Note

Go to the end to download the full example code

# Mixed-precision and accuracy settings

We test various options of KeOps regarding accuracy of computations.

## Setup

```
output_filename = "accuracy"
import importlib
import os
import time
import numpy as np
import torch
from matplotlib import pyplot as plt
use_cuda = torch.cuda.is_available()
D = 3
```

Benchmark specifications:

```
MAXTIME = 10 if use_cuda else 1 # Max number of seconds before we break the loop
REDTIME = (
2 if use_cuda else 0.2
) # Decrease the number of runs if computations take longer than 2s...
# Number of samples that we'll loop upon
NS = [
100,
200,
500,
1000,
2000,
5000,
10000,
20000,
50000,
100000,
200000,
500000,
1000000,
]
```

Synthetic dataset.

```
def generate_samples(N, D, device, lang):
"""Create point clouds sampled non-uniformly on a sphere of diameter 1."""
if lang == "torch":
if device == "cuda":
torch.cuda.manual_seed_all(123)
else:
torch.manual_seed(123)
x = torch.rand((N, D), device=device, dtype=torch.float64)
y = torch.rand((N, D), device=device, dtype=torch.float64)
# Draw a random source signal:
b = torch.randn((N, 1), device=device, dtype=torch.float64)
else:
np.random.seed(1234)
x = np.random.randn(*((N, D)))
y = np.random.randn(*((N, D)))
b = np.random.randn(*((N,)))
return x, y, b
```

Define a simple RBF product, using the `pykeops.torch.LazyTensor`

wrapper:

```
from pykeops.torch import LazyTensor
def conv_lazytensor(x, y, b, dtype, dtype_acc, sum_scheme, use_fast_math):
backend = "GPU" if use_cuda else "CPU"
x_i = LazyTensor(x.unsqueeze(-2)) # (M, 1, D)
y_j = LazyTensor(y.unsqueeze(-3)) # (1, N, D)
K_ij = ((x_i - y_j) ** 2).sum(-1) # (M, N, 1)
S_ij = K_ij * b.unsqueeze(-3) # (M, N, 1) * (1, N, 1)
return S_ij.sum(
dim=1,
backend=backend,
dtype_acc=dtype_acc,
sum_scheme=sum_scheme,
use_fast_math=use_fast_math,
)
```

## Benchmarking loops

```
def benchmark(
Routine, dev, N, D, loops, lang, dtype, dtype_acc, sum_scheme, use_fast_math
):
"""Times a convolution on an N-by-N problem, and evaluate accuracy."""
device = torch.device(dev)
x_, y_, b_ = generate_samples(N, D, device, lang)
if dtype == "float16":
torch_dtype = torch.float16
if dtype == "float32":
torch_dtype = torch.float32
elif dtype == "float64":
torch_dtype = torch.float64
x, y, b = x_.to(torch_dtype), y_.to(torch_dtype), b_.to(torch_dtype)
# We simply benchmark a convolution
N0 = min(N, 100)
Routine(
x[:N0, :], y[:N0, :], b[:N0, :], dtype, dtype_acc, sum_scheme, use_fast_math
) # Warmup run, to compile and load everything
# timings
if loops > 0:
code = "out = Routine( x, y, b, dtype, dtype_acc, sum_scheme, use_fast_math ) "
t_0 = time.perf_counter() # Actual benchmark --------------------
if use_cuda:
torch.cuda.synchronize()
for i in range(loops):
exec(code, locals())
if use_cuda:
torch.cuda.synchronize()
elapsed = time.perf_counter() - t_0 # ---------------------------
elapsed /= loops
print(
"timing of {:3} NxN convolution(s), with N ={:7}: {:3}x{:3.6f}s".format(
loops, N, loops, elapsed / loops
)
)
else:
elapsed = np.NaN
# accuracy
ind = torch.randperm(y.shape[0])
M = min(
N, 1000
) # we evaluate accuracy on a subsample of outputs only because computations with full precisions are slow.
