Processor Architecture

Abstract

A LIW processor comprises multiple execution units. The multiple execution units of the processor are divided into groups, and an input instruction word can contain instructions for one execution unit in each of the groups. The processor is optimised for use in signal processing operations, in that the multiple execution units of the processor are divided into groups which do not place significant restrictions on the desirable uses of the processor, because it has been determined that, in signal processing applications, it is not usually necessary for certain execution units to operate simultaneously. These execution units can therefore be grouped together, in such a way that only one of them can operate at a particular time, without significantly impacting on the operation of the device. An array is formed from multiple interconnected processors of this type.

Description

BACKGROUND

This invention relates to a processor architecture, and in particular to a processor architecture which is particularly useful in signal processing applications.

Modern high-performance wireless communications systems require digital processors which can provide billions of compute operations per second to achieve acceptable performance, for example to carry out operations such as filtering, equalisation and decoding functions. Increasingly these very high processing demands are satisfied by the use of multiple execution units (such as arithmetic logic units (ALUs), multipliers, address generators etc.) which can operate in parallel within a single processor cycle, and can thus increase the aggregate number of operations which can be completed per cycle.

One architectural approach which has been developed, in order to allow parallel operation of multiple execution units, is the Long Instruction Word (LIW) architecture. In this approach, instructions for each of a number of execution units are concatenated into one “long instruction word” which can be executed in a single processor cycle. Typically, in implementations of this approach, a bit field within the long instruction is reserved for an instruction for each of the execution units, regardless of whether a particular execution unit will be active within any one processor cycle. This has the disadvantageous effect that it creates excessively long instruction words, which can contain a lot of redundant information for execution units that are not active. The end result is a larger and more costly design.

SUMMARY OF THE INVENTION

The present invention relates to an alternative implementation of an LIW processor.

According to a preferred embodiment of the present invention, there is provided a processor which comprises multiple execution units. The multiple execution units of the processor are divided into groups, and an input instruction word can contain instructions for one execution unit in each of the groups.

In a further preferred embodiment of the invention, the processor is optimised for use in signal processing operations, in that the multiple execution units of the processor are divided into groups which do not place significant restrictions on the desirable uses of the processor. That is, it has been determined that, in signal processing applications, it is not usually necessary for certain execution units to operate simultaneously.

These execution units can therefore be grouped together, in such a way that only one of them can operate at a particular time, without significantly impacting on the operation of the device.

According to a further aspect of the present invention, there is provided an array, comprising a plurality of interconnected processors, wherein each of the processors comprises multiple execution units as defined above.

BRIEF DESCRIPTION OF THE DRAWINGS

For a better understanding of the present invention, and to show how it may be put into effect, reference will now be made to the accompanying drawings, in which:

FIG. 1 is a block schematic diagram of a processor array according to an aspect of the present invention;

FIG. 2 is a block schematic diagram of a processor within the processor array of FIG. 1, according to another aspect of the present invention;

FIG. 3 is an overview of the format of an instruction word for use in the processor of FIG. 2;

FIG. 4 illustrates in more detail the format of a part of the instruction word shown in FIG. 3;

FIG. 5 illustrates the operation of a second part of the instruction word shown in FIG. 3;

FIG. 6 illustrates the operation of a third part of the instruction word shown in FIG. 3.

DETAILED DESCRIPTION

FIG. 1 is a block schematic diagram of a processor array, as generally described in WO02/50624. The array is made up of array elements 20, which are interconnected by buses and switches.

The array architecture includes first bus pairs 30, shown running horizontally in FIG. 1, each pair including a respective first bus 32 carrying data from left to right in FIG. 1 and a respective second bus 36 carrying data from right to left.

The array architecture includes second bus pairs 40, shown running vertically in FIG. 1, each pair including a respective third bus 42 carrying data upwards in FIG. 1 and a respective fourth bus 46 carrying data downwards.

In FIG. 1, each diamond connection 50 represents a switch, which connects an array element 20 to a respective bus 32, 36. The array further includes a switch matrix 55 at each intersection of a first and second bus pair 30, 40. The data buses, and the switches and switch matrices, therefore allow data to be switched from one array element to another for processing, as required.

