ImageCachelineSize causes memory object allocation failure on Mali-G78

[w1049]

Environment:

Device: Huawei Mate 40 Pro
SoC: Kirin9000
GPU: Mali-G78

[INFO] fetched device report
[INFO]     7GB global memory with 2MB cache consists of 64B cachelines
[INFO]     images up to [65536, 65536] texels are supported
[INFO]     24 SMs with 512 logical threads in each

ImageCachelineSize test failed with log:

[INFO] [ImageCachelineSize]
[INFO]     initialized table for aspect 'ImageCachelineSize'
[INFO]     already know that 'LogicThreadCount' from aspect 'Device' is 512
[INFO]     already know that 'MaxImageWidth' from aspect 'Device' is 65536
[INFO]     already know that 'MaxImageHeight' from aspect 'Device' is 65536
[ERROR]     application threw an exception
[ERROR]     memory object allocation failure
[ERROR]     application cannot continue
[INFO]     saved configuration to 'ArchProbe.json'
[INFO]     saved report to 'ArchProbeReport.json'
FORTIFY: pthread_mutex_lock called on a destroyed mutex (0x73f9be9850)
[ERROR]     captured SIGABRT! progress is saved

In app.cpp, Line 504-506:

    cl::ImageFormat img_fmt(CL_RGBA, CL_FLOAT);
    cl::Image2D in_img =
      env.create_img_2d(0, img_fmt, MAX_IMG_WIDTH, MAX_IMG_HEIGHT);

On my device, both MAX_IMG_WIDTH and MAX_IMG_HEIGHT are 65536. The size of in_img will be 65536*65536*4=17179869184B=16GB, which exceeds the size of global memory. I changed the parameters to 16384*16384 and the error disappeared.

[PENGUINLIONG]

This implementation is for paper result reproduction only and we might not put much effort in such engineering issue now and in future. Sorry about that.