Perftest: keep CUDA context creation ABI-stable

cuGetProcAddress resolves versioned driver entry points from the
requested CUDA version, not from the headers used to compile perftest.
Request the CUDA 3.2 cuCtxCreate ABI explicitly and keep the function
pointer plus call site on the three-argument cuCtxCreate_v2 signature,
even when CUDA 13 headers expose a newer default cuCtxCreate prototype.

The stale configure-generated CUDA_VER macro is removed because source
now uses CUDA_VERSION from cuda.h for compile-time checks. The existing
CUDA_VER_* constants remain local names for explicit driver-entry ABI
request versions.

Signed-off-by: Zelong Yue <yuezelong@bytedance.com>

Author: Yue-ByteDance

configure.ac

@@ -369,7 +369,6 @@ if test "$cuda_found" = "yes"; then
 	AC_DEFINE_UNQUOTED([CUDA_PATH], "$cuda_h_path" , [Enable CUDA feature])
 	AC_CHECK_LIB([cuda], [cuMemGetHandleForAddressRange], [HAVE_CUDA_CUMEMGETHANDLEFORADDRESSRANGE=yes], [HAVE_CUDA_CUMEMGETHANDLEFORADDRESSRANGE=no])
 	cuda_toolkit_version=`grep "define CUDA_VERSION" $cuda_h_path | cut -d' ' -f3`
-	AC_DEFINE_UNQUOTED([CUDA_VER], [$cuda_toolkit_version], [Define CUDA_VER])
 	AC_TRY_LINK([
 	#include <$cuda_h_path>],
 	[int x = CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD|CU_DEVICE_ATTRIBUTE_DMA_BUF_SUPPORTED;],

src/cuda_loader.c

@@ -10,7 +10,7 @@ CUresult (*p_cuDeviceGetCount)(int *) = NULL;
 CUresult (*p_cuDeviceGet)(CUdevice *, int) = NULL;
 CUresult (*p_cuDeviceGetAttribute)(int *, CUdevice_attribute, CUdevice) = NULL;
 CUresult (*p_cuDeviceGetName)(char *, int, CUdevice) = NULL;
-CUresult (*p_cuCtxCreate)(CUcontext *, unsigned int, CUdevice) = NULL;
+CUresult (*p_cuCtxCreate_v2)(CUcontext *, unsigned int, CUdevice) = NULL;
 CUresult (*p_cuDevicePrimaryCtxRetain)(CUcontext *, CUdevice) = NULL;
 CUresult (*p_cuCtxSetCurrent)(CUcontext) = NULL;
 CUresult (*p_cuCtxDestroy)(CUcontext) = NULL;
@@ -25,7 +25,7 @@ CUresult (*p_cuMemcpyDtoD)(CUdeviceptr, CUdeviceptr, size_t) = NULL;
 CUresult (*p_cuMemGetHandleForAddressRange)(void *, void *, size_t, CUmemRangeHandleType, unsigned int) = NULL;
 #endif
 CUresult (*p_cuDriverGetVersion)(int* driverVersion) = NULL;
-#if CUDA_VER >= 12000
+#if CUDA_VERSION >= 12000
 CUresult (*p_cuGetProcAddress)(const char* symbol, void** pfn, int  cudaVersion, uint64_t flags, CUdriverProcAddressQueryResult* symbolStatus) = NULL;
 #else
 CUresult (*p_cuGetProcAddress)(const char* symbol, void** pfn, int  cudaVersion, uint64_t flags) = NULL;
@@ -34,7 +34,7 @@ CUresult (*p_cuMemAllocManaged)(CUdeviceptr* dptr, size_t bytesize, unsigned int
 CUresult (*p_cuCtxSynchronize) (void) = NULL;
 
