int main() {
int numClockCycles = 0;
Xil_ICacheEnable();
Xil_DCacheEnable();
print("---Entering main---\n\r");
print("---Trial Name \t Trial # \t Clock Cycles---\n\r");
int i = 0;
// Extra Method contains an interrupt routine which is set to go off at timed intervals
extra_method();
for( i=0; i < NUMBER_OF_TRIALS; i++) {
//TIMER RESET CODE
//Turn off the timer
XTmrCtr_SetControlStatusReg(XPAR_TMRCTR_0_BASEADDR, 1, 0);
//Put a zero in the load register
XTmrCtr_SetLoadReg(XPAR_TMRCTR_0_BASEADDR, 1, 0);
//Copy the load register into the counter register
XTmrCtr_SetControlStatusReg(XPAR_TMRCTR_0_BASEADDR, 1, XTC_CSR_LOAD_MASK);
//Enable (start) the timer
XTmrCtr_SetControlStatusReg(XPAR_TMRCTR_0_BASEADDR, 1, XTC_CSR_ENABLE_TMR_MASK);
//END TIMER RESET CODE
//blinkLED(int numberOfBlinks);
//offLED();
//onLED();
//sevenSegment();
//printLongerStrings(); // Write this function
//printShortStrings(); // Write this function
//printfShortStrings(); // Write this function
//xil_printfShortStrings(); // Write this function
//intAddAndMultiply(); // Write this function
//floatAddAndMultiply(); // Write this function
numClockCycles = XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, 1);
xil_printf("print \t%d\t%d\n", i,numClockCycles );
}
return 0;
}
// 这里InputData是图像数据，inputData[r][c],r行c列，这里跟各权重模板是一致的
void cnnff(CNN* cnn,float** inputData)
{
// 第一层的传播
int i, j, r, c;
// 第一层卷积层输出数据（C1）
nSize mapSize = {cnn->C1.mapSize, cnn->C1.mapSize};
nSize inSize = {cnn->C1.inputWidth, cnn->C1.inputHeight};
nSize outSize = {cnn->S2.inputWidth, cnn->S2.inputHeight};
float **wholeKernel = (float**)malloc(6 * sizeof(float*));
for (i = 0; i < 6; i++)
wholeKernel[i] = (float*)malloc(25 * sizeof(float));
int cov1layerstart = XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, TIMER_COUNTER_0);
for(j = 0;j < (cnn->C1.inChannels);j++)
{
// 卷积
int m, n, k;
int l;
for (m = 0; m < 6; m++)
{
l = 0;
for (k = 0; k < 5; k++)
{
for (n = 0; n < 5; n++)
{
wholeKernel[m][l] = cnn->C1.mapData[j][m][4-k][4-n];
l++;
}
}
}
float** mapout = cov_layer1_6(wholeKernel,mapSize,inputData,inSize,valid);
// 求和
k = 0;
for (i = 0; i < 6; i++)
for (m = 0; m < 24; m++)
for (n = 0; n < 24; n++)
cnn->C1.v[i][m][n] += mapout[i][24*m+n];
}
for(i = 0;i < (cnn->C1.outChannels);i++)
for(r = 0;r < outSize.r;r++)
for(c = 0;c < outSize.c;c++)
// sigmoid function
cnn->C1.y[i][r][c] = activation_Sigma(cnn->C1.v[i][r][c], cnn->C1.basicData[i]);
// for(i=0;i<(cnn->C1.outChannels);i++)
// {
// for(j=0;j<(cnn->C1.inChannels);j++)
// {
// float** mapout=cov(cnn->C1.mapData[j][i],mapSize,inputData,inSize,valid);
// addmat(cnn->C1.v[i],cnn->C1.v[i],outSize,mapout,outSize);
// for(r=0;r<outSize.r;r++)
// free(mapout[r]);
// free(mapout);
// }
// for(r=0;r<outSize.r;r++)
// for(c=0;c<outSize.c;c++)
// cnn->C1.y[i][r][c]=activation_Sigma(cnn->C1.v[i][r][c],cnn->C1.basicData[i]);
// }
int cov1layerend = XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, TIMER_COUNTER_0);
xil_printf("%d cycles spent on cov2\n", cov1layerend - cov1layerstart);
// 第二层的输出传播S2，采样层
outSize.c = cnn->C3.inputWidth;
outSize.r = cnn->C3.inputHeight;
inSize.c = cnn->S2.inputWidth;
inSize.r = cnn->S2.inputHeight;
for(i = 0;i < (cnn->S2.outChannels);i++)
{
// pooling的类型是取平均
if(cnn->S2.poolType == AvePool)
avgPooling(cnn->S2.y[i], outSize, cnn->C1.y[i], inSize, cnn->S2.mapSize);
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////
