1 / 56

Session 4

Session 4. DSP Hardware Lab Optimized Convolution : Assembly Language, C, DSPLIB. Agenda. DDS Solution Optimizing the FIR Filter Convolution FIR filter as case study Vanilla C Basic Optimizer Hand Coded Assembly Language Library Routines Tuned / Optimized C. Goals.

ronalee
Download Presentation

Session 4

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Session 4 DSP Hardware Lab Optimized Convolution : Assembly Language, C, DSPLIB

  2. Agenda • DDS Solution • Optimizing the FIR Filter Convolution • FIR filter as case study • Vanilla C • Basic Optimizer • Hand Coded Assembly Language • Library Routines • Tuned / Optimized C

  3. Goals • Get Comfortable with the architecture of the processor • Be able to write simple assembly language for the C6713 • Be familiar with the libraries that typically exist, how to use them, and what restrictions to look for. • Be able to write C code which can effectively use the architecture of the processor • Overall : Have a toolbox of techniques to use in order to speed up troublesome code

  4. DDS Solution / Performance

  5. Recorded Data (440Hz) P=8 Recorded Data : 440 Hz , P=8

  6. Recorded Data (440Hz) P=8

  7. Audio • These differences are noticeable! • Techniques for increasing SFDR are important here since increasing table length is not always practical • Interpolating between points • Adding random values to the fractional bits of the phase accumulator reduces periodicity of error due to phase truncation, • Other methods P=8 P=16 P=3 Perfect

  8. Recorded Data (440Hz) P=8 original dithered

  9. Optimizing Methods

  10. Review : Computing Convolution • Hold Last N samples x[n], x[n-1]… • Filter characterized by the impulse response h[N] • For each new input sample x[n], compute convolution : • y[n] = h[0] * x[n] + h[1]*x[n-1] + … + h[N-1]*x[n-N+1] • Circular Buffer • Lets proceed to use Convolution Sum as a simple example for the wide variety of performance that can be seen simply by different coding methods

  11. Circular Buffers in C • To implement the circular buffers in C, we will need to perform the address arithmetic manually, including: • Keeping track of the next index to write to • Manually “looping” from the end of the circular buffer to the beginning • This technique can be sped up a number of ways: • Only allowing the number of taps to be a power of 2 means that address calculations can avoid the complexity of a modulo operation or a branch • Double-buffering the input with two circular buffers means that through “unrolling” the processing can access the memory in a linear fashion Write next value here n-6 n-7 n n-1 n-2 n-3 n-4 n-5 start finish Write next value here… …and here n-6 n-7 n n-1 n-2 n-3 n-4 n-5 n-6 n-7 n n-1 n-2 n-3 n-4 n-5 start finish

  12. Test Case : “Double Circular Buffer” • Using “double circular buffer” technique will allow us to simplify our convolution to a simple dot product calculation on linear vectors. • Good example for class Write next value here… …and here n-6 n-7 n n-1 n-2 n-3 n-4 n-5 n-6 n-7 n n-1 n-2 n-3 n-4 n-5 start finish • How can we do this as fast as possible? • Faster = More Taps = Higher Performance • Faster = More CPU cycles available for other things • Faster = Lower Power • Allows same functionality with lower CPU clock rates • Option 1 : Pure C • Option 2 : C with Hand Written Assembly Subroutine • Option 3 : DSPLib (Really just Option 2 where someone else wrote it) • Option 4 : Optimized C • Lets do some benchmarks!

  13. Benchmarks 1024 taps = 288 us At 48kHz, limits the realizable filter to about 70 taps Gut instinct – is this reasonable? float convolve( float * hptr, Int16 * data, int newdataindex, int numtaps) { int i; float dout = 0.0; Int16 *dptr = data+newdataindex; for (i=0;i<numtaps;i++) dout+= *hptr-- * *dptr--; return dout; }

  14. Each VLIW (very long instruction word) describes an operation for each of the 8 functional units.

  15. Datapaths / Functional Units

  16. Convolution Calculation • Processor runs at 225 MHz • This is 225 million instructions per second • That is 4687 instructions per 1/48kHz • With multiple functional units being capable • Doesn’t seem right that we can only do 70 taps! • What could be some issues? • We know we could do better with assembly right?

  17. Proposed Design Flow • Write Program Entirely in C • Understandable code • Fast development • Portable code • Use profiling tools / common sense to decide where to optimize by hand • Hand optimized assembly • Linear Assembly (use optimizing assembler) • Prefab library routines • Tuning C Next few slides are going to focus on writing a function in assembly language that can be called from C. To do that, we need to understand how C uses the processor and registers so we don’t mess it up!

