Integrating the int64_t GEMM in FBGEMM into PyTorch Linear op (#30143)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30143
We would like to integrate the int64 GEMM in FBGEMM into PyTorch. This brings ~4x speedup for the Linear op with LongTensor.
Benchmark code:
```
from __future__ import absolute_import, division, print_function, unicode_literals
import time
import torch
torch.set_num_threads(1)
print("M, N, K, GOPS/sec")
for M in range(128, 1025, 128):
N = M
K = M
x = torch.LongTensor(M, K)
w = torch.LongTensor(K, N)
NITER = 20
# Test torch.nn.functional.linear
s = time.time()
for _ in range(NITER):
torch.nn.functional.linear(x, w)
# Z = x @ w
elapsed_per_iter_linear = (time.time() - s) / NITER
print(
"{}, {}, {}, {:0.2f}".format(M, N, K, 2.0 * M * N * K / elapsed_per_iter_linear / 1e9)
)
```
Before this PR:
```
M, N, K, GOPS/sec
128, 128, 128, 2.31
256, 256, 256, 2.49
384, 384, 384, 2.54
512, 512, 512, 2.57
640, 640, 640, 2.46
768, 768, 768, 2.59
896, 896, 896, 2.59
1024, 1024, 1024, 2.61
```
After this PR:
```
(base) [root@rtptest10054.frc2 ~/jhuang_test/int64_gemm]# python torch_linear.py
M, N, K, GOPS/sec
128, 128, 128, 5.35
256, 256, 256, 8.34
384, 384, 384, 9.03
512, 512, 512, 9.22
640, 640, 640, 9.55
768, 768, 768, 9.73
896, 896, 896, 9.82
1024, 1024, 1024, 9.63
```
ghstack-source-id: 94308012
Test Plan: CI
Reviewed By: jspark1105
Differential Revision: D18610019
fbshipit-source-id: f830660927b2666db34427d9de51db011f80f766
pytorch
b154a8cf - Integrating the int64_t GEMM in FBGEMM into PyTorch Linear op (#30143)
