SlideShare a Scribd company logo

Single instruction multiple data

Download as PPT, PDF
Syed Zaid Irshad
Syed Zaid Irshad

Single Instruction Multiple Data (SIMD) is an approach to improve performance by replicating data paths rather than control. Vector processors apply the same operation to all elements of a vector in parallel. The ILLIAC IV was an early SIMD computer from 1972 with 64 processing elements. Vector processors store vectors in registers and apply the same instruction to all elements simultaneously. The Cray-1 was an influential vector supercomputer from 1978 that used vector registers and optimized memory access for vectors. Vectorization improves performance by performing the same operation on multiple data elements with a single instruction.

Engineering
1 of 66
Downloaded 65 times
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Most read
19
20
21
22
23
24
25
26
27
Most read
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
Most read
Single Instruction Multiple Data Another approach to ILP and performance
Outline • Array Processors / “True” SIMD • Vector Processors • Multimedia Extensions in modern instruction sets
SIMD: Motivation • Let’s start with an example: • ILLIAC IV, U of Illinois, 1972 (prototype) • Reasoning: How to Improve Performance • Rely on Faster Circuits • Cost/circuit increases with circuit speed • At some point, cost/performance unfavorable • Concurrency: • Replicate Resources • Do more per cycle
SIMD: Motivation contd.• Replication to the extreme: Multi-processor • Very Felixible, but costly • Do we need all this flexibility? • There are middle-ground designs were only parts are replicated CU ALU MEM Uniprocessor replicate CU ALU MEM CU ALU MEM CU ALU MEM Multiprocessor
SIMD: Motivation Contd. • Recall: • Part of architecture is understanding application needs • Many Apps: • for i = 0 to infinity • a(i) = b(i) + c • Same operation over many tuples of data • Mostly independent across iterations
SIMD Architecture • Replicate Datapath, not the control • All PEs work in tandem • CU orchestrates operations CU PE MEM PE MEM PE MEM ALU μCU regs
ILLIAC IV • Goal: • 1 Gops/sec • 256 PEs as four partitions of 64 PEs • What was built • 0.2 Minsts/sec (we’ll talk about peak performance as ops) • 64 PEs • Prototype due date 1972
ILLIAC IV CU PE PMEM PE PMEM PE PMEM I/O Proc
ILLIAC IV Processing Element (PE) • 64-bit numbers, float or fixed point • Multiples of smaller numbers that add up to 64-bits • Today’s multimedia extensions • PMEM: One local memory module per PE • 2K x 64-bits • 188ns access / 350ns cycle (includes conflict resolution) • 100K components per PE
PE Contd. • PE mode: Active or Inactive, CU sets mode • All PEs operate in lock-step • Routing insts to move data from PE to PE • The CU can execute instructions while PE’s are busy • Another degree of concurrency • Datatypes • 64b float • 64b logical • 48b fixed • 32 float • 24 fixed • 8 fixed
Peak Compute Bandwidth • 64 PEs • Each can perform: • 1 64b, 2 32b, or 4 8b operations • Or, in total: • 64 elems, 128 elems, or 512 elems • Peak: • 150M 64b ops/sec up to 10G 32b ops/sec • The last figure is for integer ops • Each int op takes 66ns (4 per PE in parallel)
Control Unit (CU) • A simple CPU • Can execute instructions w/o PE intervention • Coordinates all PEs • 64 64b registers, D0-D63 • 4 64b Accumulators A0-A3 • Ops: • Integer ops • Shifts • Boolean • Loop control • Index PMem D0 D63 A0 A3 A1 A2 ALU CU
Processing Element (PE)• 64 bit regs • A: Accumulator • B: 2nd operand for binary ops • R: Routing – Inter-PE Communication • S: Temporary • X: Index for PMEM 16bits • D: mode 8bits • Communication: • PMEM only from local PE • Amongst PE with R A S B R ALU PEi X D 0 1 2043 PMEMi PEi-1 PEi+1 PEi-8 PEi+8
Datapaths • CU Bus: Insts and Data from PMEM to CU in 8 words • CDB: Broadcast to all PEs • E.g., constants for adds • Routing Network: amongst R registers • Mode: To activate/de-activate PEs CU PE PMEM PE PMEM PE PMEM Control Unit Bus Mode Common Data Bus Routing
Routing Network 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 0 8 16 24 32 40 48 56 7 15 23 31 39 47 55 63 56 57 58 59 60 61 62 63 0 1 2 3 4 5 6 7 12 19 20 21 28 i-8 i+8 i+1i-1
Using ILLIAC IV: Example #2 • DO 10 I = 1 TO 64 10 C(I) = A(I) + B(I) • LDA a + 2 load A(i) into A (same a per PMEM) • ADDRN a + 1 add B(i) into A • STA a store A into C(i) C(1) A(1) B(1) PMEM1 a C(2) A(2) B(2) PMEM2 C(64) A(64) B(64) PMEM64
Using ILLIAC IV: Example #2 • DO 10 I = 2 TO 64 • 10 A(I) = B(I) + A(I-1) • Expand into: • A(N) = A(1) + Sum B(i) [i = 2 to N] • We get: • DO 10 N=2 TO 64 • S = S + B(N) • 10 A(N) = S
Using ILLIAC IV: Example #2 contd. 1. Enable all PEs 2. All load A from a 3. i = 0 4. All R = A (including those inactive) 5. All route R to PE(2^i) to the right 6. j = 2^i – 1 7. Disable all PEs 1 through j 8. A = A + R  R contains a partial sum of many A(i) 9. i = i + 1 10. if i < lg(64) goto 4 11. Enable All PEs 12. All store A at (a + 1)
Using ILLIAC IV: Example #2 contd. • Initial State: • PMEM(1)[a] = A(1) • PMEM(1+i)[a] = B(i+1) • For example, at PE1 • STEP 1: A = A(1) • from PE2 we get B(2) • STEP 2: A = A(1) + B(2) • from PE4 we get B(4) + B(5) • STEP 3: A = A(1) + B(2) + B(4) + B(5) • From PE8 we get B(8) + B(7) + B(12) + B(13)
Vector Processors SIMD over time
Vector Processors • Vector Datatype • Apply same operation on all elements of the vector • No dependences amongst elements • Same motivation as SIMD
Properties of Vector Processors • One Vector instruction implies lots of work • Fewer instructions • Each result independent of previous result • Multiple operations in parallel • Simpler design; no need for dependence checks • Higher clock rate • Compiler must help • Fewer Branches • Memory access pattern per vector inst known • Prefetching effect • Amortize mem latency • Can exploit high-bandwidth mem system • Less/no need for data caches
