4.5.14. 3D GPU Performance Test

4.5.14.1. Test Principles

3D GPU performance testing is primarily used to evaluate GPU performance under various computing and rendering tasks. Depending on different GPU application scenarios, common tests include:

OpenGLES (rendering performance test): Use glmark2 for testing, evaluating GPU graphics rendering capabilities such as geometry processing, texture filling, and shader performance.
OpenCL (computing performance test): Use clpeak for testing, evaluating GPU performance in general-purpose computing (GPGPU) tasks such as floating-point computation, memory bandwidth, and integer computation.

4.5.14.2. Test Preparation

Executable test programs are included by default in the BSP image. You can cd to the following path to check:

cd /app/platform_samples/chip_base_test/11_gpu_3d_test/
ls clpeak
clpeak  Readme.md source-code-patch

ls glmark2
bin  data  Readme.md  run.sh  source-code-patch

For custom images, copy the directory from the BSP source code path /app/platform_samples/chip_base_test/11_gpu_3d_test to the device.

4.5.14.3. Test Methods

Rendering Performance Test Method

Connect a display to observe the 3D GPU rendering output:
- The program defaults to HDMI display.
- To switch to DSI, modify the command in the run.sh script to: ./bin/glmark2 --data-path ./data -c DSI
Execute run.sh
```
cd glmark2
./run.sh
```
Output results as follows:

Warning: DRM_CAP_ASYNC_PAGE_FLIP not supported, falling back to 'mailbox' mode for SwapInterval(0).
=======================================================
    glmark2 2023.01
=======================================================
    OpenGL Information
    GL_VENDOR:      Vivante Corporation
    GL_RENDERER:    Vivante GC8000L
    GL_VERSION:     OpenGL ES 3.1 V6.4.14.9.674707
    Surface Config: buf=32 r=8 g=8 b=8 a=8 depth=24 stencil=0 samples=0
    Surface Size:   1920x1080 fullscreen
=======================================================
[build] use-vbo=false: FPS: 105 FrameTime: 9.588 ms
[build] use-vbo=true: FPS: 121 FrameTime: 8.324 ms
[texture] texture-filter=nearest: FPS: 121 FrameTime: 8.326 ms
[texture] texture-filter=linear: FPS: 121 FrameTime: 8.332 ms
[texture] texture-filter=mipmap: FPS: 121 FrameTime: 8.332 ms
[shading] shading=gouraud: FPS: 121 FrameTime: 8.325 ms
[shading] shading=blinn-phong-inf: FPS: 121 FrameTime: 8.333 ms
[shading] shading=phong: FPS: 120 FrameTime: 8.334 ms
[shading] shading=cel: FPS: 121 FrameTime: 8.332 ms
[bump] bump-render=high-poly: FPS: 121 FrameTime: 8.332 ms
[bump] bump-render=normals: FPS: 120 FrameTime: 8.334 ms
[bump] bump-render=height: FPS: 121 FrameTime: 8.332 ms
[effect2d] kernel=0,1,0;1,-4,1;0,1,0;: FPS: 91 FrameTime: 11.097 ms
[effect2d] kernel=1,1,1,1,1;1,1,1,1,1;1,1,1,1,1;: FPS: 52 FrameTime: 19.438 ms
[pulsar] light=false:quads=5:texture=false: FPS: 120 FrameTime: 8.385 ms
[desktop] blur-radius=5:effect=blur:passes=1:separable=true:windows=4: FPS: 44 FrameTime: 23.077 ms
[desktop] effect=shadow:windows=4: FPS: 56 FrameTime: 18.121 ms
[buffer] columns=200:interleave=false:update-dispersion=0.9:update-fraction=0.5:update-method=map: FPS: 60 FrameTime: 16.675 ms
[buffer] columns=200:interleave=false:update-dispersion=0.9:update-fraction=0.5:update-method=subdata: FPS: 60 FrameTime: 16.670 ms
[buffer] columns=200:interleave=true:update-dispersion=0.9:update-fraction=0.5:update-method=map: FPS: 90 FrameTime: 11.113 ms
[ideas] speed=duration: FPS: 73 FrameTime: 13.862 ms
[jellyfish] <default>: FPS: 91 FrameTime: 11.066 ms
[terrain] <default>: FPS: 10 FrameTime: 106.800 ms
[shadow] <default>: FPS: 120 FrameTime: 8.337 ms
[refract] <default>: FPS: 27 FrameTime: 37.329 ms
[conditionals] fragment-steps=0:vertex-steps=0: FPS: 121 FrameTime: 8.329 ms
[conditionals] fragment-steps=5:vertex-steps=0: FPS: 90 FrameTime: 11.122 ms
[conditionals] fragment-steps=0:vertex-steps=5: FPS: 120 FrameTime: 8.343 ms
[function] fragment-complexity=low:fragment-steps=5: FPS: 120 FrameTime: 8.336 ms
[function] fragment-complexity=medium:fragment-steps=5: FPS: 105 FrameTime: 9.603 ms
[loop] fragment-loop=false:fragment-steps=5:vertex-steps=5: FPS: 120 FrameTime: 8.337 ms
[loop] fragment-steps=5:fragment-uniform=false:vertex-steps=5: FPS: 120 FrameTime: 8.335 ms
[loop] fragment-steps=5:fragment-uniform=true:vertex-steps=5: FPS: 120 FrameTime: 8.345 ms
=======================================================
                                  glmark2 Score: 97
=======================================================

