Publication number | USRE39121 E1 |
Publication type | Grant |
Application number | US 10/366,502 |
Publication date | Jun 6, 2006 |
Filing date | Feb 13, 2003 |
Priority date | Nov 29, 1996 |
Fee status | Paid |
Also published as | CN1193771A, CN1210647C, CN1512317A, CN100356316C, DE69720922D1, DE69720922T2, EP0845741A2, EP0845741A3, EP0845741B1, EP1306752A1, US5974540, US6237084 |
Publication number | 10366502, 366502, US RE39121 E1, US RE39121E1, US-E1-RE39121, USRE39121 E1, USRE39121E1 |
Inventors | Toru Morikawa, Nobuo Higaki, Akira Miyoshi, Keizo Sumida |
Original Assignee | Matsushita Electric Industrial Co., Ltd. |
Export Citation | BiBTeX, EndNote, RefMan |
Patent Citations (24), Non-Patent Citations (4), Referenced by (2), Classifications (14), Legal Events (5) | |
External Links: USPTO, USPTO Assignment, Espacenet | |
More than one reissue application has been filed for the reissue of U.S. Pat. No. 6,237,084. The reissue application are application Nos. 10/366,502 and 11/016,920, all of which are divisional reissues of U.S. Pat. No. 6,237,084.
This is a divisional application of U.S. Ser. No. 08/980,676 now U.S. Pat. No. 5,974,540 filed Dec. 1, 1997.
1. Field of the Invention
The present invention relates to a processor that performs processing according to instruction sequences that are stored in a ROM or the like.
2. Background of the Invention
In recent years, there has been a visible increase in the use of application software that can interactively reproduce various kinds of data, such as video data, still image data, and audio data, that have been compressed according to techniques such as frame encoding, field encoding, or motion compensation. As such software has been developed, there has been increasing demand for multimedia-oriented processors that can efficiently execute the software. These multimedia-oriented processors are processors designed with a special architecture to facilitate programming, such as the compression and decompression of video and audio data. The high-speed processing required for handling video data is the matrix multiplication of compressed data that has N*N matrix elements with coefficient data that also has N*N matrix elements. Representative examples of compressed data that has N*N matrix elements are the luminescence block composed of 16*16 luminescence elements, the blue color difference block (Cb block) composed of 8*8 color difference elements, and the red color difference block (Cr block) composed of 8*8 color difference elements used in MPEG (Moving Pictures Experts Group) techniques. The matrix multiplication for compressed data referred to here is performed very frequently when executing the approximation calculations for an inverse DCT (Discrete Cosine Transform) in image compression methods such as MPEG and JPEG (Joint Photographic Experts Group).
The following is a description of conventional multimedia-oriented processors that can perform high-speed matrix multiplication. The basic architecture of conventional multimedia-oriented processors is provided with a sum-product result register (hereinafter simply referred to as an MCR register) as hardware, and is provided with an instruction set that includes a “MOV MCR,**” transfer instruction for transferring a sum-product value.
An example of the hardware construction of a conventional multimedia-oriented processor is shown in FIG. 1. As shown in
When a conventional multimedia-oriented processor is used, however, positive correction saturation operations for amending the sum-product value pose many difficulties for programmers.
Positive conversion processing refers to the conversion of a sum-product value that is a negative value into either zero or a positive value. Normally, compressed data is expressed as a coded relative value that reflects the relation of the present value to the preceding and succeeding values. As a result, there are many cases when the sum of products for each element in the compressed data and the corresponding coefficients is a negative value. Most reproduction-related hardware, such as displays and speakers, however is only able to process uncoded data, so that when the sum-product values are to be reproduced, it is first necessary to perform positive conversion processing.
Saturation calculation processing refers to processing that sets all values that exceed a given range (or, in other words, which are “saturated”) at a predetermined value. This is to say, when an element that includes an erroneous bit generated during transfer is used in a sum-product calculation as part of the sum-product processing for compressed data, there is an increase in the probability of the sum-product value exceeding a value that can be expressed by the stated number of bits. Since most reproduction-related hardware is only physically capable of reproducing uncoded data with a fixed valid number of bits, such as eight bits, saturation processing is required to convert the sum-product value into a value that can be expressed using the valid number of bits.
It has been conventional practice to perform this kind of positive value conversion processing and saturation calculation processing by converting the-sum-product value using a subroutine that corrects the sum-product value. An example of a subroutine that corrects the sum-product value is explained below. In this example, the register width and the calculation width of the calculation unit are 32 bits, with the width of the MCR being 32 bits, and the sum-product value being expressed as a coded 16-bit integer. The data that can be handled by the reproduction-related hardware needs to be expressed using uncoded 8-bit integers. This subroutine is set as using the data register D0 for storing the calculation result. Each instruction is expressed using two operands, with the left and right operands being respectively called the first and the second operands. The second operand is used both to indicate the transfer address of a transfer instruction and the storage address of an arithmetical instruction.
Describing the above instructions in order, Instruction 1, “MOV MCR,D0”, transfers the stored value of the MCR register into the data register D0. Instruction 2, “CMP 0xFFFF_{—}8000,D0”, compares the value in the data register with the immediate “0xFFFF_{—}8000”, where “0x” shows that the value is given in hexadecimal. This comparison is performed by subtracting the immediate “0xFFFF_{—}8000” given in the first operand from the stored value of the data register D0 given in the second operand.
The sixteenth bit of the immediate “0xFFFF_{—}8000” in Instruction 2 is the code bit used for a 16-bit coded integer, so that when the stored value of the data register D0 is greater that the immediate “0xFFFF_{—}8000”, this shows that the value stored in the MCR is a negative number.
On the other hand, when the stored value of the D0 register is less than “0xFFFF_{—}8000”, this shows that the value stored by the MCR is a positive number. If this number is a positive number, a carry is performed and the carry flag in the flag register is set.
The letter “B” in the “BCC” in Instruction 3 stands for “Branch”, while the letters “CC” stand for “Carry Clear”.
When the comparison in Instruction 2 finds that the stored value of the register D0 is less than the immediate “0xFFFF_{—}8000”, a branch is performed to Instruction 6 which has the label “CARRY”. Conversely, when the comparison in Instruction 2 finds that the stored value of the register D0 is greater than the immediate “0xFFFF_{—}8000”, Instruction 4, “MOV 0x0000_{—}0000,D0” transfers the value zero into the register D0, amending the sum-product value to zero. After this amendment, the unconditional branch “BRA END” in Instruction 5 is performed to transfer the processing to the “END” label, thereby completing the positive conversion processing.
The processing described above is performed when the stored value of the register D0 is negative. The following is a description of the processing performed when the stored value of the register D0 is greater than the immediate “0xFFFF_{—}8000”. In such a case, Instruction 6, “CMP 0x0000_{—}00F,D0” compares the stored value of the register D0 with the immediate “0x0000_{—}00FF”. This comparison is performed by subtracting the immediate “0x0000_{—}00FF” given in the first operand from the stored value of the data register D0 given in the second operand. When the stored value of the D0 register is smaller than the immediate “0x0000_{—}00FF”, a carry is performed and the carry flag in the flag register is set.
The letters “CS” in Instruction 7, “BCS END”, stand for “Carry Set”, so that when the carry flag is set, a branch is performed to the label “END” from Instruction 7.
When the carry flag is not set, no branch is performed in Instruction 7 and processing advances to Instruction 8, “MOV 0x0000_{—}00FF,D0”, where the immediate “0x0000_{—}00FF” is transferred into the register D0 to amend the calculation result to “0x0000_{—}00FF”, thereby completing the saturation calculation processing.
