Patent Grant
7424505
1. Field of Invention
In particular, the invention relates to the field of computer systems. More specifically, the invention relates to the area of packed data operations.
2. Description of Related Art
In typical computer systems, processors are implemented to operate on values represented by a large number of bits (e.g., 64) using instructions that produce one result. For example, the execution of an add instruction will add together a first 64-bit value and a second 64-bit value and store the result as a third 64-bit value. However, multimedia applications (e.g., applications targeted at computer supported cooperation (CSC—the integration of teleconferencing with mixed media data manipulation), 2D/3D graphics, image processing, video compression/decompression, recognition algorithms and audio manipulation) require the manipulation of large amounts of data which may be represented in a small number of bits. For example, graphical data typically requires 8 or 16 bits and sound data typically requires 8 or 16 bits. Each of these multimedia application requires one or more algorithms, each requiring a number of operations. For example, an algorithm may require an add, compare and shift operation.
To improve efficiency of multimedia applications (as well as other applications that have the same characteristics), prior art processors provide packed data formats. A packed data format is one in which the bits typically used to represent a single value are broken into a number of fixed sized data elements, each of which represents a separate value. For example, a 64-bit register may be broken into two 32-bit elements, each of which represents a separate 32-bit value. In addition, these prior art processors provide instructions for separately manipulating each element in these packed data types in parallel. For example, a packed add instruction adds together corresponding data elements from a first packed data and a second packed data. Thus, if a multimedia algorithm requires a loop containing five operations that must be performed on a large number of data elements, it is desirable to pack the data and perform these operations in parallel using packed data instructions. In this manner, these processors can more efficiently process multimedia applications.
However, if the loop of operations contains an operation that cannot be performed by the processor on packed data (i.e., the processor lacks the appropriate instruction), the data will have to be unpacked to perform the operation. For example, if the multimedia algorithm requires an add operation and the previously described packed add instruction is not available, the programmer must unpack both the first packed data and the second packed data (i.e., separate the elements comprising both the first packed data and the second packed data), add the separated elements together individually, and then pack the results into a packed result for further packed processing. The processing time required to perform such packing and unpacking often negates the performance advantage for which packed data formats are provided. Therefore, it is desirable to incorporate in a computer system a set of packed data instructions that provide all the required operations for typical multimedia algorithms. However, due to the limited die area on today's general purpose microprocessors, the number of instructions which may be added is limited. Therefore, it is desirable to invent instructions that provide both versatility (i.e. instructions which may be used in a wide variety of multimedia algorithms) and the greatest performance advantage.
One prior art technique for providing operations for use in multimedia algorithms is to couple a separate digital signaling processor (DSP) to an existing general purpose processor (e.g., The Intel® 486 manufactured by Intel Corporation of Santa Clara, Calif.). The general purpose processor allocates jobs that can be performed using packed data (e.g., video processing) to the DSP.
One such prior art DSP includes a multiply accumulate instruction that adds to an accumulation value the results of multiplying together two values. (see Kawakami, Yuichi, et al., “A Single-Chip Digital Signal Processor for Voiceband Applications”, IEEE International Solid-State Circuits Conference, 1980, pp. 40–41). An example of the multiply accumulate operation for this DSP is shown below in Table 1, where the instruction is performed on the data values A1 and B1 accessed as Source1 and Source2, respectively.
One limitation of this prior art instruction is its limited efficiency—i.e., it only operates on 2 values and an accumulation value. For example, to multiply and accumulate two sets of 2 values requires the following 2 instructions performed serially: 1) multiply accumulate the first value from the first set, the first value from the second set, and an accumulation value of zero to generate an intermediate accumulation value; 2) multiply accumulate the second value from the first set, the second value from the second set, and the intermediate accumulation value to generate the result.
Another prior art DSP includes a multiply accumulate instruction that operates on two sets of two values and an accumulation value (See “Digital Signal Processor with Parallel Multipliers”, U.S. Pat. No. 4,771,379—referred to herein as the “Ando et al.” reference). An example of the multiply accumulate instruction for this DSP is shown below in Table 2, where the instruction is performed on the data values A1, A2, B1 and B2 accessed as Source 1–4, repectively.
Using this prior art technique, two sets of 2 values are multiplied and then added to an accumulation value in one instruction.
This multiply accumulate instruction has limited versatility because it always adds to the accumulation value. As a result, it is difficult to use the instruction for operations other than multiply accumulate. For example, the multiplication of complex numbers is commonly used in multimedia applications. The multiplication of two complex number (e.g., r1 i1 and r2 i2) is performed according to the following equation:
Real Component=r1·r2−i1·i2
Imaginary Component=r1·i2+r2·i1
This prior art DSP cannot perform the function of multiplying together two complex numbers using one multiply accumulate instruction.
The limitations of this multiply accumulate instruction can be more clearly seen when the result of such a calculation is needed in a subsequent multiplication operation rather than an accumulation. For example, if the real component were calculated using this prior art DSP, the accumulation value would need to be initialized to zero in order to correctly compute the result. Then the accumulation value would again need to be initialized to zero in order to calculate the imaginary component. To perform another complex multiplication on the resulting complex number and a third complex number (e.g., r3, i3), the resulting complex number must be rescaled and stored into the acceptable memory format and the accumulation value must again be initialized to zero. Then, the complex multiplication can be performed as described above. In each of these operations the ALU, which is devoted to the accumulation value, is superfluous hardware and extra instructions are needed to re-initialize this accumulation value. These extra instructions would otherwise have been unnecessary.