out = Routine(
x[:M, :], y[ind, :], b[ind, :], dtype, dtype_acc, sum_scheme, use_fast_math
)
ref_out = Routine(x_[:M, :], y_, b_, "float64", "float64", "kahan_scheme", False)
mean_err = (
(out.double() - ref_out.double()).abs().mean() / ref_out.double().abs().mean()
).item()
mean_err = float("NaN") if mean_err == 0 else mean_err
max_err = (
(out.double() - ref_out.double()).abs().max() / ref_out.double().abs().mean()
).item()
max_err = float("NaN") if max_err == 0 else max_err
print(
"accuracy of an MxN convolution, with M = {}, N ={:7}: mean err={:.1e}, max err={:.1e}".format(
M, N, mean_err, max_err
)
)
return elapsed, mean_err, max_err
def bench_config(
Routine, backend, dev, lang, dtype, dtype_acc, sum_scheme, use_fast_math
):
"""Times a convolution for an increasing number of samples."""
print(
"Backend : {}, Device : {}, dtype : {}, dtype_acc : {}, sum_scheme : {}, use_fast_math : {} -------------".format(
backend, dev, dtype, dtype_acc, sum_scheme, use_fast_math
)
)
times = []
mean_errs = []
max_errs = []
try:
Nloops = [100, 10, 1, 0]
nloops = Nloops.pop(0)
for n in NS:
elapsed, mean_err, max_err = benchmark(
Routine,
dev,
n,
D,
nloops,
lang,
dtype,
dtype_acc,
sum_scheme,
use_fast_math,
)
times.append(elapsed)
mean_errs.append(mean_err)
max_errs.append(max_err)
if nloops > 0:
if (nloops * elapsed > MAXTIME) or (
nloops * elapsed > REDTIME / 10 and nloops > 1
):
nloops = Nloops.pop(0)
except RuntimeError:
print("**\nMemory overflow !")
except IndexError:
print("**\nToo slow !")
fill_nans = (len(NS) - len(times)) * [np.nan]
return times + fill_nans, mean_errs + fill_nans, max_errs + fill_nans
def full_bench(title, routines):
backends = [backend for (_, backend, _, _, _, _, _) in routines]
print("Benchmarking : {} ===============================".format(title))
lines_times = [NS]
lines_mean_errs = [NS]
lines_max_errs = [NS]
for routine, backend, lang, dtype, dtype_acc, sum_scheme, use_fast_math in routines:
res = bench_config(
routine,
backend,
"cuda" if use_cuda else "cpu",
lang,
dtype,
dtype_acc,
sum_scheme,
use_fast_math,
)
lines_times.append(res[0])
lines_mean_errs.append(res[1])
lines_max_errs.append(res[2])
benches_times = np.array(lines_times).T
benches_mean_errs = np.array(lines_mean_errs).T
benches_max_errs = np.array(lines_max_errs).T
for ind_benches, benches in enumerate(
(benches_times, benches_mean_errs, benches_max_errs)
):
# Creates a pyplot figure:
plt.figure(figsize=(12, 8))
linestyles = [
"o-",
"s-",
"^-",
"<-",
">-",
"v-",
"1-",
"+-",
"*-",
"x-",
"p-",
"d-",
]
for i, config in enumerate(routines):
plt.plot(
benches[:, 0],
benches[:, i + 1],
linestyles[i],
linewidth=2,
label='config = "{}"'.format(config[3:]),
)
plt.xlabel("Number of samples")
if ind_benches == 0:
plt.title("Runtimes for {} in dimension {}".format(title, D))
plt.ylabel("Seconds")
elif ind_benches == 1:
plt.title("Mean errors for {} in dimension {}".format(title, D))
plt.ylabel("Relative mean error")
elif ind_benches == 2:
plt.title("Max errors for {} in dimension {}".format(title, D))
plt.ylabel("Relative max error")
plt.yscale("log")
plt.xscale("log")
plt.legend(loc="upper left")
plt.grid(True, which="major", linestyle="-")
plt.grid(True, which="minor", linestyle="dotted")
true_vals = benches[:, 1:].flatten()
true_vals = true_vals[np.isfinite(true_vals)]
if ind_benches == 0:
plt.axis([NS[0], NS[-1], true_vals.min(), MAXTIME])
else:
plt.axis([NS[0], NS[-1], true_vals.min(), 100 * true_vals.max()])
plt.tight_layout()
# Save as a .csv to put a nice Tikz figure in the papers:
header = "Npoints " + " ".join(backends)
os.makedirs("output", exist_ok=True)
np.savetxt(
"output/" + output_filename + "_" + str(ind_benches) + ".csv",
benches,
fmt="%-9.5f",
header=header,
comments="",
)
```

## KeOps

```
routines = [
(
conv_lazytensor,
"float16, direct_sum",
"torch",
"float16",
"float16",
"direct_sum",
True,
),
(
conv_lazytensor,
"float16, block_sum",
"torch",
"float16",
"float16",
"block_sum",
True,
),
(
conv_lazytensor,
"float16, kahan_scheme",
"torch",
"float16",
"float16",
"kahan_scheme",
True,
),
(
conv_lazytensor,
"float16, float32 acc",
"torch",
"float16",
"float32",
"block_sum",
True,
),
(
conv_lazytensor,
"float32, direct_sum",
"torch",
"float32",
"float32",
"direct_sum",
True,
),
(
conv_lazytensor,
"float32, block_sum",
"torch",
"float32",
"float32",
"block_sum",
True,
),