Computing Performance Test Method

Execute the command:
```
cd clpeak
./clpeak
```

Output results as follows:

root@buildroot:/userdata# ./clpeak

Platform: Vivante OpenCL Platform
Device: Vivante OpenCL Device GC8000L.6214.0148
	Driver version  : OpenCL 3.0 V6.4.14.9.674707 (Linux ARM64)
	Compute units   : 1
	Clock frequency : 996 MHz

	Global memory bandwidth (GBPS)
	float   : 4.15
	float2  : 7.60
	float4  : 11.27
	float8  : 10.08
	float16 : 6.33

	Single-precision compute (GFLOPS)
	float   : 7.96
	float2  : 15.86
	float4  : 31.50
	float8  : 31.20
	float16 : 30.96

	Half-precision compute (GFLOPS)
	half   : 15.90
	half2  : 31.65
	half4  : 62.74
	half8  : 62.14
	half16 : 60.41

	No double precision support! Skipped

	Integer compute (GIOPS)
	int   : 7.96
	int2  : 15.81
	int4  : 15.75
	int8  : 15.60
	int16 : 15.51

	Integer compute Fast 24bit (GIOPS)
	int   : 7.96
	int2  : 15.81
	int4  : 15.75
	int8  : 15.60
	int16 : 15.51

	Integer char (8bit) compute (GIOPS)
	char   : 15.86
	char2  : 31.31
	char4  : 61.07
	char8  : 59.73
	char16 : 58.65

	Integer short (16bit) compute (GIOPS)
	short   : 15.86
	short2  : 31.31
	short4  : 61.07
	short8  : 59.72
	short16 : 58.65

	Transfer bandwidth (GBPS)
	enqueueWriteBuffer              : 2.30
	enqueueReadBuffer               : 1.39
	enqueueWriteBuffer non-blocking : 4.40
	enqueueReadBuffer non-blocking  : 4.09
	enqueueMapBuffer(for read)      : 4260.88
		memcpy from mapped ptr        : 0.12
	enqueueUnmap(after write)       : 5390.27
		memcpy to mapped ptr          : 4.31

	Kernel launch latency : 50.17 us

4.5.14.4. Test Metrics

Rendering Performance Metrics

Higher FPS (frames per second) values in the test results indicate better GPU performance in that test scenario. Below is a line-by-line analysis of the test results:

