INPUT OPERAND CONTROL IN DATA PROCESSING SYSTEMS

This invention relates to data processing systems. More particularly, this invention relates to data processing systems having a plurality of registers for storing data words to be manipulated by an arithmetic logic unit operating under control of program instruction words.

It is known to provide program instructions that specify whether an operation is to be performed on input operands contained in one register or an input operand contained in two registers but treated as a single input operand. Viewed from one aspect the present invention provides apparatus for data processing, said apparatus comprising: a plurality of registers for storing data words to be manipulated, each of said registers having at least an N-bit capacity; and an arithmetic logic unit responsive program instruction words to perform arithmetic logic operations specified by said program instruction words; wherein said arithmetic logic unit is responsive to at least one program instruction word that includes:

(i) a source register bit field specifying a source register of said plurality of registers storing an input operand data word for said program instruction word;

(ii) an input operand size flag specifying whether said input operand data word has an N-bit size or an (N/2)-bit size; and (iii) when said input size flag specifies an (N/2)-bit size, a high/low location flag indicating in which of high order bit positions of said source register and low order bit positions of said source register said input operand data word is located. There has been a trend in data processing to increase the datapath widths of the systems. Early systems had 8-bit datapaths. These then developed to 16-bit datapaths and it is now common to have 32-bit and 64-bit datapaths. With these increases in datapath widths, the registers within the data processing systems have also increased to have matching widths. The present invention recognises that when the data words to be manipulated are smaller than the datapath widths then using a full register to

store these words is wasteful of the register resources of the device This is particularly the case in load/store architecture machine in which all data to be manipulated must be in a register and for which you wish to reduce the number of times you need to fetch data from the cache or main memory The invention recognises the above consideration and provides the solution of using an input operand size flag and a high/low location flag to indicate the input operand size and in which poπion of the register it is stored In this way, a single register can hold more than one input operand so more efficiently utilising the register resources of the device ard yet those input operands may be separately manipulated The advantages of the present invention are enhanced further when an N-bit data bus links the data storage device to the register In this case, the data bus may be used to transfer two operands at a time, so more efficiently using the bus bandwidth and reducing the possibility of performance bottleneck

In preferred embodiments of the invention said arithmetic logic umt is responsive to at least one parallel operation program instruction word that performs separate arithmetic logic operations upon a first ( /2)-bιt input operand data word and a second (N/2Vbιt input operand data word stored within respective high order bit positions and low order bit positions of a single source register

The provision of parallel operation program instruction worαs allows two independent calculations to be performed by the arithmetic logic umt making full use of us N-bit αatapath capabilities even though the inDUt operands are smaller in size than the maximum data path width. This increases the data processing capabilities of the system considerably whilst not incurring significant extra overhead

One modification that may be made is that in which said arithmetic logic umt has a signal path that functions as a carry chain between bit positions in arithmetic logic operations and, when executing a parallel operation program instruction word said signal path is broken between said first (N/2)-bιt input operand data word and said second (N/2)-bιt input operand data word

Whilst the parallel operation program instructions could take many forms, it is preferred that said parallel operation program instruction word performs the arithmetic logic operation of one of

0) a parallel add in which two parallel (N72)-bιt additions are performed.

j (ii) a parallel subtract in which two parallel (N/2)-bιt subtractions are performed; (iii) a parallel shift in which two parallel (N/2)-bιt shift operations are performed: and (iv) a parallel add/subtract in which an (N/2)-bιt add and an (N/2)-bit subtraction are performed in parallel.

A further refinement of the invention is that when said input size flag specifies an N-bit size, said high/low location flag indicates whether those bits stored in said high order bit positions should be moved to said lower order bit positions and those bits stored said low order bit positions should be moved to said h gh order bit positions prior to use as an N-bit input operand data word

This feamre is particularly useful during transform operations A particularly effective hardware implementation of this functionality comprising at least one multiplexer responsive to said high/low location flag for selecting for supply to the low order (N/2)-bιts of said datapath an (N/2)-bit input operand data word stored in one of high order bit positions of said source register and low order bit positions of said source register.

In order to deal with signed arithmetic without undue complication it is preferred to provide a circuit for sign extending an (N/2)-bιt input operand data word prior to input to said N-bit datapath

Vieweα from another aspect the present invention provides a method of processing data , said method comprising the steps of. storing data words to be manipulated in a plurality of registers, each of said registers having at least an N-bit capacity; and in response program instruction words, performing arithmetic logic operations specified by said program instruction words; wherein at least one program instruction word includes:

(i) a source register bit field specifying a source register of said plurality of registers storing an input operand data word for said program instruction word;

(ii) an input operand size flag specifying whether said input operand data word has an N-bit size or an (N/2Vbιt size; and

(iii) when said input size flag specifies an (N/2)-bit size, a high/low location flag indicating in which of high order bit positions of said source register and low order bit positions of said source register said input operand data word is located.

An embodiment of the invention will now be described, by way of example only, with reference to the accompanying drawings in which

Figure 1 illustrates the high level configuration of a digital signal processing apparatus, Figure 2 illustrates the input buffer of register configuration of a coprocessor;

Figure 3 illustrates the datapath through the coprocessor; Figure 4 illustrates a mutlipiexmg circuit for read high or low order bits from a register;

Figure 5 is a block diagram illustrating register remapping logic used by the coprocessor in preferred embodiments;

Figure 6 illustrates in more detail the register remapping logic shown m Figure 5. and

Figure 7 is a table illustrating a Block Filter Algorithm

The s>stem described below is concerned with digital signal processing (DSP). DSP can take many forms, but may typically be considered to be processing that requires the high speed (real time) processing of large volumes of data This data typically represents some analogue physical signal. A good example of DSP is that used in digital mobile telephones in which radio signals are received and transmitted that require decoding and encoding (typically using convolution, transform and correlation operations) to and from an analogue sound signal Another example is disk driver controllers m which the signals recovered from the disk heads are processed to yield head tracking control.

In the context of the above, there follows a description of a digital signal processing system based upon a microprocessor core (in this case an ARM core from the range of microprocessors designed by Advanced RISC Machines Limited of

Cambridge, United Kingdom) cooperating with a coprocessor. The interface of the microprocessor and the coprocessor and the coprocessor architecture itself are specifically configured to provide DSP functionality. The microprocessor core will be referred to as the ARM and the coprocessor as the Piccolo The ARM and the Piccolo will typically be fabricated as a single integrated circuit that will often include other elements (e.g. on-chip DRAM, ROM. D to A and A to D convenors etc.) as part of an ASIC

Piccolo is an ARM coprocessor, it therefore executes part of the ARM instruction set. The ARM coprocessor instructions allow ARM to transfer data between Piccolo and memory (using Load Coprocessor, LDC and Store Coprocessor, STC. instructions), and to transfer ARM registers to and from Piccolo (using move :o coprocessor. MCR. and move from coprocessor, MRC, instructions). One way of viewing the synergistic interaction of the ARM and Piccolo is that ARM acts as a powerful address generator for Piccolo data, with Piccolo being left free to perform DSP operations requiring the real time handling of large volumes of data to produce corresponding real time results Figure 1 illustrates the ARM 2 and Piccolo 4 with the ARM 2 issuing control signals to the Piccolo 4 to control the transfer of data words to and from Piccolo 4 An instruction cache 6 stores the Piccolo program instruction words that are required by Piccolo 4 A single DRAM memory 8 stores all the data and instruction words required by both the ARM 2 and Piccolo 4 The ARM 2 is responsible for addressing the memory 8 and controlling all data transfers The arrangement with only a single memory 8 and one set of data and address buses is less complex and expensive than the typical DSP approach that requires multiple memories and buses with high bus bandwidths.

Piccolo executes a second instruction stream (the digital signal processing program instruction words) from the instruction cache 6, which controls the Piccolo datapath. These instructions include digital signal processing type operations, ^" or example Multiply-Accumulate, and control flow instructions, for example zero overhead loop instructions. These instructions operate on data which is held in Piccolo registers 10 (see Figure 2) This data was earlier transferred from memory 8 by ihe ARM 2. The instructions are streamed from the instruction cache 6; the instruction cache 6 drives the data bus as a full bus master A small Piccolo instruction cache 6 will be a 4 line, 16 words per line direct mapped cache (64 instructions) In some implementations, it may be worthwhile to make the instruction cache bigger

Thus two tasks are run independently - ARM loading data, and Piccolo processing I: This allows sustained single cycle data processing on 16 bit data.

Piccolo has a data input mechanism (illustrated in Figure 2) that allows the ARM to prefetch sequential data, loading the data before it is required by Piccolo. Piccolo can

access the loaded data m any order, automatically refilling its register as the old data is used for the last time (all instructions have one bit per source operand to indicate that the source register should be refilled) This input mechanism is termed the reorder buffer and comprises an input buffer 12 Every value loaded into Piccolo (via an LDC or MCR see below) carries with it a tag Rn specifying which register the value is destined for The tag Rn is stored alongside the data word in the input buffer When a register is accessed via a register selecting circuit 14 and the instruction specifies the data register is to be refilled, the register is marked as empty by asserting a signal E The register is then automatically refilled by a refill control circuit 16 using the oldest loaded value destined for that register within the input buffer 12 The reorder buffer holds 8 tagged values The input buffer 12 has a form similar to a FIFO except that data words can be extracted from tne centre of the queue after which later stored words will be passed along to fill the space Accordingly, the data words furthest from the input are the oldest and this can be used to decide which data word should be used to refill a register when the input buffer 12 holds two data words with the correct tag Rn

Piccolo outputs data by storing it in an output buffer 18 FIFO) as shown in Figure 3 Data is written to the FIFO sequential!) . and read out to memory 8 m the same order bv ARM The output buffer 18 holds 8 32 bit values Piccolo connects to ARM via the coprocessor interface (CP Control signals of

Figure 1 ) On execution of an ARM coprocessor instruction Piccolo can either execute the instruction, cause the ARM to wait until Piccolo is ready before executing the instruction or refuse to execute the instruction In the last case ARM will take an undefined instruction exception The most common coprocessor instructions mat Piccolo will execute are LDC and STC. which respectively load and store data words to and from the memory 8 via the data bus with ARM generating all addresses It is these instructions which load data into the reorder buffer, and store data from trie output buffer 18 Piccolo will stall the ARM on an LDC if there is not enough room in the input reorder buffer to load in the data and on an STC if there is insufficient data in the output buffer to store l e the data the ARM is expecting is not in tne output buffer 18 Piccolo also executes ARM/Coprocessor register transfers to allow ARM to access Piccolo s special

registers

Piccolo fetches its own instructions from memory to control the Piccolo datapath illustrated in Figure 3 and to transfer data from the reorder buffer to registers and from registers to the output buffer 18 The arithmetic logic unit of the Piccolo that executes these instructions has a multiplier/adder circuit 20 that performs multiplies, adds, subtracts, multiple-accumulates, logical operations, shifts and rotates There is also provided in the datapath an accumulate/decumulate circuit 22 and a scale/saturate circuit 24

The Piccolo instructions are initially loaded from memory into the instruction cache 6, where Piccolo can access them without needing access back to the main memory

Piccolo cannot recover from memory aborts Therefore if Piccolo is used in a Miτual memory system, all Piccolo data must be in physical memory througnout the

Piccolo task This is not a significant limitation given the real time nature of Piccolo tasks e g real time DSP If a memory abort occurs Piccolo will stop and set a flag in a status register S2

Figure 3 shows the overall datapath functionality of Piccolo The register bank 10 uses 3 read ports and 2 write ports One write port (the L port) is used to refill registers from the reorder buffer The output buffer 18 is updated directly from the ALU result bus 26. output from the output buffer 18 is under ARM program control

Tne .ARM coprocessor interface performs LDC (Load Coprocessor; instructions into tne reorder buffer, and STC (Store Coprocessor) instructions from the outDut buffer 18. as well as MCR and MRC (Move ARM register to/from CP register) on the register bank 10 The remaining register ports are used for the ALU Two read ports ( A and B) drive the inputs to the multiplier/adder circuit 20, the C read port is used to aπve the accumulator/decumulator circuit 22 input The remaining write port W is used to return results to the register bank 10

The multiplier 20 performs a 16 x 16 signed or unsigned multiph with an oDtional 48 bit accumulate The sealer unit 24 can provide a 0 to 31 immediate aπtnmetic or logical shift right, followed by an optional saturate The snirter and logical unit 20 can perform either a shift or a logical operation every cycle

Piccolo has 16 general purpose registers named DO-D15 or A0-A3. X0-X3, Y0-Y3, Z0-Z3. The first four registers (A0-A3) are intended as accumulators and are 48 bits wide, the extra 16 bits providing a guard against overflow during many successive calculations. The remaining registers are 32 bits wide. Each of Piccolo's registers can be treated as containing two independent 16 bit values. Bits 0 to 15 contain the low half, bits 16 to 31 contain the high half. Instructions can specify a particular 16 bit half of each register as a source operand, or they may specify the entire 32 bit register.