A further limitation of this prior art technique is that the data must be accessed through expensive multi-ported memory. This is because the multipliers are connected directly with data memories. Therefore the amount of parallelism which can be exploited is limited to a small number by the cost of the interconnection, and the fact that this interconnection is not decoupled from the instruction.
The Ando, et al. reference also describes that an alternative to this expensive interconnection is to introduce a delay for each subsequent pair of data to be multiplied. This solution diminishes any performance advantages to those provided by the solution previously shown in Table 1.
Furthermore, the notion of multi-ported memory or of pipelined accesses to memory entails the use of multiple addresses. This explicit use of one address per datum, clearly demonstrates that the critical notion of packed data is not employed in this prior art.
A method and apparatus for including in a processor instructions for performing multiply-add operations on packed data is described. In one embodiment, a processor is coupled to a memory. The memory has stored therein a first packed data and a second packed data. The processor performs operations on data elements in the first packed data and the second packed data to generate a third packed data in response to receiving an instruction. At least two of the data elements in this third packed data storing the result of performing multiply-add operations on data elements in the first and second packed data.
The invention is illustrated by way of example, and not limitation, in the figures. Like references indicate similar elements.
a illustrates in-register packed data representations according to one embodiment of the invention.
b illustrates in-register packed data representations according to one embodiment of the invention.
c illustrates in-register packed data representations according to one embodiment of the invention.
a illustrates a control signal format for indicating the use of packed data according to one embodiment of the invention.
b illustrates a second control signal format for indicating the use of packed data according to one embodiment of the invention.
In the following description, numerous specific details are set forth to provide a thorough understanding of the invention. However, it is understood that the invention may be practiced without these specific details. In other instances, well-known circuits, structures and techniques have not been shown in detail in order not to obscure the invention.
To provide a foundation for understanding the description of the embodiments of the invention, the following definitions are provided.
This application describes a method and apparatus for including in a processor instructions for performing multiply-add and multiply-subtract operations on packed data. In one embodiment, two multiply-add operations are performed using a single multiply-add instruction as shown below in Table 3a and Table 3b—Table 3a shows a simplified representation of the disclosed multiply-add instruction, while Table 3b shows a bit level example of the disclosed multiply-add instruction.
Thus, the described embodiment of the multiple-add instruction multiplies together corresponding 16-bit data elements of Source1 and Source2 generating four 32-bit intermediate results. These 32-bit intermediate results are summed by pairs producing two 32-bit results that are packed into their respective elements of a packed result. As further described later, alternative embodiment may vary the number of bits in the data elements, intermediate results, and results. In addition, alternative embodiment may vary the number of data elements used, the number of intermediate results generated, and the number of data elements in the resulting packed data. The multiply-subtract operation is the same as the multiply-add operation, except the adds are replaced with subtracts. The operation of an example multiply-subtract instruction is shown below in Table 4.
Of course, alternative embodiments may implement variations of these instructions. For example, alternative embodiments may include an instruction which performs at least one multiply-add operation or at least one multiply-subtract operation. As another example, alternative embodiments may include an instruction which performs at least one multiply-add operation in combination with at least one multiply-subtract operation. As another example, alternative embodiments may include an instruction which perform multiply-add operation(s) and/or multiply-subtract operation(s) in combination with some other operation.
Execution unit 130 is used for executing instructions received by processor 109. In addition to recognizing instructions typically implemented in general purpose processors, execution unit 130 recognizes instructions in packed instruction set 140 for performing operations on packed data formats. Packed instruction set 140 includes instructions for supporting multiply-add and/or multiply-subtract operations. In addition, packed instruction set 140 may also include instructions for supporting a pack operation, an unpack operation, a packed add operation, a packed subtract operation, a packed multiply operation, a packed shift operation, a packed compare operation, a population count operation, and a set of packed logical operations (including packed AND, packed ANDNOT, packed OR, and packed XOR) as described in “A Set of Instructions for Operating on Packed Data ,” filed on Aug. 31, 1995, application Ser. No. 521,360.
Execution unit 130 is coupled to register file 150 by internal bus 170. Register file 150 represents a storage area on processor 109 for storing information, including data. It is understood that one aspect of the invention is the described instruction set for operating on packed data. According to this aspect of the invention, the storage area used for storing the packed data is not critical. However, one embodiment of the register file 150 is later described with reference to
Another device which may be coupled to bus 101 is a hard copy device 124 which may be used for printing instructions, data, or other information on a medium such as paper, film, or similar types of media. Additionally, computer system 100 can be coupled to a device for sound recording, and/or playback 125, such as an audio digitizer coupled to a microphone for recording information. Further, the device may include a speaker which is coupled to a digital to analog (D/A) converter for playing back the digitized sounds.
Also, computer system 100 can be a terminal in a computer network (e.g., a LAN). Computer system 100 would then be a computer subsystem of a computer network. Computer system 100 optionally includes video digitizing device 126. Video digitizing device 126 can be used to capture video images that can be transmitted to others on the computer network.
In one embodiment, the processor 109 additionally supports an instruction set which is compatible with the x86 instruction set used by existing processors (such as the Pentium® processor) manufactured by Intel Corporation of Santa Clara, Calif. Thus, in one embodiment, processor 109 supports all the operations supported in the IA™—Intel Architecture, as defined by Intel Corporation of Santa Clara, Calif. (see Microprocessors, Intel Data Books volume 1 and volume 2, 1992 and 1993, available from Intel of Santa Clara, Calif.). As a result, processor 109 can support existing x86 operations in addition to the operations of the invention. While the invention is described as being incorporated into an x86 based instruction set, alternative embodiments could incorporate the invention into other instruction sets. For example, the invention could be incorporated into a 64-bit processor using a new instruction set.
