KR20000048531A - Input operand control in data processing systems - Google Patents

Abstract

PURPOSE: A data processing system is provided to more effecively use bus bandwidth by connecting a register with data storage by a n-bit data bus. CONSTITUTION: A data processing system having a plurality of registers (10) and an arithmetic logic unit (20, 22, 24) includes program instruction words having a source register bit field Sn specifying one of the registers storing an input operand data word together with an input operand size flag indicating whether the input operand has a N-bit size or (N/2)bit size together with a high/low location flag indicating which of the high order bit positions or low order bit positions stores the input operand if it is of the smaller size. The arithmetic logic unit is also able to perform parallel operation program instruction words operating independently upon (N/2)bit input operand data words stored in respective halves of a register.

Description

INPUT OPERAND CONTROL IN DATA PROCESSING SYSTEMS}

The present invention relates to a data processing apparatus. In particular, the present invention relates to a data processing apparatus having a plurality of registers for storing data words to be operated by an arithmetic logic apparatus operating under the control of a program instruction word.

The present invention provides program instructions that indicate whether any one operation is performed on an input operand contained in one register or an input operand contained in two registers, but treated as a single input operand.

In one aspect, the present invention provides a data processing apparatus, which comprises:

A plurality of registers storing data words to be manipulated, each having at least N bits of capacity,

An arithmetic logic device that performs arithmetic logic operations specified by the program command word in response to a program command word,

The operation logic device

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

(ii) an input operand size flag indicating whether the input operand data word has an N bit size or an (N / 2) bit size;

(iii) when the input size flag indicates a (N / 2) bit size, a high / low indicating whether the input operand data is located at the highest bit position of the source register or the lowest bit position of the source register; Respond to at least one program command word that includes a (high / low) position flag.

In data processing, there is a tendency to increase the data path width of the data processing apparatus. Early data processing devices have an 8-bit data path. They were then extended to 16-bit data paths, followed by 32-bit and 64-bit data paths. As these datapath widths increased, the registers in the data processing apparatus also increased to match these datapath widths. If the data words to be manipulated are smaller than the data path width, the present invention recognizes that using all registers to store these data words wastes register resources of the data processing apparatus. This is especially the case for load / storage mechanisms where all data to be manipulated must be in registers and you want to reduce the number of times data must be fetched from cache or main memory. The present invention recognizes the above considerations and provides a solution for using an input operand size flag and a high / low position flag to indicate the input operand size. As such, only one register can hold more than one input operand using the register resource of the data processing device more efficiently, and still these input operands can be manipulated individually.

The advantage of the present invention is further enhanced when an N-bit data bus connects registers and data storage. In this case, the data bus can be used to transmit two operands at the same time by more efficiently using the bus bandwidth and reducing the likelihood of performance bottlenecks.

In a preferred embodiment of the invention, the arithmetic logic unit comprises a first (N / 2) bit input operand data word and a second (N / 2) stored within each of the highest and lowest bit positions of only one source register. Respond to at least one parallel operation program instruction word that performs individual arithmetic logic operations on the bit input operand data word.

By providing a parallel operation program command word, the two operations can be performed independently by an arithmetic logic device that fully utilizes the N-bit datapath capability even if the input operand size is less than the maximum datapath width. This significantly increases the data processing capacity of the data processing device, while not incurring significant extra overhead.

One variation that can be made is that the arithmetic logic device has a signal path that functions as a transport chain between bit positions in the arithmetic logic operation, and when executing a parallel operation program command word, the signal path is the first (N / 2) interrupted between the bit input operand data word and the second (N / 2) bit input operand data word.

The parallel operation program instruction may take various forms, but the parallel operation program instruction word is

(i) parallel addition, which adds two parallel (N / 2) bits;

(ii) parallel subtraction, which performs two parallel (N / 2) bit subtractions,

(iii) a parallel shift for performing two parallel (N / 2) bit shift operations,

(iv) It is preferable to perform one of the operation logic operations of parallel addition / subtraction which performs (N / 2) bit addition and (N / 2) bit subtraction in parallel.

More specifically with respect to the present invention, when the input size flag indicates an N bit size, the high / low position flag is used to determine whether these bits stored at the highest bit position should be moved to the lowest bit position and the lowest. Indicates that these bits stored in the upper bit position should be moved to the highest bit position before being used as an N bit input operand data word.

This feature is particularly useful when converting operations. In particular, an effective hardware implementation of this function includes the (N / 2) bit input operand data word stored in one of the most significant bit position of the source register and the most significant bit position of the source register. At least one multiplexer responsive to the high / low position flag selected to feed a bit.

In order to perform signed operations without severe complexity, it is desirable to provide a circuit for sign-extending (N / 2) -bit input operand data words before inputting into the N-bit datapath.

In another aspect, the present invention provides a data processing method, the data processing method,

Storing data words to be manipulated in a plurality of registers each having at least N bit capacity;

In response to a program command word, performing an arithmetic logic operation specified by the program command word,

At least one program command word is

(i) a source program bit field indicating 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 indicating whether the input operand data word has an N bit size or an (N / 2) bit size;

(iii) when the input size flag indicates an (N / 2) bit size, a high indicating whether the input operand data word is located in one of the highest bit position of the source register and the lowest bit position of the source register; Include the / row position flag.

Embodiments of the present invention will be described with reference to the accompanying drawings.

1 shows a configuration of a digital signal processing apparatus,

Figure 2 shows the input buffer of the register configuration of the coprocessor,

3 shows the datapath passing through the coprocessor,

4 shows a multiplexing circuit for reading the highest or lowest bits from a register,

5 is a block diagram showing register remapping logic used by a coprocessor in a preferred embodiment,

FIG. 6 illustrates the register remapping logic shown in FIG. 5 in more detail.

7 is a table illustrating a block filter algorithm.

The system described later relates to digital signal processing (DSP). A DSP can take many forms, but will typically be considered a process that requires fast (real-time) processing of large amounts of data. This data typically represents some analog physical signal. A good example of a DSP is a digital mobile phone in which wireless signals are transmitted and received that require decoding to analog sound signals (typically using convolution operations, conversion operations, and correlation operations) and encoding from analog sound signals. . Another example is a disc driver controller which processes head-played signals by processing signals reproduced from the disc head.

In the context of the above, it follows the description of a digital signal processor based on a microprocessor core, in this case an ARM core from the range of microprocessors designed by Advanced RISC Machines Limited of Cambridge, United Kingdom. . The interface between the microprocessor and the coprocessor and its own coprocessor architecture is clearly configured to provide a functional DSP. The microprocessor core is called ARM and the coprocessor is called Piccolo. ARM and Piccolo will typically be fabricated as a single integrated circuit that often includes other devices (eg, on-chip DRAM, ROM, D / A and A / D converters, etc.) as part of an ASIC.

Piccolo is an ARM coprocessor, so it executes part of the ARM instruction set. ARM coprocessor instructions allow ARM to transfer data between Piccolo and memory (using the load coprocessor LDC and store coprocessor STC instructions), and move them to coprocessors, MCR instructions, coprocessor, MRC instructions. Transfer the ARM register from / to Piccolo. One way of investigating the synergistic interaction between ARM and Piccolo is that ARM acts as a powerful address generator for Piccolo data, responding by executing DSP operations that require real-time processing of large amounts of data with Piccolo freely placed. To generate real-time results.

Figure 1 shows an ARM 2 and a piccolo 4, which provide a control signal to the piccolo 4 to control the transfer of data words to / from the piccolo 4 . The instruction cache 6 accumulates the Piccolo program instruction words required by Piccolo 4. A single DRAM memory 8 accumulates all data and instruction words required by the ARM 2 and Piccolo 4. This ARM 2 is responsible for addressing the memory 8 and controlling all data transfers. Devices with only one memory 8 and one set of data buses and address buses are less complex and less expensive than typical DSP methods requiring high bus bandwidth and multiple memories and buses.

Piccolo executes a second instruction stream (digital signal processing program instruction word) from the instruction cache 6 that controls the Piccolo datapath. These instructions include digital signal processing forms of operation, such as multiply-accumulate, and control flow instructions, such as zero overhead loop instructions. These instructions operate on the data held in Piccolo register 10 (see FIG. 2). This data was initially transferred from the memory 8 by the ARM 2. This command flows out of the instruction cache 6, which drives the data bus as a full bus master. The small Piccolo instruction cache 6 would be 4 lines, with 16 words per line representing the mapped cache (64 instructions). In some implementations, it is worthwhile to make the instruction cache larger.

In this way, the two tasks execute ARM, which independently loads data, and Piccolo, which processes it. This allows a single cycle of data to be processed for 16 bit data. Piccolo has a data input mechanism (shown in Figure 2) that allows ARM to prefetch contiguous data and load that data before it is required by Piccolo. Piccolo can access the loaded data in any order, automatically recharging its register when the old data is used at the last time (every instruction is one per source operand to indicate that the source register should be refilled). Bit). This input mechanism is called an ordering buffer and has an input buffer 2. Every value loaded into Piccolo (via LDC or MCR) carries an Rn tag that indicates which register is set. The tag Rn is stored on the data word side in the input buffer. When one register is accessed through the register selector 14 and the command indicates that the data register is to be refilled, this register is marked blank by asserting the signal E. The register is automatically recharged by the recharge control circuit 16 using the oldest loaded data defined for the register in the input buffer 12. The order buffer holds eight tagged values. This input buffer 12 has a form similar to a FIFO except that data words can be extracted from the center of the queue after the later stored words have been passed to fill the space. Thus, using this to determine which data word should be used to recharge the register when the data word furthest from the input is the oldest data word and the input buffer 12 holds two data words with the correct tag Rn. You can decide.

Piccolo outputs the data after storing the data in the output buffer 18 (FIFO), as shown in FIG. Data is sequentially written to the FIFO and read out to the memory 8 in the same order by the ARM. The output buffer 18 holds eight 32-bit values.

Piccolo is connected to the ARM through the coprocessor interface (CP control signal of FIG. 1). Upon execution of an ARM coprocessor instruction, Piccolo executes the instruction and may cause ARM to wait until Piccolo is ready before executing the instruction, or may not execute the instruction. In the last case, ARM will exclude instructions that are not defined.

Most common coprocessor instructions that Piccolo executes correspond to LDCs and STCs that load and store data words in and out of memory 8 via the data bus, respectively, and the ARM generates all addresses. These instructions load data into the order buffer and store the data from the output buffer 18. If there is not enough space in the input order buffer to load into the data, Piccolo will disable the function of the ARM with the LDC, and if there is insufficient data in the output buffer to store, i.e. the data that ARM expects is in the output buffer 18. If not, the STC will disable ARM. Piccolo also executes ARM / coprocessor register transfers, allowing ARM to access Piccolo's special registers.

Piccolo fetches its own instruction from the memory to control the Piccolo data path shown in Figure 3, transfer data from the order buffer to the register, and transfer data from the register to the output buffer 8. Piccolo's arithmetic logic that executes these instructions has a multiplication / addition circuit 20 that performs multiplication, addition, subtraction, multiplication-accumulation, logic operations, shift and rotation. In addition, an accumulation / cumulative disposal circuit 22 and a scale / saturation circuit 24 are provided in the data path.

Piccolo instructions are initially loaded from memory into the instruction cache 6, where Piccolo can access those instructions without having to go back and access the main memory.

Piccolo cannot recover from memory outages. Thus, if Piccolo is used in a virtual memory system, all Piccolo must be in physical memory during the Piccolo task. Given the real-time nature of the Piccolo task, such as a real-time DSP, this is not an important limitation. If a memory interruption occurs, Piccolo will stop and set a flag in status register S2.

3 shows Piccolo's overall datapath function. The register bank 10 uses three read ports and two write ports. One write port (L port) is used to recharge the register from the order buffer. The output buffer 18 is updated directly from the ALU result bus 26 and the output from the output buffer 18 is under ARM program control. The ARM coprocessor interface loads LDC (load coprocessor) instructions in the order buffer and STC (store coprocessor) instructions from MCR and MRC (moves ARM registers to / from CP registers) as well as output buffers 18. Perform with register bank 10.

The remaining register port is used for the ALU. The two read ports A and B drive the inputs to the multiply / add circuit 20, and use the C read ports to drive the cumulative / cumulative disposal circuit 22 inputs. The remaining write port W is used to return the result to register bank 10.

The multiplier 20 performs an arbitrary 48-bit accumulation and 16x16 signed or unsigned multiplication. Scale portion 24 may provide an immediate operation of 0-31 or a logical shift to the right that comes before any saturation. The shifter and logic device 30 may perform either a shift operation or a logic operation every cycle.