Piccolo also provides for saturated arithmetic. Variants of the multiply, add and subtract instructions provide a saturated result if the result is greater than the size of the destination register. Where the destination register is a 48 bit accumulator, the value is saturated to 32 bits (i.e. there is no way to saturate a 48 bit valuej. There is no overflow detection on 48 bit registers. This is a reasonable restriction since it would take at least 65536 multiply accumulate instructions to cause an overflow. Each Piccolo register is either marked as "empty" (E flag, see Figure 2) or contains a value (it is not possible to have half of a register empty). Initially, all registers are marked as empty. On each cycle Piccolo attempts with the refill control circuit 16 to fill one of the empty registers by a value from the input reorder buffer. Alternatively if the register is written with a value from the ALU it is no longer marked as "empty". If a register is written from the ALU and at the same time there is a value waiting to be placed in the register from the reorder buffer then the result is undefined. Piccolo's execution unit will stall if a read is made to an empty, register.

The Input Reorder Buffer (ROB) sits between the coprocessor interface and

Piccolo's register bank. Data is loaded into the ROB with ARM coprocessor transfers. The ROB contains a number of 32-bit values, each with a tag indicating the Piccolo register that the value is destined for. The tag also indicates whether the data should be transferred to a whole 32-bit register or just to the bottom 16-bits of a 32-bit register. If the data is destined for a whole register, the bottom 16 bits of the entry will be transferred to the bottom half of the target register and the top 16 bits will be transferred to the top half of the register (sign extended if the target register is a 48-bit accumulator). If the data is destined for just the bottom half of a register (so called ^" Half Register ^' ), the bottom 16 bits will be transferred first.

The register tag always refers to a physical destination register, no register remapping is performed (see below regarding register remapping)

On every cycle Piccolo attempts to transfer a data entry from the ROB to the register bank as follows

-Each entry in the ROB is examined and the tags compared with the registers that .are empty, it is determined whether a transfer can be made from part or all of an entry to a register.

-From the set of entries that can make a transfer, the oldest entry is selected and its data transferred to the register bank

-The tag of this entry is updated to mark the entry as empty If only pan of the_entry was transferred, only the part transferred is marked empty

For example, if the target register is completely empty and the selected ROB entry contains data destined for a full register, the whole 32 bits are transferred and the entry is marked empty If the bottom half of the target register is empty and the ROB entry contains data destined for the bottom half of a register, the bottom 16 bits of the ROB entry are transferred to the bottom half of the target register and the bottom half of the ROB is marked as empty

The high and low 16-bιts of data in any entry can be transrerred independently If no entry contains data that can be transferred to the register bank, no transfer is made that cycle The table below describes all possible combinations of target ROB entrv and target register status.

98/12627

1 1

To summarise, the two halves of a register may be refilled independently from the ROB. The data in the ROB is either marked as destined for a whole register or as two 16-bit values destined for the bottom half of a register.

Data is loaded into the ROB using ARM coprocessor instructions. How the data

98/12627

12 is marked in the ROB depends on which ARM coprocessor instruction was used to perform the transfer. The following ARM instructions are available for filling the ROB with data:

LDP{<cond>}<16/32> <dest>, [Rn]{! }. <size>

LDP{<cond>}<16/32>W <dest>, <wrap>. [Rn] {!] , #<sιze>

LDP{<cond>} 16U <bank>, [Rn] { ' }

MPR{<cond>} <dest>, Rn

MRP{<cond>} <dest>, Rn

The following ARM instruction is provided for configuring the ROB

LDP A bank lιsτ>

The first tnree are assembled as LDCs, MPR and MRP as MCRs, LDPA is assembled as a CDP instruction.

In the above <dest> stands for a Piccolo register (A0-Z3). Rn for an ARM register, <sιze> for a constant number of bytes which must be a non zero multiple of

4 and <wrap> for a constant (1 ,2,4,8). Fields surrounded by [ } are optional. For a transfer to be able to fit into the Reorder Buffer. <sιze> must be at most 32. In many

8/12627

13 circumstances <sιze> will be smaller than this limit to avoid deadlock The <16/32> field indicates whether the data being loaded should be treated as 16-bit data, and endianess specific action taken (see below), or as 32-bit data

Note 1 : In the following text, when referring to LDP or LDP W this refers to both the 16-bit and 32-bit variants of the instructions.

Note2 A 'word' is a 32-bit chunk from memory, which may consist of two 16-bit data items or one 32-bit data item.

The LDP instruction transfers a number of data items, marking them as destined for a full register The instruction will load <sιze>/4 words from address Rn m memory, inserting them into the ROB The number of words that can be transferred is limited by the following

-The quantity <sιze> must be a non-zero multiple of 4,

-<sιze> must be less than or equal to the size of the ROB for a particular implementation (8 words in the first version, and guaranteed to be no less than this in future versions)

The first data item transferred will be tagged as destined for <dest>. the second as destined for <dest>- and so on (with wrapping from Z3 to AO) If the ' is specified then the register Rn is incremented by <sιze> afterwards If the LDP 16 variant is used, endian specific action is performed on the two

16-bit halfwords forming the 32-bit data items as they are returned from the memory system See below for more details on Big Endian and Little Endian Support

The LDPW instruction transfers a number of data items to a set of registers The first data item transferred is tagged as destined for <dest>, the next for <dest>+l, etc When <wrap> transfers have occurred, the next item transferred is tagged as destined for <dest>, and so on The <wrap> quantity is specified in halfword quantities

For LDPW, the following restrictions apply

-The quantity <sιze> must be a non-zero multiple of 4

98/12627 n- ι«w

14

-<size> must be less than or equal to the size of the ROB for a particular implementation (8 words in the first version, and guaranteed to be no less than this in future versions);

-<dest> may be one of (AO, XO, Y0, Z0};

-<wrap> may be one of {2,4,8} halfwords for LDP32W and one of { 1 ,2,4,8} halfwords for LDP16W;

-The quantity <size> must be greater than 2*<wrap>, otherwise no wrapping occurs and the LDP instruction shall be used instead.

For example, the instruction,

LDP32W XO, 2, [RO]!, #8

will load two words into the ROB, marking them as destined for the full register XO. RO will be incremented by 8. The instruction,

LDP32W XO, 4, [RO], #16

will load four words into the ROB. marking them as destined for XO, X I. XO. XI (in that order). RO will not be affected. For LDP16W, <wrap> may be specified as 1,2,4 or 8. The wrap of 1 will cause all data to be tagged as destined for the bottom half of the destination register <dest>.l. This is the 'Half Register ^' case.

For example the instruction,

LDP16W XO, 1, [R0]!. #8

will load two words into the ROB, marking them as 16-bit data destined for X0.1. RO

8/12627

15 will be incremented by 8 The instruction.

LDP16W XO, 4, [RO], #16

will behave m a similar fashion to the LDP32W examples, except for the fact that endian specific action may be performed on the data as it is returned from memory

All unused encodings of the LDP instruction may be reserved for future expansion

The LDP16U instruction is provided to support the efficient transfer of non- word aligned 16-bit data. LDP16U support is provided for registers D-J to D15 (the X, Y and Z banks) The LDP16U instruction will transfer one 32-bit word of data (containing two 16-bit data items) from memory into Piccolo Piccolo will discard the bottom 16 bits of this data and store the top 16 bits in a holding register There is a holding register for the X, Y and Z banks Once the holding register of a bank is primed, the behaviour of LDP{W} instructions is modified if the data is destined for a register in that bank The data loaded into the ROB is formed by the concatenation of the holding register and the bottom 16 bits of data being transferred by the LDP instruction The upper 16 bits of data being transferred is put into the holding register.

entry <- data 1 J holdιng_register holdιng_regιster <- data.h

This mode of operation is persistent until it is turned off a LDPA instruction The holding register does not record the destination register tag or size. These characteristics are obtained from the instruction that provides the next value of data l

Endian specific behaviour may always occur on the data returned by the memorv system There is no non 16-bit equivalent to LDP16U since it is assumed that all 32-bit data items will be word aligned in memorv

98/12627

16

The LDPA instruction is used to switch off the unaligned mode of operation initiated by a LDP16U instruction. The unaligned mode may be turned off independently on banks X. Y, Z. For example the instruction,

LDPA {X. Y}

will turn off the unaligned mode on banks X and Y. Data in the holding registers of these anks will be discarded.

Executing an LDPA on a bank which is not in unaligned mode is allowed, and will leave that bank in aligned mode.

The MPR instruction places the contents of ARM register Rn into theHOB, destined for Piccolo register <dest>. The destination register <dest> may be any full register in the range A0-Z3. For example the instruction,

MPR XO, R3

will transfer the contents of R3 into the ROB, marking the data as destined for the full register XO.

No endianess specific behaviour occurs to the data as it is transferred from ARM to Piccolo since the ARM is internally little endian.

The MPRW instruction places the contents of ARM register Rn into the ROB, marking it as two 16-bit data items destined for the 16-bit Piccolo register <dest ,l. The restrictions on <dest> are the same as those for the LDPW instructions (i.e. AO.XOΥO.ZO). For example the instruction,

MPRW XO, R3

will transfer the contents of R3 into the ROB, marking the data as 2 16-bit quantities destined for X0.1. It should be noted that as for the LDP16W case with a wrap of 1 , only the bottom half of a 32-bit register can be targeted.

As with MPR no endianess specific operations are applied to the data. LDP is encoded as:

31 30292S 272625242 2221 20 19 13 17 16 15 14 13 12 II 10 9 8 7 6 5 4 3 2 I

where PICCOLOl is Piccolo's firs: coprocessor number (currently 8). The N bit selects between LDP32 (1) and LDP16 (0).

LDPW is encoded as:

3! 30292S 272625 423222! 20 19 18 17 16 15 14 13 12 1! 10 9 8 7 6 5 4

where DEST is 0-3 for destination register A0,XO,Y0,Z0 and WRAP is 0-3 for wrap values 1.2.4,8. PICCOL02 is Piccolo's second coprocessor number (currently 9). The N bit selects between LDP32 (1) and LDP16 (0).

LDP16U is encoded as:

31 30 29 23 27 26 25 24 23 22 21 20 19 13 17 16 15 14 13 12 1 1 10 9 S 7 6 5 4 3 2 1

COND 1 10 w Rn DES P1CCOL02 00000001

where DEST is 1 -3 for the destination bank X. Y, Z.

LDPA is encoded as:

31 3029 28 272625 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2

where BANK[3 :0] is used to turn off the unaligned mode on a per bank basis. If BANK[l ] is set, unaligned mode on bank X is turned off. BAN [2] and BANI ] turn off unaligned mode on banks Y and Z if set, respectively. N.B. This is a CDP operation.

MPR is encoded as:

3! 30 29 23 27 25 25 24 23 22 21 20 19 13 17 16 15 14 13 12 11 10 9 S 7 6 5 4 3 2 1 0

MPRW is encoded as:

31 3029 23 272625 24 0 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

where DEST is 1 -3 for the destination register X0,Y0,Z0.

The output FIFO can hold up to eight 32-bit values. These are transferred frcm Piccolo by using one of the following (ARM) opcodes:

STP {<cond>] <l6/32> [ n]{ ! }, #<size>

MRP Rn

The first saves <size>/4 words from the output FIFO to the address given by the AFΛl register Rn, indexing Rn if the ! is present. To prevent deadlock, <size> must not be

greater than the size of the output FIFO (8 entries in the this implementation). If the STP16 variant is used, endian specific behaviour may occur to the data returned from the memory system.

The MRP instruction removes one word from the output FIFO and places it in ARM register Rn. As with MPR no endian specific operations are applied to the data

The ARM encoding for STP is:

3 1 30 29 23 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 1 1 10 9 8 7 6 5 4 3 2 1 0

COND 10 P | U | N | W | 0 Rn 0000 I PICCOLO l SIZE/4

where N selects between STP 2 (1) and STP 16 (0). For the definitions of the P, U and

W b s, refer to an ARM data sheet.

The ARM encoding for MRP is:

31 30292827262524232221 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 1 0

The Piccolo instruction set assumes little endian operation internally For example when accessing a 32-bit register as 16 bits halves, the lower half is assumed to occupy bits 15 to 0. Piccolo may be operating in a system with big endian memory or peripherals and must therefore take care to load 16-bit packed data in the correct manner

Piccolo (i.e. the DSP adapted coprocessor), like the ARM (e.g. the ARM7 microprocessors produced by Advanced RISC Machines Limited of Cambndge, United Kingdom), has a 'BIGEND' configuration pin which the programmer can control, perhaps with a programmable peripheral. Piccolo uses this pm to configure the input reorder buffer and output FIFO

When the ARM loads packed 16-bit data into the reorder buffer it must indicate this by using the 16-bit form of the LDP instruction. This information is combined with the state of the 'BIGEND' configuration input to place data into the holding latches and reorder buffer in the appropriate order. In particular when in big endian mode the holding register stores the bottom 16 bits of the loaded word, and is paired up with the top 16 bits of the next load. The holding register contents always end up in the bottom 16 bits of the word transferred into the reorder buffer.

