| 64 | |
| 65 | template <typename T> |
| 66 | void |
| 67 | axpyCorrectnessTest(TestParams *params) |
| 68 | { |
| 69 | cl_int err; |
| 70 | T *X, *Y; //For OpenCL implementation |
| 71 | T *blasX, *blasY;// For reference implementation |
| 72 | cl_mem bufX, bufY; |
| 73 | clMath::BlasBase *base; |
| 74 | cl_event *events; |
| 75 | T alpha; |
| 76 | |
| 77 | base = clMath::BlasBase::getInstance(); |
| 78 | |
| 79 | if ((typeid(T) == typeid(cl_double) || |
| 80 | typeid(T) == typeid(DoubleComplex)) && |
| 81 | !base->isDevSupportDoublePrecision()) { |
| 82 | |
| 83 | std::cerr << ">> WARNING: The target device doesn't support native " |
| 84 | "double precision floating point arithmetic" << |
| 85 | std::endl << ">> Test skipped" << std::endl; |
| 86 | SUCCEED(); |
| 87 | return; |
| 88 | } |
| 89 | |
| 90 | printf("number of command queues : %d\n\n", params->numCommandQueues); |
| 91 | |
| 92 | events = new cl_event[params->numCommandQueues]; |
| 93 | memset(events, 0, params->numCommandQueues * sizeof(cl_event)); |
| 94 | |
| 95 | size_t lengthX = (1 + ((params->N -1) * abs(params->incx))); |
| 96 | size_t lengthY = (1 + ((params->N -1) * abs(params->incy))); |
| 97 | |
| 98 | X = new T[lengthX + params->offBX ]; |
| 99 | Y = new T[lengthY + params->offCY ]; |
| 100 | blasX = new T[lengthX + params->offBX ]; |
| 101 | blasY = new T[lengthY + params->offCY ]; |
| 102 | |
| 103 | if((X == NULL) || (blasX == NULL) || (Y == NULL) || (blasY == NULL)) |
| 104 | { |
| 105 | ::std::cerr << "Cannot allocate memory on host side\n" << "!!!!!!!!!!!!Test skipped.!!!!!!!!!!!!" << ::std::endl; |
| 106 | deleteBuffers<T>(X, Y, blasX, blasY); |
| 107 | delete[] events; |
| 108 | SUCCEED(); |
| 109 | return; |
| 110 | } |
| 111 | |
| 112 | srand(params->seed); |
| 113 | |
| 114 | // Populate X and Y |
| 115 | randomVectors(params->N, (X+params->offBX), params->incx, (Y+params->offCY), params->incy); |
| 116 | |
| 117 | memcpy(blasX, X, (lengthX + params->offBX) * sizeof(T)); |
| 118 | memcpy(blasY, Y, (lengthY + params->offCY) * sizeof(T)); |
| 119 | alpha = convertMultiplier<T>(params->alpha); |
| 120 | |
| 121 | // Allocate buffers |
| 122 | bufX = base->createEnqueueBuffer(X, (lengthX + params->offBX)* sizeof(T), 0, CL_MEM_READ_ONLY); |
| 123 | bufY = base->createEnqueueBuffer(Y, (lengthY + params->offCY)* sizeof(T), 0, CL_MEM_READ_WRITE); |
nothing calls this directly
no test coverage detected