pytorch
48d6e2ad - Disable the mkldnn for conv2d in some special cases (#40610)

Commit View On GitHub

Commit

4 years ago

Disable the mkldnn for conv2d in some special cases (#40610) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/40610 We have benchmarked several models, which shows the native implementation of conv2d is faster mkldnn path. For group conv, the native implementation does not batch all the groups. Test Plan: ``` import torch import torch.nn.functional as F import numpy as np from timeit import Timer num = 50 S = [ # [1, 1, 100, 40, 16, 3, 3, 1, 1, 1, 1], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], # [1, 2048, 4, 2, 512, 1, 1, 1, 1, 0, 0], # [1, 512, 4, 2, 512, 3, 3, 1, 1, 1, 1], # [1, 512, 4, 2, 2048, 1, 1, 1, 1, 0, 0], [1, 3, 224, 224, 64, 7, 7, 2, 2, 3, 3, 1], [1, 64, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1], [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32], [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1], [1, 64, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1], [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32], [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1], [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32], [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 56, 56, 256, 3, 3, 2, 2, 1, 1, 32], [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1], [1, 256, 56, 56, 512, 1, 1, 2, 2, 0, 0, 1], [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32], [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32], [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1], [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32], [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 28, 28, 512, 3, 3, 2, 2, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 512, 28, 28, 1024, 1, 1, 2, 2, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1], [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32], [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 1024, 3, 3, 2, 2, 1, 1, 32], [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 14, 14, 2048, 1, 1, 2, 2, 0, 0, 1], [1, 2048, 7, 7, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 7, 7, 1024, 3, 3, 1, 1, 1, 1, 32], [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1], [1, 2048, 7, 7, 1024, 1, 1, 1, 1, 0, 0, 1], [1, 1024, 7, 7, 1024, 3, 3, 1, 1, 1, 1, 32], [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1], ] for x in range(105): P = S[x] print(P) (N, C, H, W) = P[0:4] M = P[4] (kernel_h, kernel_w) = P[5:7] (stride_h, stride_w) = P[7:9] (padding_h, padding_w) = P[9:11] X_np = np.random.randn(N, C, H, W).astype(np.float32) W_np = np.random.randn(M, C, kernel_h, kernel_w).astype(np.float32) X = torch.from_numpy(X_np) g = P[11] conv2d_pt = torch.nn.Conv2d( C, M, (kernel_h, kernel_w), stride=(stride_h, stride_w), padding=(padding_h, padding_w), groups=g, bias=True) class ConvNet(torch.nn.Module): def __init__(self): super(ConvNet, self).__init__() self.conv2d = conv2d_pt def forward(self, x): return self.conv2d(x) model = ConvNet() def pt_forward(): with torch.no_grad(): model(X) torch._C._set_mkldnn_enabled(True) t = Timer("pt_forward()", "from __main__ import pt_forward, X") print("MKLDNN pt time = {}".format(t.timeit(num) / num * 1000.0)) torch._C._set_mkldnn_enabled(False) t = Timer("pt_forward()", "from __main__ import pt_forward, X") print("TH pt time = {}".format(t.timeit(num) / num * 1000.0)) OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 python bm.py ``` output: ``` [1, 3, 224, 224, 64, 7, 7, 2, 2, 3, 3, 1] MKLDNN pt time = 5.891108009964228 TH pt time = 7.0624795742332935 [1, 64, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 1.4464975893497467 TH pt time = 0.721491202712059 [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 1.4036639966070652 TH pt time = 3.299683593213558 [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.3908068016171455 TH pt time = 2.227546200156212 [1, 64, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.226586602628231 TH pt time = 1.3865559734404087 [1, 256, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.31307839602232 TH pt time = 2.4284918047487736 [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 1.5028003975749016 TH pt time = 3.824346773326397 [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.4405963867902756 TH pt time = 2.6227117888629436 [1, 256, 56, 56, 128, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.405764400959015 TH pt time = 2.644723802804947 [1, 128, 56, 56, 128, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 1.5220053866505623 TH pt time = 3.9365867897868156 [1, 128, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.606868200004101 TH pt time = 2.5387956015765667 [1, 256, 56, 56, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 6.0041105933487415 TH pt time = 5.305919591337442 [1, 256, 56, 56, 256, 3, 3, 2, 2, 1, 1, 32] MKLDNN pt time = 1.4830979891121387 TH pt time = 7.532084975391626 [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.025687597692013 TH pt time = 2.2185291908681393 [1, 256, 56, 56, 512, 1, 1, 2, 2, 0, 0, 1] MKLDNN pt time = 3.5893129743635654 TH pt time = 2.696530409157276 [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8203356079757214 TH pt time = 2.0819314010441303 [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.8583215996623039 TH pt time = 2.7761065773665905 [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9077288135886192 TH pt time = 2.045416794717312 [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.805021796375513 TH pt time = 2.131381593644619 [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.9023251943290234 TH pt time = 2.9028950072824955 [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.1174601800739765 TH pt time = 2.275596000254154 [1, 512, 28, 28, 256, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.100480604916811 TH pt time = 2.399571593850851 [1, 256, 28, 28, 256, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.9321337938308716 TH pt time = 2.886691205203533 [1, 256, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.065785188227892 TH pt time = 2.1640316024422646 [1, 512, 28, 28, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 5.891813579946756 TH pt time = 4.2956990003585815 [1, 512, 28, 28, 512, 3, 3, 2, 2, 1, 1, 32] MKLDNN pt time = 0.9399276040494442 TH pt time = 4.7622935846447945 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.2426914013922215 TH pt time = 2.3699573799967766 [1, 512, 28, 28, 1024, 1, 1, 2, 2, 0, 0, 1] MKLDNN pt time = 3.0341636016964912 TH pt time = 2.6606030017137527 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.991385366767645 TH pt time = 2.6313263922929764 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7330256141722202 TH pt time = 3.008321188390255 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.880081795156002 TH pt time = 2.289068605750799 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9583285935223103 TH pt time = 2.6302105747163296 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7322711870074272 TH pt time = 2.8230775892734528 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8620235808193684 TH pt time = 2.4078205972909927 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.828651014715433 TH pt time = 2.616014201194048 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7084695994853973 TH pt time = 2.8024527989327908 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7884829975664616 TH pt time = 2.4237345717847347 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.89030060172081 TH pt time = 2.5852439925074577 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.724627785384655 TH pt time = 2.651805803179741 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.249914798885584 TH pt time = 2.0440668053925037 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.722136974334717 TH pt time = 2.531316000968218 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7164162024855614 TH pt time = 2.8521843999624252 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8891782090067863 TH pt time = 2.436912599951029 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.0049769952893257 TH pt time = 2.649025786668062 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7299130037426949 TH pt time = 2.67714099958539 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.799382768571377 TH pt time = 2.4427592009305954 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.0201382003724575 TH pt time = 2.6285660080611706 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.6983320042490959 TH pt time = 2.9118607938289642 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8802538104355335 TH pt time = 2.385452575981617 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9600497893989086 TH pt time = 2.594646792858839 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.5688861943781376 TH pt time = 2.5941073894500732 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7758505940437317 TH pt time = 2.336081601679325 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.6135251857340336 TH pt time = 2.3902921937406063 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.6303061917424202 TH pt time = 2.6228136010468006 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8868251852691174 TH pt time = 2.5620524026453495 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.057632204145193 TH pt time = 2.691414188593626 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7316274009644985 TH pt time = 3.14683198928833 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.2674955762922764 TH pt time = 2.602821197360754 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.0993166007101536 TH pt time = 2.609328981488943 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7257938012480736 TH pt time = 2.9255208000540733 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.3086097799241543 