Layout suite — reduction / copy / elementwise × memory layout × dtype

ratio = NumPy_ms / NumSharp_ms — >1.0 = NumSharp faster. ✅≥1.0 🟡≥0.5 🟠≥0.2 🔴<0.2. Layouts (8, harmonized with the cast subsystem): C, F (Fortran), T (transpose), strided [:, ::2], sliced (offset), negrow [::-1,:], negcol [:,::-1], bcast (stride-0). Fills the op-matrix's blind spot (it measures C-contiguous only). 100K + 1M elements, best-of-rounds.

Reduction (sum/min/max/prod, both axes)

Geomean by lay

size	C	F	T	strided	negrow	negcol	sliced	bcast
100K	1.01 ✅	1.05 ✅	1.04 ✅	0.49 🟠	0.61 🟡	0.56 🟡	0.68 🟡	0.59 🟡
1M	0.92 🟡	0.94 🟡	0.94 🟡	0.56 🟡	0.76 🟡	0.58 🟡	0.70 🟡	0.50 🟡

Geomean by dt

size	f64	f32	c128	dec	f16	i32	i64
100K	0.78 🟡	0.85 🟡	1.01 ✅	0.08 🔴	1.04 ✅	0.89 🟡	1.15 ✅
1M	0.88 🟡	1.05 ✅	1.02 ✅	0.07 🔴	1.00 ✅	0.80 🟡	1.02 ✅

Geomean by op

size	sum	min	max	prod
100K	0.72 🟡	0.61 🟡	0.61 🟡	1.06 ✅
1M	0.74 🟡	0.59 🟡	0.58 🟡	1.14 ✅

Worst 15 cells (NumSharp slowest vs NumPy)

key	NumSharp ms	NumPy ms	ratio
1M|dec|bcast|sum|ax0	5.3741	0.0657	0.01 🔴
100K|dec|bcast|sum|ax0	0.5412	0.0101	0.02 🔴
1M|dec|sliced|sum|ax0	5.4381	0.1108	0.02 🔴
100K|dec|C|sum|ax0	0.5494	0.0114	0.02 🔴
1M|dec|negrow|sum|ax0	5.4461	0.1134	0.02 🔴
100K|dec|negrow|sum|ax0	0.5371	0.0114	0.02 🔴
100K|dec|T|sum|ax1	0.5356	0.0114	0.02 🔴
1M|dec|F|sum|ax1	5.3609	0.1202	0.02 🔴
1M|dec|T|sum|ax1	5.5377	0.1270	0.02 🔴
100K|dec|sliced|sum|ax0	0.5332	0.0123	0.02 🔴
100K|dec|F|sum|ax1	0.5381	0.0126	0.02 🔴
1M|dec|C|sum|ax0	5.4777	0.1298	0.02 🔴
1M|i32|bcast|sum|ax1	4.1105	0.1203	0.03 🔴
100K|i32|bcast|sum|ax1	0.4050	0.0168	0.04 🔴
100K|dec|C|max|ax0	0.2871	0.0137	0.05 🔴

Copy / identity-ufunc (np.positive)

Geomean by lay

size	C	F	T	strided	sliced	negrow	negcol	bcast
100K	1.16 ✅	1.47 ✅	1.25 ✅	0.86 🟡	1.56 ✅	1.63 ✅	2.22 ✅	1.57 ✅
1M	2.87 ✅	2.95 ✅	2.89 ✅	1.96 ✅	2.67 ✅	2.62 ✅	3.24 ✅	2.67 ✅

Geomean by dt

size	u8	i8	i16	u16	i32	u32	i64	u64	char	f16	f32	f64	c128
100K	0.93 🟡	1.41 ✅	1.75 ✅	1.80 ✅	0.99 🟡	1.09 ✅	1.66 ✅	1.03 ✅	2.44 ✅	2.15 ✅	1.06 ✅	0.84 🟡	2.60 ✅
1M	4.51 ✅	4.77 ✅	2.15 ✅	2.11 ✅	2.11 ✅	2.22 ✅	2.61 ✅	2.64 ✅	2.15 ✅	2.14 ✅	2.14 ✅	2.57 ✅	5.31 ✅