(
conv_lazytensor,
"float32, block_sum",
"torch",
"float32",
"float32",
"block_sum",
False,
),
(
conv_lazytensor,
"float32, kahan_scheme",
"torch",
"float32",
"float32",
"kahan_scheme",
True,
),
(
conv_lazytensor,
"float32, float64 acc",
"torch",
"float32",
"float64",
"block_sum",
True,
),
(
conv_lazytensor,
"float64, direct_sum",
"torch",
"float64",
"float64",
"direct_sum",
True,
),
(
conv_lazytensor,
"float64, block_sum",
"torch",
"float64",
"float64",
"block_sum",
True,
),
(
conv_lazytensor,
"float64, kahan_scheme",
"torch",
"float64",
"float64",
"kahan_scheme",
True,
),
]
full_bench(" Matrix-Vector products", routines)
plt.show()
```

```
Benchmarking : Matrix-Vector products ===============================
Backend : float16, direct_sum, Device : cuda, dtype : float16, dtype_acc : float16, sum_scheme : direct_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000005s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.5e-03, max err=6.9e-03
timing of 100 NxN convolution(s), with N = 200: 100x0.000005s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=3.6e-03, max err=1.7e-02
timing of 100 NxN convolution(s), with N = 500: 100x0.000005s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=4.2e-03, max err=2.9e-02
timing of 100 NxN convolution(s), with N = 1000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=6.1e-03, max err=3.6e-02
timing of 100 NxN convolution(s), with N = 2000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=4.6e-03, max err=3.2e-02
timing of 100 NxN convolution(s), with N = 5000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=8.6e-03, max err=7.6e-02
timing of 100 NxN convolution(s), with N = 10000: 100x0.000008s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=1.2e-02, max err=9.0e-02
timing of 100 NxN convolution(s), with N = 20000: 100x0.000011s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=1.3e-02, max err=9.4e-02
timing of 100 NxN convolution(s), with N = 50000: 100x0.000021s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=2.0e-02, max err=1.1e-01
timing of 10 NxN convolution(s), with N = 100000: 10x0.000510s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=6.0e-02, max err=4.3e-01
timing of 10 NxN convolution(s), with N = 200000: 10x0.001445s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=9.4e-02, max err=5.9e-01
timing of 10 NxN convolution(s), with N = 500000: 10x0.008434s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=8.7e-02, max err=6.3e-01
timing of 1 NxN convolution(s), with N =1000000: 1x0.316824s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=1.9e-01, max err=1.2e+00
Backend : float16, block_sum, Device : cuda, dtype : float16, dtype_acc : float16, sum_scheme : block_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000006s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.5e-03, max err=6.9e-03
timing of 100 NxN convolution(s), with N = 200: 100x0.000005s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=3.6e-03, max err=1.6e-02
timing of 100 NxN convolution(s), with N = 500: 100x0.000005s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=2.6e-03, max err=1.5e-02
timing of 100 NxN convolution(s), with N = 1000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=4.0e-03, max err=2.1e-02
timing of 100 NxN convolution(s), with N = 2000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=1.5e-03, max err=6.7e-03
timing of 100 NxN convolution(s), with N = 5000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.2e-03, max err=1.1e-02
timing of 100 NxN convolution(s), with N = 10000: 100x0.000008s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=1.8e-03, max err=1.0e-02
timing of 100 NxN convolution(s), with N = 20000: 100x0.000011s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=2.2e-03, max err=9.3e-03
timing of 100 NxN convolution(s), with N = 50000: 100x0.000021s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=2.1e-03, max err=1.0e-02
timing of 10 NxN convolution(s), with N = 100000: 10x0.000533s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=6.3e-03, max err=3.7e-02
timing of 10 NxN convolution(s), with N = 200000: 10x0.001534s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=8.1e-03, max err=3.7e-02
timing of 10 NxN convolution(s), with N = 500000: 10x0.008915s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=7.1e-03, max err=6.4e-02
