Presentation is loading. Please wait.

Presentation is loading. Please wait.

CSE 661 - Parallel and Vector ArchitecturesVector Computers – slide 1 Vector Computers Muhamed Mudawar Computer Engineering Department King Fahd University.

Similar presentations


Presentation on theme: "CSE 661 - Parallel and Vector ArchitecturesVector Computers – slide 1 Vector Computers Muhamed Mudawar Computer Engineering Department King Fahd University."— Presentation transcript:

1 CSE Parallel and Vector ArchitecturesVector Computers – slide 1 Vector Computers Muhamed Mudawar Computer Engineering Department King Fahd University of Petroleum and Minerals

2 CSE Parallel and Vector ArchitecturesVector Computers – slide 2 + r1 r2 r3 add r3, r1, r2 SCALAR (1 operation) v1 v2 v3 + vector length addv v3, v1, v2 VECTOR (N operations)  Vector processors have high-level operations that work on linear arrays of numbers: "vectors" Vector Processing

3 CSE Parallel and Vector ArchitecturesVector Computers – slide 3 Vector Supercomputers Idealized by Cray-1, 1976: Scalar Unit + Vector Extensions  Load/Store Architecture  Vector Registers  Vector Instructions  Hardwired Control  Highly Pipelined Functional Units  Interleaved Memory System  No Data Caches  No Virtual Memory

4 CSE Parallel and Vector ArchitecturesVector Computers – slide 4 Cray-1 (1976) memory bank cycle 50 ns processor cycle 12.5 ns (80MHz)

5 CSE Parallel and Vector ArchitecturesVector Computers – slide 5 Single Port Memory 16 banks of 64-bit words + 8-bit SECDED 80MW/sec data load/store 320MW/sec instruction buffer refill 4 Instruction Buffers 64-bitx16 NIP LIP CIP (A 0 ) ( (A h ) + j k m ) 64 T Regs (A 0 ) ( (A h ) + j k m ) 64 B Regs S0 S1 S2 S3 S4 S5 S6 S7 A0 A1 A2 A3 A4 A5 A6 A7 SiSi T jk AiAi B jk FP Add FP Mul FP Recip Int Add Int Logic Int Shift Pop Cnt SjSj SiSi SkSk Addr Add Addr Mul AjAj AiAi AkAk V0 V1 V2 V3 V4 V5 V6 V7 VkVk VjVj ViVi V. Mask V. Length 64 Element Vector Registers Cray-1 (1976)

6 CSE Parallel and Vector ArchitecturesVector Computers – slide 6 Vector Instruction Set Architecture  ISA in general  Operations, Data types, Format, Accessible Storage, Addressing Modes, Exceptional Conditions  Vector ISA  Operations  Data types (Float, int, V op V, S op V)  Format  Source and Destination Operands  Length  Successor (consecutive, stride, indexed, gather/scatter)  Conditional operations  Exceptions

7 CSE Parallel and Vector ArchitecturesVector Computers – slide 7 Styles of Vector Architectures  Memory-memory vector processors: all vector operations are memory to memory  CDC Star100, Cyber203, Cyber205, 370 vector extensions  Vector-register processors: all vector operations between vector registers (except load and store)  Vector equivalent of load-store architectures  Introduced in the Cray-1  Includes all vector machines since late 1980s: Cray, Convex, Fujitsu, Hitachi, NEC  We assume vector-register for the rest of lecture

8 CSE Parallel and Vector ArchitecturesVector Computers – slide 8 Scalar Registers r0 r15 Vector Registers v0 v15 [0][1][2][VLRMAX-1] [0][1][VLR-1] Vector Arithmetic Instructions ADDV v3, v1, v2 v3 v2 v1 VLR Vector Length Register v1 Vector Load and Store Instructions LV v1, r1, r2 Base, r1Stride, r2 Memory Vector Register Vector Programming Model

9 CSE Parallel and Vector ArchitecturesVector Computers – slide 9 Instr.OperandsOperationComment ADDV V1,V2,V3V1=V2+V3vector + vector ADDSV V1,F0,V2V1=F0+V2scalar + vector MULTV V1,V2,V3V1=V2xV3vector x vector MULSV V1,F0,V2V1=F0xV2scalar x vector LV V1,R1V1=M[R1..R1+63]load, stride=1 LVWS V1,R1,R2V1=M[R1..R1+63*R2]load, stride=R2 LVI V1,R1,V2V1=M[R1+V2i,i=0..63] load, indexed CeqV VM,V1,V2VMASKi = (V1i=V2i)?comp. setmask MOV VLR,R1Vec. Len. Reg. = R1set vector length MOV VM,R1Vec. Mask = R1set vector mask Vector Instructions