Classes of Vector Processors • Memory to memory • Vectors are in memory • Load/store • Vectors are in registers • Load/store to communicate with memory • This prevailed
Historical Perspective • Mid-60s: performance concerns • SIMD processor arrays • Also fast Scalar machines • CDC 6600 • Texas Instruments ASC, 1972 • Memory to memory vector • Cray Develops CRAY-1, 1978
CRAY-1 • Fast and simple scalar processor • 80 Mhz • Vector register concept • Much simple ISA • Reduced memory pressure • Tight integration of scalar and vector units • Cylindrical design to minimize wire lengths • Freon Cooling
Physical Organization of CRAY-1
Components of Vector Processor • Scalar CPU: registers, datapaths, instruction fetch • Vector Registers: • Fixed length memory bank holding a single vector reg • Typically 8-32 Vregs, up to 8Kbits per Vreg • At least; 2 Read, 1 Write ports • Can be viewed as an array of N elements • Vector Functional Units: • Fully pipelined. New op per cycle • Typically 2 to 8 FUs: integer and FP • Multiple datapaths to process multiple elements per cycle if needed • Vector Load/Store Units (LSUs): • Fully pipelined • Multiple elems fetched/store per cycle • May have multiple LSUs • Cross-bar: • Connects FUS, LSUs and registers
CRAY-1 Organization • Simple 16-bit Reg-to-Reg ISA • Use two 16-bit to get Imm • Natural combinations of scalar and vector • Scalar bit-vectors match vector length • Gather/Scatter M-R • Cond. Merge
CRAY-1 CPU • Scalar and vector modes • 12.5 ns clock • 64-bit words • Int & FP units • 12 FUs • 8 24-bit A regs • 64 B regs (temp storage for A) • 8 64-bit S regs • 64 T regs (temp storage for S) • 64 64-elem, 64bit elem V regs
CRAY-1 CPU • Vector Length Register • Can use only a prefix of a vreg • Vector Mask Register • Can use only a subset of a vreg • Real Time Register (counts clock cycles) • Four instruction buffers • 64 16-bit parcels • 128 Basic Instructions • Interrupt Control • NO virtual memory system
Cray-1 Memory System • 1M 64b words + 8 check bits (single error correction, double error detection) • 16 banks of 64K words • 4 clocks period • 1 word per clock for B, T and Vreg • 1 word per 2 clocks for A & S • 4 words per clock for inst buffers
Instruction Format • Fields g h I j k m • Bits 0-3 4-6 7-9 10-12 13-15 16-31 • Bits cnts 4 3 3 3 3 16 • X X  opcode • Rd Rs1 Rs2 • A/S B/T
Basic Vector Instructions • Inst Operands Operation Comment • VADD.VV V1, V2, V3 V1=V2+V3 vector+vector • VADD.SV V1, R0, V2 V1=R0+V2 scalar+vector • VMUL.VVV1, V2, V3 V1=V2*V3 vector * vector • VMUL.SV V1, R0, V2 V1=R0*V2 scalar * vector • VLD V1, R0 V1=M[R0…R0+63] stride = 1 • VLDS V1, R1, R2 V1=M[R1…R1+63*R2] stride=R2 • VLDX V1, R1, V2 V1=M[R1+V2[i], i=0 to 63] gather • VST store equiv of VLD • VSTS store equiv of VLDS • VSTX V1, R1 M[R1+V2[i], i=0 to 63]=V1 scatter
Vector Memory Operations • Load/Store move groups of data between memory and registers • Addressing Modes • Unit-stride: Fastest • Non-Unit, constant stride (interleaved memory helps • Indexed (gather-scatter) • Vector equiv of register indirect • Sparse arrays • Can vectorize more loops
Vector Code Example • Y[0:63] = Y[0:63] + a * X[0:63] • LD R0, a • VLD V1, Rx Load X[] in V1 • VLD V2, Ry Load Y[] in V2 • VMUL.SV V3, R0, V1 V3 = X[]*a • VADD.VV V4, V2, V3 V4 = Y[]+X[]*a • VST Ry, V4 store in Y[]
Scalar Equivalent • LD R0, a • LI R5, 512 (offset at the end of X[]) • Loop: LD R2, 0(Rx) • MULTD R2, R0, R2 • LD R3, 0(Ry) • ADD R4, R2, R3 • ST R4, 0(Ry) • ADD Rx, Rx, 8 • ADD Ry, Ry, 8 • SUB R5, R5, 8 • BNE Loop LD R0, a VLD V1, Rx VLD V2, Ry VMUL.SV V3, R0, V1 VADD.VV V4, V2, V3 VST Ry, V4
Vector Length Register • Allows us to vectorize code where the elements do not exactly fit within the vector register • What if we need a vector of just 32 elems? • Vector length register: • Operate up to this element • Can be anything from 0 to Maximum (64 in CRAY-1) • Can also be used to support runtime vector length variability
Strip Mining • Suppose (application vector length) AVL > MVL (max vector length) • Each loop iteration handles MVL elems • Last iteration AVL MOD MVL • VL = (AVL mode MVL) • For (I=0; I<VL; I++) • Y[I] = A*X[I] + Y[I] • low = (AVL mod MVL) • VL = low • For (i=low; i < VL; i++) • Y[i] = A*X[i] + Y[i]
Optimization #1: Chaining• Subsequent vector op can be initiated as soon as a preceding vector op it depends upon produces its first result • Example • Vadd.vv v1, v2, v3 • Vadd.sv v4, v1, R0 V1(1) V1(2) V1(3) V1(4) V1(63) time Add initiated V4(1) V4(2) V4(3) V4(4) V4(63) unchained V1(1) V1(2) V1(3) V1(4) V1(63) Add initiated V4(1) V4(2) V4(3) V4(4) V4(63) chained
Optimization #2: Conditional Execution • Vector Mask Register • Bit vector: used as predicate • If 0 operation is not performed for the corresponding pair • VLD V1, Ra • VLD V2, Rb • VCMP.NEQ.VV VMR, V1, V2 • VSUB.VV V3, V2, V1 (VMR) • VST V3, Ra • For (i = 0; i < 64; i++) • if (A[i] != B[i]) A[i] = A[i] – B[i]
Optimization #3: Multi-lane Implementation • Vectors are interleaved so that multiple elems can be accessed per cycle • Replicate resources • Equivalent of Superscalar • Because of no intra-vector dependences and because inter-vector dependences are aligned (elem(i) to elem(i)) no need for inter-bank communications
Two Ways to View Vectorization • Classic Approach: Inner-loop • Think machine as having 32 vector registers with 16 elems • 1 instruction updates all elements of a vector • Vectorize single dimension array operations • A new approach: Outer-loop • Think of machine as 16 “virtual processors” each with 32 scalar registers • 1 instruction updates register in 16 VPs • Good for irregular kernels • Hardware is the same for both • These describe the compiler’s perspective
Startup Cost
Execution Cost
Multimedia extensions SIMD in modern CPUs
