dzhoshkun / cuda-learning Goto Github PK

• answer the questions about MAX_DEPTH and INIT_SUBDIV in the mandelbrot exercise

This is for trying out some of the warp-level primitives mentioned in the relevant CUDA blog post

Description

The following steps are a possible way to make video data available for GIFT-Grab-supported pipelines processing images on the GPU using the CIELAB color space using NVIDIA Performance Primitives (NPP):

1. Capture image data in I420 color format
2. Copy data from host to device
3. Convert I420 data to BGR format
4. Convert BGR data to CIELAB format
5. Do process data
6. Convert CIELAB data to BGR format
7. Convert BGR data to I420 format
8. Copy data from device back to host

The indirect conversions are because currently no direct conversion is supported by NPP.

Aim

The performance of this pipeline should be assessed by measuring the time it takes to perform steps 2, 3, 4, 6, 7, and 8 so as to know how much time is left for step 5 in a pre-defined frame rate configuration.

Variables

• Image resolution in step 1
• Capture time in step 1
• Used GPU
• CUDA version
• Frame rate

• Benchmark matrix multiplication on device memory vs. matrix multiplication using shared memory on:
  • TITAN X
  • GTX 980 Ti
• note down:
  • results
  • CUDA version
  • driver version

• load a Full HD image
• convert it to Lab color space
• cupy.argsort its a channel
  • copy it to device first
• numpy.argsort its a channel
• compare runtimes using e.g. timeit

Following the NVIDIA article on optimizing data transfers, when I run the bandwidthtest application, I get the following output on Adelie:

(ssh) dzhoshkun@gift-adelie ‹ master ● › : ~/ws/code-samples/series/cuda-cpp/optimize-data-transfers
[0] % CUDA_VISIBLE_DEVICES=0 ./bandwithtest 

Device: GeForce GTX TITAN X
Transfer size (MB): 16

Pageable transfers
  Host to Device bandwidth (GB/s): 5.442285
  Device to Host bandwidth (GB/s): 6.198576

Pinned transfers
  Host to Device bandwidth (GB/s): 6.114146
  Device to Host bandwidth (GB/s): 6.509175


(ssh) dzhoshkun@gift-adelie ‹ master ● › : ~/ws/code-samples/series/cuda-cpp/optimize-data-transfers
[0] % CUDA_VISIBLE_DEVICES=1 ./bandwithtest 

Device: GeForce GTX 980 Ti
Transfer size (MB): 16

Pageable transfers
  Host to Device bandwidth (GB/s): 7.783488
  Device to Host bandwidth (GB/s): 11.366924

Pinned transfers
  Host to Device bandwidth (GB/s): 12.176038
  Device to Host bandwidth (GB/s): 12.547878

Why should a TITAN X have half the bandwidth of a GTX 980 Ti?

Virtually the same code as #11, but using e.g. PyCuda.

The CUDA performance guidelines state:

comparing the floating-point operation throughput or memory throughput - whichever makes more sense - of a particular kernel to the corresponding peak theoretical throughput of the device indicates how much room for improvement there is for the kernel.

• select a number of typical kernels (for instance from typical CUDA examples)
• choose a GPU to run these on
• measure the corresponding throughput for each kernel
• compare the throughput to the corresponding theoretical peak throughput
• try to increase the throughput by changing the implementation

There is a simple example provided on the NVIDIA website.
I will tinker with it to understand the impact of the different configuration options when running GPU-accelerated applications.

Pre-requisite:

• https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/

As a continuation of #1:

• implement a streamed version of the add function
• call it with e.g. 22 streams (GTX 980 Ti) and check the speed-up against the other versions