The code snippit calcs a 3D vector magnitude: mag= sqrt(X*X + Y*Y + Z*Z)
// Note: blockDim.x = 300, gMem= ptr to global mem chunk
__shared__ sMem[100];
float regA;
for (j=0; j<50; j++) {
if(threadIdx.x < 3) {
regA= gMem[j];
sMem[threadIdx.x]= regA*regA; // Line A write 5193
}
__syncthreads();
if(threadIdx.x == 0) {
regA= sMem[0]; // Line B read 5197
regA+= sMem[1];
regA+= sMem[2]; // Line C read 5199
sMem[0]= sqrt(regA);
}
}
The syncthreads is preventing RAW hazards. W/o it, I get RAW & WAR hazard warnings. W/ it, I still get WAR warnings. The warning is coming from line pairs AB & AC.
WARN:(Warp Level Programming) Potential WAR hazard detected at __shared__ 0x30b in block (0, 0, 0) :
Read Thread (0, 0, 0) at 0x000000b0 in /src/trap.cu:5199:Mag(float const *, float const *, int, float*, int)
Write Thread (2, 0, 0) at 0x00000080 in /src/trap.cu:5193:Mag(float const *, float const *, int, float*, int)
Current Value : 64, Incoming Value : 66
WARN:(Warp Level Programming) Potential WAR hazard detected at __shared__ 0x307 in block (0, 0, 0) :
Read Thread (0, 0, 0) at 0x000000a8 in /src/trap.cu:5197::Mag(float const *, float const *, int, float*, int)
Write Thread (1, 0, 0) at 0x00000080 in /src/trap.cu:5193::Mag(float const *, float const *, int, float*, int)
Current Value : 67, Incoming Value : 66
Why doesn't the __synthreads(); prevent the WAR hazards?
There is no intervening __synthreads() when the code executes lines 5197 and 5199 and then loops back to line 5193.
User contributions licensed under CC BY-SA 3.0