The output FIFO may contain either packed 16-bit or 32-bit data. The programmer must use the correct form of the STP instruction so that Piccolo can ensure that the 16-bit data is provided on the correct halves of the data bus. When configured as big endian the top and bottom 16-bit halves are swapped when the 15- bit forms of STP are used.

Piccoio has 4 private registers which can only be accessed from the ARM. They are called S0-S2. They can only be accessed with MRC and MCR instructions. The oDcoαes are:

MPSR Sn, R

MRPS Rm, Sn

These opcodes transfer a 32-bit value between ARM register Rm and private register Sn. They are encoded in ARM as a coprocessor register transfer:

31 30 29 23 27 2ό 25 24 20 19 13 17 16 15 14 13 12 11 10 9 8 7 6 5 1 0

where L is 0 for the MPSR and 1 for the MRPS.

Register SO contains the Piccolo unique ID and revision code

31 30292327262524232221 20 19 13 17 16 15 14 13 12 11 10 9 7 6 5 4 3

Bιts[3 0] contain the revision number for the processor

Bιts[15 4] contain a 3 digit pan number in binary coαed decimal forma; 0x500 for Piccolo

Bιts[23.16] contain the architecture version 0x00 = Version 1

Bιts[31 :24] contain the ASCII code of an implementers trademark- 0x41 = A = .ARM

Ltd

Register SI is the Piccolo status register

31 30292327262524232221 20 19 18 17 16 15 14 13 12 II 10 9

Primary condition code flags (N,Z,C,V)

Secondary condition code flags (SN,SZ,SC,SV)

E bit Piccolo has been disabled by the ARM and has halted

U bit Piccolo encountered an UNDEFINED instruction and has halted

B bit Piccolo encountered a BREAKPOINT and has halted

H bit: Piccolo encountered a HALT instruction and has halted.

A bit. Piccolo suffered a memory abort (load, store or Piccolo instruction) and has halted.

D bit: Piccolo has detected a deadlock condition and has halted (see below)

Register S2 is the Piccolo program counter:

31 30292827262524232221 20 19 18 16 15 14 13 12 1 1 10 9 8 6 5 4 2 1 0

Prozrim Counter 0 0

Writing to the program counter will start Piccolo executing a program at that address (leaving halted state if it is halted). On reset the program counter is undefined, since Piccolo is always started by writing to the program counter

During execution Piccolo monitors the execution of instructions and the status of the coprocessor interface. If it detects that

-Piccolo has stalled waiting for either a register to be refilled or the output FIFO to have an available entry.

The coprocessor interface is busy-waiting, because of insufficient space in the ROB or insufficient items in the output FIFO.

If both of these conditions are detected Piccolo sets the D-bit in its status register, halts and rejects the ARM coprocessor instruction, causing .ARM to take the undefined instruction trap.

This detection of deadlock conditions allows a system to be constructed which can at Jeast warn the programmer that the condition has occurred and repoπ the exact

point of failure, by reading the ARM and Piccolo program counters and registers. It should be stressed that deadlock can only happen due to an incorrect program or perhaps another pan of the system corrupting Piccolo's state. Deadlock can not occur due to data starvation or 'overload'.

There are several operations available that may be used to control Piccolo from the ARM, these are provided by CDP instructions. These CDP instructions will only be accepted when the ARM is in a privileged state. If this is not the case Piccolo will reject the CDP instruction resulting in the ARM taking the undefined instruction trap. The following operations are available:

Reset

Enter State Access Mode

Enabie

Disable

Piccoio may be reset in sof ware bv usir.2 the PRESET instruction.

PRESET Clear Piccolo ^' s state

This instructtcr. :s encoded as:

il 302923272625 23 222! 20 19 IS 17 16 15 14 13 12 11 10 9 3 7 6 5 4

When this instruction is executed the following occurs:

-All registers are marked as empty (ready for refill).

-Input ROB is cleared.

-Output FIFO is cleared.

-Loop counters are reset.

-Piccolo is put into halted state (and H bit of S2 will be set).

Executing the PRESET instruction may take several cycles to complete_(2-3 for this embodiment). Wnilst it is executing, following ARM coprocessor instructions to be executed on Piccolo will be busv waited.

In state access mode, Piccolo's state may be saved and restored using STC and LDC instructions (see the below regarding accessing Piccolo state from .ARM). To enter state access mode, the PST.ATE instruction must first be executed:

PSTATE; Enter State Access Mode

This instruction is encoded as:

31 30292S 2726252423 2221 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2

COND 1110 0001 0000 0000 PICCOLOl 000 0 0000

When executed, the PSTATE instruction will:

-Halt Piccolo (if it is not already halted), setting the E bit in Piccolo's Status Register

-Configure Piccolo into its State Access Mode

Executing the PSTATE instruction may take several cycles to complete, as Piccolo's instruction pipeline must drain before it can halt. Whilst it is executing, following ARM coprocessor instructions to be executed on Piccolo will be busy waited.

The PEN.ABLE and PDISABLE instructions are used for fast context switching. When Piccolo is disabled, only private registers 0 and 1 (the ID and Status registers) are accessible, and only then from a privileged mode. Access to any other state, or any access from user moαe will cause an ARM undefined instruction exception Disabling Piccolo causes it to halt execution. When Piccoio has halted execution, it will acknowledge the fact by setting the E bit m the status register

Piccolo is enabled by executing the PENABLE instruction:

PEN ABLE, Enable Piccolo

this instruction is encodeα as

il 3029 2827262524232221 20 19 IS 17 16 15 14 13 12 10 9 8 7 6 5 4 3

Piccolo is disabled by executing the PDISABLE instruction:

PDIS.ABLE ; Disable Piccolo

This instruction is encoded as

31 30292S 2726252423 2221 20 19 13 17 16 15 14 13 12 11 10 9 8 7 6 5 I 0

When this instruction is executed, the following occurs

-Piccolo's instruction pipeline will drain.

-Piccolo will halt and the H bit in the Status register set.

The Piccolo instruction cache holds the Piccolo instructions which control the

Piccolo datapath. If present it is guaranteed to hold at least 64 instructions, staπing on a 16 word boundary. The following ARM opcode assembles into an MCR. Its action is to force the cache to fetch a line of (16) instructions staπing at the specified address (which must be on a 16- word boundary). This fetch occurs even if the cache alreaαv holds data related to this address.

PMIR Rm

Piccolo must be halted before a PMIR can be performed

The MCR encoding of this opcode is:

31 30292327262524232221 20 19 IS 17 16 15 14 13 12 11 10 9 S 7 6 5 4 3

This section discusses the Piccolo instruction set which controls the Piccolo data path. Each instruction is 32 bits long. The instructions are read from the Piccolo instruction cache.

Der oding the instruction set is quite straight forward. The top 6 bits (26 to 31) give a major opcode, with bits 22 to 25 providing a minor opcode for a few specific instructions. Bits shaded in grey are currently unused and reserved for expansioriXthey must contain the indicated value at present).

There are eleven major instruction classes. This does not fully correspond to the major opcode filed in the instruction, for ease of decoding some sub-classes.

2 2 2 2 2 2 2 2 2 1 1

The instructions in the above table have the following names:

Standard Data Operation

Logical Operation

Conditional Add/Subtract

Undefined

Shifts

Select

Undefined

Parallel Select

Multiply Accumulate

Undefined

Multiply Double

Undefined

Move Signed Immediate

Undefined

Repeat

Repeat

Register List Operations

Branch

Renamins Parameter Move

Halt/Break

The format for each class of instructions is described in detail in the following sections. The source and destination operand fields are common to most instructions and described in detail in separate sections, as is the register re-mapping.

Most instructions require two source operands; Source 1 and Source 2. Some exceptions are saturating absolute.

The Source 1 (SRCI) operand has the following 7 bit format:

18 17 16 15 14 12

The elements of the field have the following meaning:

-Size - indicates the size of operand to read (l=32-bit, 0=16-bit) -Refill - specifies that the register should be marked as empty after being read and can be refilled from the ROB.

-Register Number - encodes which of the 16 32-bit registers to read.

-Hi/Lo - For 16-bit reads indicates which half of the 32-bit register to read. For 32-bit operands, when set indicates that the two 16-bit halves of the register should be interchanged.

The register size is specified in the assembler by adding a suffix to the register number: .1 for the low 16 bits, .h for the high 16 bits or .x for 32 bits with the upper and lower sixteen bits interchanged.

The general Source 2 (SRC2) has one of the following three 12 bit formats:

10 0

Figure 4 illustrates a multiplexer arrangement responsive to the Hi Lo bit and Size bit to switch appropriate halves of the selected register to the Piccolo datapath. If the Size bit indicates 16 bits, then a sign extending circuit pads the high order bits of the datapath with 0s or Is as appropriate.

The first encoding specifies the source as being a register, the fields having the same encoding as the SRCI specifier. The SCALE field specifies a scale to be applied to the result of the ALU.

The 8-bit immediate with rotate encoding allows the generation of a 32-bit immediate which is expressible by an 8-bit value and 2-bit rotate. The following table

_>-> shows the immediate values that can be generated from the 8 -bit value XY:

The 6-bit Immediate encoding allows the use of a 6-bit unsigned immediate (range 0 to 63), together with a scale applied to the output of the ALU.

The general Source 2 encoding is common to most instruction variants. There are some exceptions to this rule which support a limited subset of the Source 2 encoding or modify it slightly:

-Select Instructions.

-Shift Instructions.

-Parallel Operations.

-Multiply Accumulate Instructions.

-Multiply Double Instructions.

Select instructions only support an operand which is a register or a 6-bit unsigned immediate. The scale is not available as these bits are used by the condition field of the instruction.

1 1 10 7 0

SRC2 SEL

Shift instructions only support an operand which is a 16-bit register or a 5-bit unsigned immediate between 1 and 31. No scale of the result is available.

1 1 10 9 8 7 6 5 4

SRC2 SHIFT

In the case of parallel operations, if a register is specified as the source of the operand, a 32-bit read must be performed. The immediate encoding has slightly different meaning for the parallel operations. This allows an immediate to be duplicated onto both 16-bit halves of a 32-bit operand. A slightly restricted range of scales are available for parallel operations.

1 ^' 10 8

SRC2 PARALLEL

If the 6-bit immediate is used then it is always duplicated onto both halves of the 32-bit quantity. If the 8-bit immediate is used it is duplicated only if the rotate indicates that the 8-bit immediate should be rotated onto the top half of the 32-bit quantity:

No scale is available for parallel select operations; the scale field shall be set to 0 for these instructions.

The multiply accumulate instructions do not allow an 8-bit rotated immediate to be specified. Bit 10 of the field is used to partly specify which accumulator to use. Source 2 is implied as a 16-bit operand.

1 1 10 9 8 1 0

SRC2 MULA

Multiply double instructions do not allow the use of a constant. Only a 16-bit register can be specified. Bit 10 of the field is used to partly specify which accumulator to use.

11 10 8 0

SRC2 MULD

Some instructions aiways imply a 32-bit operation (e g ADDADD). and in these cases the size bit shall be set to 1, with the Hi Lo bit used to optionally swap the two 16-bit halves of the 32-bit operand Some instructions alwavs imply a 16- bit operation (e g MUL) and the size bit should be set to 0 The Hi/Lo bit then selects which half of the register is used (it is assumed that the missing size bit is clear) Multiply-accumlulate instructions allow independent specification of the source accumulator and destination registers For these instructions the Size bits aie used to indicate the source accumulator, and the size bits are implied by the instruction type as 0

When a 16-bit value is read (via the A or B busses) it is automatically sign extended to a 32-bit quantity If a 48 bit register is read (via the A or B busses.), onl the bottom 32 bits appear on the bus Hence in all cases source 1 ana source 2 are convened to 32-bit values Only accumulate instructions using bus C can access the full 48 bits of an accumulator register

If the refill bit is set, the register is marked as empty after use ana will be refilled from the ROB by the usual refill mechanism (see the section on tne ROB ι Piccolo will not stall unless the register is used again as a source operanα before the refill has taken place The minimum number of cycles before the refilled data is valid (best case - the data is waiting at the head of the ROB) will be eitnε: 1 or 2 Hence it is advisable not to use the refilled data on the instruction following the refill request If use of the operand on the next two instructions can be avoided it should be, since this will prevent performance loss on deeper pipeline implementations

The refill bit is specified in the assembler by suffixing the register numbei wun a ^{' Λ} ' The section of the register marked as empty derjends on the register operand The two halves of each register may be marked for refill independently (for example XO 1 ^Λ will mark only the bottom half of XO for refill. X0 ^Λ will mark the whole of XO for refill) When the top ^' half (bits 47 16) of a 48-bit register ire

refilled, the 16-bits of data is written to bits 31 : 16 and is sign extended up to bit 47.

If an attempt is made to refill the same register twice (eg ADD X1.X0 ^Λ ,X0 ^Λ ), then only one refill takes place. The assembler should only allow the syntax ADD XI , X0,X0 ^Λ .

