Every character
is a stroke.

Write the shortest kernel that matches the reference output. Then write the fastest one. Two scorecards, one hole.

Par — length of the reference implementation
Strokes — characters in your submission
Speedup — measured against the baseline

Scorecard

6 holes · course record 68

Hole	Problem	Shape		Par	Best	Fastest	Field
01	Row SoftmaxNumerically stable, last axis	4096 × 1024	PAR 3	180	149	88.2µs	4
02	Square Matmulfp16, tensor cores allowed	1024 × 1024	PAR 5	420	388	198.6µs	3
03	LayerNorm ForwardAffine, eps inside the sqrt	8192 × 768	PAR 4	260	241	142.8µs	2
04	GELU (tanh approx)Pure elementwise	1048576	PAR 3	96	68	22.9µs	3
05	Row ArgmaxTies resolve to lowest index	2048 × 4096	PAR 4	210	194	79.3µs	2
06	Causal Conv1DDepthwise, width 4, no lookahead	32 × 512 × 2048	PAR 5	340	327	356.2µs	2
Out				1506	1367	−139	5

#	Player	Backend	Strokes	To par	Runtime	Speedup
1	geohot	tinygrad	149	−31	96.4µs	2.23×
2	nx7	triton	158	−22	103.7µs	2.07×
3	mmu	cuda	171	−9	88.2µs	2.43×
4	halide_h	metal	186	+6	121.5µs	1.77×