In one embodiment, the registers 209 are used for both packed data and floating point data. In one such embodiment, the processor 109, at any given time, must treat the registers 209 as being either stack referenced floating point registers or non-stack referenced packed data registers. In this embodiment, a mechanism is included to allow the processor 109 to switch between operating on registers 209 as stack referenced floating point registers and non-stack referenced packed data registers. In another such embodiment, the processor 109 may simultaneously operate on registers 209 as non-stack referenced floating point and packed data registers. As another example, in another embodiment, these same registers may be used for storing integer data.
Of course, alternative embodiments may be implemented to contain more or less sets of registers. For example, an alternative embodiment may include a separate set of floating point registers for storing floating point data. As another example, an alternative embodiment may including a first set of registers, each for storing control/status information, and a second set of registers, each capable of storing integer, floating point, and packed data. As a matter of clarity, the registers of an embodiment should not be limited in meaning to a particular type of circuit. Rather, a register of an embodiment need only be capable of storing and providing data, and performing the functions described herein.
The various sets of registers (e.g., the integer registers 201, the registers 209) may be implemented to include different numbers of registers and/or to different size registers. For example, in one embodiment, the integer registers 201 are implemented to store thirty-two bits, while the registers 209 are implemented to store eighty bits (all eighty bits are used for storing floating point data, while only sixty-four are used for packed data). In addition, registers 209 contains eight registers, R0 212a through R7 212h. R1 212a, R2 212b and R3 212c are examples of individual registers in registers 209. Thirty-two bits of a register in registers 209 can be moved into an integer register in integer registers 201. Similarly, a value in an integer register can be moved into thirty-two bits of a register in registers 209. In another embodiment, the integer registers 201 each contain 64 bits, and 64 bits of data may be moved between the integer register 201 and the registers 209.
At step 301, the decoder 165 receives a control signal from either the cache 160 or bus 101. Decoder 165 decodes the control signal to determine the operations to be performed.
At step 302, Decoder 165 accesses the register file 150, or a location in memory. Registers in the register file 150, or memory locations in the memory, are accessed depending on the register address specified in the control signal. For example, for an operation on packed data, the control signal can include SRC1, SRC2 and DEST register addresses. SRC1 is the address of the first source register. SRC2 is the address of the second source register. In some cases, the SRC2 address is optional as not all operations require two source addresses. If the SRC2 address is not required for an operation, then only the SRC1 address is used. DEST is the address of the destination register where the result data is stored. In one embodiment, SRC1 or SRC2 is also used as DEST. SRC1, SRC2 and DEST are described more fully in relation to
In another embodiment of the invention, any one, or all, of SRC1, SRC2 and DEST, can define a memory location in the addressable memory space of processor 109. For example, SRC1 may identify a memory location in main memory 104, while SRC2 identifies a first register in integer registers 201 and DEST identifies a second register in registers 209. For simplicity of the description herein, the invention will be described in relation to accessing the register file 150. However, these accesses could be made to memory instead.
At step 303, execution unit 130 is enabled to perform the operation on the accessed data. At step 304, the result is stored back into register file 150 according to requirements of the control signal.
Packed word 402 is sixty-four bits long and contains four word 402 data elements. Each word 402 data element contains sixteen bits of information.
Packed doubleword 403 is sixty-four bits long and contains two doubleword 403 data elements. Each doubleword 403 data element contains thirty-two bits of information.
a through 5c illustrate the in-register packed data storage representation according to one embodiment of the invention. Unsigned packed byte in-register representation 510 illustrates the storage of an unsigned packed byte 401 in one of the registers R0 212a through R7 212h. Information for each byte data element is stored in bit seven through bit zero for byte zero, bit fifteen through bit eight for byte one, bit twenty-three through bit sixteen for byte two, bit thirty-one through bit twenty-four for byte three, bit thirty-nine through bit thirty-two for byte four, bit forty-seven through bit forty for byte five, bit fifty-five through bit forty-eight for byte six and bit sixty-three through bit fifty-six for byte seven. Thus, all available bits are used in the register. This storage arrangement increases the storage efficiency of the processor. As well, with eight data elements accessed, one operation can now be performed on eight data elements simultaneously. Signed packed byte in-register representation 511 illustrates the storage of a signed packed byte 401. Note that the eighth bit of every byte data element is the sign indicator.
Unsigned packed word in-register representation 512 illustrates how word three through word zero are stored in one register of registers 209. Bit fifteen through bit zero contain the data element information for word zero, bit thirty-one through bit sixteen contain the information for data element word one, bit forty-seven through bit thirty-two contain the information for data element word two and bit sixty-three through bit forty-eight contain the information for data element word three. Signed packed word in-register representation 513 is similar to the unsigned packed word in-register representation 512. Note that the sixteenth bit of each word data element is the sign indicator.
Unsigned packed doubleword in-register representation 514 shows how registers 209 store two doubleword data elements. Doubleword zero is stored in bit thirty-one through bit zero of the register. Doubleword one is stored in bit sixty-three through bit thirty-two of the register. Signed packed doubleword in-register representation 515 is similar to unsigned packed doubleword in-register representation 514. Note that the necessary sign bit is the thirty-second bit of the doubleword data element.