If a register read is attempted before that register has been refilled, Piccolo stalls waiting for the register to be refilled. If a register is marked for refill, and the register is then updated before the refilled value is read, the result is UNPREDICTABLE (for example ADD XO, X0 ^Λ , XI is unpredictable since it marks XO for refill and then refills it by placing the sum of XO and XI into ϊϊT.

The 4-bit scale field encodes fourteen scale types:

-ASR #0, 1. 2, 3, 4, 6, 8, 10 -ASR #12 to 16 -LSL #1 Parallel Max/Min instructions do not provide a scale, and therefore the six bit constant variant of source 2 is unused (Set to 0 by assembler).

Within a REPEAT instruction register re-mapping is supported, allowing a REPEAT to access a moving ^" window' of registers without unrolling the loop. This is described in more detail in below.

Destination operands have the following 7 bit format:

Λ 11 20 19

SD HL DEST

There are ten variants of this basic encoding

Assembler Mnemonic 2D s 20 19

^" No

(No

The register number (Dx) indicates which of the 16 registers is being addressed. The Hi/Lo bit and the Size bit work together to address each 32-bit register as a pair of 16-bit registers. The Size bit defines how the appropriate flags, as defined in the instruction type, will be set. irrespective of whether a result is written to the register bank and or output FIFO. This allows the construction of compares and similar instructions. The add with accumulate class of instruction must write back the result to a register.

The following table shows the behaviour of each encoding:

In all cases the result of any operation prior to writing back to a register or inserting into the output FIFO is a 48 bit quantity. There are two cases:

If the write is of 16-bits the 48 bit quantity is reduced to a 16-bit quantity by selecting the bottom 16 bits [15:0]. If the instruction saturates then the value will be saturated into the range -2 ^Λ 15 to 2 ^Λ 15-1. The 16-bit value is then written back to the indicated register and. if the Write FIFO bit is set, to the output FIFO. If it is written to the output FIFO then it is held until the next 16-bit value is written when the values are paired up and placed into the output FIFO as a single 32-bit value.

For 32-bit writes the 48 bit quantity is reduced to a 32-bit quantity by selecting the bottom 32 bits [31 :0].

For both 32-bit and 48-bit writes, if the instruction saturates the 48-bιrvalue will be convened to a 32-bit value in the range -2 ^Λ 31-1 to 2 ^Λ 31. Following the saturation:

-If writeback to an accumulator is performed, the full 48 bits will be written. -If writeback to a 32-bit register is performed, bits [31 :0] are written.

-If writeback to the output FIFO is indicated, again bits [31 :0] will be written.

The destination size is specified in the assembler by a .1 or .h after me register number. If no register writeback is performed then the register number is unimportant, so omit the destination register to indicate no write to a register or use ^Λ to indicate a write only to the output FIFO. For example, SUB , X0, Y0 is equivalent to CMP

X0. Y0 and ADD ^Λ . X0, Y0 places the value of XO^YO into the output FIFO

If there is no room in the output FIFO for a value. Piccolo stalls waiting for space to become available. If a 16-bit value is written out. for example ADD X0.h ^A . XI , X2. then the value is latched until a second 16-bit value is written. The two values are then combined and placed into the output FIFO as a 32-bit number. The first 16-bit value written always appears in the lower half of the 32-bit word. Data entered into the output FIFO is marked as either 16 or 32-bit data, to allow endianess to be corrected

on big endian systems.

If a 32-bit value is written between two 16-bit writes then the action is undefined.

Within a REPEAT instruction register re-mapping is supported, allowing a REPEAT to access a moving 'window ^* of registers without unrolling the loop. This is described in more detail below.

In preferred embodiments of the present invention, the REPEAT instruction provides a mechanism to modify the way in which register operands are specified within a loop. Under this mechanism, the registers to be accessed are determined by a function of the register operand in the instruction and an offset into the register bank. The offset is changed in a programmable manner, preferably at the end όTeach instruction loop. The mechanism may operate independently on registers residing in the X. Y and Z banks. In preferred embodiments, this facility is not available for registers in the A bank.

The notion of a logical and physical register can be used. The instruction operands are logical register references, and these are then mapped to physical register references identifying specific Piccolo registers 10. All operations, including refilling, operate on the physical register. The register remapping only occurs on the Piccolo instruction stream side - data loaded into Piccolo is always destined for a physical register, and no remapping is performed.

The remapping mechanism will be discussed further with reference to Figure 5. which is a block diagram illustrating a number of the internal components of the Piccolo coprocessor 4. Data items retrieved by the ARM core 2 from memory are placed in the reorder buffer 12, and the Piccolo registers 10 are refilled from the reorder buffer 12 in the manner described earlier with reference to Figure 2. Piccolo instructions stored in the cache 6 are passed to an instruction decoder 50 within Piccolo 4, where they are decoded prior to being passed to the Piccolo processor core 54. The Piccolo processor core 54 includes the multiplier/adder circuit 20. the

accumulate/decumulate circuit 22, and the scale/saturate circuit 24 discussed earlier with reference to Figure 3

If the instruction decoder 50 is handling instructions forming part of an instruction loop identified by a REPEAT instruction, and the REPEAT instruction has indicated that remapping of a number of registers should take place, then the register remapping logic 52 is employed to perform the necessary remapping The register remapping logic 52 can be considered as being part of the instruction decoder 50, although it will be apparent to those skilled in the art that the register remapping logic 52 may be provided as a completely separate entity to the instruction decoder 50

An instruction will typically include one or more operands identify mg registers containing the data items required by the instruction For example, a typical instruction may include two source operands and one destination operand, idenufving ~wo registers containing data items required by the instruction, and a register in to which the result of the instruction should be placed The register remapping logic 52 receives the operands of an instruction from the instruction decoder 50 these operands identifying logical register references Based on the logical register references, the register remapping logic will determine whether remapping should or should noi be applied, and will then apply a remapping to phvsical register references as required If it is determined that remapping should not be applied, the logical register references are provided as the physical register references The preferred manner in which the remapping is performed will be discussed in more detail later

Each output physical register reference from the register remapping logic is passed to the Piccolo processor core 54, such that the processor core can then apply the instruction to the data item in the particular register 10 identified bv the physical register reference

The remapping mechanism of the preferred embodiment allows each bank of registers to be split into two sections, namely a section within which registers mav be

remapped, and a section in which registers retain their original register references without remapping In preferred embodiments, the remapped section starts at the bottom of the register bank being remapped

A number of parameters are employed by the remapping mechanism, and these parameters will be discussed in detail with reference to Figure 6. which is a block diagram illustrating how the various parameters are used by the register remapping logic 52. It should be noted that these parameters are given values that are relative to a point within the bank being remapped, this point being, for example, the bottom of the bank

The register remapping logic 52 can be considered as comprising twojnain logical blocks, namely the Remap block 56 and the Base Update block 58 The register remapping logic 52 employs a base pointer that provides an offset value to be added to the logical register reference, this base pointer value being provided to the remap block 56 by base update block 58

A BASESTART signal can be used to define the initial value of the base pointer, this for example typically being zero, although some other value may be specified This BASESTART signal is passed to multiplexor 60 within the Base Update block 58 During the first iteration of the instruction loop, the BASESTART signal is passed by the multiplexor 60 to the storage element 66, whereas for subsequent iterations of the loop, the next base pointer value is supplied by the multiplexor 60 to the storage element 66

The output of the storage element 66 is passed as the current base pointer value to the ReMap logic 56, and is also passed to one of the inputs of an adder 62 within the Base Update logic 58 The adder 62 also receives a BASEINC signal that provides a base increment value The adder 62 is arranged to increment the current base pointer value supplied by storage element 66 by the BASEINC value, and to pass the result to the modulo circuit 64

The modulo circuit also receives a BASEWRAP value, and compares this value to the output base pointer signal from the adder 62. If the incremented base pointer value equals or exceeds the BASEWRAP value, the new base pointer is wrapped round to a new offset value. The output of the modulo circuit 64 is then the next base pointer value to be stored in storage element 66. This output is provided to the multiplexor 60. and from there to the storage element 66.

However, this next base pointer value cannot be stored in the storage element 66 until a BASEUPDATE signal is received by the storage element 66 from the loop hardware managing the REPEAT instruction. The BASEUPDATE signal will be produced periodically by the loop hardware, for example each time the instruction loop is to be repeated. When a BASEUPDATE signal is received by the storage element 66. the storage element will overwrite the previous base pointer value with the next base pointer value provided by the multiplexor 60. In this manner, the base pointer value supplied to the ReMap logic 58 will change to the new base pointer value.

The physical register to be accessed inside a remapped section of a register bank is determined by the addition of a logical register reference contained within an operand of an instruction, and the base pointer value provided by the base update logic 58. This addition is performed by adder 68 and the output is passed to modulo circuit 70. In preferred embodiments, the modulo circuit 70 also receives a register wrap vaiue. and if the output signal from the adder 68 (the addition of the logical register reference and the base pointer value) exceeds the register wrap value, the result will wτap through to the bottom of the remapped region. The output of the modulo circuit 70 is then provided to multiplexor 72.

A REGCOUNT value is provided to logic 74 within Remap block 56, identifying the number of registers within a bank which are to be remapped. The logic 74 compares this REGCOUNT value with the logical register reference, and passes a control signal to multiplexor 72 dependent on the result of that comparison. The multiplexor 72 receives as its two inputs the logical register reference and the output

from modulo circuit 70 (the remapped register reference). In preferred embodiments of the present invention, if the logical register reference is less than the REGCOUNT value, then the iogic 74 instructs the multiplexor 72 to output the remapped register reference as the Physical Register Reference. If. however, the logical register reference is greater than or equal to the REGCOUNT value, then the logic 74 instructs the multiplexor 72 to output the logical register reference directly as the physical register reference.

As previously mentioned, in preferred embodiments, it is the REPEAT instruction which invokes the remapping mechanism As will be discussed m more detail later, REPEAT instructions provide four zero cycle loops in hardware These hardware loops are illustrated in Figure 5 as part of the instruction decoder 50. Each time the instruction decoder 50 requests an instruction from cache 6, the cache returns that instruction to the instruction decoder, whereupon the instruction decoder determines whether the returned instruction is a REPEAT instruction. If so, one of the hardware loops is configured to handle that REPEAT instruction

Each repeat instruction specifies the number of instructions in the loop and the numoer of times to go around the loop (which is either a constant or read from a Piccolo register) Two opcodes REPEAT and NEXT are provided for defining a hardware loop, the NEXT opcode being used merely as a delimiter and not being assembled as an instruction The REPEAT goes at the start of the loop, and NEXT delimits the end of the loop, allowing the assembler to calculate the number of instructions in the loop body In preferred embodiments, the REPEAT instruction can include remapping parameters such as the REGCOUNT, BASEINC, BASEWRAP and REGWRAP parameters to be employed by the register remapping logic 52.

A number of registers can be provided to store remapping parameters used by the register remapping logic Within these registers, a number of sets of predefined remapping parameters can be provided, whilst some registers are left for the storage of user defined remapping parameters If the remapping parameters specified with the

REPEAT instruction are equal to one of the sets of predefined remapping parameters, then the appropriate REPEAT encoding is used, this encoding causing a multiplexor or the like to provide the appropriate remapping parameters from the registers directly to the register remapping iogic. If, on the other hand, the remapping parameters are not the same as any of the sets of predefined remapping parameters, then the assembler will generate a Remapping Parameter Move Instruction (RMOV) which allows the configuration of the user defined register remapping parameters, the RMOV instruction being followed by the REPEAT instruction. Preferably, the user defined remapping parameters would be placed by the RMOV instruction in the registers left aside for storing such user defined remapping parameters, and the multiplexor would then be programmed to pass the contents of those registers to the register remapping logic.

In the preferred embodiments, the REGCOUNT. BASEINC, BASEWRAP and REGWRAP parameters take one of the values identified in the following chart:

Returning to Figure 6, an example of how the various parameters are used by the remap block 56 is as follows (in this example, the logical and physical register values are relative to the particular bank):

if (Logical Register < REGCOUNT)

Physical Register = (Logical Register + Base) MOD REGCOUNT

else

Physical Register = Logical Register

end if

At the end of the loop, before the next iteration of the loop begins, the following update to the base pointer is performed by the base update logic 58

Base = (Base -t- BASEINC) MOD BASEWRAP

At the end of a remapping loop, the register remapping will be switched off and all registers will then be accessed as physical registers In preferred emboαiments, only one remapping REPEAT will be active at any one time Loops ma\ still be nested, but only one may update the remapping -v ariables at any particular time However, it will be appreciated that, if desired, remapping repeats could be nested

To illustrate the benefits achieved with regards to code densitv as a result of employing the remapping mechanism according to the preferred embodiment of the present invention, a typical block filter algorithm will now be discussed The principles of the block filter algorithm will first be discussed with reference to Figure 7 As illustrated in Figure 7 accumulator register AO is arranged to accumulate ^r .he results of a number of multiplication operations, the multiplication operations being the multiplication of coefficient cO by data item dO. the multiplication of coefficient cl by data item dl, the multiplication of coefficient c2 by data item d2. etc Register

Al accumulates the results of a similar set of multiDlication operations, but this time the set of coefficients have been shifted such that cO is now multiplied by dl, cl is now multiplied by d2. c2 is now multiDlied bv d3, etc Likewise, register A2 accumulates the results of multiplying the data values by the coefficient shiited

another step to the right, such that cO is multiplied by d2, cl is multiplied by d3, c2 is multiplied by d4, etc. This shift, multiply, and accumulate process is then repeated with the result being placed in register A3.

