Advanced Architecture
Computer Organization and Assembly Languages Yung-Yu Chuang
with slides by S. Dandamudi, Peng-Sheng Chen, Kip Irvine, Robert Sedgwick and Kevin Wayne
Intel microprocessor history
3
Early Intel microprocessors
• Intel 8080 (1972) – 64K addressable RAM – 8-bit registers
– CP/M operating system – 5,6,8,10 MHz
– 29K transistors
• Intel 8086/8088 (1978) – IBM-PC used 8088 – 1 MB addressable RAM – 16-bit registers
– 16-bit data bus (8-bit for 8088) – separate floating-point unit (8087) – used in low-cost microcontrollers now
my first computer (1986)
4
The IBM-AT
• Intel 80286 (1982) – 16 MB addressable RAM – Protected memory
– several times faster than 8086 – introduced IDE bus architecture – 80287 floating point unit
– Up to 20MHz – 134K transistors
5
Intel IA-32 Family
• Intel386 (1985)
– 4 GB addressable RAM – 32-bit registers
– paging (virtual memory) – Up to 33MHz
• Intel486 (1989)
– instruction pipelining – Integrated FPU – 8K cache
• Pentium (1993)
– Superscalar (two parallel pipelines)
6
Intel P6 Family
• Pentium Pro (1995)
– advanced optimization techniques in microcode – More pipeline stages
– On-board L2 cache
• Pentium II (1997)
– MMX (multimedia) instruction set – Up to 450MHz
• Pentium III (1999)
– SIMD (streaming extensions) instructions (SSE) – Up to 1+GHz
• Pentium 4 (2000)
– NetBurst micro-architecture, tuned for multimedia – 3.8+GHz
• Pentium D (2005, Dual core)
IA32 Processors
• Totally Dominate Computer Market
• Evolutionary Design
– Starting in 1978 with 8086
– Added more features as time goes on
– Still support old features, although obsolete
• Complex Instruction Set Computer (CISC)
– Many different instructions with many different formats
• But, only small subset encountered with Linux programs
– Hard to match performance of Reduced Instruction Set Computers (RISC)
– But, Intel has done just that!
ARM history
• 1983 developed by Acorn computers
– To replace 6502 in BBC computers – 4-man VLSI design team
– Its simplicity comes from the inexperience team – Match the needs for generalized SoC for reasonable
power, performance and die size
– The first commercial RISC implemenation
• 1990 ARM (Advanced RISC Machine), owned by Acorn, Apple and VLSI
ARM Ltd
Design and license ARM core design but not fabricate
Why ARM?
• One of the most licensed and thus widespread processor cores in the world
– Used in PDA, cell phones, multimedia players, handheld game console, digital TV and cameras – ARM7: GBA, iPod
– ARM9: NDS, PSP, Sony Ericsson, BenQ – ARM11: Apple iPhone, Nokia N93, N800
– 90% of 32-bit embedded RISC processors till 2009
• Used especially in portable devices due to its low power consumption and reasonable
performance
ARM powered products
12
Performance boost
• Increasing clock rate is insufficient. Architecture (pipeline/cache/SIMD) becomes more significant.
In his 1965 paper, Intel co-founder Gordon Moore observed that
“the number of transistors per square inch had doubled every 18 months.
Basic architecture
Basic microcomputer design
• clock synchronizes CPU operations
• control unit (CU) coordinates sequence of execution steps
• ALU performs arithmetic and logic operations
Central Processor Unit (CPU)
Memory Storage Unit registers
ALU clock
I/O Device
#1
I/O Device
#2 data bus
control bus address bus
CU
Basic microcomputer design
• The memory storage unit holds instructions and data for a running program
• A bus is a group of wires that transfer data from one part to another (data, address, control)
Central Processor Unit (CPU)
Memory Storage Unit registers
ALU clock
I/O Device
#1
I/O Device
#2 data bus
control bus address bus
CU
Clock
• synchronizes all CPU and BUS operations
• machine (clock) cycle measures time of a single operation
• clock is used to trigger events
one cycle 1
0
• Basic unit of time, 1GHz→clock cycle=1ns
• An instruction could take multiple cycles to complete, e.g. multiply in 8088 takes 50 cycles
Instruction execution cycle
• Fetch
• Decode
• Fetch operands
• Execute
• Store output
I-1 I-2 I-3 I-4
PC program
I-1 instruction register op1
op2
memory fetch
ALU registers
write decode
execute read
write
(output)
registers
flags
program counter
instruction queue
Pipeline
Multi-stage pipeline
• Pipelining makes it possible for processor to execute instructions in parallel
• Instruction execution divided into discrete stages
S1 S2 S3 S4 S5
1
Cycles
Stages
S6
2 3 4 5 6 7 8 9 10 11 12
I-1
I-2 I-1
I-2 I-1
I-2 I-1
I-2 I-1
I-2 I-1
I-2
Example of a non- pipelined processor.
