未验证 提交 3fdbab25 编写于 作者: W Wangzheee 提交者: GitHub

cherrypick fix op_teller (#37266)

上级 71b04f61
...@@ -59,6 +59,8 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -59,6 +59,8 @@ struct SimpleOpTypeSetTeller : public Teller {
#if CUDA_VERSION >= 10020 #if CUDA_VERSION >= 10020
teller_set.insert("reshape"); teller_set.insert("reshape");
teller_set.insert("reshape2"); teller_set.insert("reshape2");
int8_teller_set.insert("reshape");
int8_teller_set.insert("reshape2");
#endif #endif
} }
...@@ -74,24 +76,54 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -74,24 +76,54 @@ struct SimpleOpTypeSetTeller : public Teller {
private: private:
// use this set for no calib int8. // use this set for no calib int8.
std::unordered_set<std::string> int8_teller_set{"mul", std::unordered_set<std::string> int8_teller_set{"mul",
"conv2d",
"matmul", "matmul",
"stack", "conv2d",
"conv2d_fusion", "conv2d_fusion",
"pool2d", "pool2d",
"relu", "relu",
"depthwise_conv2d",
"softmax", "softmax",
"sigmoid", "sigmoid",
"hard_swish",
"depthwise_conv2d",
"batch_norm", "batch_norm",
"concat",
"tanh",
"pad",
"elementwise_add", "elementwise_add",
"elementwise_mul",
"dropout",
"prelu",
"conv2d_transpose",
"depthwise_conv2d_transpose",
"leaky_relu", "leaky_relu",
"fc", "fc",
"concat", "shuffle_channel",
"swish",
"split",
"instance_norm",
"gelu",
"layer_norm",
"scale", "scale",
"elementwise_mul", "stack",
"conv2d_transpose", "transpose2",
"hard_swish"}; "transpose",
"flatten2",
"flatten",
"gather",
"gather_nd",
"yolo_box",
"roi_align",
"affine_channel",
"nearest_interp",
"anchor_generator",
"reduce_sum",
"reduce_mean",
"conv3d",
"conv3d_transpose",
"mish",
"nearest_interp_v2",
"pool3d",
"deformable_conv"};
std::unordered_set<std::string> teller_set{"mul", std::unordered_set<std::string> teller_set{"mul",
"matmul", "matmul",
"conv2d", "conv2d",
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册