Compile both raw and object variants of LTO-IR

pynvjitlink/tests/conftest.py

@@ -84,6 +84,11 @@ def device_functions_ltoir():
     return read_test_file("test_device_functions.ltoir")
 
 
+@pytest.fixture(scope="session")
+def device_functions_ltoir_object():
+    return read_test_file("test_device_functions.ltoir.o")
+
+
 @pytest.fixture(scope="session")
 def device_functions_object():
     return read_test_file("test_device_functions.o")

pynvjitlink/tests/test_pynvjitlink.py

@@ -63,8 +63,8 @@ def test_add_file(input_file, input_type, gpu_arch_flag, request):
 # We test the LTO input case separately as it requires the `-lto` flag. The
 # OBJECT input type is used because the LTO-IR container is packaged in an ELF
 # object when produced by NVCC.
-def test_add_file_lto(device_functions_ltoir, gpu_arch_flag):
-    filename, data = device_functions_ltoir
+def test_add_file_lto(device_functions_ltoir_object, gpu_arch_flag):
+    filename, data = device_functions_ltoir_object
 
     handle = _nvjitlinklib.create(gpu_arch_flag, "-lto")
     _nvjitlinklib.add_data(handle, InputType.OBJECT.value, data, filename)
@@ -123,11 +123,11 @@ def test_get_linked_cubin_link_not_complete_error(
     _nvjitlinklib.destroy(handle)
 
 
-def test_get_linked_cubin_from_lto(device_functions_ltoir, gpu_arch_flag):
-    filename, data = device_functions_ltoir
-    # device_functions_ltoir is a host object containing a fatbin containing an
-    # LTOIR container, because that is what NVCC produces when LTO is
-    # requested. So we need to use the OBJECT input type, and the linker
+def test_get_linked_cubin_from_lto(device_functions_ltoir_object, gpu_arch_flag):
+    filename, data = device_functions_ltoir_object
+    # device_functions_ltoir_object is a host object containing a fatbin
+    # containing an LTOIR container, because that is what NVCC produces when
+    # LTO is requested. So we need to use the OBJECT input type, and the linker
     # retrieves the LTO IR from it because we passed the -lto flag.
     input_type = InputType.OBJECT.value
     handle = _nvjitlinklib.create(gpu_arch_flag, "-lto")
@@ -140,11 +140,11 @@ def test_get_linked_cubin_from_lto(device_functions_ltoir, gpu_arch_flag):
     assert cubin[:4] == b"\x7fELF"
 
 
-def test_get_linked_ptx_from_lto(device_functions_ltoir, gpu_arch_flag):
-    filename, data = device_functions_ltoir
-    # device_functions_ltoir is a host object containing a fatbin containing an
-    # LTOIR container, because that is what NVCC produces when LTO is
-    # requested. So we need to use the OBJECT input type, and the linker
+def test_get_linked_ptx_from_lto(device_functions_ltoir_object, gpu_arch_flag):
+    filename, data = device_functions_ltoir_object
+    # device_functions_ltoir_object is a host object containing a fatbin
+    # containing an LTOIR container, because that is what NVCC produces when
+    # LTO is requested. So we need to use the OBJECT input type, and the linker
     # retrieves the LTO IR from it because we passed the -lto flag.
     input_type = InputType.OBJECT.value
     handle = _nvjitlinklib.create(gpu_arch_flag, "-lto", "-ptx")
@@ -154,9 +154,9 @@ def test_get_linked_ptx_from_lto(device_functions_ltoir, gpu_arch_flag):
     _nvjitlinklib.destroy(handle)
 
 
-def test_get_linked_ptx_link_not_complete_error(device_functions_ltoir, gpu_arch_flag):
+def test_get_linked_ptx_link_not_complete_error(device_functions_ltoir_object, gpu_arch_flag):
     handle = _nvjitlinklib.create(gpu_arch_flag, "-lto", "-ptx")
-    filename, data = device_functions_ltoir
+    filename, data = device_functions_ltoir_object
     input_type = InputType.OBJECT.value
     _nvjitlinklib.add_data(handle, input_type, data, filename)
     with pytest.raises(RuntimeError, match="NVJITLINK_ERROR_INTERNAL error"):

test_binary_generation/Makefile

@@ -41,11 +41,14 @@ all:
 	nvcc $(NVCC_FLAGS) $(CUBIN_FLAGS) -o $(OUTPUT_DIR)/undefined_extern.cubin undefined_extern.cu
 	nvcc $(NVCC_FLAGS) $(CUBIN_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.cubin test_device_functions.cu
 	nvcc $(NVCC_FLAGS) $(FATBIN_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.fatbin test_device_functions.cu
-	nvcc $(NVCC_FLAGS) $(LTOIR_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.ltoir test_device_functions.cu
 	nvcc $(NVCC_FLAGS) $(PTX_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.ptx test_device_functions.cu
 	nvcc $(NVCC_FLAGS) $(OBJECT_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.o test_device_functions.cu
 	nvcc $(NVCC_FLAGS) $(LIBRARY_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.a test_device_functions.cu
 
+	# Generate LTO-IR wrapped in a fatbin
+	nvcc $(NVCC_FLAGS) $(LTOIR_FLAGS) -o $(OUTPUT_DIR)/test_device_functions.ltoir.o test_device_functions.cu
+	# Generate LTO-IR in a "raw" LTO-IR container
+	python generate_raw_ltoir.py --arch sm_$(GPU_CC) -o $(OUTPUT_DIR)/test_device_functions.ltoir test_device_functions.cu
 	# We also want to test linking a .cu file; this needs no compilation,
 	# so copy it instead
 	cp test_device_functions.cu $(OUTPUT_DIR)