timing of 1 NxN convolution(s), with N =1000000: 1x0.334264s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=1.4e-02, max err=1.4e-01
Backend : float16, kahan_scheme, Device : cuda, dtype : float16, dtype_acc : float16, sum_scheme : kahan_scheme, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000006s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.0e-03, max err=3.0e-03
timing of 100 NxN convolution(s), with N = 200: 100x0.000006s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=1.6e-03, max err=4.0e-03
timing of 100 NxN convolution(s), with N = 500: 100x0.000005s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=9.4e-04, max err=4.8e-03
timing of 100 NxN convolution(s), with N = 1000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=9.4e-04, max err=4.3e-03
timing of 100 NxN convolution(s), with N = 2000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=4.6e-04, max err=1.8e-03
timing of 100 NxN convolution(s), with N = 5000: 100x0.000007s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=6.6e-04, max err=2.9e-03
timing of 100 NxN convolution(s), with N = 10000: 100x0.000009s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=4.3e-04, max err=1.9e-03
timing of 100 NxN convolution(s), with N = 20000: 100x0.000013s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=5.3e-04, max err=2.6e-03
timing of 100 NxN convolution(s), with N = 50000: 100x0.000028s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=4.1e-04, max err=1.9e-03
timing of 10 NxN convolution(s), with N = 100000: 10x0.000697s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=1.2e-03, max err=5.5e-03
timing of 10 NxN convolution(s), with N = 200000: 10x0.002033s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=1.5e-03, max err=6.9e-03
timing of 1 NxN convolution(s), with N = 500000: 1x0.118587s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=5.1e-04, max err=2.3e-03
timing of 1 NxN convolution(s), with N =1000000: 1x0.443472s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=1.6e-03, max err=6.8e-03
Backend : float16, float32 acc, Device : cuda, dtype : float16, dtype_acc : float32, sum_scheme : block_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000006s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.5e-03, max err=6.9e-03
timing of 100 NxN convolution(s), with N = 200: 100x0.000005s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=3.6e-03, max err=1.6e-02
timing of 100 NxN convolution(s), with N = 500: 100x0.000005s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=2.6e-03, max err=1.5e-02
timing of 100 NxN convolution(s), with N = 1000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=4.0e-03, max err=2.2e-02
timing of 100 NxN convolution(s), with N = 2000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=1.5e-03, max err=6.8e-03
timing of 100 NxN convolution(s), with N = 5000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.1e-03, max err=9.7e-03
timing of 100 NxN convolution(s), with N = 10000: 100x0.000008s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=1.5e-03, max err=7.3e-03
timing of 100 NxN convolution(s), with N = 20000: 100x0.000011s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=1.9e-03, max err=9.0e-03
timing of 100 NxN convolution(s), with N = 50000: 100x0.000022s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=1.4e-03, max err=7.9e-03
timing of 10 NxN convolution(s), with N = 100000: 10x0.000523s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=4.5e-03, max err=2.4e-02
timing of 10 NxN convolution(s), with N = 200000: 10x0.001474s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=4.4e-03, max err=2.2e-02
timing of 10 NxN convolution(s), with N = 500000: 10x0.008531s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=1.4e-03, max err=6.7e-03
timing of 1 NxN convolution(s), with N =1000000: 1x0.320990s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=5.5e-03, max err=3.2e-02
Backend : float32, direct_sum, Device : cuda, dtype : float32, dtype_acc : float32, sum_scheme : direct_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.4e-07, max err=5.8e-07
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=4.2e-07, max err=1.7e-06
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=4.0e-07, max err=2.4e-06
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=6.4e-07, max err=3.8e-06
