Method exec

dnn/src/rocm/convolution/forward/1x1.cpp:46–61

    return args.handle->matmul_opr()->get_workspace_in_bytes(A, B, C);
}
void ConvolutionForwardImpl::Algo1x1::exec(const ExecArgs& args) const {
    TensorND A, B, C;
    extract_matmul_layouts(args, A.layout, B.layout, C.layout);
    A.reset_ptr(args.filter_tensor->raw_ptr());
    B.reset_ptr(args.src_tensor->raw_ptr());
    C.reset_ptr(args.dst_tensor->raw_ptr());
    size_t batch = args.src_layout->shape[0];
    auto mm = args.handle->matmul_opr();
    auto strd_B = args.src_layout->stride[0] * args.src_layout->dtype.size(),
         strd_C = args.dst_layout->stride[0] * args.dst_layout->dtype.size();
    for (size_t i = 0; i < batch; ++i) {
        mm->exec(A, B, C, args.workspace);
        incr_refp(B.get_ref_ptr(), strd_B);
        incr_refp(C.get_ref_ptr(), strd_C);
    }
}

/*
 *  Funcitons to handle large batch