10 CSE Parallel and Vector ArchitecturesVector Computers – slide 10  Each result independent of previous result  Long pipeline, compiler ensures no dependencies  High clock rate  Vector instructions access memory with known pattern  Highly interleaved memory  Amortize memory latency of over ­ 64 elements  No (data) caches required! (Do use instruction cache)  Reduces branches and branch problems in pipelines  Single vector instruction implies lots of work (­ loop)  Fewer instruction fetches Properties of Vector Processors

11 CSE Parallel and Vector ArchitecturesVector Computers – slide 11 # Scalar Code LI R4, 64 loop: L.D F0, 0(R1) L.D F2, 0(R2) ADD.D F4, F2, F0 S.D F4, 0(R3) ADDIU R1, 8 ADDIU R2, 8 ADDIU R3, 8 SUBIU R4, 1 BNEZ R4, loop # Vector Code LI VLR, 64 LV V1, R1 LV V2, R2 ADDV V3, V1, V2 SV V3, R3 # C code for (i=0; i<64; i++) C[i] = A[i] + B[i]; Vector Code Example

12 CSE Parallel and Vector ArchitecturesVector Computers – slide 12  Compact  one short instruction encodes N operations  Expressive, tells hardware that these N operations:  Are independent  Use the same functional unit  Access disjoint registers  Access registers in the same pattern as previous instructions  Access a contiguous block of memory (unit-stride load/store)  Access memory in a known pattern (strided load/store)  Scalable  Can run same object code on more parallel pipelines or lanes Vector Instruction Set Advantages

13 CSE Parallel and Vector ArchitecturesVector Computers – slide 13 Components of a Vector Processor  Vector Register File  Has at least 2 read and 1 write ports  Typically 8-32 vector registers  Each holding 64 (or more) 64-bit elements  Vector Functional Units (FUs)  Fully pipelined, start new operation every clock  Typically 4 to 8 FUs: FP add, FP mult, FP reciprocal  Integer add, logical, shift (multiple of same unit)  Vector Load-Store Units (LSUs)  Fully pipelined unit to load or store a vector  May have multiple LSUs  Scalar registers  Single element for FP scalar or address  Cross-bar to connect FUs, LSUs, registers

14 CSE Parallel and Vector ArchitecturesVector Computers – slide 14 Machine Year Clock Regs Elements FUs LSUs Cray MHz86461 Cray XMP MHz86482L, 1S Cray YMP MHz8648 2L, 1S Cray C MHz Cray T MHz Conv. C MHz Conv. C MHz Fuj. VP MHz Fuj. VP MHz NEC SX/ MHz8+8K256+var168 NEC SX/ MHz8+8K256+var168 Examples of Vector Machines

15 CSE Parallel and Vector ArchitecturesVector Computers – slide 15 Vector Arithmetic Execution  Use deep pipeline (=> fast clock) to execute element operations  Simplifies control of deep pipeline because elements in vector are independent  No hazards! V1V1 V2V2 V3V3 V3 <- v1 * v2 Six stage multiply pipeline

16 CSE Parallel and Vector ArchitecturesVector Computers – slide 16 Vector Memory System  Cray-1: 16 banks  4 cycle bank busy time  Bank busy time: Cycles between accesses to same bank  12 cycle latency + BaseStride ABCDEF Vector Registers Memory Banks Address Generator

17 CSE Parallel and Vector ArchitecturesVector Computers – slide 17  Great for unit stride:  Contiguous elements in different DRAMs  Startup time for vector operation is latency of single read  What about non-unit stride?  Above good for strides that are relatively prime to 8  Bad for strides = 2, 4 and worse for strides = multiple of 8  Better: prime number of banks…! Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Unpipelined DRAM Addr+0 Addr+1 Addr+2Addr+4Addr+5Addr+3 Addr+6 Addr+7 Interleaved Memory Layout

18 CSE Parallel and Vector ArchitecturesVector Computers – slide 18 ADDV C,A,B C[1] C[2] C[0] A[3]B[3] A[4]B[4] A[5]B[5] A[6]B[6] Execution using one pipelined functional unit C[4] C[8] C[0] A[12]B[12] A[16]B[16] A[20]B[20] A[24]B[24] C[5] C[9] C[1] A[13]B[13] A[17]B[17] A[21]B[21] A[25]B[25] C[6] C[10] C[2] A[14]B[14] A[18]B[18] A[22]B[22] A[26]B[26] C[7] C[11] C[3] A[15]B[15] A[19]B[19] A[23]B[23] A[27]B[27] Execution using four pipelined functional units Vector Instruction Execution

