benchmark: batched matmul with scale factor

benchmark.py

import torch
from torch import einsum, tensor, matmul, bmm, baddbmm, empty
import time

scale=2
repeats = 10

# both einsum 0s use the same plan, so whichever batch runs first has to pay the price of warmup
# uncomment this to run a warmup before either batch runs, for fairer comparison of batch avg time
# q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
# k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
# start = time.perf_counter()
# (einsum('b i d, b j d -> b i j', q, k) * scale).max().item()
# duration = time.perf_counter()-start
# print('einsum 0 warmup took %.4f seconds' % (duration))

batch_duration = 0
for ix in range(repeats):
  q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  start = time.perf_counter()
  (einsum('b i d, b j d -> b i j', q, k) * scale).max().item()
  duration = time.perf_counter()-start
  print('einsum 0 iteration %d took %.4f seconds' % (ix, duration))
  batch_duration += duration
print('%d iterations of einsum 0 took %.4f seconds; avg %.4f secs' % (repeats, batch_duration, batch_duration/repeats))

batch_duration = 0
for ix in range(repeats):
  q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  start = time.perf_counter()
  (einsum('b n m, b m p -> b n p', q, k.transpose(1, 2)) * scale).max().item()
  duration = time.perf_counter()-start
  print('einsum 0 transposed k iteration %d took %.4f seconds' % (ix, duration))
  batch_duration += duration
print('%d iterations of einsum 0 transposed k took %.4f seconds; avg %.4f secs' % (repeats, batch_duration, batch_duration/repeats))

batch_duration = 0
for ix in range(repeats):
  q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  start = time.perf_counter()
  (matmul(q, k.transpose(1, 2)) * scale).max().item()
  duration = time.perf_counter()-start
  print('matmul iteration %d took %.4f seconds' % (ix, duration))
  batch_duration += duration
print('%d iterations of matmul took %.4f seconds; avg %.4f secs' % (repeats, batch_duration, batch_duration/repeats))

batch_duration = 0
for ix in range(repeats):
  q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  start = time.perf_counter()
  (bmm(q, k.transpose(1, 2)) * scale).max().item()
  duration = time.perf_counter()-start
  print('bmm iteration %d took %.4f seconds' % (ix, duration))
  batch_duration += duration
print('%d iterations of bmm took %.4f seconds; avg %.4f secs' % (repeats, batch_duration, batch_duration/repeats))

e = empty((1, 1, 1), device='mps')

batch_duration = 0
for ix in range(repeats):
  q = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  k = torch.rand(16, 4096, 40, dtype=torch.float, device="mps")
  start = time.perf_counter()
  baddbmm(e, q, k.transpose(1, 2), alpha=scale, beta=0).max().item()
  duration = time.perf_counter()-start
  print('baddbmm iteration %d took %.4f seconds' % (ix, duration))
  batch_duration += duration
print('%d iterations of baddbmm took %.4f seconds; avg %.4f secs' % (repeats, batch_duration, batch_duration/repeats))

thanks for this. similar results here (base M1 16GB)
torch 1.12.1:

einsum 0 iteration 0 took 0.1881 seconds
einsum 0 iteration 1 took 0.0805 seconds
einsum 0 iteration 2 took 0.0803 seconds
einsum 0 iteration 3 took 0.0799 seconds
einsum 0 iteration 4 took 0.0802 seconds
einsum 0 iteration 5 took 0.0831 seconds
einsum 0 iteration 6 took 0.0817 seconds
einsum 0 iteration 7 took 0.0814 seconds
einsum 0 iteration 8 took 0.0810 seconds
einsum 0 iteration 9 took 0.0805 seconds
10 iterations of einsum 0 took 0.9165 seconds; avg 0.0917 secs
einsum 0 transposed k iteration 0 took 0.0811 seconds
einsum 0 transposed k iteration 1 took 0.0810 seconds
einsum 0 transposed k iteration 2 took 0.0796 seconds
einsum 0 transposed k iteration 3 took 0.0803 seconds
einsum 0 transposed k iteration 4 took 0.0803 seconds
einsum 0 transposed k iteration 5 took 0.0809 seconds
einsum 0 transposed k iteration 6 took 0.0808 seconds
einsum 0 transposed k iteration 7 took 0.0801 seconds
einsum 0 transposed k iteration 8 took 0.0803 seconds
einsum 0 transposed k iteration 9 took 0.0805 seconds
10 iterations of einsum 0 transposed k took 0.8051 seconds; avg 0.0805 secs
matmul iteration 0 took 0.0996 seconds
matmul iteration 1 took 0.0830 seconds
matmul iteration 2 took 0.0823 seconds
matmul iteration 3 took 0.0844 seconds
matmul iteration 4 took 0.0856 seconds
matmul iteration 5 took 0.0834 seconds
matmul iteration 6 took 0.0823 seconds
matmul iteration 7 took 0.0826 seconds
matmul iteration 8 took 0.0828 seconds
matmul iteration 9 took 0.0821 seconds
10 iterations of matmul took 0.8483 seconds; avg 0.0848 secs
bmm iteration 0 took 0.0794 seconds
bmm iteration 1 took 0.0796 seconds
bmm iteration 2 took 0.0798 seconds
bmm iteration 3 took 0.0812 seconds
bmm iteration 4 took 0.0800 seconds
bmm iteration 5 took 0.0805 seconds
bmm iteration 6 took 0.0894 seconds
bmm iteration 7 took 0.0887 seconds
bmm iteration 8 took 0.0802 seconds
bmm iteration 9 took 0.0799 seconds
10 iterations of bmm took 0.8188 seconds; avg 0.0819 secs
baddbmm iteration 0 took 0.0535 seconds
baddbmm iteration 1 took 0.0456 seconds
baddbmm iteration 2 took 0.0457 seconds
baddbmm iteration 3 took 0.0456 seconds
baddbmm iteration 4 took 0.0454 seconds
baddbmm iteration 5 took 0.0451 seconds
baddbmm iteration 6 took 0.0458 seconds
baddbmm iteration 7 took 0.0458 seconds
baddbmm iteration 8 took 0.0457 seconds
baddbmm iteration 9 took 0.0453 seconds
10 iterations of baddbmm took 0.4635 seconds; avg 0.0463 secs

