US20140074901A1 - Bandwidth efficient instruction-driven multiplication engine - Google Patents
Bandwidth efficient instruction-driven multiplication engine Download PDFInfo
- Publication number
- US20140074901A1 US20140074901A1 US14/055,177 US201314055177A US2014074901A1 US 20140074901 A1 US20140074901 A1 US 20140074901A1 US 201314055177 A US201314055177 A US 201314055177A US 2014074901 A1 US2014074901 A1 US 2014074901A1
- Authority
- US
- United States
- Prior art keywords
- operands
- multipliers
- register
- local
- operand
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
- 238000000034 method Methods 0.000 claims abstract description 16
- 230000004044 response Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 description 11
- 230000035508 accumulation Effects 0.000 description 10
- 238000009825 accumulation Methods 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 3
- 230000004075 alteration Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- TVEXGJYMHHTVKP-UHFFFAOYSA-N 6-oxabicyclo[3.2.1]oct-3-en-7-one Chemical compound C1C2C(=O)OC1C=CC2 TVEXGJYMHHTVKP-UHFFFAOYSA-N 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000010348 incorporation Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/30003—Arrangements for executing specific machine instructions
- G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
- G06F9/3001—Arithmetic instructions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/30003—Arrangements for executing specific machine instructions
- G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
- G06F9/30036—Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline, look ahead
- G06F9/3824—Operand accessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline, look ahead
- G06F9/3824—Operand accessing
- G06F9/3826—Bypassing or forwarding of data results, e.g. locally between pipeline stages or within a pipeline stage
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline, look ahead
- G06F9/3824—Operand accessing
- G06F9/383—Operand prefetching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline, look ahead
- G06F9/3885—Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
- G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
- G06F7/52—Multiplying; Dividing
- G06F7/523—Multiplying only
- G06F7/53—Multiplying only in parallel-parallel fashion, i.e. both operands being entered in parallel
- G06F7/5324—Multiplying only in parallel-parallel fashion, i.e. both operands being entered in parallel partitioned, i.e. using repetitively a smaller parallel parallel multiplier or using an array of such smaller multipliers
Definitions
- This invention relates to digital signal processors and, more particularly, to a software programmable complex multiplication engine.
- CDMA systems require significant hardware acceleration in order to perform functions such as beamforming and path searching.
- CDMA systems often implement these algorithms directly with a dedicated ASIC or an on-chip coprocessor unit. Although this approach offers the highest potential performance, it carries significant design risks and is very inflexible to changes in standards and algorithms.
- a multiplication engine for a digital processor.
- the multiplication engine comprises a plurality of multipliers, each receiving a first operand and a second operand; a local operand register having a plurality of locations to hold the first operands for respective ones of the multipliers; a first operand bus coupled to the local operand register to supply the first operands from a compute register file to the local operand register; a second operand bus coupled to the plurality of multipliers to supply one or more of the second operands from the compute register file to respective ones of the multipliers; and a control unit responsive to a digital processor instruction to supply the first operands from the local operand register to respective ones of the multipliers, to supply the second operands from the compute register file to respective ones of the multipliers on the second operand bus and to multiply the first operands by the respective second operands in the respective multipliers, wherein one or more of the first operands in the local operand register are reused by the plurality of multipliers
- a method for performing multiplication in a digital processor.
- the method comprises providing a plurality of multipliers; providing a local operand register having a plurality of locations to hold first operands for respective ones of the multipliers; supplying the first operands from a compute register file to the local operand register on a first operand bus; supplying second operands from the compute register file to respective ones of the multipliers on a second operand bus; controlling operation of the multipliers and the local operand register in response to a digital processor instruction by supplying the first operands from the local operand register to respective ones of the multipliers, supplying the second operands from the compute register file to respective ones of the multipliers on the second operand bus and multiplying the first operands by the respective second operands in respective ones of the multipliers; and reusing one or more of the first operands in the local operand register in two or more multiplication operations by the plurality of multipliers.
- a multiplication engine for a digital processor.
- the multiplication engine comprises a multiplier circuit, a register file to supply data to the multiplier circuit, the register file including a plurality of register locations in a shift configuration, and a control circuit to load data from a data source into the is register file and to shift data in the register file after a multiply instruction is executed, wherein data in the register file is used by two or more multiply instructions.
- a method for digital processing in a multiplication engine including a multiplier circuit and a register file to supply data to the multiplier circuit.
- the method comprises loading data from a data source into the register file, executing a first multiply instruction using data in the register file, shifting data in the register file from current register locations to next register locations after execution of the first multiply instruction, and executing a second multiply instruction using the shifted data in the register file.