As mentioned previously, registers 209 may be used for both packed data and floating point data. In this embodiment of the invention, the individual programming processor 109 may be required to track whether an addressed register, R0 212a for example, is storing packed data or floating point data. In an alternative embodiment, processor 109 could track the type of data stored in individual registers of registers 209. This alternative embodiment could then generate errors if, for example, a packed addition operation were attempted on floating point data.
The following describes one embodiment of the control signal formats used by processor 109 to manipulate packed data. In one embodiment of the invention, control signals are represented as thirty-two bits. Decoder 165 may receive the control signal from bus 101. In another embodiment, decoder 165 can also receive such control signals from cache 160.
a illustrates a control signal format for indicating the use of packed data according to one embodiment of the invention. Operation field OP 601, bit thirty-one through bit twenty-six, provides information about the operation to be performed by processor 109; for example, packed addition, packed subtraction, etc. SRC1 602, bit twenty-five through twenty, provides the source register address of a register in registers 209. This source register contains the first packed data, Source1, to be used in the execution of the control signal. Similarly, SRC2 603, bit nineteen through bit fourteen, contains the address of a register in registers 209. This second source register contains the packed data, Source2, to be used during execution of the operation. DEST 605, bit five through bit zero, contains the address of a register in registers 209. This destination register will store the result packed data, Result, of the packed data operation.
Control bits SZ 610, bit twelve and bit thirteen, indicates the length of the data elements in the first and second packed data source registers. If SZ 610 equals 012, then the packed data is formatted as packed byte 401. If SZ 610 equals 102, then the packed data is formatted as packed word 402. SZ 610 equaling 002 or 112 is reserved, however, in another embodiment, one of these values could be used to indicate packed doubleword 403.
Control bit T 611, bit eleven, indicates whether the operation is to be carried out with saturate mode. If T 611 equals one, then a saturating operation is performed. If T 611 equals zero, then a non-saturating operation is performed. Saturating operations will be described later.
Control bit S 612, bit ten, indicates the use of a signed operation. If S 612 equals one, then a signed operation is performed. If S 612 equals zero, then an unsigned operation is performed.
b illustrates a second control signal format for indicating the use of packed data according to one embodiment of the invention. This format corresponds with the general integer opcode format described in the “Pentium Processor Family User's Manual,” available from Intel Corporation, Literature Sales, P.O. Box 7641, Mt. prospect, Ill., 60056-7641. Note that OP 601, SZ 610, T 611, and S 612 are all combined into one large field. For some control signals, bits three through five are SRC1 602. In one embodiment, where there is a SRC1 602 address, then bits three through five also correspond to DEST 605. In an alternate embodiment, where there is a SRC2 603 address, then bits zero through two also correspond to DEST 605. For other control signals, like a packed shift immediate operation, bits three through five represent an extension to the opcode field. In one embodiment, this extension allows a programmer to include an immediate value with the control signal, such as a shift count value. In one embodiment, the immediate value follows the control signal. This is described in more detail in the “Pentium Processor Family User's Manual,” in appendix F, pages F-1 through F-3. Bits zero through two represent SRC2 603. This general format allows register to register, memory to register, register by memory, register by register, register by immediate, register to memory addressing. Also, in one embodiment, this general format can support integer register to register, and register to integer register addressing.
As mentioned previously, T 611 indicates whether operations optionally saturate. Where the result of an operation, with saturate enabled, overflows or underflows the range of the data, the result will be clamped. Clamping means setting the result to a maximum or minimum value should a result exceed the range's maximum or minimum value. In the case of underflow, saturation clamps the result to the lowest value in the range and in the case of overflow, to the highest value. The allowable range for each data format is shown in Table 5.
As mentioned above, T 611 indicates whether saturating operations are being performed. Therefore, using the unsigned byte data format, if an operation's result=258 and saturation was enabled, then the result would be clamped to 255 before being stored into the operation's destination register. Similarly, if an operation's result=−32999 and processor 109 used signed word data format with saturation enabled, then the result would be clamped to −32768 before being stored into the operation's destination register.
In one embodiment of the invention, the SRC1 register contains packed data (Source1), the SRC2 register contains packed data (Source2), and the DEST register will contain the result (Result) of performing the multiply-add or multiply-subtract instruction on Source1 and Source2. In the first step of the multiply-add and multiply-subtract instruction, Source 1 will have each data element independently multiplied by the respective data element of Source2 to generate a set of respective intermediate results. These intermediate results are summed by pairs to generate the Result for the multiply-add instruction. In contrast, these intermediate results are subtracted by pairs to generate the Result for the multiply-subtract instruction.
In one embodiment of the invention, the multiply-add and multiply-subtract instructions operate on signed packed data and truncate the results to avoid any overflows. In addition, these instructions operate on packed word data and the Result is a packed double word. However, alternative embodiments could support these instructions for other packed data types.
At step 701, decoder 165 decodes the control signal received by processor 109. Thus, decoder 165 decodes: the operation code for a multiply-add instruction or a multiply-subtract instruction.
At step 702, via internal bus 170, decoder 165 accesses registers 209 in register file 150 given the SRC1 602 and SRC2 603 addresses. Registers 209 provide execution unit 130 with the packed data stored in the SRC1 602 register (Source1), and the packed data stored in SRC2 603 register (Source2). That is, registers 209 communicate the packed data to execution unit 130 via internal bus 170.
