|
|
@@ -283,7 +283,7 @@ __global__ void cudaMulmod(GlobalThreadState *global_ts,
|
|
|
// Check for a distinguished point
|
|
|
asm("mov.u32 %0, $xr0;" : "=r" (xlow));
|
|
|
if ((xlow & 0x000003ff) == 0 || !(order_2|order_1) && order_0 < (1<<20)) {
|
|
|
- unsigned int *ourbuffer = DPstreamWrite(a_0,a_1,a_2,b_0,b_1,b_2);
|
|
|
+ unsigned int *ourbuffer = DPstreamAlloc();
|
|
|
if (ourbuffer) {
|
|
|
CIOS_WRITE_DP(ourbuffer, a_0, a_1, a_2, b_0, b_1, b_2);
|
|
|
}
|