164 lines
3.7 KiB
C++
Executable File
164 lines
3.7 KiB
C++
Executable File
#include "Softmax.h"
|
|
#include <stdio.h>
|
|
#include "Error.h"
|
|
#include "arm_nnfunctions.h"
|
|
#include "Test.h"
|
|
|
|
/*
|
|
|
|
Tests have shown that, compared to a float32 implementation
|
|
there is an average error of 4.2 percent and standard deviation
|
|
of 0.89.
|
|
|
|
Which means that with 100 batches, 4 batches will give the wrong position
|
|
for the max.
|
|
|
|
But it depends highly of the vector dimension.
|
|
|
|
Regressions are giving:
|
|
|
|
Average error rate = -0.555548 + 0.246918 vecDim
|
|
Variance = -0.0112281 + 0.0382476 vecDim
|
|
|
|
So for vecDim = 21 we have
|
|
Average error rate = 4.6 percent
|
|
Variance for error rate = 0.8
|
|
|
|
This data is used to define the threshold for tests
|
|
|
|
*/
|
|
#define THRESHOLD 7.5
|
|
|
|
int16_t findMaxIndex(q7_t *vec_in, int length)
|
|
{
|
|
int16_t currentIndex=0;
|
|
int16_t i=1;
|
|
q7_t currentMax=vec_in[0];
|
|
|
|
while(i<length)
|
|
{
|
|
if (vec_in[i] > currentMax)
|
|
{
|
|
currentMax = vec_in[i];
|
|
currentIndex = i;
|
|
}
|
|
i++;
|
|
}
|
|
|
|
return(currentIndex+1);
|
|
}
|
|
|
|
int16_t differences(int16_t *pa,int16_t *pb, int length)
|
|
{
|
|
int16_t d=0;
|
|
int i=0;
|
|
while(i < length)
|
|
{
|
|
if (*pa != *pb)
|
|
{
|
|
d++;
|
|
}
|
|
|
|
pa++;
|
|
pb++;
|
|
i++;
|
|
}
|
|
return(d);
|
|
}
|
|
|
|
|
|
void Softmax::test_softmax_q7()
|
|
{
|
|
const q7_t *vec_in = input.ptr();
|
|
q7_t *pTmp = temp.ptr();
|
|
int16_t *pOut = output.ptr();
|
|
int16_t maxIndex;
|
|
|
|
for(int i=0; i <this->nbSamples;i++)
|
|
{
|
|
arm_softmax_q7(vec_in, this->vecDim, pTmp );
|
|
maxIndex=findMaxIndex(pTmp,this->vecDim);
|
|
*pOut++ = maxIndex;
|
|
|
|
vec_in += this->vecDim;
|
|
pTmp += this->vecDim;
|
|
}
|
|
|
|
int diff = differences(ref.ptr(),output.ptr(),this->nbSamples);
|
|
|
|
ASSERT_TRUE(100.0*diff/this->nbSamples <= THRESHOLD);
|
|
|
|
}
|
|
|
|
void Softmax::test_softmax_with_batch_q7()
|
|
{
|
|
const q7_t *vec_in = input.ptr();
|
|
q7_t *pTmp = temp.ptr();
|
|
int16_t *pOut = output.ptr();
|
|
int16_t maxIndex;
|
|
|
|
arm_softmax_with_batch_q7(vec_in, this->nbSamples,this->vecDim, pTmp );
|
|
|
|
for(int i=0; i <this->nbSamples;i++)
|
|
{
|
|
maxIndex=findMaxIndex(pTmp,this->vecDim);
|
|
*pOut++ = maxIndex;
|
|
pTmp += this->vecDim;
|
|
}
|
|
|
|
int diff = differences(ref.ptr(),output.ptr(),this->nbSamples);
|
|
|
|
ASSERT_TRUE(100.0*diff/this->nbSamples <= THRESHOLD);
|
|
|
|
}
|
|
|
|
|
|
void Softmax::setUp(Testing::testID_t id,std::vector<Testing::param_t>& paramsArgs,Client::PatternMgr *mgr)
|
|
{
|
|
|
|
switch(id)
|
|
{
|
|
case Softmax::TEST_SOFTMAX_Q7_1:
|
|
{
|
|
ref.reload(Softmax::REF1_S16_ID,mgr);
|
|
dims.reload(Softmax::DIMS1_S16_ID,mgr);
|
|
input.reload(Softmax::INPUT1_Q7_ID,mgr);
|
|
|
|
const int16_t *pDims=dims.ptr();
|
|
|
|
this->nbSamples = pDims[0];
|
|
this->vecDim = pDims[1];
|
|
}
|
|
break;
|
|
|
|
case Softmax::TEST_SOFTMAX_WITH_BATCH_Q7_2:
|
|
{
|
|
ref.reload(Softmax::REF1_S16_ID,mgr);
|
|
dims.reload(Softmax::DIMS1_S16_ID,mgr);
|
|
input.reload(Softmax::INPUT1_Q7_ID,mgr);
|
|
|
|
const int16_t *pDims=dims.ptr();
|
|
|
|
this->nbSamples = pDims[0];
|
|
this->vecDim = pDims[1];
|
|
}
|
|
break;
|
|
|
|
}
|
|
|
|
output.create(ref.nbSamples(),Softmax::OUTPUT_S16_ID,mgr);
|
|
// Used to compare bit exactness of the reference C version
|
|
// and the optimized version.
|
|
temp.create(this->vecDim*this->nbSamples,Softmax::TEMP_Q7_ID,mgr);
|
|
|
|
|
|
}
|
|
|
|
void Softmax::tearDown(Testing::testID_t id,Client::PatternMgr *mgr)
|
|
{
|
|
// Array are big so by default they are not dumped and only
|
|
// used for debug.
|
|
//output.dump(mgr);
|
|
//temp.dump(mgr);
|
|
}
|