pytorch - 💡(How to fix) Fix `F.conv2d` CPU performance regression (~4–16×) for bfloat16 inputs introduced in 2.13.0.

import torch
import torch.nn.functional as F
import time
import statistics

def bench_cpu(fn, n=30):
    fn()  # warmup
    return statistics.median(
        [(lambda t0: (fn(), time.perf_counter() - t0)[1])(time.perf_counter())
         for _ in range(n)]
    ) * 1e3

torch.manual_seed(0)

# Shape A: dilated 7×7, NCHW, bfloat16
inp_a = torch.randn(8, 16, 112, 112, dtype=torch.bfloat16)
w_a   = torch.randn(32, 16, 7, 7,    dtype=torch.bfloat16)
t_a   = bench_cpu(lambda: F.conv2d(inp_a, w_a, stride=1, padding=0, dilation=2))
print(f"Shape A: {t_a:.2f} ms")   # expect ~1.9 ms; regressed to ~29 ms on May 12

# Shape B: grouped 3×3, channels_last, bfloat16
inp_b = torch.randn(8, 256, 56, 56, dtype=torch.bfloat16).to(memory_format=torch.channels_last)
w_b   = torch.randn(512, 32, 3, 3,  dtype=torch.bfloat16).to(memory_format=torch.channels_last)
t_b   = bench_cpu(lambda: F.conv2d(inp_b, w_b, stride=1, padding=1, groups=8))
print(f"Shape B: {t_b:.2f} ms")   # expect ~3.0 ms; regressed to ~12 ms on May 12

---

PyTorch version: 2.13.0.dev20260512+cu130
Python:          3.12.3 (GCC 13.3.0)
OS:              Linux 6.14.0-37-generic x86_64 (Ubuntu, glibc 2.39)
numpy:           2.4.4
GPU:             NVIDIA GeForce RTX 5090 (sm_120)
CUDA:            13.0