Priority: P1 · Confidence: high · Effort: medium
Problem
The per-key decode + scoring logic is duplicated across all three GPU/CPU back-ends:
src/bruteforce.cu (CUDA/HIP, __device__)src/dmrcrack.cl (OpenCL kernel source)src/bruteforce.c (portable CPU)
These are hand-kept in sync. A bug fixed in one is easily missed in the others. Concrete evidence: the Hytera EP keystream fix (kiv[i] = key5[i] ^ MI[i], 40-bit MI) was applied to bruteforce.cu and bruteforce.c but silently missed in dmrcrack.cl (it kept kiv[0]=key5[0] and a 32-bit MI) until a follow-up pass. The OpenCL backend would have produced wrong Hytera keys.
include/dmr_tables.h and include/hytera_ks.h already single-source parts of this (the de-interleave tables and the Hytera keystream). The rest of the scoring kernel (de-interleave → mbe demod → extract → pack → RC4 → inter-frame Hamming + bit-freq chi²) is still triplicated.
Suggested fix
Move the scoring kernel body into a shared header included by bruteforce.cu, dmrcrack.cl and bruteforce.c, with thin macro shims to bridge the dialects (__device__/inline/__kernel, __shared__/__local, uchar/unsigned char). One implementation, compiled per back-end. Guard with a cross-backend known-answer test.
Why it matters
Eliminates the "fixed in one backend, not the others" bug class that already bit the Hytera path. Prerequisite for the broader multi-backend architecture work.
Priority: P1 · Confidence: high · Effort: medium
Problem
The per-key decode + scoring logic is duplicated across all three GPU/CPU back-ends:
src/bruteforce.cu(CUDA/HIP,__device__)src/dmrcrack.cl(OpenCL kernel source)src/bruteforce.c(portable CPU)These are hand-kept in sync. A bug fixed in one is easily missed in the others. Concrete evidence: the Hytera EP keystream fix (
kiv[i] = key5[i] ^ MI[i], 40-bit MI) was applied tobruteforce.cuandbruteforce.cbut silently missed indmrcrack.cl(it keptkiv[0]=key5[0]and a 32-bit MI) until a follow-up pass. The OpenCL backend would have produced wrong Hytera keys.include/dmr_tables.handinclude/hytera_ks.halready single-source parts of this (the de-interleave tables and the Hytera keystream). The rest of the scoring kernel (de-interleave → mbe demod → extract → pack → RC4 → inter-frame Hamming + bit-freq chi²) is still triplicated.Suggested fix
Move the scoring kernel body into a shared header included by
bruteforce.cu,dmrcrack.clandbruteforce.c, with thin macro shims to bridge the dialects (__device__/inline/__kernel,__shared__/__local,uchar/unsigned char). One implementation, compiled per back-end. Guard with a cross-backend known-answer test.Why it matters
Eliminates the "fixed in one backend, not the others" bug class that already bit the Hytera path. Prerequisite for the broader multi-backend architecture work.