CNTK/Source/Math/MatrixQuantizerCPU.cpp

113 строки
4.1 KiB
C++

#include "stdafx.h"
#include "MatrixQuantizerCPU.h"
namespace Microsoft { namespace MSR { namespace CNTK {
template <class ElemType>
MatrixQuantizerCPU<ElemType>::MatrixQuantizerCPU()
: MatrixQuantizerImpl<ElemType>(CPUDEVICE)
{
}
template <class ElemType>
void MatrixQuantizerCPU<ElemType>::QuantizeAsync(const Matrix<ElemType>& inMatrix, const Matrix<ElemType>& inResidual, QuantizedMatrix<ElemType>& outQMatrix, Matrix<ElemType>& outResidual, bool zeroThresholdFor1Bit)
{
// The outQMatrix should be on the CPU
// TODO: Support transferring the quantization output to a quantized matrix on the GPU
assert(outQMatrix.GetDeviceId() == CPUDEVICE);
size_t nBits = outQMatrix.GetNumBits();
size_t nRow = inMatrix.GetNumRows();
size_t nCol = inMatrix.GetNumCols();
// Verify that the different matrix parameters have matching dimensions
assert((outQMatrix.GetNumRows() == nRow) && (outQMatrix.GetNumCols() == nCol));
assert((inResidual.GetNumRows() == nRow) && (inResidual.GetNumCols() == nCol));
assert((outResidual.GetNumRows() == nRow) && (outResidual.GetNumCols() == nCol));
const size_t ldNbits = ValueQuantizer<ElemType>::ld(nBits);
#ifdef QUANTUSEPPL
Concurrency::parallel_for((size_t) 0, us.cols(), [&](size_t j)
#else
for (size_t j = 0; j < nCol; j++)
#endif
{
auto& qcol = *(outQMatrix.GetQuantizedColumn(j));
if (zeroThresholdFor1Bit)
{
// Explicit use of 'template' keyword is needed to compile with GCC
ColumnQuantizer<ElemType>::template ComputeRangeStatColj<true>(inMatrix.Data(), inResidual.Data(), (long) nRow, j, nBits, qcol.lower, qcol.upper);
}
else
{
// Explicit use of 'template' keyword is needed to compile with GCC
ColumnQuantizer<ElemType>::template ComputeRangeStatColj<false>(inMatrix.Data(), inResidual.Data(), (long) nRow, j, nBits, qcol.lower, qcol.upper);
}
ColumnQuantizer<ElemType> q(ldNbits, qcol.lower, qcol.upper);
if (zeroThresholdFor1Bit)
{
// Explicit use of 'template' keyword is needed to compile with GCC
q.template Quantize<true>(inMatrix.Data(), inResidual.Data(), (long) nRow, j, qcol.bits, outResidual.Data());
}
else
{
// Explicit use of 'template' keyword is needed to compile with GCC
q.template Quantize<false>(inMatrix.Data(), inResidual.Data(), (long) nRow, j, qcol.bits, outResidual.Data());
}
}
#ifdef QUANTUSEPPL
);
#endif
}
template <class ElemType>
void MatrixQuantizerCPU<ElemType>::WaitQuantizeAsyncDone()
{
// TODO: Currently this is a no-op since the actual quantization is synchronous
}
// unquantize an entire matrix, calling unquantize() for each column
template <class ElemType>
void MatrixQuantizerCPU<ElemType>::UnquantizeAsync(QuantizedMatrix<ElemType>& inQMatrix, Matrix<ElemType>& outMatrix, bool add /*= false*/)
{
// The inQMatrix and hould be on the CPU
assert(inQMatrix.GetDeviceId() == CPUDEVICE);
assert(outMatrix.GetDeviceId() == CPUDEVICE);
size_t nBits = inQMatrix.GetNumBits();
size_t nRow = inQMatrix.GetNumRows();
size_t nCol = inQMatrix.GetNumCols();
// Verify that the different matrix parameters have matching dimensions
assert((outMatrix.GetNumRows() == nRow) && (outMatrix.GetNumCols() == nCol));
const size_t ldNbits = ValueQuantizer<ElemType>::ld(nBits);
#ifdef QUANTUSEPPL
Concurrency::parallel_for((size_t) 0, us.cols(), [&](size_t j)
#else
for (size_t j = 0; j < nCol; j++)
#endif
{
const auto& qcol = *(inQMatrix.GetQuantizedColumn(j));
ColumnQuantizer<ElemType> q(ldNbits, qcol.lower, qcol.upper);
q.Unquantize(outMatrix.Data(), (long) nRow, j, qcol.bits, add);
}
#ifdef QUANTUSEPPL
);
#endif
}
template <class ElemType>
void MatrixQuantizerCPU<ElemType>::WaitUnquantizeAsyncDone()
{
// TODO: Currently this is a no-op since the actual quantization is synchronous
}
//The explicit instantiation part will make the linker happy
template class MatrixQuantizerCPU<float>;
template class MatrixQuantizerCPU<double>;
}}}