- FIG. 1 is a block diagram showing an example of a digital signal processor architecture
- FIG. 2 is a block diagram showing an example of a compute block in the digital signal processor
- FIG. 3 is a block diagram that illustrates a multiplication engine in accordance with an embodiment of the invention
- FIG. 4 is a schematic block diagram that illustrates the multiplication engine of FIG. 3 in greater detail
- FIG. 5 is a schematic block diagram that illustrates the registers involved in the operation of the multiplication engine of FIG. 3 ;
- FIG. 6 is a schematic diagram that illustrates an FIR filter calculation in accordance with an embodiment of the invention.
- FIG. 7 is a schematic diagram that illustrates a matrix multiplication calculation in accordance with an embodiment of the invention.
- FIG. 1 A block diagram of an example of a digital signal processor (DSP) 10 suitable for incorporation of the present invention is shown in FIG. 1 .
- the digital signal processor may be the TigerSharc digital signal processor manufactured and sold by Analog Devices, Inc., Norwood, Mass., for example.
- the digital signal processor 10 may include a compute X block 12 , a compute Y block 14 , an instruction sequencer 16 , memory blocks 20 , 22 , 24 , an integer ALU 30 and an I/O processor or DMA controller 32 .
- the elements of DSP 10 are interconnected by data and address buses 40 a , 40 b , 40 c and 40 d.
- the compute block includes a compute register file 50 and several computation units.
- the computation units include an ALU 52 , a multiplier 54 , a shifter 56 and an accelerator 58 .
- Compute register file 50 receives data on buses 40 a and 40 b and supplies operands to the computation units on operand buses 64 and 66 .
- the results of the computations are supplied on result buses 70 , 72 , 74 , 76 and 78 to compute register file 50 .
- the results may be written to memory from compute register file 50 or supplied to the computation units for subsequent computations.
- a multiplication engine 100 in accordance with an embodiment of the invention is shown in FIGS. 3-5 .
- Multiplication engine 100 may correspond to multiplier 54 shown in FIG. 2 , may be used in accelerator 58 , or both.
- the multiplication engine 100 includes multiplier units 110 , 112 , . . . 118 .
- multiplication engine 100 includes 16 multiplier units.
- Multiplication engine 100 further includes a local operand register 130 coupled to each of multiplier units 110 , 112 , . . . 118 .
- the outputs of multiplier units 110 , 112 , . . . 118 are supplied to accumulators 150 , 152 , . . . 158 , respectively.
- Each accumulator may include a summing unit and an accumulation register.
- each of accumulators 150 , 152 , . . . 158 includes two accumulation registers for enhanced performance.
- accumulator 150 includes a summing unit 160 and an accumulation register 170 , and may include a second accumulation register 171 ( FIG. 4 ).
- Multiplication engine 100 further includes a control unit 180 that controls the components of multiplication engine 100 in response to instructions being executed.
- Multiplication engine 100 receives operands from compute register file 50 ( FIG. 2 ) on a first operand bus 190 and a second operand bus 192 . Results are returned to compute register file 50 on a result bus 196 .
- First operand bus 190 is coupled to local operand register 130 and to each of multiplier units 110 , 112 , . . . 118 .
- Second operand bus 192 is coupled to each of multiplier units 110 , 112 , . . . 118 .
- Each of the multiplier units 110 , 112 , . . . 118 can be configured in response to an instruction being executed.
- each of the multiplier units is configured as eight multipliers of 16 bits by 2 bits.
- each of the multiplier units is configured as a single multiplier of 16 bits by 16 bits.
- Local operand register 130 provides local storage of operands used by multiplier units 110 , 112 , . . . 118 . Local operand register 130 is useful where operands are reused by the multiplier units for two or more calculations. In some configurations, the same operands are used for two or more consecutive computations by the same multiplier units. In other configurations, operands are reused by different multiplier units for consecutive computations and the operands in local operand register 130 are shifted after completion of a computation. By holding operands that are reused in local operand register 130 , transfer of data on operand buses 190 and 192 is reduced and operating efficiency is increased.
- the multiplier units of multiplication engine 100 may be configured as complex multipliers 510 , each of which receives a first operand from local operand register 130 (THR register) and real and imaginary parts of a second operand from an Rms register 512 in compute register file 50 .
- the second operand may be an input data value and the first operand may be a coefficient.
- the multiplication engine 100 further includes complex summing units 520 , each of which performs complex addition of a value output by complex multiplier 510 and a previous value.
- the output of each multiplier 510 is summed with a previous value in an accumulation register 522 to provide a current value that is placed in the accumulation register.
- multiplication engine 100 includes eight complex multipliers 510 and eight complex summing units 520 .
