mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
246f5c412e
Summary:
This reverts commit d73c830e23.
We have observed significant perf drop when training ResNext101 with multiple amd GPUs:
Before:
https://ci.pytorch.org/jenkins/job/caffe2-builds/job/py2-clang7-rocmdeb-ubuntu16.04-bench/1636/console
2 GPUs ResNext training got 150\~160 imgs/sec
4 GPUs ResNext training got 270\~280 imgs/sec
After:
https://ci.pytorch.org/jenkins/job/caffe2-builds/job/py2-clang7-rocmdeb-ubuntu16.04-bench/1637/console
Both 2 and 4 GPUs ResNext training drop to 110\~120 imgs/sec
Similar perf drop are seen on ResNet50 training jobs as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18680
Differential Revision: D14702941
Pulled By: bddppq
fbshipit-source-id: 828141805afc23f25c08d4a2eb6d4b99f817c128
66 lines
2.1 KiB
C++
66 lines
2.1 KiB
C++
#include "caffe2/operators/conditional_op.h"
|
|
#include "caffe2/core/operator.h"
|
|
#include "caffe2/core/tensor.h"
|
|
|
|
namespace caffe2 {
|
|
|
|
template <>
|
|
bool ConditionalOp<CPUContext>::RunOnDevice() {
|
|
auto& condition = Input(0);
|
|
auto& dataT = Input(1);
|
|
auto& dataF = Input(2);
|
|
|
|
// verify the inputs shape
|
|
CAFFE_ENFORCE_EQ(condition.dim(), 1);
|
|
CAFFE_ENFORCE(dataT.dim() >= 1);
|
|
CAFFE_ENFORCE(dataT.sizes()[0] == condition.sizes()[0]);
|
|
CAFFE_ENFORCE_EQ(dataT.dim(), dataF.dim());
|
|
for (size_t i = 0; i < dataT.sizes().size(); i++) {
|
|
CAFFE_ENFORCE(dataT.sizes().at(i) == dataF.sizes().at(i));
|
|
}
|
|
const auto innerSize = dataT.size_from_dim(1);
|
|
const auto innerSizeBytes = innerSize * dataT.dtype().itemsize();
|
|
CAFFE_ENFORCE(innerSize * dataF.dtype().itemsize() == innerSizeBytes);
|
|
|
|
// initialize output shape
|
|
auto* dataOut = Output(0);
|
|
const auto* condPtr = condition.template data<bool>();
|
|
dataOut->ResizeLike(dataT);
|
|
auto* outPtr = (char*)dataOut->raw_mutable_data(dataT.dtype());
|
|
|
|
// perform conditional op along first dimension
|
|
const auto* ptrT = (char*)dataT.raw_data();
|
|
const auto* ptrF = (char*)dataF.raw_data();
|
|
for (int64_t i = 0; i < condition.numel(); i++) {
|
|
auto* dst = outPtr + i * innerSizeBytes;
|
|
if (condPtr[i]) {
|
|
context_.CopyItemsSameDevice(
|
|
dataT.dtype(), innerSize, ptrT + i * innerSizeBytes, dst);
|
|
} else {
|
|
context_.CopyItemsSameDevice(
|
|
dataF.dtype(), innerSize, ptrF + i * innerSizeBytes, dst);
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
REGISTER_CPU_OPERATOR(Conditional, ConditionalOp<CPUContext>);
|
|
|
|
OPERATOR_SCHEMA(Conditional)
|
|
.NumInputs(3)
|
|
.NumOutputs(1)
|
|
.SetDoc(R"DOC(
|
|
Given a 1-D tensor of boolean values, apply conditional operator along the first
|
|
dimension of DataT and DataF and return DataO. Note, DataT and DataF must
|
|
have the exact same shape and type.
|
|
)DOC")
|
|
.Input(0, "Condition", "Boolean tensor to select DataT or DataF")
|
|
.Input(1, "DataT", "Data to use when True")
|
|
.Input(2, "DataF", "Data to use when False")
|
|
.Output(0, "DataO", "Output data after applying ConditionalOp")
|
|
.IdenticalTypeAndShapeOfInput(1);
|
|
|
|
NO_GRADIENT(Conditional);
|
|
|
|
} // caffe2
|