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=5.0e-07, max err=5.3e-06
timing of 100 NxN convolution(s), with N = 5000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=1.1e-06, max err=6.5e-06
timing of 100 NxN convolution(s), with N = 10000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=1.7e-06, max err=1.1e-05
timing of 100 NxN convolution(s), with N = 20000: 100x0.000008s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=1.7e-06, max err=1.1e-05
timing of 100 NxN convolution(s), with N = 50000: 100x0.000034s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=2.7e-06, max err=1.5e-05
timing of 10 NxN convolution(s), with N = 100000: 10x0.001018s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=6.1e-06, max err=4.3e-05
timing of 10 NxN convolution(s), with N = 200000: 10x0.003838s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=9.8e-06, max err=6.4e-05
timing of 1 NxN convolution(s), with N = 500000: 1x0.233670s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=1.1e-05, max err=1.1e-04
timing of 1 NxN convolution(s), with N =1000000: 1x0.909367s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=2.6e-05, max err=1.9e-04
Backend : float32, block_sum, Device : cuda, dtype : float32, dtype_acc : float32, sum_scheme : block_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.4e-07, max err=5.8e-07
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=4.2e-07, max err=1.8e-06
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=3.0e-07, max err=1.6e-06
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=4.3e-07, max err=2.5e-06
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=1.4e-07, max err=7.6e-07
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.8e-07, max err=1.5e-06
timing of 100 NxN convolution(s), with N = 10000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=2.2e-07, max err=1.1e-06
timing of 100 NxN convolution(s), with N = 20000: 100x0.000009s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=2.8e-07, max err=1.5e-06
timing of 100 NxN convolution(s), with N = 50000: 100x0.000033s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=2.7e-07, max err=1.8e-06
timing of 10 NxN convolution(s), with N = 100000: 10x0.000977s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=6.7e-07, max err=2.8e-06
timing of 10 NxN convolution(s), with N = 200000: 10x0.003618s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=9.1e-07, max err=4.5e-06
timing of 1 NxN convolution(s), with N = 500000: 1x0.218077s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=8.4e-07, max err=7.5e-06
timing of 1 NxN convolution(s), with N =1000000: 1x0.849226s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=2.1e-06, max err=1.5e-05
Backend : float32, block_sum, Device : cuda, dtype : float32, dtype_acc : float32, sum_scheme : block_sum, use_fast_math : False -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.4e-07, max err=5.8e-07
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=4.2e-07, max err=1.8e-06
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=3.0e-07, max err=1.6e-06
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=4.3e-07, max err=2.5e-06
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=1.4e-07, max err=7.6e-07
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.8e-07, max err=1.5e-06
timing of 100 NxN convolution(s), with N = 10000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=2.2e-07, max err=1.1e-06
timing of 100 NxN convolution(s), with N = 20000: 100x0.000009s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=2.8e-07, max err=1.5e-06
timing of 100 NxN convolution(s), with N = 50000: 100x0.000033s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=2.7e-07, max err=1.8e-06
timing of 10 NxN convolution(s), with N = 100000: 10x0.000977s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=6.7e-07, max err=2.8e-06
timing of 10 NxN convolution(s), with N = 200000: 10x0.003618s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=9.1e-07, max err=4.5e-06
timing of 1 NxN convolution(s), with N = 500000: 1x0.218071s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=8.4e-07, max err=7.5e-06
timing of 1 NxN convolution(s), with N =1000000: 1x0.846144s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=2.1e-06, max err=1.5e-05
