CHIPTOP - GPU Verification Checklist

Version 01, Jun.22, 2020

<Revision history>

Date Content of revision Version Creator

RVC/NhanLe
Jun. 22, 2020 New creation 01
Jun. 22, 2020

<Reference document>

No. Title name Version

1 RZG2L_GPU_design_spec_rev0.00_20200325.xlsx 0.00
2 RZG2L_ChipTop_Verification_Policy_v02.xlsx 02
3 RZG2L_GPU_SS_UT_Verification_Checklist_v04.xlsx 04
ARM VALIDATION CHECK
Created by: RVC/NhanLe

Target GPU
1st category
No. No.
1 GPU Function Check 1

2
3
4

6
7

8
9

10

11

12

13

14

15

16

17

18
19
 GPU
2nd category
(Test case name) No.
apb_gpu_power 1

hw_hard_reset 2
register_reset 3
register_rw 4

simple_blending_test 5

simple_bus 6
simple_compute 7

simple_shader 8
sw_hard_reset 9

sw_soft_reset 10

avg_power 11

di_dt 12

l2cache_flush_128k 13

l2cache_flush_256k 14

l2cache_flush_512k 15

l2cache_flush_64k 16

max_power 17

min_power 18
sustained_max_power 19
 3rd category
Judgment TM Name
(Test case explanation)
Power up/down shader core/core group through
control registers then check corresponding status
registers
Check all register reset values after HW reset
Check all register reset values after HW reset
Write all writable register values and read back to
check results
This test is called angry_birds_style_blending_test. It
runs a simple fragment shading test with varying
/texturing/alpha test/blending
Simple test to see if the bus works properly
Simple compute job to test basic integer operations
and load/store
A simple shader test with vertex shader
Send hard_reset GPU command and check the reset
value of the register
Send sw_reset GPU command and check the reset
value of the register
Create a small workload whose power consumption is
similar to Angry Birds
Use power override register to create architectural di/dt
vectors
Create a compute job to load the initial data into 128K
L2 cache lines, change the values, then flush them
back to the memory
Create a compute job to load the initial data into 256K
L2 cache lines, change the values, then flush them
back to the memory
Create a compute job to load the initial data into 512K
L2 cache lines, change the values, then flush them
back to the memory
Create a compute job to load the initial data into 64K
L2 cache lines, change the values, then flush them
back to the memory
Create a small workload that can trigger as many
modules as possible to achieve the highest power
consumption
Power up the GPU and do nothing to profile the
minimum power consumption
Create a workload whose power consumption is high
but sustainable
Judged by Confirmation Method Remark