Piccolo has 16 general purpose registers DO-D15 or A0-A3, X0-X3, Y0-Y3, Z0-Z3. The first four registers (A0-A3) are designated as accumulators, 48 bits wide, and an extra 16 bits to monitor for overflow during many consecutive calculations. The remaining registers are 32 bits wide.

Each of Piccolo's registers can be considered to contain two independent 16-bit values. Half of the low values are bits 0-15, and half of the high values are bits 16-31. The instructions may represent a particular 16 bit half of each register as a source operand, or may represent an entire 32 bit register.

Piccolo also provides a saturated operation. Modifications of the multiply, add, and subtract instructions provide a saturated result if the saturated result is greater than the size of the destination register. The destination register is a 48-bit accumulator, whose value is saturated to 32 bits (ie there is no way to saturate a 48-bit value). No overflow is detected for 48-bit registers. This is a reasonable restriction, because it takes at least 65536 multiply accumulate instructions to cause overflow.

Each Piccolo register is marked as "blank" (E flag, see Figure 2), or contains one value (it is not possible to have half the blank of a register). Initially, all registers are marked blank. For each cycle, Piccolo attempts to use the recharge control circuit 16 to fill one of the empty registers with a value from the input reorder buffer. Alternatively, if any one value from the ALU is written to this register, it is no longer marked as "blank". If there is a value waiting to be written to the register from the ALU and placed in the register from the order buffer, the result is not limited. Reading from the free register will stop Piccolo's execution.

An Input Reorder Buffer (ROB) is located 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 representing a Piccolo register for which a value is defined. This tag also indicates whether data should be sent to the entire 32-bit register or only to the lower 16 bits of the 32-bit register. Once the data has been set for the entire register, the bottom 16 bits of the entry will be sent to the bottom half of the target register, and the top 16 bits will be sent to the top half of the register (extended sign when the target register is a 48-bit accumulator). . If data is reserved for the bottom half of the register only, the bottom 16 bits will be sent first.

The register tag always describes the physical destination register and does not remap any registers (see below for register remapping).

Each cycle Piccolo attempts to transfer a data entry from the ROB to the register bank as follows.

-Examine each entry in the ROB and compare the tag to a register that is empty to determine whether it can transfer from some or all of the entries to the register.

From the set of entries that can be transferred, select the oldest entry and transfer its data to the register bank.

-Update the tag of this entry to mark the entry as blank. When only part of an entry is sent, only the part sent is marked blank.

For example, if this target register is completely empty and the selected ROB entry contains data defined for all registers, a total of 32 bits are transferred and the entry is marked as empty. If the bottom half of the target register is blank and the ROB entry contains data defined for the bottom half of the 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 blank.

High 16 bits and low 16 bits of data in any entry can be transmitted independently. If no entry contains data that can be transferred to the register bank, no transfer is made for that cycle. The table below describes all possible combinations of target ROB entry and target register states.

Target, Rn, state Target ROB Entry Status Blank Low half blanks High half blanks All registers, both half valid Rn.h <-entry.hRn.l <-entry.l The entry indicated a space. Rn.l <-entry.lentry.l shows a space. Rn.h <-entry.hentry.h shows a blank Full register, high half valid Rn.h <-entry.h The entry is blank. Rn.h <-entry.h The entry is blank. Full register, low half valid Rn.h <-entry.l entry shows blank Rn.l <-entry.l entry shows blank Half register, both half valid Rn.h <-entry.lentry.l shows a space Rn.h <-entry.lentry.l shows a space Half register high half valid Rn.h <-entry.h The entry is blank. Rn.h <-entry.h The entry is blank.

In summary, the two half registers will be recharged independently from the ROB. The data in the ROB is displayed as defined for all registers or as two 16-bit values defined for the lower half of the register.

Load data into the ROB using the ARM coprocessor instructions. How much data is displayed in the ROB depends on the transmission using the ARM coprocessor instructions. The following ARM instruction can be used to fill the ROB with data:

LDP {<cond>} <16/32> <dest>, [Rn] {!}, # <Size>

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

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

MPR {<cond>} <dest>, Rn

MRP {<cond>} <dest>, Rn

The following ARM instructions are provided to construct the ROB.

LDPA <bank list>

The first three are assembled as LDC, MPR and MRP as MCR, LDP are assembled as CDP instructions.

Where <dest> represents a piccolo register (A0-Z3), Rn represents an ARM register, <size> represents an integer number of bytes that must be a multiple of four, not zero, and <wrap> represents an integer (1, 2, 3, 8). Fields surrounded by {} are optional. For transfers to fit within the order buffer, <size> should be at most 32. In many situations, <size> will be less than this limit to avoid deadlocks. The <16/32> field indicates whether the loaded data should be treated as 16-bit data or as 32-bit data.

Note 1: In the following text, when referring to LDP or LDPW, this indicates a 16-bit and 32-bit variant of the instruction.

Note 2: One 'word' is a 32-bit group from memory consisting of two 16-bit data items or one 32-bit data item.

The LDP instruction transfers multiple data items, displaying data items as specified for all registers. This instruction will load <size> / 4 words from address Rn in memory, inserting words into the ROB. The number of words that can be transmitted is limited to the following.

The quantity <size> must be a multiple of four, not zero.

<size> must be less than or equal to the size of the ROB for a particular implementation (8 words in the first version, as in the first version in a future version).

The transmitted first data item is tagged as defined for <dest>, and the second data item is tagged as defined for <dest> +1 and the like (with overlapping from Z3 to A0). If! Is specified, the register Rn is later incremented by <size>.

When the LDP16 variant is used, endian special operations are performed on two 16-bit halfwords that form 32-bit data items as returned from the memory system. For more details on large endian and small endian support see:

The LDPW instruction transfers multiple data items into one set of registers. The transmitted first data item is tagged as defined for <dest>, and the next data item is tagged as defined for dest> +1 and the like. When a <wrap> transfer occurs, the next item to be sent is tagged as specified for <dest> and so on. <wrap> amount is specified as the amount of halfwords.

For LDPW, the following restrictions apply.

The quantity <size> must be a multiple of four, not zero.

<size> must be less than or equal to the size of the ROB for a particular implementation (8 words in the first version, as in the first version in a future version).

-<dest> will be one of {A0, X0, Y0, Z0}.

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

If the quantity <size> is not greater than 2 * <wrap> no wrap will occur, and the LDP instruction will be used instead.

For example, the command

LDP32W X0, 2, [R0] !, # 8

We will load two words into the ROB, displaying two words as defined for the entire register X0. R0 will be increased by eight. The command below is

LDP32W X0, 4, [R0], # 16

We will load four words into the ROB and display four words as specified (in this order) for X0, X1, X0, X1. R0 will not be affected.

For LDP16W, <wrap> will be designated as 1, 2, 4 or 8. Due to the overlap of 1, all data is tagged as specified for the lower half of the destination register <dest> .l. This is the 'half register' case.

For example, the command

LDP16W X0,1, [R0]!, # 8

Two words will be loaded into the ROB, representing two words as 16-bit data specified for X0.1. R0 will be increased by eight. The command below is

LDP16W X0,4, [R0], # 6

It will work in a similar way to the LDP32W example, except that it will perform endian special operations on the data when it is returned from memory.

All unused encodings of the LDP instruction will be specified for future expansion.

The LDP16U command is provided to support efficient transfer of non-word aligned 16-bit data. LDP16U support is provided for registers D4-D15 (X, Y, and Z banks). The LDP16U instruction will transfer one 32-bit word (two 16-bit data items) of data from memory into Piccolo. Piccolo will discard the lower 16 bits of this data and store the upper 16 bits in the retention register. There is a holding register for the X, Y and Z banks. Once the bank's holding registers are prepared, they change the behavior of the LDP {W} instruction when data is destined for the registers in this bank. The data loaded into the ROB is formed by the concatenation of the lower 16 bits of the data being transmitted by the LDP instruction and the retention register. The upper 16 bits of the data being transferred are placed in a holding register.

entry <-data.l: holding_resister

holding_register <-data.h

The mode of operation is unchanged until turned off by the LDPA instruction. This holding register does not write the destination register tag or size. These properties are obtained from the command giving the next value of data.l.

Endian special actions always occur for data returned by the memory system. Since all 32-bit data items are assumed to be words aligned in memory, there is no 16-bit equivalent to the LDP16U.

Use the LDPA instruction to switch the unaligned mode of operation initiated by the LDP16U instruction. This unaligned mode will be turned off independently for banks X, Y, and Z. For example, the following command:

LDPA {X, Y}

It will turn off the unaligned mode for banks X and Y. Data in the retention registers of these banks will be discarded.

This will allow you to run LDPA on a bank that is not in unaligned mode, leaving it in unaligned mode.

The MPR instruction places the contents of ARM register Rn in the ROB, as specified for Piccolo register <dest>. The destination register <dest> will be any previous register in the range of A0-Z3. For example, the following command:

MPR X0, R3

We will transfer the contents of R3 into the ROB and display the data as specified for all registers X0.

Since ARM has little endianness inside, no endian special behavior occurs when data is transferred from ARM to Piccolo.

When the MPRW instruction displays data as two 16-bit data items specified for the 16-bit Piccolo register <dest..l, the contents of the ARM register Rn are placed in the ROB. The limit for <dest> is the same as for LDPW instructions (ie A0, X0, Y0, Z0). For example, the following command:

MPRW X0, R3

We will transfer the contents of R3 into the ROB and display the data as two 16-bit quantities defined for X0.1. For the overlap of 1 and the LDP16W case, only the bottom half of the 32-bit register will be targeted.

As for the MPR, no endian special operation is applied to the data.

LDP is encoded as follows:

Where PICCOLO1 is Piccolo's first coprocessor number (currently 8). The N bit selects either LDP32 (1) or LDP16 (0).

The LDPW is encoded as follows.

Where DEST is 0-3 for destination registers A0, X0, Y0, Z0, and WRAP is 0-3 for overlapping values 1, 2, 4, and 8. PICCOLO2 is Piccolo's second coprocessor number (currently 9). The N bit selects either LDP32 (1) or LDP16 (0).

The LDP16U is encoded as follows.

Here, DEST is 1-3 of destination banks X, Y, and Z.

LDPA is encoded as follows.

Here, BANK [3: 0] is used to turn off the unaligned mode on a bank basis. If BANK [1] is set, the unaligned mode for bank X is turned off. If BANK [2] and BANK [3] are set respectively, turn off the unaligned mode for banks Y and Z. This is a CDP operation.

MPR is encoded as follows.

MPRW is encoded as follows:

Here, DEST is 1-3 for destination registers X0, Y0, Z0.

The output FIFO can hold up to eight 32-bit values. They are sent from Piccolo using one of the following (ARM) opcodes:

STP {<cond>} <16/32> [Rn] {!}, # <Size>

MRP Rn

If present, the first save <size> / 4 words to the address given by ARM register Rn, which indexes Rn from the output FIFO. To prevent deadlocks, <size> must be larger than the size of the output FIFO (8 entries at this time). If the STP16 variant is used, endian special behavior will occur on the data returned from the memory system.

The MRP instruction removes one word from the output FIFO and places it in the ARM register Rn. As for the MPR, no endian special operation is applied to the data.

The ARM encoding for STP is as follows:

Here, N selects one of STP32 (1) and STP16 (0). For the definition of the P, U and W bits, an ARM data sheet is described.

The ARM encoding for MRP is

The Piccolo instruction set assumes little endian operation inside. For example, suppose that when accessing a 32 bit register as a 16 bit half, the bottom half occupies bit 15-0. Piccolo works in and around systems with large endian memory, so it must be handled to load 16-bit full data in the correct way.

Piccolo (i.e., DSP retrofit coprocessor) such as ARM (e.g., the ARM7 microprocessor manufactured by Advanced RISC Machines Limited of Cambridge, United Kingdom) has a 'BIGEND' configuration pin that the programmer may possibly control with a programmable peripheral. Have Piccolo uses this pin to configure the input order buffer and the output FIFO.

When ARM loads full 16-bit data into the order buffer, it displays 16-bit data using the 16-bit form of the LDP instruction. This information is combined with the state of the 'BIGEND' configuration input to place the data in the retention latch and order buffer in the proper order. In particular, in large endian mode, the holding register stores the lower 16 bits of the loaded word and is paired with the upper 16 bits of the next load. The holding register contents always end in the lower 16 bits of the word transferred into the order buffer.

The output FIFO contains either full 16 bits or 32 bits. Since the programmer must use the correct form of the STP instruction, Piccolo can ensure that 16-bit data is provided for the correct half of the data bus. When configured with large endian, the upper and lower 16-bit halves are swapped when the 16-bit form of STP is used.

