mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-18 16:11:50 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #20800 from SamFC10:fix-int8-conv

Browse Source
This commit is contained in:
Alexander Alekhin 2021-10-03 18:52:08 +00:00
commit 603d623eda
1 changed files with 16 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
modules/dnn/src/int8layers/convolution_layer.cpp
View File

@ -159,7 +159,7 @@ public:
enum { VEC_ALIGN = 32, DFT_TYPE = CV_8S }; enum { VEC_ALIGN = 32, DFT_TYPE = CV_8S };
Mat weightsMat; Mat weightsMat;
std::vector<int> biasvec; std::vector<int> biasvec;
Mat outputMultiplier; std::vector<float> outputMultiplier;
Mat activationLUT; Mat activationLUT;
Ptr<ActivationLayerInt8> activ; Ptr<ActivationLayerInt8> activ;
@ -249,10 +249,14 @@ public:
Mat biasMat = blobs[1]; Mat biasMat = blobs[1];
biasvec.resize(numOutput+2); biasvec.resize(numOutput+2);
for(int i = 0; i < numOutput; i++ )
biasvec[i] = biasMat.at<int>(i);
outputMultiplier = blobs[2]; Mat outMult = blobs[2];
outputMultiplier.resize(numOutput+2);
for(int i = 0; i < numOutput; i++ )
{
biasvec[i] = biasMat.at<int>(i);
outputMultiplier[i] = outMult.at<float>(i);
}
} }
bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
@ -283,16 +287,17 @@ public:
for (int i = 0; i < outCn; ++i) for (int i = 0; i < outCn; ++i)
{ {
float off = outputMultiplier.at<float>(i) * output_sc; float off = outputMultiplier[i] * output_sc;
if (!w.empty()) if (!w.empty())
off *= w.at<float>(i); off *= w.at<float>(i);
if (!b.empty()) if (!b.empty())
biasvec[i] += (int)std::round(b.at<float>(i)/off); biasvec[i] += (int)std::round(b.at<float>(i)/off);
outputMultiplier.at<float>(i) = off/new_sc; outputMultiplier[i] = off/new_sc;
} }
biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1]; biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
outputMultiplier[outCn] = outputMultiplier[outCn+1] = outputMultiplier[outCn-1];
} }
class ParallelConv : public cv::ParallelLoopBody class ParallelConv : public cv::ParallelLoopBody
@ -315,7 +320,7 @@ public:
bool useAVX512; bool useAVX512;
int blk_size_cn; int blk_size_cn;
int inpZp, outZp; int inpZp, outZp;
const float* multiplier; const std::vector<float>* multiplier;
ParallelConv() ParallelConv()
: input_(0), weights_(0), output_(0), ngroups_(0), nstripes_(0), : input_(0), weights_(0), output_(0), ngroups_(0), nstripes_(0),
@ -323,7 +328,7 @@ public:
, blk_size_cn(0), inpZp(0), outZp(0), multiplier(0) , blk_size_cn(0), inpZp(0), outZp(0), multiplier(0)
{} {}
static void run( const Mat& input, Mat& output, const Mat& weights, const Mat& multipliers, static void run( const Mat& input, Mat& output, const Mat& weights, const std::vector<float>& multipliers,
const std::vector<int>& biasvec, const Mat& activLUT, const std::vector<int>& biasvec, const Mat& activLUT,
const std::vector<size_t>& kernel_size, const std::vector<size_t>& strides, const std::vector<size_t>& kernel_size, const std::vector<size_t>& strides,
const std::vector<size_t>& pads_begin, const std::vector<size_t>& pads_end, const std::vector<size_t>& pads_begin, const std::vector<size_t>& pads_end,
@ -392,7 +397,7 @@ public:
p.inpZp = inp_Zp; p.inpZp = inp_Zp;
p.outZp = out_Zp; p.outZp = out_Zp;
p.multiplier = multipliers.ptr<float>(0); p.multiplier = &multipliers;
p.ofstab_.resize(karea * ncn); p.ofstab_.resize(karea * ncn);
int* ofstab = &p.ofstab_[0]; int* ofstab = &p.ofstab_[0];
@ -501,6 +506,7 @@ public:
const int8_t* wptr_orig_ = weights_->ptr<int8_t>(); const int8_t* wptr_orig_ = weights_->ptr<int8_t>();
size_t wstep = weights_->step1(); size_t wstep = weights_->step1();
const int* biasptr_ = &biasvec_->at(0); const int* biasptr_ = &biasvec_->at(0);
const float* multptr_ = &multiplier->at(0);
const int* lutptr_ = !activLUT_->empty() ? activLUT_->ptr<int>() : 0; const int* lutptr_ = !activLUT_->empty() ? activLUT_->ptr<int>() : 0;
int* data_out0_ = output_->ptr<int>(); int* data_out0_ = output_->ptr<int>();
AutoBuffer<int8_t> rowbuf0_; AutoBuffer<int8_t> rowbuf0_;
@ -539,7 +545,7 @@ public:
int startOutCn = (subsampleIdx % ngroups)*outCn; int startOutCn = (subsampleIdx % ngroups)*outCn;
const int8_t* wptr_orig = wptr_orig_ + wstep*startOutCn; const int8_t* wptr_orig = wptr_orig_ + wstep*startOutCn;
const int* biasptr = biasptr_ + startOutCn; const int* biasptr = biasptr_ + startOutCn;
const float* multptr = multiplier + startOutCn; const float* multptr = multptr_ + startOutCn;
for( int cn0 = 0; cn0 < inpCn; cn0 += blk_size_cn ) for( int cn0 = 0; cn0 < inpCn; cn0 += blk_size_cn )
{ {