At step 703, decoder 165 enables execution unit 130 to perform the instruction. If the instruction is a multiply-add instruction, flow passes to step 714. However, if the instruction is a multiply-subtract instruction, flow passes to step 715.
In step 714, the following is performed. Source1 bits fifteen through zero are multiplied by Source2 bits fifteen through zero generating a first 32-bit intermediate result (Intermediate Result 1). Source1 bits thirty-one through sixteen are multiplied by Source2 bits thirty-one through sixteen generating a second 32-bit intermediate result (Intermediate Result 2). Source1 bits forty-seven through thirty-two are multiplied by Source2 bits forty-seven through thirty-two generating a third 32-bit intermediate result (Intermediate Result 3). Source1 bits sixty-three through forty-eight are multiplied by Source2 bits sixty-three through forty-eight generating a fourth 32-bit intermediate result (Intermediate Result 4). Intermediate Result 1 is added to Intermediate Result 2 generating Result bits thirty-one through 0, and Intermediate Result 3 is added to Intermediate Result 4 generating Result bits sixty-three through thirty-two.
Step 715 is the same as step 714, with the exception that Intermediate Result 1 Intermediate Result 2 are subtracted to generate bits thirty-one through 0 of the Result, and Intermediate Result 3 and Intermediate Result 4 are subtracted to generate bits sixty-three through thirty-two of the Result.
Different embodiments may perform the multiplies and adds/subtracts serially, in parallel, or in some combination of serial and parallel operations.
At step 720, the Result is stored in the DEST register.
In one embodiment, the multiply-add and multiply-subtract instructions can execute on multiple data elements in the same number of clock cycles as a single multiply on unpacked data. To achieve execution in the same number of clock cycles, parallelism is used. That is, registers are simultaneously instructed to perform the multiply-add/subtract operations on the data elements. This is discussed in more detail below.
Packed multiply-adder/subtractor 801 has the following inputs: Source1 [63:0] 831, Source2[63:0] 833, and Enable 880. Packed multiply-adder/subtractor 801 includes four 16×16 multiplier circuits: 16×16 multiplier A 810, 16×16 multiplier B 811, 16×16 multiplier C 812 and 16×16 multiplier D 813. 16×16 multiplier A 810 has as inputs Source1[15:0] and Source2[15:0]. 16×16 multiplier B 811 has as inputs Source1 [31:16] and Source2[31:16]. 16×16 multiplier C 812 has as inputs Source1[47:32] and Source2[47:32]. 16×16 multiplier D 813 has as inputs Source1[63:48] and Source2[63:48]. The 32-bit intermediate results generated by 16×16 multiplier A 810 and 16×16 multiplier B 811 are received by adder/subtractor 1350, while the 32-bit intermediate results generated by 16×16 multiplier C 812 and 16×16 multiplier D 813 are received by adder/subtractor 851.
Based on whether the current instruction is a multiply/add or multiply/subtract instruction, adder/subtractor 850 and adder/subtractor 851 add or subtract their respective 32-bit inputs. The output of adder/subtractor 850 (i.e., Result bits 31 through zero of the Result) and the output of adder/subtractor 851 (i.e., bits 63 through 32 of the Result) are combined into the 64-bit Result and communicated to Result Register 871.
In one embodiment, each of adder/subtractor 851 and adder/subtractor 850 are composed of four 8-bit adders/subtractors with the appropriate propagation delays. However, alternative embodiments could implement adder/subtractor 851 and adder/subtractor 850 in any number of ways (e.g., two 32-bit adders/subtractors).
To perform the equivalent of these multiply-add or multiply-subtract instructions in prior art processors which operate on unpacked data, four separate 64-bit multiply operations and two 64-bit add or subtract operations, as well as the necessary load and store operations, would be needed. This wastes data lines and circuitry that are used for the bits that are higher than bit sixteen for Source1 and Source2, and higher than bit thirty two for the Result. As well, the entire 64-bit result generated by the prior art processor may not be of use to the programmer. Therefore, the programmer would have to truncate each result.
Performing the equivalent of this multiply-add instruction using the prior art DSP processor described with reference to Table 1 requires one instruction to zero the accumulation value and four multiply accumulate instructions. Performing the equivalent of this multiply-add instruction using the prior art DSP processor described with reference to Table 2 requires one instruction to zero the accumulation value and 2-accumulate instructions.
As previously described, the prior art multiply accumulate instructions always add the results of their multiplications to an accumulation value. This accumulation value becomes a bottleneck for performing operations other than multiplying and accumulating (e.g., the accumulation value must be cleared each time a new set of operations is required which do not require the previous accumulation value). This accumulation value also becomes a bottleneck if operations, such as rounding, need to be performed before accumulation.
In contrast, the disclosed multiply-add and multiply-subtract instructions do not carry forward an accumulation value. As a result, these instructions are easier to use in a wider variety of algorithms. In addition, software pipelining can be used to achieve comparable throughput. To illustrate the versatility of the multiply-add instruction, several example multimedia algorithms are described below. Some of these multimedia algorithms use additional packed data instructions. The operation of these additional packed data instructions are shown in relation to the described algorithms. For a further description of these packed data instructions, see “A Set of Instructions for Operating on Packed Data,” filed on Aug. 31, 1995, application Ser. No. 521,360. Of course, other packed data instructions could be used. In addition, a number of steps requiring the use of general purpose processor instructions to manage data movement, looping, and conditional branching have been omitted in the following examples.