In this embodiment of the invention, at least some of the array elements 20 take the form of processors, as shown in more detail in FIG. 2. In accordance with this illustrated embodiment of the present invention, the processors 20 are adapted to make them particularly suitable for use as array elements, although the invention is also applicable to individual processors.

The processor 20 includes a 64×64 bit instruction memory 60, which contains instructions loaded into the memory to control the operation of the processor. In operation of the device, instructions are fetched from the instruction memory 60, and passed to an instruction decoder 62, where they are decoded to configure the datapaths and execution units in the processor.

In this illustrated embodiment, the processor comprises six execution units. The first available execution unit is a first Arithmetic Logic Unit (ALU) 64, which can perform a number of arithmetic and logical operations.

The second available execution unit is a communications unit 66, which is connected to the input communications bus 68 and the output communications bus 70, and is able to perform “put” and “get” operations to move data to and from the external communications buses 68, 70, and is also able to move data to and from the 15×16 bit data registers 84. The registers 84 are connected to the execution units by means of a data bus 85.

In this illustrated embodiment, the communications unit 66 is thereby optimised to support the processing performed in the array, whereby data flows from one processor 20 to another, with parts of the processing being performed at each stage.

The third available execution unit is a combined Memory Access Unit (MAU)/second ALU 72, which performs a variety of load and store operations over a bus 74 to a 64×32 bit data memory 76, and also provides a subset of the ALU operations performed by the first ALU 64.

The fourth available execution unit is a branch unit 78, which performs a number of conditional and unconditional branch operations.

The fifth available execution unit is a Multiplier Accumulator (MAC) Unit 80, which performs a variety of multiply and multiply accumulate operations with various bit widths. In an alternative embodiment of the invention, this unit may be replaced by a simpler Multiply unit.

In this illustrated embodiment of the invention, there is a sixth available execution unit in the form of an Application Specific Unit (ASU) 82. More specifically, the ASU 82 is adapted to perform a number of highly specialised operations for wireless signal processing applications, such as complex spread and complex despread, in order to support CDMA transmit and receive functionality. In an alternative embodiment of the invention, this unit may be omitted.

As is conventional, in general each execution unit is able to perform one operation in one clock cycle. However, The first ALU 64 is also able to perform a shift operation on the first operand of the basic arithmetic or logical operations. Thus, in this special case, two instructions can effectively execute simultaneously on that one execution unit.

Analysis of a wide range of signal processing applications has now led to the conclusion that it is not necessary for all of the execution units to be able to operate simultaneously. In this illustrated embodiment of the invention, the execution units are clustered into three groups, each controlled by a separate instruction in a LIW instruction.

Specifically, in this illustrated embodiment, the first group 86 includes only the first Arithmetic Logic Unit (ALU) 64; the second group 88 is made up of the communications unit 66, and the combined Memory Access Unit (MAU)/second ALU 72; and the third group 90 is made up the branch unit 78, the Multiplier Accumulator (MAC) Unit 80, and the Application Specific Unit (ASU) 82.

According to this preferred embodiment of the invention, the device is then controlled such that any one, any two, or all three of the groups 86, 88, 90 can be active at any one time, but such that no more than one of the execution units within a group can be active at any one time. Further, the instruction format is such that this can be achieved efficiently in each case.

Specifically, a long instruction word can include an instruction LIW#1 for the first group 86, an instruction LIW#2 for the second group 88, and an instruction LIW#3 for the third group 90.

FIG. 3 shows the basic structure of a long instruction word instruction, which is also explained in more detail in FIGS. 4, 5 and 6.

Thus, the long instruction word first contains a short, 3 bit, bit sequence, which indicates whether the first group 86 is active in that processor cycle and, if so, indicates what class of operation is to be performed, so that execution units and datapaths can be configured.

As shown in FIG. 4, except in the case where the first three bits of byte 0 are 000, the first group 86 is active in that processor cycle and that three bit sequence indicates what operation is to be performed by the first Arithmetic Logic Unit (ALU) 64.

Thus, when the value of the first three bits is within the range 001-100, the operation is an ALU operation with three operands, for example adding two values to give a result, with the three operands then being the register addresses of the two values to be added plus the register address in which the result is to be stored.