Piccolo has four dedicated registers that can only be accessed from ARM. These dedicated registers are called S0-S2. These dedicated registers can only be accessed with MRC and MCR instructions. The operation code is as follows.

MPSR Sn, Rm

MRPS Rm, Sn

These opcodes transfer 32-bit values between the ARM register Rm and the dedicated register Sn. These opcodes are encoded in ARM as coprocessor register transfers as follows:

Where L is 0 for MPSR and 1 for MRPS.

Register S0 contains Piccolo unique ID and revision code.

Bits [3: 0] contain the revision number for the processor.

Bits [15: 4] contain the binary coded decimal format: 0x500 to 3 digits for Piccolo.

Bits [23:16] contain component version: 0x00 = version 1.

Bits [31:24] contain the ASCII code of the Executor Trademark: 0x41 = A = ARM Ltd.

Register S1 is a Piccolo status register.

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

Secondary Condition Code Flags (SN, SZ, SC, SV)

E bit: Piccolo was disabled by the ARM and stopped.

U bit: Piccolo stopped in collision with UNDEFINED command.

Bit B: Piccolo crashed with a BREAKPOINT.

H bit: Piccolo was stopped in conflict with the HALT instruction.

A bit: Piccolo was stopped by allowing a memory abort (load, store or Piccolo instruction).

D bit: Piccolo was stopped by detecting a deadlock condition (see below).

Register S2 is the Piccolo program counter as follows.

When writing to the program counter, Piccolo will start executing the program at that address. Since Piccolo is always started by writing to the program counter, the program counter at the time of reset is not specified.

At run time, Piccolo monitors the execution of instructions and the status of the coprocessor interface. If it is detected,

Piccolo does not wait for either register or output FIFO to be recharged to have an available entry.

The coprocessor interface is waiting for use 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, stops and rejects the ARM coprocessor instruction, causing the ARM to trap an indeterminate instruction.

By reading the ARM and Piccolo program counters and registers, a system can be configured that can at least warn the programmer that a condition has been caused by the detection of a deadlock condition, and a system can be configured that can report the exact point of failure. Deadlock can only be caused by an incorrect program or another part of the system that modulates Piccolo's state. Deadlock cannot occur due to lack of data or 'overload'.

There are several operations available for controlling Piccolo from ARM, and these operations are provided by CDP instructions. These CDP instructions will only be accepted when the ARM is in privileged state. Otherwise, Piccolo will reject CDP instructions from ARM that take indeterminate instruction traps.

The following operations are available.

_ Reset

_ Enter status access mode

_ Enable

_ Disable

Piccolo will be reset in software using the PRESET command.

PRESET; Clear Piccolo Status

This command is encoded as follows:

When this command is executed, the following occurs:

All registers are marked with a blank (prepare to recharge).

-The input ROB is cleared.

-Output FIF0 is cleared.

The loop counter is reset.

Piccolo remains stopped (and the H bit of S2 will be set).

By executing the PRESET command some cycles will be completed (2-3 for this embodiment). While the instruction is executing, the next ARM coprocessor instruction to be executed for Piccolo will be waiting for use.

In state access mode, Piccolo's state will be saved and restored using the STC and LDC instructions (see below for accessing Piccolo state from the ARM). To enter the status access mode, the PSTATE command must be executed first.

PSTATE; Enter status access mode

This command is encoded as follows:

When the command is executed, the PSTATE command

-Stop Piccolo (if not yet stopped), setting the E bit in Piccolo's status register.

-Configure Piccolo to its state access mode.

Because Piccolo's instruction pipeline must be removed before Piccolo can stop, several cycles will be completed by executing the PSTATE instruction. When the instruction is executed, the next ARM coprocessor instruction to be executed for Piccolo will be queued.

The PENABLE and PDISABLE commands are used for quick contact switching. When Piccolo is disabled, only dedicated registers 0 and 1 (ID and status registers) can be accessed from privileged mode. Access to any other state or access from user mode excludes ARM indeterminate instructions. Piccolo stops running when Piccolo is disabled. When Piccolo stops executing, Piccolo will acknowledge that by setting the E bit in the status register.

Piccolo is enabled by executing the PENABLE instruction.

PENABLE; Piccolo Enable

This command is encoded as follows:

Piccolo is disabled by executing the PDISABLE command.

PDISABLE; Piccolo Disable

This command is encoded as follows:

When this command is executed, the following occurs:

Piccolo's command pipeline will be removed.

Piccolo will stop, and the H bit in the status register will be set.

The Piccolo instruction cache holds Piccolo instructions that control the Piccolo datapath. If present, it holds at least 64 instructions, ensuring that it starts at 16 word boundaries. By its action, the cache fetched a line of (16) instructions starting at the specified address. This fetch occurs even if the cache already holds data associated with this address.

PMIR Rm

Piccolo must be stopped before PMIR can be performed.

The MCR encoding of this opcode is as follows.

This section discusses the Piccolo instruction set, which controls the Piccolo data path. Each instruction is 32 bits. This instruction is read from the Piccolo instruction cache.

The decoding of the instruction set goes completely straight. The upper six bits 26-31 give bits 22-25 to the significant opcodes, which provide insignificant opcodes for some special instructions. Bits shaded in grey are currently unused for expansion (these bits must contain the indicated value).

There are seven important command classes. This does not fully correspond to the main opcode field in the instruction to facilitate decoding some subset.

The commands in the table have the following names.

Standard data operation

Logical operation

Condition addition / subtraction

Indeterminate

shift

Selection

Indeterminate

Parallel selection

Multiplication

Indeterminate

Double multiplication

Indeterminate

Move code immediately

Indeterminate

repeat

repeat

Register list operations

Branch

Rename Parameter

Stop / Stop

The format for each class of commands is described in detail in the following sections. The source and destination operand fields are common to most instructions, as is register remapping, and are described in detail in the division section.

Most instructions require two source operands, source 1 and source 2. Some exceptions are absolute oversupply.

The Source 1 (SRC1) operand has the following 7 bit format.

The elements of the field have the following meanings.

Size indicates the size of the operand for reading (1 = 32 bits, 0 = 16 bits).

-Recharge- specifies that the register should be marked blank after being read and can be refilled from the ROB.

Register number encodes any of the 16 32-bit registers for reading.

Hi / Lo represents half of the 32-bit register for reading for 16-foot reading. For 32-bit operands, if set, indicates that half of the two 16-bit registers must be interchanged.

size Hi / Lo Part of the register being accessed 0 0 Low 16-bit 0 One 16 bit high One 0 32 bits One One I swapped 32 bits, half.

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

Typical source 2 (SRC2) has one of the following three 12-bit formats.

4 shows the multiplexer placement in response to the Hi / Lo bit and the size bit to convert the appropriate half of the selected register into the Piccolo datapath. If this size bit represents 16 bits, the sign extension circuit embeds the most significant bit of the data path with the sign 0 or 1.

The first encoding designates the source as a register, and the field has the same encoding as the SRC1 specifier. The SCALE field specifies the scale that should be applied to the result of the ALU.

8-bit immediates of cyclic encoding can result in 32-bit immediates that can be represented by 8-bit values and 2-bit cyclics. The following table shows the immediate values that can be generated from the 8-bit value XY.

cycle Immediate 0 0x000000XY One 0x0000XY00 10 0x00XY0000 11 0xXY000000

6-bit immediate encoding can use 6-bit unsigned immediates (range 0-63), along with the scale applied to the output of the ALU.

Common source 2 encodings are mostly common to instruction variants. There are some exceptions to the rules that support or slightly modify a limited subset of source 2 encoding.

-Select command.

Shift command.

Parallel operation.

-Multiply accumulate instruction.

-Multiplication double instruction.

The select instruction supports only one register or 6-bit unsigned immediate operand. This scale is not available because these bits are used by the condition field of the instruction.

The shift instruction supports only 16-bit registers or 5-bit unsigned immediate operands between 1 and 31. No scale of the result is available.

In the case of parallel operations, if a register is specified as the source of the operand, a 32 bit read must be performed. This immediate encoding has a slightly different meaning for parallel operations. This allows immediate copying on both 16-bit halves of a 32-bit operand. A slightly limited range of scales is available for parallel computations.

If 6-bit immediates are used, they are always copied over both halves of the 32-bit quantity. If 8-bit immediates are used, 8-bit immediates are copied only if this cycle indicates that 8-bit immediates must be cycled on the upper half of the 32-bit quantity.

cycle Immediate 0 0x000000XY One 0x0000XY00 10 0x00XY00XY 11 0xXY00SY00

No scale is available for parallel select operations, and this scale field will be set to zero for these instructions.

An 8-bit circular immediate cannot be specified by a multiply accumulate instruction. Bit 10 of the field is used to specify which accumulator to use in part. Source 2 is implied as a 16 bit operand.

Multiply double instructions cannot use integers. Only 16-bit registers can be specified. Bit 10 of the field is used to partially specify which accumulator to use.

Some instructions always include 32-bit operations (eg ADDADD), in which case the size bit will be set to 1 and the Hi / Lo bits are used to selectively exchange two 16-bit halves of a 32-bit operand. . Some instructions always include 16-bit operations (eg, MUL), and the size bit should be set to zero. Hi / Lo selects which half register is used (assuming the missing size bits are cleared). Multiply accumulate instructions allow independent specification of the source accumulator and destination register. For these instructions, the size accumulator is used to indicate the source accumulator, and the size bits are implied as zero by the instruction type.

When a 16-bit value is read (via the A or B bus), it is automatically code extended to a 32-bit quantity. If a 48-bit register is read (via the A or B bus), only the bottom 32 bits appear on the bus. Thus, in all cases, Source 1 and Source 2 are converted to 32 bit values. Using bus C, only an accumulate instruction can access all 48 bits of an accumulator register.

If the recharge bit is set, the register will be blank after use and will be filled from the ROB with a normal refill mechanism (see part on the ROB). Piccolo will not stop unless the register is used again as the source operand before it is refilled. The minimum number of cycles (in the best case-the data is waiting at the head of the ROB) before the replenished data is in effect will be either 1 or 2. Thus, it is wise not to use refilled data in commands that follow a recharge request. If the use of operands can be avoided for the next two instructions, this will prevent performance loss for deeper pipeline execution.

The refile bit is specified in assembly by appending a '^' to the end of the register number. The portion of the register marked as empty depends on the register operand. Two halves of each register will be displayed independently for recharging (eg, X0.1 ^ will only display the bottom half of X0 for recharging, and X0 ^ will display the entirety of X0 for recharging). If the upper 'half' of the 48-bit register (bit 47:16) is refilled, 16 bits of data are written into bits 31:16 and sign extended to bit 47.

If an attempt is made to recharge the same register twice (eg ADD X1, X0 ^, X0 ^), only one recharge occurs. The assembler should only allow syntax ADD X1, X0, X0 ^.

If a register read is attempted before the register is refilled, Piccolo does not wait for the register to be refilled. If the register for recharging is displayed and the register is updated before the refilled value is read, the result is UNPREDICTABLE (e.g. ADD X0, X0 ^, X1 cannot predict, because ADD X0, X0 ^, X1 Since X0 is indicated for this recharging, the sum of X0 and X1 is placed in ADD X0, X0 ^, X1 to recharge ADD X0, X0 ^, X1).

The 4-bit scale field encodes 14 scale types.

-ASR # 0, 1, 2, 3, 4, 6, 8, 10

-ASR # 12-16

-LSL # 1

Since the parallel Max / Min instruction does not provide one scale, a constant 6-bit variant of source 2 is not used (set to zero by the assembler).

Register remapping is supported within the REPEAT instruction, allowing REPEAT to access the move 'window' of the register without rotating the loop. This is explained in more detail below.

The destination operand has the following 7 bit format.

There are ten variations of this default encoding.

The register number Dx indicates which of the 16 registers is addressed. The Hi / Lo bit and the size bit are a pair of 16-bit registers that work simultaneously with addressing each 32-bit register. Regardless of whether the result is written to the register bank and / or output FIFO, this size bit defines how the appropriate flag is set, as defined by the instruction form. This allows the construction of comparisons and the construction of similar commands. Additions with the accumulating class of instructions must write the result back to the register.

The following table shows the operation of each encoding.