Multimedia ISA Extensions • Intel’s MMX • The Basics • Instruction Set • Examples • Integration into Pentium • Relationship to vector ISAs • AMD’s 3DNow! • Intel’s ISSE (a.k.a. KNI)
MMX: Basics • Multimedia applications are becoming popular • Are current ISAs a good match for them? • Methodology: • Consider a number of “typical” applications • Can we do better? • Cost vs. performance vs. utility tradeoffs • Net Result: Intel’s MMX • Can also be viewed as an attempt to maintain market share • If people are going to use these kind of applications we better support them
Multimedia Applications • Most multimedia apps have lots of parallelism: • for I = here to infinity • out[I] = in_a[I] * in_b[I] • At runtime: • out[0] = in_a[0] * in_b[0] • out[1] = in_a[1] * in_b[1] • out[2] = in_a[2] * in_b[2] • out[3] = in_a[3] * in_b[3] • ….. • Also, work on short integers: • in_a[i] is 0 to 256 for example (color) • or, 0 to 64k (16-bit audio)
Observations • 32-bit registers are wasted • only using part of them and we know • ALUs underutilized and we know • Instruction specification is inefficient • even though we know that a lot of the same operations will be performed still we have to specify each of the individually • Instruction bandwidth • Discovering Parallelism • Memory Ports? • Could read four elements of an array with one 32-bit load • Same for stores • The hardware will have a hard time discovering this • Coalescing and dependences
MMX Contd. • Can do better than traditional ISA • new data types • new instructions • Pack data in 64-bit words • bytes • “words” (16 bits) • “double words” (32 bits) • Operate on packed data like short vectors • SIMD • First used in Livermore S-1 (> 20 years)
MMX:Example Up to 8 operations (64bit) go in parallel  Potential improvement: 8x  In practice less but still good Besides another reason to think your machine is obsolete
Data Types
MMX: Instruction Set • 57 new instructions • Integer Arithmetic • add/sub/mul • multiply add • signed/unsigned • saturating/wraparound • Shifts • Compare (form mask) • Pack/Unpack • Move • from/to memory • from/to registers
Arithmetic • Conventional: Wrap-around • on overflow, wrap to -1 • on underflow, wrap to MAXINT • Think of digital audio • What happens when you turn volume to the MAX? • Similar for pictures • Saturating arithmetic: • on overflow, stay at MAXINT • on underflow, stat at MININT • Two flavors: • unsigned • signed
Operations • Mult/Add • Compares • Conversion • Interpolation/Transpose • Unpack (e.g., byte to word) • Pack (e.g., word to byte)
Matrix Transpose 4x4 • That’s for the first two rows m33 m32 m31 m30 m13 m12 m11 m10 m23 m22 m21 m20 m03 m02 m01 m00 punpcklwd punpcklwd m31 m21 m30 m20 m11 m01 m10 m00 punpckhdq punpckldq m31 m21 m11 m01 m30 m20 m10 m00 m03 m02 m01 m00 m13 m12 m11 m10 m23 m22 m21 m20 m33 m32 m31 m30 m30 m20 m10 m00 m31 m21 m11 m01 m33 m22 m12 m02 m33 m23 m13 m03
Examples • Image Composting • A and B images fade-in and fade-out • A * fade + B * (1 - fade), OR • (A - B) * fade + B • Image Overlay • Sprite: e.g., mouse cursor • Spite: normal colors + transparent • for i = 1 to Sprite_Length • if A[I] = clear_color then • Out_frame[I] = C[I] • else Out_frame[I] = A[I] • Matrix Transpose • Covert from row major to column major • Used in JPEG
Chroma Keying • for (i=0; i<image_size; i++) • if (x[i] == Blue) new_image[i] =y[i] • else new_image[i] = x[i];
Chroma Keying Code • Movq mm3, mem1 • Load eight pixels from persons’ image • Movq mm4, mem2 • Load eight pixels from the background image • Pcmpeqb mm1, mm3 • Pand mm4, mm1 • Pandn mm1, mm3 • Por mm4, mm1
Integration into Pentium • Major issue: OS compatibility • Create new registers? • Share registers with FP • Existing OSes will save/restore • Use 64-bit datapaths • Pipe capable of 2 MMX IPC • Separate MEM and Execute stage
“Recent” Multimedia Extensions • Intel MMX: integer arithmetic only • New algorithms -> new needs • Need for massive amounts of FP ops • Solution? MMX like ISA but for FP not only integer • Example: AMD’s 3DNow! • New data type: • 2 packed single-precision FP • 2 x 32-bits • sign + exponent + significant • New instructions • Speedup potential: 2x
AMD’s 3DNow! • 21 new instructions • Average: motivated by MPEG • Add, Sub, Reverse Sub, Mul • Accumulate • (A1, A2) acc (B1, B2) = (B1 + B2, A1 + A2) • Comparison (create mask) • Min, Max (pairwise) • Reciprocal and SQRT, • Approximation: 1st step and other steps • Prefetch • Integer from/to FP conversion • All operate on packed FP data • sign * 2^(mantissa - 127) * exponent
Recent Extensions Cont. • Intel’s ISSE • very similar to AMD’s 3DNow! • But has separate registers • Lessons? • Applications change over time • Careful when introducing new instructions • How useful are they? • Cost? • LEGACY: are they going to be useful in the future? • Everyone has their own Multimedia Instruction set these days • read handout
Intel’s SSE • Multimedia/Internet? • 70 new instructions • Major Types: • SIMD-FP 128-bit wide 4 x 16 bit FP • Data movement and re-organization • Type conversion • Int to Fp and vice versa • Scalar/FP precision • State Save/Restore • New SSE registers not like MMX • Memory Streaming • Prefetch to specified hierarchy level • New Media • Absolute Diff, Rounded AVG, MIN/MAX
Altivec (PowerPC Mmedia Ext) • 128-bit registers • 8, 16, or 32 bit data types • Scalar or single-precision FP • 162 Instructions • Saturation or Modulo arithmetic • Four operand Instructions • 3 sources, 1 target
Altivec Design Process • Look at Mmedia Kernel • Justify new instructions • Video • 8bit int LowQ, 16-bit int HighQ • Audio • 16bit int LowQ, SP FP HighQ • Image Processing • 8bit int LowQ, 16bit Int HighQ • 3D Graphics • 16bit int LowQ, SP FP HighQ • Speech Recog. • 16bit int Low Q, Sp FP HighQ • Communications/Crypto • 8-bit or 16bit unsigned int

