Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Use single precision in gain calculation, use pointers instead of span. #8051

Merged
merged 1 commit into from Jul 12, 2022
Merged

Use single precision in gain calculation, use pointers instead of span. #8051

merged 1 commit into from Jul 12, 2022

Conversation

RAMitchell
Copy link
Member

Removing spans in favour of raw pointers decreases register pressure in the GPU kernel significantly.

Use of double/single precision in gain calculations is inconsistent. Prefer single precision.

@trivialfis
Copy link
Member

Do you have some benchmarks for accuracy for small datasets? Preferably with CPU implementation as well.

@RAMitchell
Copy link
Member Author

Here are 2 gbm-bench runs for both CPU and GPU. The accuracies are identical.

Before PR 
{
"airline": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.8434307346593626,
      "Accuracy": 0.7182532274977227,
      "Log_Loss": 0.5297073384219716,
      "Precision": 0.6536453477079318,
      "Recall": 0.8642811257864607
    },
    "train_time": 966.4464671728201,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.8431610435789277,
      "Accuracy": 0.718068859889482,
      "Log_Loss": 0.5300302877835903,
      "Precision": 0.6534910973662341,
      "Recall": 0.8641185139771076
    },
    "train_time": 91.14509701775387,
    "train_time_std": 0.0
  }
},
"bosch": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.6902547809021118,
      "Accuracy": 0.955129883843717,
      "Log_Loss": 0.24913303252016256,
      "Precision": 0.04193227091633466,
      "Recall": 0.29543859649122806
    },
    "train_time": 57.79425460193306,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.6911762239977932,
      "Accuracy": 0.9580316789862724,
      "Log_Loss": 0.24111224289242209,
      "Precision": 0.04423262289814715,
      "Recall": 0.28982456140350876
    },
    "train_time": 12.88072164868936,
    "train_time_std": 0.0
  }
},
"covtype": {
  "xgb-cpu": {
    "accuracy": {
      "Accuracy": 0.9398466476769103,
      "F1": 0.9397222738018455,
      "Precision": 0.9398793000566292,
      "Recall": 0.9398466476769103
    },
    "train_time": 39.06878037704155,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "Accuracy": 0.9379620147500495,
      "F1": 0.9378270737037167,
      "Precision": 0.9380157400229653,
      "Recall": 0.9379620147500495
    },
    "train_time": 18.243315340019763,
    "train_time_std": 0.0
  }
},
"epsilon": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.9477093412445664,
      "Accuracy": 0.87034,
      "Log_Loss": 0.3009770990117234,
      "Precision": 0.8434001151196687,
      "Recall": 0.9092783505154639
    },
    "train_time": 1550.9631590410136,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.9481207911778409,
      "Accuracy": 0.87118,
      "Log_Loss": 0.2998701572805944,
      "Precision": 0.8448715324936278,
      "Recall": 0.9090381343208888
    },
    "train_time": 46.500129331834614,
    "train_time_std": 0.0
  }
},
"fraud": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.9654261008004777,
      "Accuracy": 0.9995611109160493,
      "Log_Loss": 0.003519674551653179,
      "Precision": 0.9506172839506173,
      "Recall": 0.7857142857142857
    },
    "train_time": 2.6741700717248023,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.965037059421405,
      "Accuracy": 0.9995611109160493,
      "Log_Loss": 0.00356269826075694,
      "Precision": 0.9397590361445783,
      "Recall": 0.7959183673469388
    },
    "train_time": 1.332988286856562,
    "train_time_std": 0.0
  }
},
"higgs": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.8399291373085628,
      "Accuracy": 0.7342790909090909,
      "Log_Loss": 0.5217814797614634,
      "Precision": 0.6912109500847503,
      "Recall": 0.9014453459278047
    },
    "train_time": 112.69107929291204,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.8393949180683775,
      "Accuracy": 0.7335804545454545,
      "Log_Loss": 0.5225674019070321,
      "Precision": 0.6904788642120271,
      "Recall": 0.9015628188704419
    },
    "train_time": 17.52672170335427,
    "train_time_std": 0.0
  }
},
"year": {
  "xgb-cpu": {
    "accuracy": {
      "MeanAbsError": 6.220576286315918,
      "MeanSquaredError": 79.77742004394531,
      "MedianAbsError": 4.271484375
    },
    "train_time": 16.023101320955902,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "MeanAbsError": 6.22576379776001,
      "MeanSquaredError": 79.91011810302734,