The problem with the sum-product value amendment process described above lies in the considerable increase in code size caused by the insertion of the above eight instructions for one amendment of a sum-product value. When the program is written into a ROM to embed the software into the information processing apparatus, the required amount of installed ROM will have to need to be increased by an amount equal to this increase in code size, leading to an increase in manufacturing cost. A large number of manufacturers of domestic appliances such as digital video players, electronic notebooks, and word processors seek to improve on their rivals' products by using their own decompression processing programs, although the installation of such decompression processing programs presently has the drawback of increasing costs by increasing the required amount of ROM, making such installation problematic.
There is also the problem that since eight instructions need to be executed to correct one sum-product value, there is a large increase in processing time. When, as shown in
The same sum-product processing and positive conversion saturation calculation processing is required to obtain the other matrix multiplication result elements H21, H31, H41, H51, H61, H71, H81, . . . , and since there are 64 elements in the coefficient matrix Gij (where i,j=1,2,3,4,5 . . . 8), the sum-product value amending subroutine for positive conversion saturation calculation processing needs to be performed 64 times. This sum-product value amending subroutine includes branch instructions (as Instructions 3, 5, and 7), so that when this sum-product value amending subroutine is executed, branches will occur regardless of whether negative values or saturation occur, so that the 64 iterations of the subroutine will not be performed smoothly. When attempts are made to improve the processing speed of the sum-product operation by introducing pipeline processing to the processor, the execution of the stated three branch instructions will result in a noticeable drop in processing efficiency.
In order to increase the speed of the matrix multiplication, it is possible to install a specialized circuit for performing matrix multiplication. However, if all of the matrix multiplications are performed by a specialized circuit, there would be a vast increase in hardware, and the processor characteristic known as versatility, whereby the processor executes a variety of processes in accordance with the program written by the programmer, is lost. If the versatility of the processor is lost, there is the risk that the processor will not be able to respond to programmers' wishes, and so will not, for example, be able to execute an original decompression processing program.
It is a primary object of the present invention to provide a processor that can perform a rounding process made up of a positive conversion process and a saturation calculation process at high speed, while minimizing the increase in code size caused by the rounding process.
The stated object can be achieved by a processor that successively decodes and executes instructions in an instruction sequence, the instruction sequence including instructions that indicate a storage address of a value used in an operation, the processor including: a detecting unit for detecting whether a next instruction to be decoded includes an operation content indication showing that the next instruction is a correction instruction and, if present, reading the operation content indication; and a rounding unit for rounding, when the detecting unit has detected an operation content indication showing that the next instruction is a correction instruction, a coded m-bit integer stored at a storage address indicated by the instruction to a value expressed as an uncoded s-bit integer (where s<m).
With the stated construction, the processing for rounding values is performed once each time a correction instruction is detected out of the instruction sequence, so that the rounding process can be executed by the programmer writing only one instruction.
As the rounding process is performed according to one correction instruction, the execution time for one execution of the rounding process is extremely short. When the rounding of calculated values is required very often, such as when decompressing data, there will not be a significant increase in the time taken by the decompression processing.
Since the rounding process can be performed by simply executing a correction instruction, when the processor attempts to perform a sum-products operation at high speed through pipeline processing, thee will be no confusion in the pipeline. Accordingly, the code size of the instruction sequence can be reduced and the execution of the instruction sequence made faster by adding a small amount of hardware to the processor.
The stated object can also be achieved by a processor that successively decodes and executes instructions in an instruction sequence, the instruction sequence including instructions that indicate a storage address of a value to be used in an operation, the processor including: a first detecting unit for detecting whether a next instruction to be decoded includes an indication showing that the instruction has a calculation performed; a second detecting unit for detecting whether the next instruction to be decoded includes an indication showing that calculation is to be performed and that rounding is-to be performed on a calculation result; a calculating unit for performing, when the first detecting unit detects that the next instruction includes an indication showing that the instruction has a calculation performed, a calculation using an m-bit integer in accordance with the indication; and a rounding unit for rounding, when the second detecting unit has detected that the next instruction to be decoded includes an indication showing that rounding is to be performed, a calculation result of a calculation that uses an m-bit integer to a value expressed as an uncoded s-bit integer (where s<m).
With the stated construction, correction instructions for performing a rounding process of a coded calculation result are provided, so that the two processes composed of a calculation process and a rounding process can be performed in a single step. As a result, positive conversion saturation calculation processing is performed in the same step as the calculation processing, so that the effective number of steps taken the positive conversion saturation calculation processing is zero.
These and other objects, advantages and features of the invention will become apparent from the following description thereof taken in conjunction with the accompanying drawings which illustrate a specific embodiment of the invention. In the drawings:
First Embodiment
The following is an explanation of the first embodiment of the present invention with reference to the drawings.
The RAM 10 stores the compressed data Fij (i,j=1,2,3,4,5 . . . 8) composed of 8*8 matrix elements and coefficient data Gji (i,j=1,2,3,4,5 . . . 8) composed of 8*8 matrix elements. When a fetch address for the ith row and jth column is outputted to the address bus 17, the data indicated by the outputted address is outputted to the data bus 18. When the operation execution apparatus 14 calculates the ith row and jth column element of the multiplication result matrix Hij (i,j=1,2,3,4,5 . . . 8) for the multiplication of the compressed data Fij composed of 8*8 matrix elements and the 8*8 matrix coefficients Gij, and the ith row and jth column address is outputted to the address bus 17 as the write address, the ith row and jth column element transferred to the data bus 18 is written into the storage area indicated by the outputted address. The multiplication result matrix Hij is used in an approximation calculation using an inverse DCT, is subjected to predetermined processing, and is used by the reproduction-related hardware (not illustrated). It should be noted here that Fij and Gji are expressed as 8-bit coded integers whose the eighth bit counted from the LSB (least significant bit) is used as the code bit.
ROM 11 stores a decompression processing program for the compressed data stored in the RAM 10, so that when a read address is outputted from the instruction fetch circuit 12, the instruction indicated by the read address in the stored decompression processing program is outputted to the data bus 18. The decompression processing program stored by the ROM 11 is composed of a variety of instructions, such as transfer instructions, arithmetic instructions, and branch instructions. Of these, the arithmetic instructions can be roughly classified into arithmetic calculation instructions, sum-product function multiplication instructions, and logic operation instructions. Arithmetic calculation instructions, such as addition instructions, subtraction instructions, and multiplication instructions each have a first operand and second operand. The first operand has two addressing modes that are an indication of an immediate and an indirect indication of a register. On the other hand, an indirect indication of a register is the only possible addressing mode for the second operand, although the second operand also doubles as an indication of the storage address for the calculation result. For the example of the addition instruction “ADD imm,D1”, the first operand is the immediate value imm, while the second operand is the register D1. Since the second operand indicates the storage address for the calculation result, the calculation result of the addition instruction “ADD D0,D1” stores the result of the addition of the values in registers D0 and D1 in register D1.
In the present embodiment, the decompression processing program includes a matrix multiplication subroutine that generates the matrix multiplication result Hij with 8*8 elements by multiplying the compressed data Fij composed of 8*8 matrix elements by the coefficient data Gij composed of 8*8 matrix elements. It should be especially noted that this matrix multiplication subroutine is mainly composed of sum-product function multiplication instructions “MACCB Dm,Dn” and positive conversion saturation calculation instructions “MCSST Dm”.
