I saw that you posted your submission before I was able to respond. So, yes, we will allow you to amend that submission. We will contact you off line.
Iâve not seen this happen before. This function has been called many thousands of times before in this program, so it is something specific with this call. This can hang if the address being referenced is outside of the address map in the select logic. If this is happening you will see a message in the logic simulation transcript stating âaddressing errorâ.
If I am reading this correctly, the reference is in the softmax function call. The time consumed by the softmax function is minimal (and will be a rounding error in the power consumed). And the softmax is the final function to be performed. So you can sum the times up to this point and assume this is the time needed to compute the prediction.
Can you send me your neural network definition? Send it to Russell.klein@siemens.com. I will see if I can reproduce the problem.
To keep you progressing, just assume that this function took zero time. And look at the sum of the times to that point.
Thanks, Russ
The rule states that âarea used by the accelerators and the memory needed to hold the weights and any intermediate valuesâ.
Could you specify on how the area would be counted. I assume for each accelerator we count the total area reported by catapult after âextractâ stage. But how about the memory?
Thanks.
Hello.
I have run into the same issue. Has there been a solution or should I just stop the simulation and continue with the results at the time of the stall?
Hi Arisg,
Your inference time will be very close to the time that the last layer reported completing, so you can just use this.
If you want to run the simulation to completion, the bug is in the file catapult_conv2d.cpp in the cpp directory. On line 344 is the definition of the variable âinput_image_memâ. This is an array with a first dimension hard-coded to 10. One or more of your convolution layers has more than 10 channels. This array needs to be set to the greatest number of channels in any of your convolutional layers. Or you can just set it to 200 (or some large number thatâs not too large, but bigger than the channel count you might use).
I have fixed this, but rolling it out means that everyones VM will be reset, causing them to lose all work in progress. And I donât want to do that.
Regards, Russ
Hi Arisg,
Welcome to the Hackathon!
The area score for memory will be 0.4 square microns per bit needed for storing weights, features, and intermediate results.
If you look at the end of the file memory.h in the include file, there is a define for âMEMORY_SIZEâ. This is the total memory area needed, in words. Multiply this by the âWORD_SIZEâ you used for quantizing the network (the default is 32, if you did not change it). The multiply that. by 0.4 to get the area consumed by memory.
Let me know if you have further questions.
Regards, Russ