      "MedianAbsError": 4.2696533203125
    },
    "train_time": 6.9804930170066655,
    "train_time_std": 0.0
  }
}
}
After PR 
{
"airline": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.8434307346593626,
      "Accuracy": 0.7182532274977227,
      "Log_Loss": 0.5297073384219716,
      "Precision": 0.6536453477079318,
      "Recall": 0.8642811257864607
    },
    "train_time": 994.9771266668104,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.8431610435789277,
      "Accuracy": 0.718068859889482,
      "Log_Loss": 0.5300302877835903,
      "Precision": 0.6534910973662341,
      "Recall": 0.8641185139771076
    },
    "train_time": 89.69093904690817,
    "train_time_std": 0.0
  }
},
"bosch": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.6902547809021118,
      "Accuracy": 0.955129883843717,
      "Log_Loss": 0.24913303252016256,
      "Precision": 0.04193227091633466,
      "Recall": 0.29543859649122806
    },
    "train_time": 55.15471205022186,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.6911762239977932,
      "Accuracy": 0.9580316789862724,
      "Log_Loss": 0.24111224289242209,
      "Precision": 0.04423262289814715,
      "Recall": 0.28982456140350876
    },
    "train_time": 12.608841334935278,
    "train_time_std": 0.0
  }
},
"covtype": {
  "xgb-cpu": {
    "accuracy": {
      "Accuracy": 0.9398466476769103,
      "F1": 0.9397222738018455,
      "Precision": 0.9398793000566292,
      "Recall": 0.9398466476769103
    },
    "train_time": 40.34531668201089,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "Accuracy": 0.9379620147500495,
      "F1": 0.9378270737037167,
      "Precision": 0.9380157400229653,
      "Recall": 0.9379620147500495
    },
    "train_time": 17.709761895705014,
    "train_time_std": 0.0
  }
},
"epsilon": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.9477093412445664,
      "Accuracy": 0.87034,
      "Log_Loss": 0.3009770990117234,
      "Precision": 0.8434001151196687,
      "Recall": 0.9092783505154639
    },
    "train_time": 1647.445112537127,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.9481207911778409,
      "Accuracy": 0.87118,
      "Log_Loss": 0.2998701572805944,
      "Precision": 0.8448715324936278,
      "Recall": 0.9090381343208888
    },
    "train_time": 45.50282173091546,
    "train_time_std": 0.0
  }
},
"fraud": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.9654261008004777,
      "Accuracy": 0.9995611109160493,
      "Log_Loss": 0.003519674551653179,
      "Precision": 0.9506172839506173,
      "Recall": 0.7857142857142857
    },
    "train_time": 2.592053124215454,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.965037059421405,
      "Accuracy": 0.9995611109160493,
      "Log_Loss": 0.00356269826075694,
      "Precision": 0.9397590361445783,
      "Recall": 0.7959183673469388
    },
    "train_time": 1.2642242829315364,
    "train_time_std": 0.0
  }
},
"higgs": {
  "xgb-cpu": {
    "accuracy": {
      "AUC": 0.8399291373085628,
      "Accuracy": 0.7342790909090909,
      "Log_Loss": 0.5217814797614634,
      "Precision": 0.6912109500847503,
      "Recall": 0.9014453459278047
    },
    "train_time": 109.59541429579258,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "AUC": 0.8393949180683775,
      "Accuracy": 0.7335804545454545,
      "Log_Loss": 0.5225674019070321,
      "Precision": 0.6904788642120271,
      "Recall": 0.9015628188704419
    },
    "train_time": 17.26801946386695,
    "train_time_std": 0.0
  }
},
"year": {
  "xgb-cpu": {
    "accuracy": {
      "MeanAbsError": 6.220576286315918,
      "MeanSquaredError": 79.77742004394531,
      "MedianAbsError": 4.271484375
    },
    "train_time": 17.15528222732246,
    "train_time_std": 0.0
  },
  "xgb-gpu": {
    "accuracy": {
      "MeanAbsError": 6.22576379776001,
      "MeanSquaredError": 79.91011810302734,
      "MedianAbsError": 4.2696533203125
    },
    "train_time": 6.902157460339367,
    "train_time_std": 0.0
  }
}
}

@trivialfis
Copy link
Member

It's weird that the CPU implementation seems somehow slower.

@RAMitchell
Copy link
Member Author

I wouldn't read too much into it. The CPU runs have very high variance on my dual socket machine.

@RAMitchell RAMitchell merged commit 0bdaca2 into dmlc:master Jul 12, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@trivialfis trivialfis trivialfis approved these changes

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

None yet
2 participants
@RAMitchell @trivialfis