When the value of the first three bits is within the range 101-110, the operation is a load or store operation between the data memory and a nominated register or register pair.

When the value of the first three bits is 111, the operation is an ALU operation with two operands, one operand, or no operands, for example nop.

In any of these cases, the fourth bit then indicates whether an extension byte is to be used, as will be described in more detail below. The remaining four bits of byte 0, and the eight bits of byte 1, then indicate the operands or opcode values, depending on the value of the first three bits of byte 0, as shown in FIG. 4. More specifically, where FIG. 4 says that four of these bits represent an operand, they define the address, within the registers 84, from which the first ALU 64 should retrieve the respective operand on which it will perform the defined operation.

When the first three bits of byte 0 are not 000, and thus the first group 86 is active in that processor cycle, the fourth bit must be set to “1”, and the extension byte must be used, if either the second group 88 or the third group 90 is active.

In the case where the first three bits of byte 0 are 000, the first group 86 is not active in that processor cycle, and byte 0 of the long instruction word then contains further short bit sequences, which indicate whether the second group 88 and third group 90 are active and, if so, what class of operation is to be performed.

Thus, there is a 3 bit sequence Lcode 2 relating to the second group 88, as shown in FIG. 5, and a 2 bit sequence Lcode 3 relating to the second group 90, as shown in FIG. 6.

If either or both of the second group 88 and third group 90 is active, then additional bytes LIW#2108 provide required information to allow the second group 88 to perform the intended function, and additional bytes LIW#3110 provide required information to allow the third group 90 to perform the intended function.

In the case where the first three bits of byte 0 are not 000, and an LIW#1 instruction or “short” Memory Access operation is to be executed, the extension byte must be used if either or both of the second group 88 and third group 90 is active. If so, the extension byte carries Lcode2 and Lcode3, and additional bytes LIW#2108 and LIW#3110 contain the required information to allow the relevant group to perform the intended function.

The extension byte also carries a 2-bit extension opcode “ex op”, which allows more possible instructions for ALU#0. The extension byte also includes a 1 bit flag, S. If set, the flag S indicates the presence of a shift operation on the ALU first operand. In that case, an additional byte following the extension byte is used to define whether the shift is logical or arithmetic, to the left or right, and how many bits are shifted (4-bit value).

The instruction set architecture supports the use of short constants (which, in this illustrated embodiment, are 4 bits long) and long constants (which, in this illustrated embodiment, are 16 bits long). As shown in FIG. 4, operands are generally 4 bits long, and one of these 4-bit operands normally refers to one of the registers 84, but it can alternatively be used to indicate a 4-bit constant value. Where it is required to use a longer constant value, the operand value “15” is used to direct the instruction decoder 62 to take the value in the 16-bit field 112, which in that case appears at the end of the long instruction word instruction, as a 16-bit constant value. No useful information is therefore stored at the register address “15” (R15). Thus, writing to R15 is used to indicate that an operation result should be discarded.

It can therefore be seen that the encoded instruction word is organized on byte boundaries. It can further be seen from FIGS. 3-6 that an individual LIW instruction can be between 1 byte (the special case where none of the groups is active, and there are no LIW#1, no LIW#2 and no LIW#3 instructions) and 9 bytes in length. The instruction decoder 62 can therefore support any combination of instruction lengths within a single 64-bit instruction word and can tolerate LIW instructions which are contained in successive 64-bit instruction words.

In one embodiment of the invention, the length of any single LIW instruction cannot exceed 8 bytes. However, in other embodiments of the invention, this maximum length can be set to any desired value. This restriction results in a small number of combinations of LIW#1, LIW#2 and LIW#3 instructions which cannot be supported because they exceed this length. These illegal combinations are trapped by the Instruction Decode block 62, resulting in the setting of an Illegal Instruction flag. Preferably, a compiler and assembler operating to support the processor architecture should also intercept disallowed instruction combinations at compile time.

