Ceva Dsp Programming

Use this skill for CEVA DSP kernel development, intrinsics selection, and optimization work in L1SW.

Scope

• Target architecture: CEVA-XC22 / XC6
• Main codebase area: app/dsp/
• Common use cases:
  • explain what a CEVA intrinsic does
  • explain CEVA Vec-C naming or intrinsic families
  • choose intrinsics for a kernel or hot loop
  • optimize CEVA SIMD code
  • map an algorithm to CEVA vector operations
  • answer mode-bit or configuration-register questions
  • route vector, complex, semi-complex, or width-specific CEVA operations to the right reference file
  • analyze an existing CEVA DSP kernel in L1SW

Working Rules

• Do not load the entire reference/ directory.
• Do not load reference/README.md unless the task is broad category discovery.
• First inspect the user request and the relevant code.
• Then read only the minimum number of reference files needed for the task.
• Prefer reference files whose names directly match the operation or intrinsic family in the request.
• If the request names a specific intrinsic or operation, load that exact reference file first.
• If the request contains a filename-like operation name, map it to the exact reference filename before using any broader category routing.
• Normalize user wording when matching filenames: ignore case, and treat spaces, hyphens, and underscores as interchangeable.
• If the request is about an existing kernel, search the code in app/dsp/ first, then load only the matching reference docs.

Recommended Workflow

Existing kernel analysis

1. Read the user request carefully and identify the operation family.
2. Inspect the relevant CEVA code in the workspace, usually under app/dsp/.
3. Map any filename-like request terms to exact reference filenames.
4. Load only the most relevant reference file or files from ceva-dsp-programming/reference/.
5. If needed, load one summary file for cross-checking:
   • reference/quick-reference.md
   • reference/CEVA_DSP_Intrinsics_Summary.md
6. Provide the answer or implementation using intrinsics that already appear in the L1SW codebase when possible.

Greenfield intrinsic selection

1. Read the user request carefully and identify the operation family.
2. Map any filename-like request terms to exact reference filenames.
3. Load the smallest matching reference file set.
4. Search app/dsp/ for similar intrinsics usage before recommending a final choice.
5. If needed, load one summary file for cross-checking.

Exact Filename Routing

Prefer exact filenames over category routing whenever the request is specific enough.

• vector pack 72 bits -> reference/vectorPack72-bits.md
• vector shift 72 bit -> reference/vectorShift72-bit.md
• vector exponent long 72 bit -> reference/vectorExponentLong72-bit.md
• semi complex 32x16 -> reference/semi-complex32x16.md
• semi complex 32x16 with exponent -> reference/semi-complex32x16WithExponent.md
• complex 32x16 -> reference/complex32x16.md
• load predicate -> reference/loadPredicate.md
• store predicate register -> reference/storePredicateRegister.md

When both a generic file and a vector-specific file seem applicable:

• prefer the vector* file for vector-form questions
• prefer the width-specific file for 40-bit, 64-bit, or 72-bit questions
• prefer the WithExponent variant when the request explicitly mentions exponent handling

Routing Guide

Use the smallest matching set of files.

By operation family

• Load and store:
  • reference/load.md
  • reference/store.md
  • reference/loadPredicate.md
  • reference/storePredicateRegister.md
• Arithmetic:
  • reference/addition.md
  • reference/subtract.md
  • reference/absolute.md
  • reference/negate.md
  • reference/round.md
• Multiply and MAC:
  • reference/multiply.md
  • reference/multiply-accumulate.md
  • reference/multiply-subtract.md
  • reference/genericMultiply.md
  • reference/genericMultiply-accumulate.md
• Shift, scale, exponent:
  • reference/shift.md
  • reference/scale.md
  • reference/exponent.md
  • reference/normalize.md
• Compare, predicate, masking:
  • reference/compare.md
  • reference/predicateRegisterManipulation.md
  • reference/and.md
  • reference/or.md
  • reference/xor.md
• Pack, unpack, transpose, permute:
  • reference/pack.md
  • reference/unpack.md
  • reference/transpose.md
  • reference/permute.md
• Floating point:
  • reference/floatArithmetic.md
  • reference/floatMultiplication.md
  • reference/floatDivision.md
  • reference/floatNormalize.md

By data shape or complex form

• 16x16, 32x16, 32x32 patterns:
  • reference/16x16.md
  • reference/32x16.md
  • reference/32x32.md
• Complex:
  • reference/complex16x16.md
  • reference/complex32x16.md
  • reference/complex32x32.md
• Semi-complex:
  • reference/semi-complex16x16.md
  • reference/semi-complex16x32.md
  • reference/semi-complex32x16.md
  • reference/semi-complex32x32.md
• With exponent handling:
  • load the matching WithExponent reference file

By mode-register and configuration questions

• Saturation, rounding, overflow, multiplication mode bits:
  • reference/saturationModeBits.md
  • reference/roundingModeBits.md
  • reference/overflowModeBits.md
  • reference/multiplicationModeBits.md
• Division, floating-point, store pre-shift, automatic mode bits:
  • reference/divisionModeBits.md
  • reference/floatingPointModeBits.md
  • reference/storePre-shiftModeBits.md
  • reference/automicOperationsModeBits.md

L1SW-Specific Guidance

• Prefer intrinsics that are already used in L1SW, especially under app/dsp/.
• When suggesting an optimization, preserve the kernel's numerical intent, scaling, and saturation behavior.
• Check whether the implementation depends on:
  • vector width and lane layout
  • Q-format or exponent tracking
  • VCU0 and VCU1 split
  • MODV0, MODV2, and related mode bits
  • load and store alignment and post-modify addressing
• Do not recommend generic CEVA intrinsics blindly if there is no evidence they are used or supported in this codebase.

Answer Expectations

• Explain why a chosen intrinsic fits the operand shape and data flow.
• Call out scaling, rounding, overflow, and accumulator width when relevant.
• For optimization tasks, compare the new approach against the current kernel structure.
• For implementation tasks, prefer minimal, verifiable changes.

Fallbacks

• If the exact intrinsic is unknown, search the workspace for similar usage before proposing one.
• If multiple reference files may apply, read the most specific file first and only then add a broader summary file.
• If the request is broad, use reference/README.md only as an index, not as the primary source.