stillwater-sc / universal / 30029179195
85%
master: 84%

LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran
Jobs 1
Files 703
Run time 2min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 84.793% (+0.05%) from 84.744%
30029179195

push

github

web-flow 
feat(elreal): Phase 9 MVP -- block-shape study, dot-product benchmark, sweep oracle (#1185)

* feat(elreal): Phase 9 MVP -- block-shape design study, dot-product benchmark, sweep oracle (#933)

Kicks off Phase 9 (evaluation vs qd + hardware block-shape study, dissertation
5.1) with the MVP deliverables:

benchmark/performance/arithmetic/elreal/performance.cpp
  - memory footprint per block shape (half/bfloat16/float/double);
  - convergence: blocks needed to reach 50/100/200/320 decimal digits for
    pi (Machin), e (Taylor), sqrt2 (Newton), vs the 320-digit references;
  - time-to-first-block latency of the transcendental generators;
  - stream-wise ZBCL dot-product throughput (N = 16/64/256);
  - precision ceiling vs qd (elreal unbounded; qd fixed ~64 digits).
  Narrow hosts are exercised best-effort: each cell is guarded so a host that
  overruns its exponent range reports what it achieved rather than aborting.

elastic/elreal/oracle/sweep.cpp
  - randomised sweep of elreal ZBCL arithmetic against the exact dyadic-rational
    oracle (mpfr is out of scope -- Universal is dependency-free): add()/sub()
    exactness to the bit, mul() agreement to a high digit floor, over hosts
    {double, float}. Standard MANUAL_TESTING + REGRESSION_LEVEL skeleton.

docs/algorithmic-details/elreal-block-shape-study.md
  - written-up findings with the measured tables + first recommendation.

Key findings: double is the only host that reaches the high-precision regime
(300+ digits in ~20 blocks); float/bfloat16 saturate near ~35 digits because the
series arithmetic degrades in the narrow host (not a storage limit -- the wide
integer<256> exponent means narrow blocks save no memory); half does not
converge on the division-based generators (the known fp16 exponent-range
blocker). Geometric predicates, cancellation-stress sums, and the full
recommendation matrix are deferred as Phase 9 follow-ups.

CMake: benchmark_elreal_performance target + el_oracle_sweep target wired... (continued)

50111 of 59098 relevant lines covered (84.79%)

7019601.0 hits per line

Coverage Regressions

Lines Coverage ∆ File
1
80.58
 -0.22% include/sw/universal/number/posit1/specialized/posit_16_1.hpp
Jobs
ID Job ID Ran Files Coverage
1 30029179195.1 703
84.79
 GitHub Action Run
Source Files on build 30029179195