For example, 80386.
Many wasted cycles.
Pipelined execution
• More efficient use of cycles, greater throughput of instructions: (80486 started to use pipelining)
S1 S2 S3 S4 S5
1
Cycles
Stages
S6
2 3 4 5 6 7
I-1 I-2 I-1
I-2 I-1 I-2 I-1
I-2 I-1 I-2 I-1
I-2
For k stages and n instructions, the number of
required cycles is:
k + (n – 1) compared to k*n
• Pipelining requires buffers
– Each buffer holds a single value
– Ideal scenario: equal work for each stage
• Sometimes it is not possible
• Slowest stage determines the flow rate in the entire pipeline
Pipelined execution Pipelined execution
• Some reasons for unequal work stages
– A complex step cannot be subdivided conveniently – An operation takes variable amount of time to
execute, e.g. operand fetch time depends on where the operands are located
• Registers
• Cache
• Memory
– Complexity of operation depends on the type of operation
• Add: may take one cycle
• Multiply: may take several cycles
• Operand fetch of I2 takes three cycles
– Pipeline stalls for two cycles
• Caused by hazards
– Pipeline stalls reduce overall throughput
Pipelined execution Wasted cycles (pipelined)
• When one of the stages requires two or more clock cycles, clock cycles are again wasted.
S1 S2 S3 S4 S5
1
Cycles
Stages
S6
2 3 4 5 6 7
I-1 I-2 I-3
I-1 I-2 I-3
I-1 I-2 I-3
I-1
I-2 I-1 I-1 8
9
I-3 I-2 I-2 exe
10 11
I-3 I-3 I-1
I-2
I-3
For k stages and n instructions, the number of required cycles is:
k + (2n – 1)
Superscalar
A superscalar processor has multiple execution pipelines. In the following, note that Stage S4 has left and right pipelines (u and v).
S1 S2 S3 u S5
1
Cycles
Stages
S6
2 3 4 5 6 7
I-1 I-2 I-3 I-4
I-1 I-2 I-3 I-4
I-1 I-2 I-3 I-4
I-1
I-3 I-1
I-2 I-1 v
I-2
I-4 S4
8 9
I-3 I-4
I-2 I-3
10 I-4
I-2
I-4 I-1
I-3
For k states and n instructions, the number of required cycles is:
k + n
Pentium: 2 pipelines Pentium Pro: 3
Pipeline stages
• Pentium 3: 10
• Pentium 4: 20~31
• Next-generation micro-architecture: 14
• ARM7: 3
Hazards
• Three types of hazards
– Resource hazards
• Occurs when two or more instructions use the same resource, also called structural hazards
– Data hazards
• Caused by data dependencies between instructions, e.g. result produced by I1 is read by I2
– Control hazards
• Default: sequential execution suits pipelining
• Altering control flow (e.g., branching) causes problems, introducing control dependencies
Data hazards
add r1, r2, #10 ; write r1 sub r3, r1, #20 ; read r1
fetch decode reg ALU wb
fetch decode stall reg ALU wb
Data hazards
• Forwarding: provides output result as soon as possible
fetch decode reg ALU wb
fetch decode stall reg ALU
add r1, r2, #10 ; write r1 sub r3, r1, #20 ; read r1
wb
Data hazards
• Forwarding: provides output result as soon as possible
fetch decode reg ALU wb
fetch decode stall reg ALU add r1, r2, #10 ; write r1
sub r3, r1, #20 ; read r1
fetch decode stall reg ALU wb
wb
Control hazards
fetch decode reg ALU wb
fetch decode reg ALU wb
fetch decode reg ALU wb fetch decode reg ALU
fetch decode reg ALU bz r1, target
add r2, r4, 0 ...
target: add r2, r3, 0
wb
Control hazards
• Braches alter control flow
– Require special attention in pipelining
– Need to throw away some instructions in the pipeline
• Depends on when we know the branch is taken
• Pipeline wastes three clock cycles – Called branch penalty
– Reducing branch penalty
• Determine branch decision early
Control hazards
• Delayed branch execution
– Effectively reduces the branch penalty
– We always fetch the instruction following the branch
• Why throw it away?