Encoding Register write FIFO record V FLAG One Record all registers No record 32-bit overflow 2 Record all registers 32-bit recording 32-bit overflow 3 Writes Low 16 Bits to Dx.1 No record 16-bit overflow 4 Writes Low 16 Bits to Dx.1 Low 16-bit recording 16-bit overflow 5 Write low 16 bits to Dx.h No record 16-bit overflow 6 Write low 16 bits to Dx.h Low 16-bit recording 16-bit overflow 7 No record No record 16-bit overflow 8 No record No record 32-bit overflow 9 No record Low 16-bit recording 16-bit overflow 10 No record 32-bit recording 32-bit overflow

In all cases, the result of any operation before returning to the register and writing or inserting into the output FIFO is a 48-bit quantity. There are two cases:

If the recording is 16 bits, the 48 bit amount is reduced to the 16 bit amount by selecting the lower 16 bits [15: 0]. If this command saturates, the value will saturate within the range -2 ^ 15-2 ^ 15-1. The 16-bit value is written back to the indicated register and, if the write FIFO bit is set, writes back to the output FIFO. When a 16 bit value is written to the output FIFO, when the value is placed in pairs within the output FIFO as a single 32 bit value, it is retained until the next 16 bit value is written.

For 32-bit recording, the 48-bit amount is reduced to the 32-bit amount by selecting the lower 32 bits [31: 0].

For 32-bit and 48-bit writes, if the command saturates, the 48-bit value will be converted to 32 bits in the range -2 ^ 31-1 to 2 ^ 31. After being saturated, i.e.

If it is written back to the accumulator, all 48 bits will be recorded.

When writing back to the 32-bit register, bits [31: 0] are written.

When writing back to the output FIFO, bits [31: 0] will be written again.

The destination size is specified in the assembler by a.l or .h after the register number. If no register is written back, the register number is not important, so omit the destination register to indicate no write to the register, or use ^ to indicate write only to the output FIFO. For example, SUB, X0, Y0 is equal to CMP, X0, Y0 and ADD ^, and X0, Y0 places the value of X0 + Y0 in the output FIFO.

If there is no space in the output FIFO for one value, Piccolo does not wait for the space to become available.

If a 16-bit value writes, for example, ADD X0.h ^, X0, X2, the value is latched until the second 16-bit value is written. The two values are combined and placed in the output FIFO as a 32-bit number. The first 16-bit value recorded appears in the lower half of the 32-bit word. Data input into the output FIFO is represented by either 16-bit data or 32-bit data, so that endianness can be modified on large endian systems.

If a 32-bit value is written between two 16-bit writes, the operation is undefined.

Register remapping is supported within the REPEAT instruction so that REPEAT can access the move 'window' of the register without rotating the loop. This is explained in more detail below.

In a preferred embodiment of the present invention, the REPEAT instruction provides a mechanism for changing the manner in which register operands are specified in one loop. Under this mechanism, the register to be accessed is determined by the function of the register operand in the instruction and the offset in the register bank. This offset is changed in a programmable manner, preferably at the end of each instruction loop. This mechanism will work regardless of the registers present in the X, Y and Z banks. In the preferred embodiment, this facility is not available for registers in the A bank.

The concept of logical and physical registers can be used. The instruction operands are logical register references, which are mapped to physical register references that identify a particular Piccolo register 10. All operations, including refills, operate on physical registers. Only register remapping occurs on the Piccolo instruction stream side, and the data loaded into Piccolo is always set to a physical register, and no remapping is done.

The remapping mechanism will be described with reference to FIG. 5, which is a block diagram illustrating a number of internal components of the Piccolo coprocessor 4. The data item retrieved by the ARM core 2 from the memory is placed in the order buffer 12, and the Piccolo register 10 is refilled from the order buffer 12 in the manner previously described with reference to FIG. Piccolo instructions stored in cache 6 are passed to instruction decoder 50 in Piccolo 4, where these Piccolo instructions are decoded before being passed to Piccolo processor core 54. Piccolo processor core 54 includes multiplication / addition circuit 20, accumulation / cumulative disposal circuit 22, and scale / saturation circuit 24 previously described with reference to FIG.

If the instruction decoder 50 is processing instructions that form part of the instruction loop identified by the REPEAT instruction, and the REPEAT instruction has indicated that remapping of multiple registers should occur, register remapping logic 52 is disabled. To perform the necessary remapping. Although it will be apparent to those skilled in the art that the register remapping logic 52 will be provided to the instruction decoder 50 as a completely separate entity, the register remapping logic 52 may be implemented by the instruction decoder 50. It can be considered part.

The instruction will typically include one or more operands that identify the register containing the data item requested by this instruction. For example, a representative instruction would include two source operands and one destination operand, identifying two registers containing the data items requested by the instruction and either register in which the result of the instruction should be placed. Register remapping logic 52 receives operands of either instruction from instruction decoder 50, which identify logical register references. Based on this logical register reference, register remapping logic will determine whether or not lapping should be applied and apply remapping to the physical register reference as required. If it is determined that remapping should not be applied, a logical register reference is provided as the physical register reference. The preferred method of remapping will be described in more detail later.

Each physical register reference output from the register remapping logic is passed to the Piccolo coprocessor core 54 so that the processor core can apply instructions to data items in the special register 10 identified by the physical register reference.

The remapping mechanism of the preferred embodiment allows each bank of registers to be separated into two parts, one where the registers are remapped, and one where the registers hold their original register references without remapping. In a preferred embodiment, this remapped portion starts at the bottom of the register bank being remapped.

The remapping mechanism uses a number of parameters, and these parameters will be described with reference to FIG. 6, which is a block diagram illustrating that the register remapping logic 52 uses various parameters. These parameters are given values associated with a point in the bank that is being remapped, which is, for example, the bottom of the bank.

Register remapping logic 52 may be considered to have two main logical blocks, namely remap block 56 and base update block 58. Register remapping logic 52 uses a base pointer that provides an offset value that should be added to the logical register reference, which base pointer value is provided to remap block 56 by base update block 58.

The BASESTART signal can be used to define the initial value of the base pointer, even if some other value is specified, which is typically 0, for example. This BASESTART signal is sent to the multiplexer 60 in the base update block 58. On the first iteration of the instruction loop, the BASESTART signal is transmitted to the memory element 66 by the multiplexer, but for the next iteration of the loop, the next base pointer value is sent to the memory element 66 by the multiplexer 60. Supplied.

The output of the storage element 66 is passed to the remap logic 56 as the current base pointer value and to one of the inputs of the adder 62 in the base update logic 58. Adder 62 also receives a BASEINC signal providing a base increment value. This adder 62 is arranged to increase the current base pointer value supplied by the storage element 66 by the BASEINC value, and to deliver the result to the modulo circuit 64.

The modulo circuit also receives the BASEWRAP value and compares this value with the output base pointer signal from adder 62. If the incremented base pointer value is equal to or exceeds the BASEWRAP value, the new base pointer is overwritten with the new offset value. The output of the modulo circuit 64 is the next base pointer value to be stored in the memory element 66. This output is provided to the multiplexer 60 and is provided from the multiplexer 60 to the storage element 66.

However, this next base pointer value cannot be stored in memory 66 until the BASEUPDATE signal is received by memory 66 from the loop hardware processing the REPEAT instruction. The BASEUPDATE signal will be generated periodically by the loop hardware, eg every time the instruction loop is repeated. When the BASEUPDATE signal is received by memory 66, the memory will overwrite the previous base pointer value with the next base pointer value provided by multiplexer 60. As such, the base pointer value provided to the ReMap logic 58 will be changed to the new base pointer value.

The physical register to be accessed in the remapped portion of the register bank is determined by the addition of the logical register reference contained within the operand of the instruction and the base pointer value provided by the base update logic 58. This addition is done by an adder and the output is passed to the modulo circuit 70. In a preferred embodiment, the modulo circuit 70 also receives a register overlap value, and if the output signal from the adder 68 (addition of logical register reference and base pointer value) exceeds the register overlap value, the result is Will overlap to the bottom of the remapped region. The output of the modulo circuit 70 is provided to the multiplexer 72.

The REGCOUNT value is provided to logic 74 in Remap block 56 that identifies the number of registers in the bank to be remapped. Logic 74 compares this REGCOUNT value with a logical register reference and transmits a control signal to multiplexer 72 depending on the comparison result. Multiplexer 72 receives as its two inputs a logical register reference and an output from modulo circuit 70 (a remapped register reference). In a preferred embodiment of the present invention, if the logical register reference is less than the REGCOUNT value, logic 74 instructs the multiplexer to output the remapped register reference as a physical register reference. However, if the logical register reference is greater than or equal to the REGCOUNT value, logic 74 instructs multiplexer 72 to output the logical register reference directly as a physical register reference.

As previously explained, in the preferred embodiment, this is a REPEAT command that appeals to the rewrapping mechanism. As described in more detail later, the REPEAT instruction provides four zero cycle loops in hardware. These hardware loops are illustrated in FIG. 5 as part of the instruction decoder 50. Each time the instruction decoder 50 requests an instruction from the cache 6, the cache returns the instruction to the instruction decoder, so that the instruction decoder determines whether the returned instruction is a REPEAT instruction. If so, configure one of the hardware loops to process the REPEAT instruction.

Each repeat instruction traverses the loop (either constant or read from the Piccolo register) by specifying the number and number of instructions in the loop. Two opcodes REPEAT and NEXT are provided to define a hardware loop, and the NEXT opcode is simply used as a delimiter and is not assembled as either instruction. Because REPEAT is placed at the beginning of the loop, and NEXT delimits the end of the loop, the assembler can calculate the number of instructions in the loop body. In a preferred embodiment, the REPEAT instruction includes remapping parameters such as REGCOUNT, BASEINC, BASEWRAP and REGWRAP parameters that the register remapping logic 52 should use.

Multiple registers may be provided to store the remapping parameters used by the register remapping logic. Within these registers, a number of predetermined sets of remapping parameters can be provided, but the user leaves some registers for storage of defined remapping parameters. If the remapping parameter specified in the REPEAT instruction is equal to one of the predetermined set of remapping parameters, then the appropriate REPEAT encoding is used, whereby the multiplexer or the like provides the appropriate remapping parameter with register remapping logic directly from this register. On the other hand, if the remapping parameter is not equal to any of the predetermined set of remapping parameters, the assembler issues a Remapping Parameter Move Instruction (RMOV) that allows the user to configure the prescribed register remapping parameter. Will occur, and this RMOV instruction comes before the REPEAT instruction. Preferably, the user-defined remapping parameters will be placed by RMOV instructions in registers placed next to those user-defined remapping parameters, and the multiplexer passes the contents of these registers to register remapping logic. Will be programmed to.

In a preferred embodiment, the REGCOUNT, BASEINC, BASEWARP and REGWRAP parameters have one of the values identified in the chart below.

parameter Explanation REGCOUNT It identifies and remaps the number of 16-bit registers and has values 0, 2, 4, and 8. Registers below REGCOUNT are remapped, and registers above REGCOUNT or such as REGCOUNT are accessed directly. BASEINC This defines how many 16-bit registers the base pointer is incremented at the end of each loop iteration. In a preferred embodiment, it has a value of 1, 2 or 4, although in fact it may have other values, including negative values where appropriate. BASEWRAP This determines the upper limit of the base calculation. The base overlap coefficients have values 2, 4, and 8. REGWRAP This determines the upper limit of the remap calculation. The register overlap coefficients have values 2, 4, and 8. REGWRAP will be selected to equal REGCOUNT.

Returning to FIG. 6, an example of how the various parameters are used by remap block 56 is as follows (in this example, logical and physical register values are associated with a 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, the update to the next base pointer is made by the base update logic 58 before the iteration of the next loop begins.

Base = (Base + BASEINC) MOD BASEWRAP

At the end of the remapping loop, register remapping will be switched, after which all registers will be accessed as physical registers. In a preferred embodiment, only one remapping REPEAT will be active at any time. The loop can still be nested, but only one can update the remapping variable at any particular time. However, it will be appreciated that if desired, remapping iterations can be nested.

To illustrate the benefits obtained with respect to code density as a result of using the remapping mechanism according to a preferred embodiment of the present invention, a representative block filter algorithm will be described. First, the principle of the block filter algorithm will be described with reference to FIG. As shown in Fig. 7, the accumulator register A0 is arranged to accumulate the results of a plurality of multiplication operations, which multiplication of the data item d0 and the coefficient c0, the multiplication of the data item d1 and the coefficient c1, the data item d2 and Multiplication of the coefficient c2; Register A1 accumulates the result of a set of similar multiplication operations, but at this time the set of coefficients is shifted so that c0 is multiplied by d1, c1 is multiplied by d2, and c2 is multiplied by d3. Similarly, register A2 accumulates the result of multiplying the coefficient value and the data value by shifting another step to the right, so co is multiplied by d2, and c1 is multiplied by d3, and c2 is multiplied by d4. This shift, multiplication, and accumulation process is repeated and the result is placed in register A3.