- the complex multipliers correspond to the multiplier units 110 , 112 , . . . 118 shown in FIG. 3
- the complex summing units correspond to summing units 160 , 162 , . . . 168 shown in FIG. 3
- the accumulation registers 522 correspond to accumulation registers 170 , 172 , . . . 178 shown in FIG. 3 .
- Rms registers 512 may be located in compute register file 50 ( FIG. 2 ).
- Local operand register 130 and accumulation registers 522 may be located in close proximity to complex multipliers 510 .
- Local operand register 130 is configured to perform shift operations as described below.
- local operand register 130 includes eight operands, each having 16 bits, including 8 bits real and 8 bits imaginary.
- Control circuit 180 ( FIG. 3 ) causes operands to be loaded into local operand register 130 from compute register file 50 when operands are needed for multiply instructions. The operands in local operand register 130 are used to execute a multiply instruction as shown in FIG. 5 . After the multiply instruction has been executed, control circuit 180 causes the operands in local operand register 130 to be shifted to the right. In the case of 16-bit operands, the operands are shifted 16 bits to the right by control circuit 180 . In addition, a new operand value is loaded from compute register file 50 to local operand register 130 .
- a second multiply instruction is executed with the shifted operands in local operand register 130 . This process can be repeated until all computations have been completed. In each calculation, the operands contained in local operand register 130 are multiplied by operands supplied from the Rms registers 512 in compute register file 50 . If necessary, a new set of operands can be loaded into local operand register 130 by control circuit 180 . Thus, two or more multiply instructions can be executed without reloading local operand register 130 with a complete set of new operands after each multiply instruction.
- the multiplication engine can be used to execute an FIR filter instruction as follows.
- TRsh+ Rms**THRo
- THR7h Rss (j)(clr)(sho)(mhl))(shl)
- the instruction makes 16 complex multiplications and afterwards shifts the contents of local operand register 130 (THR 7:0) by 16 bits to the right and updates the THR7h location by the new short word data from the compute register file.
- the data d15:d0 is stored in the THR7:0 (each data number is 16b only).
- the coefficient c0 is being loaded by Rms and multiplied by all the data numbers.
- the data in the THR is shifted to the right and d16 is loaded to the THR7h.
- the data in the THR7:0 inhabits data numbers d16:d1 and is ready for step 2.
- the multiplication engine can be used to execute a matrix instruction (broadcast) as follows.
Abstract
Description
- This application claims priority based on Provisional Application Ser. No. 60/879,760, filed Jan. 10, 2007, which is hereby incorporated by reference in its entirety.
- This invention relates to digital signal processors and, more particularly, to a software programmable complex multiplication engine.
- Advanced wireless networks require significant hardware acceleration in order to perform functions such as beamforming and path searching. To address these data processing requirements, CDMA systems often implement these algorithms directly with a dedicated ASIC or an on-chip coprocessor unit. Although this approach offers the highest potential performance, it carries significant design risks and is very inflexible to changes in standards and algorithms.
- These and other algorithms usually involve multiplication operations. One of the limiting factors in a high performance multiplication engine is the rate at which data can be supplied to the engine from a register file or a memory. The speed and width of data buses can be increased, but at the expense of chip area and power dissipation. A further factor is that data is often reused in digital signal processing algorithms, such as FIR digital filters.
- Accordingly, there is a need for improved multiplication engines and multiplication methods.
- According to a first aspect of the invention, a multiplication engine is provided for a digital processor. The multiplication engine comprises a plurality of multipliers, each receiving a first operand and a second operand; a local operand register having a plurality of locations to hold the first operands for respective ones of the multipliers; a first operand bus coupled to the local operand register to supply the first operands from a compute register file to the local operand register; a second operand bus coupled to the plurality of multipliers to supply one or more of the second operands from the compute register file to respective ones of the multipliers; and a control unit responsive to a digital processor instruction to supply the first operands from the local operand register to respective ones of the multipliers, to supply the second operands from the compute register file to respective ones of the multipliers on the second operand bus and to multiply the first operands by the respective second operands in the respective multipliers, wherein one or more of the first operands in the local operand register are reused by the plurality of multipliers in two or more multiplication operations.
- According to a second aspect of the invention, a method is provided for performing multiplication in a digital processor. The method comprises providing a plurality of multipliers; providing a local operand register having a plurality of locations to hold first operands for respective ones of the multipliers; supplying the first operands from a compute register file to the local operand register on a first operand bus; supplying second operands from the compute register file to respective ones of the multipliers on a second operand bus; controlling operation of the multipliers and the local operand register in response to a digital processor instruction by supplying the first operands from the local operand register to respective ones of the multipliers, supplying the second operands from the compute register file to respective ones of the multipliers on the second operand bus and multiplying the first operands by the respective second operands in respective ones of the multipliers; and reusing one or more of the first operands in the local operand register in two or more multiplication operations by the plurality of multipliers.