• Place a useful instruction to execute
• This is called delay slot
add R2,R3,R4 branch target sub R5,R6,R7 . . .
branch target add R2,R3,R4
sub R5,R6,R7 . . .
Delay slot
Branch prediction
• Three prediction strategies
– Fixed
• Prediction is fixed
– Example: branch-never-taken
» Not proper for loop structures
– Static
• Strategy depends on the branch type – Conditional branch: always not taken – Loop: always taken
– Dynamic
• Takes run-time history to make more accurate predictions
Branch prediction
• Static prediction
– Improves prediction accuracy over Fixed
Instruction type Instruction Distribution
(%)
Prediction:
Branch taken?
Correct prediction
(%) Unconditional
branch
70*0.4 = 28 Yes 28
Conditional branch
70*0.6 = 42 No 42*0.6 = 25.2
Loop 10 Yes 10*0.9 = 9
Call/return 20 Yes 20
Overall prediction accuracy = 82.2%
Branch prediction
• Dynamic branch prediction
– Uses runtime history
• Takes the past n branch executions of the branch type and makes the prediction
– Simple strategy
• Prediction of the next branch is the majority of the previous n branch executions
• Example: n = 3
– If two or more of the last three branches were taken, the prediction is “branch taken”
• Depending on the type of mix, we get more than 90%
prediction accuracy
Branch prediction
• Impact of past n branches on prediction accuracy
Type of mix
n Compiler Business Scientific 0 64.1 64.4 70.4 1 91.9 95.2 86.6 2 93.3 96.5 90.8 3 93.7 96.6 91.0 4 94.5 96.8 91.8 5 94.7 97.0 92.0
10 Predict branch
01 Predict no branch
Branch prediction
00 Predict no branch
11 Predict branch branch
branch no
branch
no branch
branch
no branch no
branch branch
Multitasking
• OS can run multiple programs at the same time.
• Multiple threads of execution within the same program.
• Scheduler utility assigns a given amount of CPU time to each running program.
• Rapid switching of tasks
– gives illusion that all programs are running at once – the processor must support task switching
– scheduling policy, round-robin, priority
Cache
SRAM vs DRAM
Tran. Access Needs
per bit time refresh? Cost Applications SRAM 4 or 6 1X No 100X cache memories DRAM 1 10X Yes 1X Main memories,
frame buffers
Central Processor Unit (CPU)
Memory Storage Unit registers
ALU clock
I/O Device
#1 I/O Device
#2 data bus
control bus address bus
CU
The CPU-Memory gap
The gap widens between DRAM, disk, and CPU speeds.
1 10 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000
1980 1985 1990 1995 2000
year
ns
Disk seek time DRAM access time SRAM access time CPU cycle time
register cache memory disk
Access time (cycles)
1 1-10 50-100 20,000,000
Memory hierarchies
• Some fundamental and enduring properties of hardware and software:
– Fast storage technologies cost more per byte, have less capacity, and require more power (heat!).
– The gap between CPU and main memory speed is widening.
– Well-written programs tend to exhibit good locality.
• They suggest an approach for organizing memory and storage systems known as a memory hierarchy.
Memory system in practice
Larger, slower, and cheaper (per byte) storage devices
registers on-chip L1 cache (SRAM)
main memory (DRAM)
local secondary storage (virtual memory) (local disks)
remote secondary storage
(tapes, distributed file systems, Web servers) off-chip L2
cache (SRAM) L0:
L1:
L2:
L3:
L4:
L5:
Smaller, faster, and more expensive (per byte) storage devices
Reading from memory
• Multiple machine cycles are required when reading from memory, because it responds much more slowly than the CPU (e.g.33 MHz). The wasted clock cycles are called wait states.
Processor Chip L1 Data 1 cycle latency
16 KB 4-way assoc Write-through
32B lines L1 Instruction 16 KB, 4-way
32B lines
Regs. L2 Unified
128KB--2 MB 4-way assoc Write-back Write allocate
32B lines
MemoryMain Up to 4GB
Pentium III cache hierarchy
Cache memory
• High-speed expensive static RAM both inside and outside the CPU.