  18. Strategy 1 C Callable Assembly Language Functions Required reading : http://focus.ti.com/lit/ug/spru733a/spru733a.pdf Chapter 3 will be most helpful with Lab 4

  19. Programming In Assembly • Benefit : we can use some of the hardware features that we have no way of expressing in C • AMR for circular addressing • Other control bits for special features (ex. Saturating arithmetic) • Benefit : we can also be generally more efficient than the code we write in C (perhaps) • Drawback : spru733.pdf • Compromise : Write parts that need to be fast in ASM, rest in C for simplicity. • We need to know “the rules” of register usage..etc. so that we can co-exist with the code generated by the C compiler. • These rules are well documented.

  20. Register Usage (from spru187) Where do arguments come to your asm, and how it returns values back to C You are responsible for Preserving their value

  21. Datapaths / Functional Units

  22. Hand Optimized Assembly • Knows exactly what each unit is doing and when • Seeks to optimize parallelization • 8 parallel operations at once can be launched from the pipeline scheduler • Understands the pipeline • Multiple clocks required to execute the full part of each instruction • Seeks to minimize waiting • Uses the most efficient instructions for the desired operations

  23. Example: Calling an assembly function comment label instruction Example 3.3: Factorial of a Number Using C Calling an Assembly Function Images from Digital Signal Processing and Applications with the C6713 and C6416 DSK, Rulph Chassaing

  24. Example Instruction Spru733 has complete list of instructions

  25. Example Instruction

  26. Indirect Addressing

  27. Pipeline Stages During the execution phase, sources are read. If data is available on the next cycle, then the delay slot is 0. If data is available, 2 clocks later, the delay slot is 1, etc. Images from Digital Signal Processing and Applications with the C6713 and C6416 DSK, Rulph Chassaing

  28. Delay Slot From SPRU189f

  29. More Assembly Instructions Images from Digital Signal Processing and Applications with the C6713 and C6416 DSK, Rulph Chassaing

  30. Hand Optimized Assembly • Knows exactly what each unit is doing and when • Seeks to optimize parallelization • 8 parallel operations at once can be launched from the pipeline scheduler • Understands the pipeline • Multiple clocks required to execute the full part of each instruction • Seeks to minimize waiting • Uses the most efficient instructions for the desired operations • This is hard!

  31. Strategy 2 DSPLIB Functions Reference : http://focus.ti.com/lit/ug/spru657c/spru657c.pdf (programmer’s ref) http://focus.ti.com/lit/an/spra947a/spra947a.pdf (examples using DSPLIB)

  32. DSPLIB • Catalog of hand tuned assembly language functions distributed freely by TI • Allows advertising benchmarks for popular functions • Highly optimized • Generally a good choice Example : DSPLIB FIR filter Cycles (4*floor((nh−1)/2)+14)*(ceil(nr/4)) + 8 e.g., nh=10, nr=100, cycles=758 cycles So 1024 taps, 100 samples = 526*4*25+8 = 5268

  33. DSPLIB From SPRU657b

  34. DSPLIB From SPRU657b

  35. Typical Function (Dot Product) IMPORTANT: Note these!

  36. Dotprod (cont)

  37. Strategy 3 Tuning / Optimizing C Code

  38. Setting Compiler Optimization Include support for debugging (obviously nice ‘till production) Compartments across which To optimize -O2 : Time for 1024 taps = 22 us!

  39. Optimization • Optimization should almost always be used in production code • Correctly written code should still work with the optimizer on • Incorrectly written code may behave differently in the optimized code than in unoptimized, but this is not a good reason for its non-use • Optimizer can make debug much tougher. • Statements executed out of order • Variables don’t necessarily exist • Code that you think is running gets “optimized away” if you don’t use its result. • Nice balance is to debug functions with optimizer off, optimize them, and leave them alone.