- According to a third aspect of the invention, a multiplication engine is provided for a digital processor. The multiplication engine comprises a multiplier circuit, a register file to supply data to the multiplier circuit, the register file including a plurality of register locations in a shift configuration, and a control circuit to load data from a data source into the is register file and to shift data in the register file after a multiply instruction is executed, wherein data in the register file is used by two or more multiply instructions.
- According to a fourth aspect of the invention, a method is provided for digital processing in a multiplication engine including a multiplier circuit and a register file to supply data to the multiplier circuit. The method comprises loading data from a data source into the register file, executing a first multiply instruction using data in the register file, shifting data in the register file from current register locations to next register locations after execution of the first multiply instruction, and executing a second multiply instruction using the shifted data in the register file.
- For a better understanding of the present invention, reference is made to the accompanying drawings, which are incorporated herein by reference and in which:
-
FIG. 1 is a block diagram showing an example of a digital signal processor architecture; -
FIG. 2 is a block diagram showing an example of a compute block in the digital signal processor; -
FIG. 3 is a block diagram that illustrates a multiplication engine in accordance with an embodiment of the invention; -
FIG. 4 is a schematic block diagram that illustrates the multiplication engine ofFIG. 3 in greater detail; -
FIG. 5 is a schematic block diagram that illustrates the registers involved in the operation of the multiplication engine ofFIG. 3 ; -
FIG. 6 is a schematic diagram that illustrates an FIR filter calculation in accordance with an embodiment of the invention; and -
FIG. 7 is a schematic diagram that illustrates a matrix multiplication calculation in accordance with an embodiment of the invention. - A block diagram of an example of a digital signal processor (DSP) 10 suitable for incorporation of the present invention is shown in
FIG. 1 . The digital signal processor may be the TigerSharc digital signal processor manufactured and sold by Analog Devices, Inc., Norwood, Mass., for example. Thedigital signal processor 10 may include acompute X block 12, acompute Y block 14, aninstruction sequencer 16,memory blocks integer ALU 30 and an I/O processor orDMA controller 32. The elements of DSP 10 are interconnected by data andaddress buses - An example of
compute blocks FIG. 2 . The compute block includes acompute register file 50 and several computation units. The computation units include an ALU 52, amultiplier 54, ashifter 56 and anaccelerator 58.Compute register file 50 receives data onbuses operand buses result buses register file 50. The results may be written to memory fromcompute register file 50 or supplied to the computation units for subsequent computations. - A
multiplication engine 100 in accordance with an embodiment of the invention is shown inFIGS. 3-5 .Multiplication engine 100 may correspond tomultiplier 54 shown inFIG. 2 , may be used inaccelerator 58, or both. Themultiplication engine 100 includesmultiplier units FIGS. 3-5 ,multiplication engine 100 includes 16 multiplier units.Multiplication engine 100 further includes alocal operand register 130 coupled to each ofmultiplier units multiplier units accumulators accumulators accumulator 150 includes asumming unit 160 and anaccumulation register 170, and may include a second accumulation register 171 (FIG. 4 ).Multiplication engine 100 further includes acontrol unit 180 that controls the components ofmultiplication engine 100 in response to instructions being executed. -
Multiplication engine 100 receives operands from compute register file 50 (FIG. 2 ) on afirst operand bus 190 and asecond operand bus 192. Results are returned to computeregister file 50 on aresult bus 196.First operand bus 190 is coupled tolocal operand register 130 and to each ofmultiplier units Second operand bus 192 is coupled to each ofmultiplier units - Each of the
multiplier units -
Local operand register 130 provides local storage of operands used bymultiplier units Local operand register 130 is useful where operands are reused by the multiplier units for two or more calculations. In some configurations, the same operands are used for two or more consecutive computations by the same multiplier units. In other configurations, operands are reused by different multiplier units for consecutive computations and the operands inlocal operand register 130 are shifted after completion of a computation. By holding operands that are reused inlocal operand register 130, transfer of data onoperand buses - As shown in