– Level-1 cache: inside the CPU – Level-2 cache: outside the CPU
• Cache hit: when data to be read is already in cache memory
• Cache miss: when data to be read is not in cache memory. When? compulsory, capacity and conflict.
• Cache design: cache size, n-way, block size, replacement policy
Caching in a memory hierarchy
0 1 2 3
4 5 6 7
8 9 10 11
12 13 14 15
Larger, slower, cheaper Storage device at level k+1 is partitioned into blocks.
Data is copied between levels in block-sized transfer units
8 9 14 3
Smaller, faster, more Expensive device at level k caches a subset of the blocks from level k+1 level k
level k+1
4 4
4 10
10
10
Request 14
Request 12
General caching concepts
• Program needs object d, which is stored in some block b.
• Cache hit
– Program finds b in the cache at level k. E.g., block 14.
• Cache miss
– b is not at level k, so level k cache must fetch it from level k+1.
E.g., block 12.
– If level k cache is full, then some current block must be replaced (evicted). Which one is the “victim”?
• Placement policy:where can the new block go? E.g., b mod 4
• Replacement policy:which block should be evicted? E.g., LRU
9 3
0 1 2 3
4 5 6 7
8 9 10 11
12 13 14 15
level k
level k+1
14 14
12 14
4*
124*
12
0 1 2 3
Request 12
4*4*
12
Locality
• Principle of Locality: programs tend to reuse data and instructions near those they have used recently, or that were recently referenced themselves.
– Temporal locality: recently referenced items are likely to be referenced in the near future.
– Spatial locality:items with nearby addresses tend to be referenced close together in time.
• In general, programs with good locality run faster then programs with poor locality
• Locality is the reason why cache and virtual memory are designed in architecture and operating system. Another example is web browser caches recently visited webpages.
Locality example
• Data
– Reference array elements in succession (stride-1 reference pattern):
– Reference sum each iteration:
• Instructions
– Reference instructions in sequence:
– Cycle through loop repeatedly:
sum = 0;
for (i = 0; i < n; i++) sum += a[i];
return sum;
Spatial locality
Spatial locality Temporal locality
Temporal locality
Locality example
• Being able to look at code and get a qualitative sense of its locality is important. Does this function have good locality?
int sum_array_rows(int a[M][N]) {
int i, j, sum = 0;
for (i = 0; i < M; i++) for (j = 0; j < N; j++)
sum += a[i][j];
return sum;
} stride-1 reference pattern
Locality example
• Does this function have good locality?
int sum_array_cols(int a[M][N]) {
int i, j, sum = 0;
for (j = 0; j < N; j++) for (i = 0; i < M; i++)
sum += a[i][j];
return sum;
} stride-N reference pattern
Blocked matrix multiply performance
• Blocking (bijk and bikj) improves performance by a factor of two over unblocked versions (ijk and jik)
– relatively insensitive to array size.
0 10 20 30 40 50 60
25 50 75 100 125
150 175
200 225
250 275
300 325
350 375
400 Array size (n)
Cycles/iteration
kji jki kij ikj jik ijk
bijk (bsize = 25) bikj (bsize = 25)
Cache-conscious programming
• make sure that memory is cache-aligned
• Split data into hot and cold (list example)
• Use union and bitfields to reduce size and increase locality
SIMD
56
SIMD
• MMX (Multimedia Extension) was introduced in 1996 (Pentium with MMX and Pentium II).
• Intel analyzed multimedia applications and found they share the following characteristics:
– Small native data types (8-bit pixel, 16-bit audio) – Recurring operations
– Inherent parallelism
57
SIMD
• SIMD (single instruction multiple data)
architecture performs the same operation on multiple data elements in parallel
• PADDW MM0, MM1
58
SISD/SIMD/Streaming
59
MMX
• After analyzing a lot of existing applications such as graphics, MPEG, music, speech recognition, game, image processing, they found that many multimedia algorithms
execute the same instructions on many pieces of data in a large data set.
• Typical elements are small, 8 bits for pixels, 16 bits for audio, 32 bits for graphics and general computing.
• New data type: 64-bit packed data type. Why 64 bits?
– Good enough – Practical
60
MMX data types
61
MMX instructions
• 57 MMX instructions are defined to perform the parallel operations on multiple data elements packed into 64-bit data types.
• These include add, subtract, multiply, compare, and shift, data conversion, 64-bit data move, 64-bit logical operation and multiply-add for multiply- accumulate operations.
• All instructions except for data move use MMX registers as operands.
• Most complete support for 16-bit operations.