TH pt time = 2.544360812753439 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.0537622049450874 TH pt time = 2.6343842037022114 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7194169983267784 TH pt time = 2.9009717889130116 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.6461398042738438 TH pt time = 2.3600555770099163 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.6328082010149956 TH pt time = 2.415131386369467 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.6832938082516193 TH pt time = 2.6299685798585415 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9594415985047817 TH pt time = 2.509857602417469 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.956229578703642 TH pt time = 2.691046390682459 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7222409918904305 TH pt time = 2.938339803367853 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9467295855283737 TH pt time = 2.4219116009771824 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.0215882137417793 TH pt time = 2.7782391756772995 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.719242412596941 TH pt time = 2.8529402054846287 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8062099777162075 TH pt time = 2.9951974004507065 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.1621821969747543 TH pt time = 2.5330167822539806 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.690075010061264 TH pt time = 2.5531245954334736 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.832614816725254 TH pt time = 2.339891381561756 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7835668064653873 TH pt time = 2.513139396905899 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7026367820799351 TH pt time = 2.796882800757885 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.6479675993323326 TH pt time = 2.4971639923751354 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.9846629686653614 TH pt time = 2.4657804146409035 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.5969022028148174 TH pt time = 2.697007991373539 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7602720074355602 TH pt time = 2.4498093873262405 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.963611613959074 TH pt time = 2.6310251839458942 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7004458084702492 TH pt time = 2.9164502024650574 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.887732572853565 TH pt time = 2.4575488083064556 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8350806050002575 TH pt time = 2.23197178915143 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.5626789852976799 TH pt time = 2.704860605299473 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.6168799959123135 TH pt time = 2.2481359727680683 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.5654693879187107 TH pt time = 2.2636358067393303 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.6836861930787563 TH pt time = 2.825192976742983 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7971909940242767 TH pt time = 2.471243590116501 [1, 1024, 14, 14, 512, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.8480279818177223 TH pt time = 2.553586605936289 [1, 512, 14, 14, 512, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.7191735878586769 TH pt time = 2.6465672068297863 [1, 512, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 2.7811027877032757 TH pt time = 2.457349617034197 [1, 1024, 14, 14, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 5.434317365288734 TH pt time = 4.639615211635828 [1, 1024, 14, 14, 1024, 3, 3, 2, 2, 1, 1, 32] MKLDNN pt time = 0.9400106035172939 TH pt time = 2.9971951991319656 [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 4.494664408266544 TH pt time = 3.478870000690222 [1, 1024, 14, 14, 2048, 1, 1, 2, 2, 0, 0, 1] MKLDNN pt time = 4.8432330042123795 TH pt time = 3.6410867795348167 [1, 2048, 7, 7, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 4.779010973870754 TH pt time = 3.4093930013477802 [1, 1024, 7, 7, 1024, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.8385192044079304 TH pt time = 3.0921380035579205 [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 3.9088409766554832 TH pt time = 3.130124807357788 [1, 2048, 7, 7, 1024, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 4.0072557888925076 TH pt time = 2.977220807224512 [1, 1024, 7, 7, 1024, 3, 3, 1, 1, 1, 1, 32] MKLDNN pt time = 0.8867520093917847 TH pt time = 3.1505179964005947 [1, 1024, 7, 7, 2048, 1, 1, 1, 1, 0, 0, 1] MKLDNN pt time = 4.118196591734886 TH pt time = 3.46621660515666 ``` Reviewed By: dzhulgakov Differential Revision: D22250817 fbshipit-source-id: c9dc61b633e11a378a05810d711a696effd7f02b

Author

lly-zero-one

Committer

facebook-github-bot

Parents

ce3ba3b9

pytorch 48d6e2ad - Disable the mkldnn for conv2d in some special cases (#40610)

Commit

pytorch
48d6e2ad - Disable the mkldnn for conv2d in some special cases (#40610)