|
@@ -26,128 +26,17 @@
|
|
* Number of flops of Gemm
|
|
* Number of flops of Gemm
|
|
*/
|
|
*/
|
|
|
|
|
|
-//#define USE_PERTURBATION 1
|
|
|
|
-
|
|
|
|
-
|
|
|
|
-#ifdef USE_PERTURBATION
|
|
|
|
-#define PERTURBATE(a) ((starpu_drand48()*2.0f*(AMPL) + 1.0f - (AMPL))*(a))
|
|
|
|
-#else
|
|
|
|
-#define PERTURBATE(a) (a)
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
-static double cpu_chol_task_11_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/1000.0f*0.894/0.79176);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cpu_chol_task_11_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
-static double cuda_chol_task_11_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/50.0f/10.75/5.088633/0.9883);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cuda_chol_task_11_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
-static double cpu_chol_task_21_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/7706.674/0.95/0.9965);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cpu_chol_task_21_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
-static double cuda_chol_task_21_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/50.0f/10.75/87.29520);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cuda_chol_task_21_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
-static double cpu_chol_task_22_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/50.0f/10.75/8.0760);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cpu_chol_task_22_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
-static double cuda_chol_task_22_cost(starpu_buffer_descr *descr)
|
|
|
|
-{
|
|
|
|
- uint32_t n;
|
|
|
|
-
|
|
|
|
- n = starpu_matrix_get_nx(descr[0].handle);
|
|
|
|
-
|
|
|
|
- double cost = (((double)(n)*n*n)/50.0f/10.75/76.30666);
|
|
|
|
-
|
|
|
|
-#ifdef STARPU_MODEL_DEBUG
|
|
|
|
- printf("cuda_chol_task_22_cost n %d cost %e\n", n, cost);
|
|
|
|
-#endif
|
|
|
|
-
|
|
|
|
- return PERTURBATE(cost);
|
|
|
|
-}
|
|
|
|
-
|
|
|
|
struct starpu_perfmodel_t chol_model_11 = {
|
|
struct starpu_perfmodel_t chol_model_11 = {
|
|
- .per_arch = {
|
|
|
|
- [STARPU_CPU_DEFAULT] = { .cost_model = cpu_chol_task_11_cost },
|
|
|
|
- [STARPU_CUDA_DEFAULT] = { .cost_model = cuda_chol_task_11_cost }
|
|
|
|
- },
|
|
|
|
.type = STARPU_HISTORY_BASED,
|
|
.type = STARPU_HISTORY_BASED,
|
|
.symbol = "chol_model_11"
|
|
.symbol = "chol_model_11"
|
|
};
|
|
};
|
|
|
|
|
|
struct starpu_perfmodel_t chol_model_21 = {
|
|
struct starpu_perfmodel_t chol_model_21 = {
|
|
- .per_arch = {
|
|
|
|
- [STARPU_CPU_DEFAULT] = { .cost_model = cpu_chol_task_21_cost },
|
|
|
|
- [STARPU_CUDA_DEFAULT] = { .cost_model = cuda_chol_task_21_cost }
|
|
|
|
- },
|
|
|
|
.type = STARPU_HISTORY_BASED,
|
|
.type = STARPU_HISTORY_BASED,
|
|
.symbol = "chol_model_21"
|
|
.symbol = "chol_model_21"
|
|
};
|
|
};
|
|
|
|
|
|
struct starpu_perfmodel_t chol_model_22 = {
|
|
struct starpu_perfmodel_t chol_model_22 = {
|
|
- .per_arch = {
|
|
|
|
- [STARPU_CPU_DEFAULT] = { .cost_model = cpu_chol_task_22_cost },
|
|
|
|
- [STARPU_CUDA_DEFAULT] = { .cost_model = cuda_chol_task_22_cost }
|
|
|
|
- },
|
|
|
|
.type = STARPU_HISTORY_BASED,
|
|
.type = STARPU_HISTORY_BASED,
|
|
.symbol = "chol_model_22"
|
|
.symbol = "chol_model_22"
|
|
};
|
|
};
|