NumKong/include/numkong/dot/README.md at main · rese3c/NumKong · GitHub
Skip to content

README.md

Latest commit

 

History

History
443 lines (396 loc) · 50.4 KB

README.md

File metadata and controls

443 lines (396 loc) · 50.4 KB
Raw

Vector-Vector Dot Products in NumKong

NumKong implements dot products for every numeric type supported by the library, as a core building block of higher-level functionality for vectors and higher rank tensors.

Dot product for real numbers and integers is defined as:

$$ \text{dot}(a, b) = \sum_{i=0}^{n-1} a_i \cdot b_i $$

For complex numbers, the dot product expands via the distributive property of complex multiplication:

$$ \text{dot}(a, b) = \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,re} - a_{i,im} \cdot b_{i,im}) + j \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,im} + a_{i,im} \cdot b_{i,re}) $$

The conjugate dot product negates the imaginary part of $b$:

$$ \text{vdot}(a, b) = \sum_{i=0}^{n-1} a_i \cdot \bar{b_i} = \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,re} + a_{i,im} \cdot b_{i,im}) + j \sum_{i=0}^{n-1} (a_{i,im} \cdot b_{i,re} - a_{i,re} \cdot b_{i,im}) $$

Where $\bar{b_i}$ is the complex conjugate of $b_i$. Reformulating as Python pseudocode for interleaved real/imaginary scalar arrays:

def dot_real(a: List[number], b: List[number]) -> number:
    return sum(ai * bi for ai, bi in zip(a, b))

def dot_complex(a: List[number], b: List[number]) -> Tuple[number, number]:
    a_re, a_im = a[0::2], a[1::2]
    b_re, b_im = b[0::2], b[1::2]
    ab_re = sum(ar * br - ai * bi for ar, ai, br, bi in zip(a_re, a_im, b_re, b_im))
    ab_im = sum(ar * bi + ai * br for ar, ai, br, bi in zip(a_re, a_im, b_re, b_im))
    return ab_re, ab_im

def vdot_complex(a: List[number], b: List[number]) -> Tuple[number, number]:
    a_re, a_im = a[0::2], a[1::2]
    b_re, b_im = b[0::2], b[1::2]
    ab_re = sum(ar * br + ai * bi for ar, ai, br, bi in zip(a_re, a_im, b_re, b_im))
    ab_im = sum(ai * br - ar * bi for ar, ai, br, bi in zip(a_re, a_im, b_re, b_im))
    return ab_re, ab_im

Input & Output Types

Real and integer dot products:

Input Type Output Type Description
f64 f64 64-bit IEEE 754 double precision
f32 f32 32-bit IEEE 754 single precision
f16 f32 16-bit IEEE 754 half precision, widened output
bf16 f32 16-bit brain float, widened output
e4m3 f32 8-bit Float8: 4 exponent, 3 mantissa bits
e5m2 f32 8-bit Float8: 5 exponent, 2 mantissa bits
e2m3 f32 8-bit MX format: 2 exponent, 3 mantissa bits
e3m2 f32 8-bit MX format: 3 exponent, 2 mantissa bits
i8 i32 8-bit signed integers
u8 u32 8-bit unsigned integers
i4 i32 4-bit signed integers, packed nibble pairs
u4 u32 4-bit unsigned integers, packed nibble pairs
u1 u32 1-bit binary packed octets, popcount of AND

Complex dot products (both dot and vdot):

Input Type Output Type Description
f64c f64c 64-bit complex pairs
f32c f32c 32-bit complex pairs
f16c f32c 16-bit complex pairs, widened output
bf16c f32c 16-bit brain complex pairs, widened output

Optimizations

Compensated Arithmetic for Large Floats

nk_dot_f64_serial uses Neumaier compensated summation — tracking a correction term adjusted by magnitude comparison at each step. nk_dot_f64_haswell, nk_dot_f64_skylake, nk_dot_f64_sve implement the Dot2 algorithm by Ogita, Rump, and Oishi: TwoProd via FMA captures the rounding error of each product exactly, and a TwoSum chain propagates it through the accumulator. On SVE, the final horizontal reduction uses svtbl to extract upper halves at each tree level, applying TwoSum at every stage. The serial path uses Neumaier because it processes one element at a time and can cheaply branch on magnitudes. Dot2 avoids those branches entirely — TwoProd and TwoSum are pure arithmetic with no comparisons, mapping naturally to wide SIMD where branching per lane is impossible.

Lookup Tables for Mini-Floats

nk_dot_e2m3_haswell, nk_dot_e3m2_haswell, nk_dot_e2m3_skylake, nk_dot_e3m2_skylake encode 32 MX format values into scaled integers via dual 16-entry LUTs loaded into vector registers. The low 4 magnitude bits index VPSHUFB, bit 4 selects between the lower and upper table via blending, and the results feed into VPMADDUBSW + VPMADDWD chains with a final $\div 256$ scaling. The Sapphire-specific MX implementation in sapphire.h replaces this with a single 64-entry signed LUT via VPERMUTEX2VAR, where the sign bit naturally selects between positive and negative tables. That path accumulates in native Float16 via VFMADD_PH and flushes to Float32 every 128 elements to avoid overflow.