If register remapping in accordance with the preferred embodiment of the present invention is not employed, then the following instruction loop will be required to perform the block filter instruction:

;car- with 4 new data values

ZERO {AO-A3} ; Zero the accumulators

REPEAT Zl ; Zl= (numoer of coeffs/4)

co the next four coefficients, on the first t me around.

aO += d0*c0+dl*clχd2*c2+d3*c3

al += dl*c0+d2*cl+d3*c2+d4*c3

a2 += d2*c0+d3*cl+d4*c2+d5*c3

a3 += d3*c0+d4*cl+d5*c2+d6*c3

MU A AO, X0.1", Y0.1 , AO , aO += dO*cO, and load d4

MULA Al, XO.h , YO.1 , Al ; ^s al += dl*cO

MULA A2, Xl.l , Y0.1 , A2 ; a2 += d2*c0

MULA A3, XI.h , Y0.1 ^* , A3 ; a3 += d3*c0, and load C4

MULA AO, XO.h ^* , YO.h , AO ; aO += dl*cl, and load d5

MULA Al, Xl.l , YO.h , Al ; al += d2*cl

MULA A2, Xl.h , YO.h , A2 ; a2 += d3*Cl

MULA A3, X0.1 , YO.h ^* , A3 ; a3 += d4*cl, and load c5

MULA AO, Xl.l ^* , Yl.l , AO ; aO += d2*c2, and load d6

MULA Al, Xl.h , Yl.l , Al ; al += d3*c2

MULA A2, X0.1 , Yl.l , A2 ; a2 += d4*c2

MULA A3, XO.h , Yl.l ^* , A3 ; a3 += dS*c2, and load c6

MULA AO, Xl.h", Yl.h , AO ; aO += d3*c2, and load d?

MULA Al, X0.1 , Yl.h , Al ; al += d4*c3

MULA A2, XO.h , Yl.h , A2 ; a2 += d5*c3

MULA A3, Xl.l , Yl.h ^* , A3 ; a3 += d6*c3, and load c7

NEXT

In this example, the data values are placed in the X bank of registers and the coefficient values are placed in the Y bank of registers. As a first step, the four

accumulator registers AO, Al, A2, and A3 are set to zero. Once the accumulator registers have been reset, an instruction loop is then entered, which is delimited by the REPEAT and NEXT instructions. The value Zl identifies the number of times that the instruction loop should be repeated, and for the reasons that will be discussed later, this will actually be equal to the number of coefficients (cO, cl, c2. etc.) divided by

4.

The instruction loop comprises 16 multiply accumulate instructions (MULA), which, after the first iteration through the loop, will result in the registers AO. Al, A2, A3 including the result of the calculations shown in the above code between the REPEAT and the first MULA instruction. To illustrate how the multiply accumulate instructions operate, we will consider the first four MULA instructions. The first instruction multiplies the data value within the first, or lower. 16 bits of the X bank register zero with the lower 16 bits within Y bank register zero, and adds the result to the accumulator register AO. At the same time the lower 16 bits of the X bank register zero are marked by a refill bit, this indicating that that part of the register can now be refilled with a new data vaiue. It is marked in this way, since as will be apparent from Figure 7, once data item dO has been multiplied by the coefficient cO (this being represented by the first MULA instruction), then dO is no longer required for the rest of the block filter instruction and so can be replaced by a new data value.

The second MULA instruction then multiplies the second, or higher 16 bits of the X bank register zero with the lower 16 bits of the Y bank register zero (this representing the multiplication dl x cO shown in Figure 7). Similarly, the third and fourth MULA instructions represent the multiplications d2 x cO, and d3 x cO, respectively. As will be apparent from Figure 7, once these four calculations have been performed, coefficient CO is no longer required and so the register Y0.1 is marked by a refill bit to enable it to be overwritten with another coefficient (c4).

The next four MULA instructions represent the calculations dlxc l . d2xcl, d3xcl . and d4xcl. respectively. Once the calculation dlxcl has been performed, the

register XO.h is marked by a refill bit since dl is no longer required. Similarly, once all four calculations have been performed, the register YO.h is marked for refilling, since the coefficient cl is no longer needed. Similarly, the next four MULA instructions correspond to the calculations d2xc2, d3xc2, d4xc2, and d5xc2. whilst the final four calculations correspond to the calculations d3xc3, d4xc3. d5xc3, and d6xc3.

Since, m the above described embodiment, registers are not remappable, each multiplication operation has to be reproduced explicitly with the specific register required being designated in the operands. Once the sixteen MULA instructions have been performed, the instruction loop can be repeated for coefficients c4 to c7 and data items d4 to d 10 Also, because the loop acts on four coefficient values per iteration, then the number of coefficient values must be a multiple of four and the computation Zl = no. of coeffs/4 must be calculated.

By employing the remapping mechanism in accordance with the preferred embodiment of the present invention, the instruction loop can be dramatically reduced, such that it now only includes 4 multiply accumulate instructions, rather than the 16 multiply accumulate instructions that were otherwise required Using the remapping mechanism, the code can now be written as follows

start witr. 4 new data values

ZERO ( A0 -A3 » , Zero the accumulators

REPEAT Zl , X-+ n4 w4 r4 , Y++ n4 w4 r4 , Zl= (number c f coef f icients

Remapping s applied to the X and Y ba s

Four 16 c:: registers in these banks are remappec

_; The base pointer for both banks is incremented by one on each

,- iteration of the loop.

_; The base pointer wraps when it reaches the fourth register m the

; bank.

MULA AO, X0.1 ^* , Y0.1 , AO ; aO += dO*cO, and load d4

MULA Al, XO.h, YO.1 , Al ; al += dl*cO

MULA A2, Xl.l, YO.1 , A2 ; a2 += d2*c0

MULA A3, Xl.h, Y0.1 ^* , A3 ; a3 += d3*c0, and load c4

NEXT ; go round loop and advance remapping

As before, the first step is to set the four accumulator registers A0-A3 to zero. Then, the instruction loop is entered, delimited by the REPEAT and NEXT opcodes. The REPEAT instruction has a number of parameters associated therewith, which are as follows:

X— : indicates that BASEINC is '1' for the X Bank of registers

n4 : indicates that REGCOUNT is '4' and hence the first four X Bank registers

X0.1 to Xl .h are to be remapped

w4 : indicates that BASEWRAP is '4' for the X Bank of registers

r4 : indicates that REGWRAP is '4' for the X Bank of registers

Y-H- : indicates that BASEINC is ' 1 ' for the Y Bank of registers

n4 : indicates that REGCOUNT is '4' and hence the first four Y Bank registers

Y0.1 to Yl .h are to be remapped

w4 : indicates that BASEWRAP is '4' for the Y Bank of registers

r4 : indicates that REGWRAP is '4' for the Y Bank of registers

It should also be noted that now the value Zl is equal to the number of coefficients, rather than being equal to the number of coefficients/4 as in the prior art example.

For the first iteration of the instruction loop, the base pointer value is zero, and so there is no remapping. However, next time the loop is executed, the base pointer value will be ^" 1 ' for both the X and Y banks, and so the operands will be mapped as follows:

X0.1 becomes XO.h

XO.h becomes Xl .l

Xl.l becomes Xl.h

Xl .h becomes X0.1 (since BASEWRAP is '4')

Y0.1 becomes YO.h

YO.h becomes Yl.l

Yl .l becomes Yl .h

Yl.h becomes Y0.1 (since BASEWRAP is '4')

Hence, it can be seen that on the second iteration, the four MULA instructions actually perform the calculations indicated by the fifth to eight MULA instructions in the example discussed earlier that does not include the remapping of the present invention. Similarly, the third and fourth iterations through the loop perform the calculations formerly performed by the ninth to twelfth, and thirteenth to sixteenth MULA instructions of the prior art code.

Hence, it can be seen that the above code performs exactly the same block filter algorithm as the prior art code, but improves code density within the loop body by a factor of four, since only four instructions need to be provided rather than the sixteen required by the prior art.

By employing the register remapping technique in accordance with preferred embodiments of the present invention, the following benefits can be realised:

1. It improves code density;

2. It can in certain situations hide the latency from marking a register being as empty to that register being refilled by Piccolo's reorder buffer. This could be achieved by unrolling loops, at the cost of increased code size;

3. It enables a variable number of registers to be accessed - by varying the number of loop iterations performed the number of registers accessed may be varied; and

4 It can ease algorithm development For suitable algorithms, the programmer can produce a piece of code for the nth stage of the algorithm, then use register remapping to apply the formula to a sliding set of data

It will be apparent that certain changes can be made to the above described register remapping mechanism without departing from the scope of the present invention For example, it is possible for the bank of registers 10 to provide more physical registers than can be specified by the programmer in an instruction operand

Whilst these extra registers cannot be accessed directly, the register remapping mechanism can make these registers available For example consider the example discussed earlier where the X bank of registers has four 32 bit registers available to the programmer, and hence eight 16 bit registers can be specified by logical register references It is possible for the X bank of registers to actualh consist of, for example, six 32 bit registers, in which case there will be four additional 16 bit registers not directly accessible to the programmer However, these extra four registers can be made available by the remapping mechanism therebv providing additional registers for the storage of data items

The following assembler syntax may will be used

» means iogical shift right or shift left if the shift operand is negative (see <lscale> below )

-» means arithmetic shift right or shift left if the shift operand is negative

(see <scale> below )

RORmeans Rotate Right

SAT(aιmeans the saturated value of a (saturated to 16 or 32 bits αepenαing

on the size of the destination register) Specificallv to saturate to lo

bits, any value greater than +0x7fff is replaced by +0x7fff and

any value less than -0x8000 is replaced by -0x8000. Saturation

to 32 bits is similar with extremes +0x7fffffff and -0x80000000. If the

destination register is 48 bits the saturation is still at 32 bits.

Source operand 1 can be one of the following formats:

<srcl>will be used a shorthand for [Rn!Rn.l|Rn.h|Rn.x][ ^Λ ]. In other words all 7 bits of the source specifier are valid and the register is read_as a 32-bit value (optionally swapped) or a 16-bit value sign extended. For an accumulator only the bottom 32 bits are read. The ^Λ specifies register refill.

<srcl_16>is short for [Rn.l|Rn.h][ ^Λ ]. Only 16-bit values can be read.

<srcl_32>is short for [RnjRn.x][ ^Λ ]. Only a 32-bit value can be read, with the upper and lower halves optionally swapped.

Source operand 2 can be one of the following formats:

<src2>will be a shorthand for three options: a source register of the form [Rn!Rn.l|Rn.h|Rn.x][ ^Λ ], plus a scale (<scale>) of the final result. an optionally shifted eight bit constant (<immed_8>), but no scale of the final result. a six bit constant (<immed_6>), plus a scale (<scale>) of the final result. <src2_maxmin> is the same as <src2> but a scale is not permitted. <src2_shift> shift instructions provide a limited subset of <src2>. See above. for details.

<src2_par> as for <src2_shift>

For instructions which specify a third operand:

<acc> is short for any of the four accumulator registers [A0'A1!A2|A3].

All 48 bits are read. No refill can be specified. The destination register has the format:

<dest> which is short for [RnjRn.l!Rn.hj.I| ][ ^Λ ] With no "." extension the full register is written (48 bits in the case of an accumulator). In the case where no write back to the register is required, the register used is unimportant. The assembler supports the omission-of a destination register to indicate that write back is not required or ".1" to indicate that no writeback is required but flags should be set as though the result is a 16-bit quantity ^Λ denotes that the value is written to the output FIFO

<scale> represents a number of arithmetic scales There are fourteen available scales:

ASR #0, 1. 2, 3. 4, 6, 8, 10 ASR #12 to 16 LSL #1 <ιmmed 8> stands for a unsigned 8-bit immediate value. This consists of a bvte rotated left bv a shift of 0, 8, 16 or 24. Hence values

OxYZOOOOOO.

OxOOYZOOOO, OxOOOOYZOO and OxOOOOOOYZ can be encoded for any YZ. The rotate is encoded as a 2 bit quantity. <ιmm_6> Stands for an unsigned 6-bit immediate. <PARAMS> is used to specify register re-mapping and has the following format: <BANK><BASEINC> n<RENUMBER> w<BASE\VRAP>

<BANK> can be [XjYjZ] <BASEINC> can be [-H-J+ 1 |+2|+4] <RENUMBER> can be [O!2| ;8] <BASEWRAP> can be [2J4J8]

The expression <cond> is shorthand for any one of the following condition codes. Note that the encoding is slightly different from the ARM since the unsigned

LS and HI codes have been replaced by more useful signed overflow/underflow tests.

