#include <string.h>
#include <stdlib.h>
#include <stdio.h>
#include <hip/hip_runtime.h>
#include <device/device_config.h>
#include <device/hip/check.h>
#include "cdtp_kernel.h"
#include <common/neko_log.h>
Go to the source code of this file.
|
template<const int > |
int | tune_cdtp (void *dtx, void *x, void *dr, void *ds, void *dt, void *dxt, void *dyt, void *dzt, void *w3, int *nel, int *lx) |
|
void | hip_cdtp (void *dtx, void *x, void *dr, void *ds, void *dt, void *dxt, void *dyt, void *dzt, void *w3, int *nel, int *lx) |
|
◆ CASE
Value: case LX: \
if(autotune[LX] == 0 ) { \
autotune[LX]=tune_cdtp<LX>(dtx,
x, \
} else if (autotune[LX] == 1 ) { \
CASE_1D(LX); \
} else if (autotune[LX] == 2 ) { \
CASE_KSTEP(LX); \
} \
break
__global__ void const T *__restrict__ const T *__restrict__ dr
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ ds
__global__ void const T *__restrict__ x
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dt
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ w3
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dzt
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dyt
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dxt
◆ CASE_1D
Value: hipLaunchKernelGGL( HIP_KERNEL_NAME(cdtp_kernel_1d<real, LX, 1024> ), \
HIP_CHECK(hipGetLastError());
◆ CASE_KSTEP
Value: hipLaunchKernelGGL( HIP_KERNEL_NAME(cdtp_kernel_kstep<real, LX> ), \
HIP_CHECK(hipGetLastError());
◆ CASE_LARGE
Value: case LX: \
CASE_KSTEP(LX); \
break
◆ hip_cdtp()
void hip_cdtp |
( |
void * |
dtx, |
|
|
void * |
x, |
|
|
void * |
dr, |
|
|
void * |
ds, |
|
|
void * |
dt, |
|
|
void * |
dxt, |
|
|
void * |
dyt, |
|
|
void * |
dzt, |
|
|
void * |
w3, |
|
|
int * |
nel, |
|
|
int * |
lx |
|
) |
| |
Fortran wrapper for device hip \( D^T X \)
Definition at line 58 of file opr_cdtp.hip.
◆ tune_cdtp()
template<const int >
int tune_cdtp |
( |
void * |
dtx, |
|
|
void * |
x, |
|
|
void * |
dr, |
|
|
void * |
ds, |
|
|
void * |
dt, |
|
|
void * |
dxt, |
|
|
void * |
dyt, |
|
|
void * |
dzt, |
|
|
void * |
w3, |
|
|
int * |
nel, |
|
|
int * |
lx |
|
) |
| |