Algebraic Domain Shifting

nk_dot_i8_icelake, nk_dot_u8_icelake work around VPDPBUSD requiring UInt8 × Int8 operands. For Int8 × Int8, one operand is XORed with 0x80 to shift to unsigned, and the correction $128 \cdot \sum b_i$ is computed via VPSADBW, which runs on port 5 and avoids contention with DPBUSD on ports 0-1. nk_dot_i4_icelake extends this to packed nibbles using the identity $(a'-8)(b'-8) = a' b' - 8(a'+b') + 64$ — two VPDPBUSD calls handle low and high nibbles separately, with SAD-based correction. nk_dot_i8_v128relaxed, nk_dot_u8_v128relaxed face an even tighter constraint: WASM's i32x4_relaxed_dot_i8x16_i7x16_add computes Int8 × Int7, so the sign bit of one operand must be masked off entirely. For Int8 × Int8, the sign bit of $b$ is cleared to produce a 7-bit value, and a windowed correction $-128 \cdot \sum_{b_i < 0} a_i$ is accumulated in Int16 and flushed every 127 iterations to prevent overflow. For UInt8 × UInt8, $b$ is XORed with 0x80 to shift into signed range, same as Ice Lake, with the correction $128 \cdot \sum a_i$ computed via pairwise widening adds.

Octave Decomposition for E4M3 via VNNI

nk_dot_e4m3_icelake splits the 4-bit E4M3 exponent into 2 "octave" bits (top) and 2 "remainder" bits (bottom). The bottom 5 bits (2 remainder + 3 mantissa) map via VPERMB to u8 integers in [0, 120] — identical structure to the E2M3 $\times 16$ LUT. A subnormal fixup replaces LUT entries for magnitude < 8 with $2 \times \text{mantissa}$ via a second masked VPERMB, avoiding VPADDB on the VPDPBUSD execution ports. Sign is computed via VPTERNLOGD with immediate 0x14, fusing (a \oplus b) \wedge \lnot \text{0x7F} in one instruction. The 4 octave bins per operand produce $4 \times 4 = 16$ VPDPBUSD cross-products accumulated into 7 registers grouped by octave sum $k = o_a + o_b \in [0, 6]$. Each accumulator is scaled by $2^{4k-20}$ — an exact power of two, introducing no rounding. This processes 64 E4M3 bytes per iteration in u8, doubling the element density of the BF16 upcast path.

Widening Fusion Through BFloat16 on x86

nk_dot_e5m2_genoa converts FP8 values to BF16, then accumulates via VDPBF16PS, reusing Genoa's BF16 dot-product instruction for FP8 types. Each VDPBF16PS fuses two BF16 multiply-adds per 32-bit lane at 6-cycle throughput. On Skylake-X–class CPUs without BF16 dot-product hardware, nk_dot_e4m3_skylake / nk_dot_e5m2_skylake (and their Haswell twins nk_dot_e4m3_haswell / nk_dot_e5m2_haswell) instead route through the Giesen-style FP8 → F16 fake-bit-pattern cast, widen via VCVTPH2PS, and accumulate in F32 with two independent FMA chains reducing into a single register — avoiding the 3-chain scheduler-stall of the BF16 algebraic form on kernels without native BF16 FMA. nk_dot_bf16c_genoa uses the same instruction for complex BF16, preparing operands with VPSHUFB for lane swapping and VPXORD with 0x80000000 for sign flips before feeding into VDPBF16PS.

Deferred Sign-Flip in Complex Dot Products

The Haswell BFloat16Complex/Float16Complex/Float32Complex kernels compute $\sum (a_r b_r - a_i b_i)$ without per-pair subtraction. Instead, two accumulators collect interleaved products $[a_r b_r, a_i b_i, \ldots]$ and $[a_r b_i, a_i b_r, \ldots]$, and a post-loop XOR flips the sign of every odd lane to produce the subtraction. This gives one FMA per accumulator per iteration, but each lane grows $O(n)$ while the true result is $O(\sqrt{n})$. The Float32Complex kernel absorbs this via Float64 accumulators; Genoa's VDPBF16PS and ARM's FMLSL pair terms naturally. For BFloat16Complex/Float16Complex on Haswell the accumulator is Float32, so the $O(\log n)$ precision loss from lane separation is visible in max ULP at large $n$, though mean ULP remains low.

Widening Fusion Through Float16 on Arm

nk_dot_f16_neonfhm, nk_dot_f16c_neonfhm use the ARMv8.4-FHM instructions FMLAL/FMLSL, which fuse FP16-to-FP32 conversion with multiply-accumulate in a single operation. vfmlalq_low_f16 and vfmlalq_high_f16 process the lower and upper 4 elements of an 8-wide FP16 vector respectively. For complex dot products, FMLSL provides the subtraction path $a_{re} b_{im} - a_{im} b_{re}$ without a separate negate step.

Widening Chains on RISC-V

nk_dot_i8_rvv, nk_dot_u8_rvv use vwmul for Int8 × Int8 → Int16 widening multiply followed by vwadd to widen-accumulate into Int32 — a two-stage chain that naturally prevents overflow. nk_dot_bf16_rvvbf16 uses the Zvfbfwma extension's vfwmaccbf16 for fused BFloat16 × BFloat16 → Float32 widening multiply-accumulate. nk_dot_e4m3_rvvbf16, nk_dot_e5m2_rvvbf16 convert Float8 to BFloat16 via 256-entry LUTs, then feed the same vfwmaccbf16 path.

Performance

The following performance tables are produced by manually re-running nk_test and nk_bench included internal tools to measure both accuracy and throughput at different input shapes. The input size is controlled by the NK_DENSE_DIMENSIONS environment variable and set to 256, 1024, and 4096 elements. The throughput is measured in gb/s as the number of bytes read per second amortized for a large batch of vector pairs. Accuracy is reported as mean ULP (units in last place) unless noted otherwise — the average number of representable floating-point values between the result and the exact answer. Rows marked 🧩 use external BLAS baselines rather than NumKong kernels. Each kernel runs for at least 20 seconds per configuration. Benchmark threads are pinned to specific cores; on machines with heterogeneous core types (e.g., Apple P/E cores), only the fastest cores are used. Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.

Intel Sapphire Rapids

Native

Kernel 256 1024 4096
f64c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
dot_f64c_with_blas 🧩 29.1 gb/s, 25 ulp 27.6 gb/s, 97 ulp 13.6 gb/s, 32 ulp
vdot_f64c_with_blas 🧩 29.1 gb/s, 18 ulp 27.9 gb/s, 17 ulp 15.3 gb/s, 25 ulp
nk_dot_f64c_serial 5.45 gb/s, 3.9 ulp 6.49 gb/s, 9.0 ulp 6.84 gb/s, 2.9 ulp
nk_vdot_f64c_serial 5.47 gb/s, 4.6 ulp 6.41 gb/s, 1.6 ulp 6.76 gb/s, 2.2 ulp
nk_dot_f64c_skylake 23.8 gb/s, 0 ulp 23.4 gb/s, 0 ulp 11.8 gb/s, 0 ulp
nk_vdot_f64c_skylake 23.6 gb/s, 0 ulp 23.7 gb/s, 0 ulp 11.6 gb/s, 0 ulp
f32c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
dot_f32c_with_blas 🧩 28.7 gb/s, 8.6 ulp 29.8 gb/s, 13 ulp 15.8 gb/s, 19 ulp
vdot_f32c_with_blas 🧩 29.2 gb/s, 11 ulp 30.2 gb/s, 14 ulp 15.7 gb/s, 21 ulp
nk_dot_f32c_serial 9.46 gb/s, 0 ulp 9.82 gb/s, 0 ulp 9.71 gb/s, 0 ulp
nk_vdot_f32c_serial 9.64 gb/s, 0 ulp 9.95 gb/s, 0 ulp 10.1 gb/s, 0 ulp
nk_dot_f32c_haswell 22.4 gb/s, 0 ulp 22.2 gb/s, 0 ulp 12.6 gb/s, 0 ulp
nk_vdot_f32c_haswell 22.4 gb/s, 0 ulp 21.8 gb/s, 0 ulp 14.7 gb/s, 0 ulp
nk_dot_f32c_skylake 25.6 gb/s, 0 ulp 27.2 gb/s, 0 ulp 17.0 gb/s, 0 ulp
nk_vdot_f32c_skylake 27.8 gb/s, 0 ulp 27.4 gb/s, 0 ulp 18.8 gb/s, 0 ulp
bf16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16c_serial 0.628 gb/s, 0.1 ulp 0.627 gb/s, 2.3 ulp 0.626 gb/s, 7.9 ulp
nk_vdot_bf16c_serial 0.622 gb/s, 0.2 ulp 0.624 gb/s, 2.1 ulp 0.627 gb/s, 11.2 ulp
nk_dot_bf16c_haswell 21.5 gb/s, 0.1 ulp 18.5 gb/s, 1.3 ulp 18.4 gb/s, 3.4 ulp
nk_vdot_bf16c_haswell 21.9 gb/s, 0.8 ulp 19.0 gb/s, 2.0 ulp 18.5 gb/s, 4.5 ulp
nk_dot_bf16c_genoa 37.9 gb/s, 0 ulp 30.3 gb/s, 1.1 ulp 29.5 gb/s, 2.8 ulp
nk_vdot_bf16c_genoa 36.1 gb/s, 0.7 ulp 30.2 gb/s, 1.2 ulp 30.2 gb/s, 3.3 ulp
f16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16c_serial 2.02 gb/s, 14.4 ulp 2.00 gb/s, 27.3 ulp 2.02 gb/s, 34.0 ulp
nk_vdot_f16c_serial 1.67 gb/s, 15.0 ulp 1.64 gb/s, 26.3 ulp 1.64 gb/s, 34.2 ulp
nk_dot_f16c_haswell 23.9 gb/s, 12.7 ulp 19.4 gb/s, 22.3 ulp 19.3 gb/s, 40.1 ulp
nk_vdot_f16c_haswell 24.0 gb/s, 11.1 ulp 20.0 gb/s, 17.4 ulp 17.1 gb/s, 29.2 ulp
f64 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
dot_f64_with_blas 🧩 27.8 gb/s, 6.9 ulp 30.1 gb/s, 9.3 ulp 15.7 gb/s, 20 ulp
nk_dot_f64_serial 4.28 gb/s, 2.2 ulp 4.39 gb/s, 2.0 ulp 4.42 gb/s, 3.3 ulp
nk_dot_f64_haswell 24.2 gb/s, 0 ulp 25.7 gb/s, 0 ulp 18.3 gb/s, 0 ulp
nk_dot_f64_skylake 29.0 gb/s, 0 ulp 28.6 gb/s, 0 ulp 24.9 gb/s, 0 ulp
f32 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
dot_f32_with_blas 🧩 47.8 gb/s, 14 ulp 30.7 gb/s, 14 ulp 29.7 gb/s, 15 ulp
nk_dot_f32_serial 11.0 gb/s, 0 ulp 11.2 gb/s, 0 ulp 11.5 gb/s, 0 ulp
nk_dot_f32_haswell 30.5 gb/s, 0 ulp 23.9 gb/s, 0 ulp 24.4 gb/s, 0 ulp
nk_dot_f32_skylake 44.2 gb/s, 0 ulp 29.8 gb/s, 0 ulp 30.0 gb/s, 0 ulp
bf16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16_serial 0.633 gb/s, 0 ulp 0.630 gb/s, 0.5 ulp 0.638 gb/s, 5.4 ulp
nk_dot_bf16_haswell 39.3 gb/s, 0 ulp 25.5 gb/s, 0.2 ulp 20.2 gb/s, 25.3 ulp
nk_dot_bf16_skylake 62.7 gb/s, 0 ulp 30.2 gb/s, 0.2 ulp 29.5 gb/s, 2.3 ulp
nk_dot_bf16_genoa 88.8 gb/s, 0 ulp 29.7 gb/s, 0.2 ulp 31.2 gb/s, 2.2 ulp
f16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16_serial 1.31 gb/s, 11.5 ulp 1.32 gb/s, 33.7 ulp 1.30 gb/s, 59.7 ulp
nk_dot_f16_haswell 31.3 gb/s, 7.0 ulp 22.8 gb/s, 14.0 ulp 19.8 gb/s, 29.8 ulp
nk_dot_f16_skylake 54.9 gb/s, 6.2 ulp 31.7 gb/s, 8.6 ulp 30.9 gb/s, 22.8 ulp
e5m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e5m2_serial 1.90 gb/s, 0 ulp 1.07 gb/s, 0 ulp 1.08 gb/s, 0 ulp
nk_dot_e5m2_haswell 4.92 gb/s, 0 ulp 4.95 gb/s, 0 ulp 4.80 gb/s, 0 ulp
nk_dot_e5m2_skylake 6.20 gb/s, 0 ulp 6.36 gb/s, 0 ulp 6.25 gb/s, 0 ulp
nk_dot_e5m2_genoa 12.1 gb/s, 0 ulp 12.6 gb/s, 0 ulp 12.6 gb/s, 0 ulp
e4m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e4m3_serial 0.762 gb/s, 0 ulp 0.424 gb/s, 0 ulp 0.420 gb/s, 0 ulp
nk_dot_e4m3_haswell 3.78 gb/s, 0 ulp 3.77 gb/s, 0 ulp 3.75 gb/s, 0 ulp
nk_dot_e4m3_skylake 5.10 gb/s, 0 ulp 5.16 gb/s, 0 ulp 5.21 gb/s, 0 ulp
nk_dot_e4m3_icelake 13.2 gb/s, 0 ulp 14.9 gb/s, 0 ulp 14.7 gb/s, 0 ulp
e3m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e3m2_serial 1.47 gb/s, 0 ulp 1.05 gb/s, 0 ulp 1.04 gb/s, 0 ulp
nk_dot_e3m2_haswell 12.0 gb/s, 0 ulp 12.2 gb/s, 0 ulp 12.2 gb/s, 0 ulp
nk_dot_e3m2_skylake 21.6 gb/s, 0 ulp 23.1 gb/s, 0 ulp 23.2 gb/s, 0 ulp
nk_dot_e3m2_icelake 23.1 gb/s, 0 ulp 24.3 gb/s, 0 ulp 23.9 gb/s, 0 ulp
e2m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e2m3_serial 1.87 gb/s, 0 ulp 1.25 gb/s, 0 ulp 1.96 gb/s, 0 ulp
nk_dot_e2m3_haswell 20.5 gb/s, 0 ulp 20.4 gb/s, 0 ulp 19.3 gb/s, 0 ulp
nk_dot_e2m3_skylake 35.7 gb/s, 0 ulp 33.2 gb/s, 0 ulp 30.7 gb/s, 0 ulp
nk_dot_e2m3_icelake 58.0 gb/s, 0 ulp 46.0 gb/s, 0 ulp 31.5 gb/s, 0 ulp
nk_dot_e2m3_alder 29.9 gb/s, 0 ulp 30.8 gb/s, 0 ulp 29.1 gb/s, 0 ulp
i8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i8_serial 16.9 gb/s 16.8 gb/s 15.6 gb/s
nk_dot_i8_haswell 43.2 gb/s 35.8 gb/s 29.1 gb/s
nk_dot_i8_skylake 52.9 gb/s 36.5 gb/s 28.5 gb/s
nk_dot_i8_icelake 64.0 gb/s 46.2 gb/s 26.8 gb/s
nk_dot_i8_alder 42.8 gb/s 40.4 gb/s 31.1 gb/s
u8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u8_serial 16.9 gb/s 16.5 gb/s 15.8 gb/s
nk_dot_u8_haswell 47.7 gb/s 37.7 gb/s 29.1 gb/s
nk_dot_u8_skylake 48.7 gb/s 32.6 gb/s 27.5 gb/s
nk_dot_u8_icelake 68.4 gb/s 46.9 gb/s 30.2 gb/s
nk_dot_u8_alder 42.1 gb/s 41.8 gb/s 31.6 gb/s
i4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i4_serial 9.37 gb/s 11.8 gb/s 11.8 gb/s
nk_dot_i4_haswell 8.22 gb/s 8.53 gb/s 8.23 gb/s
nk_dot_i4_icelake 24.3 gb/s 36.3 gb/s 25.5 gb/s
u4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u4_serial 10.6 gb/s 12.0 gb/s 11.9 gb/s
nk_dot_u4_haswell 15.0 gb/s 16.4 gb/s 14.3 gb/s
nk_dot_u4_icelake 48.1 gb/s 64.4 gb/s 30.9 gb/s
u1 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u1_serial 3.92 gb/s 5.04 gb/s 4.97 gb/s
nk_dot_u1_haswell 14.7 gb/s 43.2 gb/s 69.4 gb/s
nk_dot_u1_icelake 17.9 gb/s 68.8 gb/s 110 gb/s

WASM

Measured with Wasmtime v42 (Cranelift backend).

Kernel 256 1024 4096
f64c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64c_serial 1.76 gb/s, 5.3 ulp 2.43 gb/s, 3.3 ulp 0.28 gb/s, 2.9 ulp
nk_vdot_f64c_serial 1.02 gb/s, 3.5 ulp 2.44 gb/s, 5.5 ulp 0.15 gb/s, 2.2 ulp
nk_dot_f64c_v128relaxed 2.80 gb/s, 37.8 ulp 3.01 gb/s, 34.9 ulp 0.21 gb/s, 167 ulp
nk_vdot_f64c_v128relaxed 2.06 gb/s, 20.1 ulp 2.87 gb/s, 51.4 ulp 0.04 gb/s, 57.2 ulp
f32c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32c_serial 1.26 gb/s, 0 ulp 1.74 gb/s, 0 ulp 0.08 gb/s, 0 ulp
nk_vdot_f32c_serial 1.13 gb/s, 0 ulp 1.78 gb/s, 0 ulp 0.21 gb/s, 0 ulp
nk_dot_f32c_v128relaxed 1.62 gb/s, 0 ulp 1.92 gb/s, 0 ulp 0.20 gb/s, 0 ulp
nk_vdot_f32c_v128relaxed 1.66 gb/s, 0 ulp 1.69 gb/s, 0 ulp 0.13 gb/s, 0 ulp
bf16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16c_serial 1.00 gb/s, 0.1 ulp 2.29 gb/s, 1.7 ulp 0.08 gb/s, 7.9 ulp
nk_vdot_bf16c_serial 0.581 gb/s, 0.1 ulp 0.919 gb/s, 2.9 ulp 0.30 gb/s, 11.2 ulp
f16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16c_serial 1.39 gb/s, 12.6 ulp 0.759 gb/s, 22.2 ulp 0.23 gb/s, 34 ulp
nk_vdot_f16c_serial 1.11 gb/s, 14.4 ulp 0.828 gb/s, 41.8 ulp 0.02 gb/s, 34.2 ulp
f64 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64_serial 1.53 gb/s, 3.1 ulp 1.72 gb/s, 2.5 ulp 0.20 gb/s, 3.3 ulp
nk_dot_f64_v128relaxed 2.62 gb/s, 3.2 ulp 2.11 gb/s, 3.6 ulp 0.28 gb/s, 3.8 ulp
f32 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32_serial 1.95 gb/s, 0 ulp 1.89 gb/s, 0 ulp 0.28 gb/s, 0 ulp
nk_dot_f32_v128relaxed 0.083 gb/s, 0 ulp 1.61 gb/s, 0 ulp 1.37 gb/s, 0 ulp
bf16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16_serial 2.90 gb/s, 0 ulp 2.27 gb/s, 0.5 ulp 0.22 gb/s, 5.2 ulp
nk_dot_bf16_v128relaxed 0.521 gb/s, 0 ulp 2.30 gb/s, 0.3 ulp 0.30 gb/s, 2.4 ulp
f16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16_serial 0.648 gb/s, 13.5 ulp 0.712 gb/s, 32 ulp 0.08 gb/s, 59.7 ulp
nk_dot_f16_v128relaxed 1.58 gb/s, 7.0 ulp 1.05 gb/s, 30.8 ulp 0.09 gb/s, 65.1 ulp
e5m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e5m2_serial 1.27 gb/s, 0 ulp 0.679 gb/s, 0 ulp 0.10 gb/s, 0 ulp
nk_dot_e5m2_v128relaxed 0.970 gb/s, 0 ulp 0.955 gb/s, 0 ulp 0.17 gb/s, 0 ulp
e4m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e4m3_serial 0.312 gb/s, 0 ulp 0.342 gb/s, 0 ulp 0.12 gb/s, 0 ulp
nk_dot_e4m3_v128relaxed 1.05 gb/s, 0 ulp 0.721 gb/s, 0 ulp 0.30 gb/s, 0 ulp
e3m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e3m2_serial 0.565 gb/s, 0 ulp 0.552 gb/s, 0 ulp 0.06 gb/s, 0 ulp
nk_dot_e3m2_v128relaxed 0.670 gb/s, 0 ulp 2.91 gb/s, 0 ulp 0.24 gb/s, 0 ulp
e2m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e2m3_serial 0.584 gb/s, 0 ulp 0.661 gb/s, 0 ulp 0.07 gb/s, 0 ulp
nk_dot_e2m3_v128relaxed 2.69 gb/s, 0 ulp 0.131 gb/s, 0 ulp 0.09 gb/s, 0 ulp
i8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i8_serial 1.17 gb/s 1.17 gb/s 0.29 gb/s
nk_dot_i8_v128relaxed 1.71 gb/s 0.896 gb/s 0.24 gb/s
u8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u8_serial 1.16 gb/s 0.658 gb/s 0.30 gb/s
nk_dot_u8_v128relaxed 0.873 gb/s 0.997 gb/s 0.15 gb/s
i4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i4_serial 0.217 gb/s 0.226 gb/s 0.28 gb/s
nk_dot_i4_v128relaxed 1.53 gb/s 2.87 gb/s 0.24 gb/s
u4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u4_serial 0.303 gb/s 0.250 gb/s 0.003 gb/s
nk_dot_u4_v128relaxed 0.126 gb/s 2.70 gb/s 0.08 gb/s
u1 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u1_serial 1.95 gb/s 1.53 gb/s 0.09 gb/s
nk_dot_u1_v128relaxed 0.548 gb/s 1.88 gb/s 0.13 gb/s

Apple M5

Native

Kernel 256 1024 4096
f64c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64c_serial 8.02 gb/s, 5 ulp 7.30 gb/s, 3 ulp 7.25 gb/s, 9.7 ulp
nk_vdot_f64c_serial 8.29 gb/s, 4.2 ulp 7.53 gb/s, 3.3 ulp 7.38 gb/s, 3.3 ulp
nk_dot_f64c_neon 23.7 gb/s, 0 ulp 21.6 gb/s, 0 ulp 21.3 gb/s, 0 ulp
nk_vdot_f64c_neon 23.6 gb/s, 0 ulp 21.8 gb/s, 0 ulp 20.9 gb/s, 0 ulp
f32c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32c_serial 27.8 gb/s, 0 ulp 24.6 gb/s, 0 ulp 23.2 gb/s, 0 ulp
nk_vdot_f32c_serial 27.2 gb/s, 0 ulp 24.0 gb/s, 0 ulp 22.6 gb/s, 0 ulp
nk_dot_f32c_neon 22.8 gb/s, 0 ulp 18.2 gb/s, 0 ulp 16.9 gb/s, 0 ulp
nk_vdot_f32c_neon 22.7 gb/s, 0 ulp 17.5 gb/s, 0 ulp 16.7 gb/s, 0 ulp
bf16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16c_serial 15.6 gb/s, 0.2 ulp 12.5 gb/s, 2.8 ulp 12.5 gb/s, 15.8 ulp
nk_vdot_bf16c_serial 15.9 gb/s, 0.2 ulp 12.9 gb/s, 2.6 ulp 11.7 gb/s, 11.4 ulp
nk_dot_bf16c_neonbfdot 26.3 gb/s, 0.1 ulp 18.5 gb/s, 2 ulp 17.6 gb/s, 8.8 ulp
nk_vdot_bf16c_neonbfdot 26.5 gb/s, 0.1 ulp 18.2 gb/s, 1.8 ulp 17.3 gb/s, 8.8 ulp
f16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16c_serial 15.8 gb/s, 20.8 ulp 13.0 gb/s, 64.1 ulp 12.3 gb/s, 73.1 ulp
nk_vdot_f16c_serial 15.8 gb/s, 24.8 ulp 13.0 gb/s, 31.9 ulp 12.3 gb/s, 137 ulp
nk_dot_f16c_neonhalf 26.1 gb/s, 3.0 ulp 18.4 gb/s, 6.5 ulp 16.8 gb/s, 20.5 ulp
nk_vdot_f16c_neonhalf 26.1 gb/s, 34.9 ulp 18.5 gb/s, 40.7 ulp 17.0 gb/s, 73.1 ulp
nk_dot_f16c_neonfhm 25.3 gb/s, 3.0 ulp 17.1 gb/s, 6.5 ulp 15.9 gb/s, 20.5 ulp
nk_vdot_f16c_neonfhm 25.0 gb/s, 31.4 ulp 17.0 gb/s, 38.6 ulp 15.8 gb/s, 67.6 ulp
f64 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64_serial 8.11 gb/s, 2.4 ulp 8.13 gb/s, 175 ulp 8.09 gb/s, 2.7 ulp
nk_dot_f64_neon 44.2 gb/s, 0 ulp 42.3 gb/s, 0 ulp 38.4 gb/s, 0 ulp
f32 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32_serial 23.3 gb/s, 0 ulp 15.8 gb/s, 0 ulp 14.6 gb/s, 0 ulp
nk_dot_f32_neon 46.4 gb/s, 0 ulp 38.0 gb/s, 0 ulp 34.8 gb/s, 0 ulp
bf16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16_serial 12.4 gb/s, 0 ulp 8.59 gb/s, 0.9 ulp 7.36 gb/s, 6 ulp
nk_dot_bf16_neon 39.0 gb/s, 3.7 ulp 27.2 gb/s, 3.7 ulp 19.9 gb/s, 3.7 ulp
nk_dot_bf16_neonbfdot 70.8 gb/s, 0 ulp 60.8 gb/s, 0.6 ulp 47.8 gb/s, 4.5 ulp
f16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16_serial 12.0 gb/s, 19 ulp 8.33 gb/s, 31.1 ulp 7.11 gb/s, 57.8 ulp
nk_dot_f16_neon 35.7 gb/s, 33.4 ulp 25.8 gb/s, 37.4 ulp 21.3 gb/s, 23.1 ulp
nk_dot_f16_neonfhm 48.7 gb/s, 14.9 ulp 27.5 gb/s, 26.7 ulp 18.8 gb/s, 39.9 ulp
e5m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e5m2_serial 3.80 gb/s, 0 ulp 3.41 gb/s, 0 ulp 3.41 gb/s, 0 ulp
nk_dot_e5m2_neon 19.0 gb/s, 0 ulp 13.2 gb/s, 0 ulp 10.5 gb/s, 0 ulp
nk_dot_e5m2_neonfhm 25.6 gb/s, 0 ulp 15.3 gb/s, 0 ulp 9.55 gb/s, 0 ulp
nk_dot_e5m2_neonbfdot 3.65 gb/s, 0 ulp 3.82 gb/s, 0 ulp 3.68 gb/s, 0 ulp
e4m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e4m3_serial 1.74 gb/s, 0 ulp 1.72 gb/s, 0 ulp 1.71 gb/s, 0 ulp
nk_dot_e4m3_neon 4.44 gb/s, 0 ulp 4.51 gb/s, 0 ulp 4.57 gb/s, 0 ulp
nk_dot_e4m3_neonfhm 10.1 gb/s, 0 ulp 8.51 gb/s, 0 ulp 7.96 gb/s, 0 ulp
nk_dot_e4m3_neonbfdot 3.59 gb/s, 0 ulp 3.68 gb/s, 0 ulp 3.64 gb/s, 0 ulp
e3m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e3m2_serial 2.51 gb/s, 0 ulp 2.33 gb/s, 0 ulp 2.24 gb/s, 0 ulp
nk_dot_e3m2_neonsdot 20.5 gb/s, 0 ulp 20.7 gb/s, 0 ulp 20.1 gb/s, 0 ulp
e2m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e2m3_serial 2.54 gb/s, 0 ulp 2.27 gb/s, 0 ulp 2.29 gb/s, 0 ulp
nk_dot_e2m3_neonsdot 47.3 gb/s, 0 ulp 47.5 gb/s, 0 ulp 43.4 gb/s, 0 ulp
i8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i8_serial 115 gb/s 102 gb/s 92.3 gb/s
nk_dot_i8_neonsdot 92.8 gb/s 87.4 gb/s 59.9 gb/s
u8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u8_serial 110 gb/s 99.2 gb/s 94.9 gb/s
nk_dot_u8_neonsdot 92.5 gb/s 86.6 gb/s 59.5 gb/s
i4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i4_serial 23.0 gb/s 24.4 gb/s 24.2 gb/s
nk_dot_i4_neonsdot 58.2 gb/s 44.7 gb/s 30.4 gb/s
u4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u4_serial 25.3 gb/s 27.2 gb/s 26.9 gb/s
nk_dot_u4_neonsdot 67.3 gb/s 47.4 gb/s 29.4 gb/s
u1 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u1_serial 7.01 gb/s 7.63 gb/s 7.19 gb/s
nk_dot_u1_neon 33.3 gb/s 64.6 gb/s 88.0 gb/s

WASM

Measured with Wasmtime v43 (Cranelift backend).