The following is a description of the generation of the matrix Hij with 8*8 elements by multiplying the compressed data Fij (i,j=1,2,3,4,5 . . . 8) composed of 8*8 matrix elements by the coefficient data Gji (i,j=1,2,3,4,5 . . . 8) also composed of 8*8 matrix elements. When doing so, the calculation shown in Equation 1 below is necessary to calculate the 1^{st }row, 1^{st }column element H11 of the matrix Hij.
H11=G11*F11+G12*F21+G13*F31+G14*F41+G15*F51+G16*F61+G17*F71+G18*F81 Equation 1
This Equation 1 has Gji*Fij (i,j=1,2,3,4,5 . . . 8) as its elemental operations and is a compound operation which finds an algebraic sum of the elemental operations. The calculation of the respective elemental operations and the calculation of the algebraic sum are expressed in the present matrix multiplication subroutine by a loop statement (this loop statement being called a “sum-product loop”) that repeatedly has a sum-product function multiplication instruction “MACCB Dm,Dn” performed.
As a result of the transfer instructions mentioned above, the address registers A0, A1, and A2 are used to indicate the read addresses for the compressed data, the coefficient data, and the matrix multiplication result data which are each composed of 8*8 elements.
Instruction 4, “MOV INIT,D2”, sets the initial value INIT for the number of iterations into the data register D2, while instruction 5, “MOV (A0),D0”, has the coefficient data Fji read from the address indicated by the address register A0 transferred to the data register D0. Instruction 6, “MOV (A1),D1”, has the compressed data Gij read from the address indicated by the address register A1 transferred to the data register D1. Instruction 7, “MACCB D0,D1”, is a multiplication instruction with a sum-product function that uses the data register D0, the data register D1, and the sum-product result register 6. Instruction 8, “CMP NUMBER,D2” is an instruction which performs an upper limit check on the number of iterations by subtracting the total number of iterations NUMBER from the number of iterations stored in the data register D2.
Instruction 9, “BCS LP1_NEXT” is a conditional branch instruction that branches to instruction 13 “ADD 1,D2” with the label LP1_NEXT when the carry flag is ON as a result on the subtraction performed in instruction 8 “CMP NUMBER, D2”.
Instruction 10, “MCSST D1” is an instruction that performs conversion to a positive value and saturation calculation processing (hereinafter referred to as “positive conversion saturation calculation processing”) when the carry flag is OFF.
Instruction 13, “INC D2” is an addition instruction that increments the number of iterations stored in the data register D2 by 1, while instruction 14, “INC A0”, is an addition instruction that increments the read address for coefficient data stored in the address register A0 by 1. Instruction 15, “ADD ROW1,A1”, is an addition instruction that adds one row number ROW1 to the read address of the coefficient data stored in the address register A1. Instruction 16, “BRA LP1_START”, is an unconditional branch instruction that branches to the label LP1_START.
The label LP1_START is attached to instruction 5, “MOV (A0),D0”, with the instruction sequence from instruction 5 to instruction 12, “BRA LP1_END” calculating one elemental operation, with the algebraic sum of the calculation results of all iterations of this instruction sequence being found.
The label LP1_NEXT is attached to instruction 13, “ADD 1,D2”, with the instruction sequence from instruction 13 to instruction 16, “BRA LP1_START” moving the read address of the compressed data to a next row and the read address of the coefficient data to the next column when the matrix multiplication of one row of elements by one column of elements has been completed. At the same time, the number of iterations stored in the data register D2 is incremented by 1.
The field indicating the read address of the multiplier and the field indicating the read address of the multiplicand can each be set at one of “00”, “01”, “10”, and “11”, thereby indicating one of data register D0, data register D1, data register D2, and the sum-product result register 6 as a read address for the multiplier Gji or the multiplicand Fij.
The one-bit field for indicating the calculated content of the elementary operation shows the content of the calculation of the elementary operation performed for the multiplier Gji and the multiplicand Fij. When “1” is written into this field, the multiplication “Gji*Fij” of the multiplier Gij and the multiplicand Fij is indicated as the content of the elementary operation on the multiplier Gji and the multiplicand Fij.
When the one-bit field indicating the storage address of a sum-product value is set at “1”, this indicates that the MCR (the sum-product result register 6 which is described later) is set as the storage address of the sum-product value. When the one-bit field indicating the calculated content of the algebraic sum is set at “1”, this shows that the algebraic sum is set so that the multiplication result “Gji*Fij” is added to the sum-product value stored in the sum-product result register 6.
When the sum-product operation “G11*F11+G12*F21+G13*F31 . . . ” is performed, it should be noted that a bit error when transferring the element Fij of the compressed data can result in a sum-product value ““G11*F11+G12*F21+G13*F31 . . . ” which is beyond a reproducible range for the reproduction-related hardware. Since this risk exists, the matrix multiplication subroutine performs the positive conversion saturation calculation instruction “MCSST Dm” after the loop processing repeating the sum-product function multiplication instruction “MACCB D0,D1” has been completed, so that positive conversion saturation calculation processing is performed for the sum-product value.
The format of the operation codes for the positive conversion saturation calculation instruction “MCSST Dm” is shown in FIG. 7. As shown in
As described above, the decompression processing program stored in the ROM 11 is such that the elementary operations and the algebraic sum calculation that compose the compound operation that is required by the matrix multiplication subroutine are performed by a sum-product function multiplication instruction “MACCB Dm,Dn”, so that the algorithm is very compact. Since only this operation needs to be performed by the processor, the memory area of the ROM 11 that is used by the decompression processing program is extremely small.
This completes the description of the instruction sequences stored in the ROM 11, so that the following explanation will instead focus on the constructional elements of the processor shown in FIG. 3.
The instruction fetch circuit 12 shown in
The decoder 13 has an instruction buffer for accumulating the plurality of instructions that are read from the data bus 18 and an instruction register for holding an instruction to be decoded, out of the plurality of instructions accumulated in the instruction buffer. The decoder 13 decodes the instruction stored in the instruction buffer and has the operation execution apparatus 14 perform the necessary control to have the decoded instruction executed. Of the control operations mentioned here, special attention should be paid to (1) register output control, (2) calculation execution control, and (3) constant generation control. These are described in more detail below.
(1) Register output control refers to a controlling of the operation execution apparatus 14 to output a stored value of a register indicated by either the first or second operand in an arithmetic calculation instruction, a logic operation instruction, or a sum-product function multiplication instruction. (2) Calculation execution control refers to a controlling of the operation execution apparatus 14 to execute the calculation indicated by an arithmetic calculation instruction, a logic operation instruction, or a sum-product function multiplication instruction. (3) Constant generation control refers to a controlling of the operation execution apparatus 14 to generate a maximum value or zero for performing a positive conversion saturation calculation instruction. The (1) register output control and the (2) calculation execution control are performed when an arithmetic calculation instruction, a logic operation instruction, or a sum-product function multiplication instruction is decoded by the decoder 13, while the (1) register output control and the (3) constant generation control are performed when a positive conversion saturation calculation instruction is decoded by the decoder 13. Occurrences of (2) calculation execution control only happen when a sum-product function multiplication function is decoded, while occurrences of (3) constant generation control only happen when a positive conversion saturation calculation instruction is decoded, so that the (2) calculation execution control and the (3) constant generation control are mutually exclusive.
The address bus 17 has a bit width of 32 bits and is used to transfer the compressed data Fji, the coefficient data Gij, and the matrix multiplication data Hij when data is outputted by the RAM 10.
The operation execution apparatus 14 includes a register file and an ALU circuit, and performs calculation according to control by the decoder 13.
