Halide multi-pipeline CUDA initialization bug

.gitignore

hl
process
initialize_halide.o
visualize_halide.o
initialize_halide.h
visualize_halide.h

README.md

Linking multiple Halide-compiled modules together into a single app, they seem to be sharing/clobbering some global state which should be per-module state. As a result, only the first module’s compiled PTX is ever actually getting loaded, so attempts to grab subsequent kernels fail since they were indeed never loaded.

hl.cpp

#include <Halide.h>
using namespace Halide;

int main (int argc, char const *argv[])
{
    {
    	Var x("x"), y("y");
    	Func output("initialize");

    	output(x,y) = cast<uint16_t>(25535);

       Target target = get_target_from_environment();
        if(target.features & (Target::CUDA | Target::OpenCL)) {
     output.compute_root().cuda_tile(x, y, 16, 16);
    	} else {
     output.compute_root();
    	}

       output.compile_to_file("initialize_halide" );
    }

    {
        ImageParam input(UInt(16), 2);
    	Var x("x"), y("y"), c("c");
    	Func output("visualize");

    	//output(x,y,c) = input(x,y); // don't even use input for now!
    	output(x,y,c) = select(c == 0, cast<uint16_t>(20000),
    cast<uint16_t>(60000)); // just a constant

       Target target = get_target_from_environment();
    	if(target.features & (Target::CUDA | Target::OpenCL)) {
     output.compute_root().cuda_tile(x, y, 16, 16 );
    	} else {
     output.compute_root();
    	}

       output.compile_to_file("visualize_halide", input );
    }	

    return 0;
}

Makefile

HALIDE_DIR ?= ~/Halide

hl: hl.cpp
	g++ -I$(HALIDE_DIR)/include/ hl.cpp $(HALIDE_DIR)/bin/$(BUILD_PREFIX)/libHalide.a -ldl -ohl

visualize_halide.o: hl
	HL_TARGET=x86-64-osx-cuda-gpu_debug ./hl

process: process.cpp visualize_halide.o
	g++ -I$(HALIDE_DIR)/apps/support/ *.o process.cpp -lpng -L/usr/local/cuda/lib -lcuda -oprocess

process.cpp

#include "initialize_halide.h"
#include "visualize_halide.h"
#include <static_image.h>
#include <image_io.h>

int main (int argc, char const *argv[])
{
    const int width = 256;
    const int height = 256;

    // For rendering an output
    Image<uint16_t> render(width, height, 3);

    // Initialize test
    Image<uint16_t> test(width, height);
    initialize_halide(test); // initialize with a constant grey value (result not used in visualize!)
    // ^^^^^^^^ if I comment this out, then visualize_halide works!

    visualize_halide(test, render);
    //render.copy_to_host(); // does not matter if I use it or not
    save(render, "visual.png"); // garbled output
    
    return 0;
}