int main(int argc, char **argv) {
ImageParam in(UInt(8), 3, "input");
Var x("x"), y("y"), c("c"), x1, y1;
Func f("f"), g("g"), h("h"), k("k");
f(x, y, c) = cast<uint8_t>(255 - in(x, y, c));
g(x, y, c) = cast<uint8_t>(2*in(x, y, c));
h(x, y, c) = f(x, y, c) + g(x, y, c);
k(x, y, c) = f(x, y, c) - g(x, y, c);
f.reorder(c, x, y);
g.reorder(c, x, y);
h.reorder(c, x, y);
k.reorder(c, x, y);
//g.compute_with(f, y);
f.gpu_tile(x, y, x1, y1, 16, 16);
g.gpu_tile(x, y, x1, y1, 16, 16);
h.gpu_tile(x, y, x1, y1, 16, 16);
k.gpu_tile(x, y, x1, y1, 16, 16);
Halide::Target target = Halide::get_host_target();
target.set_feature(Halide::Target::CUDA, true);
Pipeline({f, g, h, k}).compile_to_object("build/generated_fct_fusiongpu_ref.o", {in}, "fusiongpu_ref", target);
Pipeline({f, g, h, k}).compile_to_lowered_stmt("build/generated_fct_fusiongpu_ref.txt", {in}, Halide::Text, target);
return 0;
}
int main(int argc, char* argv[]) {
ImageParam in{UInt(8), 3, "input"};
Func RGB2Gray{"RGB2Gray"};
Var x, y, c, x0, y0, x1, y1;
const Expr yuv_shift = cast<uint32_t>(14);
const Expr R2Y = cast<uint32_t>(4899);
const Expr G2Y = cast<uint32_t>(9617);
const Expr B2Y = cast<uint32_t>(1868);
RGB2Gray(x, y) = cast<uint8_t>(CV_DESCALE( (in(x, y, 2) * B2Y
+ in(x, y, 1) * G2Y
+ in(x, y, 0) * R2Y),
yuv_shift));
RGB2Gray.compute_root();
RGB2Gray.gpu_tile(x, y, x0, y0, x1, y1, 16, 16);
Halide::Target target = Halide::get_host_target();
target.set_feature(Target::Feature::CUDA, true);
RGB2Gray.compile_to_object("build/generated_fct_cvtcolorgpu_ref.o", {in}, "cvtcolorgpu_ref", target);
RGB2Gray.compile_to_lowered_stmt("build/generated_fct_cvtcolorgpu_ref.txt", {in}, Text, target);
return 0;
}
void compileHalide(const std::vector<Mat> &outputs, Ptr<BackendNode>& node, int targetId)
{
#ifdef HAVE_HALIDE
CV_Assert(!node.empty());
Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs.back();
int outW, outH, outC, outN;
Halide::Var x("x"), y("y"), c("c"), n("n");
getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);
top.bound(x, 0, outW).bound(y, 0, outH)
.bound(c, 0, outC).bound(n, 0, outN);
Halide::Target target = Halide::get_host_target();
target.set_feature(Halide::Target::NoAsserts);
if (targetId == DNN_TARGET_OPENCL)
{
target.set_feature(Halide::Target::OpenCL);
}
CV_Assert(target.supported());
top.compile_jit(target);
#endif // HAVE_HALIDE
}
int main(int argc, char **argv) {
Halide::Func theFunc = getFunction();
if (argc >= 3) {
std::vector<Halide::Argument> arguments = theFunc.infer_arguments();
Halide::Target target = Halide::get_target_from_environment();
target.set_feature(Halide::Target::Feature::UserContext);
theFunc.compile_to_object(argv[1] + std::string(".o"), arguments, argv[2], target);
return 0;
}
return 1;
}
HalideBackendWrapper::HalideBackendWrapper(int targetId, const cv::Mat& m)
: BackendWrapper(DNN_BACKEND_HALIDE, targetId)
{
managesDevMemory = true;
buffer = wrapToHalideBuffer(m);
if (targetId == DNN_TARGET_CPU)
{
return;
}
else if (targetId == DNN_TARGET_OPENCL)
{
Halide::Target t = Halide::get_host_target();
t.set_feature(Halide::Target::OpenCL);
buffer.copy_to_device(t);
}
else
CV_Error(Error::StsNotImplemented, "Unknown target identifier");
}
int main(int argc, char* argv[]) {
ImageParam in(UInt(8), 3, "input");
ImageParam kernelX(Float(32), 1, "kernelx");
ImageParam kernelY(Float(32), 1, "kernely");
Func gaussiangpu("gaussiangpu"), gaussiangpu_x("gaussiangpu_x");
Var x("x"), y("y"), c("c");
Expr e = 0.0f;
for (int i = 0; i < 5; ++i) {
e += cast<float>(in(x + i, y, c)) * kernelX(i);
}
gaussiangpu_x(x, y, c) = e;
Expr f = 0.0f;
for (int j = 0; j < 5; ++j) {
f += gaussiangpu_x(x, y + j, c) * kernelY(j);
}
gaussiangpu(x, y, c) = cast<uint8_t>(f);
// gaussiangpu_x.reorder(c, x, y);
gaussiangpu.reorder(c, x, y);
Var x_inner, y_inner, x_outer, y_outer, tile_index;
// // gaussiangpu.tile(x, y, x_outer, y_outer, x_inner, y_inner, 8, 8)
// // .fuse(x_outer, y_outer, tile_index)
// // .compute_root();
// // .parallel(x_outer);
// gaussiangpu_x.compute_root();
// gaussiangpu_x.gpu_tile(x, y, x_inner, y_inner, x_outer, y_outer, 32, 32);
gaussiangpu.gpu_tile(x, y, x_inner, y_inner, x_outer, y_outer, 32, 32);
Halide::Target target = Halide::get_host_target();
target.set_feature(Target::Feature::CUDA, true);
gaussiangpu.compile_to_object("build/generated_fct_gaussiangpu_ref.o", {in, kernelX, kernelY}, "gaussiangpu_ref", target);
gaussiangpu.compile_to_lowered_stmt("build/generated_fct_gaussiangpu_ref.txt", {in, kernelX, kernelY}, Text, target);
return 0;
}