How to correctly config BF16-Datatype in Gemmini

[haoywu-intel]

Hello, I have built this system with chipyard 1.9.1 and gemmini v0.7.1, the example simulations (both spike and verilator) in README work well on my machine.

Now, I'm trying to modify the datatypes. I've tried the default int_8 (GemminiConfigs.defaultConfig) and FP32 (GemminiFPConfigs.FP32DefaultConfig) configs. With tiny modifications on tiled_matmul_ws.c, I can get correct matrix multiplication results for both.

However, when I set the datatype to BF16 (GemminiFPConfigs.BF16Default8Config), the gemmini seems using the uint16 type to calculate the matrix instead of BF16 type. and I noticed the following typedef in gemmini_params.h:
typedef uint16_t elem_t;

Anyone has any comments? What else I need to do to support BF16 type matrix multiplication?

[hansa31]

Have you found a solution for this?