General Rendering Tests

Test Item	FPS	Description
use-vbo=false	105	Performance without using VBO (Vertex Buffer Object)
use-vbo=true	121	Using VBO reduces CPU-GPU data transfer and improves FPS
texture-filter=nearest/linear/mipmap	121	Texture sampling methods have almost no impact on performance
shading=gouraud/phong/blinn-phong-inf/cel	120~121	Shader computational complexity has minimal impact
bump-render=high-poly/normals/height	120~121	Techniques like normal mapping have little noticeable impact

Complex Computation Tests

Test Item	FPS	Description
effect2d (convolution kernel filtering)	91 / 52	2D computational tasks such as Gaussian blur show lower performance
desktop (window blur/shadow)	44 / 56	Complex UI effects impact GPU processing capability
buffer (large-scale data update)	60 / 90	Data transfer mode affects performance
ideas (particle system)	73	High object complexity within the scene
jellyfish (jellyfish simulation)	91	Tests GPU computation and animation rendering capability
terrain (terrain rendering)	10	High terrain complexity, insufficient GPU processing power
refract (refraction)	27	Refraction computation is complex, resulting in lower performance

Computational Tests

Test Item	FPS	Description
conditionals (branching computation)	90~121	Strong GPU capability in executing conditional branching computations
function (fragment computation with increasing complexity)	105~120	Fragment computation tasks executed well
loop (loop computation)	120	High efficiency in GPU execution of loop computations

Computing Performance Metrics

Memory Bandwidth (GBPS)

float   : 4.15
float2  : 7.60
float4  : 11.27
float8  : 10.08
float16 : 6.33

Memory bandwidth first increases and then decreases with vector size, with float4 achieving the highest bandwidth performance.
This may be influenced by cache structure or memory access patterns.

Single-Precision Compute Performance (GFLOPS)

float   : 7.96
float2  : 15.86
float4  : 31.50
float8  : 31.20
float16 : 30.96

Single-precision floating-point compute performance increases with vector size but stabilizes at float8 and above.
Low compute capability may result from a limited number of GPU compute units (only 1 Compute Unit).

Half-Precision Compute Performance (GFLOPS)

half   : 15.90
half2  : 31.65
half4  : 62.74
half8  : 62.14
half16 : 60.41

Half-precision compute performance is higher than single-precision, as expected.
In applications such as AI inference, half-precision compute is typically used to improve throughput.

Integer Compute Performance (GIOPS)

int   : 7.96
int2  : 15.81
int4  : 15.75
int8  : 15.60
int16 : 15.51

Integer compute performance is similar to single-precision floating-point compute.
May be limited by SIMD compute units and cache architecture.

Transfer Bandwidth (GBPS)

enqueueWriteBuffer              : 2.30
enqueueReadBuffer               : 1.39
enqueueWriteBuffer non-blocking : 4.40
enqueueReadBuffer non-blocking  : 4.09
enqueueMapBuffer(for read)      : 4260.88
  memcpy from mapped ptr        : 0.12
enqueueUnmap(after write)       : 5390.27
  memcpy to mapped ptr          : 4.31

Low memory access efficiency (enqueueWriteBuffer only 2.30 GBPS).
Non-blocking transfers show significant improvement over blocking transfers; data transfer strategy optimization is recommended.
Map/Unmap mechanism is extremely fast, but memcpy is affected by CPU performance, resulting in lower actual throughput.

Kernel Launch Latency

Kernel launch latency : 50.17 us

High kernel launch latency, possibly due to driver scheduling or OpenCL runtime overhead.
Could create performance bottlenecks in real-time computing applications (e.g., computer vision); reducing kernel call frequency is recommended.

4.5.14.5. Common Issues

Rendering performance test fails to execute

Issue description: Error log as follows

glmark2# ./run.sh
start insmod display module related drivers ...
Opened DRM module 'vs-drm'
Error: Failed to find a suitable connector
Error: main: Could not initialize canvas
./run.sh: line 13:  1482 Segmentation fault      (core dumped) ./bin/glmark2 --data-path ./data

Problem analysis: No suitable display connector was found.
Solution: Connect an HDMI or DSI display.