19 CSE Parallel and Vector ArchitecturesVector Computers – slide 19 Vector Unit Structure Lane Functional Unit Vector Registers Memory Subsystem Elements 0, 4, 8, … Elements 1, 5, 9, … Elements 2, 6, 10, … Elements 3, 7, 11, …

20 CSE Parallel and Vector ArchitecturesVector Computers – slide  Vector register file  Each register is an array of elements  Size of each register determines maximum vector length  Vector length register determines vector length for a particular operation  Multiple parallel execution units = “lanes”  Sometimes called “pipelines” or “pipes” Vector Unit Implementation

21 CSE Parallel and Vector ArchitecturesVector Computers – slide 21 LaneVector register elements striped over lanes [0] [8] [16] [24] [1] [9] [17] [25] [2] [10] [18] [26] [3] [11] [19] [27] [4] [12] [20] [28] [5] [13] [21] [29] [6] [14] [22] [30] [7] [15] [23] [31] T0 Vector Microprocessor (1995) See intro.html intro.html

22 CSE Parallel and Vector ArchitecturesVector Computers – slide 22 for (i=0; i < N; i++) C[i] = A[i] + B[i]; load add store load add store Iter. 1 Iter. 2 Scalar Sequential Code Vectorization is a massive compile-time reordering of operation sequencing  requires extensive loop dependence analysis Vector Instruction load add store load add store Iter. 1 Iter. 2 Vectorized Code Time Automatic Code Vectorization

23 CSE Parallel and Vector ArchitecturesVector Computers – slide 23  Problem: Vector registers have fixed length  What to do if Vector Length > Max Vector Length?  Stripmining: generate code such that each vector operation is done for a size ≤ MVL  First loop iteration: do short piece (n mod MVL)  Remaining iterations: VL = MVL index = 0; /* start at index 0 */ VL = (n mod MVL) /* find the odd size piece */ while (n > 0) { /* do vector instructions on VL elements */ n = n – VL; index = index + VL; VL = MVL /* reset the length to max */ } Vector Stripmining

24 CSE Parallel and Vector ArchitecturesVector Computers – slide 24 ANDI R1, N, 63 # N mod 64 MOV VLR, R1 # Do remainder loop: LV V1, RA SLL R2, R1, 3 # Multiply by 8 ADDU RA, RA, R2 # Advance pointer LV V2, RB ADDU RB, RB, R2 ADDV V3, V1, V2 SV V3, RC ADDU RC, RC, R2 SUBU N, N, R1 # Subtract elements LI R1, 64 MOV VLR, R1 # Reset full length BGTZ N, loop # Any more to do? for (i=0; i

25 CSE Parallel and Vector ArchitecturesVector Computers – slide 25  Two components of vector startup penalty  Functional unit latency: time through pipeline  Dead time: time before another vector instruction can use same pipelined functional unit RXXXW RXXXW RXXXWRXXXX RXXXX RXXXXRXXXXRXXXW RXXXW RXXXW Functional Unit Latency = 1 Read cycle, 3 eXecute cycles, and 1 Write cycle Dead Time First Vector Instruction 2nd Vector Instruction Vector Startup Dead Time

26 CSE Parallel and Vector ArchitecturesVector Computers – slide 26 Dead Time and Short Vectors Cray C90, Two lanes 4-cycle dead time Maximum efficiency 94% with 128 element vectors T0, Eight lanes No dead time 100% efficiency with 8 element vectors No dead time 4-cycle dead time 64 cycles active Dead time reduces the maximum efficiency Can be eliminated with better control logic and fully pipelining instruction startup

27 CSE Parallel and Vector ArchitecturesVector Computers – slide 27  Vector version of register bypassing  Introduced with Cray-1 Memory V1V1 Load Unit Mult. V2V2 V3V3 Chain Add V4V4 V5V5 Chain LV v1 MULV v3,v1,v2 ADDV v5,v3,v4 Vector Chaining

28 CSE Parallel and Vector ArchitecturesVector Computers – slide 28 Vector Chaining Advantage  With chaining, can start dependent instruction as soon as first result appears Load Mul Add Load Mul Add Time  Without chaining, must wait for last element of result to be written before starting dependent instruction

