static void *do_init(CUdevice dev, int flags, int *ret) {
cuda_context *res;
CUcontext ctx;
unsigned int fl = CU_CTX_SCHED_AUTO;
int i;
CHKFAIL(NULL);
if (flags & GA_CTX_SINGLE_THREAD)
fl = CU_CTX_SCHED_SPIN;
if (flags & GA_CTX_MULTI_THREAD)
fl = CU_CTX_SCHED_YIELD;
err = cuDeviceGetAttribute(&i, CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING, dev);
CHKFAIL(NULL);
if (i != 1)
FAIL(NULL, GA_UNSUPPORTED_ERROR);
err = cuCtxCreate(&ctx, fl, dev);
CHKFAIL(NULL);
res = cuda_make_ctx(ctx, 0);
if (res == NULL) {
cuCtxDestroy(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
res->flags |= flags;
/* Don't leave the context on the thread stack */
cuCtxPopCurrent(NULL);
return res;
}
static void *cuda_init(int ord, int flags, int *ret) {
CUdevice dev;
CUcontext ctx;
cuda_context *res;
static int init_done = 0;
unsigned int fl = CU_CTX_SCHED_AUTO;
if (ord == -1) {
/* Grab the ambient context */
err = cuCtxGetCurrent(&ctx);
CHKFAIL(NULL);
res = cuda_make_ctx(ctx, DONTFREE);
if (res == NULL) {
FAIL(NULL, GA_IMPL_ERROR);
}
res->flags |= flags;
return res;
}
if (!init_done) {
err = cuInit(0);
CHKFAIL(NULL);
init_done = 1;
}
err = cuDeviceGet(&dev, ord);
CHKFAIL(NULL);
if (flags & GA_CTX_SINGLE_THREAD)
fl = CU_CTX_SCHED_SPIN;
if (flags & GA_CTX_MULTI_THREAD)
fl = CU_CTX_SCHED_YIELD;
err = cuCtxCreate(&ctx, fl, dev);
CHKFAIL(NULL);
res = cuda_make_ctx(ctx, 0);
res->flags |= flags;
if (res == NULL) {
cuCtxDestroy(ctx);
FAIL(NULL, GA_IMPL_ERROR);
}
/* Don't leave the context on the thread stack */
cuCtxPopCurrent(NULL);
return res;
}
static void *cuda_init(int ord, int flags, int *ret) {
CUdevice dev;
cuda_context *res;
static int init_done = 0;
if (ord == -2) {
CUcontext ctx;
/* Grab the ambient context */
err = cuCtxGetCurrent(&ctx);
CHKFAIL(NULL);
/* If somebody made a context, then the api is initialized */
init_done = 1;
res = cuda_make_ctx(ctx, DONTFREE);
if (res == NULL) {
FAIL(NULL, GA_IMPL_ERROR);
}
res->flags |= flags;
return res;
}
if (!init_done) {
err = cuInit(0);
CHKFAIL(NULL);
init_done = 1;
}
if (ord == -1) {
int i, c;
err = cuDeviceGetCount(&c);
CHKFAIL(NULL);
for (i = 0; i < c; i++) {
err = cuDeviceGet(&dev, i);
CHKFAIL(NULL);
res = do_init(dev, flags, NULL);
if (res != NULL)
return res;
}
FAIL(NULL, GA_NODEV_ERROR);
} else {
err = cuDeviceGet(&dev, ord);
CHKFAIL(NULL);
return do_init(dev, flags, ret);
}
}
