fix(ops/jit): skip lookup include path when nvcc executable not found

GitOrigin-RevId: f5e7dce1c52fe099de8e90684bbe5b0566a5115a

imperative/python/test/unit/core/test_subgraph.py

 import functools
+import os
+import platform
+import subprocess
+import sys
 
 import numpy as np
 import pytest
@@ -152,3 +156,56 @@ def test_subgraph_jit_backward():
         y1 = x1 * x1
         y2 = mul(x2, x2)
         gm.backward(y2)
+
+
+@pytest.mark.skipif(
+    platform.system() != "Linux", reason="jit fusion is only available on Linux",
+)
+def test_subgraph_jit():
+    prog = """
+import megengine
+import numpy as np
+from megengine.core.tensor.utils import subgraph_fn
+
+# 3 * 4 * 5 > MEGDNN_MAX_NDIM
+x_np = np.random.rand(3, 4, 5).astype("float32")
+x1 = megengine.Tensor(x_np)
+x2 = megengine.Tensor(x_np)
+
+@subgraph_fn(
+    "Mul",
+    dtype=x1.dtype,
+    device=x1.device,
+    nr_inputs=2,
+    gopt_level=None,
+    jit_fusion=True,
+    custom_grad=True,
+)
+def mul(inputs, f, c):
+    x, y = inputs[0:2]
+    z = f("*", x, y)
+    (dz,) = yield (z,)
+    dx = f("*", dz, y)
+    dy = f("*", dz, x)
+    yield (dx, dy)
+
+y, = mul(x1, x2)
+
+# ensure execution
+y.numpy()
+"""
+    env = dict(os.environ)
+    if "PATH" in env:
+        # remove nvcc from environ["PATH"]
+        path = env["PATH"]
+        paths = path.split(os.pathsep)
+        paths = [
+            path
+            for path in paths
+            if not (os.path.isdir(path) and "nvcc" in os.listdir(path))
+        ]
+        path = os.pathsep.join(paths)
+        env["PATH"] = path
+    # previous program may be stored in persistent cache
+    env["MGE_FASTRUN_CACHE_TYPE"] = "MEMORY"
+    subprocess.check_call([sys.executable, "-c", prog], env=env)

src/core/impl/utils/cuda_helper.cpp

@@ -17,6 +17,8 @@ using namespace mgb;
 #include <unistd.h>
 #endif
 
+namespace {
+
 #ifndef PATH_MAX
 #define PATH_MAX 4096
 #endif
@@ -32,6 +34,7 @@ using namespace mgb;
 #define PATH_SPLITER   '\\'
 #define ENV_PATH       "Path"
 #define NVCC_EXE       "nvcc.exe"
+
 void* dlopen(const char* file, int) {
     return static_cast<void*>(LoadLibrary(file));
 }
@@ -108,15 +111,15 @@ std::string find_file_in_envs_with_intmd(
 std::string get_nvcc_root_path() {
     auto nvcc_root_path = find_file_in_envs_with_intmd({ENV_PATH}, NVCC_EXE);
     if (nvcc_root_path.empty()) {
-        mgb_throw(
-                MegBrainError,
-                "nvcc not found. Add your nvcc to your environment Path");
+        return {};
     } else {
         auto idx = nvcc_root_path.rfind(PATH_SPLITER);
         return nvcc_root_path.substr(0, idx + 1);
     }
 }
 
+}  // namespace
+
 std::vector<std::string> mgb::get_cuda_include_path() {
 #if MGB_CUDA
     std::vector<std::string> paths;
@@ -131,13 +134,15 @@ std::vector<std::string> mgb::get_cuda_include_path() {
 
     // 2. use nvcc path
     auto nvcc_path = get_nvcc_root_path();
-    auto cudart_header_path = nvcc_path + ".." + PATH_SPLITER + "include" +
-                              PATH_SPLITER + "cuda_runtime.h";
-    //! double check path_to_nvcc/../include/cuda_runtime.h exists
-    auto ret = check_file_exist(cudart_header_path.c_str(), F_OK);
-    if (ret == 0) {
-        paths.emplace_back(nvcc_path + "..");
-        paths.emplace_back(nvcc_path + ".." + PATH_SPLITER + "include");
+    if (!nvcc_path.empty()) {
+        auto cudart_header_path = nvcc_path + ".." + PATH_SPLITER + "include" +
+                                  PATH_SPLITER + "cuda_runtime.h";
+        //! double check path_to_nvcc/../include/cuda_runtime.h exists
+        auto ret = check_file_exist(cudart_header_path.c_str(), F_OK);
+        if (ret == 0) {
+            paths.emplace_back(nvcc_path + "..");
+            paths.emplace_back(nvcc_path + ".." + PATH_SPLITER + "include");
+        }
     }
 
     // 3. use libcudart.so library path
@@ -161,4 +166,4 @@ std::vector<std::string> mgb::get_cuda_include_path() {
 #else
     mgb_throw(MegBrainError, "cuda disabled at compile time");
 #endif
-}
\ No newline at end of file
+}