More Related Content

What's hot (20)

PDF
Array Processor
Anshuman Biswal
 
PPTX
Introduction to parallel processing
Page Maker
 
PDF
Course outline of parallel and distributed computing
National College of Business Administration & Economics ( NCBA&E)
 
PPTX
Addressing mode Computer Architecture
Haris456
 
PPT
Unit 3-pipelining &amp; vector processing
vishal choudhary
 
PPTX
Parallel computing and its applications
Burhan Ahmed
 
PDF
Describr the features of pentium microppr
edwardkiwalabye1
 
PDF
Computer Organization
anishgoel
 
PPTX
Pentium processor
salmanjilani
 
PPT
Flynn's Classification parallel processing
JacquesMark
 
PPTX
Data warehouse system and its concepts
Gaurav Garg
 
PPTX
Parallel Processors (SIMD)
Ali Raza
 
PDF
Pipelining and ILP (Instruction Level Parallelism)
Dr. A. B. Shinde
 
PPT
EMBEDDED OS
AJAL A J
 
PPTX
VIRTUAL MEMORY
Kamran Ashraf
 
PPTX
Segmentation
Md. Tanvir Hossain
 
PPTX
INSTRUCTION LEVEL PARALLALISM
Kamran Ashraf
 
PPT
11 instruction sets addressing modes
Sher Shah Merkhel
 
PPT
Process threads operating system.
Reham Maher El-Safarini
 
PPTX
Page replacement algorithms
Piyush Rochwani
 
Array Processor
Anshuman Biswal
 
Introduction to parallel processing
Page Maker
 
Course outline of parallel and distributed computing
National College of Business Administration & Economics ( NCBA&E)
 
Addressing mode Computer Architecture
Haris456
 
Unit 3-pipelining &amp; vector processing
vishal choudhary
 
Parallel computing and its applications
Burhan Ahmed
 
Describr the features of pentium microppr
edwardkiwalabye1
 
Computer Organization
anishgoel
 
Pentium processor
salmanjilani
 
Flynn's Classification parallel processing
JacquesMark
 
Data warehouse system and its concepts
Gaurav Garg
 
Parallel Processors (SIMD)
Ali Raza
 
Pipelining and ILP (Instruction Level Parallelism)
Dr. A. B. Shinde
 
EMBEDDED OS
AJAL A J
 
VIRTUAL MEMORY
Kamran Ashraf
 
Segmentation
Md. Tanvir Hossain
 
INSTRUCTION LEVEL PARALLALISM
Kamran Ashraf
 
11 instruction sets addressing modes
Sher Shah Merkhel
 
Process threads operating system.
Reham Maher El-Safarini
 
Page replacement algorithms
Piyush Rochwani
 

Similar to Single instruction multiple data (20)

PPT
chapter4.ppt
Rakesh Pogula
 
PPTX
Computer Architecture Vector Computer
Haris456
 
PPTX
Data-Level Parallelism in Vector, SIMD, and GPU Architectures.pptx
aliali240367
 
PPTX
Caqa5e ch4
Aravindharamanan S
 
PPTX
Data-Level Parallelism in Microprocessors
Dilum Bandara
 
PPTX
05 instruction set design and architecture
Waqar Jamil
 
PPTX
Ca unit v 27 9-2020
Thyagharajan K.K.
 
PPT
Happy To Use SIMD
Wei-Ta Wang
 
PPTX
Computer Organisation Part 4
Krishna Kumar Bohra
 
PPTX
chapter 1 -Basic Structure of Computers.pptx
janani603976
 
PPT
Instruction Set Architecture (ISA)
Gaditek
 
PDF
Computer organization basics
Deepak John
 
PPT
Performance Tuning by Dijesh P
PlusOrMinusZero
 
PPT
Al2ed chapter2
Abdullelah Al-Fahad
 
PPTX
Modern CPUs and Caches - A Starting Point for Programmers
Yaser Zhian
 
PPTX
Coa swetappt copy
sweta_pari
 
PPTX
Parallel Computing
Mohsin Bhat
 
PPTX
Advanced computer architecture
krishnaviswambharan
 
PPTX
x86 architecture
i i
 
PPT
Chapter 4 the processor
s9007912
 
chapter4.ppt
Rakesh Pogula
 
Computer Architecture Vector Computer
Haris456
 
Data-Level Parallelism in Vector, SIMD, and GPU Architectures.pptx
aliali240367
 
Caqa5e ch4
Aravindharamanan S
 
Data-Level Parallelism in Microprocessors
Dilum Bandara
 
05 instruction set design and architecture
Waqar Jamil
 
Ca unit v 27 9-2020
Thyagharajan K.K.
 
Happy To Use SIMD
Wei-Ta Wang
 
Computer Organisation Part 4
Krishna Kumar Bohra
 
chapter 1 -Basic Structure of Computers.pptx
janani603976
 
Instruction Set Architecture (ISA)
Gaditek
 
Computer organization basics
Deepak John
 
Performance Tuning by Dijesh P
PlusOrMinusZero
 
Al2ed chapter2
Abdullelah Al-Fahad
 
Modern CPUs and Caches - A Starting Point for Programmers
Yaser Zhian
 
Coa swetappt copy
sweta_pari
 
Parallel Computing
Mohsin Bhat
 
Advanced computer architecture
krishnaviswambharan
 
x86 architecture
i i
 
Chapter 4 the processor
s9007912
 
Ad

More from Syed Zaid Irshad (20)