1) Multiplication of Complex Numbers
The disclosed multiply-add instruction can be used to multiply two complex numbers in a single instruction as shown in Table 6a. As previously described, the multiplication of two complex number (e.g., r1 i1 and r2 i2) is performed according to the following equation:
Real Component=r1·r2−i1·i2
Imaginary Component=r1·i2+r2·i1
If this instruction is implemented to be completed every clock cycle, the invention can multiply two complex numbers every clock cycle.
As another example, Table 6b shows the instructions used to multiply together three complex numbers.
2) Multiply Accumulation Operations
The disclosed multiply-add instructions can also be used to multiply and accumulate values. For example, two sets of four data elements (A1-4 and B1-4) may be multiplied and accumulated as shown below in Table 7. In one embodiment, each of the instructions shown in Table 7 is implemented to complete each clock cycle.
If the number of data elements in each set exceeds 8 and is a multiple of 4, the multiplication and accumulation of these sets requires fewer instructions if performed as shown in table 8 below.
As another example, Table 9 shows the separate multiplication and accumulation of sets A and B and sets C and D, where each of these sets includes 2 data elements.
As another example, Table 10 shows the separate multiplication and accumulation of sets A and B and sets C and D, where each of these sets includes 4 data elements.
3) Dot Product Algorithms
Dot product (also termed as inner product) is used in signal processing and matrix operations. For example, dot product is used when computing the product of matrices, digital filtering operations (such as FIR and IIR filtering), and computing correlation sequences. Since many speech compression algorithms (e.g., GSM, G.728, CELP, and VSELP) and Hi-Fi compression algorithms (e.g., MPEG and subband coding) make extensive use of digital filtering and correlation computations, increasing the performance of dot product increases the performance of these algorithms. The dot product of two length N sequences A and B is defined as:
Performing a dot product calculation makes extensive use of the multiply accumulate operation where corresponding elements of each of the sequences are multiplied together, and the results are accumulated to form the dot product result.
The dot product calculation can be performed using the multiply-add instruction. For example if the packed data type containing four sixteen-bit elements is used, the dot product calculation may be performed on two sequences each containing four values by:
For vectors with more than just a few elements the method shown in Table 10 is used and the final results are added together at the end. Other supporting instructions include the packed OR and XOR instructions for initializing the accumulator register, the packed shift instruction for shifting off unwanted values at the final stage of computation. Loop control operations are accomplished using instructions already existing in the instruction set of processor 109.
4) Discrete Cosign Transform Algorithms
Discrete Cosine Transform (DCT) is a well known function used in many signal processing algorithms. Video and image compression algorithms, in particular, make extensive use of this transform.
In image and video compression algorithms, DCT is used to transform a block of pixels from the spatial representation to the frequency representation. In the frequency representation, the picture information is divided into frequency components, some of which are more important than others. The compression algorithm selectively quantizes or discards the frequency components that do not adversely affect the reconstructed picture contents. In this manner, compression is achieved.
There are many implementations of the DCT, the most popular being some kind of fast transform method modeled based on the Fast Fourier Transform (FFT) computation flow. In the fast transform, an order N transform is broken down to a combination of order N/2 transforms and the result recombined. This decomposition can be carried out until the smallest order 2 transform is reached. This elementary 2 transform kernel is often referred to as the butterfly operation. The butterfly operation is expressed as follows:
X=a*x+b*y
Y=c*x−d*y
where a, b, c and d are termed the coefficients, x and y are the input data, and X and Y are the transform output.
The multiply-add allows the DCT calculation to be performed using packed data in the following manner:
In some situations, the coefficients of the butterfly operation are 1. For these cases, the butterfly operation degenerates into just adds and subtracts that may be performed using the packed add and packed subtract instructions.
An IEEE document specifies the accuracy with which inverse DCT should be performed for video conferencing. (See, IEEE Circuits and Systems Society, “IEEE Standard Specifications for the Implementations of 8×8 Inverse Discrete Cosine Transform,” IEEE Std. 1180–1990, IEEE Inc. 345 East 47th St., NY, N.Y. 10017, USA, Mar. 18, 1991). The required accuracy is met by the disclosed multiply-add instruction because it uses 16-bit inputs to generate 32-bit outputs.
In this manner, the described multiply-add instruction can be used to improve the performance of a number of different algorithms, including algorithms that require the multiplication of complex numbers, algorithms that require transforms, and algorithms that require multiply accumulate operations. As a result, this multiply-add instruction can be used in a general purpose processor to improve the performance of a greater number algorithms than the described prior art instructions.
While the described embodiment uses 16-bit data elements to generate 32-bit data elements, alternative embodiments could use different sized inputs to generate different sized outputs. In addition, while in the described embodiment Source1 and Source 2 each contain 4 data elements and the multiply-add instruction performs two multiply-add operations, alternative embodiment could operate on packed data having more or less data elements. For example, one alternative embodiment operates on packed data having 8 data elements using 4 multiply-adds generating a resulting packed data having 4 data elements. While in the described embodiment each multiply-add operation operates on 4 data elements by performing 2 multiplies and 1 addition, alternative embodiments could be implemented to operate on more or less data elements using more or less multiplies and additions. As an example, one alternative embodiment operates on 8 data elements using 4 multiplies (one for each pair of data elements) and 3 additions (2 additions to add the results of the 4 multiplies and 1 addition to add the results of the 2 previous additions).
While the invention has been described in terms of several embodiments, those skilled in the art will recognize that the invention is not limited to the embodiments described. The method and apparatus of the invention can be practiced with modification and alteration within the spirit and scope of the appended claims. The description is thus to be regarded as illustrative instead of limiting on the invention.
