How expensive are explicit conversions with convert_T() in OpenCL?

Asked Feb 16 '17 at 09:27

Active Feb 16 '17 at 10:23

Viewed 106 times

With reference to this question, is it possible that the gain in performance obtained by vectorized operations is being offset by the explicit conversions with convert_T()? Note that the default type of the variable is unsigned char. I am using OpenCL 2.0. My GPU is Intel HD Graphics 530 (Gen9).

Will it make a difference between convert_int4() and convert_short4()?

edited May 23 '17 at 12:24

Community

asked Feb 16 '17 at 09:27

Harsh Wardhan

2,110
10
36
51

it could be hardware dependent. if hardware has it, fast, and if drivers emulate it, slow. – huseyin tugrul buyukisik Feb 16 '17 at 09:44
Yeah, your best bet is to check what kind of assembly your platform generates – Jovasa Feb 16 '17 at 09:52
You could try your own bit hack magic but would only be fast on an fpga. – huseyin tugrul buyukisik Feb 16 '17 at 10:20
2

Well float to int may need some conversion. But int2short is just shift operation (or a noop if it is later used somewhere and the HW can read the LSB of the register). It all depends on the case. – DarkZeros Feb 16 '17 at 18:11
@DarkZeros which is used for (float to int / int to float)? FP32 logic units or integer logic units or completely different but just a few special units? – huseyin tugrul buyukisik Feb 17 '17 at 18:36

How expensive are explicit conversions with convert_T() in OpenCL?

0 Answers0