Backend : float32, kahan_scheme, Device : cuda, dtype : float32, dtype_acc : float32, sum_scheme : kahan_scheme, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=8.6e-08, max err=3.2e-07
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=1.2e-07, max err=4.2e-07
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=1.2e-07, max err=6.3e-07
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=1.3e-07, max err=7.1e-07
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=5.4e-08, max err=3.6e-07
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=9.0e-08, max err=5.0e-07
timing of 100 NxN convolution(s), with N = 10000: 100x0.000007s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=6.3e-08, max err=3.2e-07
timing of 100 NxN convolution(s), with N = 20000: 100x0.000011s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=6.5e-08, max err=3.3e-07
timing of 100 NxN convolution(s), with N = 50000: 100x0.000047s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=5.0e-08, max err=2.7e-07
timing of 10 NxN convolution(s), with N = 100000: 10x0.001465s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=1.4e-07, max err=6.6e-07
timing of 10 NxN convolution(s), with N = 200000: 10x0.005598s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=1.4e-07, max err=6.3e-07
timing of 1 NxN convolution(s), with N = 500000: 1x0.342779s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=5.5e-08, max err=3.2e-07
timing of 1 NxN convolution(s), with N =1000000: 1x1.336671s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=2.0e-07, max err=1.0e-06
Backend : float32, float64 acc, Device : cuda, dtype : float32, dtype_acc : float64, sum_scheme : block_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=1.4e-07, max err=5.8e-07
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=4.2e-07, max err=1.8e-06
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=3.0e-07, max err=1.6e-06
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=4.3e-07, max err=2.5e-06
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=1.4e-07, max err=7.6e-07
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.6e-07, max err=1.5e-06
timing of 100 NxN convolution(s), with N = 10000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=1.8e-07, max err=1.0e-06
timing of 100 NxN convolution(s), with N = 20000: 100x0.000009s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=2.5e-07, max err=1.3e-06
timing of 100 NxN convolution(s), with N = 50000: 100x0.000033s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=1.8e-07, max err=7.4e-07
timing of 10 NxN convolution(s), with N = 100000: 10x0.000977s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=5.1e-07, max err=2.8e-06
timing of 10 NxN convolution(s), with N = 200000: 10x0.003622s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=5.0e-07, max err=2.8e-06
timing of 1 NxN convolution(s), with N = 500000: 1x0.218161s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=1.8e-07, max err=8.1e-07
timing of 1 NxN convolution(s), with N =1000000: 1x0.846391s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=6.9e-07, max err=3.7e-06
Backend : float64, direct_sum, Device : cuda, dtype : float64, dtype_acc : float64, sum_scheme : direct_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=2.1e-16, max err=1.3e-15
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=8.7e-16, max err=3.7e-15
timing of 100 NxN convolution(s), with N = 500: 100x0.000004s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=8.1e-16, max err=4.7e-15
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=1.2e-15, max err=6.0e-15
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=9.2e-16, max err=6.2e-15
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=2.0e-15, max err=1.4e-14
timing of 100 NxN convolution(s), with N = 10000: 100x0.000008s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=3.1e-15, max err=1.8e-14
timing of 100 NxN convolution(s), with N = 20000: 100x0.000012s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=3.1e-15, max err=2.1e-14
timing of 100 NxN convolution(s), with N = 50000: 100x0.000063s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=4.6e-15, max err=3.4e-14
timing of 10 NxN convolution(s), with N = 100000: 10x0.001977s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=1.1e-14, max err=6.4e-14
