How bcast works

[hipudding]

I saw a testcase in test-backend-ops to test bcast with ADD, MUL and DIV. As far as I know, bcast only works for different dims, and all dims should be same or 1. But this case cast is adding tensor a(16,10,10,10) with b(32,10,10,10). I read the source code, but can't undersand clearly, could some one tell me how it works?

Does it bcast tensor a(16,10,10,10) to (32,10,10,10), and copy first "0-16" to "17-31"?

[slaren]

It repeats the second tensor to match the dimension of the first. Eg. ggml_add(a, b) is equivalent to ggml_add(a, ggml_repeat(b, a)).