torch 1.14.0.dev20221121:

einsum 0 iteration 0 took 0.6219 seconds
einsum 0 iteration 1 took 0.1123 seconds
einsum 0 iteration 2 took 0.1147 seconds
einsum 0 iteration 3 took 0.1134 seconds
einsum 0 iteration 4 took 0.1126 seconds
einsum 0 iteration 5 took 0.1126 seconds
einsum 0 iteration 6 took 0.1122 seconds
einsum 0 iteration 7 took 0.1137 seconds
einsum 0 iteration 8 took 0.1137 seconds
einsum 0 iteration 9 took 0.1126 seconds
10 iterations of einsum 0 took 1.6395 seconds; avg 0.1640 secs
einsum 0 transposed k iteration 0 took 0.1133 seconds
einsum 0 transposed k iteration 1 took 0.1125 seconds
einsum 0 transposed k iteration 2 took 0.1123 seconds
einsum 0 transposed k iteration 3 took 0.1128 seconds
einsum 0 transposed k iteration 4 took 0.1128 seconds
einsum 0 transposed k iteration 5 took 0.1127 seconds
einsum 0 transposed k iteration 6 took 0.1129 seconds
einsum 0 transposed k iteration 7 took 0.1128 seconds
einsum 0 transposed k iteration 8 took 0.1131 seconds
einsum 0 transposed k iteration 9 took 0.1126 seconds
10 iterations of einsum 0 transposed k took 1.1277 seconds; avg 0.1128 secs
matmul iteration 0 took 0.0979 seconds
matmul iteration 1 took 0.0800 seconds
matmul iteration 2 took 0.0808 seconds
matmul iteration 3 took 0.0812 seconds
matmul iteration 4 took 0.0794 seconds
matmul iteration 5 took 0.0792 seconds
matmul iteration 6 took 0.0794 seconds
matmul iteration 7 took 0.0796 seconds
matmul iteration 8 took 0.0790 seconds
matmul iteration 9 took 0.0791 seconds
10 iterations of matmul took 0.8157 seconds; avg 0.0816 secs
bmm iteration 0 took 0.0765 seconds
bmm iteration 1 took 0.0956 seconds
bmm iteration 2 took 0.0757 seconds
bmm iteration 3 took 0.0763 seconds
bmm iteration 4 took 0.0773 seconds
bmm iteration 5 took 0.0764 seconds
bmm iteration 6 took 0.0768 seconds
bmm iteration 7 took 0.0768 seconds
bmm iteration 8 took 0.0767 seconds
bmm iteration 9 took 0.0763 seconds
10 iterations of bmm took 0.7844 seconds; avg 0.0784 secs
baddbmm iteration 0 took 0.4508 seconds
baddbmm iteration 1 took 0.0462 seconds
baddbmm iteration 2 took 0.0442 seconds
baddbmm iteration 3 took 0.0445 seconds
baddbmm iteration 4 took 0.0442 seconds
baddbmm iteration 5 took 0.0453 seconds
baddbmm iteration 6 took 0.0450 seconds
baddbmm iteration 7 took 0.0452 seconds
baddbmm iteration 8 took 0.0448 seconds
baddbmm iteration 9 took 0.0453 seconds
10 iterations of baddbmm took 0.8555 seconds; avg 0.0855 secs