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

; Start with 4 new data values

ZERO {A0-A3}; Zero the accumulator.

REPEAT Z1; Z1 = (number of coefficients / 4)

; Initially, the following four coefficients are processed:

; a0 + = d0 * c0 + d1 * c1 + d2 * c2 + d3 * c3

; a1 + = d1 * c0 + d2 * c1 + d3 * c2 + d4 * c3

; a2 + = d2 * c0 + d3 * c1 + d4 * c2 + d5 * c3

; a3 + = d3 * c0 + d4 * c1 + d5 * c2 + d6 * c3

MULA A0, X0.l ^, Y0.l, A0; a0 + = d0 * c0, and d4 load

MULA A1, X0.h, Y0.l, Al; a1 + = d1 * c0

MULA A2, X1.l, Y0.l, A2; a2 + = d2 * c0

MULA A3, X1.h, Y0.l ^, A3; a3 + = d3 * c0, and c4 load

MULA A0, X0.h ^, Y0.h, A0; a0 + = load d1 * c1 and d5

MULA A1, X1.l, Y0.h, A1; a1 + = d2 * c1

MULA A2, X1.h, Y0.h, A2; a2 + = d3 * c1

MULA A3, X0.l, Y0.h ^, A3; a3 + = d4 * c1, and c5 load

MULA A0, X1.l ^, Y1.l, A0; load a0 + = d2 * c2 and d6

MULA A1, X1.h, Y1.l, A1; a1 + = d3 * c2

MULA A2, X0.l, Y1.l, A2; a2 + = d4 * c2

MULA A3, X0.h, Y1.l ^, A3; a3 + = d5 * c2, and c6 load

MULA A0, X1.h ^, Y1.h, A0; a0 + = load d3 * c3 and d7

MULA A1, X0.l, Y1.h, A1; a1 + = d4 * c3

MULA A2, X0.h, Y1.h, A2; a2 + = d5 * c3

MULA A3, X1.l, Y1.h ^, A3; load a3 + = d6 * c3 and c7

NEXT

In this example, the data value is placed in the X bank of the register and the count value is placed in the Y bank of the register. As a first step, four accumulator registers A0, A1, A2 and A3 are set to zero. Once the accumulator registers are reset, an instruction loop is input and bounded by the REPEAT and NEXT instructions. The value Z1 identifies the number of times the instruction loop should be repeated, and for reasons described later, this will actually be equal to the coefficient (c0, c1, c2, etc.) / 4.

The instruction loop has 16 multiply accumulate instructions (MULAs), which are registers A0 and A1 that contain the result of the computation shown in the code between REPEAT and the first MULA instruction after the first iteration through the loop. , A2, A3. To illustrate how the multiply accumulate instructions work, we will consider the first four MULA instructions. The first instruction multiplies the data value in the first or lower 16 bits of the X bank register zero by the lower 16 bits in the Y bank register zero and adds the result to the accumulator register A0. At the same time, the lower 16 bits of the X bank register zero are marked with refill bits, which indicates that some of the registers can be refilled with new data values. As is apparent from FIG. 7, once data item d0 has been multiplied by the coefficient c0 (this is indicated by the first MULA instruction), d0 is no longer required for the remainder of the block filter instruction, so it is replaced by a new data value. Can be.

The second MULA instruction multiplies the second or upper 16 bits of the X bank register zero by the lower 16 bits of the Y bank register zero (this represents d1 × c0 shown in FIG. 7). Similarly, the third and fourth MULA instructions represent multiplications d2xc0 and d3xc0, respectively. As shown in Fig. 7, once these four calculations have been made, the coefficient C0 is no longer required, so the register Y0.l can be marked with the refill bit, and overwritten with another coefficient c4.

The next four MULA instructions represent the calculations d1 × c1, d2 × c1, d3 × c1, and d4 × c1, respectively. Once the calculation d1xc1 has been performed, the register X0.h is indicated by the refill bit because d1 is no longer required. Similarly, once all four calculations have been made, the register Y0.h is marked for recharging because the coefficient c1 is no longer required. Similarly, the next four MULA instructions correspond to calculations d2 × c2, d3 × c2, d4 × c2, d5 × c2, and the last four calculations correspond to calculations d3 × c3, d4 × c3, d5 × c3, d6 × c3. Corresponds.

In the above-described embodiments, since the registers cannot be remapped, each multiplication operation must be explicitly reproduced with the specific register required in the operand. Once 16 MULA instructions have been issued, the instruction loop can be repeated for coefficients c4 to c7 and data items d4 to d10. Also, since this loop acts on four coefficient values per iteration, the number of coefficient values must be a multiple of four, and the calculation Z1 = the number of coefficients / 4 must be calculated.

By using the remapping mechanism according to the preferred embodiment of the present invention, the instruction loop can be dramatically reduced, including four multiply accumulate instructions rather than the sixteen multiply accumulate instructions that were otherwise required. Using the remapping mechanism, the code can be written as follows.

; Start with 4 new data values

Zero {A0-A3}; Zero the accumulator.

Repeating Z1, X ++ n4 w4 r4, Y ++ n4 w4 r4; Z1 = (number of coefficients)

; Remapping is applied to the X and Y banks.

; Four 16-bit registers in these banks are remapped.

; The base points of both banks are increased by one for each.

; Loop iteration

; When the base pointer reaches four registers in the bank,

; Base pointers overlap

MULA A0, X0.l ^, Y0.l, A0; a0 + = d0 * c0, and d4 load

MULA A1, X0.h, Y0.l, A1; a1 + = d1 + c0

MULA A2, X1.l, Y0.l, A2; a2 + = d2 + co

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

NEXT; Iterate over the loop and remap it.

As before, the first step will set four accumulator registers A0-A3 to zero. Then, an instruction loop is input and bounded by the REPEAT and NEXT opcodes. This REPEAT instruction has a number of parameters associated with it, and these parameters are as follows.

X ++: BASEINC indicates '1' for the X bank of registers.

n4: REGCOUNT is '4', thus indicating that the first four X bank registers X0.l to X1.h are 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 ++: Indicates that BASEINC is '1' for the Y bank of registers.

n4: REGCOUNT is '4', thus indicating that the first four Y bank registers Y0.l to Y1.h are 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.

Note that the value Z1 is equal to the number of coefficients, rather than the number of coefficients / 4, as in the conventional example.

For the first iteration of the instruction loop, the base pointer value is zero, so there is no remapping. However, in the next loop execution, since the base pointer value is '1' for the X and Y banks, the operand will be mapped as follows.

X0.l becomes X0.h.

X0.h becomes X1.l.

X1.l becomes X1.h.

X1.h becomes X0.l (since BASEWRAP is '4').

Y0.l becomes Y0.h.

Y0.h becomes Y1.l.

Y1.l becomes Y1.h.

Y1.h becomes Y0.l (since BASEWRAP is '4').

Thus, for the second iteration, the four MULA instructions actually perform the calculation indicated by the fifth to eighth MULA instructions in the previously described example that do not actually include the remapping of the present invention. Similarly, the third and fourth iterations through the loop perform calculations previously made by the ninth through twelfth, and thirteenth through sixteenth MULA instructions of the conventional code.

Thus, since only four instructions should be provided rather than the sixteen previously required, it is understood that the code performs exactly the same block filter algorithm as the conventional code, but improves the code density in the loop body by four factors. Able to know.

By using the register remapping technique according to the preferred embodiment of the present invention, the following advantages can be realized.

This improves code density.

2. This can hide the latency from indicating that the register is empty under certain circumstances to the register being refilled by the Piccolo recorder buffer. This can be accomplished without traversing the loop at the cost of increased code size.

3. This allows a variable of the register to be accessed, and by changing the number of loop iterations done, the number of registers accessed may be changed.

4. This can facilitate algorithm development. For a suitable algorithm, the programmer can generate a portion of the code for the nth step of the algorithm and then apply that approach to an uncertain set of data using register remapping.

It will be apparent that certain changes may be made to the register remapping mechanism described above without departing from the scope of the present invention. For example, a bank of registers 10 may provide more physical registers than can be specified by the programmer in the instruction operand. These extra registers cannot be accessed directly, but the register remapping mechanism can make these registers available. For example, consider the above example where the X bank of registers has four 32-bit registers available to the programmer, so that eight 16-bit registers can be specified by physical register references. The X bank of registers may actually consist of, for example, six 32-bit registers, in which case there will be four additional 16-bit registers that are not directly accessible to the programmer. However, these extra four registers may be made available by the remapping mechanism, thereby providing additional registers for storage of data items.

The following assembler syntax will be used.

>> means logical right shift or left shift when the shift operand is negative (see <lscale> below).

->> means an arithmetic right shift or left shift when the shift operand is negative (see <scale> below).

ROR means Rotate Right.

SAT (a) means the saturated value of a (depending on the size of the destination register

16 or 32 bit saturated value). Specifically, to saturate to 16 bits, + 0x7fff

Any value greater than is replaced by + 0x7fff, and any value less than -0x8000

Replaced by -0x8000 Saturation to 32 bits results in the final + 0x7fffffff and

Similar to -0x80000000. If the destination register is 48 bits, it is still 32 bits

Saturated at

Source operand 1 may be in one of the following formats:

<srcl> is a shorthand for [Rn: Rn.l: Rn.h: Rn.x] [^]. In other words,

All 7-bit source specifiers are valid and registers are (optionally exchanged)

Read as a 32-bit value or a signed extended 16-bit value. About accumulator

Only the lower 32 bits are read. ^ Stands for register recharge.

<srcl_16> is a short form of [Rn.l: Rn.h] [^]. Only 16 bit values can be read.

<srcl_32> is shorthand for [Rn: Rn.x] [^]. Only 32-bit value exchanged arbitrarily

And the bottom half.

Source operand 2 may be in one of the following formats:

<src2> will be a shorthand for the three options below.

_ Of the form [Rn: Rn.l: Rn.h: Rn.x] [^] plus the scale of the final result

Source register.

_ Arbitrarily shifted 8-bit constant (<immed_8>), without scale of the final result.

6-bit constant (<immed_6>), plus the scale of the final result (<scale>).

<src2_maxmin> is the same as <src2>, but no scaling is performed.

The <src2_shift> shift instruction provides a limited subset of <src2>. Detailed

See above.

<src2_par> is as far as <src2_shift> is concerned

For an instruction specifying a third operand:

<acc> is an abbreviation of any of four accumulator registers [A0: A1: A2: A3].

All 48 bits are read. No recharge can be specified.

The destination register has the following format.

<dest> is shorthand for [Rn: Rn.l: Rn.h: .l:] [^]. "." Without expansion, full cash

The master is written (48 bits for accumulator). How to get back to the register

If no writing is required, the register used is not important. Assembling

The blur is a number to indicate that it is not required to go back and write

New register, or writing back is not required, but the result is 16

Generate a ".l" to indicate that the flag should be set even if the bit amount is

Omit. ^ Indicates that value is written to the output FIFO.

<scale> represents a number of arithmetic scales. There are 14 scales available

All.

ASR # 0, 1, 2, 3, 4, 6, 8, 10

ASR # 12-16

LSS # 1

<immed_8> represents an unsigned 8-bit immediate. This is 0, 8, 16 or 24

It consists of bytes cycled to the left by a shift. Thus, the value 0xYZ000000,

0x00YZ0000, 0x0000YZ00 and 0x000000YZ can be encoded for any YZ

All. This cycle is encoded as a 2-bit quantity.

<imm_6> represents an unsigned 6-bit immediate.

<PARAMS> is used to specify register remapping and has the following format:

<BANK> <BASEINC> n <RENUMBER> w <BASEWRAP>

<BANK> can be [X: Y: Z].

<BASEINC> may be [++: + 1: +2: +4].

<RENUMBER> may be [0: 2: 4: 8].

<BASEWRAP> may be [2: 4: 8].

The expression <cond> is a shorthand for any of the following condition codes. Note that encoding is slightly different from ARM because unsigned LS and HI codes have been replaced by more useful sign overflow / underflow tests. The V and N flags are set for Piccolo differently than ARM, so the conversion to flag task in conditional testing is not the same as ARM.