62
Saturation arithmetic
wrap-around saturating
• Useful in graphics applications.
• When an operation overflows or underflows, the result becomes the largest or smallest possible representable number.
• Two types: signed and unsigned saturation
63
Keys to SIMD programming
• Efficient data layout
• Elimination of branches
64
Application: frame difference
A B
|A-B|
65
Application: frame difference
A-B B-A
(A-B) or (B-A)
66
Application: frame difference
MOVQ mm1, A //move 8 pixels of image A MOVQ mm2, B //move 8 pixels of image B MOVQ mm3, mm1 // mm3=A
PSUBSB mm1, mm2 // mm1=A-B PSUBSB mm2, mm3 // mm2=B-A POR mm1, mm2 // mm1=|A-B|
67
Data-independent computation
• Each operation can execute without needing to know the results of a previous operation.
• Example, sprite overlay
for i=1 to sprite_Size if sprite[i]=clr
then out_color[i]=bg[i]
else out_color[i]=sprite[i]
• How to execute data-dependent calculations on
several pixels in parallel. 68
Application: sprite overlay
69
Application: sprite overlay
MOVQ mm0, sprite MOVQ mm2, mm0 MOVQ mm4, bg MOVQ mm1, clr PCMPEQW mm0, mm1 PAND mm4, mm0 PANDN mm0, mm2 POR mm0, mm4
70
Other SIMD architectures
• Graphics Processing Unit (GPU): nVidia 7800, 24 pipelines (8 vector/16 fragment)
Impacts on programming
• You need to be aware of architecture issues to write more efficient programs (such as cache- aware).
• Need parallel thinking for better utilizing
parallel features of processors.
RISC v.s. CISC
Trade-offs of instruction sets
• Before 1980, the trend is to increase instruction complexity (one-to-one mapping if possible) to bridge the gap. Reduce fetch from memory.
Selling point: number of instructions, addressing modes. (CISC)
• 1980, RISC. Simplify and regularize instructions to introduce advanced architecture for better performance, pipeline, cache, superscalar.
high-level language machine code semantic gap
compiler C, C++
Lisp, Prolog, Haskell…
RISC
• 1980, Patternson and Ditzel (Berkeley),RISC
• Features
– Fixed-length instructions – Load-store architecture – Register file
• Organization
– Hard-wired logic
– Single-cycle instruction – Pipeline
• Pros: small die size, short development time, high performance
• Cons: low code density, not x86 compatible
RISC Design Principles
• Simple operations
– Simple instructions that can execute in one cycle
• Register-to-register operations
– Only load and store operations access memory – Rest of the operations on a register-to-register basis
• Simple addressing modes
– A few addressing modes (1 or 2)
• Large number of registers
– Needed to support register-to-register operations – Minimize the procedure call and return overhead
RISC Design Principles
• Fixed-length instructions
– Facilitates efficient instruction execution
• Simple instruction format
– Fixed boundaries for various fields
• opcode, source operands,…
CISC and RISC
• CISC – complex instruction set
– large instruction set
– high-level operations (simpler for compiler?) – requires microcode interpreter (could take a long
time)
– examples: Intel 80x86 family
• RISC – reduced instruction set
– small instruction set
– simple, atomic instructions
– directly executed by hardware very quickly
– easier to incorporate advanced architecture design – examples: ARM (Advanced RISC Machines) and DEC
Alpha (now Compaq), PowerPC, MIPS
CISC and RISC
CISC
(Intel 486) RISC (MIPS R4000)
#instructions 235 94
Addr. modes 11 1
Inst. Size (bytes) 1-12 4
GP registers 8 32
Why RISC?
• Simple instructions are preferred
– Complex instructions are mostly ignored by compilers
• Due to semantic gap
• Simple data structures
– Complex data structures are used relatively infrequently
– Better to support a few simple data types efficiently
• Synthesize complex ones
• Simple addressing modes
– Complex addressing modes lead to variable length instructions
• Lead to inefficient instruction decoding and scheduling
Why RISC? (cont’d)
• Large register set
– Efficient support for procedure calls and returns
• Patterson and Sequin’s study
– Procedure call/return: 1215% of HLL statements
» Constitute 3133% of machine language instructions
» Generate nearly half (45%) of memory references
– Small activation record
• Tanenbaum’s study
– Only 1.25% of the calls have more than 6 arguments – More than 93% have less than 6 local scalar variables – Large register set can avoid memory references