This presemt patent application is a Continuation of application Ser. No. 08/522,067, filed Aug. 31, 1995, now U.S. Pat. No. 6,385,634. Application Ser. No. 08/960,413, titled “Apparatus for Performing Multiply-Add Operations on Packed Data,” filed Oct. 29, 1997, now U.S. Pat. No. 5,983,256, which is a Continuation of Ser. No. 08/551,196, filed Oct. 31, 1995, abandoned, which is a Continuation of Ser. No. 08/522,067, filed Aug. 31, 1995, pending. application Ser. No. 08/606,212, titled “Apparatus for Performing Multiply-Add Operations on Packed Data,” filed Feb. 23, 1996, now U.S. Pat. No. 6,035,316, which is a CIP of Ser. No. 08/522,067, filed Aug. 31, 1995, pending. application Ser. No. 08/554,625, titled “An Apparatus for Performing Multiply-Subtract Operations on Packed Data,” filed Nov. 6, 1995, now U.S. Pat. No. 5,721,892, which is a Continuation of Ser. No. 08/521,803, filed Aug. 31, 1995, abandoned.
| Number | Name | Date | Kind |
|---|---|---|---|
| 3641736 | Dwyer | Feb 1972 | A |
| 3675001 | Singh | Jul 1972 | A |
| 3711692 | Batcher | Jan 1973 | A |
| 3723715 | Chen et al. | Mar 1973 | A |
| 3900724 | McIver | Aug 1975 | A |
| 3941990 | Rabasse et al. | Mar 1976 | A |
| 4161784 | Cushing et al. | Jul 1979 | A |
| 4344151 | White | Aug 1982 | A |
| 4393468 | New | Jul 1983 | A |
| 4418383 | Doyle et al. | Nov 1983 | A |
| 4498177 | Larson | Feb 1985 | A |
| 4542476 | Nagafuji et al. | Sep 1985 | A |
| 4677582 | Nagafuji | Jun 1987 | A |
| 4707800 | Montrone et al. | Nov 1987 | A |
| 4760525 | Webb | Jul 1988 | A |
| 4760545 | Inagami et al. | Jul 1988 | A |
| 4761753 | Izumisawa | Aug 1988 | A |
| 4771379 | Ando et al. | Sep 1988 | A |
| 4811269 | Hirose et al. | Mar 1989 | A |
| 4876660 | Owen et al. | Oct 1989 | A |
| 4890218 | Bram | Dec 1989 | A |
| 4945479 | Rusterholz et al. | Jul 1990 | A |
| 4956801 | Priem et al. | Sep 1990 | A |
| 4972362 | Elkind et al. | Nov 1990 | A |
| 4985848 | Pfeiffer et al. | Jan 1991 | A |
| 4989168 | Kuroda et al. | Jan 1991 | A |
| 5001662 | Baum | Mar 1991 | A |
| 5032865 | Schlunt | Jul 1991 | A |
| 5047975 | Patti et al. | Sep 1991 | A |
| 5081698 | Kohn | Jan 1992 | A |
| 5095457 | Jeong | Mar 1992 | A |
| 5111422 | Ullrich | May 1992 | A |
| 5126964 | Zurawski | Jun 1992 | A |
| 5161247 | Murakami et al. | Nov 1992 | A |
| 5187679 | Vassiliadis et al. | Feb 1993 | A |
| 5189636 | Patti et al. | Feb 1993 | A |
| 5193167 | Sites et al. | Mar 1993 | A |
| 5210711 | Rossmere et al. | May 1993 | A |
| 5222037 | Taniguchi | Jun 1993 | A |
| 5227994 | Ohki Mitsuharu | Jul 1993 | A |
| 5241492 | Girardeau, Jr. | Aug 1993 | A |
| 5241493 | Chu et al. | Aug 1993 | A |
| 5262976 | Young et al. | Nov 1993 | A |
| 5268995 | Diefendorff et al. | Dec 1993 | A |
| 5293558 | Narita et al. | Mar 1994 | A |
| 5303355 | Gergen et al. | Apr 1994 | A |
| 5311508 | Buda et al. | May 1994 | A |
| 5321644 | Schibinger | Jun 1994 | A |
| 5325320 | Chiu | Jun 1994 | A |
| 5327369 | Ashkenazi | Jul 1994 | A |
| 5327543 | Miura et al. | Jul 1994 | A |
| 5390135 | Lee et al. | Feb 1995 | A |
| 5410670 | Hansen et al. | Apr 1995 | A |
| 5420815 | Nix et al. | May 1995 | A |
| 5426598 | Hagihara | Jun 1995 | A |
| 5426783 | Norrie et al. | Jun 1995 | A |
| 5442799 | Murakami et al. | Aug 1995 | A |
| 5457805 | Nakamura | Oct 1995 | A |
| 5487022 | Simpson et al. | Jan 1996 | A |
| 5487159 | Byers et al. | Jan 1996 | A |
| 5500811 | Corry | Mar 1996 | A |
| 5506865 | Weaver, Jr. | Apr 1996 | A |
| 5509129 | Guttag et al. | Apr 1996 | A |
| 5515520 | Hatta et al. | May 1996 | A |
| 5517438 | Dao-Trong et al. | May 1996 | A |
| 5528529 | Seal | Jun 1996 | A |
| 5541865 | Ashkenazi | Jul 1996 | A |
| 5576983 | Shiokawa | Nov 1996 | A |