29 CSE Parallel and Vector ArchitecturesVector Computers – slide 29 Vector Instruction Parallelism Can overlap execution of multiple vector instructions Example: 32 elements per vector register and 8 lanes load Load Unit mul Multiply Unit add Add Unit load time Instruction issue Complete 24 operations/cycle while issuing 1 short instruction/cycle muladd

30 CSE Parallel and Vector ArchitecturesVector Computers – slide 30  Vector Execution Time depends on:  Vector length, data dependences, and structural hazards  Initiation rate  Rate at which a vector unit consumes vector elements  Typically, initiation rate = number of lanes  Execution time of a vector instruction = VL / Initiation Rate  Convoy  Set of vector instructions that can execute in same clock  No structural or data hazards (similar to VLIW concept)  Chime  Execution time of one convoy  m convoys take m chimes = approximately m x n cycles  If each chime takes n cycles and no overlapping convoys Vector Execution Time

31 CSE Parallel and Vector ArchitecturesVector Computers – slide 31 LVV1, Rx; Load vector X MULVSV2, V1, F0; vector-Scalar multiply LVV3, Ry; Load vector Y ADDVV4, V2, V3; Add vectors SVRy, V4; Store result in vector Y  4 Convoys => 4 Chimes 1.LV 2.MULVS, LV 3.ADDV 4.SV Suppose VL=64 For 1 Lane:Chime = 64 cycles For 2 Lanes:Chime = 32 cycles For 4 Lanes:Chime = 16 cycles Example on Convoys and Chimes

32 CSE Parallel and Vector ArchitecturesVector Computers – slide 32  Vector startup comes from pipeline latency  Important source of overhead, so far ignored  Startup time = depth of pipeline  Increases the effective time to execute a convoy  Time to complete a convoy depends  Vector startup, vector length, number of lanes OperationStart-up penalty (from CRAY-1) Vector load/store12 cycles Vector multiply7 cycles Vector add6 cycles Startup penalty for load/store can be very high (100 cycles) Vector Startup

33 CSE Parallel and Vector ArchitecturesVector Computers – slide 33  Consider same example with 4 convoys  Vector length = n  Assume Convoys don’t overlays  Show the time of each convoy assuming 1 lane ConvoyStart timeFirst resultLast result 1. LV n 2. MULVS, LV12 + n12 + n n 3. ADDV24 + 2n24 + 2n n 4. SV n30 + 3n n Example on Vector Startup  Total cycles = n (with extra 42 startup cycles)

34 CSE Parallel and Vector ArchitecturesVector Computers – slide 34  Suppose: MULVV1, V2, V3 ADDVV4, V1, V5; RAW dependence  Chaining: Allow a vector operation to start as soon as the individual elements of the vector source operands become available. Forward individual elements of a vector register.  Dependent instructions can be placed in the same convoy (if no structural hazard) MULTV ADDV Total = 141 cycles 128/141 = 0.91 Flops/cycle Unchained = 2 convoys Total = 77 cycles 1.66 Flops/cycle MULTV ADDV Chained = 1 convoy Vector Chaining

35 CSE Parallel and Vector ArchitecturesVector Computers – slide 35 Vector Stride  Adjacent elements are not sequential in memory do 10 i = 1,100 do 10 j = 1,100 A(i,j) = 0.0 do 10 k = 1,100 10A(i,j) = A(i,j) + B(i,k) * C(k,j)  Either B or C accesses are not adjacent  800 bytes between adjacent vector elements  Stride: distance separating elements that are to be merged into a single vector  Caches do unit stride  LVWS (load vector with stride) instruction  Think of addresses per vector element

36 CSE Parallel and Vector ArchitecturesVector Computers – slide 36 Memory Addressing Modes  Load/store operations move groups of data between registers and memory  Three types of vector addressing  Unit stride  Contiguous block of information in memory  Fastest: always possible to optimize this  Non-unit (constant) stride  Harder to optimize memory system for all possible strides  Prime number of data banks makes it easier to support different strides at full bandwidth  Indexed (gather-scatter)  Vector equivalent of register indirect  Good for sparse arrays of data  Increases number of programs that vectorize

37 CSE Parallel and Vector ArchitecturesVector Computers – slide 37 Vector Scatter/Gather Want to vectorize loops with indirect accesses for (i=0; i