PDF
Data Structures & Algorithms - Spring 2025.pdf
Syed Zaid Irshad
 
PDF
Operating System.pdf
Syed Zaid Irshad
 
PDF
DBMS_Lab_Manual_&_Solution
Syed Zaid Irshad
 
PPTX
Data Structure and Algorithms.pptx
Syed Zaid Irshad
 
PPTX
Design and Analysis of Algorithms.pptx
Syed Zaid Irshad
 
PPTX
Professional Issues in Computing
Syed Zaid Irshad
 
PDF
Reduce course notes class xi
Syed Zaid Irshad
 
PDF
Reduce course notes class xii
Syed Zaid Irshad
 
PDF
Introduction to Database
Syed Zaid Irshad
 
PDF
C Language
Syed Zaid Irshad
 
PDF
Flowchart
Syed Zaid Irshad
 
PDF
Algorithm Pseudo
Syed Zaid Irshad
 
PDF
Computer Programming
Syed Zaid Irshad
 
PDF
ICS 2nd Year Book Introduction
Syed Zaid Irshad
 
PDF
Security, Copyright and the Law
Syed Zaid Irshad
 
PDF
Computer Architecture
Syed Zaid Irshad
 
PDF
Data Communication
Syed Zaid Irshad
 
PDF
Information Networks
Syed Zaid Irshad
 
PDF
Basic Concept of Information Technology
Syed Zaid Irshad
 
PDF
Introduction to ICS 1st Year Book
Syed Zaid Irshad
 
Data Structures & Algorithms - Spring 2025.pdf
Syed Zaid Irshad
 
Operating System.pdf
Syed Zaid Irshad
 
DBMS_Lab_Manual_&_Solution
Syed Zaid Irshad
 
Data Structure and Algorithms.pptx
Syed Zaid Irshad
 
Design and Analysis of Algorithms.pptx
Syed Zaid Irshad
 
Professional Issues in Computing
Syed Zaid Irshad
 
Reduce course notes class xi
Syed Zaid Irshad
 
Reduce course notes class xii
Syed Zaid Irshad
 
Introduction to Database
Syed Zaid Irshad
 
C Language
Syed Zaid Irshad
 
Flowchart
Syed Zaid Irshad
 
Algorithm Pseudo
Syed Zaid Irshad
 
Computer Programming
Syed Zaid Irshad
 
ICS 2nd Year Book Introduction
Syed Zaid Irshad
 
Security, Copyright and the Law
Syed Zaid Irshad
 
Computer Architecture
Syed Zaid Irshad
 
Data Communication
Syed Zaid Irshad
 
Information Networks
Syed Zaid Irshad
 
Basic Concept of Information Technology
Syed Zaid Irshad
 
Introduction to ICS 1st Year Book
Syed Zaid Irshad
 
Ad

Recently uploaded (20)

PPTX
artificial intelligence applications in Geomatics
NawrasShatnawi1
 
PPTX
Day2 B2 Best.pptx
helenjenefa1
 
PPTX
The Role of Information Technology in Environmental Protectio....pptx
nallamillisriram
 
DOCX
CS-802 (A) BDH Lab manual IPS Academy Indore
thegodhimself05
 
PPTX
Introduction to Design of Machine Elements
PradeepKumarS27
 
PPTX
Heart Bleed Bug - A case study (Course: Cryptography and Network Security)
Adri Jovin
 
PPTX
Introduction to Neural Networks and Perceptron Learning Algorithm.pptx
Kayalvizhi A
 
PPTX
VITEEE 2026 Exam Details , Important Dates
SonaliSingh127098
 
PPTX
Solar Thermal Energy System Seminar.pptx
Gpc Purapuza
 
PPT
PPT2_Metal formingMECHANICALENGINEEIRNG .ppt
Praveen Kumar
 
PPTX
Mechanical Design of shell and tube heat exchangers as per ASME Sec VIII Divi...
shahveer210504
 
PPTX
Hashing Introduction , hash functions and techniques
sailajam21
 
PPTX
原版一样(Acadia毕业证书)加拿大阿卡迪亚大学毕业证办理方法
Taqyea
 
PDF
MAD Unit - 1 Introduction of Android IT Department
JappanMavani
 
PPTX
Types of Bearing_Specifications_PPT.pptx
PranjulAgrahariAkash
 
PDF
Pressure Measurement training for engineers and Technicians
AIESOLUTIONS
 
PDF
Biomechanics of Gait: Engineering Solutions for Rehabilitation (www.kiu.ac.ug)
publication11
 
PDF
PORTFOLIO Golam Kibria Khan — architect with a passion for thoughtful design...
MasumKhan59
 
PPTX
Element 11. ELECTRICITY safety and hazards
merrandomohandas
 
PPTX
Worm gear strength and wear calculation as per standard VB Bhandari Databook.
shahveer210504
 
artificial intelligence applications in Geomatics
NawrasShatnawi1
 
Day2 B2 Best.pptx
helenjenefa1
 
The Role of Information Technology in Environmental Protectio....pptx
nallamillisriram
 
CS-802 (A) BDH Lab manual IPS Academy Indore
thegodhimself05
 
Introduction to Design of Machine Elements
PradeepKumarS27
 
Heart Bleed Bug - A case study (Course: Cryptography and Network Security)
Adri Jovin
 
Introduction to Neural Networks and Perceptron Learning Algorithm.pptx
Kayalvizhi A
 
VITEEE 2026 Exam Details , Important Dates
SonaliSingh127098
 
Solar Thermal Energy System Seminar.pptx
Gpc Purapuza
 
PPT2_Metal formingMECHANICALENGINEEIRNG .ppt
Praveen Kumar
 
Mechanical Design of shell and tube heat exchangers as per ASME Sec VIII Divi...
shahveer210504
 
Hashing Introduction , hash functions and techniques
sailajam21
 
原版一样(Acadia毕业证书)加拿大阿卡迪亚大学毕业证办理方法
Taqyea
 
MAD Unit - 1 Introduction of Android IT Department
JappanMavani
 
Types of Bearing_Specifications_PPT.pptx
PranjulAgrahariAkash
 
Pressure Measurement training for engineers and Technicians
AIESOLUTIONS
 
Biomechanics of Gait: Engineering Solutions for Rehabilitation (www.kiu.ac.ug)
publication11
 
PORTFOLIO Golam Kibria Khan — architect with a passion for thoughtful design...
MasumKhan59
 
Element 11. ELECTRICITY safety and hazards
merrandomohandas
 
Worm gear strength and wear calculation as per standard VB Bhandari Databook.
shahveer210504
 