It should be especially noted here that the construction is such that the instruction fetch circuit 12 performs the fetch stage, the decoder 13 the decoding stage, and the operation execution apparatus 14 the operation execution stage, the memory write stage, and the register write stage. These five stages are realized by a five-stage pipeline process. The instruction fetch circuit 12 starts to fetch another instruction once an instruction has entered the decoding stage performed by the decoder 13, and so does not wait for the execution of the present and preceding instructions to be completed. In the same way, the decoder 13 does not wait for the execution of the present and preceding instructions to be completed, and so starts to decode a new instruction once a decoded instruction has entered the operation execution stage performed by the operation execution apparatus 14. By performing such processing, the processor processes the instruction sequence stored in the ROM 11 according to a five-stage pipeline process composed of an instruction fetch stage, a decoding stage, an execution stage, a memory access stage, and a register write stage, as shown in FIG. 12A.
This completes the description of the constructional elements of the processor. The following description will focus on the internal construction of the operation execution apparatus 14. As shown in
The register file 1 is composed of four 32-bit data registers D0-D3, and three 32-bit address registers A0-A2. During (1) register output control, when one or two register names are indicated by the decoder 13, the register file 1 has the stored values of the registers with the indicated register names outputted via paths C2, C3 to the first internal bus 15 and the second internal bus 16. The register file 1 also holds the value transferred on the data bus 18 which it receives via the path C1.
The first internal bus 15 is 32 bits wide and transfers a 32-bit stored value outputted by the register file 1 to the ALU circuit 2.
The second internal bus 16 is also 32 bits wide and transfers a 32-bit stored value outputted by the register file 1 to the ALU circuit 2.
The code extension circuit 4 performs code extension when the stored value of a data register transferred from the register file 1 via the first internal bus 15 is a negative number. In the present embodiment, multipliers and multiplicands are defined as coded 8-bit numbers, so that the code extension circuit 4 performs code extension of 8-bit negative numbers. As one example, when the stored value transferred via the first internal bus 15 is the 8-bit negative value “0x0000_{—}0080”, the code in the 8^{th }bit is extended to the 9^{th }through 32^{nd }bits so that the value “0xFFFF_FF80” is outputted to the ALU circuit 2.
The code extension circuit 5 performs code extension when the stored value of a data register transferred from the register file 1 via the second internal bus 16 is a negative number. In the present embodiment, multipliers and multiplicands are defined as coded 8-bit numbers, so that the code extension circuit 5 performs code extension of 8-bit negative numbers. The method used for code extension is the same as for the code extension circuit 4.
The ALU circuit 2 is composed of an addition unit, a multiplication unit, and a barrel shifter that are all 32-bits wide, and performs calculations according to the (2) calculation execution control indicated by the decoder 13. Since the input terminals of the ALU circuit 2 are connected to the output terminals of the code extension circuit 4 and the sum-product result register 6, the calculation performed according to the (2) calculation execution control is performed using the 32-bit stored value of a register which is outputted by the code extension circuit 4 and the 32-bit stored value of a register which is outputted by the sum-product result register 6.
When a sum-product function multiplication instruction “MACCB Dm,Dn” is decoded, the ALU circuit 2 performs multiplication of the 32-bit stored value of a register outputted by the code extension circuit 4 and the 32-bit stored value of a register outputted by the code extension circuit 5 and outputs a 64-bit multiplication result. The ALU circuit 2 also adds the 32-bit value transferred on the path P1 to the lower-order 32 bits of the 64-bit multiplication result and outputs a 32-bit addition result on the path P2.
It should be noted here that when the sum-product function multiplication instruction “MACCB D0,DL1” is decoded, the stored values of the read address registers indicated by the first and second operands of the sum-product function multiplication instruction “MACCB D0,D1” will be transferred on the first internal bus 15 and the second internal bus 16, so the ALU circuit 2 will perform the multiplication of the stored value of the register D0 and the stored value of the register D1. Also, when the sum-product function multiplication instruction “MACCB D0,D1” is decoded, the stored value of the sum-product result register 6 will be transferred on the path P1, so that the multiplication result of the data register D0 and the data register D1 will be added to the stored value of the sum-product result register 6 which has been outputted to the path P1. The result of this addition is then outputted on the path P2.
The sum-product result register 6 stores the sum-product value which has hitherto been accumulated, and outputs its stored value on the path P1 every time a sum-product function multiplication instruction is decoded. When the stored value on the path P1 and the multiplication result have been added by the ALU circuit 2, the addition result is outputted on the path P2, with this value being latched by the sum-product result register 6 and stored as the updated sum-product value. It should be noted here that the sum-product result register 6 stores the result of the multiplication by the ALU circuit 2 of the multiplier Gji and the multiplicand Fij as a coded 16-bit value.
The positive conversion saturation calculation circuit 3 rounds the stored value of the sum-product result register 6 expressed as a coded 16-bit value to a positive 8-bit integer.
The middle part of
The lower part of
In
It should be noted here that the range of values that can be used the sum-product value is a range of 15-bit positive and negative values to avoid the totaling of rounding errors. In more detail, when the sum-product value is found from the multiplication result of the multiplier Gji and the multiplicand Fij which are both coded 8-bit values, if the multiplication result were to be rounded to eight bits every time because the range of the multiplication matrix element Hij is eight bits, the rounding error would increase every time multiplication is performed. To avoid such increases in rounding error, the sum-product result register 6 sets the sum-product result as 16 bits and the positive conversion saturation calculation is only performed when the multiplication of one row by one column has been completed.
The stored value of the sum-product result register 6 shown by the bar in the center of
The positive conversion saturation calculation circuit 3 rounds the range of values shown by the symbol y2 (negative values) to the value “0x0000_{—}0000”.
The internal construction of the positive conversion saturation calculation circuit 3 is shown inside the broken line y8 in FIG. 4. As shown in
When the instruction read by the instruction fetch circuit 12 and decoded by the decoder 13 is a positive conversion saturation calculation instruction “MCSST”, the constant generator 21 generates a maximum positive value which, in accordance with the content of the positive conversion-saturation calculation width field, is an 8-bit uncoded value, a 16-bit uncoded value, or a 24-bit uncoded value. When the generation of an 8-bit uncoded value is indicated by the positive conversion-saturation calculation width field, the constant generator 21 generates the 32-bit coded value “0x0000_{—}00FF”, which is the maximum value for an uncoded 8-bit value, and outputs it to the multiplexer 24. When the generation of a 16-bit uncoded value is indicated by the positive conversion-saturation calculation width field, the constant generator 21 generates the 32-bit coded value “0x0000_FFFF”, which is the maximum value for an uncoded 16-bit value, and outputs it to the multiplexer 24. Similarly when the generation of a 24-bit uncoded value is indicated by the positive conversion-saturation calculation width field, the constant generator 21 generates the 32-bit coded value “0x00FF_FFFF”, which is the maximum value for an uncoded 24-bit value, and outputs it to the multiplexer 24.
The comparator 22 compares the magnitude of the value held by the sum-product result register 6 with the magnitude of maximum value outputted by the constant generator 21. This comparison is performed by subtracting the value held by the constant generator 21 from the maximum value outputted by the constant generator 21 and detecting whether a carry has occurred as a result of the subtraction. When an 8-bit value is indicated by the positive conversion-saturation calculation width field, the stored value of the sum-product result register 6 is subtracted from the 32-bit coded integer “0x0000_{—}00FF”, which is the maximum value for an uncoded 8-bit value. When a 16-bit value is indicated by the positive conversion-saturation calculation width field, the stored value of the sum-product result register 6 is subtracted from the 32-bit coded integer “0x0000_FFFF”, which is the maximum value for an uncoded 16-bit value. Similarly, when a 24-bit value is indicated by the positive conversion-saturation calculation width field, the stored value of the sum-product result register 6 is subtracted from the 32-bit coded integer “0x00FF_FFFF”, which is the maximum value for an uncoded 24-bit value.