The V and N flags are set differently on Piccolo than on the ARM so the translation from condition testing to flag checking is not the same as the ARM either.

Last result was zero.

Last result was non zero.

Used after a shift MAX operation.

Last result was negative.

Last result was positive Signed overflow/saturation on last result No overflow/saturation on last result Overflow positive on last result. Overflow negative on last result

Since Piccolo deals with signed quantities, the unsigned LS and HI conditions have been dropped and replaced by VP and VN which describe the direction of any overflow. Since the result of the ALU is 48 bits wide, MI and LT now perform the

same function, similarly PL and GE. This leaves 3 slots for future expansion.

All operations are signed unless otherwise indicated.

The primary and secondary condition codes each consist of:

N - negative.

Z - zero.

C - carry/unsigned overflow.

V - signed overflow.

Arithmetic instructions can be divided into two types; parallel and 'full width ^" . The 'full width' instructions only set the primary flags, whereas the parallel operators set the primary and secondary flags based on the upper and lower 16-bit halves of the result.

The N. Z and V flags are calculated based on the full ALU result, after the scale has been applied but prior to being written to the destination. .An ASR will always reduce the number of bits required to store the result, but an ASL would increase it. To avoid this Piccolo truncates the 48-bit result when an ASL scale is applied, to limit the number of bits over which zero detect and overflow must carried out.

The N flag is calculated presuming signed arithmetic is being carried out. This is because when overflow occurs, the most significant bit of the result is either the C flag or the N flag, depending on whether the input operands are signed or unsigned.

The V flag indicates if any loss of precision occurs as a result of writing the result to the selected destination. If no write-back is selected a 'size' is still implied, and the overflow flag is set correctly. Overflow can occur when:

-Writing to a 16-bit register when the result is not in the range -2 ^Λ 15 to 2 ^Λ 15-1. -Writing to a 32-bit register when the result is not in the range -2 ^Λ 31 to 2 ^Λ 31-1.

Parallel add subtract instructions set the N. Z and V flags independently on the upper and lower halves of the result.

When writing to an accumulator the V flag is set as if writing to a 32-bit register. This is to allow saturating instructions to use accumulators as 32-bit registers.

The saturating absolute instruction (SABS) also sets the overflow flag if the absolute value of the input operand would not fit in designated destination. The Carry flag is set by add and subtract instructions and is used as a 'binary' flag by the MAX/MIN, SABS and CLB instructions. All other instructions, including multiply operations preserve the Carry flag(s).

For add and subtract operations the Carry is that which is generated by either bit 31 or bit 15 or the result, based on whether the destination is 32 or 16-bits wide. The standard arithmetic instructions can be divided up into a number types, depending on how the flags are set:

In the case of Add and Subtract instructions, if the N bit is set. then all flags are preserved. If the N bit is not set then the flags are updated as follows:

Z is set if the full 48 bit result was 0. N is set if the full 48 bit result had bit 47 set (was negative).

V is set if either:

The destination register is 16-bit and the signed result will not fit into a 16-bit register (not in the range -2 ^Λ 15<=x<2 ^Λ 15)

The destination register is a 32/48 bit register and the signed result will not fit into 32 bits.

If <dest> is a 32 or 48 bit register then the C flag is set if there is a carry out of bit 31 when summing <srcl> and <src2> or if no borrow occurred from bit 31 when subtracting <src2> from <srcl> (the same carry value you would expect on the

ARM). If <dest> is a 16-bit register then the C flag is set if there is a carry out of bit 15 of the sum.

The secondary flags (SZ, SN, SV, SC) are preserved.

In the case of instructions which either carry out a multiplication or accumulate from a 48-bit register.

Z is set Λ me full 48 bit result was 0 N ^* is set if the full 48 bit result haα bit 47 set (was negative) V is set if eπner (1) the destination register is 16-bit and the signeα result will not fit into a 16-Dit register (not in the range -2 ^Λ 15<=x<2 ^Λ 15) or (2) the αestmation register is a 32'48 bit register and the signed result will not fit into

C is preserved The seconcary flags (SZ, SN, SV , SC) are preserved

ihe otπe: instructions, including logical operations, parallel aααs and subtracts max ar.α mm, sr fts etc are covered below

Tne Acα and Subtract instructions add or suDtract two registers, scaie me rest..: and then store back to a register The operands are treatea as signec v lues Flag -ideating far tne non-saturating variants is optional and mav oe surjpresseα c aocer.cins an X :o tne end of the instruction

31 30292S 2" 25252-i 23222! 20 19 IS 17 16 15 14 13 12 11 10 9 6 5 4

OPC DES1 S I SRC I SRC2

OPC specifies "le voe of instruction

Action (OPO

100N0 dest = (srcl + src2) (-» scale) (, N)

1 10N0 dest = (srcl - src2) (-» scale) (, N)

10001 dest = SAT((srcl + src2) (-» scale))

1 1001 dest = SAT((srcl - src2) (-» scale)) OH IO dest = (src2 - srcl) (-» scale)

01 111 dest = SAT((src2 - srcl) (-» scale))

101N0 dest = (srcl + src2 + Carry) (-» scale) (, N)

11 1N0 dest = (srcl - src2 + Carry - 1) (-» scale) (, N)

Mnemonics:

100N0 ADD{N} <dest>, <srcl>, <src2> >.<scale>}

1 10N0 SUB{N} <dest>, <srcl>, <src2> {,<scale>}

10001 SADD <dest>, <srcl>, <src2> {.<sca!e>}

1 1001 SSUB <dest>, <srcl>, <src2> (,<scale>) OH IO RSB <dest>, <srcl>, <src2> {,<scale>}

01 111 SRSB <dest>, <srcl>, <src2> {,<scale>}

101N0 ADC{N} <dest>, <srcl>, <src2> {,<scale>}

1 11N0 SBC(N} <dest>, <srcl>, <src2> {,<scale>}

The assembler supports the following opcodes

CMP <srcl>, <src2> CMN <srcl>, <src2>

CMP is a subtract which sets the flags with the register write disabled. CMN is an add which sets the flags with register write disabled.

Flags:

These have been discussed above.

Reasons for inclusion"

ADC is useful for inserting carry into the bottom of a register following a shift'MAX/MT operation It is also used to do a 32/32 bit divide It also provides for extendeα precision adds. The addition of an N bit gives finer control of the flags, in particular the carry This enables a 32 '32 bit division a: 2 cycles oer bit

Saturated adds ana subtracts are neeαed for G 729 etc

Incrementing' αecrement g counters RSB is useful for calculating sr fts is a common operation) A saturated RSB is neeαeα for sarurateα negation t useα :r. G 729)

Add/subtract accumulate instructions perform acdition inc suctractioπ w m accumulation an scaling/saturation Unlike tne multiply accumulate instructions tr.e accumulator number cannot be specified independently of the destination register Tne oottom two oits of the destination register give ;r.e numoer. ac , o: e -3 :.". accumulator to accumulate into. Hence .ADDA X0.X1.X1.A0 and ADD A A3.X 1 ,X2.A3 are valid, but ADDA XI ,X 1 ,X2,A0 is not With this ciass of instruction, tne result must be written baci to a register - the no writeback encoαings of :ne destination field are not allowed

3, 30 29 2S 27 26 25 24 23 22 2! 20 19 13 17 16 15 14 13 2 11 10 9 I 0

OPC specifies the type of instruction. In the following ace is (DEST[l :0j). The Sa bit indicates saturation.

Action (OPC):

0 dest = (SAT}(acc + (srcl + src2)) {-» scale} 1 dest = {SAT}(acc + (srcl - src2)) {-» scale}

Mnemonics: S}ADDA <dest>, <srcl>, <src2>, <acc> {,<scale>} {S}SUBA <dest>, <srcl>, <src2>, <acc> {,<scale>}—

An S before the command indicates saturation.

Flags: See above.

Reasons for inclusion:

The ADDA (add accumulate) instruction is useful for summing two words of an array of integers with an accumulator (for instance to find their average) per cycle. The

SUBA (subtract accumulate) instruction is useful in calculating the sum of the differences (for correlation); it subtracts two separate values and adds the difference to a third register.

Addition with rounding can be done by using <dest> different from <acc>. For example, X0=(X1+X2+16384)»15 can be done in one cycle by keeping 16384 in AO.

Addition with a rounding constant can be done by ADDA X0,X1,#16384,A0.

For a bit exact implementation of: sum of ((a_i * b_i)»k) (quite common - used in TrueSpeech)

the standard Piccolo code would be: MUL tl , a_0, b_0, ASR k ADD ans, ans, tl MUL ι2, a_l, b_l , ASR#k ADD ans, ans, t2

There are two problems with this code; it is too long and the adds are not to -S-bu precision so guard bits can't be used. A better solution is to use .ADDA

MUL tl , a_0, b_0, ASR k MUL , a_l , b_l . ASR k

.A DA ans, tl , t2, ans

This gives a 25% speed increase and retains -8-bit accuracy

Add/Subtract m Parallel instructions perform addition and subtraction on two signed 16-bit quantities held in pairs in 32-bit registers The primary condition code flags are set from the result of the most significant 16 bus, the secondary' flags are updated from tne least significant half. Only 32-bit registers can oe specified as tne source for these instructions, although the values can be halfworc swapped. The individual halves of each register are treated as signed values. The calculations and sca ng are done with no loss of precision. Kence ADD. DD XO, XI . X2, ASR=1 will produce the correct averages in the upper and lower haives of XO. Optional sararatioπ is provided for each instruction for which the Sa bit must be set.

31 30 29 2S 2 ^" 25 25 24 23 222! 20 19 IS 17 16 15 14 13 12 !! 10 9 S ! 0

OPC defines the operation.

Action (OPC):

000 dest.h = (srcl .h + src2.h) -» {scale}, dest.l = (srcl .l + src2.1) -» {scale}

001 dest.h = (srcl .h + src2.h) -» {scale}, dest.l = (srcl .l - src2.1) -» {scale}

100 dest.h = (srcl.h - src2.h) -» {scale}, dest.l = (srcl.l + src2.1) -» {scale}

101 dest.h = (srcl.h - src2.h) -» {scale} , dest.l = (srcl .l - src2.1) -» {scale}

Each sum/difference is independently saturated if the Sa bit is set.

<dest>, <srcl_32>, <src2_32> {,<scale>}

<dest>, <srcl_32>, <src2_32> {,<scale>}

<dest>, <srcl_32>, <src2_32> {,<scale>} <dest>, <srcl_32>, <src2_32> {,<scale>}

An S before the command indicates saturation.

The assembler also supports

CM CMN <dest>, <srcl_32>, <src2_32> {,<scale>} CMNCMP <dest>, <srcl_32>, <src2_32> {,<scale>} CMPCMN <dest>, <srcl_32>, <src2_32> {,<scale>} CMPCMP <dest>, <srcl_32>. <src2_32> {,<scale>}

senerated bv the standard instructions with no write-back.

68

Flags

C is set if there is a carry out of bit 15 wnen adding the two upoer sixteen bit halves Z is set if the sum of the upper sixteen bit halves is 0 N is set if the sum of tne upper sixteen bit halves is negative

V is set if the signed 17 bit sum of the upper sixteen bit halves will not fit into 16 bits (post scale) SZ, SN, SV and SC are set similarlv for the lower 16-bu halves

Reason for inclusion

The parallel .Ada and Suotract instructions are useful for Derformmg operations on complex numoers neid in a single 32 -bit register They are used m tne FFT kernel It is aiso useful for simple addition subtraction of vectors of 16-bu data allowing two elements to oe processed per c cιe

The Brancn i conα ona!) instruction allows conditional changes m control flow PiccoiO may ta e tnree cvcles to execute a taken brancn

3C 2 ^Q 2S 2" 262: 13 222 20 19 IS 12 1 10 9

H I I I OC 000 IMMEDUTE 16 COND

Action

Brancn ov offset if <cond> hoiαs according to tne primarv flags

The offset is a signed 16-bιt numcer of worαs At the moment tne range

of the offset is restricted to -32768 to -32767 words

69

The address calculation performed is

target address = branch instruction address - ^ά - OFFSE 1

Mnemonics:

3<cor.ά> < .s.i.ιac:or. Iabel>

Flass:

Unaffected.

Reasons for inclusion:

Highly usefui in most routines.

Conditional Add or Subtract instructions conditionally add or subtract s:c2 to srcl .

31 30 29 2S 27 25 25 24 22 21 20 19 IS 17 16 15 U 13 1 1 10 9 S 1 0

OFC specities the type of instruction.

Action (OPC):

0 if (carry set) temp=src 1 -src2 else temp=srcl+src2

dest = temp {-» scale}

1 if (carry set) temp=srcl-src2 else temp=srcl+src2

dest = temp {-» scale} BUT if scale is a shift left

then the new value of carry (from srcl-src2 or

srcl+src2) is shifted into the bottom.

Mnemonics:

0 CAS <dest>, <srcl>, <src2> {,<scaie>}

1 CASC <dest>, <srcl>, <src2> {,<scale>}

Flags:

See above.