FIG. 5 , the multiplier units ofmultiplication engine 100 may be configured ascomplex multipliers 510, each of which receives a first operand from local operand register 130 (THR register) and real and imaginary parts of a second operand from anRms register 512 incompute register file 50. For example, the second operand may be an input data value and the first operand may be a coefficient. Themultiplication engine 100 further includes complex summingunits 520, each of which performs complex addition of a value output bycomplex multiplier 510 and a previous value. In particular, the output of eachmultiplier 510 is summed with a previous value in anaccumulation register 522 to provide a current value that is placed in the accumulation register. In the embodiment ofFIG. 5 ,multiplication engine 100 includes eightcomplex multipliers 510 and eight complex summingunits 520. The complex multipliers correspond to themultiplier units FIG. 3 , the complex summing units correspond to summingunits FIG. 3 , and the accumulation registers 522 correspond to accumulation registers 170, 172, . . . 178 shown inFIG. 3 . - In the embodiment of
FIG. 5 , Rms registers 512 may be located in compute register file 50 (FIG. 2 ).Local operand register 130 and accumulation registers 522 may be located in close proximity tocomplex multipliers 510.Local operand register 130 is configured to perform shift operations as described below. - In the embodiment of
FIG. 5 ,local operand register 130 includes eight operands, each having 16 bits, including 8 bits real and 8 bits imaginary. Control circuit 180 (FIG. 3 ) causes operands to be loaded into local operand register 130 fromcompute register file 50 when operands are needed for multiply instructions. The operands inlocal operand register 130 are used to execute a multiply instruction as shown inFIG. 5 . After the multiply instruction has been executed,control circuit 180 causes the operands inlocal operand register 130 to be shifted to the right. In the case of 16-bit operands, the operands are shifted 16 bits to the right bycontrol circuit 180. In addition, a new operand value is loaded fromcompute register file 50 tolocal operand register 130. Then, a second multiply instruction is executed with the shifted operands inlocal operand register 130. This process can be repeated until all computations have been completed. In each calculation, the operands contained inlocal operand register 130 are multiplied by operands supplied from the Rms registers 512 incompute register file 50. If necessary, a new set of operands can be loaded intolocal operand register 130 bycontrol circuit 180. Thus, two or more multiply instructions can be executed without reloadinglocal operand register 130 with a complete set of new operands after each multiply instruction. - The multiplication engine can be used to execute an FIR filter instruction as follows.
- TRsh+=Rms**THRo, THR7h=Rss (j)(clr)(sho)(mhl))(shl)
-
-
- Rms—is an input single short coefficient.
- THRo—is an octal register data—inhabits 16 data numbers.
- THR7h—is the msb short in the THR7:0 registers.
- Rss—is a short operand which is loaded into the msb THR.
- j—for conjugate multiplication option.
- clr—clears the TR accumulators.
- sho—for real 8 bit multiplications.
- mhl—high/low Rms
- shl—high/low Rss
- The instruction makes 16 complex multiplications and afterwards shifts the contents of local operand register 130 (THR 7:0) by 16 bits to the right and updates the THR7h location by the new short word data from the compute register file.
- Example (from
FIG. 6 ): -
- Step1: TR15:0=c0**[d15 . . . d0]=c0**d15, c0**d14, c0**d13, . . . , c0**d3, c0**d2, c0**d1, c0**d0
- The data d15:d0 is stored in the THR7:0 (each data number is 16b only). The coefficient c0 is being loaded by Rms and multiplied by all the data numbers. Then the data in the THR is shifted to the right and d16 is loaded to the THR7h. Thus, the data in the THR7:0 inhabits data numbers d16:d1 and is ready for
step 2. -
-
R31:28=q[k2+=0x4]; R27:24=-q[j2+=0x4]; // 1,2.load coefficients to R31:24 xR3:0=[k0+=0x4]; yR3:0 = [j0+=0x4];; // 1.load d7:d0 to xR3:0 // 2.load d39:d32 to yR3:0 xR7:4=[k0+=0x4]; yR7:4 = [j0+=0x4];; if 1.load d15:d8 to xR7:4 // 2.load d47:d40 to yR7:4 xR3:0=[k0+=0x4]; yR3:0 = [j0+=0x4]; THR3:0 = R3:0;; // 1.load d23:d16 to xR7:4 // 2.load d55:d48 to yR7:4 // 3.load data to THR3:0 THR7:4 = R7:4;; // 3.load data to THR3:0 xR7:4=[k0+=0x4]; yR7:4 = [j0+=0x4]; TR15:0+= THR7:0**R24l, THR7h = R0l (bc)(clr);; // Octal_data**single_tap (broadcast) xR11:8=[k0+=0x4]; yR11:8 = [j0+=0x4]; TR15:0+= THR7:0**R24h, THR7h = R0h (bc);; xR15:12=[k0+=0x4];yR15:12 = [j0+=0x4]; TR15:0+= THR7:0**R25l, THR7h = R1l (bc);; TR15:0+= THR7:0**R25h, THR7h = R1h (bc);; TR15:0+= THR7:0**R26l, THR7h = R2l (bc);; TR15:0+= THR7:0**R26h, THR7h = R2h (bc);; TR15:0+= THR7:0**R27l, THR7h = R3l (bc);; TR15:0+= THR7:0**R27h, THR7h = R3h (bc);; TR15:0+= THR7:0**R28l, THR7h = R4l (bc);; TR15:0+= THR7:0**R28h, THR7h = R4h (bc);; TR15:0+= THR7:0**R29l, THR7h = R5l (bc);; TR15:0+= THR7:0**R29h, THR7h = R5h (bc);; TR15:0+= THR7:0**R30l, THR7h = R6l (bc);; TR15:0+= THR7:0**R30h, THR7h = R6h (bc);; TR15:0+= THR7:0**R31l, THR7h = R7l (bc);; TR15:0+= THR7:0**R31h, THR7h = R7h (bc);; R23:20 = TR3:0;; q[k1+=0x4] = xR23:20; q[j1+=0x4] = yR23:20; R23:20 = TR7:4;; q[k1+=0x4] = xR23:20; q[j1+=0x4] = yR23:20; R23:20 = TR11:8;; q[k1+=0x4] = xR23:20; q[j1+=0x4] = yR23:20; R23:20 =TR15:12;; q[k1+=0x4] = xR23:20; q[j1+=0x4] = yR23:20;; - The multiplication engine can be used to execute a matrix instruction (broadcast) as follows.