timing of 10 NxN convolution(s), with N = 200000: 10x0.007663s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=1.8e-14, max err=1.4e-13
timing of 1 NxN convolution(s), with N = 500000: 1x0.472913s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=2.1e-14, max err=1.3e-13
timing of 1 NxN convolution(s), with N =1000000: 1x1.843502s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=5.0e-14, max err=3.1e-13
Backend : float64, block_sum, Device : cuda, dtype : float64, dtype_acc : float64, sum_scheme : block_sum, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=2.1e-16, max err=1.3e-15
timing of 100 NxN convolution(s), with N = 200: 100x0.000003s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=8.7e-16, max err=3.7e-15
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=5.6e-16, max err=2.8e-15
timing of 100 NxN convolution(s), with N = 1000: 100x0.000003s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=7.7e-16, max err=3.5e-15
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=2.5e-16, max err=1.5e-15
timing of 100 NxN convolution(s), with N = 5000: 100x0.000005s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=4.8e-16, max err=2.6e-15
timing of 100 NxN convolution(s), with N = 10000: 100x0.000007s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=4.0e-16, max err=2.2e-15
timing of 100 NxN convolution(s), with N = 20000: 100x0.000012s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=5.0e-16, max err=3.0e-15
timing of 100 NxN convolution(s), with N = 50000: 100x0.000062s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=4.8e-16, max err=2.5e-15
timing of 10 NxN convolution(s), with N = 100000: 10x0.001976s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=1.3e-15, max err=5.8e-15
timing of 10 NxN convolution(s), with N = 200000: 10x0.007664s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=1.6e-15, max err=7.9e-15
timing of 1 NxN convolution(s), with N = 500000: 1x0.473102s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=1.5e-15, max err=1.3e-14
timing of 1 NxN convolution(s), with N =1000000: 1x1.841738s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=3.7e-15, max err=2.3e-14
Backend : float64, kahan_scheme, Device : cuda, dtype : float64, dtype_acc : float64, sum_scheme : kahan_scheme, use_fast_math : True -------------
timing of 100 NxN convolution(s), with N = 100: 100x0.000004s
accuracy of an MxN convolution, with M = 100, N = 100: mean err=7.9e-17, max err=4.5e-16
timing of 100 NxN convolution(s), with N = 200: 100x0.000004s
accuracy of an MxN convolution, with M = 200, N = 200: mean err=1.3e-16, max err=6.0e-16
timing of 100 NxN convolution(s), with N = 500: 100x0.000003s
accuracy of an MxN convolution, with M = 500, N = 500: mean err=1.2e-16, max err=5.5e-16
timing of 100 NxN convolution(s), with N = 1000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 1000: mean err=1.3e-16, max err=5.5e-16
timing of 100 NxN convolution(s), with N = 2000: 100x0.000004s
accuracy of an MxN convolution, with M = 1000, N = 2000: mean err=4.5e-17, max err=4.3e-16
timing of 100 NxN convolution(s), with N = 5000: 100x0.000006s
accuracy of an MxN convolution, with M = 1000, N = 5000: mean err=6.9e-17, max err=4.4e-16
timing of 100 NxN convolution(s), with N = 10000: 100x0.000010s
accuracy of an MxN convolution, with M = 1000, N = 10000: mean err=4.6e-17, max err=4.5e-16
timing of 100 NxN convolution(s), with N = 20000: 100x0.000016s
accuracy of an MxN convolution, with M = 1000, N = 20000: mean err=5.6e-17, max err=3.8e-16
timing of 100 NxN convolution(s), with N = 50000: 100x0.000088s
accuracy of an MxN convolution, with M = 1000, N = 50000: mean err=4.2e-17, max err=3.6e-16
timing of 10 NxN convolution(s), with N = 100000: 10x0.002785s
accuracy of an MxN convolution, with M = 1000, N = 100000: mean err=1.2e-16, max err=7.7e-16
timing of 1 NxN convolution(s), with N = 200000: 1x0.108043s
accuracy of an MxN convolution, with M = 1000, N = 200000: mean err=1.3e-16, max err=7.9e-16
timing of 1 NxN convolution(s), with N = 500000: 1x0.665048s
accuracy of an MxN convolution, with M = 1000, N = 500000: mean err=4.3e-17, max err=4.7e-16
timing of 1 NxN convolution(s), with N =1000000: 1x2.605435s
accuracy of an MxN convolution, with M = 1000, N =1000000: mean err=1.7e-16, max err=9.7e-16
```

**Total running time of the script:** (0 minutes 38.054 seconds)