 int load_cuda_function(void **func_ptr, const char *func_name, int version) {
-    #if CUDA_VER >= 12000
+    #if CUDA_VERSION >= 12000
     CUresult res = p_cuGetProcAddress(func_name, func_ptr, version, 0, NULL);
     #else
     CUresult res = p_cuGetProcAddress(func_name, func_ptr, version, 0);
@@ -69,7 +69,8 @@ int load_cuda_library(void) {
         { (void**)&p_cuDeviceGet,                 "cuDeviceGet",                  CUDA_VER_2_0  },
         { (void**)&p_cuDeviceGetAttribute,        "cuDeviceGetAttribute",         CUDA_VER_2_0  },
         { (void**)&p_cuDeviceGetName,             "cuDeviceGetName",              CUDA_VER_2_0  },
-        { (void**)&p_cuCtxCreate,                 "cuCtxCreate",                  CUDA_VER_3_2  },
+        /* CUDA_VER_3_2 selects the cuCtxCreate_v2 ABI across CUDA 11-13. */
+        { (void**)&p_cuCtxCreate_v2,              "cuCtxCreate",                  CUDA_VER_3_2  },
         { (void**)&p_cuDevicePrimaryCtxRetain,    "cuDevicePrimaryCtxRetain",     CUDA_VER_7_0  },
         { (void**)&p_cuCtxSetCurrent,             "cuCtxSetCurrent",              CUDA_VER_4_0  },
         { (void**)&p_cuCtxDestroy,                "cuCtxDestroy",                 CUDA_VER_4_0  },

src/cuda_loader.h

@@ -30,7 +30,7 @@ extern CUresult (*p_cuDeviceGetCount)(int *);
 extern CUresult (*p_cuDeviceGet)(CUdevice *, int);
 extern CUresult (*p_cuDeviceGetAttribute)(int *, CUdevice_attribute, CUdevice);
 extern CUresult (*p_cuDeviceGetName)(char *, int, CUdevice);
-extern CUresult (*p_cuCtxCreate)(CUcontext *, unsigned int, CUdevice);
+extern CUresult (*p_cuCtxCreate_v2)(CUcontext *, unsigned int, CUdevice);
 extern CUresult (*p_cuDevicePrimaryCtxRetain)(CUcontext *, CUdevice);
 extern CUresult (*p_cuCtxSetCurrent)(CUcontext);
 extern CUresult (*p_cuCtxDestroy)(CUcontext);
@@ -47,7 +47,7 @@ extern CUresult (*p_cuMemGetHandleForAddressRange)(void *, void *, size_t, CUmem
 extern CUresult (*p_cuDriverGetVersion)(int* driverVersion);
 extern CUresult (*p_cuCtxSynchronize) (void);
 extern CUresult (*p_cuMemAllocManaged)(CUdeviceptr* dptr, size_t bytesize, unsigned int  flags);
-#if CUDA_VER >= 12000
+#if CUDA_VERSION >= 12000
 extern CUresult (*p_cuGetProcAddress)(const char* symbol, void** pfn, int  cudaVersion, uint64_t flags, CUdriverProcAddressQueryResult* symbolStatus);
 #else
 extern CUresult (*p_cuGetProcAddress)(const char* symbol, void** pfn, int  cudaVersion, uint64_t flags);

src/cuda_memory.c

@@ -94,9 +94,11 @@ static int init_gpu(struct cuda_memory_ctx *ctx)
 	printf("[pid = %d, dev = %d] device name = [%s]\n", getpid(), ctx->cuDevice, name);
 	printf("creating CUDA Ctx\n");
 
-	error = p_cuCtxCreate(&ctx->cuContext, CU_CTX_MAP_HOST, ctx->cuDevice);
+	/* Create context */
+	error = p_cuCtxCreate_v2(&ctx->cuContext, CU_CTX_MAP_HOST, ctx->cuDevice);
+
 	if (error != CUDA_SUCCESS) {
-		printf("cuCtxCreate() error=%d\n", error);
+		printf("cuCtxCreate_v2() error=%d\n", error);
 		return FAILURE;
 	}