There is one situation where alignment of LIW instructions must be restricted further, and that is in the case of branch destinations. The architecture relies on an instruction being decoded every processor cycle and therefore it is necessary that a branch destination is aligned at the beginning of a 64-bit instruction word. The instruction decoder 62 interprets an all-0 byte instruction (equivalent to “no LIW#1, no LIW#2, no LIW#3”) as a “new line” and will fetch the next 64-bit instruction word. Thus the compiler and assembler can use the “new line” instruction at the end of an instruction sequence immediately prior to a branch destination, in order to ensure 64-bit alignment of the instruction at the branch destination.

The long instruction word format therefore has the property that the length LIW_inst of the long instruction word is independent of the total number of execution units. Rather, it is determined by the maximum number of execution units which can be active in a single processor cycle. In the illustrated embodiment, a maximum of three execution units out of the six available can be active in a single LIW instruction/processor cycle, and the maximum length of a single LIW instruction is limited to 64 bits.

Further, the length LIW_inst of the long instruction word can vary, from one instruction to the next, depending on the number of active execution units within a given cycle. Thus, in many instruction cycles, it is likely that LIW_inst will be less than 64 bits.

Also, multiple instructions can be packed into the 64 bit wide instruction memory 60, usually without the need for alignment to word boundaries, and the instructions can overrun a 64-bit instruction word boundary into the following instruction word.

Taken together, these factors mean that the result is object code which is compact and highly efficient in both high-throughput signal processing applications with multiple parallel operations per cycle and lower-throughput, more complex control operations.

This reduces hardware complexity significantly and, more importantly, reduces the complexity of the compiler/assembler required to support the architecture.

Claims

1. A long instruction word processor comprising a plurality of execution units; and means for decoding instruction words containing sub-instructions, wherein: the said execution units are divided into a plurality of groups with each of the groups containing at least one execution unit, a single instruction word contains a sub-instruction for one of the execution units in each of the groups, and a single instruction word contains a respective sub-instruction for a respective execution unit in at least one of the plurality of groups.

2. The processor of claim 1, further comprising an instruction memory having a fixed width, wherein the a length of an instruction word depends on the number of sub-instructions in the instruction word.

3. The processor of claim 2, wherein the fixed width of the instruction memory defines rows of the instruction memory and the means for decoding instruction words comprises means for decoding instruction words which cross row boundaries of the instruction memory.

4. The processor of claim 1, wherein a first of the plurality of groups comprises a first arithmetic logic unit and a second of the plurality of groups comprises a communications unit, the communications unit being adapted for transferring data from an incoming data bus of the processor and for transferring data to an outgoing data bus of the processor.

5. The processor of claim 4, wherein the communications unit is further adapted for transferring data from internal registers of the processor and for transferring data to the internal registers of said processor.

6. The processor of claim 4, wherein the second of the groups further comprises a second arithmetic logic unit combined with a memory access unit, the memory access unit being adapted for transferring data from an internal memory of said processor and for transferring data to the internal memory of said processor.

7. The processor of claim 4, further comprising a third group of execution units.

8. The processor of claim 7, wherein the third group of execution units comprises a branch unit adapted for performing branch operations.

9. The processor of claim 7, wherein the third group of execution units comprises a multiplier unit.

10. The processor of claim 7, wherein the third group of execution units comprises a multiplier accumulator unit.

11. The processor of claim 7, wherein the third group of execution units comprises a unit for performing signal processing operations.

12. The processor of claim 1, further comprising: an instruction memory having a fixed width wherein: the execution units are divided into a first group and at least a second group, an instruction word, containing sub-instructions for an execution unit in the first group, has a first length, an instruction word, containing sub-instructions for an execution unit in the second group, has a second length greater than said first length, and the instruction memory is adapted to store instruction words of the first length and instruction words of said second length.

13. A processor array, comprising. an array of long instruction word processors; and a plurality of communications buses interconnecting the long instruction word processors wherein each of the long instruction word processors further comprises: a plurality of execution units, including at least an arithmetic logic unit and a communications unit is adapted to transfer data from one of the communications buses or onto one of the communications buses; and means for decoding instruction words containing sub-instructions, wherein: the execution units are divided into a plurality of groups each containing at least one execution unit with the arithmetic logic unit and the communications unit being in different groups, a single instruction word contains a sub-instruction for one of the execution units in each of said groups, and a single instruction word contains a respective sub-instruction for a respective execution unit in at least one of the plurality of groups.