Worst 15 cells (NumSharp slowest vs NumPy)

key	NumSharp ms	NumPy ms	ratio
100K|i64|strided|pos	0.0257	0.0110	0.43 🟠
100K|f64|strided|pos	0.0232	0.0105	0.45 🟠
100K|u64|strided|pos	0.0240	0.0109	0.45 🟠
100K|i64|C|pos	0.0441	0.0208	0.47 🟠
100K|f32|strided|pos	0.0202	0.0103	0.51 🟡
100K|i32|strided|pos	0.0200	0.0109	0.54 🟡
100K|i64|T|pos	0.0376	0.0208	0.55 🟡
100K|u8|negrow|pos	0.0419	0.0232	0.56 🟡
100K|u8|sliced|pos	0.0410	0.0230	0.56 🟡
100K|u8|bcast|pos	0.0409	0.0231	0.56 🟡
100K|f64|C|pos	0.0418	0.0242	0.58 🟡
100K|f64|F|pos	0.0388	0.0227	0.58 🟡
100K|u32|strided|pos	0.0180	0.0112	0.62 🟡
100K|f64|T|pos	0.0364	0.0230	0.63 🟡
100K|u64|T|pos	0.0358	0.0290	0.81 🟡

Elementwise (add/mul/neg/abs/sqrt/less/copy)

Geomean by lay

size	C	F	T	strided	sliced	negrow	negcol	bcast
100K	0.54 🟡	0.75 🟡	0.68 🟡	0.53 🟡	0.84 🟡	0.81 🟡	1.16 ✅	0.84 🟡
1M	1.57 ✅	1.54 ✅	1.53 ✅	1.15 ✅	1.66 ✅	1.65 ✅	1.80 ✅	1.67 ✅

Geomean by dt

size	f64	f32	c128	f16	i32	i64
100K	0.58 🟡	0.53 🟡	1.18 ✅	0.75 🟡	0.79 🟡	0.81 🟡
1M	1.91 ✅	1.59 ✅	1.74 ✅	0.96 🟡	1.55 ✅	1.82 ✅

Geomean by op

size	add	mul	neg	abs	sqrt	less	copy
100K	0.97 🟡	0.93 🟡	0.71 🟡	0.70 🟡	0.94 🟡	0.61 🟡	0.50 🟡
1M	1.81 ✅	1.80 ✅	2.12 ✅	1.66 ✅	1.55 ✅	0.69 🟡	1.82 ✅

Worst 15 cells (NumSharp slowest vs NumPy)

key	NumSharp ms	NumPy ms	ratio
100K|f64|strided|abs	0.0481	0.0075	0.16 🔴
100K|f64|C|copy	0.0700	0.0112	0.16 🔴
100K|f64|strided|neg	0.0502	0.0082	0.16 🔴
100K|f64|C|abs	0.0653	0.0113	0.17 🔴
100K|f32|C|abs	0.0310	0.0056	0.18 🔴
100K|f16|C|copy	0.0169	0.0031	0.18 🔴
100K|f64|C|mul	0.0690	0.0129	0.19 🔴
100K|f64|C|neg	0.0649	0.0129	0.20 🔴
100K|f64|C|add	0.0638	0.0130	0.20 🟠
100K|f16|negrow|copy	0.0178	0.0038	0.21 🟠
100K|f32|C|mul	0.0316	0.0069	0.22 🟠
100K|i32|bcast|copy	0.0224	0.0050	0.22 🟠
100K|f32|C|add	0.0311	0.0069	0.22 🟠
100K|f32|T|neg	0.0270	0.0060	0.22 🟠
100K|f32|T|add	0.0302	0.0068	0.23 🟠