When a carry is detected as the result of the subtraction described above and the latched value is judged to exceed the maximum value, the comparator 22 outputs the logic value “1” to the multiplexer 24. Conversely, when the value held by the sum-product result register 6 is judged to be equal to or below the maximum value, the comparator 22 outputs the logic value “0” to the multiplexer 24.
The polarity judging unit 23 judges whether the code bit of the value stored by the sum-product result register 6 is “ON”. Here, depending on the content the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST”, an 8-bit uncoded value, a 16-bit uncoded value, or a 24-bit uncoded value is indicated, so that the position of the code bit will change. As a result, the polarity judging unit 23 changes the bit which is to be judged in accordance with the indication in the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST”.
When the indication in the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST” is for a 24-bit uncoded value, the polarity judging unit 23 judges whether the 24^{th }bit from the LSB side is “ON”, while when the indication in the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST” is for a 16-bit uncoded value, the polarity judging unit 23 judges whether the 16^{th }bit from the LSB side is “ON”. Similarly, when the indication in the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST” is for an 8-bit uncoded value, the polarity judging unit 23 judges whether the 8^{th }bit from the LSB side is “ON”. This judgement refers to a judgement of whether the sum-product value held by the sum-product result register 6 is expressed as a negative number when values are expressed in accordance with the indication given in the positive conversion-saturation calculation width field of the positive conversion saturation calculation instruction “MCSST1”. When the value is a negative value, the polarity judging unit 23 outputs the logic value “1” to the multiplexer 24. Conversely, when the value is zero or a positive value, the polarity judging unit 23 outputs the logic value “0” to the multiplexer 24.
The zero generator 25 generates the integer “0x0000_{—}0000” when the decoded instruction is a positive conversion saturation calculation instruction “MCSST”.
The multiplexer 24 selects and outputs one of the maximum value generated by the constant generator 21, the zero value “0x0000_{—}0000” generated by the zero generator 25, and the sum-product value held by the sum-product result register 6, in accordance with the combination of the logic values outputted by the comparator 22 and the polarity judging unit 23.
If the logic value outputted by the comparator 22 is set as the logic value x and the logic value outputted by the polarity judging unit 23 is set as the logic value y, the correspondence between the combinations of these logic values and the output value of the multiplexer 24 can be expressed by the truth table shown in FIG. 9. It should be noted here that the example truth table shown in
As shown in
When the output value of the comparator 22 is “1” and the output value of the polarity judging unit 23 is “0”, the multiplexer 24 outputs the maximum value “0x0000_{—}00FF” generated by the constant generator 21.
When the output value of the comparator 22 is “0” and the output value of the polarity judging unit 23 is “1”, the multiplexer 24 outputs the zero value “0x0000_{—}0000” generated by the zero generator 25. When the output value of the comparator 22 is “1” and the output value of the polarity judging unit 23 is “1”, the multiplexer 24 outputs the zero value “0x0000_{—}0000” generated by the zero generator 25.
In
The multiplication of the 32-bit held value “0x0000_{—}007F” of the data register D0 outputted by the code extension circuit 4 and the 32-bit value “0xFFFF_FF80” extended by the code extension circuit 5 is performed by the code extension circuit 5, and the lower 32-bits “0xFFFF_C080” of the 64-bit multiplication result “0xFFFF_FFFF_FFFF_C080” are outputted to the sum-product result register 6. When the positive conversion saturation calculation instruction “MCSST” has been decoded, the sum-product result register 6 outputs its held value to the positive conversion saturation calculation circuit 3, which judges that the 32-bit value “0xFFFF_C080” is a coded 16-bit negative number. As a result, the positive conversion saturation calculation circuit 3 outputs the 8-bit zero value “0x0000_{—}0000” to the data bus 18 so that this zero value is held by the data register D1 in the register file 1.
The following is a description of the operation of the processor constructed as described above. A transfer instruction included in the matrix multiplication subroutine is first written into the instruction buffer of the decoder 13 by the instruction fetch circuit 12 and the instruction “MOV(A0), D0” is decoded by the decoder 13. This instruction 5:“MOV(A0),D0” is a transfer instruction that indicates a data read for the RAM 10 using indirect register referencing that indicates the read address using the address register A0. As a result, an element on the first row and first column (F11) of the compressed data matrix Fij that is composed of 8*8 matrix elements stored in the RAM 10 is transferred to the data register D0. The following instruction, instruction 6:“MOV(A1),D0” similarly writes an element on the first row and first column (G11) of the coefficient data matrix Gji that is composed of 8*8 matrix elements into the data register D1 in the register file 1 via the data bus 18.
The next instruction in the matrix multiplication subroutine is instruction 7 which is the sum-product function multiplication instruction “MACCB D0,D1”. This is fetched by the instruction fetch circuit 12 and written into the instruction buffer of the decoder 13, before being decoded by the decoder 13. When the decoder 13 decodes the sum-product function multiplication instruction “MACCB D0,D1”, the held values F11 and G11 of the data register D0 and the data register D1 are transferred to the first internal bus 15 and the second internal bus 16.
When the sum-product function multiplication instruction “MACCB D0,D1” is decoded, the first internal bus 15 and the second internal bus 16 transfer the held values F11 and G11 of the read address registers indicated by the first and second operands of the sum-product function a multiplication instruction “MACCB D0,D1”. These values are then outputted by the code extension circuit 4 and the code extension circuit 5 into the ALU circuit 2, where the multiplication of the held value of the data register D0 and the held value of the data register D1 is performed. The multiplication result “F11*G11” is then transferred to the sum-product result register 6 and is held by the sum-product result register 6.
Once the sum-product result register 6 has stored the multiplication result “F11*G11”, a branch is performed to the label “LP1_NEXT” due to the execution of instruction 8: “CMP NUMBER,D2” and “instruction 9:“BCS LP1_NEXT”, so that the instruction 13:“ADD 1,D2” is decoded. This instruction 13:“ADD 1,D2” increments the number of iterations. After instruction 13:“ADD 1,D2”, instruction 14:“ADD 1,A0” and instruction 15:“ADD ROW,A1” are executed, so that read addresses of the ROM 11 are advanced to the next column and row. As a result of the incrementing in these instructions, the read address of the compressed data is advanced to the 2^{nd }row, 1^{st }column element and the read address of the coefficient data Gji is advanced to the 1^{st }column, 2^{nd }row element.
After the read addresses have been incremented, the following instruction, instruction 16:“BRA LP1_START” is decoded. The branch address of instruction 16:“BRA LP1_START” is instruction 5:“MOV(A0),D0” which has label “LP1_START” attached, so that the branch in instruction 16 has instruction 5:“MOV(A),D0” and instruction 6:“MOV(A1),D1” re-executed.
As a result of these transfer instructions, the 2^{nd }row, 1^{st }column element F21 in the compressed data in the RAM 10 is transferred into the data register D0, while the 1^{st }row, 2^{nd }column element G12 in the compressed data is transferred into the data register D1 in the register file 1 via the data bus 18.
