The grouped convolution path in Generate_GPU_ALPAKA of ROperator_Conv.hxx produces incorrect numerical output.
gemm_n is set to total output channels at the top of the function and used directly in per-group matmul calls, g_out_offset strides, groupFOffset, and groupBiasElements but it is never divided by fAttrGroup.
PR #22 catches this issue
The grouped convolution path in
Generate_GPU_ALPAKAofROperator_Conv.hxxproduces incorrect numerical output.gemm_nis set to total output channels at the top of the function and used directly in per-group matmul calls,g_out_offsetstrides,groupFOffset, andgroupBiasElementsbut it is never divided byfAttrGroup.PR #22 catches this issue