-
- 1) TRsh+=Rmd**Rnd (j)(clr)(sho)(f)
- 2) TRsh+=Rmd**THRd (j)(clr)(sho)(f)
- 3) TRsh+=Rmq**Rn (j)(clr)(sho)(f)
- 4) TRsh+=Rmq**THR (j)(clr)(sho)(f)
- 5) TRsh+=(Rmq,Rnq)**THRs (j)(clr)(sho)(f)(ns)
-
-
- Rmd—holds 4 data numbers, in total 64 bits operand.
- Rnd,THRd—holds 4 coefficients numbers, in total 64 bits operand.
- From
FIG. 7 one may observe that in order to calculate the 4×4 matrix multiplication we need to make a 4 matrix accumulations. While each one of those matrix is the dot products of: - That means that in order to calculate matrix A we don't need to load 16 data numbers and 16 coefficients and then to multiply them, but we can bring only 4 data numbers and 4 coefficients and multiply each of the data numbers by each of the coefficients to accept the dot product. That method utilizes all the 16 multipliers and saves bus bandwidth.
-
-
xyR9:8 = q[k0+=0x4]; xyR11:10 = q[j0+=0x4];; // 1,2. xR11:8 = S(3,A1),S(2,A1),S(1,A1),S(0,A1),S(3,A0),S(2,A0),S(1,A0),S(0,A0) // yR11:8 = S(7,A1),S(6,A1),S(5,A1),S(4,A1),S(7,A0),S(6,A0),S(5,A0),S(4,A0) xyR1:0 = q[k1+=0x4]; yR3:2 = q[j1+=0x4];; // 1,2. xr3:0 = h(1,3),h(1,2),h(1,1),h(1,0),h(0,3),h(0,2),h(0,1),h(0,0) // yr3:0 = h(5,3),h(5,2),h(5,1),h(5,0),h(4,3),h(4,2),h(4,1),h(4,0) xyR13:12 = q[k0+=0x4]; xyR15:14 = q[j0+=0x4];; // 1,2. xR15:12 = S(3,A3),S(2,A3),S(1,A3),S(0,A3),S(3,A2),S(2,A2),S(1,A2),S(0,A2) // yR15:12 = S(7,A3),S(6,A3),S(5,A3),S(4,A3),S(7,A2),S(6,A2),S(5,A2),S(4,A2) // Start 2 Matrix calculation xyR7:4 = q[k1+=0x4]; xyR7:4 = c[j1+=0x4]; TR15:0+= R17:16 ** R1:0 (clr)(bc);; xyR11:8 = q[k0+=0x4]; xyR11:8 = q[j0+=0x4]; TR15:0+= R19:18 ** R3:2 (bc);; xR3:0 = q[k1+=0x4]; yR3:0 = q[j1+=0x4]; TR15:0+= R21:20 ** R5:4 (bc);; xR15:12 = q[k0+=0x4]; yR15:12 = q[j0+=0x4]; TR15:0+= R23:22 ** R7:6 (bc);; // End of first 2 Start next 2 Matrix calculation xR7:4 = q[k1+=0x4]; yR7:4 = q[j1+=0x4]; TR31:16+= R17:16 ** R1:0 (clr)(bc); R27:24= TR3:0;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; TR31:16+= R19:18 ** R3:2 (bc); R27:24= TR7:4;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; TR31:16+= R21:20 ** R5:4 (bc); R27:24= TR3:0;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; TR31:16+= R23:22 ** R7:6 (bc); R27:24= TR7:4;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; R27:24= TR3:0;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; R27:24= TR7:4;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; R27:24= TR3:0;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24; R27:24= TR7:4;; [k2+=0x4] = xR27:24; [j2+=0x4] = yR27:24;; // End of 4 Matrix calculation - Having thus described several aspects of at least one embodiment of this invention, it is to be appreciated various alterations, modifications, and improvements will readily occur to those skilled in the art. Such alterations, modifications, and improvements are intended to be part of this disclosure, and are intended to be within the spirit and scope of the invention. Accordingly, the foregoing description and drawings are by way of example only.