In the matrix multiplication subroutine, these transfer instructions are followed by the sum-product function multiplication instruction “MACCB D0,D1”, so that this sum-product function multiplication instruction “MACCB D0,D1” is written into the internal buffer of the decoder 13 by the instruction fetch circuit 12 and is decoded by the decoder 13. As a result of the decoding, the held values of the data register D0 and the data register D1 are transferred to the first internal bus 15 and the second internal bus 16.
When the sum-product function multiplication instruction “MACCB D0,D1” is decoded, the first internal bus 15 and the second internal bus 16 transfer the held values F11 and G11 of the read address registers indicated by the first and second operands of the sum-product function multiplication instruction “MACCB D0,D1”. These values are the matrix elements G12 and F21, so that the multiplication of the held value G12 of the data register D0 and the held value F21 of the data register D1 is performed by the ALU circuit 2. At this point, the sum-product result register 6 holds the value “G11*F11” which is the total of the elemental operations thusfar performed in the sum-product calculation. When the sum-product function multiplication instruction “MACCB D0,D1” is decoded, the sum-product result register 6 outputs the held value onto the path P1.
Since the held value of the same-product result register 6 is outputted onto the path P1, the multiplication result “G12*F21” of the data register D0 and the data register D1 is added to the held value of the sum-product result register 6 on the path P1. The result of the addition is then outputted onto the path P2.
Once the addition of the held value “G11*F11” and the multiplication result “G12*F21” has been performed by the ALU circuit 2, the addition result “G11*F11+G12*F21” is outputted onto the path P2, so that the sum-product result register 6 holds this addition result as the sum-product value “G11*F11+G12*F21”.
The processing described above is repeated for all of the elements on the first row of the coefficient data Gji and all of the elements in the first column of the compressed data Fij, so that the sum-product value is calculated for “G11*F11+G12*F21+G13*F31+G14*F41+G15*F51+G16*F61+G17*F71+G18*F81” and stored in the sum-product result register 6. Here, should there be a bit error during the transfer of the element 31 from the RAM 10, there is the risk that the sum-product value held by the sum-product result register 6 will be a value (such as “0x0000_{—}78FF”) that clearly exceeds the range of values that can be reproduced by the reproduction-related hardware.
After this, the next loop statement in the ROM 11, the positive conversion saturation calculation instruction “MCSST D1”, is written into the internal buffer of the decoder 13, this positive conversion saturation calculation instruction “MCSST D1” is decoded by the decoder 13.
When the positive conversion saturation calculation instruction “MCSST D1” is decoded by the decoder 13, the held value “0x000_{—}78FF” of the sum-product result register 6 is outputted onto the path P1. After this value has been outputted to the P1, the comparator 22 is activated by the decoder 13. The comparator 22 compares the held value of the sum-product result register 6 with the 32-bit coded integer “0x0000_{—}00FF” to see which is larger. Here, since the held value of the sum-product result register 6 “0x000_{—}78FF” exceeds the 32-bit coded integer 0x0000_{—}00FF”, the comparator 22 outputs the logic value “1” to the multiplexer 24.
The polarity judging unit 23 judges whether the 16^{th }bit counting from the LSB side in the value held by the sum-product result register 6 is “ON”. This judgement equates to a judgement as to whether the held sum-product value of the sum-product result register 6 is a negative number. The held value “0x000_{—}78FF” expressed in binary is “0000_{—}0000 0000_{—}0000 0111_{—}1000 1111_{—}1111”, so that the 16^{th }bit counting from the LSB side can be seen to be “0”. As a result, the logic value “0” is outputted to the multiplexer 24.
In the present case, the maximum value “0x0000_{—}00FF” and the zero value “0x0000_{—}0000” are generated by the constant generator 21 and the zero generator 25, and the multiplexer 24 selectively outputs one of the maximum value, the zero value, and the held value of the sum-product result register 6 in accordance with the combination of the logic values outputted by the comparator 22 and the polarity judging unit 23. In the present example, the output of the comparator 22 is “1” and the output-of the polarity judging unit 23 is “0”, so that the multiplexer 24 outputs the maximum value “0x0000_{—}00FF” to the data bus 18.
According to control by the decoder 13, the selected maximum value outputted to the data bus 18 is transferred to the data register D1 that is indicated by the operand of the positive conversion saturation calculation instruction “MCSST D1”, and is held by the data register D1. This held value is then written into the RAM 10 as the element H11 for the 1^{st }row, 1^{st }column of the multiplication result matrix Hij.
When the sum-product has been completed for all of the elements in the first column of the compressed data matrix Fij and the elements on the first row of the coefficient matrix Gji, the sum-product processing is performed for the elements in the second column of the compressed data matrix Fij and the elements on the first row of the coefficient matrix Gji. When the calculation of “G11*F12+G12*F22+G13*F32+G14*F42+G15*F52+G16*F62+G17*F72+G18*F82” has been completed, the sum-product value is held by the sum-product result register 6.
Here, if there is a bit error when transferring the element F32 from the RAM 10, the sum-product value held by the sum-product result register 6 ends up at a negative value “0x0000_{—}86FF” that cannot be reproduced by the reproduction-related hardware.
After this, the next loop statement in the ROM 11, the positive conversion saturation calculation instruction “MCSST D1”, is written into the internal buffer of the decoder 13, this positive conversion saturation calculation instruction “MCSST D1” is decoded by the decoder 13.
When the positive conversion saturation calculation instruction “MCSST D1” is decoded by the decoder 13, the held value “0x000_{—}86FF” of the sum-product result register 6 is outputted onto the path P1. After this value has been outputted to the P1, the comparator 22 is activated by the decoder 13. The comparator 22 compares the held value of the sum-product result register 6 with the 32-bit coded integer “0x0000_{—}00FF” to see which is larger. Here, since the held value of the sum-product result register 6 “0x000_{—}86FF” exceeds the 32-bit coded integer “0x0000_{—}00FF”, the comparator 22 outputs the logic value “1” to the multiplexer 24.
The polarity judging unit 23 judges whether the 16^{th }bit counting from the LSB side in the value held by the sum-product result register 6 is “ON”. This judgement equates to a judgement as to whether the held sum-product value of the sum-product result register 6 is a negative number. The held value “0x000_{—}86FF” expressed in binary is “0000_{—}0000 0000_{—}0000 1000_{—}0110 1111_{—}1111”, so that the 16^{th }bit counting from the LSB side can be seen to be “1”. As a result, the logic value “1” is outputted to the multiplexer 24.
In the present case, the maximum value “0x0000_{—}00FF” and the zero value “0x0000_{—}0000” are generated by the constant generator 21 and the zero generator 25, and the multiplexer 24 selectively outputs one of the maximum value, the zero value, and the held value of the sum-product result register 6 in accordance with the combination of the logic values outputted by the comparator 22 and the polarity judging unit 23. In the present example, the output of the comparator 22 is “1” and the output of the polarity judging unit 23 is “1”, so that the multiplexer 24 outputs the zero value “0x0000_{—}0000” to the data bus 18.
According to control by the decoder 13, the selected zero value outputted to the data bus 18 is transferred to the data register D1 that is indicated by the operand of the positive conversion saturation calculation instruction “MCSST D1”, and is held by the data register D1. This held value is then written into the RAM 10 as the element H12 for the 1^{st }row, 2^{nd }column of the multiplication result matrix Hij.
By repeating the above processing and writing in the remaining elements in the matrix multiplication table, the matrix multiplication table is written into the RAM 10, and by using the result of this matrix multiplication as the result of an approximation calculation of an inverse DCT, the decompression processing of compressed data can be performed.