Single instruction multiple data

  • 1. Single Instruction Multiple Data Another approach to ILP and performance
  • 2. Outline • Array Processors / “True” SIMD • Vector Processors • Multimedia Extensions in modern instruction sets
  • 3. SIMD: Motivation • Let’s start with an example: • ILLIAC IV, U of Illinois, 1972 (prototype) • Reasoning: How to Improve Performance • Rely on Faster Circuits • Cost/circuit increases with circuit speed • At some point, cost/performance unfavorable • Concurrency: • Replicate Resources • Do more per cycle
  • 4. SIMD: Motivation contd.• Replication to the extreme: Multi-processor • Very Felixible, but costly • Do we need all this flexibility? • There are middle-ground designs were only parts are replicated CU ALU MEM Uniprocessor replicate CU ALU MEM CU ALU MEM CU ALU MEM Multiprocessor
  • 5. SIMD: Motivation Contd. • Recall: • Part of architecture is understanding application needs • Many Apps: • for i = 0 to infinity • a(i) = b(i) + c • Same operation over many tuples of data • Mostly independent across iterations
  • 6. SIMD Architecture • Replicate Datapath, not the control • All PEs work in tandem • CU orchestrates operations CU PE MEM PE MEM PE MEM ALU μCU regs
  • 7. ILLIAC IV • Goal: • 1 Gops/sec • 256 PEs as four partitions of 64 PEs • What was built • 0.2 Minsts/sec (we’ll talk about peak performance as ops) • 64 PEs • Prototype due date 1972
  • 9. ILLIAC IV Processing Element (PE) • 64-bit numbers, float or fixed point • Multiples of smaller numbers that add up to 64-bits • Today’s multimedia extensions • PMEM: One local memory module per PE • 2K x 64-bits • 188ns access / 350ns cycle (includes conflict resolution) • 100K components per PE
  • 10. PE Contd. • PE mode: Active or Inactive, CU sets mode • All PEs operate in lock-step • Routing insts to move data from PE to PE • The CU can execute instructions while PE’s are busy • Another degree of concurrency • Datatypes • 64b float • 64b logical • 48b fixed • 32 float • 24 fixed • 8 fixed
  • 11. Peak Compute Bandwidth • 64 PEs • Each can perform: • 1 64b, 2 32b, or 4 8b operations • Or, in total: • 64 elems, 128 elems, or 512 elems • Peak: • 150M 64b ops/sec up to 10G 32b ops/sec • The last figure is for integer ops • Each int op takes 66ns (4 per PE in parallel)
  • 12. Control Unit (CU) • A simple CPU • Can execute instructions w/o PE intervention • Coordinates all PEs • 64 64b registers, D0-D63 • 4 64b Accumulators A0-A3 • Ops: • Integer ops • Shifts • Boolean • Loop control • Index PMem D0 D63 A0 A3 A1 A2 ALU CU
  • 13. Processing Element (PE)• 64 bit regs • A: Accumulator • B: 2nd operand for binary ops • R: Routing – Inter-PE Communication • S: Temporary • X: Index for PMEM 16bits • D: mode 8bits • Communication: • PMEM only from local PE • Amongst PE with R A S B R ALU PEi X D 0 1 2043 PMEMi PEi-1 PEi+1 PEi-8 PEi+8
  • 14. Datapaths • CU Bus: Insts and Data from PMEM to CU in 8 words • CDB: Broadcast to all PEs • E.g., constants for adds • Routing Network: amongst R registers • Mode: To activate/de-activate PEs CU PE PMEM PE PMEM PE PMEM Control Unit Bus Mode Common Data Bus Routing
  • 15. Routing Network 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 0 8 16 24 32 40 48 56 7 15 23 31 39 47 55 63 56 57 58 59 60 61 62 63 0 1 2 3 4 5 6 7 12 19 20 21 28 i-8 i+8 i+1i-1
  • 16. Using ILLIAC IV: Example #2 • DO 10 I = 1 TO 64 10 C(I) = A(I) + B(I) • LDA a + 2 load A(i) into A (same a per PMEM) • ADDRN a + 1 add B(i) into A • STA a store A into C(i) C(1) A(1) B(1) PMEM1 a C(2) A(2) B(2) PMEM2 C(64) A(64) B(64) PMEM64
  • 17. Using ILLIAC IV: Example #2 • DO 10 I = 2 TO 64 • 10 A(I) = B(I) + A(I-1) • Expand into: • A(N) = A(1) + Sum B(i) [i = 2 to N] • We get: • DO 10 N=2 TO 64 • S = S + B(N) • 10 A(N) = S
  • 18. Using ILLIAC IV: Example #2 contd. 1. Enable all PEs 2. All load A from a 3. i = 0 4. All R = A (including those inactive) 5. All route R to PE(2^i) to the right 6. j = 2^i – 1 7. Disable all PEs 1 through j 8. A = A + R  R contains a partial sum of many A(i) 9. i = i + 1 10. if i < lg(64) goto 4 11. Enable All PEs 12. All store A at (a + 1)
  • 19. Using ILLIAC IV: Example #2 contd. • Initial State: • PMEM(1)[a] = A(1) • PMEM(1+i)[a] = B(i+1) • For example, at PE1 • STEP 1: A = A(1) • from PE2 we get B(2) • STEP 2: A = A(1) + B(2) • from PE4 we get B(4) + B(5) • STEP 3: A = A(1) + B(2) + B(4) + B(5) • From PE8 we get B(8) + B(7) + B(12) + B(13)
  • 21. Vector Processors • Vector Datatype • Apply same operation on all elements of the vector • No dependences amongst elements • Same motivation as SIMD
  • 22. Properties of Vector Processors • One Vector instruction implies lots of work • Fewer instructions • Each result independent of previous result • Multiple operations in parallel • Simpler design; no need for dependence checks • Higher clock rate • Compiler must help • Fewer Branches • Memory access pattern per vector inst known • Prefetching effect • Amortize mem latency • Can exploit high-bandwidth mem system • Less/no need for data caches
  • 23. Classes of Vector Processors • Memory to memory • Vectors are in memory • Load/store • Vectors are in registers • Load/store to communicate with memory • This prevailed
  • 24. Historical Perspective • Mid-60s: performance concerns • SIMD processor arrays • Also fast Scalar machines • CDC 6600 • Texas Instruments ASC, 1972 • Memory to memory vector • Cray Develops CRAY-1, 1978
  • 25. CRAY-1 • Fast and simple scalar processor • 80 Mhz • Vector register concept • Much simple ISA • Reduced memory pressure • Tight integration of scalar and vector units • Cylindrical design to minimize wire lengths • Freon Cooling
