int THProcessYUYV(THNETWORK *network, unsigned char *image, int width, int height, float **results, int *outwidth, int *outheight)
{
THFloatTensor *out;
THFloatStorage *st;
#ifdef CUDNN
if(network->net->cuda)
THError("This function is not supported with CUDNN");
#endif
#ifdef OPENCL
if(network->net->cuda)
THError("This function is not supported with OpenCL");
#endif
st = THFloatStorage_new(width * height * 3);
yuyv2fRGB(image, st->data, width*height, width, width, height, network->mean, network->std);
THFloatTensor *t = THFloatTensor_new();
t->storage = st;
t->nDimension = 3;
t->size[0] = 3;
t->size[1] = height;
t->size[2] = width;
t->stride[0] = width * height;
t->stride[1] = width;
t->stride[2] = 1;
out = forward(network->net, t);
THFloatTensor_free(t);
*results = out->storage->data;
if(out->nDimension >= 3)
{
*outwidth = out->size[out->nDimension - 1];
*outheight = out->size[out->nDimension - 2];
} else *outwidth = *outheight = 1;
return THFloatTensor_nElement(out);
}
void THFloatTensor_resizeAs(THFloatTensor *tdst, THFloatTensor *tsrc)
{
if(tsrc == tdst)
return;
long nelemsrc = THFloatTensor_nElement(tsrc);
tdst->nDimension = tsrc->nDimension;
memcpy(tdst->size, tsrc->size, sizeof(tsrc->size));
memcpy(tdst->stride, tsrc->stride, sizeof(tsrc->stride));
if(!tdst->storage)
tdst->storage = THFloatStorage_new(nelemsrc);
else if(nelemsrc != THFloatTensor_nElement(tdst))
{
if(tdst->storage)
tdst->storage->data = realloc(tdst->storage->data, sizeof(*tdst->storage->data) * nelemsrc);
else tdst->storage = THFloatStorage_new(nelemsrc);
}
}
void THFloatTensor_resize1d(THFloatTensor *t, long size0)
{
t->nDimension = 1;
t->size[0] = size0;
t->stride[0] = 1;
if(!t->storage)
t->storage = THFloatStorage_new(size0);
}
void THFloatTensor_resize3d(THFloatTensor *t, long size0, long size1, long size2)
{
t->nDimension = 3;
t->size[0] = size0;
t->size[1] = size1;
t->size[2] = size2;
t->stride[2] = 1;
t->stride[1] = size2;
t->stride[0] = size1 * size2;
if(!t->storage)
t->storage = THFloatStorage_new(size0 * size1 * size2);
}
void THFloatTensor_resize(THFloatTensor *t, long *size, int nDimension)
{
int i;
long stride = 1;
t->nDimension = nDimension;
memcpy(t->size, size, nDimension * sizeof(*t->size));
for(i = nDimension - 1; i >= 0; i--)
{
t->stride[i] = stride;
stride *= t->size[i];
}
if(!t->storage)
t->storage = THFloatStorage_new(stride);
}
int THProcessImages(THNETWORK *network, unsigned char **images, int batchsize, int width, int height, int stride, float **results, int *outwidth, int *outheight, int bgr)
{
int i;
THFloatTensor *out, *t = 0;
THFloatStorage *st;
#ifdef CUDNN
if(network->net->cuda)
{
#ifdef HAVEFP16
if(floattype == CUDNN_DATA_HALF)
{
st = THCudaStorage_new(batchsize * (width * height * 3));
for(i = 0; i < batchsize; i++)
cuda_rgb2half((unsigned short *)st->data + i * (width * height * 3), images[i], width, height, stride, network->mean, network->std, bgr);
} else
#endif
{
st = THCudaStorage_new(batchsize * width * height * 3);
for(i = 0; i < batchsize; i++)
cuda_rgb2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std, bgr);
}
} else
#endif
#ifdef OPENCL
if(network->net->opencl)
t = OpenCL_LoadImage(images[0], width, height, stride, network->mean, network->std, bgr);
else
#endif
{
st = THFloatStorage_new(batchsize * width * height * 3);
if(bgr)
#pragma omp parallel for if(batchsize>1) private(i)
for(i = 0; i < batchsize; i++)
bgr2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std);
else
#pragma omp parallel for if(batchsize>1) private(i)
for(i = 0; i < batchsize; i++)
rgb2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std);
}
if(!t)
{
t = THFloatTensor_new();
t->storage = st;
if(batchsize == 1)
{
t->nDimension = 3;
t->size[0] = 3;
t->size[1] = height;
t->size[2] = width;
t->stride[0] = width * height;
t->stride[1] = width;
t->stride[2] = 1;
} else {
t->nDimension = 4;
t->size[0] = batchsize;
t->size[1] = 3;
t->size[2] = height;
t->size[3] = width;
t->stride[0] = 3 * width * height;
t->stride[1] = width * height;
t->stride[2] = width;
t->stride[3] = 1;
}
}
#ifdef CUDNN
if(network->net->cuda)
{
out = forward(network->net, t);
if(network->out)
THFloatTensor_free(network->out);
#ifdef HAVEFP16
if(floattype == CUDNN_DATA_HALF)
network->out = THFloatTensor_newFromHalfCudaTensor(out);
else
#endif
network->out = THFloatTensor_newFromCudaTensor(out);
out = network->out;
} else
#endif
#ifdef OPENCL
if(network->net->opencl)
{
out = forward(network->net, t);
if(network->out)
THFloatTensor_free(network->out);
#ifdef HAVEFP16
if(cl_datasize == 2)
network->out = THFloatTensor_newFromHalfOpenCLImageTensor(out);
else
#endif
network->out = THFloatTensor_newFromOpenCLImageTensor(out);
out = network->out;
} else
#endif
out = forward(network->net, t);
THFloatTensor_free(t);
*results = out->storage->data;
if(out->nDimension >= 3)
{
*outwidth = out->size[out->nDimension - 1];
*outheight = out->size[out->nDimension - 2];
} else *outwidth = *outheight = 1;
return THFloatTensor_nElement(out);
}