outSize.c = cnn->S4.inputWidth;
outSize.r = cnn->S4.inputHeight;
inSize.c = cnn->C3.inputWidth;
inSize.r = cnn->C3.inputHeight;
mapSize.c = cnn->C3.mapSize;
mapSize.r = cnn->C3.mapSize;
// for(i = 0;i < (cnn->C3.outChannels);i++) // 12
// {
// for(j = 0;j < (cnn->C3.inChannels);j++) //6
// {
// float** mapout = cov(cnn->C3.mapData[j][i], mapSize, cnn->S2.y[j], inSize, valid);
// addmat(cnn->C3.v[i],cnn->C3.v[i], outSize, mapout, outSize);
// }
//
//// for(j = 0;j < (cnn->C3.inChannels);j++)
//// {
//// mapKernel[j] = cnn->C3.mapData[j][i];
//// float** mapout = cov_layer3(cnn->C3.mapData[j][i], mapSize, cnn->S2.y[j], inSize, valid);
////
//// addmat(cnn->C3.v[i],cnn->C3.v[i],outSize,mapout,outSize);
//// }
//
// for(r = 0;r < outSize.r;r++)
// for(c = 0;c < outSize.c;c++)
// cnn->C3.y[i][r][c] = activation_Sigma(cnn->C3.v[i][r][c],cnn->C3.basicData[i]);
// }
// float **wholeKernel = (float**)malloc(6*sizeof(float*));
// for (i = 0; i < 6; i++)
// wholeKernel[i] = (float*)malloc(25 * sizeof(float));
int cov2layerstart = XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, TIMER_COUNTER_0);
for(j = 0;j < (cnn->C3.inChannels);j++)
{
int m, n, k;
int l;
// convert 6 kernel to vector
for (m = 0; m < 6; m++)
{
l = 0;
for (k = 0; k < 5; k++)
{
for (n = 0; n < 5; n++)
{
wholeKernel[m][l] = cnn->C3.mapData[j][m][4-k][4-n];
l++;
}
}
}
float** mapout = cov_layer3_6(wholeKernel, mapSize, cnn->S2.y[j], inSize, valid);
// add all mapout to the v
for (i = 0; i < 6; i++)
for (m = 0; m < 8; m++)
for (n = 0; n < 8; n++)
cnn->C3.v[i][m][n] += mapout[i][8*m+n];
for(i = 0; i < 6;i++)
free(mapout[i]);
free(mapout);
for (m = 0; m < 6; m++)
{
l = 0;
for (k = 0; k < 5; k++)
{
for (n = 0; n < 5; n++)
{
wholeKernel[m][l] = cnn->C3.mapData[j][m+6][4-k][4-n];
l++;
}
}
}
mapout = cov_layer3_6(wholeKernel, mapSize, cnn->S2.y[j], inSize, valid);
for (; i < 12; i++)
for (m = 0; m < 8; m++)
for (n = 0; n < 8; n++)
cnn->C3.v[i][m][n] += mapout[i-6][8*m+n];
for(i = 0; i < 6;i++)
free(mapout[i]);
free(mapout);
}
for (i = 0; i < 6; i++)
free(wholeKernel[i]);
free(wholeKernel);
for (i=0; i < cnn->C3.outChannels; i++)
for(r=0;r<outSize.r;r++)
for(c=0;c<outSize.c;c++)
cnn->C3.y[i][r][c]=activation_Sigma(cnn->C3.v[i][r][c],cnn->C3.basicData[i]);
////////////////////////////////////////////////////////////////////////////////////////////////////
int cov2layerend = XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, TIMER_COUNTER_0);
xil_printf("%d cycles spent on cov2\n", cov2layerend - cov2layerstart);
// 第四层的Pooling层
inSize.c=cnn->S4.inputWidth;
inSize.r=cnn->S4.inputHeight;
outSize.c=inSize.c/cnn->S4.mapSize;
outSize.r=inSize.r/cnn->S4.mapSize;
for(i=0;i<(cnn->S4.outChannels);i++)
{
if(cnn->S4.poolType == AvePool)
avgPooling(cnn->S4.y[i],outSize,cnn->C3.y[i],inSize,cnn->S4.mapSize);
}
// 输出层O5的处理
// 首先需要将前面的多维输出展开成一维向量
float O5inData[192];
for(i = 0;i < (cnn->S4.outChannels);i++)
for(r = 0;r < outSize.r;r++)
for(c = 0;c < outSize.c;c++)
O5inData[i*outSize.r*outSize.c+r*outSize.c+c]=cnn->S4.y[i][r][c];
nSize nnSize = {cnn->O5.inputNum, cnn->O5.outputNum};
nnff(cnn->O5.v, O5inData, cnn->O5.wData, cnn->O5.basicData, nnSize);
// 计算每一个数字的概率
for(i = 0;i < cnn->O5.outputNum;i++)
cnn->O5.y[i] = activation_Sigma(cnn->O5.v[i],cnn->O5.basicData[i]);
}
int counter_method(){
return XTmrCtr_GetTimerCounterReg(XPAR_TMRCTR_0_BASEADDR, 1);
}