Claims (21)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US14/055,177 US20140074901A1 (en) | 2007-01-10 | 2013-10-16 | Bandwidth efficient instruction-driven multiplication engine |
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US87976007P | 2007-01-10 | 2007-01-10 | |
US12/008,334 US8589469B2 (en) | 2007-01-10 | 2008-01-10 | Bandwidth efficient instruction-driven multiplication engine |
US14/055,177 US20140074901A1 (en) | 2007-01-10 | 2013-10-16 | Bandwidth efficient instruction-driven multiplication engine |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/008,334 Continuation US8589469B2 (en) | 2007-01-10 | 2008-01-10 | Bandwidth efficient instruction-driven multiplication engine |
Publications (1)
Publication Number | Publication Date |
---|---|
US20140074901A1 true US20140074901A1 (en) | 2014-03-13 |
Family
ID=39742723
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/008,334 Expired - Fee Related US8589469B2 (en) | 2007-01-10 | 2008-01-10 | Bandwidth efficient instruction-driven multiplication engine |
US14/055,177 Abandoned US20140074901A1 (en) | 2007-01-10 | 2013-10-16 | Bandwidth efficient instruction-driven multiplication engine |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/008,334 Expired - Fee Related US8589469B2 (en) | 2007-01-10 | 2008-01-10 | Bandwidth efficient instruction-driven multiplication engine |
Country Status (1)
Country | Link |
---|---|
US (2) | US8589469B2 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
IT202000006298A1 (en) * | 2020-03-25 | 2021-09-25 | St Microelectronics Srl | DATA PROCESSING PROCESS, MAC CIRCUIT, DSP SYSTEM AND CORRESPONDING IT PRODUCT |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2010096119A1 (en) * | 2008-10-29 | 2010-08-26 | Adapteva Incorporated | Variable instruction width digital signal processor |
US10824938B2 (en) * | 2017-04-24 | 2020-11-03 | Intel Corporation | Specialized fixed function hardware for efficient convolution |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5222035A (en) * | 1990-05-28 | 1993-06-22 | Hitachi, Ltd. | Digital filter circuit |
US5442580A (en) * | 1994-05-25 | 1995-08-15 | Tcsi Corporation | Parallel processing circuit and a digital signal processer including same |
US5594679A (en) * | 1993-03-31 | 1997-01-14 | Sony Corporation | Adaptive video signal processing apparatus |
US6524430B1 (en) * | 1999-09-21 | 2003-02-25 | Jusung Engineering Co., Ltd. | Apparatus for fabricating a semiconductor device |
US7013321B2 (en) * | 2001-11-21 | 2006-03-14 | Sun Microsystems, Inc. | Methods and apparatus for performing parallel integer multiply accumulate operations |
US7353244B2 (en) * | 2004-04-16 | 2008-04-01 | Marvell International Ltd. | Dual-multiply-accumulator operation optimized for even and odd multisample calculations |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB8401808D0 (en) * | 1984-01-24 | 1984-02-29 | Int Computers Ltd | Binary multiplication |
US5072418A (en) * | 1989-05-04 | 1991-12-10 | Texas Instruments Incorporated | Series maxium/minimum function computing devices, systems and methods |
US4939687A (en) * | 1988-11-01 | 1990-07-03 | General Electric Company | Serial-parallel multipliers using serial as well as parallel addition of partial products |
US5001661A (en) * | 1990-01-23 | 1991-03-19 | Motorola, Inc. | Data processor with combined adaptive LMS and general multiplication functions |
US5113364A (en) * | 1990-10-29 | 1992-05-12 | Motorola, Inc. | Concurrent sticky-bit detection and multiplication in a multiplier circuit |
US5880981A (en) * | 1996-08-12 | 1999-03-09 | Hitachi America, Ltd. | Method and apparatus for reducing the power consumption in a programmable digital signal processor |
GB2323187B (en) * | 1997-03-14 | 2001-09-19 | Nokia Mobile Phones Ltd | Data processing circuit using both edges of a clock signal |
US6289417B1 (en) * | 1998-05-18 | 2001-09-11 | Arm Limited | Operand supply to an execution unit |
-
2008
- 2008-01-10 US US12/008,334 patent/US8589469B2/en not_active Expired - Fee Related
-
2013
- 2013-10-16 US US14/055,177 patent/US20140074901A1/en not_active Abandoned
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5222035A (en) * | 1990-05-28 | 1993-06-22 | Hitachi, Ltd. | Digital filter circuit |