0000 EQ Z = 0 The final result is 0.

0001 NE Z = 1 Last result is not zero.

0010 CS C = 1 Use after shift / MAX operation.

0011 CC C = 0

0100 MI / LT N = 1 Final result is negative.

0101 PL / GE N = 0 The final result is positive.

0110 VS V = 1 Saturation for sign overflow / final result.

0111 VC V = 0 No overflow / saturation for final result.

1000 VP V = 1 & N = 0 Positive overflow for final result.

1001 VN V = 1 & N = 1 Negative overflow for the final result.

1010 pending

1011 pending

1100 GT N = 0 & Z = 0

1101 LE N = 1: Z = 1

1110 AL

1111 hold

Since Piccolo handles the sign quantity, the unsigned LS and HI conditions have been replaced with VP and VN describing the direction of some overflow. Because the result of the ALU is 48 bits, the MI and LT perform the same function similarly to the PL and GE. This leaves three slots for expansion of the pending.

If not indicated otherwise, all actions are signed.

The first and second condition codes consist of the following:

N-minus

Z-zero

C-carry / unsigned overflow

V-sign overflow

Operation instructions can be separated into two forms: parallel and 'full width'. Only the 'full width' instruction sets the first flag, and parallel operation sets the first and second flags based on the top and bottom 16-bit halves of the result.

The N, Z and V flags are calculated based on the results of all ALUs after they are applied except before the scale is written to the destination. ASR will always reduce the number of bits required to store the result, but ASL will increase it. To avoid this piccolo truncation, when ASL scale is applied, 0 detection and overflow should be performed on the 48 bit result to limit the number of bits.

The N flag is calculated and estimation code arithmetic is performed. 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 or not the input operand is signed.

The V flag indicates that if there is a loss of accuracy, the loss occurs as a result of writing the result to the selected destination. If no rewriting is selected, the 'size' is still involved and the overflow flag is set correctly. Overflow is

-When the result is not in the range of -2 ^ 15 to 2 ^ 15-1, writing to a 16-bit register.

Occurs when writing to a 32-bit register when the result is not in the range -2 ^ 31 to 2 ^ 31-1.

The parallel add / sub instruction sets the N, Z, and V flags independently of the top and bottom half of the result.

When writing to the accumulator, the V flag is set as if writing to a 32-bit register. This allows the saturation instruction to use the accumulator as a 32-bit register.

The Saturation Absolute Instruction (SABS) also sets an overflow flag if the absolute value of the input operand matches the specified destination.

The carry flag is set by the add and subtract instructions, and the MAX / MIN. Used as a 'binary' flag by SABS and CLB instructions. All other instructions, including multiplication operations, preserve the carry flag.

For add and subtract operations, the carry is generated by either bit 31 or bit 15 or the result, based on whether the destination is 32 bits or 16 bits.

The standard arithmetic instructions can be divided into numbers according to the flag setting method.

In the case of add and subtract instructions, if N bits are set, all flags are preserved. If the N bit is not set, the flag is updated as follows.

Z is set when all 48-bit results are zero.

N is set when all 48-bit results have 47 sets of bits (negative).

V is either the destination register is 16 bits and the sign result does not fit within the 16 bit register (not in the range -2 ^ 15 <= x <2 ^ 15), or the destination register is a 32/48 bit register. It is set when the sign result does not fit within 32 bits.

If <dest> is a 32 or 48-bit register, the C flag borrows from bit 31 if there is a carry of bit 31 when adding <src1> and <src2> or subtracting <src2> from <src1>. If neither occurs, it is set (you will expect the same carry value for the ARM). If <dest> is a 16-bit register, the C flag is set if there is a carry of bit 15 of the sum.

The second flag (SZ, SN, SV, SC) is saved.

For these instructions, multiply or accumulate from a 48-bit register.

Z is set when all 48-bit results are zero.

N is set when the entire 48-bit result has 47 sets of bits (negative)

V is either (1) if the destination register is 16 bits and the sign result does not fit within the 16-bit register (not in the range -2 ^ 15 <= x <2 ^ 15), or (2) the destination register is It is a 32 / 48-bit register and is set when the sign result does not fit 32 bits.

C is preserved.

The second flag SZ, SN, SV, SC is saved.

Other instructions, including logic operations, parallel addition and subtraction, maximum and minimum, shift, and the like, are included below.

Add and subtract instructions add or subtract two registers, scale the result, and then return to the register and store it. Operands are treated as signed values. Flag updates for unsaturated variants are optional and will be suppressed by adding N to the end of the instruction.

OPC specifies the type of command.

Action (OPC):

100N0 dest = (src1 + src2) (->> scale) (, N)

110N0 dest = (src1-src2) (->> scale) (, N)

10001 dest = SAT ((src1 + src2) (->> scale))

11001 dest = SAT (src1-src2) (->> scale)

01110 dest = (src2-src1) (->> scale)

01111 dest = SAT ((src2-src1) (->> scale))

101N0 dest = (src1 + src2 + Carry) (->> scale) (, N)

111N0 dest = (src1-src2 + Carry-1) (->> scale) (, N)

Reminiscent:

100N0 ADD {N} <dest>, <src1>, <src2> {, <scale>}

110N0 SUB {N} <dest>, <src1>, <src2> {, <scale>}

10001 SADD <dest>, <src1>, <src2> {, <scale>}

11001 SSUB <dest>, <src1>, <src2> {, <scale>}

01110 RSB <dest>, <src1>, <src2> {, <scale>}

01111 SRSB <dest>, <src1>, <src2> {, <scale>}

101N0 ADC {N} <dest>, <src1>, <src2> {, <scale>}

111N0 SBC {N} <dest>, <src1>, <src2> {, <scale>}

The assembler supports the following opcodes:

CMP <src1>, <src2>

CMN <src1>, <src2>

CMP is a subtraction that sets a flag where register write is not possible. The CMN is an addition for setting a flag in which register writing is not possible.

flag:

These are described above.

Reasons for inclusion:

The ADC is useful for inserting a carry into the bottom of a register following shift / MAX / MIN operation. It is also used for 32/32 bit division. It also prepares for extended precision additions. The addition of N bits represents finer control of the flag, in particular carry. This allows 32/32 bit division in 2 cycles per bit.

Saturated addition and subtraction are required for G.729, etc.

Increasing / decreasing the counter, RSB is useful for calculating shifts (X = 32-x is a common operation). Saturated RSB is required for saturated negation (used in G729).

Add / Subtract Accumulation instructions perform addition and subtraction with accumulation and scaling / saturation. Unlike the multiply accumulate instructions, the accumulator number cannot be specified regardless of the destination register. The lower two bits of the destination register indicate the number of acc accumulators to accumulate within the accumulator. Thus, ADDA X0, X1, X2, A0 and ADDA A3, X1, X2, A3 are valid, but ADDA X1, X1, X2, A0 are invalid. By the class of this instruction, the result must be written back to the register and no writeback encoding of any destination field is allowed.

OPC specifies the type of command. In the following acc is (DEST [1: 0]). The Sa bit represents saturation.

Action (OPC):

0 dest = {SAT} (acc + (src1 + src2)) {->> scale}

1 dest = {SAT} (acc + (src1-src2)) {->> scale}

Reminiscent:

0 {S} ADDA <dest>, <src1>, <src2>, <acc> {, <scale>}

1 {S} SUBA <dest>, <src1>, <src2>, <acc> {, <scale>}

S before the command indicates saturation.

flag:

See above

Reasons for inclusion:

The Add Accumulate (ADDA) instruction is useful to sum two words of an array of integers with an accumulator (eg, to find their average) every cycle. The SUBA (Subtraction Accumulation) instruction is useful for calculating the sum of differences (for correlations), and it subtracts two divided values and adds the difference with the third register.

Addition can be done with rounding by using <dest> different from <acc>. For example, X0 = (X1 + X2 + 16384) >> 15 can be done in one cycle by keeping 16384 within A0. Addition with rounding constants can be done by ADDA X0, X1, # 16384, A0.

For the correct execution of the bits of (sum of (a_i * b_i) >> k) (completely used for Common-TrueSpeech), the standard Piccolo code is as follows.

MUL t1, a_0, b_0, ASR # k

ADD ans, ans, t1

MUL t2, a_1, b_1, ASR # k

ADD ans, ans, t2

There are two problems with this code. One is that it is too long, and the other is that there is no addition to 48-bit precision, so guard bits cannot be used. A better solution is to use ADDA.

MUL t1, a_0, b_0, ASR # k

MUL t2, a_1, b_1, ASR # k

ADDA ans, t1, t2, ans

This gives a 25% speed increase and maintains 48-bit accuracy.

Addition / subtraction in parallel instructions performs addition and subtraction on two signed 16-bit quantities held in pairs in a 32-bit register. The first condition code flag is set from the result of the top 16 bits, and the second flag is updated from the bottom half. Although values can be halfword exchanged, only 32-bit registers can be specified as the source for these instructions. Each half of each register is treated as a sign value. Calculation and scaling is done without any loss of accuracy. Thus, ADDADD X0, X1, X2, ASR # 1 will produce an accurate average at the top and bottom half of X0. Random saturation is prepared for each instruction for which the Sa bit should be set.

OPC defines the behavior.

Action (OPC):

000 dest.h = (src1.h + src2.h)->> {scale},

dest.l = (src1.l + src2.l)->> {scale}

001 dest.h = (src1.h + src2.h)->> {scale},

dest.l = (src1.l-src2.l)->> {scale}

100 dest.h = (src1.h-src2.h)->> {scale},

dest.l = (src1.l + src2.l)->> {scale}

101 dest.h = (src1.h-src2.h)->> {scale},

dest.l = (src1.l-src2.l)->> {scale}

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

Reminiscent:

000 {S} ADDADD <dest>, <src1_32>, <src2_32> {, <scale>}

001 {S} ADDSUB <dest>, <src1_32>, <src2_32> {, <scale>}

100 {S} SUBADD <dest>, <src1_32>, <src2_32> {, <scale>}

101 {S} SUBSUB <dest>, <src1_32>, <src2_32> {, <scale>}

S before the command indicates saturation.

The assembler also supports the following generated by standard instructions without going back and writing.

CMNCMN <dest>, <src1_32>, <src2_32> {, <scale>}

CMNCMP <dest>, <src1_32>, <src2_32> {, <scale>}

CMPCMN <dest>, <src1_32>, <src2_32> {, <scale>}

CMPCMP <dest>, <src1_32>, <src2_32> {, <scale>}

flag:

C is set by performing bit 15 when adding two upper 16 bit halves.

Z is set if the sum of the top 16 bits half is zero.

N is set if the upper 16 bits half is negative.

V is set if the sum of the sign 17 bits of the top 16 bits half does not fit within 16 bits.

(Post-scale).

SZ, SN, SV, and SC are set similarly for the lower 16 bit half.

Reasons for inclusion:

Parallel add and subtract instructions are useful for performing operations on complex numbers held in 32-bit registers. These commands are used in the FFT kernel. It is also useful for simple addition / subtraction of vectors of 16-bit data so that two devices can be processed cycle by cycle.

Branch (conditional) instructions allow conditional changes in the control flow.

Piccolo takes three cycles to execute the acquired branch.

action:

Branch by offset if <cond> holds according to the first flag.

This offset is the 16-bit number of words in the word. The current offset ranges from -32768 to

Limited to +32767 words.

The address operation performed

Target address = branch instruction address + 4 + OFFSET.

Reminiscent:

B <cond> <destination_label>

flag:

Not affected

Reasons for inclusion:

Very useful in most routines.

Conditional addition or subtraction instructions conditionally add or subtract src2 through src1.

OPC specifies the type of command.

Action (OPC):

0 (carrying setting) temp = src1 + src2 if other temp = src1-src2

dest = temp {->> scale}

1 (carrying setting) temp = src1 + src2 if other temp = src1-src2

dest = temp {->> scal} But if the scale is left shifted,

The new value of the carry (from src1-src2 or src1 + src2) is shifted downward.

Reminiscent:

0 CAS <dest>, <src1>, <src2> {, <scale>}

1 CASC <dest>, <src1>, <src2> {, <scale>}

flag:

See above

Reasons for inclusion:

Effective division code can be constructed by conditional addition or subtraction instructions.

Example 1: Divide a 32-bit unsigned value at X0 by a 16-bit unsigned value at X1 (under the assumption that X0 <(X1 << 16) and X1.h = 0).

LSL X1, X1, # 15; Divisor Shift Up

SUB X1, X1, # 0; Carry flag set

REPEAT # 16

CASC X0, X0, X1, LSL # 1

NEXT

At the end of the loop X0.l holds the quotient of division. The rest can be recovered from X0.h depending on the value of the carry.

Example 2: With the initial termination, divide the 32-bit positive value at X1 by the 32-bit positive value at X0.

MOV X2, # 0; Clear share

LOG Z0, X0; The number of bits X0 can be shifted.

LOG Z1, X1; The number of bits X1 can be shifted.

SUBS Z0, Z1, Z0; X1 shifts up so 1s are matched.