38 CSE Parallel and Vector ArchitecturesVector Computers – slide 38 Vector Scatter/Gather Scatter example: for (i=0; i

39 CSE Parallel and Vector ArchitecturesVector Computers – slide 39 Memory Banks  Most vector processors support large number of independent memory banks  Memory banks are need for the following reasons  Multiple Loads/Stores per cycle  Memory bank cycle time > CPU cycle time  Ability to load/store non-sequential elements  Multiple processor sharing the same memory  Each processor generates its stream of load/store instructions

40 CSE Parallel and Vector ArchitecturesVector Computers – slide 40 Example on Memory Banks  The Cray T90 has a CPU cycle = ns  The cycle of the SRAM in memory system = 15 ns  Cray T90 can support 32 processors  Each processor is capable of generating 4 loads and 2 stores per CPU clock cycle  What is the number of memory banks required to allow all CPUs to run at full memory bandwidth  Solution:  Maximum number of memory references per cycle 32 CPUs x 6 references per cycle = 192  Each SRAM busy is busy for 15 / = 6.92 ≈ 7 cycles  To handle 192 requests per cycle requires 192 x 7 = 1344 memory banks  Cray T932 actually has 1024 memory banks

41 CSE Parallel and Vector ArchitecturesVector Computers – slide 41 Problem: Want to vectorize loops with conditional code: for (i=0; i0) then A[i] = B[i] Solution: Add vector mask registers  Vector version of predicate registers, 1 bit per element  Vector operation becomes NOP at elements where mask bit is 0 Code example: CVM# Turn on all bits in Vector Mask LV vA, rA# Load entire A vector SGTV vA, 0 # Set bits in mask register where A>0 LV vA, rB# Load B vector into A under mask SV vA, rA# Store A back to memory under mask Vector Conditional Execution

42 CSE Parallel and Vector ArchitecturesVector Computers – slide 42 Vector Masks  Vector masks have two important uses  Conditional execution and arithmetic exceptions  Alternative is conditional move/merge  More efficient than conditional moves  No need to perform extra instructions  Avoid exceptions  Downside is:  Extra bits in instruction to specify the mask register  For multiple mask registers  Extra interlock early in the pipeline for RAW hazards

43 CSE Parallel and Vector ArchitecturesVector Computers – slide 43 Masked Vector Instructions C[4] C[5] C[1] Write data port A[7]B[7] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 Density-Time Implementation Scan mask vector and Execute only elements with Non-zero masks C[1] C[2] C[0] A[3]B[3] A[4]B[4] A[5]B[5] A[6]B[6] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 Write data portWrite Enable A[7]B[7]M[7]=1 Simple Implementation Execute all N operations Turn off result writeback according to mask

44 CSE Parallel and Vector ArchitecturesVector Computers – slide 44  Compress:  Packs non-masked elements from one vector register contiguously at start of destination vector register  Population count of mask vector gives packed vector length  Used for density-time conditionals and for general selection  Expand: performs inverse operation M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 A[3] A[4] A[5] A[6] A[7] A[0] A[1] A[2] M[3]=0 M[4]=1 M[5]=1 M[6]=0 M[2]=0 M[1]=1 M[0]=0 M[7]=1 B[3] A[4] A[5] B[6] A[7] B[0] A[1] B[2] Expand A[7] A[1] A[4] A[5] Compress A[7] A[1] A[4] A[5] Compress/Expand Operations

45 CSE Parallel and Vector ArchitecturesVector Computers – slide 45 Problem: Loop-carried dependence on reduction variables sum = 0; for (i=0; i1) Vector Reductions

46 CSE Parallel and Vector ArchitecturesVector Computers – slide 46  Multimedia Processing: compress, graphics, audio, images  Standard benchmark kernels: Matrix Multiply, FFT, Sort  Lossy Compression: JPEG, MPEG video and audio  Lossless Compression: Zero removal, RLE, Differencing, LZW  Cryptography: RSA, DES/IDEA, SHA/MD5  Speech and handwriting recognition  Operating systems/Networking: memcpy, parity, checksum  Databases: hash/join, data mining, image/video serving  Language run-time support: stdlib, garbage collection Vector Applications

47 CSE Parallel and Vector ArchitecturesVector Computers – slide 47  “…media processing will become the dominant force in computer architecture & microprocessor design.”  “... new media-rich applications... involve significant real-time processing of continuous media streams, and make heavy use of vectors of packed 8-, 16-, and 32-bit integer and FP”  Needs include high memory BW, high network BW, continuous media data types, real-time response, fine grain parallelism  “How Multimedia Workloads Will Change Processor Design”, Diefendorff & Dubey, IEEE Computer (9/97) New Architecture Direction?

