Question on RTL test bench googlenet_conv2_3x3_int16

nvdla / hw

RTL, Cmodel, and testbench for NVDLA

Other

1.71k stars 565 forks source link

Question on RTL test bench googlenet_conv2_3x3_int16 #259

Open silvaurus opened 5 years ago

silvaurus commented 5 years ago

Hi!

Could you tell me why this Googlenet testbench has an input of 19x19x64? Standard Googlenet has 56x56x64. And also, why the output dump 0x2da80 bytes (187008)? Since 187008/2 = 93504 = 487 192. 487 is a prime number. Standard Googlenet has 56 56 * 192.

silvaurus commented 5 years ago

Question solved. 0x2da80 is because of the output line stride and surface stride., which adds several lines of 0s. After getting rid of those lines of 0s, output goes back to 1919192. See here http://nvdla.org/hw/format.html.

Why using 19 instead of 56 is unknown. Maybe because this is a Cifar10 model. Which the input for conv2 should be 32/2 = 16.