Frequently Asked Questions
What is Dahlia?
Dahlia is a typed imperative programming language for designing FPGA accelerators. It targets High-Level Synthesis toolchains such as Vivado HLS as its compilation target. Dahlia aims to reduce pitfalls of HLS programming, from simple interface issues which are not checked by vendor tools to complex and unsuspecting behavior due to interaction between different optimizations.
What is predictable accelerator design?
High performance HLS designs require tuning hardware optimizations such as loop unrolling, pipelining, memory banking etc. These parameters have subtle interactions with each other and design constraints of FPGA programming.
For example, when a parallel loop tries to read five values from a memory that can only service two reads per cycle, the HLS compiler will multiplex the reads and spread them across multiple cycles. This will result in a design that consumes more area for the parallel loop body with suboptimal improvement in performance.
This is unpredictable design–FPGA designs should improve latency when loop bodies run in parallel. Dahlia statically reasons about memory use and rejects designs that don’t behave predictably.
Why design a new language?
Dahlia implements its own language constructs and semantics to cleanly reason about HLS programs and explore new ideas. Concepts such as logical time steps and affine reasoning are not commonly supported in modern programming languages. Implementing our own compiler and type checker simplifies these implementation challenges.
How does Dahlia work?
Hardware resources are reused through time-multiplexing—the compiler separates resource use in time by reasoning about use locations and latencies and guarantees that they are never used in a conflicting manner. In order to model this, Dahlia extends affine types with a notion of time-sensitivity.
At a high level, Dahlia guarantees that complex time-multiplexing is represented in the source program. By default, Dahlia rejects programs with complex memory access patterns that require extensive compiler transformations to be implemented as hardware. Instead of relying on the compiler to automatically generate additional hardware, Dahlia requires the programmer to explicitly choose complex transformations.
Dahlia is built and maintained by the CAPRA research group at Cornell University.
As a part of the Dahlia project, have built and open-sourced several research artifacts: