
mtl5 5.9.7 — build flags: native_fast_gemm, highway_simd
x86_64, 16 thread(s), baseline f64

== dot ===============================================================
dtype              n         s/op    GFLOP/s     vs f64
-------------------------------------------------------
f64             1000    2.347e-07      8.523 >     1.0x
f64            10000    1.549e-06     12.910 >     1.0x
f64           100000    1.739e-05     11.500       1.0x
f32             1000    1.336e-07     14.971 >     0.6x
f32            10000    8.074e-07     24.770 >     0.5x
f32           100000    6.990e-06     28.610       0.4x
cfloat32        1000    3.085e-04      0.006 >  1314.8x
cfloat32       10000    3.188e-03      0.006 >  2057.8x
cfloat32      100000    3.158e-02      0.006    1815.9x
posit32         1000    1.199e-03      0.002 >  5110.9x
posit32        10000    1.219e-02      0.002 >  7869.2x
posit32       100000    1.210e-01      0.002    6959.4x
lns32           1000    9.966e-04      0.002 >  4246.9x
lns32          10000    1.004e-02      0.002 >  6480.3x
lns32         100000    1.005e-01      0.002    5781.6x
takum32         1000    1.393e-04      0.014 >   593.8x
takum32        10000    1.384e-03      0.014 >   893.6x
takum32       100000    1.400e-02      0.014     805.2x
dd_cascade      1000    9.325e-05      0.021 >   397.4x
dd_cascade     10000    9.376e-04      0.021 >   605.2x
dd_cascade    100000    9.620e-03      0.021     553.1x
td_cascade      1000    1.550e-04      0.013 >   660.5x
td_cascade     10000    1.530e-03      0.013 >   987.9x
td_cascade    100000    1.531e-02      0.013     880.1x
qd_cascade      1000    2.108e-04      0.009 >   898.4x
qd_cascade     10000    2.049e-03      0.010 >  1322.7x
qd_cascade    100000    2.056e-02      0.010    1181.9x
  > baseline is within 20x of per-call overhead — ratio is a LOWER bound; prefer the larger sizes

== gemv ==============================================================
dtype              n         s/op    GFLOP/s     vs f64
-------------------------------------------------------
f64               64    3.663e-06      2.236 >     1.0x
f64              256    4.747e-05      2.761       1.0x
f64              512    2.213e-04      2.370       1.0x
f32               64    3.588e-06      2.283 >     1.0x
f32              256    4.456e-05      2.941       0.9x
f32              512    1.987e-04      2.638       0.9x
cfloat32          64    1.304e-03      0.006 >   356.0x
cfloat32         256    2.094e-02      0.006     441.1x
cfloat32         512    8.340e-02      0.006     376.9x
posit32           64    4.980e-03      0.002 >  1359.5x
posit32          256    7.978e-02      0.002    1680.8x
posit32          512    3.192e-01      0.002    1442.6x
lns32             64    4.149e-03      0.002 >  1132.8x
lns32            256    6.648e-02      0.002    1400.5x
lns32            512    2.657e-01      0.002    1200.8x
takum32           64    6.359e-04      0.013 >   173.6x
takum32          256    1.002e-02      0.013     211.1x
takum32          512    3.890e-02      0.013     175.8x
dd_cascade        64    3.486e-04      0.023 >    95.2x
dd_cascade       256    5.648e-03      0.023     119.0x
dd_cascade       512    2.231e-02      0.023     100.8x
td_cascade        64    6.724e-04      0.012 >   183.6x
td_cascade       256    1.038e-02      0.013     218.6x
td_cascade       512    4.105e-02      0.013     185.6x
qd_cascade        64    9.025e-04      0.009 >   246.4x
qd_cascade       256    1.411e-02      0.009     297.2x
qd_cascade       512    5.585e-02      0.009     252.4x
  > baseline is within 20x of per-call overhead — ratio is a LOWER bound; prefer the larger sizes

== gemm ==============================================================
dtype              n         s/op    GFLOP/s     vs f64
-------------------------------------------------------
f64               32    9.016e-06      7.269 >     1.0x
f64               64    4.559e-05     11.500       1.0x
f64              128    1.500e-04     27.968       1.0x
f32               32    6.529e-06     10.037 >     0.7x
f32               64    3.159e-05     16.598       0.7x
f32              128    1.417e-04     29.604       0.9x
cfloat32          32    8.840e-03      0.007 >   980.5x
cfloat32          64    7.198e-02      0.007    1578.8x
cfloat32         128    6.030e-01      0.007    4020.5x
posit32           32    3.892e-02      0.002 >  4316.3x
posit32           64    3.138e-01      0.002    6882.2x
posit32          128     too slow          -          -
lns32             32    3.340e-02      0.002 >  3704.4x
lns32             64    2.665e-01      0.002    5845.7x
lns32            128     too slow          -          -
takum32           32    5.117e-03      0.013 >   567.6x
takum32           64    4.060e-02      0.013     890.5x
takum32          128    3.236e-01      0.013    2157.6x
dd_cascade        32    2.865e-03      0.023 >   317.7x
dd_cascade        64    2.324e-02      0.023     509.8x
dd_cascade       128    1.855e-01      0.023    1236.9x
td_cascade        32    5.627e-03      0.012 >   624.1x
td_cascade        64    4.425e-02      0.012     970.6x
td_cascade       128    3.490e-01      0.012    2327.1x
qd_cascade        32    7.466e-03      0.009 >   828.1x
qd_cascade        64    5.924e-02      0.009    1299.5x
qd_cascade       128    4.677e-01      0.009    3118.3x
  > baseline is within 20x of per-call overhead — ratio is a LOWER bound; prefer the larger sizes

== lu ================================================================
dtype              n         s/op    GFLOP/s     vs f64
-------------------------------------------------------
f64               32    6.859e-06      3.185       1.0x
f64               64    4.106e-05      4.256       1.0x
f64              128    2.874e-04      4.865       1.0x
f32               32    6.832e-06      3.198       1.0x
f32               64    4.387e-05      3.984       1.1x
f32              128    2.777e-04      5.035       1.0x
cfloat32          32    3.670e-03      0.006     535.0x
cfloat32          64    2.553e-02      0.007     621.7x
cfloat32         128    1.884e-01      0.007     655.5x
posit32           32    1.491e-02      0.001    2174.1x
posit32           64    1.102e-01      0.002    2683.4x
posit32          128    8.531e-01      0.002    2968.9x
lns32             32    1.096e-02      0.002    1597.5x
lns32             64    8.959e-02      0.002    2181.9x
lns32            128    7.267e-01      0.002    2528.8x
takum32           32    1.570e-03      0.014     228.9x
takum32           64    1.264e-02      0.014     307.7x
takum32          128    1.009e-01      0.014     351.2x
dd_cascade        32    8.578e-04      0.025     125.1x
dd_cascade        64    6.675e-03      0.026     162.6x
dd_cascade       128    5.218e-02      0.027     181.6x
td_cascade        32    1.150e-02      0.002    1676.1x
td_cascade        64    9.482e-02      0.002    2309.2x
td_cascade       128    7.704e-01      0.002    2680.9x
qd_cascade        32    2.895e-02      0.001    4220.4x
qd_cascade        64    2.416e-01      0.001    5884.8x
qd_cascade       128    1.991e+00      0.001    6929.1x

== qr ================================================================
dtype              n         s/op    GFLOP/s     vs f64
-------------------------------------------------------
f64               32    2.577e-05      1.696       1.0x
f64               64    1.737e-04      2.013       1.0x
f64              128    1.489e-03      1.878       1.0x
f32               32    2.450e-05      1.783       1.0x
f32               64    1.743e-04      2.005       1.0x
f32              128    1.380e-03      2.026       0.9x
cfloat32          32    1.095e-02      0.004     424.9x
cfloat32          64    7.625e-02      0.005     439.1x
cfloat32         128    5.729e-01      0.005     384.8x
posit32           32    4.380e-02      0.001    1700.0x
posit32           64    3.153e-01      0.001    1815.7x
posit32          128     too slow          -          -
lns32             32    2.408e-02      0.002     934.4x
lns32             64    1.883e-01      0.002    1084.6x
lns32            128    1.463e+00      0.002     982.4x
takum32           32    4.454e-03      0.010     172.9x
takum32           64    3.333e-02      0.010     191.9x
takum32          128    2.579e-01      0.011     173.2x
dd_cascade        32    2.918e-03      0.015     113.2x
dd_cascade        64    2.165e-02      0.016     124.6x
dd_cascade       128    1.661e-01      0.017     111.6x
td_cascade        32    3.762e-02      0.001    1460.1x
td_cascade        64    2.947e-01      0.001    1696.9x
td_cascade       128     too slow          -          -
qd_cascade        32    9.624e-02      0.000    3735.3x
qd_cascade        64    7.564e-01      0.000    4355.4x
qd_cascade       128     too slow          -          -