On the other hand, when the carry flag is set at “OFF”, the decoding stage of instruction 10:“MCSST D1” is performed at the same time as the execution stage of instruction 9:“BCS LP1_NEXT”. After this, the execution stage of instruction 10:“MCSST D1” is performed at the same time as the memory access stage of instruction 9:“BCS LP1_NEXT”. The positive conversion saturation calculation processing for the matrix multiplication result of one row of elements by one column of elements is performed when the instruction located before it is in the memory access stage, so that the processing can be seen to be performed without confusion in the pipeline.
In this way, even if the processor provided in the positive conversion saturation calculation circuit 3 needs to perform the matrix multiplication of one row of elements and one column of elements with a very high frequency, the positive conversion saturation calculation processing is expressed in the machine language program as a single instruction, so that there is no confusion in the pipeline. As a result, the processor can operate at high speed.
With the present embodiment described above, the positive conversion saturation calculation processing of sum-product values is performed by subjecting the sum-product value accumulated in the sum-product result register 6 to positive conversion saturation calculation processing, so that application programs for matrix approximation calculations required by decompression processing of video data and audio data can be easily coded using a remarkably small code size. Since there is a large reduction in code size, a large reduction-can be made in the amount of ROM that needs to be installed to store the program.
The positive conversion saturation calculation processing for the sum-product value is such that the positive correction processing and the saturation calculation processing are performed at the same time for the held value of the sum-product result register 6, so that the processing is performed at high speed. The positive conversion saturation calculation processing needs to be performed every time one row of elements is multiplied by one column of elements so that when 8 rows are multiplied by 8 rows, 64 executions of the positive conversion saturation calculation processing are necessary. However, since the positive correction processing and the saturation calculation processing are performed smoothly, each execution of the positive conversion saturation calculation processing is completed in a very short time. If the positive conversion saturation calculation processing is completed in a short time, the decompression processing for image data and audio data that require the matrix multiplication of a large amount of data can be performed at high speed.
Since the positive conversion saturation calculation processing performed by the positive conversion saturation calculation circuit 3 does not include branch instructions, the processor can perform high-speed pipeline processing without the risk of confusion in the pipeline. By executing such a high-performance pipeline, matrix multiplication can be performed at an improved speed.
Since positive conversion saturation calculation processing is performed without installing a specialized circuit for matrix multiplication, there is no loss in versatility for the processor. Accordingly, should a user wish to control the processor according to an original decompression processing program, this is still possible.
Applied Example for the First Embodiment
In this example, one of the data registers D0 to D2 is indicated as the read address for the positive conversion saturation calculation processing according to the positive conversion saturation calculation instruction “MCSST”, with the sum-product result register 6 being indicated as the storage address for the calculation. In this example, the instruction format of the positive conversion saturation calculation instruction “MCSST” is shown in FIG. 13. As shown in
By writing one of “11”, “00”, “10”, and “01” into the storage address indication field, one of the sum-product result register 6, the data register D0, the data register D1, and the data register D2 can be indicated as the storage address.
The instruction format of this positive conversion saturation calculation instruction “MCSST” has been amended so the internal construction of the operation execution apparatus 14 shown in
The path C6 is a path for transferring the held value of the data register D0, the data register D1, or the data register D2 on the second internal bus 16 to the positive conversion saturation calculation circuit 3.
In the same way, the path C7 is a path for transferring the held value of the data register D0, the data register D1, or the data register D2 on the first internal bus 15 to the positive conversion saturation calculation circuit 3.
The selector 30 outputs one of the held value of the data register D0, the data register D1, or the data register D2 transferred on the path C6 or C7, or the held value of the sum-product result register 6 to the comparator 22 in the positive conversion saturation calculation circuit 3 based on an indication of the storage address field in the positive conversion saturation calculation instruction “MCSST”.
The path C8 is a path for transferring the processing result of the positive conversion saturation calculation circuit 3 from the data bus 18, to which it has been transferred from the positive conversion saturation calculation circuit 3 via the path C4, to the sum-product result register 6.
By making the simple addition described above, the functioning of the positive conversion saturation calculation instruction “MCSST” can be extended in the present embodiment.
Second Embodiment
The second embodiment of the present invention executes positive conversion saturation calculation processing for a multiplication result when multiplication is performed by the ALU circuit 2. To perform positive conversion saturation calculation processing for a multiplication result, the second embodiment is constructed as shown in
A “MULBSST Dm,Dn” instruction is a multiplication instruction that indicates that the multiplication result should be further subjected to positive conversion saturation calculation processing. In other words, multiplication is performed using the lower 8 bits of the Dm register and the Dn register, and the positive conversion saturation calculation circuit 3 is then instructed to perform positive conversion saturation calculation processing on the coded 16-bit multiplication result.
By writing one of “01”, “10”, and “11” into the positive conversion saturation calculation processing field, it is possible to specify that the positive conversion saturation calculation processing with a rounding width of a 24-bit positive number, a 16-bit positive number, or an 8-bit positive number.
By writing one of “00”, “01”, “10”, and “11” into the multiplier read address indicating field and the multiplicand read address indicating field, any of the data register D0, the data register D1, the data register D2, and the sum-product result register 6 can be indicated as the read address register for the multiplier Fij and the multiplicand Gji.
When executing the positive conversion saturation calculation function multiplication instruction “MULBSST Dm,Dn”, the register file 1 outputs the held values of the registers with the register names indicated by the first and second operands. The ALU circuit 2 then multiplies the values of registers Dm and Dn and outputs the multiplication result. The same positive conversion saturation calculation processing as in the first embodiment is then performed on the multiplication result by the positive conversion saturation calculation circuit 3, and the result of the positive conversion saturation calculation processing is stored in the register indicated by the second operand of the positive conversion saturation calculation function multiplication instruction.
The following is an explanation of the operation of the above processor based on a matrix multiplication subroutine. First, a transfer instruction included in the matrix multiplication subroutine is written into the instruction buffer of the decoder 13 by the instruction fetch circuit 12, and the fetched transfer instruction is decoded by the decoder 13. As a result, the first row, first column element (F11) of the compressed data Fij which is composed of 8*8 matrix elements stored in the RAM 10 is transferred in the data register D0 and the first row, first column element (G11) of the coefficient data Gji which is also composed of 8*8 matrix elements is transferred into the data register D1.
In the matrix multiplication subroutine, the positive conversion saturation calculation function multiplication instruction “MULBSST D0,D1” follows the transfer instruction, so that this instruction is next fetched by the instruction fetch circuit 12 and written into the instruction buffer of the decoder 13, before being decoded by the decoder 13. When the decoder 13 decodes the positive conversion saturation calculation function multiplication instruction “MULBSST D0,D1”, the values F11 and G11 held by the data register D0 and the data register D1 are transferred to the first internal bus 15 and the second internal bus 16.
When the positive conversion saturation calculation function multiplication instruction “MULBSST D0,D1” is decoded, the held values F11, G11 of the read address instructions indicated by the first and second operands of the positive conversion saturation calculation function multiplication instruction “MULBSST D0,D1” are transferred onto the first internal bus 15 and the second internal bus 16, so that held value of the data register D0 and the held value of the data register D1 are multiplied by the ALU circuit 2, with the multiplication result being outputted onto the path P1. Here, however, a bit error occurs for F11, so that the multiplication result becomes “0x0000_{—}78FF”, which is a value that cannot be expressed using one byte.
