How do I call runtime functions from Halide::Runtime::Buffer?

[mcourteaux]

I'm looking into an optimization that involves calling to halide_buffer_copy(). This function is part of the runtime and is defined in the device_interface.cpp file, which gets processed by the LLVM_Runtime_Linker.

However, this function is not always available as the runtime somehow is not always linked at link-time, but can be only produced later, when JIT-ing. Is there a nice way to call some of the common "Univseral CPP Initmods" functions from within Halide::Runtime::Buffer? It's sort of counterintuitive that the runtime is not available from within one of the runtime headers.

---

To avoid an A-B problem, what I was trying to do is replace all of this in Halide::Runtime::Buffer::copy_from():

Halide/src/runtime/HalideBuffer.h

Lines 1432 to 1454 in 9ba1829

	if (T_is_void ? (type().bytes() == 1) : (sizeof(not_void_T) == 1)) {
	using MemType = uint8_t;
	auto &typed_dst = reinterpret_cast<Buffer<MemType, Dims, InClassDimStorage> &>(dst);
	auto &typed_src = reinterpret_cast<Buffer<const MemType, D2, S2> &>(src);
	typed_dst.for_each_value([&](MemType &dst, MemType src) { dst = src; }, typed_src);
	} else if (T_is_void ? (type().bytes() == 2) : (sizeof(not_void_T) == 2)) {
	using MemType = uint16_t;
	auto &typed_dst = reinterpret_cast<Buffer<MemType, Dims, InClassDimStorage> &>(dst);
	auto &typed_src = reinterpret_cast<Buffer<const MemType, D2, S2> &>(src);
	typed_dst.for_each_value([&](MemType &dst, MemType src) { dst = src; }, typed_src);
	} else if (T_is_void ? (type().bytes() == 4) : (sizeof(not_void_T) == 4)) {
	using MemType = uint32_t;
	auto &typed_dst = reinterpret_cast<Buffer<MemType, Dims, InClassDimStorage> &>(dst);
	auto &typed_src = reinterpret_cast<Buffer<const MemType, D2, S2> &>(src);
	typed_dst.for_each_value([&](MemType &dst, MemType src) { dst = src; }, typed_src);
	} else if (T_is_void ? (type().bytes() == 8) : (sizeof(not_void_T) == 8)) {
	using MemType = uint64_t;
	auto &typed_dst = reinterpret_cast<Buffer<MemType, Dims, InClassDimStorage> &>(dst);
	auto &typed_src = reinterpret_cast<Buffer<const MemType, D2, S2> &>(src);
	typed_dst.for_each_value([&](MemType &dst, MemType src) { dst = src; }, typed_src);
	} else {
	assert(false && "type().bytes() must be 1, 2, 4, or 8");
	}

with one simple call to halide_copy_buffer() which is actually a quite-optimized procedure for this task (as opposed to the for_each_value([&]....) stuff):

halide_buffer_copy(nullptr, src, dst.buf.device_interface, dst);

[abadams]

The for_each_value stuff, at the time it was written at least, autovectorizes nicely so it's a lot more efficient than it looks.

Are you using Halide as a JIT compiler? If you're using it as an AOT compiler that function should be linked. If it has been dead-stripped, we may be annotating it incorrectly. In general the global runtime functions are only available from JIT via methods on Internal::JITSharedRuntime. Unfortunately this is both in the internal namespace and is just a few memoization cache functions right now.

The device API functions are available from JIT-land via the functions in DeviceInterface.h

Halide::Runtime::Buffer may be used from JIT-land or AOT-land, so unfortunately it can't depend on things in the runtime other than those accessed via the device interface pointer.

[mcourteaux]

Okay, so the answer is: not possible. Thanks for the quick reply!