Kernel 256 1024 4096
f64c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64c_serial 4.17 gb/s, 3.8 ulp 6.03 gb/s, 3.9 ulp 6.37 gb/s, 3.2 ulp
nk_vdot_f64c_serial 6.00 gb/s, 3.8 ulp 6.55 gb/s, 3.4 ulp 6.83 gb/s, 15.1 ulp
nk_dot_f64c_v128relaxed 46.7 gb/s, 26 ulp 38.2 gb/s, 42 ulp 40.5 gb/s, 88 ulp
nk_vdot_f64c_v128relaxed 46.1 gb/s, 22.8 ulp 39.8 gb/s, 37.3 ulp 39.9 gb/s, 43.6 ulp
f32c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32c_serial 20.9 gb/s, 0 ulp 21.3 gb/s, 0 ulp 22.5 gb/s, 0 ulp
nk_vdot_f32c_serial 19.9 gb/s, 0 ulp 21.5 gb/s, 0 ulp 22.4 gb/s, 0 ulp
nk_dot_f32c_v128relaxed 22.4 gb/s, 0 ulp 20.3 gb/s, 0 ulp 19.8 gb/s, 0 ulp
nk_vdot_f32c_v128relaxed 21.9 gb/s, 0 ulp 20.3 gb/s, 0 ulp 19.9 gb/s, 0 ulp
bf16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16c_serial 10.3 gb/s, 0.1 ulp 11.6 gb/s, 2.5 ulp 11.6 gb/s, 10 ulp
nk_vdot_bf16c_serial 10.7 gb/s, 0.2 ulp 11.7 gb/s, 2.1 ulp 11.6 gb/s, 11.4 ulp
f16c ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16c_serial 3.76 gb/s, 13 ulp 3.83 gb/s, 20 ulp 3.86 gb/s, 90 ulp
nk_vdot_f16c_serial 3.81 gb/s, 13.9 ulp 3.89 gb/s, 35.5 ulp 3.85 gb/s, 42.4 ulp
f64 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f64_serial 7.26 gb/s, 2.4 ulp 7.45 gb/s, 2.6 ulp 7.96 gb/s, 2.2 ulp
nk_dot_f64_v128relaxed 38.7 gb/s, 2.6 ulp 42.0 gb/s, 3.2 ulp 43.9 gb/s, 2.6 ulp
f32 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f32_serial 19.0 gb/s, 16 ulp 14.6 gb/s, 69 ulp 14.0 gb/s, 104 ulp
nk_dot_f32_v128relaxed 20.4 gb/s, 0 ulp 18.9 gb/s, 0 ulp 18.7 gb/s, 0 ulp
bf16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_bf16_serial 9.53 gb/s, 0 ulp 7.42 gb/s, 0.6 ulp 7.20 gb/s, 5.9 ulp
nk_dot_bf16_v128relaxed 41.9 gb/s, 0 ulp 28.3 gb/s, 0.4 ulp 21.5 gb/s, 3.7 ulp
f16 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_f16_serial 3.31 gb/s, 16 ulp 3.63 gb/s, 26 ulp 3.66 gb/s, 53 ulp
nk_dot_f16_v128relaxed 11.4 gb/s, 9.0 ulp 11.2 gb/s, 23 ulp 12.0 gb/s, 39 ulp
e5m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e5m2_serial 3.02 gb/s, 0 ulp 2.95 gb/s, 0 ulp 3.16 gb/s, 0 ulp
nk_dot_e5m2_v128relaxed 3.47 gb/s, 0 ulp 3.45 gb/s, 0 ulp 3.48 gb/s, 0 ulp
e4m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e4m3_serial 0.978 gb/s, 0 ulp 0.893 gb/s, 0 ulp 0.936 gb/s, 0 ulp
nk_dot_e4m3_v128relaxed 2.78 gb/s, 0 ulp 2.75 gb/s, 0 ulp 2.78 gb/s, 0 ulp
e3m2 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e3m2_serial 3.13 gb/s, 0 ulp 2.95 gb/s, 0 ulp 3.16 gb/s, 0 ulp
nk_dot_e3m2_v128relaxed 12.1 gb/s, 0 ulp 11.9 gb/s, 0 ulp 12.6 gb/s, 0 ulp
e2m3 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_e2m3_serial 2.99 gb/s, 0 ulp 3.00 gb/s, 0 ulp 3.17 gb/s, 0 ulp
nk_dot_e2m3_v128relaxed 20.4 gb/s, 0 ulp 20.6 gb/s, 0 ulp 21.7 gb/s, 0 ulp
i8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i8_serial 22.2 gb/s 19.6 gb/s 17.8 gb/s
nk_dot_i8_v128relaxed 42.0 gb/s 49.0 gb/s 49.7 gb/s
u8 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u8_serial 23.0 gb/s 19.9 gb/s 17.8 gb/s
nk_dot_u8_v128relaxed 29.3 gb/s 33.0 gb/s 35.1 gb/s
i4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_i4_serial 0.990 gb/s 0.923 gb/s 0.985 gb/s
nk_dot_i4_v128relaxed 15.2 gb/s 17.9 gb/s 19.2 gb/s
u4 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u4_serial 0.992 gb/s 0.933 gb/s 0.988 gb/s
nk_dot_u4_v128relaxed 30.2 gb/s 32.1 gb/s 33.8 gb/s
u1 ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░ ░░░░░░░░░░░░░░░░░░░░░░░░
nk_dot_u1_serial 5.26 gb/s 5.80 gb/s 6.48 gb/s
nk_dot_u1_v128relaxed 21.2 gb/s 47.4 gb/s 67.3 gb/s