BLT div_end; The response is 0 since X1> X0.

LSL X1, X1, Z0; Match the main ones.

ADD Z0, Z0, # 1; The number of tests to do.

SUBS Z0, Z0, # 0; Set carry.

REPEAT Z0

CAS X0, X0, X1, LSL # 1

ADCN X2, X2, X2

NEXT

div_end

At the end, X2 holds a quotient and the rest can be recovered from X0.

Data can be normalized by a count leading bit instruction.

action:

dest is set to the number of spaces, and the value in src1 is bit 31, unlike bit 30

Should be trimmed to the left and shifted to the left. This means that src1 is -1 or 0

, A value in the range 0-30 except for the special case in which 31 is returned.

Reminiscent:

CLB <dest>, <src1>

flag:

Z is set when the result is zero.

N is cleared.

C is set when src1 is either -1 or 0.

V is preserved.

Reasons for inclusion:

Steps needed for standardization.

Stop and breakpoint instructions are provided to stop Piccolo execution.

OPC specifies the type of command.

Action (OPC):

0 Piccolo execution is stopped, stop bit is set in Piccolo Status register

Is set.

1 Piccolo execution is stopped, and the break bit is set in the Piccolo Status register.

Is set, the ARM notifies that the breakpoint has been reached

Interrupted.

Reminiscent:

0 HALT

1 BREAK

flag:

Not affected

Logic operation instructions perform logic operations on 32-bit or 16-bit registers. Operands are treated as unsigned values.

OPC encodes the logical operation to be performed.

Action (OPC):

00 dest = (src1 & src2) {->> scale}

01 dest = (src1: src2) {->> scale}

10 dest = (src1 & src2) {->> scale}

11 dest = (src1 ^ src2) {->> scale}

Reminiscent:

00 AND <dest>, <src1>, <src2> {, <scale>}

01 ORR <dest>, <src1>, <src2> {, <scale>}

10 BIC <dest>, <src1>, <src2> {, <scale>}

11 EOR <dest>, <src1>, <src2> {, <scale>}

The assembler supports the following opcodes:

TST <src1>, <src2>

TEQ <src1>, <src2>

TST is an AND that cannot be written to the register. TEQ is a register not writeable EOR.

flag:

Z is set when the result is all zeros.

N, C, and V are preserved.

SZ, SN, SC, SV are preserved.

Reasons for inclusion:

Speech compression algorithms use a full bitfield to encode information. Bit masking instructions help extract / pack these fields.

Max and Min operation commands perform the maximum and minimum operation.

OPC specifies the type of command.

Action (OPC):

0 dest = (src1 <= src2)? src1: src2

1 dest = (src1> src2)? src1: src2

Reminiscent:

0 MIN <dest>, <src1>, <src2>

1 MAX <dest>, <src1>, <src2>

flag:

Z is set if the result is zero.

N is set if the result is negative.

C is set for Max if src2> = src1 (if dest = src1)

For Min, it is set if src2> = src1 (if dest = src2).

V is preserved.

Reason for inclusion:

To find the strength of a signal, many algorithms scan one sample to find the minimum / maximum of the absolute value of the sample. MAX and MIN operation is very important for this. Depending on whether you want to find the first or last maximum in the signal, the operands src1 and src2 can be exchanged.

MAX X0, X0, # 0 will clip down to convert X0 to a positive number.

MIN X0, X0, # 255 will clip X0 above. This is useful for graphics processing.

Max and Min operations in parallel instructions perform maximum and minimum operations on parallel 16-bit data.

OPC specifies the type of command.

Action (OPC):

0 dest.l = (src1.l <= src2.l)? src1.l: src2.l

dest.h = (src1.h <= src2.h)? src1.h: src2.h

1 dest.l = (src1.l> src2.l)? src1.l: src2.l

dest.h = (src1.h> src2.h)? src1.h: src2.h

Reminiscent:

0 MINMIN <dest>, <src1>, <src2>

1 MAXMAX <dest>, <src1>, <src2>

flag:

Z is set if the upper 16 bits of the result are zero.

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

C is set for Max if src2.h> = src1.h (if dest = src1)

For Min, it is set if src2.h> = src1.h (for dest + src2)

V is preserved.

SZ, SN, SC, and SV are set similarly for the lower 16 bits half.

Reasons for inclusion:

About 32 bit Max and Min.

A long-travel immediate operation instruction allows the register to be set to any sign 16-bit, sign-extended value. Two of these instructions can set the 32-bit register to some value (by accessing the upper and lower halves in turn). See the select operation for moving between registers.

Reminiscent:

MOV <dest>, # <imm_16>

The assembler uses this MOV instruction to provide non-interlocking NOP operation, ie NOP equals MOV, # 0.

flag:

The flags are not affected.

Reason for inclusion:

Initialize register / counter.

The multiply accumulate operation instruction performs signed multiplication with cumulative or cumulative disposal, scaling, and saturation.

The field OPC specifies the type of instruction.

Action (OPC):

00 dest = (acc + (src1 * src2)) {->> scale}

01 dest = (acc- (src1 * src2)) {->> scale}

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

do.

Reminiscent:

00 {S} MULA <dest>. <src1_16>, <src2_16>, <acc> {, <scale>}

01 {S} MULS <dest>, <src1_16>, <src2_16>, <acc> {, <scale>}

S before the command indicates saturation.

flag:

See above section

Reasons for inclusion:

One cycle duration MULA is required for the FIR code. MULS is used for FFT butterflies. MULA is also useful for recursive multiplication. For example, A0 = (X0 * X1 + 16384) >> 15 can be done in one cycle by holding 16284 in another accumulator (eg A1). Different <dest> and <acc> are also required for the FFT kernel.

Multiplication The double operation instruction performs a signed multiplication, doubling the result before cumulative or cumulative disposal, scaling and saturation.

OPC specifies the type of command.

Action (OPC):

0 dest = SAT ((acc + SAT (2 * src1 * src2)) {->> scale})

1 dest = SAT ((acc-SAT (2 * src1 * src2)) {->> scale})

Reminiscent:

0 SMLDA <dest>, <src1_16>, <src2_16>, <acc> {, <scale>}

1 SMLDS <dest>, <src1_16>, <src2_16>, <acc> {, <scale>}

flag:

See above section

Reason for inclusion:

MLD instructions are required for G.729 and other algorithms using fractional part operations. Most DSPs provide a fractional part mode that allows one bit left shift at the output of the multiplier before accumulating or recording back. Support this when certain instructions provide more programming flexibility. Some of the G-series default behaviors have equivalent names.

L_mus => SMLDS

L_mac => SMLDA

They use the saturation of the multiplier when shifting left by one bit. Without loss of precision, if a fractional multiply-accumulate sequence is required, MULA can be used with the sum maintained in 1.15 format. If necessary, left shift and saturation can be used at the end to switch to 1.15 format.

The multiply operation instruction performs signed multiplication and arbitrary scaling / saturation. The source register (16 bits only) is considered as a signed number.

OPC specifies the type of command.

Action (OPC):

0 dest = (src1 * src2) {->> scale}

1 dest = SAT ((src1 * src2) {->> scale}

Reminiscent:

0 MUL <dest>, <src1_16>, <src2>, {, <scale>}

1 SMUL <dest>, <src1_16>, <src2> {, <scale>}

flag:

See above section

Reason for inclusion:

Many processes require signed and saturated multiplication.

The register list operation is used to perform an action on a set of registers. Empty and zero instructions are provided to reset the selection of registers before or during the routine. Provide an output command to store the contents of a list of registers into an output FIFO.

OPC specifies the type of command.

Action (OPC):

For 000 (k = 0; k <16; k ++)

If bit k in the register list is set, register k is empty

Is displayed.

For 001 (k = 0; k <16; k ++)

If bit k in the register list is set, register k contains zero

Is set.

010 Uncertain

011 indeterminate

For 100 (k = 0; k <16; k ++)

If bit k in the register list is set, (resister k->> scal) is output

Written to the FIFO.

For 101 (k = 0; k <16; k ++)

If bit k in the register list is set, (resister k->> scal) is output

Written to the FIFO, register k is marked empty.

For 110 (k = 0; k <16; k ++)

If bit k in the register list is set, the register k->> scal (SAT)

Is written to the output FIFO.

For 111 (k = 0; k <16; k ++)

If bit k in the register list is set, the register k->> scal (SAT)

Written to the output FIFO, register k is marked blank.

Reminiscent:

000 EMPTY <register_list>

001 ZERO <register_list>

010 unused

011 unused

100 OUTPUT <register_list> {, <scale>}

101 OUTPUT <register_list> ^ {, <scale>}

110 SOUTPUT <register_list> {, <scale>}

111 SOUTPUT <register_list> ^ {, <scale>}

flag:

Not affected

Yes:

EMPTY {A0, A1, X0-X3}

ZERO {Y0-Y3}

OUTPUT {X0-Y1} ^

The assembler also supports this syntax.

OUTPUT Rn

In this case, it will output one register using the MOV ^, Rn command.

The EMPTY instruction will stop until all registers that are empty contain valid data (that is, not empty).

The register list operation does not have to be used within the remapping REPEAT loop.

The OUPUT instruction can only specify up to eight registers for output.

Reason for inclusion:

After the routine is finished, the next routine receives data from the ARM, thinking that all registers are empty. The EMPTY command is needed to do this. Before performing FIR or other filters, all accumulators and partial results must be zero. The ZERO command helps this. Both are designed to improve code density by exchanging a series of single register moves. The OUTPUT instruction is included to improve code density by exchanging a series of MOV ^ and Rn instructions.

The remapping parameter move command RMOV is provided to allow the user to configure the prescribed register remapping parameters.

Command decoding is as follows.

Each PARAMS field consists of the following entries.

The meaning of these entries is described below.

parameter Explanation RENUMBER The number of 16-bit registers for remapping have values 0, 2, 4, 8. The registers below RENUMBER are remapped, and these are accessed directly. BASEINC The amount of base pointer is increased at the end of each loop. Has the value 1, 2, or 4. BASEWRAP The base overlap coefficients have values 2, 4, and 8.

Reminiscent:

RMOV <PARAMS>, [<PARAMS>]

The <PARAMS> field has the following format:

<PARAMS> :: = <BANK> <BASEINC> n <RENUMBER>

w <BASEWRAP>

<BANK> :: = [X: Y: Z]

<BASEINC> :: = [++: + 1: +2: +4]

<RENUMBER> :: = [0: 2: 4: 8]

<BASEWRAP> :: = [2: 4: 8]

If the RMOV command is used while remapping is active, the action is UNPREDICTABLE.

flag:

Not affected

The repeat instruction provides four zero cycle loops in hardware. The REPEAT instruction limits the new hardware loop. Piccolo uses hardware loop 0 for the first REPEAT instruction and hardware loop 1 for the REPEAT instruction nested within the first repeat instruction or the like. The REPEAT loop must be nested correctly. If you attempt to nest a loop at a depth greater than four, the behavior is unpredictable.

Each REPEAT instruction cycles through the loop by specifying the number and number of instructions in the loop (which immediately follows the REPEAT instruction) (either one is a constant or read from the Piccolo register).

If the number of instructions in the loop is small (1 or 2), Piccolo takes an extra cycle to set up the loop up.

When the loop count is registered, only the lower 16 bits are significant, and even though the number is assumed to be unsigned, 32 bit access is included (S1 = 1). If the loop count is zero, the action of the loop is not limited. A copy of the loop count can be used to immediately reuse (or recharge) the register without nutrition to the loop.

The REPEAT instruction provides a mechanism to change how register operands are specified in a loop. The encoding of the REPEAT by the number of register-designated loops will be described in detail.

REPEAT encoding with a fixed number of loops:

The RFIELD operand specifies which of the 16 remapping parameter constructs to use inside the loop.

RFIELD Remapping behavior 0 No remapping is done One Custom remapping 2..15 Remapping Configuration TBD Preset

The assembly provides two opcodes, REPEAT and NEXT, which define the hardware loop. Because REPEAT is placed at the beginning of the loop, and NEXT delimits at the end of the loop, the assembler can count the number of instructions in the loop body. For REPEAT, you must specify the number of loops only as either a constant or a register. For example,

REPEAT X0

MULA A0, Y0.l, Z0.l, A0

MULA A0, Y0.h ^, Z0.h ^, A0

NEXT

This will execute two MULA instructions X0 times. Also, this

REPEAT # 10

MULA A0, X0 ^, Y0 ^, A0

NEXT

10 multiplication accumulations will be performed.

The assembler specifies the following syntax.