  40. Tools for Further Optimization We chose an extreme example, and one that is simple to understand. There will be many cases where C code does not optimize as smoothly, and raw assembly seems far better. • One reason that Assembly code performs better is that often the person writing assembly is able to make assumptions that the compiler is not allowed: • Locations of data • Alignment of parameters • Number of iterations through loops…etc • (remember DSPLIB? It levied some of these restrictions on the calling program) • Another is that often the person writing the assembly becomes aware of the bottlenecks, and can change the requirements / assumptions in order to remove those bottlenecks

  41. Tools for Further Optimization C Compiler tries to have free flow of information between programmer and compiler in order to more duplicate these benefits of assembly language programming. • Information FROM Programmer TO Compiler • “Pragma” and other keywords allow programmer to tell the compiler some ways in which the code will be used. • In other words, the user can place restrictions on the function which may increase its speed. • Information FROM Compiler TO Programmer • Compiler tries to communicate the bottlenecks and difficulties in its compilation by: • Describing the various time spent in loops • Showing the assembly code that it generates • Providing comments with that assembly to enable programmer to understand what it is doing. • “Compiler Consultant”

  42. Some Tips After this slide, much will be processor specific – however, these tips are pretty general for processors with long pipelines. • Effectively using the capabilities of the DSP absolutely requires making as much use of parallelism as possible. (note : this is hard with the capability to execute 8 instructions in parallel) • Branches can be expensive • Don’t process large data arrays in-place if possible • Compiler can more easily optimize if it knows that its instructions won’t affect the input array. Otherwise parallelism becomes impossible • Interactively engage the compiler by inspecting its assembly code. • Always keep and idea of how fast something SHOULD go in your mind. When things deviate greatly from this, investigate!

  43. Source : SPRA666.PDF • Guide to many of these things in far more detail than presented here • Please Read! • Lab session can be arranged to talk about some of these things with the instructors, since lecture will not be comprehensive • Try some examples!

  44. Basic Guidelines

  45. Compiler Consultant • Operates on loops • Goal : make simple linear code more pipelined. • Premise was that compiler can only do this sort of optimization if it has the right information. • “Effective software pipelining requires a great deal of loop analysis information. In many cases, not all of the information required is available in the source code. Compiler Consultant analyzes the loops in the application in order to determine what information may be missing, and then issues advice on how to address the information gap.”

  46. Pseudocode Example loop: load operate store bdec loop, count ; decrement count, branch if != 0 Cycles = 4*count load(n) load(n+1) || operate(n) load(n+2) || operate(n+1) || store(n) loop: load(n+3) || operate(n+2) || store(n+1) || bdec loop, count(n) operate(n+3) || store(n+2) store(n+3) Cycles = count + 2 Source : SPRAA14

  47. Example float test (float *in1, float *in2, int n) { float sum = 0; for (i=0;i<n;i++) { sum += *in1 * *in2 ; *in2++ = *in1++; } return sum; }

  48. Example Assembly Comments ;*----------------------------------------------------------------------------* ;* SOFTWARE PIPELINE INFORMATION ;* ;* Loop source line : 467 ;* Loop opening brace source line : 468 ;* Loop closing brace source line : 471 ;* Known Minimum Trip Count : 1 ;* Known Max Trip Count Factor : 1 ;* Loop Carried Dependency Bound(^) : 7 ;* Unpartitioned Resource Bound : 4 ;* Partitioned Resource Bound(*) : 5 ;* Resource Partition: ;* A-side B-side ;* .L units 0 1 ;* .S units 2 0 ;* .D units 3 5* ;* .M units 2 1 ;* .X cross paths 1 2 ;* .T address paths 4 4 ;* Long read paths 1 2 ;* Long write paths 0 0 ;* Logical ops (.LS) 1 1 (.L or .S unit) ;* Addition ops (.LSD) 2 1 (.L or .S or .D unit) ;* Bound(.L .S .LS) 2 1 ;* Bound(.L .S .D .LS .LSD) 3 3 ;* ;* Searching for software pipeline schedule at ... ;* ii = 7 Unsafe schedule for irregular loop ;* ii = 7 Unsafe schedule for irregular loop ;* ii = 7 Did not find schedule ;* ii = 8 Schedule found with 2 iterations in parallel ;* Done

  49. Consultant Advice • Function: testLines: 468-472Analysis: • Dependence constraints increased minimum cycles/iteration • from 5 to 7. • Advice: • Alias • Options can be met. Example Problem: The compiler cannot determine whether in1(C:\CCStudio_v3.3\examples\dsk6713\csl\mcbsp\mcbsp1\main_mcbsp1.c:464) might access the same memory locations as in2(C:\CCStudio_v3.3\examples\dsk6713\csl\mcbsp\mcbsp1\main_mcbsp1.c:464) within the loop. Suggestion: Restrict qualify the definition of in1, if the safety criteria can be met. 1st Alternate Suggestion: Restrict qualify the definition of in2, if the safety criteria Also gave advice about not using –g for debug info generation. That didn’t make a difference

  50. Restrict keyword

More Related