Skip to content

This work presents the RTL design from our paper published in the journal TCSVT: 'Parallelized RDOQ Algorithm and Fully Pipelined Hardware Architecture for AVS3 Video Coding.'

Notifications You must be signed in to change notification settings

DOUDIU/AVS3-RDOQ

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

54 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AVS3-RDOQ

Introduction

Pending...

Project Structure

The tree map of this hardware architecture is shown as below:

  • result #test vectors
  • RTL #all RTL code
    • sims_1 #simulation files
    • sources_1 #source code
    • constrs_1 #timing constraint file

Tips

The simulation file in this project uses relative paths to read the test vectors from the result file. When establishing the simulation project, pay attention to whether the path to the read file is placed correctly.

Version Record

V0.0 ,单尺寸流水
时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.1 单尺寸RDOQ流水结构 898543 49121 0 392 33.1
pre_quant 31460 31855 0 32 177
ocd 1019248 36430 0 382 33.1
lnpd 3299 3435 0 0 189.25
时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.3 单尺寸RDOQ流水结构 413446 49309 0 400
pre_quant 40216 5534 0 0 188.2
ocd 254836 40527 0 400 47.58
lnpd 2767 3187 0 0 189.25
时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.6 单尺寸RDOQ流水结构 108330 23364 0 368
pre_quant 9278 16207 0 32
ocd 94603 19085 0 288 110
lnpd 3143 2584 0 0
时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.7 单尺寸RDOQ流水结构 93994 30111 0 320 109.72
pre_quant 9481 11692 0 32 350.2
ocd 82127 14429 0 288 110
lnpd 2371 3950 0 0
v0.1 ,单尺寸流水

1.ocd优化, 优化计算式,提前计算err值

2.pre_quant优化,减少乘法器位宽

3.Reconfirm the bit width of each variable.

4.Reduce the array dCurrCost_tem by half.

5.Reduce the array rdoq_est_lastby half, [2 6 12 2]->[1 6 12 2].

时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.10 单尺寸RDOQ流水结构 76846 24659 0 352 109.72(s)
pre_quant 6446 9225 0 32 350.2(v)
ocd 67397 12967 0 320 64.11(v)
lnpd 2989 2427 0 0 189.25(v)

1.Replace the multiplication operation with addition and shift operations in "temp_coef_abs_d2" for values between 2 and 9.

2.Reduce the bit width of "i64Delta" to reduce the number of DSPs used in the calculation of "err" by half.

时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.12 单尺寸RDOQ流水结构 79711 24259 0 224 126.85(s)
pre_quant 6374 9879 0 32 350.2(v)
ocd 69714 14757 0 192
lnpd 2985 2427 0 0 189.25(v)
v1.0 ,多并行流水
时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.16 多并行RDOQ流水结构 78330 21965 0 224 113.57(s)
pre_quant 6523 7349 0 32 335.6(s)
ocd 76350 15857 0 192 113.6(s)
lnpd 2854 2538 0 0 394.1(s)
v1.1,多并行流水

Delete unnecessary delay operations in the PQ module because there are many unused signals passing through this module. (The deleted signal was provided by the testbench file to the next stage OCD.)

时间 备注 LUT FF RAM(bits) DSP Fmax(MHz)
2023.4.21 多并行RDOQ流水结构 72017 23440 0 224 116.27(s)
pre_quant 6115 3289 0 32 335.6(s)
ocd 63559 16644 0 192 113.6(s)
lnpd 2338 3476 0 0 394.1(s)

About

This work presents the RTL design from our paper published in the journal TCSVT: 'Parallelized RDOQ Algorithm and Fully Pipelined Hardware Architecture for AVS3 Video Coding.'

Topics

Resources

Stars

Watchers

Forks

Packages

No packages published