REPEAT #iterations [, <PARAMS>]

Specifies the remapping parameter to use for REPEAT. If the required remapping parameter is equal to one of the predetermined set of parameters, then the appropriate REPEAT encoding is used. Without it, the assembler will generate an RMOV that comes before the REPEAT instruction to load the user-defined parameters. See above for details on the RMOV command and the remapping parameter format.

If the number of loop iterations is zero, the action of REPEAT is UNPREDICTABLE.

If the number of command fields is set to 0, the action of REPEAT is UNPREDICTABLE.

A loop consisting of only one instruction that is a branch will have the UNPREDICTABLE action.

Branches within the boundaries of the REPEAT loop and outside the boundaries of the loop are UNPREDICTABLE.

The saturated absolute instruction calculates the absolute value of saturated source 1.

action:

dest = SAT ((SRC1> = 0)? SRC1: -SRC1). This value is always saturated.

Specifically, 0x80000000 is 0x7fffffff and NOT 0x80000000 !.

Reminiscent:

SABS <dest>, <src1>

flag:

Z is set if the result is zero.

N is preserved.

C is set if src1 <0 (if dest = -src1)

V is set if saturation has occurred.

Reason for inclusion:

Useful for many DSP applications.

The select operation (conditional move) functions to conditionally move either source 1 or source 2 into the destination register. The choice is always like moving. There is also parallel operation for use after parallel addition / subtraction.

Both source operands are read by an instruction of the reason for execution, and if either source operand is empty, the instruction will stop regardless of whether the operand is exactly required.

OPC specifies the type of command.

Action (OPC):

00 If <cond> holds for the first flag, dest = src1

Or dest = src2.

01 If <cond> holds for the first flag, dest.h = src1.h

Or dest.h = src2.h,

If <cond> holds for the second flag, dest.l = src1.l

Or det.l = src2.l

10 If <cond> holds for the first flag, dest.h = src1.h

Or dest.h = src2.h

If <cond> fails for the second flag, dest.l = src1.l

Or dest.l = src2.l

11 hold

Mnemonic

00 SEL <cond> <dest>, <src1>, <src2>

01 SELTT <cond> <dest>, <src1>, <src2>

10 SELTF <cond> <dest>, <src1>, <src2>

11 unused

If the register is marked for recharging, the register is unconditionally recharged. The assembler also provides the following mnemonic.

MOV <cond> <dest>, <src1>

SELFT <cond> <dest>, <src1>, <src2>

SELFF <cond> <dest>, <src1>, <src2>

MOV <cond> A, B is the same as SEL <cond> A, B, A. SELFT and SELFF are obtained by using SELTF and SELTT by exchanging src1 and src2.

flag:

All flags will be saved and a sequence of selections will be made.

Reason for inclusion:

Used to simply determine inline without resorting to branches. It is used by the Viterbi algorithm and when scanning vectors for the sample or the largest device.

The shift operation instruction provides left, right logical shifts, right arithmetic shifts, and cycles by a certain amount. The shift amount is considered to be a signed integer from -128 to +127 obtained from the lower 8 bits of the register contents or an immediate value in the range of +1 to +31. Negative positive shifts cause the shift in the opposite direction by ABS (shift amount).

Since the input operand is a sign extended to 32 bits, and the resulting 32 bit output is a sign extended to 48 bits before writeback, writing to a 48 bit register works remarkably.

OPC specifies the type of command.

Action (OPC):

00 dest = (src2> = 0)? src1 << src2: src1 >> -src2

01 dest = (src2> = 0)? src1 >> src2: src1 << -src2

10 dest = (src2> = 0)? src1->> src2: src1 << -src2

11 dest = (src2> = 0)? src1 ROR src2: src1 ROL -src2

Reminiscent:

00 ASL <dest>, <src1>, <src2_16>

01 LSR <dest>, <src1>, <src2_16>

10 ASR <dest>, <src1>, <src2_16>

11 ROR <dest>, <src1>, <src2_16>

flag:

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 in ARM).

The behavior of register-specified shifts is as follows:

LSL by 32 has result 0, and C is set to bit 0 of src1.

LSL over 32 has result 0, C set to 0

LSR by 32 has result 0 and C is set to bit 31 of src1.

-LSR by more than 32 has result 0, C is set to 0.

ASR greater than or equal to 32 has the result of filling with bit 31 of src1, and C is equal to bit 31 of src1.

ROR by 32 has the same result as src1 and C is set to bit 31 of src1.

Since ROR by n greater than 32 will give the same result and will perform as ROR by n-32, we will subtract 32 from n repeatedly until the amount is in the range of 1 to 32.

Reasons for inclusion:

Multiplication / division by many 2. Bit and field extraction. Serial register.

Unqualified commands are set out in the instruction set listing. Their execution will cause Piccolo to stop execution, set the U bit in the status register, and disable itself (even if the E bit in the control register is cleared). This allows any future extension of the instruction set to be trapped and emulated arbitrarily for existing execution.

To access Piccolo state from ARM: Use the state access mode to observe / change Piccolo's state. This mechanism is provided for the following two purposes.

-Context switch

Debug

Piccolo is placed in state access mode by executing the PSTATE command. This mode allows all Piccolo states to be saved and restored in a sequence of STC and LDC instructions. Once in state access mode, the use of Piccolo coprocessor ID PICCOLO1 can be changed to access Piccolo's state. There are seven banks of Piccolo state. All data in a particular bank can be loaded and stored in a single LDC or STC.

Bank 0: Dedicated Registers

One 32-bit word containing the value of the Piccolo ID register (read only).

One 32-bit word containing the status of the control register.

One 32-bit word containing the status of the status register.

One 32-bit word containing the status of the program counter.

Bank 1: General Register (GPR).

16 32-bit words containing general register state.

Bank 2: accumulator

4 32-bit words containing the upper 32 bits of the accumulator register (GPR status and N.B.copy are necessary for restoration purposes and will involve another write enable for the register bank in other ways).

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

One 32-bit word indicating that the register is marked for refill (2 bits for each 32-bit register).

8 32-bit words (7 7-bit entries stored within bits 7-0) containing the status of the ROB tag.

Three 32-bit words (bits 17-0) containing the state of the unaligned ROB latch

One 32-bit word indicating that the slot in the output shift register contains valid data (bit 4 represents blanks and bit 3-0 encodes the number of entries used).

One 32-bit word (bit 17-0) containing the state of the output FIFO hold latch.

Bank 4: ROB Input Data.

8 32-bit data values.

Bank 5: Output FIFO Data.

8 32-bit data values.

Bank 6: Loop Hardware.

Four 32-bit words containing the loop start address.

Four 32-bit words containing the loop end address.

Four 32-bit words (bits 15-0) containing a loop count.

One 32-bit word containing user-defined remapping parameters and other remapping states.

The LDC instruction is used to load Piccolo state when Piccolo is in state access mode. The BANK field specifies that the bank is loaded.

The following sequence will load all Piccolo states from the address in register R0.

LDP B0, [R0], # 16! ; Dedicated register

LDP B1, [R0], # 64! ; Load general purpose registers

LDP B2, [R0], # 16! ; Accumulator rod

LDP B3, [R0], # 56! ; Load Register / ROB / FIFI State

LDP B4, [R0], # 32! ; Load ROB Data

LDP B5, [R0], # 32! ; Load output FIFO data

LDP B6, [R0], # 52! ; Load loop hardware

The STC instruction is used to store Piccolo state when Piccolo is in state access mode. The BANK field specifies that the bank is stored.

The following sequence will store all Piccolo states at an address in register R0.

STP B0, [R0], # 16! ; Save private register

STP B1, [R0], # 64! ; Save general purpose registers

STP B2, [R0], # 16! ; Save Accumulator

STP B3, [R0], # 56! ; Save Register / ROB / FIFO Status

STP B4, [R0], # 32! ; Save ROB Data

STP B5, [R0], # 32! ; Save output FIFO data

STP B6, [R0], # 52! ; Save loop hardware

Debug mode-Piccolo is supported by ARM and must respond to the same debug mechanism: software by Demon and Angel, and hardware with ICE embedded. There are several mechanisms for debugging the following Piccolo system.

ARM instruction breakpoint.

Data breakpoint (watchpoint).

-Piccolo command breakpoints.

-Piccolo software breakpoints.

ARM instruction and data breakpoints are handled by an 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 may be configured to breakpoint ARM and Piccolo.

Software breakpoints are handled by Piccolo instructions (stop or break) that cause Piccolo to stop execution, enter debug mode (B bits in the status register set), and (as Piccolo was disabled with the PDISABLE instruction). Disable itself. The program counter is in a valid state so that the address of the breakpoint can be restored. Piccolo will no longer execute the command.

A single stepping piccolo will be done by setting a breakpoint after a breakpoint on the Piccolo instruction stream.

Software Debug-The basic function provided by Piccolo is the ability to load and save all state of memory through coprocessor instructions when in state access mode. This allows the debugger to save all state to memory, read and / or update it, and restore it to Piccolo. The Piccolo storage state mechanism will be non-destructive, ie the action of storing Piccolo's state will not modulate any internal state of Piccolo. This means that Piccolo can be resumed after dumping its state without restoring it again.

The mechanism for finding the state of the Piccolo cache will be determined.

Hardware Debug-Hardware debug will be facilitated by a scan chain to Piccolo's coprocessor interface. Piccolo will then be in state access mode and will have the state examined / changed through the scan chain.

The Piccolo Status register contains a single bit to indicate that it has executed a breakpoint instruction. When a breakpointed instruction is executed, Piccolo sets the B bit in the status register and stops executing. To query Piccolo, the debugger must enable Piccolo and put it in state access mode by writing to its control register before the next access can occur.

4 shows the multiplexer placement in response to the Hi / Lo bit and the size bit to convert the appropriate half of the selected register into the Piccolo datapath. If the size bit represents 16 bits, then the sign extension circuit embeds the most significant bit of the datapath with 0 or 1 as appropriate.

Claims (11)

A plurality of registers storing data words to be manipulated, each having at least N bits of capacity, A data processing apparatus having an arithmetic logic device that performs arithmetic logic operations specified by the program command word in response to a program command word, The operation logic device, (i) a source register bit field indicating 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 indicating whether the input operand data word has an N bit size or an (N / 2) bit size; (iii) when the input size flag indicates an (N / 2) bit size, a high indicating that the input operand data word is located in either the highest bit position of the source register or the lowest bit position of the source register. And responsive to at least one program command word comprising a / row position flag. The method of claim 1, And an N-bit data bus for transferring a data word between the data storage device and the plurality of registers. The method of claim 2, And an input buffer for receiving a data word from said N-bit data bus and for supplying said N-bit data word to said plurality of registers. The method according to any of the preceding claims, The arithmetic logic is separate for the first (N / 2) bit input operand data word and the second (N / 2) bit input operand data word stored in the highest and lowest bit positions of only one source register, respectively. And responsive to at least one parallel operation program command word for performing arithmetic logic operations. The method of claim 4, wherein Said arithmetic logic device having a signal path functioning as a conveying chain between bit positions in an arithmetic logic operation. The method according to any one of claims 4 and 5, The parallel operation program command word, (i) parallel addition, which adds two parallel (N / 2) bits; (ii) parallel subtraction, which performs two parallel (N / 2) bit subtractions, (iii) a parallel shift for performing two parallel (N / 2) bit shift operations, (iv) an arithmetic logic operation of one of parallel addition / subtraction which performs (N / 2) bit addition and (N / 2) bit subtraction. The method according to any of the preceding claims, When the input size flag indicates an N bit size, whether these bits stored in the highest bit position should be moved to the least significant bit position, and those bits stored in the least significant bit position are used as the N bit input operand data word. And the high / low position flag indicates whether it should be moved to the highest bit position before. The method according to any of the preceding claims, And said computational logic device has an N-bit data path therethrough. The method of claim 8, Selecting to supply (N / 2) bit input operand data words stored in either the highest bit position of the source register or the lowest bit position of the source register to the lowest (N / 2) bits of the datapath. And at least one multiplexer responsive to a high / low position flag. The method according to any one of claims 8 and 9, And a circuit for sign-extending the (N / 2) -bit input operand data word before it is input to the N-bit data path. Storing data words to be manipulated in a plurality of registers each having at least N bit capacity; A data processing method comprising the step of performing an arithmetic logic operation designated by the program command word in response to a program command word, At least one program command word is (i) a source register bit field indicating 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 indicating whether the input operand data word has an N bit size or an (N / 2) bit size; (iii) when the input size flag indicates an (N / 2) bit size, a high indicating whether the input operand data word is located in one of the highest bit position of the source register and the lowest bit position of the source register; Data processing method comprising a / row position flag.