| 5579253 | Lee et al. | Nov 1996 | A |
| 5590365 | Ide et al. | Dec 1996 | A |
| 5594679 | Iwata | Jan 1997 | A |
| 5606677 | Balmer et al. | Feb 1997 | A |
| 5636351 | Lee | Jun 1997 | A |
| 5642306 | Mennemeier | Jun 1997 | A |
| 5651121 | Davies | Jul 1997 | A |
| 5673427 | Brown et al. | Sep 1997 | A |
| 5680161 | Lehman et al. | Oct 1997 | A |
| 5701508 | Glew et al. | Dec 1997 | A |
| 5721697 | Lee | Feb 1998 | A |
| 5721892 | Peleg et al. | Feb 1998 | A |
| 5734874 | Van Hook et al. | Mar 1998 | A |
| 5742840 | Hansen et al. | Apr 1998 | A |
| 5765037 | Morrison et al. | Jun 1998 | A |
| 5778419 | Hansen et al. | Jul 1998 | A |
| 5793661 | Dulong et al. | Aug 1998 | A |
| 5793879 | Benn et al. | Aug 1998 | A |
| 5794060 | Hansen et al. | Aug 1998 | A |
| 5794061 | Hansen et al. | Aug 1998 | A |
| 5802336 | Peleg et al. | Sep 1998 | A |
| 5805913 | Guttag et al. | Sep 1998 | A |
| 5806049 | Petruzzi | Sep 1998 | A |
| 5809321 | Hansen et al. | Sep 1998 | A |
| 5819101 | Peleg et al. | Oct 1998 | A |
| 5822603 | Hansen et al. | Oct 1998 | A |
| 5835392 | Dulong et al. | Nov 1998 | A |
| 5852726 | Lin et al. | Dec 1998 | A |
| 5859997 | Peleg et al. | Jan 1999 | A |
| 5870619 | Wilkinson et al. | Feb 1999 | A |
| 5880983 | Elliott et al. | Mar 1999 | A |
| 5883824 | Lee et al. | Mar 1999 | A |
| 5884071 | Kosaraju | Mar 1999 | A |
| 5909552 | Jensen et al. | Jun 1999 | A |
| 5936872 | Fischer et al. | Aug 1999 | A |
| 5938756 | Van Hook et al. | Aug 1999 | A |
| 5953241 | Hansen et al. | Sep 1999 | A |
| 5983256 | Peleg et al. | Nov 1999 | A |
| 5983257 | Dulong et al. | Nov 1999 | A |
| 6006318 | Hansen et al. | Dec 1999 | A |
| 6018351 | Mennemeier et al. | Jan 2000 | A |
| 6035316 | Peleg et al. | Mar 2000 | A |
| 6035387 | Hsu et al. | Mar 2000 | A |
| 6041403 | Parker et al. | Mar 2000 | A |
| 6122725 | Roussel et al. | Sep 2000 | A |
| 6154831 | Thayer et al. | Nov 2000 | A |
| 6178482 | Sollars | Jan 2001 | B1 |
| 6192467 | Abdallah et al. | Feb 2001 | B1 |
| 6230253 | Roussel et al. | May 2001 | B1 |
| 6230257 | Roussel et al. | May 2001 | B1 |
| 6233671 | Abdallah et al. | May 2001 | B1 |
| 6260137 | Fleck et al. | Jul 2001 | B1 |
| 6295599 | Hansen et al. | Sep 2001 | B1 |
| 6370558 | Guttag et al. | Apr 2002 | B1 |
| 6378060 | Hansen et al. | Apr 2002 | B1 |
| 6385634 | Peleg et al. | May 2002 | B1 |
| 6470370 | Fischer et al. | Oct 2002 | B2 |
| 6502117 | Golliver et al. | Dec 2002 | B2 |
| 6584482 | Hansen et al. | Jun 2003 | B1 |
| 6643765 | Hansen et al. | Nov 2003 | B1 |
| 6725356 | Hansen et al. | Apr 2004 | B2 |
| 7085795 | Debes et al. | Aug 2006 | B2 |
| 20020059355 | Peleg et al. | May 2002 | A1 |
| Number | Date | Country |
|---|---|---|
| 3636106 | Apr 1997 | DE |
| 19914617 | Oct 1999 | DE |
| 0 308 124 | Mar 1989 | EP |
| 2563349 | Oct 1985 | FR |
| 2339040 | Dec 2000 | GB |
| 3-268024 | Mar 1990 | JP |
| 6-266554 | Sep 1994 | JP |
| WO9708608 | Mar 1997 | WO |
| WO9722921 | Jun 1997 | WO |
| WO9722923 | Jun 1997 | WO |
| WO9722924 | Jun 1997 | WO |
| WO9723821 | Jul 1997 | WO |
| WO99507403 | Oct 1999 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 20020059355 A1 | May 2002 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 08960413 | Oct 1997 | US |
| Child | 09989736 | US | |
| Parent | 08606212 | Feb 1996 | US |
| Child | 09989736 | US | |
| Parent | 08554625 | Nov 1995 | US |
| Child | 09989736 | US | |
| Parent | 08551196 | Oct 1995 | US |
| Child | 09989736 | US | |
| Parent | 08522067 | Aug 1995 | US |
| Child | 08551196 | US | |
| Parent | 08521803 | Aug 1995 | US |
| Child | 08522067 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 08522067 | US | |
| Child | 08522067 | US |
