2000-11-06 2005-10-25 Broadcom Corporation Reconfigurable processing system and method
US20030014457A1 (en) * 2001-07-13 2003-01-16 Motorola, Inc. Method and apparatus for vector processing
US7376812B1 (en) * 2002-05-13 2008-05-20 Tensilica, Inc. Vector co-processor for configurable and extensible processor architecture
US6963341B1 (en) * 2002-06-03 2005-11-08 Tibet MIMAR Fast and flexible scan conversion and matrix transpose in a SIMD processor

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060227966A1 (en) * 2005-04-08 2006-10-12 Icera Inc. (Delaware Corporation) Data access and permute unit
US7933405B2 (en) * 2005-04-08 2011-04-26 Icera Inc. Data access and permute unit
US8719323B2 (en) 2010-06-16 2014-05-06 Lsi Corporation Methods for efficient state transition matrix based LFSR computations
US9281846B2 (en) 2013-07-31 2016-03-08 Globalfoundries Inc Turbo encoding on a parallel processor

Also Published As

Publication number Publication date Type
US7302627B1 (en) 2007-11-27 grant

Similar Documents

Publication Publication Date Title
US5606677A (en) Packed word pair multiply operation forming output including most significant bits of product and other bits of one input
US5923871A (en) Multifunctional execution unit having independently operable adder and multiplier
US5644522A (en) Method, apparatus and system for multiply rounding using redundant coded multiply result
US5995748A (en) Three input arithmetic logic unit with shifter and/or mask generator
Tremblay et al. VIS speeds new media processing
US6173394B1 (en) Instruction having bit field designating status bits protected from modification corresponding to arithmetic logic unit result
US5590350A (en) Three input arithmetic logic unit with mask generator
US5781457A (en) Merge/mask, rotate/shift, and boolean operations from two instruction sets executed in a vectored mux on a dual-ALU
US5446651A (en) Split multiply operation
US6098163A (en) Three input arithmetic logic unit with shifter
US5884069A (en) Computer and a method of operating a computer to combine data values within a singularly addressable data string
US6067613A (en) Rotation register for orthogonal data transformation
US5922066A (en) Multifunction data aligner in wide data width processor
US5640578A (en) Arithmetic logic unit having plural independent sections and register storing resultant indicator bit from every section
US5687336A (en) Stack push/pop tracking and pairing in a pipelined processor
US6295599B1 (en) System and method for providing a wide operand architecture
US7197625B1 (en) Alignment and ordering of vector elements for single instruction multiple data processing
US20100274988A1 (en) Flexible vector modes of operation for SIMD processor
US5634065A (en) Three input arithmetic logic unit with controllable shifter and mask generator
US6230180B1 (en) Digital signal processor configuration including multiplying units coupled to plural accumlators for enhanced parallel mac processing
US7516308B2 (en) Processor for performing group floating-point operations
US6026484A (en) Data processing apparatus, system and method for if, then, else operation using write priority
US5805913A (en) Arithmetic logic unit with conditional register source selection
US20030105793A1 (en) Long instruction word controlling plural independent processor operations
US6115812A (en) Method and apparatus for efficient vertical SIMD computations