1. 11 1 SPEX: A Programming Language for Software Defined Radio Yuan Lin, Robert Mullenix, Mark Woh, Scott Mahlke, Trevor Mudge, Alastair Reid 1, and Krisztián Flautner 1 The University of Michigan at Ann Arbor 1 ARM, Ltd

2. 22 2 ACAL – University of Michigan SDR Hardware Platform  Handset SDR has steep computational requirements (>40GOPS) with tight power budget (<500mW)  Heterogeneous multiprocessor solutions common  Examples Include:  SODA  Philips EVP  TI OMAP  IBM CELL* (not a mobile solution but very similar)  Themes:  Targets high-throughput signal-processing domains  VLIW SIMD  Control- and data- focused cores

3. 33 3 ACAL – University of Michigan SODA System Architecture for 3G  Based on our prior work:  4 PEs  Scalar, SIMD pipelines  32 elements wide  Scratchpad memories  ARM controller processor  Handles control, DMA  Feasible approach:  3W, 26.6 mm 2 at 180nm  ~0.5W, 6.7 mm 2 projected for 90nm

4. 44 4 ACAL – University of Michigan Programming SDR Platforms  Programming DSPs already tough  Multiprocessor architectures make a tough problem tougher  Want to achieve high performance with high productivity  Software needs to advance along with hardware  C not sufficient  Can express protocols, but awkward and inefficient  Rediscovering parallelism is challenging  Want to decouple algorithm design from implementation  No well-defined concept of time

5. 55 5 ACAL – University of Michigan Algorithm Characteristics  Digital communication protocols hierarchical  Abstracted as a series of connected kernels  Can isolate and optimize individually  Operations frequently involve matrix computations  Have real-time constraints with static control flow W-CDMA Protocol Operations

6. 66 6 ACAL – University of Michigan Desired SDR Language Features  Plenty of Parallelism  Kernels vectorizible  Pipeline the stream  Interleave concurrent tasks  Give compiler control  Express the algorithms & constraints, not run-time behavior  Static decisions should be made at compile time  (e.g. scheduling, PE assignment, memory management)  Support for Timing Models  Absolute timing primitives prevent drifting  Periodic and relative timing constraints

7. 77 7 ACAL – University of Michigan SPEX – Language Extension for SDR  Two levels: data/control separation  Kernel SPEX - the data plane  Algorithm kernel descriptions, timing unaware  C + Matlab operators + DSP fixed-point arithmetic  System SPEX - the control plane  Wireless protocol system descriptions  C + Inter-kernel communications + timing constraints

8. 88 8 ACAL – University of Michigan Kernel SPEX  Atomic Building Blocks  Non-preemptible  Ignorant of timing constraints  Maintains local state  Features  Templated definitions  Member functions  Matlab-like vector support  SystemC-like data types template kernel FIR { vector z; vector coeff; void set_coeff(vector c) { coeff = c; } void run(channel inbuf, channel outbuf) { int i; T in, out; for (i = 0; i < BSIZE; i++) { in = inbuf.pop(); z += coeff * in; out = z[0]; outbuf.push(out); z = (z(1:TAPS-1),0); } };

9. 99 9 ACAL – University of Michigan System SPEX  Synchronous primitives  A set of timing and concurrent primitives for expressing real-time execution  Modeled after real-time languages  Stream primitives  A set of streaming primitives for expressing streaming computation  Modeled after the synchronous dataflow model and its variations

10. 10 ACAL – University of Michigan Parallel execution of instructions within each scope Synchronous Example – WCDMA Real-time clock Absolute timing assertion void wcdma() { clock clk; at (clk % wcdma_frame == 0) {... adcfir(ch1); chan_est(ch1, ch2, num_fingers); parallel { bch(ch1, bch_done); if (dch_mode) { dch(ch1, ch2, num_fingers, dch_done); }... parallel { wait(clk % bch_deadline == 0); if (dch_mode) wait(clk % dch_deadline == 0); }

11. 11 ACAL – University of Michigan SDR Compilation Strategy

12. 12 ACAL – University of Michigan gcc a.out + libraries Functional debugging path Kernel SPEX Compilation Flow Virtual Kernel C SPEX frontend Kernel SPEX  Frontend removes “syntactic sugar”  Templates instantiated  Matlab features mapped to function calls  Virtual Kernel C  Infinite vector length assumed  Robust set of operators  Can be linked with special libraries and compiled with gcc to verify functional correctness  Physical Kernel C  Vector length bounded by actual SIMD width  Restricted to machine operators V-to-P translation VLIW backend Physical Kernel C SODA assembly Compilation path

13. 13 ACAL – University of Michigan System SPEX Compilation Frontend

14. 14 ACAL – University of Michigan System SPEX Task Compilation  Stream IR with dataflow primitives  i.e. push(), pop(), peek()  Step 1: Dataflow rate-matching  Insert buffers between nodes  Add loops to match the rate  Step 2: Initial resource allocation  Processor assignments  memory allocation and DMA transfer  Step 3: Control-data split  Break the task into independent threads 1 2 3

15. 15 ACAL – University of Michigan System SPEX Real-time Optimization  Hierarchical constraint scheduling  Each task is treated as a single node  Guarantees all nodes are schedulable through compiler optimizations  Non-preemptive multi- processor scheduler  Static processor assignments  Static task execution ordering  Dynamic execution timing  Iterative optimization if constraints not met  Re-compile each task with system profiling

16. 16 ACAL – University of Michigan Summary  Multiprocessor architectures makes handset SDR feasible, but complicates software  Need better language to map algorithm to hardware  SPEX capitalizes on domain properties  C and Matlab based  Control and data separation  Kernels exploit massive data parallelism  Systems can pipeline kernels and interleave tasks  Compile system and kernels independently  Provide multiple paths to ensure robust debugging

17. 17 ACAL – University of Michigan Questions?

18. 18 ACAL – University of Michigan Stream Example – DCH Channel void DCH(channel ADC_in, channel searcher_in, int num_fingers, channel & to_MAC, signal & done) { channel ch1[max_rake_finger]; channel ch2; stream { for (int i = 0; i < num_fingers; i++) rake(ADC_in, searcher_in[i], ch1[i]); combiner(ch1, ch2); viterbi(ch2, to_MAC); } done = true; } Stream computation within the scope channels and signals can be declared either as function arguments or as local variables Channel merging done in combiner function

19. 19 ACAL – University of Michigan System SPEX Compilation