markingramuk / ocl-compile-benchmark

Test project to demonstrate poor compilation performance in AMD OpenCL driver

Compilation times for the Addition kernel:

__kernel void Addition(__global float* input1, __global float* input2, __global float* output)  
{  
	int index = get_global_id(0);  
	output[index] = input1[index] + input2[index];  
}

AMD Radeon PRO W6800 (OpenCL 2.0 AMD-APP (3354.13)):

Run	Duration (ms)
1	467.382
2	106.683
3	106.861
4	107.879
5	107.799
6	104.983
7	107.775
8	105.926
9	107.524
10	107.95
-	-
Average	143.076

AMD Radeon RX 5700 XT (OpenCL 2.0 AMD-APP (3188.4)):

Run	Duration (ms)
1	2729.53
2	1254.58
3	1242.51
4	1278.31
5	1202.36
6	1269.07
7	1237.95
8	1267.08
9	1252
10	1268.22
-	-
Average	1400.16

AMD Radeon (TM) R9 390 Series (OpenCL 2.0 AMD-APP (3110.7)):

Run	Duration (ms)
1	282.61
2	26.6979
3	27.009
4	27.0421
5	26.917
6	28.2667
7	29.3081
8	27.0859
9	25.7717
10	29.2355
-	-
Average	52.9943

Intel(R) Core(TM) i9-7940X CPU @ 3.10GHz (OpenCL 2.1 (Build 0)):

Run	Duration (ms)
1	47.1116
2	19.7415
3	18.9358
4	19.5806
5	19.0173
6	19.5361
7	19.1919
8	19.4186
9	19.511
10	19.8893
-	-
Average	22.1934

Nvidia GeForce GTX 1650 (OpenCL 1.2 CUDA):

Run	Duration (ms)
1	470.605
2	0.3794
3	0.3591
4	0.2699
5	0.2626
6	0.2632
7	0.2584
8	0.2573
9	0.2595
10	0.2619
-	-
Average	47.3176

Note: The Nvidia driver appears to cache compiled kernels, leading to a more impressive average.

Intel(R) UHD Graphics 630 (OpenCL 2.1 NEO ):

Run	Duration (ms)
1	647.542
2	319.691
3	315.673
4	322.019
5	317.625
6	319.044
7	319.101
8	364.771
9	347.076
10	382.742
-	-
Average	365.528