Loran L
Loran L's Blog

Follow

Loran L's Blog

Follow
Vector registers, SIMD, SWAR,Vector Chaining. Any Difference?

Vector registers, SIMD, SWAR,Vector Chaining. Any Difference?

osd600

Loran L's photo
Loran L
·Dec 16, 2021·

3 min read

SIMD, single instruction multiple data is a very broad term. It is a concept. SIMD coud be implemented in parallel, or sequential; On one thread, or on multiple SIMT(single instruction multiple thread); On the GPU or on the CPU.

Is there some kind of magical vector register, that simultaneously can add vector 'a' to vector 'b', and store result in vector 'c'?


560px-Non-SIMD_cpu_diagram1.svg.png


560px-SIMD_cpu_diagram1.svg.png


  setvli  $10        # Set vector length VL=10
  vload   v1, a      # 10 loads from a
  vload   v2, b      # 10 loads from b
  vadd   v3, v1, v2  # 10 adds
  vstore v3, c       # 10 stores into c
  ret

Compared to regular implementation with a loop

; Hypothetical RISC machine
; add 10 numbers in a to 10 numbers in b, storing results in c
; assume a, b, and c are memory locations in their respective registers
  move  $10, count   ; count := 10
loop:
  load  r1, a
  load  r2, b
  add   r3, r1, r2   ; r3 := r1 + r2
  store r3, c
  add   a, a, $4     ; move on
  add   b, b, $4
  add   c, c, $4
  dec   count        ; decrement
  jnez  count, loop  ; loop back if count is not yet 0
  ret

source

There are SIMD instructions in most of the architectures. x64, Aarch and so on. Does it mean in code snippet one, there is no iteration done and everything is added right away in parallel? Or is iteration is done in hardware, on transistor level.

Let's investigate .

Predicated SIMD

  1. it uses masking 2.associative Processor - These receive the one (same) instruction but in each parallel processing unit an independent decision is made, based on data local to the unit, as to whether to perform the execution or whether to skip it.
  2. allows to use variable length vectors, because it has prediction capabilities, to decide if it needs to continue processing data, or ready to stop.
  3. predicate masks allow parallel if/then/else constructs without branches

Packed SIMD

  1. Also known as SWAR, simd within a register
  2. ARM NEON was inspired by this
  3. form of a group of registers and instructions
  4. instructions explicitly intended to perform parallel operations across data that is stored in the independent subwords or fields of a register

Pure Vectors

  1. characterized by pipelined functional units that accept a sequential stream of array or vector elements
  2. I think it is related to vector chaining

The way I process information above

Predicated: parallel execution, more like multithreading. Does not happen at the same physical place. Both data and instructions in different place, but parallel.

Packed: Not sure, but this one looks more like everything is happening in one place, and with true parallelism. Perhaps like on the picture above, where one 32bit register acts like four 8bit registers.

Pure Vectors: calculations happen in one place, but pipelined (in order, one by one). So here like hardware accelerated 'for' loop for future reference

 
Share this