case LX: \
{ \
STR(cdtp_kernel_lx##LX), &err); \
CL_CHECK(clSetKernelArg(kernel, 0,
sizeof(cl_mem), (
void *) &dtx)); \
CL_CHECK(clSetKernelArg(kernel, 1,
sizeof(cl_mem), (
void *) &
x)); \
CL_CHECK(clSetKernelArg(kernel, 2,
sizeof(cl_mem), (
void *) &
dr)); \
CL_CHECK(clSetKernelArg(kernel, 3,
sizeof(cl_mem), (
void *) &
ds)); \
CL_CHECK(clSetKernelArg(kernel, 4,
sizeof(cl_mem), (
void *) &
dt)); \
CL_CHECK(clSetKernelArg(kernel, 5,
sizeof(cl_mem), (
void *) &
dxt)); \
CL_CHECK(clSetKernelArg(kernel, 6,
sizeof(cl_mem), (
void *) &
dyt)); \
CL_CHECK(clSetKernelArg(kernel, 7,
sizeof(cl_mem), (
void *) &
dzt)); \
CL_CHECK(clSetKernelArg(kernel, 8,
sizeof(cl_mem), (
void *) &
w3));
\
kernel, 1, NULL, &global_item_size, \
&local_item_size, 0, NULL, NULL)); \
} \
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