42 __global__ void set_array(float *a, float value, size_t len)  in set_array()
44   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in set_array()
51 __global__ void set_array_double(double *a, double value, size_t len)  in set_array_double()
53   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in set_array_double()
60 __global__ void STREAM_Copy(float *a, float *b, size_t len)  in STREAM_Copy()
62   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Copy()
69 __global__ void STREAM_Copy_double(double *a, double *b, size_t len)  in STREAM_Copy_double()
71   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Copy_double()
78 __global__ void STREAM_Copy_Optimized(float *a, float *b, size_t len)  in STREAM_Copy_Optimized()
85   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Copy_Optimized()
89 __global__ void STREAM_Copy_Optimized_double(double *a, double *b, size_t len)  in STREAM_Copy_Optimized_double()
96   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Copy_Optimized_double()
100 __global__ void STREAM_Scale(float *a, float *b, float scale, size_t len)  in STREAM_Scale()
102   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Scale()
109 __global__ void STREAM_Scale_double(double *a, double *b, double scale, size_t len)  in STREAM_Scale_double()
111   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Scale_double()
118 __global__ void STREAM_Scale_Optimized(float *a, float *b, float scale, size_t len)  in STREAM_Scale_Optimized()
125   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Scale_Optimized()
129 __global__ void STREAM_Scale_Optimized_double(double *a, double *b, double scale, size_t len)  in STREAM_Scale_Optimized_double()
136   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Scale_Optimized_double()
140 __global__ void STREAM_Add(float *a, float *b, float *c, size_t len)  in STREAM_Add()
142   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Add()
149 __global__ void STREAM_Add_double(double *a, double *b, double *c, size_t len)  in STREAM_Add_double()
151   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Add_double()
158 __global__ void STREAM_Add_Optimized(float *a, float *b, float *c, size_t len)  in STREAM_Add_Optimized()
165   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Add_Optimized()
169 __global__ void STREAM_Add_Optimized_double(double *a, double *b, double *c, size_t len)  in STREAM_Add_Optimized_double()
176   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Add_Optimized_double()
180 __global__ void STREAM_Triad(float *a, float *b, float *c, float scalar, size_t len)  in STREAM_Triad()
182   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Triad()
189 __global__ void STREAM_Triad_double(double *a, double *b, double *c, double scalar, size_t len)  in STREAM_Triad_double()
191   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Triad_double()
198 __global__ void STREAM_Triad_Optimized(float *a, float *b, float *c, float scalar, size_t len)  in STREAM_Triad_Optimized()
205   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Triad_Optimized()
209 …al__ void STREAM_Triad_Optimized_double(double *a, double *b, double *c, double scalar, size_t len)  in STREAM_Triad_Optimized_double()
216   size_t idx = threadIdx.x + blockIdx.x * blockDim.x;  in STREAM_Triad_Optimized_double()
221 bool STREAM_Copy_verify(float *a, float *b, size_t len)  in STREAM_Copy_verify()
223   size_t idx;  in STREAM_Copy_verify()
237 bool STREAM_Copy_verify_double(double *a, double *b, size_t len)  in STREAM_Copy_verify_double()
239   size_t idx;  in STREAM_Copy_verify_double()
253 bool STREAM_Scale_verify(float *a, float *b, float scale, size_t len)  in STREAM_Scale_verify()
255   size_t idx;  in STREAM_Scale_verify()
269 bool STREAM_Scale_verify_double(double *a, double *b, double scale, size_t len)  in STREAM_Scale_verify_double()
271   size_t idx;  in STREAM_Scale_verify_double()
285 bool STREAM_Add_verify(float *a, float *b, float *c, size_t len)  in STREAM_Add_verify()
287   size_t idx;  in STREAM_Add_verify()
301 bool STREAM_Add_verify_double(double *a, double *b, double *c, size_t len)  in STREAM_Add_verify_double()
303   size_t idx;  in STREAM_Add_verify_double()
317 bool STREAM_Triad_verify(float *a, float *b, float *c, float scalar, size_t len)  in STREAM_Triad_verify()
319   size_t idx;  in STREAM_Triad_verify()
333 bool STREAM_Triad_verify_double(double *a, double *b, double *c, double scalar, size_t len)  in STREAM_Triad_verify_double()
335   size_t idx;  in STREAM_Triad_verify_double()
353 PetscErrorCode printResultsReadable(float times[][NTIMES], size_t);
855 PetscErrorCode printResultsReadable(float times[][NTIMES], const size_t bsize)  in printResultsReadable()