std::vector<Halide::Buffer<> > halideBuffers(const std::vector<Ptr<BackendWrapper> >& ptrs)
{
std::vector<Halide::Buffer<> > vec;
vec.reserve(ptrs.size());
for (const Ptr<BackendWrapper>& ptr : ptrs)
{
vec.push_back(halideBuffer(ptr));
}
return vec;
}
HalideBackendWrapper::HalideBackendWrapper(const Ptr<BackendWrapper>& base,
const MatShape& shape)
: BackendWrapper(DNN_BACKEND_HALIDE, base->targetId)
{
managesDevMemory = false;
Halide::Buffer<float> baseBuffer = halideBuffer(base);
buffer = Halide::Buffer<float>((float*)baseBuffer.raw_buffer()->host,
getBufferShape(shape));
if (baseBuffer.has_device_allocation())
{
buffer.raw_buffer()->device = baseBuffer.raw_buffer()->device;
buffer.raw_buffer()->device_interface = baseBuffer.raw_buffer()->device_interface;
buffer.set_device_dirty();
}
else
{
buffer.set_host_dirty(); // Indicate that data is on CPU.
CV_Assert(targetId == DNN_TARGET_CPU);
}
}
virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
{
#ifdef HAVE_HALIDE
int inW, inH, inC, inN;
int minN = std::max(dstRanges[0].start, 0);
int minC = std::max(dstRanges[1].start, 0);
int minY = std::max(dstRanges[2].start, 0);
int minX = std::max(dstRanges[3].start, 0);
Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
getCanonicalSize(inputBuffer, &inW, &inH, &inC, &inN);
Halide::Var x("x"), y("y"), c("c"), n("n");
Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
Halide::Func padded =
Halide::BoundaryConditions::constant_exterior(inputBuffer, paddingValue);
top(x, y, c, n) = padded(x - minX, y - minY, c - minC, n - minN);
return Ptr<BackendNode>(new HalideBackendNode(top));
#endif // HAVE_HALIDE
return Ptr<BackendNode>();
}
HalideBackendWrapper::HalideBackendWrapper(const Ptr<BackendWrapper>& base,
const MatShape& shape)
: BackendWrapper(DNN_BACKEND_HALIDE, base->targetId)
{
int w, h, c, n;
getCanonicalSize(shape, &w, &h, &c, &n);
Halide::Buffer<float> baseBuffer = halideBuffer(base);
buffer = Halide::Buffer<float>((float*)baseBuffer.raw_buffer()->host,
{w, h, c, n});
if (baseBuffer.has_device_allocation())
{
buffer.raw_buffer()->device = baseBuffer.raw_buffer()->device;
buffer.raw_buffer()->device_interface = baseBuffer.raw_buffer()->device_interface;
buffer.set_device_dirty();
}
else
{
buffer.set_host_dirty(); // Indicate that data is on CPU.
CV_Assert(targetId == DNN_TARGET_CPU);
}
}
virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
{
#ifdef HAVE_HALIDE
Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
int inW, inH, inC, inN;
getCanonicalSize(inputBuffer, &inW, &inH, &inC, &inN);
if (inW != 1 || inH != 1)
CV_Error(cv::Error::StsNotImplemented,
"Halide backend for SoftMax with spatial size "
"more than 1x1 is not implemented");
Halide::Var x("x"), y("y"), c("c"), n("n");
Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
Halide::Func expInput("expInput");
Halide::RDom r(0, inW, 0, inH, 0, inC);
expInput(x, y, c, n) = exp(inputBuffer(x, y, c, n));
Halide::Expr globalSum = sum(expInput(r.x, r.y, r.z, n));
top(x, y, c, n) = expInput(x, y, c, n) / globalSum;
return Ptr<BackendNode>(new HalideBackendNode(top));
#endif // HAVE_HALIDE
return Ptr<BackendNode>();
}
