Fixed bugs in CPU version of FFT

platforms/opencl/src/OpenCLFFT3D.cpp

@@ -210,8 +210,7 @@ cl::Kernel OpenCLFFT3D::createKernel(int xsize, int ysize, int zsize) {
     replacements["ZSIZE"] = OpenCLExpressionUtilities::intToString(zsize);
     replacements["M_PI"] = OpenCLExpressionUtilities::doubleToString(M_PI);
     replacements["COMPUTE_FFT"] = source.str();
-    if (loopRequired)
-        replacements["LOOP_REQUIRED"] = "1";
+    replacements["LOOP_REQUIRED"] = (loopRequired ? "1" : "0");
     cl::Program program = context.createProgram(context.replaceStrings(OpenCLKernelSources::fft, replacements));
     cl::Kernel kernel(program, "execFFT");
     kernel.setArg(3, zsize*sizeof(mm_float2), NULL);

platforms/opencl/src/kernels/fft.cl

@@ -14,7 +14,7 @@ __kernel void execFFT(__global const float2* restrict in, __global float2* restr
     for (int index = get_group_id(0); index < XSIZE*YSIZE; index += get_num_groups(0)) {
         int x = index/YSIZE;
         int y = index-x*YSIZE;
-#ifdef LOOP_REQUIRED
+#if LOOP_REQUIRED
         for (int z = get_local_id(0); z < ZSIZE; z += get_local_size(0))
             data0[z] = in[x*(YSIZE*ZSIZE)+y*ZSIZE+z];
 #else