  • 27. Components of Vector Processor • Scalar CPU: registers, datapaths, instruction fetch • Vector Registers: • Fixed length memory bank holding a single vector reg • Typically 8-32 Vregs, up to 8Kbits per Vreg • At least; 2 Read, 1 Write ports • Can be viewed as an array of N elements • Vector Functional Units: • Fully pipelined. New op per cycle • Typically 2 to 8 FUs: integer and FP • Multiple datapaths to process multiple elements per cycle if needed • Vector Load/Store Units (LSUs): • Fully pipelined • Multiple elems fetched/store per cycle • May have multiple LSUs • Cross-bar: • Connects FUS, LSUs and registers
  • 28. CRAY-1 Organization • Simple 16-bit Reg-to-Reg ISA • Use two 16-bit to get Imm • Natural combinations of scalar and vector • Scalar bit-vectors match vector length • Gather/Scatter M-R • Cond. Merge
  • 29. CRAY-1 CPU • Scalar and vector modes • 12.5 ns clock • 64-bit words • Int & FP units • 12 FUs • 8 24-bit A regs • 64 B regs (temp storage for A) • 8 64-bit S regs • 64 T regs (temp storage for S) • 64 64-elem, 64bit elem V regs
  • 30. CRAY-1 CPU • Vector Length Register • Can use only a prefix of a vreg • Vector Mask Register • Can use only a subset of a vreg • Real Time Register (counts clock cycles) • Four instruction buffers • 64 16-bit parcels • 128 Basic Instructions • Interrupt Control • NO virtual memory system
  • 31. Cray-1 Memory System • 1M 64b words + 8 check bits (single error correction, double error detection) • 16 banks of 64K words • 4 clocks period • 1 word per clock for B, T and Vreg • 1 word per 2 clocks for A & S • 4 words per clock for inst buffers
  • 32. Instruction Format • Fields g h I j k m • Bits 0-3 4-6 7-9 10-12 13-15 16-31 • Bits cnts 4 3 3 3 3 16 • X X  opcode • Rd Rs1 Rs2 • A/S B/T
  • 33. Basic Vector Instructions • Inst Operands Operation Comment • VADD.VV V1, V2, V3 V1=V2+V3 vector+vector • VADD.SV V1, R0, V2 V1=R0+V2 scalar+vector • VMUL.VVV1, V2, V3 V1=V2*V3 vector * vector • VMUL.SV V1, R0, V2 V1=R0*V2 scalar * vector • VLD V1, R0 V1=M[R0…R0+63] stride = 1 • VLDS V1, R1, R2 V1=M[R1…R1+63*R2] stride=R2 • VLDX V1, R1, V2 V1=M[R1+V2[i], i=0 to 63] gather • VST store equiv of VLD • VSTS store equiv of VLDS • VSTX V1, R1 M[R1+V2[i], i=0 to 63]=V1 scatter
  • 34. Vector Memory Operations • Load/Store move groups of data between memory and registers • Addressing Modes • Unit-stride: Fastest • Non-Unit, constant stride (interleaved memory helps • Indexed (gather-scatter) • Vector equiv of register indirect • Sparse arrays • Can vectorize more loops
  • 35. Vector Code Example • Y[0:63] = Y[0:63] + a * X[0:63] • LD R0, a • VLD V1, Rx Load X[] in V1 • VLD V2, Ry Load Y[] in V2 • VMUL.SV V3, R0, V1 V3 = X[]*a • VADD.VV V4, V2, V3 V4 = Y[]+X[]*a • VST Ry, V4 store in Y[]
  • 36. Scalar Equivalent • LD R0, a • LI R5, 512 (offset at the end of X[]) • Loop: LD R2, 0(Rx) • MULTD R2, R0, R2 • LD R3, 0(Ry) • ADD R4, R2, R3 • ST R4, 0(Ry) • ADD Rx, Rx, 8 • ADD Ry, Ry, 8 • SUB R5, R5, 8 • BNE Loop LD R0, a VLD V1, Rx VLD V2, Ry VMUL.SV V3, R0, V1 VADD.VV V4, V2, V3 VST Ry, V4
  • 37. Vector Length Register • Allows us to vectorize code where the elements do not exactly fit within the vector register • What if we need a vector of just 32 elems? • Vector length register: • Operate up to this element • Can be anything from 0 to Maximum (64 in CRAY-1) • Can also be used to support runtime vector length variability
  • 38. Strip Mining • Suppose (application vector length) AVL > MVL (max vector length) • Each loop iteration handles MVL elems • Last iteration AVL MOD MVL • VL = (AVL mode MVL) • For (I=0; I<VL; I++) • Y[I] = A*X[I] + Y[I] • low = (AVL mod MVL) • VL = low • For (i=low; i < VL; i++) • Y[i] = A*X[i] + Y[i]
  • 39. Optimization #1: Chaining• Subsequent vector op can be initiated as soon as a preceding vector op it depends upon produces its first result • Example • Vadd.vv v1, v2, v3 • Vadd.sv v4, v1, R0 V1(1) V1(2) V1(3) V1(4) V1(63) time Add initiated V4(1) V4(2) V4(3) V4(4) V4(63) unchained V1(1) V1(2) V1(3) V1(4) V1(63) Add initiated V4(1) V4(2) V4(3) V4(4) V4(63) chained
  • 40. Optimization #2: Conditional Execution • Vector Mask Register • Bit vector: used as predicate • If 0 operation is not performed for the corresponding pair • VLD V1, Ra • VLD V2, Rb • VCMP.NEQ.VV VMR, V1, V2 • VSUB.VV V3, V2, V1 (VMR) • VST V3, Ra • For (i = 0; i < 64; i++) • if (A[i] != B[i]) A[i] = A[i] – B[i]
  • 41. Optimization #3: Multi-lane Implementation • Vectors are interleaved so that multiple elems can be accessed per cycle • Replicate resources • Equivalent of Superscalar • Because of no intra-vector dependences and because inter-vector dependences are aligned (elem(i) to elem(i)) no need for inter-bank communications
  • 42. Two Ways to View Vectorization • Classic Approach: Inner-loop • Think machine as having 32 vector registers with 16 elems • 1 instruction updates all elements of a vector • Vectorize single dimension array operations • A new approach: Outer-loop • Think of machine as 16 “virtual processors” each with 32 scalar registers • 1 instruction updates register in 16 VPs • Good for irregular kernels • Hardware is the same for both • These describe the compiler’s perspective
  • 46. Multimedia ISA Extensions • Intel’s MMX • The Basics • Instruction Set • Examples • Integration into Pentium • Relationship to vector ISAs • AMD’s 3DNow! • Intel’s ISSE (a.k.a. KNI)