When the decoder 13 has decoded the positive conversion saturation calculation function multiplication instruction “MULBSST D0,D1”, the decoder 13 also activates the positive conversion saturation calculation circuit 3. As a result, the comparator 22 compares the magnitude of the held value of the sum-product result register 6 with the 32-bit coded integer “0x0000_{—}00FF”. Here, since the held value of the sum-product result register 6 exceeds the maximum value “0x0000_{—}00FF” for the held value of the sum-product result register 6, the comparator 22 outputs the logic value “1” to the multiplexer 24.
The polarity judging unit 23 judges whether the sixteenth bit of the value held by the sum-product result register 6 is “ON”. This refers to a judgement as to whether the value held by the sum-product result register 6 is a negative number. When expressed in binary, the held number “0x0000_{—}78FF” is “0000_{—}0000 0000_{—}0000 0111_{—}1000 1111_{—}1111”, so that the sixteenth bit can be seen to be “1”. As a result, the polarity judging unit 23 outputs the logic value “0” to the multiplexer 24.
In the present case, the constant generator 21 generates the maximum value “0x0000_{—}00FF” and the zero generator 25 generates the zero value “0x0000_{—}0000”. The multiplexer 24 selects and outputs one of the maximum value, the zero value, and the held value of the sum-product result register 6 in accordance with the combination of the logic values outputted by the comparator 22 and the polarity judging unit 23. In the present example, the output of the comparator 22 is “1” and the output of the polarity judging unit 23 is “0”, so that the multiplexer 24 outputs the maximum value “0x0000_{—}00FF” generated by the constant generator 21 to the data bus 18.
According to control by the decoder 13, the value outputted to the data bus 18 is transferred to the data register D0 indicated by the operand of the positive conversion saturation calculation instruction “MULBSST D0”, and is held by the data register D0.
By means of the second embodiment described above, a calculation instruction that performs saturation calculation processing and positive conversion processing on the coded calculation result is provided, so that three types of processing composed of calculation processing, positive conversion processing, and saturation calculation processing can be performed in one step, meaning that positive conversion saturation calculation processing is performed in the same step as the calculation processing. As a result, the effective number of execution steps required by positive conversion saturation calculation processing is reduced to zero.
It should be noted here that this second embodiment has been described as performing a rounding process for an uncoded 8-bit width, although the maximum number can be freely set at any positive integer.
Although the present invention has been fully described by way of examples with reference to accompanying drawings, it is to be noted that various changes and modifications will be apparent to those skilled in the art. Therefore, unless such changes and modifications depart from the scope of the present invention, they should be construed as being included therein.
Cited Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US4935890 * | Jan 3, 1989 | Jun 19, 1990 | Kabushiki Kaisha Toshiba | Format converting circuit for numeric data |
US4945507 * | Jun 12, 1989 | Jul 31, 1990 | Nec Corporation | Overflow correction circuit |
US5235533 * | May 11, 1992 | Aug 10, 1993 | Intel Corporation | Store rounding in a floating point unit |
US5251166 * | Apr 7, 1992 | Oct 5, 1993 | Nec Corporation | Redundant binary type digital operation unit |
US5402368 * | Dec 9, 1993 | Mar 28, 1995 | Fujitsu Limited | Computing unit and digital signal processor using the same |
US5448509 * | Dec 8, 1993 | Sep 5, 1995 | Hewlett-Packard Company | Efficient hardware handling of positive and negative overflow resulting from arithmetic operations |
US5504697 * | Dec 21, 1994 | Apr 2, 1996 | Nec Corporation | Limiter circuit producing data by use of comparison in effective digit number of data |
US5508951 * | Nov 10, 1994 | Apr 16, 1996 | Matsushita Electric Industrial Co., Ltd. | Arithmetic apparatus with overflow correction means |
US5696709 * | Mar 31, 1995 | Dec 9, 1997 | International Business Machines Corporation | Program controlled rounding modes |
US5801977 * | Apr 7, 1997 | Sep 1, 1998 | Hewlett-Packard Company | System and method for clipping integers |
US5812439 * | Oct 10, 1995 | Sep 22, 1998 | Microunity Systems Engineering, Inc. | Technique of incorporating floating point information into processor instructions |
US5889980 * | Sep 3, 1997 | Mar 30, 1999 | International Business Machines Corporation | Mode independent support of format conversion instructions for hexadecimal and binary floating point processing |
US5917740 * | Jun 24, 1997 | Jun 29, 1999 | Sun Microsystems, Inc. | Apparatus for reducing a computational result to the range boundaries of a signed 16-bit integer in case of overflow |
US5974540 * | Dec 1, 1997 | Oct 26, 1999 | Matsushita Electric Industrial Co., Ltd. | Processor which can favorably execute a rounding process composed of positive conversion and saturated calculation processing |
US6029184 * | Jun 17, 1997 | Feb 22, 2000 | Sun Microsystems, Inc. | Method of performing unsigned operations with signed instructions in a microprocessor |
US6058410 * | Dec 2, 1996 | May 2, 2000 | Intel Corporation | Method and apparatus for selecting a rounding mode for a numeric operation |
EP0657804A1 | Dec 5, 1994 | Jun 14, 1995 | Hewlett-Packard Company | Overflow control for arithmetic operations |
EP0768169A1 | Apr 26, 1996 | Apr 16, 1997 | Asahi Lite Optical Co., Ltd. | Method of manufacturing plastic lenses |
GB2300054A | Title not available | |||
JPH07182141A | Title not available | |||
JPH07210368A * | Title not available | |||
JPH08272591A | Title not available | |||
JPS5856032A * | Title not available | |||
WO1996017292A1 | Dec 1, 1995 | Jun 6, 1996 | Intel Corporation | Microprocessor with compare operation of composite operands |
Reference | ||
---|---|---|
1 | * | "Low-Power Multimedia RISC," by K. Nadehara, 8207 IEEE Micro 15 (1995) Dec., No. 6. |
2 | * | "Subword Parallelism with MAX-2," by R. Lee, IEEE Micro Aug. 1, 1996, vol. 16, No. 4. |
3 | Lee, Ruby B., "Subword Parallelism with Max-2", IEEE Micro, US, IEEE, Inc., New York, vol. 16, No. , (Aug. 1, 1996), pp. 51-59, XP000596513. | |
4 | Nadehara, Kouhei, et al., "Low-Power Multimedia RISC", IEEE Micro, US, IEEE Inc., New York, vol. 15, No. 6, (Dec. 1, 1995), pp. 20-29, XP)))538227, ISSN: 0272-1732. |
Citing Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US7340498 * | May 28, 2003 | Mar 4, 2008 | Lg Electronics Inc. | Apparatus and method for determining fixed point in mobile communication system |
US20040047436 * | May 28, 2003 | Mar 11, 2004 | Park Tan Joong | Apparatus and method for determining fixed point in mobile communication system |
U.S. Classification | 712/221, 708/551, 708/552 |
International Classification | G06F7/57, G06F7/544, G06F9/302, G06F7/38 |
Cooperative Classification | G06F7/57, G06F7/5443, G06F7/49921, G06F9/3001 |
European Classification | G06F7/57, G06F9/30A1A, G06F7/544A |
Date | Code | Event | Description |
---|---|---|---|
Oct 23, 2008 | FPAY | Fee payment | Year of fee payment: 8 |
Sep 18, 2012 | CC | Certificate of correction | |
Sep 28, 2012 | FPAY | Fee payment | Year of fee payment: 12 |
Feb 23, 2015 | AS | Assignment | Owner name: GODO KAISHA IP BRIDGE 1, JAPAN Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:PANASONIC CORPORATION;REEL/FRAME:035003/0660 Effective date: 20141119 |
May 17, 2016 | CC | Certificate of correction |