Skip to content

Commit

Permalink
[mpact][benchmark] manual sum of squares benchmark (#65)
Browse files Browse the repository at this point in the history
* [mpact][benchmark] manual sum of squares benchmark

This introduces a "manual" benchmark where we can put
some benchmarking code but without negatively adding
more load on the regular benchmark suite times.

* use 4K instead of 1K

* lint

* undo edits
  • Loading branch information
aartbik authored Jul 30, 2024
1 parent 77bb923 commit f37a36c
Show file tree
Hide file tree
Showing 2 changed files with 68 additions and 0 deletions.
9 changes: 9 additions & 0 deletions benchmark/python/manual/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
### Benchmarks run by hand

These benchmarks are not run as part of MPACT's regular testing or benchmarking.
To run an individual test, build the MPACT compiler, cd into this directory,
and then simply run a benchmark as follows:

```shell
python <benchmark-name>.py
```
59 changes: 59 additions & 0 deletions benchmark/python/manual/sum_of_sq.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,59 @@
import torch
import numpy as np
import time

from mpact.mpactbackend import mpact_jit_compile, mpact_jit_run
from mpact_benchmark.utils.tensor_generator import generate_tensor


def runbench_eager(tag, sp, net, x, num_iters=1000):
net(x) # warmup
checksum = 0
start = time.time()
for i in range(num_iters):
res = net(x).item()
checksum = checksum + res
end = time.time()
time_ms = (end - start) * 1000 / num_iters
print("%s : %.2f : %8.4f ms. : checksum=%d" % (tag, sp, time_ms, checksum))


def runbench_mpact(tag, sp, net, x, num_iters=1000):
invoker, fn = mpact_jit_compile(net, x)
mpact_jit_run(invoker, fn, x) # warmup
checksum = 0
start = time.time()
for i in range(num_iters):
res = mpact_jit_run(invoker, fn, x)
checksum = checksum + res
end = time.time()
time_ms = (end - start) * 1000 / num_iters
print("%s : %.2f : %8.4f ms. : checksum=%d" % (tag, sp, time_ms, checksum))


class SqSumNet(torch.nn.Module):
def forward(self, x):
# TODO: make this work too: return (x ** 2).sum()
return (x * x).sum()


net = SqSumNet()
h = 1024 * 4
w = 1024 * 4

for d in range(0, 101, 10):
sparsity = 1.0 - (d / 100.0)
x = generate_tensor(
seed=0, shape=(h, w), sparsity=sparsity, dtype=np.float32, drange=(1.0, 1.0)
)

# Note, we don't have binary-valued sparse tensors in PyTorch
# so we are using csr. For now, we have to hack the
# "explicitVal=1.0:f32"
# into the MLIR sparse tensor type to make optimize it fully.
s = x.to_sparse_csr()

runbench_eager("PyTorch (dense) ", sparsity, net, x)
runbench_mpact("MPACT (dense) ", sparsity, net, x)
runbench_eager("PyTorch (sparse)", sparsity, net, s)
runbench_mpact("MPACT (sparse)", sparsity, net, s)

0 comments on commit f37a36c

Please sign in to comment.