Reasons for inclusion:

The Conditional Add or Subtract instruction enables efficient divide code to be constructed.

Example 1 : Divide the 32-bit unsigned value in XO by the 16-bit unsigned value in XI (with the assumption that XO < (Xl«16) and Xl .h=0).

71

LSL XI, XI, #15 ; shift up divisor

SUB XI, XI, #0 ; set the carry flag

REPEAT #16

CASC XO, XO, X1 , LSL#1

NEXT

At the end of the loop XO.l holds the quotient of the divide. The remainder can be recovered from XO.h depending on the value of carry.

Example 2: Divide the 32-bit positive value in XO by the 32-bit positive value in XI, with early termination.

At the end, X2 holds the quotient and the remainder can be recovered from X0. The Count Leading Bits instruction allows data to be normalised.

31 3029232726252-1 23 2221 20 19 IS 17 16 15 U 13 12 I 1 10 9 8 7 6

Action: dest is set to the number of places the value in src l must be shifted left in order for bit 31 to differ from bit 30. Tnis is a value in the range 0-30 except in the special cases where src ! is. either - i or 0 where 1 is returned.

Mnemonic:

CLB <dest>, <src l>

^" lags:

Σ is set if the result is zero. is cleared.

C is set if src l is either - 1 or 0.

V is preserved.

Reasons for inclusion:

Ster needed for normalisation.

Halt and Breakpoint instructions are provided for stopping Piccolo exec.

;i 3029232" 252524 23 222! 20 19 IS 17 16 15 14 13 12 10 9 S 1 0

1 m n ! 1 OP 00000000000000000000000

OPC soeciπes tne tvpe of instruction

Action (OPC)

0 Piccolo execution is stopnea and tne Halt bit is set in tne Piccolo status register

1 Piccolo execution is stopped, the Breax bit is set in the Piccolo status register and the .ARM is interrupted to say mat a breakpoint has been reached

Mnemonics

0 HALT

1 BREAK

Fiass

Unaffected

Logical Operation instructions perform a logical operation on a 32 or 16-bu register The ocerands are treated as unsisned values

31 302923 27 2625 24 23 21 20 19 18 17 16 15 14 13 12 11 10 9 8 6 5 1 0

OPC encodes the logical operation to perform.

Action (OPC):

00 est = (srcl & src2) {-» scale}

01 dest = (sr l ', src2) {-» scale}

10 est = (srci & ~src2) {-» scale}

11 dest = (sr l ^Λ src2) {-» scale}

Mnemonics:

00 AND <dest , <srcl>, <src2> {,<scale>} 01 OFil <desι , <sr l>. <src2> {,<scale>} 10 3IC <dest>, <srcl>, <src2> {,<scale>}

11 FOR 'αest>, <srcl>, <src2> (,<scaie>)

The assembler suppoπs the following opcodes

TST <srcl>, <src2> TEQ <src!>, <src2>

n

TST :s an AND with the register write disabled TEQ is an EOR with the registe write disabled.

Flags-

Z is set if the result is all zeros

N. C, V are preserved

SZ, SN, SC. SV are preserved

Reasons for inclusion.

Speecr. compression algorithms use packed bitfields for encoding information. Bit ask-ng instructions help for extracting packing these fields.

I D Max and Min Operation instructions perform maximum and minimum operations.

31 3029 23 27 2525 24 23 2221 20 19 18 17 16 15 14 13 12 11 10 9 6 5 4 I 0

20

OPC specifies the type of instruction. Action (OPC).

0 dest = (srcl <= src2) ? srcl : src2

1 dest = (srcl > src2) ? srcl : src2

Mnemonics:

0 MIN <dest>, <srcl>, <src2>

1 MAX <dest>, <srcl>, <src2>

Flass:

Z is set if the result is zero

N is set if the result is negative

C For Max: C is set if src2>=srcl (dest=srcl case)

For Min: C is set if src2>=srcl (dest=src2 case) V preserved.

Reasons for inclusion:

In order to find the strength of a signal many algorithms scan a sample to find the minimum/maximum of the absolute value of the samples. The MAX and MIN operations are invaluable for this. Depending on whether you wish to find the first or last maximum in the signal the operands src 1 and src2 can be swapped around.

MAX XO, XO, #0 will convert XO to a positive number with clipping below.

MIN XO. XO, #255 will clip XO above. This is useful for graphics processing.

Max and Min Operations in Parallel instructions perform maximum, and

minimum operations on parallel 16-bit data.

3029 2S 2726 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 6 5 2 1 0

OPC specifies the type of instruction.

Action (OPC;

.st.l = (srcl.l <= src2.1) ? srcl.l src \

dest.h = (srcl.h <= src2.h) sr: :l.h sr i.n

dest.l = (srcl.l > src2.1) ? srcl.l : src2.1

dest.h = (srcl.h > src2.h) ? srcl.h : src2.h

Mnemonics:

0 MΓNMTN <dest>, <srcl>, <src2>

MAXMAX <dest>, <srcl>, <src2>

Flaas

is set if the upper 16 bits of the result is zero

N is set if the upper 16 bits of the result is negative

For Max. C is set if src2 h>=srcl h (dest=srcl case)

For Min. C is set if src2.h>=src 1.h (dest=src2 case)

reserved

SZ,SN.SC.SV are set similarly for the lower 16-bit halves

Reasons for inclusion

Move Long Immediate Operation instructions allow a register to oe set ^* o r signed 16-bu. sign extended vaiue Two of these instructions can set a 32-b,: register to any value (by accessing the high and low half in sequence) For moves between registers see me select operations.

30 29 2S 2= 25 24 23 22 2' 20 19 18 i 15 14 13 12 1 1 10 9 S 1 0

Mnemonics

The assembler will provide a non-interlocking NOP operation using is MOV instruction, i.e. NOP is equivalent to MOV , =0.

Fiags

Flags are unaffected.

Reasons for inclusion:

Initialising registers/counters

Multiply Accumulate Operation instructions perform signed multiplication w. accu uiation or ae-accumuiation, scaling and saturation

11 3029 2S 2725 25 24 23 22 2! 20 19 IS 17 16 15 14 10 9 8 2 1 0

ι he field OPC :pec:f:ss me rype of instruction

Action (OPC)

00 csst ^■ (ace - (srcl * src2)) (-» scale}

01 aest = (ace - (srcl * src2)) {-» scale]

In each case the result is saturated before being written to tr. destination if the Sa bit is set.

Mnemonics:

00 {S}MULA <dest>, <srcl_16>, <src2_16>, <acc> { ,<scaie>}

01 (SJ-MULS <dest>, <src l_16>, <src2_16>, <acc> {,<scale> }

.An S before the command indicates saturation

nags

See section above

Reasons for inclusion:

A one cycle sustained MULA is required for FIR. code MULS is used in me FFT butter: ^" . A MUL is aiso useful for multiply with rounding For exarr.r.e A0=(X0 " X 1 — 163 S-)» 15 can be done in once cycle by holding 16384 m another accumulator ( l for example). Different <dest and <acc> is also required for tne FFT kernel

Multiply Double Operation instructions perform signed multiplication, doubling the result prior to accumulation or de-accumulation, scahng and saturation.

30 29 2S 27 26 25 24 23 22 21 20 19 I S 16 15 14 13 12 1 1 10 9 S

OPC specifies me tvoe of instruction

Action lOPQ

cest = SAT(facc - SATι ^' 2 * src l * src2)) {-» scale] )

1 dest = SAT((acc - SAT(2 * srcl * src2)) {-» scale})

Mnemonics:

0 SMLDA <dest>, <srcl_16>, <src2_16>, <acc> {,<scale>}

1 SMLDS <dest , <srcl_16>, <src2_16>, <acc> {,<scale>} Flags:

See section above.

Reasons for inclusion:

The MLD instruction is required for G.729 and other algorithms which use fractional arithmetic. Most DSPs provide a fractional mode which enables a left shift of one bit at the output of the multiplier, prior to accumulation or writeback. Suppoπing this as a specific instruction provides more programming flexibility. The name equivalents for some of the G series basic operations are:

L_msu => SMLDS L_mac => SMLDA

These make use of the saturation of the multiplier when left shifting by one bit. If a sequence of fractional multiply-accumulates is required, with no loss of precision, MULA can be used, with the sum maintained in 33.14 format. A left shift and saturate can be used at the end to convert to 1.15 format, if required.

Multiply Operation instructions perform signed multiplication, and optional scaling/saturation. The source registers (16-bit only) are treated as signed numbers.

302928 252524 23 2221 20 19 16 15 14 13 12 11 10 9 S

OPC specifies the type of instruction.

action (OPC).

0 ces: = ( _^ srcl " src2) {-» scale}

1 dest = SAT((src l * src2) {-» scale] )

Mnemonics'

MUL <dest>, <src l 16>, <src2> { ,<scale>]

SMUL <dest>, <srcl 16>, <src2> <scale>'

Flags

iee section aDOve

Reasons for inclusion'

Signed and saturated multiplies are required by many processes

Register List Operations are used to perform actions on a set of registers. The Empty and Zero instructions are provided for reseuing a selection of registers prior to. or in between routines The Output instruction is provided to store tne contents of a list of registers to the output FIFO

302923 27 25 25 24 23 22 21 20 19 IS 17 16 15 14 13 12 11 10 9 1 0

OPC specifies the type of instruction.

Action (OPC)

000 for ( ^" κ=0; <16 k— ) '.f bit k of the register list is set then register k is marxed as being empty.

001 for (k=0; k<16 →) if bit k of the register list is set then register k is set

:o contain 0. 010 Undefined

01 1 Undefined 100 for (k=0; k<16 k— ) if bit k of the register list is set then

(register k -» scale) is wrmen to the output FIFO. 101 for (k=0; k<16, k— ) if bit k of the register list is set then

(register k -» scale) is written to the output FIFO and register k is marked as being empty. if bit k of the register list is set then

S.AT(register k -» scale) is wriπen to the output FIFO

1 1 1 for (k=0; k<16 k-÷) if bit k of the register list is set then SAT(register k -» scale) is wrmen to the output FIFO and register k is marked as being empty.

98/12627

84

Flags:

Unaffected

Examples:

EMPTY {A0. Al, X0-X3}

ZERO {Y0-Y3}

OUTPUT {X0-Y1 } ^A

The assembler will also support the syntax

OUTPUT Rn

In which case it will output one register using a MOV ^Λ . Rn instruction.

The EMPTY instruction will stall until all registers to be empties contain valid data

d e are not empty)

Register list operations must not be used within re-mappmg REPEAT loops

Tne OUPUT instruction can only specif.' up to eight registers to output

Reasons for inclusion

After a routine has finished, the next routine expects all registers to oe empty so it can rece:'. e αata from the ARM. .An EMPTY instruction is needed to accompuss- this

Before performing a FIR or other filter all accumulators and partial results need to be zercea The ZERO instruction helps witn this. Both, are designed to improve code dens:?/ by replacing a series of smgie register moves The OUTPUT instruction is mciuced to improve code density by replacing a seπes of MOV ^Λ , Rn mstructions

.A Remapping Parameter Move Instruction RMOV is provided to allow m configuration of the user cefined register re-mapp g parameters

The instruction encoding is as follows.

: 30292S 27 26 25 24 23 22 21 20 19 IS 17 16 15 14 13 12 II 10 9 S 7 6 5 3 2

H i l l 101 00 ZPARAMS YPARA S XP ARAMS

Ξacn P.ARAMS field is comprised of the following entnes:

The meaning of these entries is described below

PARAMETER DESCRIPTION

RENUMBER Number of 16-bit registers to ' perform re-mapping on, may take the values 0, 2, 4, 8.

Registers below RENUMBER are re-mapped, those above are accessed directlv.

BASEINC The amount the base pointer is incremented at the end of each loop. May take the values 1 , 2. o:

4.

BASEWRAP The base wrapping modulus may take the values 2. 4. 8.

Mnemonics:

RMOV <PARAMS>, [<P.ARAMS>]

Tne <PARAMS> field has me following format:

<PARAMS> ::= ^• <BANK><B.ASEINC> n<RENUMBER> w<BASEWRAP>

98/12627

87

<BANK> ::= [XjYjZ]

<BASEINC> ::= [-H-j+l!+2|+4]

<RENUMBER> ::= [0;2|4!8]

<BASEWRAP> ::= [2j4|8]

If the RMOV instruction is used whilst re-mapping is active, the behaviour is UNPREDICTABLE.

Flags:

Unaffected

Repeat Instructions provide four zero cycle loops in hardware. The REPEAT instruction defines a new hardware loop. Piccolo uses hardware loop 0 for the first REPEAT instruction, hardware loop 1 for a REPEAT instruction nested within the first repeat instruction and so on. The REPEAT instruction does not need to specify which loop is being used. REPEAT loops must be strictly - nested. If an attempt is made to nest loops to a depth greater than 4 then the behaviour is unpredictable.

Each REPEAT instruction specifies the number of instructions in the loop (which immediately follows the REPEAT instruction) and the number of times to go around the loop (which is either a constant or read from a Piccolo register).