48 CSE Parallel and Vector ArchitecturesVector Computers – slide 48 Multimedia Extensions  Very short vectors added to existing ISAs for micros  Intel MMX: reuses 64-bit FP registers  Split into 2x32b, 4x16b or 8x8b  Add, Subtract in parallel: 8x8b, 4x16b, 2x32b  Optional signed/unsigned saturate if overflow  Shifts (sll,srl, sra) in parallel: 8x8b, 4x16b, 2x32b  Multiply, Multiply-Add in parallel: 4x16b  Compare =, > in parallel: 8x8b, 4x16b, 2x32b  Sets field to 0s (false) or 1s (true); removes branches  Pack/Unpack  Convert 32b ↔ 16b, 16b ↔ 8b  Pack saturates (set to max) if number is too large

49 CSE Parallel and Vector ArchitecturesVector Computers – slide 49 Multimedia Extensions – cont’d  Newer designs have 128-bit registers (Altivec, SSE2)  Limited instruction set:  No vector length control  No strided load/store or scatter/gather  Unit-stride loads must be aligned to 64/128-bit boundary  Limited vector register length:  Superscalar dispatch to keep multiply/add/load units busy  Loop unrolling to hide latencies increases register pressure  Trend towards full vector support in microprocessors

50 CSE Parallel and Vector ArchitecturesVector Computers – slide 50  Emotion Engine: 6.2 GFLOPS, 75 million polygons per second  Graphics Synthesizer: 2.4 Billion pixels per second  Claim: Toy Story realism brought to games! Sony Playstation 2000

51 CSE Parallel and Vector ArchitecturesVector Computers – slide 51  Sample Vector Unit  2-wide VLIW  Includes Microcode Memory  High-level instructions like matrix-multiply  Emotion Engine:  Superscalar MIPS core  Vector Coprocessor Pipelines  RAMBUS DRAM interface Playstation 2000 – Continued

52 CSE Parallel and Vector ArchitecturesVector Computers – slide 52  CMOS Technology  500 MHz CPU, fits on single chip  SDRAM main memory (up to 64GB)  Scalar unit  4-way superscalar, out-of-order and speculative  64KB I-cache and 64KB data cache  Vector unit  8 foreground VRegs + 64 background VRegs  256x64-bit elements/VReg  1 multiply, 1 divide, 1 add/shift, 1 logical, 1 mask unit  8 lanes (8 GFLOPS peak, 16 FLOPS/cycle)  1 load & store unit, 32x8 byte accesses/cycle, 32 GB/s bandwidth  SMP structure  8 CPUs connected to memory through crossbar  256 GB/s shared memory bandwidth (4096 interleaved banks) A Modern Vector Super Computer: NEC SX-6 (2003)

53 CSE Parallel and Vector ArchitecturesVector Computers – slide 53 Vector  One instruction fetch, decode, dispatch per vector  Structured register accesses  Smaller code for high performance, less power in instruction cache misses  Bypass data cache  One TLB lookup per group of loads or stores  Move only necessary data across chip boundary Single-issue Scalar  One instruction fetch, decode, dispatch per operation  Arbitrary register accesses adds area and power  Loop unrolling and software pipelining for high performance increases instruction cache footprint  All data passes through cache, waste power if no temporal locality  One TLB lookup per load or store  Off-chip access in whole cache lines Scalar versus Vector Energy Efficiency

54 CSE Parallel and Vector ArchitecturesVector Computers – slide 54 Vector  Control logic grows linearly with issue width  N ops issued per cycle  2 ) circuitry  Vector unit switches off when not in use  Vector instructions expose parallelism without speculation  Software control of vector masks or compress/expand for conditionals Superscalar  Control logic grows quadratically with issue:  N ops issued per cycle  2 ) circuitry  Control logic consumes energy regardless of available parallelism  Speculation wastes energy to increase visible parallelism Superscalar Energy Efficiency

55 CSE Parallel and Vector ArchitecturesVector Computers – slide 55 Summary  Vector is a model for exploiting Data Parallelism  If code is vectorizable, then simpler hardware, more energy efficient, and better real-time model than Out-of-order machines  Design issues include number of lanes, number of functional units, number of vector registers, length of vector registers, exception handling, and conditional operations  Fundamental design issue is memory bandwidth  With virtual address translation and caching  Will multimedia popularity revive vector architectures?


Download ppt "CSE 661 - Parallel and Vector ArchitecturesVector Computers – slide 1 Vector Computers Muhamed Mudawar Computer Engineering Department King Fahd University."

Similar presentations


Ads by Google