  • 47. MMX: Basics • Multimedia applications are becoming popular • Are current ISAs a good match for them? • Methodology: • Consider a number of “typical” applications • Can we do better? • Cost vs. performance vs. utility tradeoffs • Net Result: Intel’s MMX • Can also be viewed as an attempt to maintain market share • If people are going to use these kind of applications we better support them
  • 48. Multimedia Applications • Most multimedia apps have lots of parallelism: • for I = here to infinity • out[I] = in_a[I] * in_b[I] • At runtime: • out[0] = in_a[0] * in_b[0] • out[1] = in_a[1] * in_b[1] • out[2] = in_a[2] * in_b[2] • out[3] = in_a[3] * in_b[3] • ….. • Also, work on short integers: • in_a[i] is 0 to 256 for example (color) • or, 0 to 64k (16-bit audio)
  • 49. Observations • 32-bit registers are wasted • only using part of them and we know • ALUs underutilized and we know • Instruction specification is inefficient • even though we know that a lot of the same operations will be performed still we have to specify each of the individually • Instruction bandwidth • Discovering Parallelism • Memory Ports? • Could read four elements of an array with one 32-bit load • Same for stores • The hardware will have a hard time discovering this • Coalescing and dependences
  • 50. MMX Contd. • Can do better than traditional ISA • new data types • new instructions • Pack data in 64-bit words • bytes • “words” (16 bits) • “double words” (32 bits) • Operate on packed data like short vectors • SIMD • First used in Livermore S-1 (> 20 years)
  • 51. MMX:Example Up to 8 operations (64bit) go in parallel  Potential improvement: 8x  In practice less but still good Besides another reason to think your machine is obsolete
  • 53. MMX: Instruction Set • 57 new instructions • Integer Arithmetic • add/sub/mul • multiply add • signed/unsigned • saturating/wraparound • Shifts • Compare (form mask) • Pack/Unpack • Move • from/to memory • from/to registers
  • 54. Arithmetic • Conventional: Wrap-around • on overflow, wrap to -1 • on underflow, wrap to MAXINT • Think of digital audio • What happens when you turn volume to the MAX? • Similar for pictures • Saturating arithmetic: • on overflow, stay at MAXINT • on underflow, stat at MININT • Two flavors: • unsigned • signed
  • 55. Operations • Mult/Add • Compares • Conversion • Interpolation/Transpose • Unpack (e.g., byte to word) • Pack (e.g., word to byte)
  • 56. Matrix Transpose 4x4 • That’s for the first two rows m33 m32 m31 m30 m13 m12 m11 m10 m23 m22 m21 m20 m03 m02 m01 m00 punpcklwd punpcklwd m31 m21 m30 m20 m11 m01 m10 m00 punpckhdq punpckldq m31 m21 m11 m01 m30 m20 m10 m00 m03 m02 m01 m00 m13 m12 m11 m10 m23 m22 m21 m20 m33 m32 m31 m30 m30 m20 m10 m00 m31 m21 m11 m01 m33 m22 m12 m02 m33 m23 m13 m03
  • 57. Examples • Image Composting • A and B images fade-in and fade-out • A * fade + B * (1 - fade), OR • (A - B) * fade + B • Image Overlay • Sprite: e.g., mouse cursor • Spite: normal colors + transparent • for i = 1 to Sprite_Length • if A[I] = clear_color then • Out_frame[I] = C[I] • else Out_frame[I] = A[I] • Matrix Transpose • Covert from row major to column major • Used in JPEG
  • 58. Chroma Keying • for (i=0; i<image_size; i++) • if (x[i] == Blue) new_image[i] =y[i] • else new_image[i] = x[i];
  • 59. Chroma Keying Code • Movq mm3, mem1 • Load eight pixels from persons’ image • Movq mm4, mem2 • Load eight pixels from the background image • Pcmpeqb mm1, mm3 • Pand mm4, mm1 • Pandn mm1, mm3 • Por mm4, mm1
  • 60. Integration into Pentium • Major issue: OS compatibility • Create new registers? • Share registers with FP • Existing OSes will save/restore • Use 64-bit datapaths • Pipe capable of 2 MMX IPC • Separate MEM and Execute stage
  • 61. “Recent” Multimedia Extensions • Intel MMX: integer arithmetic only • New algorithms -> new needs • Need for massive amounts of FP ops • Solution? MMX like ISA but for FP not only integer • Example: AMD’s 3DNow! • New data type: • 2 packed single-precision FP • 2 x 32-bits • sign + exponent + significant • New instructions • Speedup potential: 2x
  • 62. AMD’s 3DNow! • 21 new instructions • Average: motivated by MPEG • Add, Sub, Reverse Sub, Mul • Accumulate • (A1, A2) acc (B1, B2) = (B1 + B2, A1 + A2) • Comparison (create mask) • Min, Max (pairwise) • Reciprocal and SQRT, • Approximation: 1st step and other steps • Prefetch • Integer from/to FP conversion • All operate on packed FP data • sign * 2^(mantissa - 127) * exponent
  • 63. Recent Extensions Cont. • Intel’s ISSE • very similar to AMD’s 3DNow! • But has separate registers • Lessons? • Applications change over time • Careful when introducing new instructions • How useful are they? • Cost? • LEGACY: are they going to be useful in the future? • Everyone has their own Multimedia Instruction set these days • read handout
  • 64. Intel’s SSE • Multimedia/Internet? • 70 new instructions • Major Types: • SIMD-FP 128-bit wide 4 x 16 bit FP • Data movement and re-organization • Type conversion • Int to Fp and vice versa • Scalar/FP precision • State Save/Restore • New SSE registers not like MMX • Memory Streaming • Prefetch to specified hierarchy level • New Media • Absolute Diff, Rounded AVG, MIN/MAX
  • 65. Altivec (PowerPC Mmedia Ext) • 128-bit registers • 8, 16, or 32 bit data types • Scalar or single-precision FP • 162 Instructions • Saturation or Modulo arithmetic • Four operand Instructions • 3 sources, 1 target
  • 66. Altivec Design Process • Look at Mmedia Kernel • Justify new instructions • Video • 8bit int LowQ, 16-bit int HighQ • Audio • 16bit int LowQ, SP FP HighQ • Image Processing • 8bit int LowQ, 16bit Int HighQ • 3D Graphics • 16bit int LowQ, SP FP HighQ • Speech Recog. • 16bit int Low Q, Sp FP HighQ • Communications/Crypto • 8-bit or 16bit unsigned int