If the number of instructions in the loop is small (1 or 2) then Piccolo may take extra cycles to set the loop up.

If the loop count is register-specified, a 32-bit access is implied (Sl = l), though only the bottom 16 bits are significant and the number is considered to be tmsigned. If the loop count is zero, then the action of the loop is undefined. A copy

of the ioop count is taken so the register can be immediately reused (or even refilled) without affecting the loop.

The REPEAT instruction provides a mechanism to modify the way in which register operands are specified within a loop. The details are described above

Encoding of a REPEAT with a register speciiied number of loops:

31 302928272525 24 23 2221 20 19 18 17 16 15 14 13 12 11 10 1 0

Encoding of REPEAT with a fixed number of loops:

31 30 29 28 27 25 25 24 23 22 21 20 19 18 17 16 15 14 13 12 1 1 10 9 6 5 I 0

1 1 1 10 RFIELD_4 sLOOPS 13 SΓNSTRUCTIONS

the RFIELD operand specifies which of 16 re-mapping parameter configurations to use inside the ioon.

98/12627

89

The assembler provides two opcodes REPEAT and NEXT for defining a hardware loop. The REPEAT goes at the start of the loop and the NEXT delimits the end of the loop, allowing the assembler to calculate the number of instructions in the loop body. For the REPEAT it is only necessary to specify the number of loops either as a constant or register. For example:

REPEAT XO

MULA AO, Y0.1 , Z0.1 . AO MULA AO. Y0.h ^Λ , Z0.h ^Λ , AO NEXT

This will execute the two MULA instructions XO times. Aiso, REPEAT #10

MULA AO, X0 ^Λ , Y0 ^Λ , AO

NEXT

will perform 10 multiply accumulates. The assembler supports the syntax:

REPEAT ^iterations [. <PARAMS>]

To specify the re-mapping parameters to use for the REPEAT. If the required re¬ mapping parameters are equal to one of the predefined set of parameters, then the appropriate REPEAT encoding is used. If it is not then the assembler will generate an RMOV to load the user defined parameters, followed by a REPEAT instruction. See the section above for details of the RMOV instruction and the re-mappmg parameters format. ϊ: the number of iterations for a loop is 0 then the action of REPEAT is UNPREDICTABLE.

If me number of instructions field is set to 0 then the action of REPEAT is UNPREDICTABLE.

.A ioop consisting of only one instruction, with that instruction being a branch wii ^' . have UNPREDICTABLE behaviour. Branches within the bounds of a REPE.AT loop that branch outside the bounds of that loo? are UNPREDICTABLE.

The Saturating .Absolute instruction calculates the saturated absolute of source 1.

3; 30292S 2 ^' 252524 23 2221 20 19 1817 16 15 14 13 12 11 10 9 3 7 6 5

Action: i dest = SAT((src l >= 0) ? src l : -src l ). The value is alwavs

saturated. In particular the absolute value of 0x80000000 is 0x7fffffff and NOT 0x80000000!

Mnemonic:

SABS <dest>, <srcl>

Flags:

Z is set if the result is zero.

N is preserved.

C is set of srcl <0 (dest=-srcl case)

V is set if saturation occured.

Reasons for inclusion:

Useful in many DSP applications.

Select Operations (Conditional Moves) serve to conditionally move either source 1 or source 2 into the destination register. A select is always equivalent to a move. There are also parallel operations for use after parallel adds/subtracts.

Note that both source operands may be read by the instruction for implementation reasons and so if either one is empty the instruction will stall, irrespective of whether the operand is strictly required.

302928 2726 25 24 23 2221 20 19 18 17 16 15 14 13 12 II 10 9 3 2 I 0

OPC specifies the type of instruction.

Action (OPC)

00 If <cond> holds for primary flags then c.sr=srcl

01 If <cond> holds for the primary flags then cest h=srcl .h

else cest.h=src2.h.

If <cond> holds for the secondary flags then cest.l=src l .l

else cest.l=src2.

10 If <cond> holds for the primary flags then dest.h=srcl.h

eise dest.h=src2.h,

If <cond> fails for the secondary flags then dest.l=srcl .l

else dest.l=src2.1.

1 1 Reserved

Mnemonics

00 SEL<cond> <dest>, <srcl>, <src2>

01 SELTT<cond> <dest>, <srcl>, <src2>

10 SELTF<cond> <dest>, <srcl>, <src2>

1 1 Unused

If a register is marked for refill, it is unconditionally refilled. The assembler also provides the mnemonics:

MOV<cond> <dest>, <srcl>

SELFT<cond> <dest>, <srcl>, <src2>

SELFF <cond> <dest> <srcl>. <src2>

MOV<cond> A,B is equivalent to SEL<cond> A. B, A. SELFT and SELFF are obtained by swapping srcl and src2 and using SELTF, SELTT.

Flags:

All flags are preserved so that a sequence of selects may be performed.

Reasons for inclusion:

Used for making simple decisions inline without having to resort to a branch. Used by Viterbi algorithms and when scanning a sample or vector for the largest element.

Snift Operation instructions provide left and right logical shifts, right arithmetic shifts, and rotates oy a specified amount The shift amount is considered to be a signed integer between -128 and -127 taken from the bottom 8 bits of me register contents or an immediate in the range -1 to -^31 A shift of a negative amount causes a snift in the opposite direction by ABS(shιft amount)

The input operanαs are sign extended to 32-bits, the resulting 32-bit output is sign extended to 48-bits before write back so that a write to a 48-bit register behaves sensiDiv

il 30 2827262524 23 222120 19 IS 1716 15 13 1 1 10 9 6 5 1 0

OPC specifies me tvpe of instruction.

Action (OPO

00 dest = (src2>=0) ? srcl « src2 : srcl » -src2

01 dest = (src2>=0) ? srcl » src2 : srcl « -src2

10 dest = (src2>=0) ? srcl -» src2 : srcl « -src2

1 1 dest = (src2>=0) ? srcl ROR src2 : srcl ROL -src2

Mnemonics:

00 ASL <dest>, <srcl>, <src2_16>

01 LSR <dest>, <srcl>, <src2_16> 10 ASR <dest>, <srcl>, <src2_16>

1 1 ROR <dest>. <srcl>, <src2_16>

Flags:

Z is set if the result is zero. N is set if the result is negative

V is preserved

C is set to the value of the last bit shifted out (as on the .ARM)

The behaviour of register specified shifts is: -LSL by 32 has result zero. C set to bit 0 of srcl .

-LSL by more than 32 has result zero. C set to zero.

-LSR by 32 has result zero, C set to bit 31 of srcl .

-LSR by more than 32 has result zero, C set to zero.

-ASR by 32 or more has result filled with and C equal to bit 31 of srcl . -ROR by 32 has result equal to srcl and C set to bit 31 of srcl.

-ROR by n where n is greater than 32 will give the same result and carry out as

ROR by n-32; therefore repeatedly subtract 32 from n until the amount is in the range 1 to 32 and see above.

Reasons for inclusion:

Multiplication/ ^' division by a power of 2. Bit and field extraction. Serial registers.

Undefined Instructions are set out above in the instruction set listing. Their execution will cause Piccolo to halt execution, and set the U bit in the status register, and disable itself (as if the E bit in the control register was cleared). This allows any future extensions of the instructions set to be trapped and optionally emulated on existing implementations.

Acessing Piccolo State from ARM is as follows. State access mode is used to observe/modify the state of Piccolo. This mechanism is provided for two purposes: -Context Switch. -Debug.

Piccolo is put in state access mode by executing the PSTATE instruction. This mode allows all Piccolo state to be saved and restored with a sequence of STC and LDC instructions. When put into state access mode, the use of the Piccolo coprocessor ID PICCOLOl is modified to allow the state of Piccolo to be accessed.

There are 7 banks of Piccolo state. All the data in a particular bank can be loaded and stored with a single LDC or STC.

Bank 0: Private registers. - 1 32-bit word containing the value of the Piccolo ID Register (Read Only).

- I 32-bit word containing the state of the Control Register.

- 1 32-bit word containing the state of the Status Register.

- 1 32-bit word containing the state of the Program Counter.

Bank 1 : Genera! Purpose registers (GPR).

- 16 32-bit words containing the general purpose register state.

Bank 2: Accumulators.

- 4 32-bit words containing the top 32-bits of the accumulator registers (N.B. duplication with GPR state is necessary for restoration purposes - would imply

another write enable on the register bank otherwise).

Bank 3: Register/Piccolo ROB/Output FIFO Status.

- 1 32-bit word indicating which registers are marked for refill (2 bits for each 32- bit register).

- 8 32-bit words containing the state of the ROB tags (8 7-bit items stored in bits 7 to O).

- 3 32-bit words containing the state of the unaligned ROB latches (bits 17 to 0)

- 1 32-bit word indicating which slots in the output shift register contain valid data (bit 4 indicates empty, bits 3 to 0 encode the number of used entries).

- 1 32-bit word containing the state of the output FIFO holding latch (bits 17 to 0).

Bank 4. ROB Input Data.

- 8 32-bit data values.

Bank 5: Output FIFO Data.

- 8 32-bit data values

Bank 6: Loop Hardware. - 4 32-bit words containing the loop start addresses.

- 4 32-bit words containing the loop end addresses.

- 4 32-bit words containing the loop count (bits 15 to 0).

- 1 32-bit word containing user defined re-mapping parameters and other re¬ mapping state.

The LDC instruction is used to load Piccolo state when Piccolo is in state

access mode The BANK field specifies which bank is being loaded.

30 29 23 27 26 25 24 20 19 18 16 15 14 13 12 1 1 10 9 6 5 1 0

COND BASE BANK PICCOLOl OFFSET

t he following sεαuence will load all Piccolo state from the address in register RO

LDP BO, [RO]- -16 ' , private registers

LDP B l , [RO], =6^ ' , load general purpose registers

LDP B2, ΓROI , =16 ' , load accumulators

LDP B3. [RO], =56 ' , load RegisterTLOBTΪFO status

LDP B-i, [R0], =. load ROB data

LDP B5, [RO], =32 ' . load output FIFO data

LDP 36. [RO]. =52 ' , load loop hardware

The S i C instruction is used to store Piccolo state when Piccolo is in state access mode The BANK field specifies which bank is being stored

5, 3029 28 272625 24 23 2221 20 19 IS 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

COND W BASE BANK PICCOLO l OFFSET

The following seαuence will store al! Piccolo state to the address in register RO

STP BO, [RO], #16 ! save private registers

STP Bl , [RO], #64 ! save general purpose registers

STP B2, [RO], #16 ! save accumulators

STP B3, [RO], #56 ! save Register/ROB/FIFO status

STP B4, [RO], #32 ! save ROB data

STP B5, [RO], #32 ! save output FIFO data

STP B6, [RO], #52 ! save loop hardware

Debug Mode - Piccolo needs to respond to the same debug mechanisms as supported by ARM i.e. software through Demon and Angel, and hardware with

Embedded ICE. There are several mechanisms for debugging a Piccolo system:

-ARM instruction breakpoints.

-Data breakpoints (watchpoints).

-Piccolo instruction breakpoints. -Piccolo software breakpoints.

ARM instruction and data breakpoints are handled by the ARM Embedded ICE module: Piccolo instruction breakpoints are handled by the Piccolo Embedded ICE module: Piccolo software breakpoints are handled by the Piccolo core. The hardware breakpoint system will be configurable such that both the ARM and

Piccolo will be breakpointed.

Software breakpoints are handled by a Piccolo instruction (Halt or Break) causing Piccolo to halt execution, and enter debug mode (B bit in the status register set), and disable itself (as if Piccolo had been disabled with a PDISABLE instruction).

The program counter remains valid, allowing the address of the breakpoint to be recovered. Piccolo will no longer execute instructions.

Single stepping Piccolo will be done by setting breakpoint after breakpoint on the

Piccolo instruction stream

Software Debug - The basic functionality provided by Piccolo is the ability to load and save all state to memory via coprocessor instructions when in state access mode. This allows a debugger to save all state to memory, read and/or update it, and restore it to Piccolo. The Piccolo store state mechanism will be non¬ destructive, that is the action of storing the state of Piccolo will not corrupt any of Piccolo's internal state. This means that Piccolo can be restarted after dumping its state without restoring it again first. The mechanism to find the status of the Piccolo cache is to be determined.

Hardware Debug - Hardware debug will be facilitated by a scan chain on Piccolo's coprocessor interface. Piccolo may then be put into state access mode and have its state examined/modified via the scan chain.

The Piccolo Status register contains a single bit to indicate that it has executed a breakpointed instruction. When a breakpointed instruction is executed, Piccolo sets the B bit in the Status register, and halts execution. To be able to interrogate Piccolo, the debugger must enable Piccolo and put it into state access mode by writing to its control register before subsequent accesses can occur.

Figure 4 illustrates a multiplexer arrangement responsive to the Hi/Lo bit and Size bit to switch appropriate halves of the selected register to the Piccolo datapath. If the Size bit indicates 16 bits, then a sign extending circuit pads the high order bits of the datapath with 0s or Is as appropriate.