US5594679A (en) * | 1993-03-31 | 1997-01-14 | Sony Corporation | Adaptive video signal processing apparatus |
US5442580A (en) * | 1994-05-25 | 1995-08-15 | Tcsi Corporation | Parallel processing circuit and a digital signal processer including same |
US6524430B1 (en) * | 1999-09-21 | 2003-02-25 | Jusung Engineering Co., Ltd. | Apparatus for fabricating a semiconductor device |
US7013321B2 (en) * | 2001-11-21 | 2006-03-14 | Sun Microsystems, Inc. | Methods and apparatus for performing parallel integer multiply accumulate operations |
US7353244B2 (en) * | 2004-04-16 | 2008-04-01 | Marvell International Ltd. | Dual-multiply-accumulator operation optimized for even and odd multisample calculations |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
IT202000006298A1 (en) * | 2020-03-25 | 2021-09-25 | St Microelectronics Srl | DATA PROCESSING PROCESS, MAC CIRCUIT, DSP SYSTEM AND CORRESPONDING IT PRODUCT |
Also Published As
Publication number | Publication date |
---|---|
US20080222226A1 (en) | 2008-09-11 |
US8589469B2 (en) | 2013-11-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8271571B2 (en) | Microprocessor | |
US7797363B2 (en) | Processor having parallel vector multiply and reduce operations with sequential semantics | |
EP1623307B1 (en) | Processor reduction unit for accumulation of multiple operands with or without saturation | |
US8307196B2 (en) | Data processing system having bit exact instructions and methods therefor | |
JPH10187438A (en) | Method for reducing transition to input of multiplier | |
JP2001516916A (en) | Data processing device with digital signal processing capability | |
KR101395260B1 (en) | Microprocessor and method for enhanced precision sum-of-products calculation on a microprocessor | |
US20140074901A1 (en) | Bandwidth efficient instruction-driven multiplication engine | |
US6606700B1 (en) | DSP with dual-mac processor and dual-mac coprocessor | |
US11755320B2 (en) | Compute array of a processor with mixed-precision numerical linear algebra support | |
US8909687B2 (en) | Efficient FIR filters | |
JPS62118474A (en) | Vector division device | |
KR20080050226A (en) | Modular multiplication device and method for designing modular multiplication device | |
US20060224652A1 (en) | Instruction set processor enhancement for computing a fast fourier transform | |
US9582473B1 (en) | Instruction set to enable efficient implementation of fixed point fast fourier transform (FFT) algorithms | |
EP0992917B1 (en) | Linear vector computation | |
EP1936492A1 (en) | SIMD processor with reduction unit | |
CN113485751A (en) | Method for performing Galois field multiplication, arithmetic unit and electronic device | |
JP2012205298A (en) | Digital signal processor and program | |
US7590677B2 (en) | Processor with summation instruction using overflow counter | |
WO2008077803A1 (en) | Simd processor with reduction unit | |
WO2020084721A1 (en) | Computation processing device and computation processing device control method | |
EP1197874A1 (en) | Signal processor and product-sum operating device for use therein with rounding function | |
JP3723311B2 (en) | Parallel processor | |
JP2003330914A (en) | Vector calculation processing device, vector calculation method and vector calculation program |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: ANALOG DEVICES TECHNOLOGY, BERMUDA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:YANOVITCH, BARUCH;REEL/FRAME:032155/0406 Effective date: 20131107 |
|
AS | Assignment |
Owner name: ANALOG DEVICES GLOBAL, BERMUDA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ANALOG DEVICES TECHNOLOGY;REEL/FRAME:034757/0690 Effective date: 20141031 |
|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- FAILURE TO RESPOND TO AN OFFICE ACTION |
|
AS | Assignment |
Owner name: ANALOG DEVICES GLOBAL UNLIMITED COMPANY, BERMUDA Free format text: CHANGE OF NAME;ASSIGNOR:ANALOG DEVICES GLOBAL;REEL/FRAME:059090/0740 Effective date: 20161130 |
|
AS | Assignment |
Owner name: ANALOG DEVICES INTERNATIONAL UNLIMITED COMPANY, IRELAND Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ANALOG DEVICES GLOBAL UNLIMITED COMPANY;REEL/FRAME:059102/0062 Effective date: 20181105 |