rezarezvan.com
Toggle menu
Blog Lecture Notes Research Thoughts

Part 9 - Summary

EDA333 Date: May 25, 2023 5 min read

Performance and MIPS

CPU Time for a program:

CPU Time=InstructionsProgram⋅ Clock cyclesInstruction⋅ SecondsClock cycle\text{CPU Time} = \frac{\text{Instructions}}{\text{Program}} \cdot\ \frac{\text{Clock cycles}}{\text{Instruction}} \cdot\ \frac{\text{Seconds}}{\text{Clock cycle}} CPU Time=IC⋅ CPI⋅ Tc\text{CPU Time} = IC \cdot\ CPI \cdot\ T_c

Geometric mean:

∏i=1nExecution time ratioin\sqrt[n]{\prod_{i = 1}^{n} \text{Execution time ratio}_i}

MIPS Cheat Sheet:

Register convention:

$zero - Always 0

$v0-1 - Result registers

$a0-3 - Argument registers

$t0-7 - Temporary registers

$s0-7 - Content registers, save for later use

$sp - Stack pointer

$ra - Return address

Arithmetic:

add     $t0, $t0, $t1   #t0 = t0 + t1
addi    $t0, $t0, 5     #t0 = t0 + 5
addu    $t0, $t0, $t1   #t0 = t0 + t1 (unsigned)
addiu   $t0, $t0, 5     #t0 = t0 + 5  (unsigned)




sub     $t0, $t0, $t1   #t0 = t0 - t1
subi    $t0, $t0, 5     #t0 = t0 - 5
subu    $t0, $t0, $t1   #t0 = t0 - t1 (unsigned)
subiu   $t0, $t0, 5     #t0 = t0 - 5  (unsigned)

Memory:

lw $s0, 0($a0)  # s0 = MEM[a0] (word)
sw $s0, 0($a0)  # MEM[a0] = $s0 (word)




l.d $s0, 0($a0)  # s0, s1 = MEM[a0] (Double word)
s.d $s0, 0($a0)  # MEM[a0] = $s0, $s1 (Double word)

Logical:

sll $t0, $t1, 2     # t0 := t1 * 4
srl $t0, $t1, 2     # t0 := t1 / 4


and $t0, $t1, $t2   # t0 := t1 & t2
andi $t0, $t1, 2   # t0 := t1 & 2


or $t0, $t1, $t2   # t0 := t1 | t2
ori $t0, $t1, 2   # t0 := t1 | 2


nor $t0, $t1, $zero # := ~(t1)

Conditionals:

beq rs, rt, L1 # if(rs == rt); Jump to L1
bne rs, rt, L1 # if(rs != rt); Jump to L1
j L1           # Jump to L1

Remember:

  • Incrementing with 1 and not 4!
  • MIPS uses byte addressing

Pipeline

  1. Fetch Instruction. PC → Instruction memory.

  2. Decode the instruction and read from registers.

  3. Execute the instruction.

    • Arithmetic/logical computation.
    • Computation of effective memory address.
    • Computation of jump address/conditional address.

  4. Read / Write from / to memory for load/store instructions.

  5. Write back the result into the result register (RF).

Speedup=Tc∥Non-pipelined versionTc∥Pipelined version≈Number of pipeline stages\textbf{Speedup} = \frac{T_{c \quad \| \quad \text{Non-pipelined version}}}{T_{c \quad \| \quad \text{Pipelined version}}} \approx \text{Number of pipeline stages}
  • Structural hazards
    • A stage is currently busy doing an operation
  • Data hazards
    • An instruction that depends on a earlier instruction.
  • Control hazards
    • The condition and potential address of a jump has not yet by the instruction fetch.

Caches

hx=Hit rate=Percentage of hits in memory x(1−hx)=mx=Miss rate=Percentage of misses in memory xh_x = \textbf{Hit rate} = \text{Percentage of hits in memory x} \newline (1 - h_x) = m_x = \textbf{Miss rate} = \text{Percentage of misses in memory x} Tx=Hit time=Acces time for memory x(T2−T1)=Miss penalty=Miss penalty for a miss in memory 1T_x = \textbf{Hit time} = \text{Acces time for memory x} \newline (T_2 - T_1) = \textbf{Miss penalty} = \text{Miss penalty for a miss in memory 1}

Average Memory Access Time (AMAT):

AMAT=h1 T1+m1 T2=h1 T1+(1−h1)T2\text{AMAT} = h_1\ T_1 + m_1\ T_2 = h_1\ T_1 + (1 - h_1)T_2 Memory sizeCache size=N\frac{\text{Memory size}}{\text{Cache size}} = N \newline Tag bits=log2(N)\text{Tag bits} = log_2(N)

Prep for exam

  • MIPS
    • Calculate the amount of access calls to data cache (read and writes) and instruction cache.
  • Pipleline
    • TODO: Understand each stage in depth
    • Learn how to do a pipeline diagram
    • Iterate and go over optimizations and how to avoid stalls.
  • Performance calculations
    • CPI=CPIbase+CPImissCPI = CPI_{base} + CPI_{miss}
    • CPImiss=CPID miss+CPII missCPI_{miss} = CPI_{D\ miss} + CPI_{I\ miss}
  • Memory and cache
    • Iterate and go over all the different types, and how to calculate size/amount of bits.
    • Go over how to calculate how long time/amount of time it takes for certain misses, with certain sizes etc.

Notes

Page table size = Number of virtual pages ×\times Size of each page table entry

Number of virtual pages = Total virtual address space / Page size Number of virtual pages = (Number of processes ×\times Size of each virtual address space) / Page size

Cache Offset: If block size is B bytes, offset = log2(B).

Cache Index: Given by how many blocks/sets there are.

If direct mapped, meaning 1 block per set:

Number of blocks = Total Cache Size / Sizer per block

Number of sets = (Total Cache Size / Sizer per block) / Associativity level

Cache Index = log2(Number of blocks/sets)

Cache Tag: If A is the total amount of address bits, tag = A - offset - index.

A = tag + offset + index

TLB Index: If there are E pages, index = log2(E).

TLB Tag: Tag is the rest of bits. So, if we have V virtual address bits, tag = V - index.

Calculate the size of each block entry. = Number of status bits (valid, dirt etc). + Tag bits + Data bits (Usually block size (in bytes) * 8).

Calculate the size of the cache. = Number of status bits (valid, dirt etc). + Tag bits + Data bits (Usually block size (in bytes) * 8).