1 of 5

(B) K=0

…

…

(C) K=1

…

…

…

…

(E) Cin=0Wf =0

…

…

…

…

…

…

MAC

MAC

MAC

MAC

MAC

MAC

(F) Cin=0,Wf =1

(G) Cin=0,Wf =2

(H) Cin=1,Wf =0

…

…

MAC

(F) Cin=1,Wf =1

…

…

MAC

(G) Cin=1,Wf =2

GEMM

mGEMM

2 of 5

Cin

co

wf

co

wo

Filter

Input

Output

…

…

Cin

…

…

…

…

…

…

(B)

(C)

(D)

(A)

co

co

Filter

Input

Output

…

…

Cin

(A) GEMM

(B) mGEMM

co

(a) Im2col + GEMM

(b) GEMM Plus

wf

Input

Filter

Output

n

…

wo + wf - 1

Input

Filter

Output

m

k

Duplication & Shift

k

…

…

…

…

…

ci

ci

ci

ci

Duplication & Shift

ci

co

…

…

ci

(b) k=0

Im2col + GEMM

mGEMM

co

wo + wf - 1

…

(c) k=1

(d) k=2

(e) k=3

(a)

Input

Output

(g) ci=0, wf=0

(h) ci=0, wf=1

(i) ci=0, wf=2

(j) ci=1, wf=0

(f)

wf

ci

…

ci

Filter

MAC

MAC

MAC

m

n

k

k

wf

wf

MAC

MAC

MAC

MAC

MAC

Duplicated

& Shifted

3 of 5

눈금자, 눈금선을 .33” 에 맞추면 그리기 수월합니다

M

N

K

Output (C)

Filter (A)

Max Flop/s

Ours

Operational Intensity (Flop/Byte)

Throughput (Flop/s)

XNNPACK

OpenBlas

ARMNN

Cout

Wfil

Cin

Cin

Hfil

Hout

Cout

…

Output

Hin

Win

Wout

…

Input

Filter

Filter

Main

Memory

L2

L1

Register

Wfil

Hfil

Cin

Cout

Wout

Hout

Cout

Cin

Win

Hin

…

Wout Iteration

Input

Output

Filter

K

Input (B)

N Iteration

i,k

i,j

k,j

4 of 5

눈금자, 눈금선을 .33” 에 맞추면 그리기 수월합니다

ci

Input

Output

ci

Memory

Shared Last Level Cache

L1 Cache

(CPU0)

Hout xWout

Streamed from memory

Filter

​

Streamed into memory

Input

5 of 5

눈금자, 눈금선을 .33” 에 맞추면 그리기 수월합니다

Cout

B x Hout xWout

Output (C)

Filter (A)

Input (B)

Cin x Hfil xWfil

Cout x Hfil xWfil

B x Hout xWout

Output (C)

Filter (A)

Input (B)

Cin

Cin

kn2col

im2col

Cout

B x Hout xWout

Cin

Cin

Hfil xWfil

Cin x Hfil xWfil

Filter (A)

Input (B)

Output (C)

Cout (M)

B x Hout xWout (N)

Cin x Hfil xWfil (K)

Output

Input

Cin x Hfil xWfil (K)

Filter

Wfil

Hfil

Cin

Cout (M)

B x Hout xWout (N)

Filter

Cin x Hfil xWfil (K)

Wfil

Hfil

Cin

Output

Input

(A)

(B)

(C)

Wout

Cin x Hfil xWfil

Cin x Hfil xWfil

Cin x Hfil xWfil (K)