A03 Solutions

Reference answers with reasoning. Try the ../exercises/ yourself first - the learning is in the struggle, not the answer key.

This directory has its own Makefile:

make all GPU_ARCH=gfx950      # build every solution for your GPU
make run                      # build + run all; every one must PASS
make clean

#	File	Notes
1	01_fix_divergence_solution.cpp	t < n/2 → 2t, else 2(t-n/2)+1 makes every wavefront class-uniform. Explains the coalescing it trades away and when that trade flips.
2	02_warp_reduce_solution.cpp	Ladder starts at warpSize/2; cross-wavefront combine uses the computed warps_per_block. Includes the __shfl_down vs __shfl_down_sync portability note.
3	03_occupancy_math_solution.md	Full four-limiter math, why -maxrregcount is the wrong knob here, and the non-monotonic block-size/occupancy table.

The two programs both print PASSED and a measured ratio. Solution 1's speedup and solution 2's bandwidth are device-dependent — the invariants are that solution 1 is meaningfully faster than the divergent kernel and that solution 2 reports ratio 1.000x on both a 32-lane warp and a 64-lane wavefront.